The central innovation is not a single machine-learning model but an “intelligent service hub” that coordinates several task-specific algorithms. Load forecasting used a long short-term memory network with two recurrent layers, while a deep neural network handled binary fault classification and a random forest estimated whether equipment required maintenance. A convolutional neural network extracted visual features, natural-language processing transformed maintenance records into numerical representations, and a weighted late-fusion module combined the results. The authors assigned 50% of the fusion weight to sensor and time-series features, 25% to text and 25% to images or video, reflecting the continuous availability of electrical measurements and the contextual value of reports and visual inspections.

Information moves through the framework in a closed loop. Smart meters, Internet of Things sensors, cameras, drones, operational logs and inspection records are collected, cleaned and synchronized before being sent to specialized processing modules. The service hub then combines those inputs with a knowledge base containing grid topology, operating rules, maintenance records, constraints and safety limits. Its outputs can include load forecasts, fault classifications, equipment-failure probabilities, maintenance schedules, power-rerouting recommendations, renewable-energy dispatch, demand-response adjustments and cybersecurity alerts. Results from the resulting actions are fed back into the system so that predictions and control decisions can be revised as operating conditions change.

The experimental dataset contained 20,000 time-ordered samples assembled from public smart-grid resources. It included electrical measurements, environmental variables, renewable-generation information, maintenance records and fault labels, with text and visual inputs included for equipment monitoring. Because the measurements were time dependent, the researchers used the earliest 80% for training, the next 10% for validation and hyperparameter tuning, and the final 10% as an independent test set. They also conducted walk-forward validation with five sequential folds to assess performance under a more operationally realistic sequence of training on earlier observations and testing on later ones. Ten independent runs were reported with means, standard deviations and 95% confidence intervals.

For fault detection, the deep neural network produced probability scores for faulty and non-faulty states, which were assessed across decision thresholds using a receiver operating characteristic curve. The resulting ROC-AUC was 0.970 ± 0.012, with a 95% confidence interval from 0.961 to 0.979. The predictive-maintenance random forest combined historical maintenance data with equipment age, temperature, vibration, loading conditions and previous faults. On the 2,000-sample test set, it correctly classified 1,348 normal cases and 562 maintenance-required cases, while misclassifying 52 normal cases and missing 38 maintenance-required cases. Its precision for the maintenance-required class was 91.53%, recall was 93.67% and the F1-score was 92.59%.

The same predictive system attained 95.5% overall accuracy and 94.98% balanced accuracy, with a Matthews correlation coefficient of 0.894. The authors also evaluated a Weibull reliability model to represent failures associated with early life, random events or aging, using available component-failure histories where possible. Its initial scale and shape parameters were based on engineering assumptions when records were unavailable, and the researchers said they examined parameter sensitivity. In the operational configuration, components with a predicted failure probability above a validation-selected threshold of 0.60 were marked as requiring maintenance.

p>The load-forecasting module used 24-step sliding windows to predict the next demand or equipment-risk state. Across the reported repeated experiments, it achieved a MAPE of 4.8% ± 0.5%, with a 95% confidence interval of 4.4% to 5.2%. In an example daily profile, demand ranged from about 350 MW at night to 450 MW at midday. In another comparison, consumption ranged from 280 to 320 MW while generation ranged from 300 to 340 MW, illustrating the mismatch the optimization system was intended to manage by adjusting conventional generation, renewable use, storage and flexible demand. The paper describes this as a benchmark evaluation and does not report deployment in an operating utility.

For energy optimization, the framework worked over a 24-step horizon, maintained a 10% reserve margin and allowed demand-response adjustments of up to 15% of non-critical controllable load. The experimental baseline included conventional generation scheduling and reactive adjustment, while the optimized condition jointly coordinated load forecasts, renewable-generation predictions, storage, power allocation and demand response. Under identical test-horizon conditions, total modeled operating cost fell from $12,000 to $8,400. The largest reported saving was in conventional generation, which declined from $9,200 to $6,450 as renewable generation was prioritized and dispatch was adjusted; smaller reductions appeared in grid losses, renewable curtailment, peak demand and storage-control costs.

An ablation analysis indicated that the unified design mattered more than any isolated component. Removing sensor fusion produced the largest degradation among the component-removal tests, lowering fault-detection AUC from 0.970 to 0.931, predictive-maintenance accuracy from 95.5% to 91.6%, load-forecasting MAPE from 4.8% to 6.7%, and cost reduction from 30.0% to 23.5%. Removing energy optimization lowered the modeled savings to 16.2%, while a sensor-only model reached an AUC of 0.926 and 22.7% cost reduction. A broader ablation table reported an even larger cost reduction of 18.6% without multimodal fusion and 21.8% without the service hub, although the paper does not reconcile every numerical difference among its tables and figures.

Human oversight is built into the proposed control hierarchy. Routine, lower-risk actions—such as load balancing, energy allocation and maintenance scheduling—can be automated, whereas severe faults, major voltage deviations or cybersecurity alerts are routed to an operator dashboard for approval. The cybersecurity design contemplates false-data injection, replay attacks, denial of service, unauthorized commands and poisoned training data, using checks such as timestamp verification, role-based access, command authentication, anomaly filtering and cross-modal consistency. These safeguards remain architectural proposals within the evaluated framework. The study itself identifies noisy or incomplete sensor data, renewable intermittency, model generalization to larger grids and extreme conditions, cybersecurity exposure, implementation expense and regulatory standardization as unresolved limitations. Large-scale field testing with live streaming data will therefore be required before the reported benchmark gains can be treated as demonstrated utility-scale benefits.

Subject of Research: Evaluation of an autonomous, operator-controllable AI framework for multimodal smart-grid management

Article Title: An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing

Article References: Zeng, F., Zheng, H., Li, C., Rong, J., & Nong, Y. (2026). An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing. Frontiers in Energy Research, 14, 1910556. https://doi.org/10.3389/fenrg.2026.1910556

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1910556

Keywords: smart grids, artificial intelligence, multimodal data, fault detection, predictive maintenance, load forecasting, energy optimization, renewable energy, demand response, grid cybersecurity, artificial, intelligence

The study, published in Frontiers in Artificial Intelligence by Parth Yogeshbhai Purohit, Thomas Feuerle, and Peter Hecker, tackles a problem that has long frustrated airline operations: the information needed to prepare for an inbound flight often exists before the aircraft lands, but it lives in disconnected systems and separate analytical views. Maintenance risk indicators, catering trolley availability, gate congestion forecasts, and passenger-assistance requests each arrive through their own channels. The team’s contribution is not a new prediction algorithm but an operational integration — a modular architecture that converts these heterogeneous inputs into readiness scores, alert levels, and prioritized action queues that an OCC operator can review and act upon before the aircraft ever reaches the stand.

At the heart of the system sits a deliberately transparent, rule-based fusion layer. Rather than training a machine-learning model to combine signals — an approach the authors rejected because no labeled dataset of real OCC decisions exists — the researchers designed an explicit penalty scheme. Readiness is computed on a 0-to-100 scale, starting from a perfect score and subtracting non-negative penalties for maintenance urgency, predicted failure probability, inventory shortage risk, the most severe predicted process bottleneck, gate congestion, arrival delay, and unconfirmed assistance requests. Crucially, the scheme is non-compensatory in the protective direction: a benign domain can never offset a severe one, so an aircraft with critical maintenance risk keeps its full penalty even if catering supplies are perfectly stocked. When multiple domains are simultaneously severe, an additional compounding term escalates the penalty super-additively, reflecting the intuition that stacked stressors are worse than the sum of their parts.

Every penalty is retained and displayed individually in a readiness breakdown, so an operator can always see which domain drove a score. Alert levels escalate if either the readiness channel or a separate criticality channel crosses a threshold — a doubly conservative design. The output is an action queue that names specific teams and priorities: maintenance dispatch with spare-part reservation for urgent cases, inventory top-ups, bottleneck-specific coordination, assistance pre-positioning, gate buffer allocation, and a standing operator-review item. The authors chose this inspectable rule set precisely because its coefficients are explicit heuristics that can be challenged, re-elicited, or eventually replaced by a learned layer once genuine operational decision data become available.

To evaluate the architecture, the team ran a deterministic sweep of 1,000 synthetic inbound cases across six operational regimes: low-risk routine, maintenance-heavy, inventory-heavy, congestion-and-delay-heavy, assistance-readiness, and composite multi-risk. A fixed random seed makes every input and output reproducible. The sweep produced a mean readiness of 47.4 and a median of 50.6, with 370 cases flagged Critical, 284 Priority, and 188 Routine. The regimes behaved as designed: routine cases averaged a readiness of 82.3, maintenance-heavy cases dropped to 26.2, and composite multi-risk cases plunged to 8.8 — a clean differentiation confirmed by bootstrap confidence intervals and a multi-seed robustness check showing regime means stable to within half a point.

The most revealing results came from ablation experiments, in which one signal family at a time was neutralized. Removing maintenance influence raised mean readiness by 13.7 points; removing bottleneck influence raised it by 10.0 points; inventory removal added 6.7 points, and assistance just 1.4. The ordering was stable across resampling. In plain terms, maintenance and bottleneck evidence dominate how the fused system prioritizes, while inventory and assistance act as smaller modifiers. For OCC practice, the authors suggest this means pre-arrival attention and data-quality effort are most leveraged on maintenance-risk and process-bottleneck signals — though they are careful to frame this as a hypothesis generated under synthetic conditions, not a validated operational finding.

Notably, the team reports its own system’s imperfections. Under a strict zero-tolerance monotonicity criterion, six of eight sanity checks passed, but two maintenance-related checks failed: readiness rose by at most 0.4 to 0.5 points at a single grid point in the flat region before the maintenance urgency class crossed from OK to PLAN, after which readiness fell correctly by more than 16 points. Rather than relaxing the threshold to mask the deviations, the authors report them openly, attributing them to the threshold-driven interaction between trained modules and the fusion layer. They discuss — and ultimately decline to impose — hard monotonic guards, noting that strict monotonicity would forbid legitimate non-monotone interactions, such as a moderate stressor triggering an earlier, better-resourced intervention that genuinely leaves an aircraft better prepared.

The paper is equally candid about the limits of its evidence. No real airline, airport, or station-operation dataset was used; the modules are trained on synthetic labels; and no OCC user study, deployment, certification, or measured operational savings is claimed. Every frequency in the results — the 370 Critical alerts, the mean readiness of 47.4 — is an artifact of the deliberately stressed test population, not an operational rate. The authors even warn that synthetic evaluation can quietly validate the assumptions it was built on, meaning the ablation ordering quantifies signal influence within the fusion rather than validated importance in practice.

Perhaps the most thought-provoking sections reach beyond engineering into safety science. Drawing on complexity-informed accident theory, the authors argue that a decision-support layer is an intervention into a socio-technical system, not merely a technical artifact. They identify foreseeable risks: attention and resources could be displaced toward scored maintenance risk, creating vulnerabilities in weakly scored domains like passenger handling; threshold boundaries could acquire procedural force, with an aircraft one readiness point above a cutoff treated categorically differently; and because every aircraft is scored by the same rule set, a mis-specified coefficient could move an entire fleet’s advisories in the same direction at once — a common-mode failure the system itself could introduce.

The researchers also flag the danger of the illusion of control: a single displayed readiness number invites interpretation as a measurement rather than an aggregation of assumptions, and a system that quantifies maintenance and congestion but not crew fatigue or informal coordination will bias attention toward what it can count. Their proposed countermeasures are organizational — training that foregrounds uncertainty, interfaces that preserve provenance and breakdowns, and auditing of override rates. The current interface, they note, lacks the operator-feedback channel that would enable learning from disagreements, which they identify as the most important gap. The decisive next steps, the authors conclude, are not more synthetic analysis but confrontation with real turnaround records and an OCC user study — the tests that will determine whether differentiated readiness scores correspond to operationally meaningful distinctions in the chaotic, time-compressed world of live airline coordination.

Subject of Research: Human-in-the-loop decision-support architecture for pre-arrival aircraft turnaround coordination in airline Operations Control Centers

Article Title: Pre-arrival OCC decision support for aircraft turnaround coordination: a human-in-the-loop architecture and scenario-based evaluation

Article References: Purohit, P. Y., Feuerle, T., & Hecker, P. (2026). Pre-arrival OCC decision support for aircraft turnaround coordination: a human-in-the-loop architecture and scenario-based evaluation. Frontiers in Artificial Intelligence, 9, Article 1977034. https://doi.org/10.3389/frai.2026.1977034

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1977034

Keywords: aircraft turnaround, operations control center, decision support, human-in-the-loop, predictive maintenance, bottleneck forecasting, inventory readiness, passenger assistance, synthetic scenario evaluation, aviation machine learning, rule-based fusion, airline operations

The study, authored by Mohammad Waseem of Jamia Millia Islamia in New Delhi, takes an unusually broad view of the field. Rather than focusing on a single function, it critically assesses 120 research articles and review papers to build a system-level picture of how artificial intelligence and machine learning are transforming battery management across state estimation, cell balancing, thermal management, fault diagnosis, predictive maintenance, intelligent charging, edge computing, and cloud connectivity. According to the review, this integrated perspective is what previous surveys have lacked, and it reveals both how far the technology has come and where its most stubborn bottlenecks remain.

At the heart of every battery management system, or BMS, lies a deceptively simple question: what is the battery actually doing right now? Two quantities dominate the answer. The state of charge, or SoC, describes how much energy remains in the pack relative to its nominal capacity, while the state of health, or SoH, captures how much the battery has degraded over its lifetime. Neither can be measured directly. Both must be inferred from voltage, current, and temperature readings using computational algorithms, a task made harder by the fact that lithium-ion cells behave in strongly nonlinear ways that shift with temperature, aging, and driving patterns.

Conventional BMS architectures have long relied on model-based techniques such as coulomb counting, equivalent circuit models, and Kalman filtering to estimate these hidden states. These methods have served the industry well, but the review highlights a fundamental weakness: they struggle to represent the nonlinear electrochemical behavior of lithium-ion batteries under varying operating conditions. As battery packs in modern electric vehicles grow larger and more complex, the amount of real-time sensor data they generate has outstripped what traditional models can meaningfully process, and subtle degradation patterns slip through the cracks.

This is where machine learning enters the picture. Artificial neural networks have been widely applied to state of charge estimation because of their ability to capture complex nonlinear battery dynamics, and experimental studies cited in the review show that ANN-based models can achieve significantly improved accuracy compared with conventional estimation techniques, particularly under dynamic operating conditions. Support vector machines and ensemble learning methods have proven effective at detecting battery faults and abnormal operating conditions at early stages, flagging problems before they escalate into safety hazards. The trade-off is real, however: these models demand large training datasets and substantial computational resources, which can complicate deployment on the embedded processors that actually live inside a vehicle.

Deep learning has pushed the frontier further. Long short-term memory networks, convolutional neural networks, and deep neural networks have been extensively explored for predicting battery degradation and remaining useful life, the parameter known as RUL. LSTM networks are particularly well suited to this task because they excel at analyzing time-series data and capturing the temporal dependencies that characterize how batteries age. The review notes that studies published in leading energy journals have demonstrated that deep learning models can achieve high prediction accuracy and improved robustness compared with traditional analytical models, though the price is longer training times and heavier computational demands.

Beyond estimation, artificial intelligence is reshaping how batteries are charged and protected. Reinforcement learning algorithms learn optimal charging policies through interaction with the battery system and its environment, dynamically adjusting charging rates based on real-time inputs such as state of charge, state of health, temperature, and grid conditions. The goal is to balance charging speed against battery longevity while avoiding degradation mechanisms like lithium plating and thermal stress. The review reports that reinforcement learning approaches have shown improved energy efficiency and reduced battery aging compared with traditional charging methods, although they require extensive training cycles and carefully designed reward functions to perform stably.

Thermal management represents another critical frontier. Because lithium-ion battery performance and safety are tightly coupled to temperature, improper thermal regulation can accelerate degradation and, in the worst case, trigger thermal runaway, a failure mode characterized by uncontrolled heat generation and potential fire. Smart BMS designs now incorporate per-cell or per-module temperature sensing using thermistors and integrated sensing devices, feeding real-time thermal data into control units that dynamically activate cooling or heating mechanisms and adjust charge and discharge limits. Convolutional neural networks and autoencoders have proven especially effective at detecting thermal anomalies and hidden fault patterns in sensor data, giving the BMS an early warning system that rule-based logic alone cannot match.

The review also spotlights two emerging hardware and infrastructure trends that could define the next generation of battery management. The first is FPGA-based edge implementation, in which field-programmable gate arrays and embedded processors execute machine learning models with the low latency required for safety-critical real-time control. Recent work on FPGA-based intelligent battery management units demonstrates dynamic power distribution techniques, though the review cautions that these platforms bring higher implementation cost and design complexity. The second is IoT-enabled cloud integration, in which distributed sensors transmit battery data to centralized platforms where machine learning models perform fleet-level optimization, predictive diagnostics, and over-the-air updates. Prototype systems integrating these technologies have demonstrated improved reliability and predictive maintenance capabilities, but they introduce new concerns around cybersecurity, communication latency, and data management at scale.

Perhaps the review’s most consequential conclusion is that no single artificial intelligence technique is universally optimal. Artificial neural networks and LSTM models deliver high estimation accuracy for state of charge and state of health, convolutional architectures excel at fault diagnosis, and support vector machines and random forests offer robust performance with lower computational requirements that suit embedded systems. The most promising pathway forward, the author argues, lies in hybrid approaches that combine model-based methods with data-driven algorithms, leveraging the physical understanding embedded in equivalent circuit models and Kalman filters alongside the pattern-recognition power of machine learning. Future directions identified in the review include physics-informed machine learning, digital twins, edge intelligence, and cloud-connected battery ecosystems, all of which point toward battery management systems that are not merely protective but genuinely predictive and autonomous.

Subject of Research: AI and machine learning applications in smart battery management systems for electric vehicles

Article Title: AI/ML enabled smart battery management systems in electric vehicles

Article References: Waseem, M. (2026). AI/ML enabled smart battery management systems in electric vehicles. Frontiers in Batteries and Electrochemistry, 5, Article 1883201. https://doi.org/10.3389/fbael.2026.1883201

Image Credits: AI Generated

DOI: 10.3389/fbael.2026.1883201

Keywords: battery management system, electric vehicles, lithium-ion batteries, machine learning, artificial intelligence, state of charge, state of health, deep learning, thermal management, predictive maintenance, FPGA, IoT

The research team, led by Kesheng Wang of Shandong Huayu University of Technology together with colleagues at the North China Institute of Aerospace Engineering and Space Engineering University, built GSDD-Net on top of the YOLO family of one-stage object detectors, which predict object classes and bounding boxes directly from feature maps rather than working through a slower region-proposal stage. That choice matters for factories, where inspection systems must keep pace with production lines in real time. But off-the-shelf detectors struggle with gear defects, which vary enormously in scale and often occupy only a handful of pixels in an image. The team’s answer was not one big idea but four targeted architectural changes, each aimed at a specific failure mode of existing models.

The first change targets a deceptively mundane step: downsampling. Neural networks shrink images as they process them, and conventional strided convolutions throw away fine detail along the way, exactly the edge contours and texture cues that distinguish a faint scratch from ordinary machining marks. GSDD-Net replaces standard downsampling with a module called ADown, which splits the input features into two branches, sends one through convolution and the other through max pooling, and concatenates the results. By combining average-style global context with locally salient responses, ADown preserves weakly salient defect information while cutting computational cost. In the team’s ablation tests on the GID dataset, ADown alone reduced the parameter count from 2.6 million to 2.1 million while still nudging accuracy upward.

The second component, CSP-EDLAN, reworks the feature extraction and fusion stages in both the backbone and the neck of the network. It folds together three ideas: cross-stage partial connections that give gradients an additional route to flow during training, multi-gradient path aggregation that merges features from different branches, and DualConv units that pair a grouped spatial convolution with a pointwise convolution so channels can exchange information cheaply. The result is richer multi-scale feature representations for defects of diverse sizes and shapes, with less redundant computation. Among all four modules tested individually, CSP-EDLAN delivered the largest accuracy gain, lifting the key mAP50 metric by 1.5 percentage points over the YOLO11n baseline.

The third innovation confronts the messy optics of real factory floors. Metallic reflections, grease, and non-uniform illumination generate spurious responses that can swamp the weak signal of a small defect. To handle this, the researchers swapped the attention block in the baseline network for a new module called C2TSSA, built on token statistical self-attention. Instead of computing attention by comparing every pair of image tokens, an operation whose cost grows quadratically, C2TSSA derives attention weights from token-level statistics such as the mean and covariance of the features, capturing global context with only linear computational growth. The module strengthens responses consistent with the surrounding gear structure while suppressing isolated responses caused by reflections and background textures, making subtle defects stand out against their surroundings.

The fourth component, LQEHead, tackles a mismatch that plagues many modern detectors: the classification branch and the box-regression branch work separately, so a high confidence score does not always mean the predicted box actually sits on the defect. This matters acutely for small flaws, where even a slight box displacement sharply reduces overlap with the true region. LQEHead extracts stability statistics from the predicted bounding-box distributions and converts them into a signed correction that is added to the classification logit before the final activation. Predictions with reliable localization get a confidence boost; unstable ones are dialed down. The design preserves weak but well-placed detections while filtering out spurious responses from background edges and glare.

Put together, these four modules produced striking results on two benchmarks. On the GID dataset, a set of 2,000 industrial gear images covering chipping and two types of black-skin defects, GSDD-Net achieved an mAP50 of 73.0 percent and an mAP50-95 of 38.0 percent, beating the YOLO11n baseline by 3.1 and 2.1 percentage points respectively. Meanwhile the model shrank: parameters fell from 2.6 million to 1.9 million, computational cost dropped from 6.3 to 5.9 GFLOPs, and inference speed climbed from 93.7 to 105.7 frames per second. On a second, publicly available gear defect dataset featuring break, missing-tooth, and scratch defects, GSDD-Net reached 97.8 percent mAP50 and 69.6 percent mAP50-95, again improving on the baseline while using fewer parameters.

Comparisons against a broad field of contemporary detectors underscore the balance the team struck. Larger and heavier models such as RT-DETR-l, which carries 32 million parameters and needs more than 100 GFLOPs, failed to match GSDD-Net’s accuracy while running far slower at 38.8 frames per second. Newer YOLO variants including YOLOv12n, YOLOv13n, and YOLO26n each made different trade-offs between speed and accuracy but none matched GSDD-Net’s combined mAP scores on the GID dataset at comparable cost. Class-by-class analysis showed consistent gains: the hardest category, chipping, improved by 4.3 percentage points in AP50, and the stricter AP50-95 metric rose for every defect class, indicating better localization, not just better recognition. Grad-CAM heatmaps provided visual confirmation, showing GSDD-Net concentrating its attention squarely on defect regions while the baseline scattered its responses across irrelevant background.

The authors are careful to frame these numbers as results under their evaluated conditions, noting that deep-learning outcomes can fluctuate with random initialization, data augmentation, and GPU nondeterminism, and that all compared models were trained under identical settings to ensure fairness. They also acknowledge limitations: the datasets cover a limited set of defect categories and imaging conditions, and real factories may present harsher variation. Future work, they write, should expand datasets, explore semi-supervised and synthetic data generation to ease the burden of annotation, apply pruning and quantization for deployment on edge hardware, and even integrate the detector with large language models to build an intelligent inspection agent.

Still, the broader significance is clear. Quality control in intelligent manufacturing has been squeezed between two bad options: accurate but slow and power-hungry models, or fast but less reliable ones. By showing that task-oriented design, a pooling scheme that preserves detail, an aggregation module that enriches features, a statistical attention mechanism that reads the whole scene cheaply, and a detection head that knows how sure it should be, GSDD-Net demonstrates that precision and efficiency need not be enemies. For production lines where a missed scratch on a gear tooth today can mean a seized gearbox tomorrow, that balance could translate directly into fewer failures, less downtime, and safer machines.

Subject of Research: Lightweight deep learning framework for detecting gear surface defects in industrial images

Article Title: GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments

Article References: Wang, K., Zhang, H., Yang, T., Shen, P., & Zhao, Y. (2026). GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments. Frontiers in Mechanical Engineering, 12, Article 1983450. https://doi.org/10.3389/fmech.2026.1983450

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1983450

Keywords: GSDD-Net, gear defect detection, YOLO, deep learning, machine vision, industrial inspection, object detection, self-attention, lightweight models, predictive maintenance, feature extraction, surface defects

A team at Cranfield University’s Integrated Vehicle Health Management Centre, led by Chengwei Wang together with Ip-Shing Fan and Stephen P. King, has now demonstrated a way to manufacture that missing experience artificially. In a study published in Frontiers in Aerospace Engineering, they built a digital twin of an aircraft EMA, a physics-based virtual replica capable of simulating thousands of failure scenarios on demand. The framework, implemented in MATLAB Simulink and Simscape, generated a fully labelled dataset of 70,000 simulation runs covering healthy operation, three single fault modes, and three compound fault combinations. The result is a reproducible benchmark that aerospace researchers can use to train and validate health monitoring algorithms without ever degrading a physical actuator.

The digital twin concept traces its lineage to NASA’s Apollo program, where identical vehicles on Earth mirrored the spacecraft in orbit to help troubleshoot problems millions of miles away. Modernised by Michael Grieves in the early 2000s, the digital twin has evolved into a multi-physics simulation continuously informed by data from its physical counterpart. In prognostics and health management, twins serve two roles: as real-time mirrors for operational monitoring, and as simulation platforms for generating data. For new systems like EMAs, the second role is the urgent one. Machine learning diagnostics are hungry for labelled examples of faults, yet rare failure modes are precisely the events that historical records never capture. Simulation-generated data offers a legitimate foundation, especially when the underlying model is grounded in validated physics.

The Cranfield framework is organised as a four-layer architecture. At the top, an input definition layer supplies voltage and load profiles that mimic realistic flight control demands. Beneath it sits the electromechanical system layer, the heart of the model, which represents the DC motor with physical parameters including internal resistance, inductance, back-EMF constant, and torque constant, alongside a screw-nut transmission built from Simscape Multibody elements that capture nonlinear stiffness, backlash gaps, and damping. A fault injection layer then modifies these parameters dynamically during each run, and a data acquisition layer records voltage, current, power, displacement, velocity, and fault flags at 100 Hz, producing 3,000 data points per signal per 30-second cycle.

To anchor the simulation in reality, the team modelled a 30-second flap extension cycle inspired by the inboard flap of an Airbus A350XWB, using aerodynamic calculations to derive realistic mechanical loads of 8, 10, and 12 kilonewtons across four operational phases. The electrical architecture follows the standard 28-volt DC aerospace power bus, and the motor-gear-leadscrew topology aligns with validated EMA designs reported in the literature. Crucially, the model’s behavioural fidelity was checked against experimental testbeds: when mechanical backlash is injected, the simulated motor current surges above 150 amps while rotor speed dips sharply, reproducing the characteristic current-spike and velocity-drop signature that engineers observe in real jamming and binding events.

The fault catalogue covers three representative single modes and their combinations. Voltage disturbances, labelled FM1, involve a 10-second deviation drawn from a uniform distribution between 24.0 and 29.4 volts, causing torque underproduction and delayed positioning. Mechanical backlash, FM2, is simulated with an effect-based approach: rather than modelling the micro-geometry of gear wear, the framework injects a reverse torque of 50, 100, or 150 newton-metres into the drivetrain, capturing the diagnostic signature at a fraction of the computational cost. Resistance change, FM3, models cable degradation by shifting circuit resistance to one of four levels between 0.3 and 1.0 ohms, altering current draw and heating patterns. Compound faults FM4 through FM6 combine two or all three of these perturbations simultaneously, and because they act within the multi-domain dynamic equations, their interactions emerge naturally from the system’s physics rather than from artificial signal mixing.

The resulting signal behaviours reveal why compound faults are so diagnostically treacherous. Under voltage disturbance alone, displacement tracking falls short during high-load phases and velocity rise times stretch noticeably. Backlash produces stair-stepped displacement lags and oscillatory velocity during direction reversals. Resistance increase elevates RMS current and average power throughout the cycle. But in the triple-fault scenario FM6, the current waveform becomes a chaotic superposition: large spikes driven by backlash events ride on a baseline modulated by concurrent voltage and resistance faults, while the controller’s attempts to compensate for one disturbance are repeatedly disrupted by another. The study shows that fault signatures are not static; at high resistance, backlash spikes are suppressed and unstable, while at low resistance they are sharp and pronounced, meaning any diagnostic system must cope with severity-dependent symptoms.

The dataset construction protocol was designed with machine learning hygiene in mind. For each of the seven operating conditions, 10,000 independent simulation runs were executed, giving strict class balance across all categories. Fault onset time, magnitude, and duration were randomised independently in every run, and only a single 15-second analysis window was extracted per cycle, with no overlapping segments, eliminating the risk of data leakage between training and validation subsets. Every sample carries structured metadata, including fault mode, magnitude, injection timing, load profile identifier, and random seed, so the entire dataset can be regenerated deterministically. Twenty statistical and frequency-domain features were then extracted, and Principal Component Analysis confirmed dense, well-separated clusters for the fault classes, with UMAP visualisation reinforcing the same conclusion.

Classification experiments quantified the dataset’s diagnostic value. Using a Random Forest classifier with three physically meaningful variables, mean current, rotor angular velocity, and worm torque output, accuracy approached 100 percent even with Gaussian measurement noise up to 5 percent added to all signals, while a simpler linear discriminant analysis benchmark still exceeded 96 percent. Restricting the model to current alone collapsed accuracy to roughly 83 percent, with systematic confusion between electrically dominated faults and their mechanically coupled compound counterparts. That degradation is itself informative: it demonstrates the dataset is not trivially separable and that diagnostic power genuinely arises from multi-domain electromechanical interactions, validating the physics consistency of the twin itself.

The authors are candid about the framework’s boundaries. Thermal dynamics, including winding heating and thermal-mechanical coupling, are not yet modelled; the fault injections represent effect-level perturbations rather than progressive wear laws; parameter uncertainty and manufacturing tolerances remain unquantified; and the twin currently operates open-loop without assimilating real operational data. Each of these gaps maps onto a planned extension, from physics-informed degradation models enabling remaining useful life estimation to Bayesian updating and Kalman filtering for closed-loop recalibration. Even in its present form, however, the framework fills a pressing industrial need: a safe, scalable, and fully traceable source of fault-labelled actuator data that bridges the gap between expensive physical test rigs and the algorithms that will one day keep electric aircraft flying. As actuation goes all-electric, twins like this one may become the training grounds where the next generation of aircraft health monitoring is born.

Subject of Research: Digital twin-based fault simulation for aircraft electromechanical actuator health monitoring

Article Title: A digital twin-based fault simulation framework for aircraft electromechanical actuators: method and demonstration

Article References: Wang, C., Fan, I.-S., & King, S. P. (2026). A digital twin-based fault simulation framework for aircraft electromechanical actuators: method and demonstration. Frontiers in Aerospace Engineering, 5, Article 1708392. https://doi.org/10.3389/fpace.2026.1708392

Image Credits: AI Generated

DOI: 10.3389/fpace.2026.1708392

Keywords: digital twin, electromechanical actuator, More Electric Aircraft, fault diagnosis, prognostics and health management, synthetic dataset, MATLAB Simulink, Simscape, machine learning, aerospace engineering, fault injection, predictive maintenance

The stakes could hardly be higher. Lithium-ion batteries have become the backbone of modern electrification, powering everything from smartphones and wearables to frequency-regulating grid farms that stabilize intermittent solar and wind output. Global production is racing toward multi-terawatt-hour capacity as nations pursue carbon neutrality. But every cell carries an internal clock. Capacity fade, growth of the solid electrolyte interphase that consumes active lithium, metallic lithium plating during fast charging or cold operation, micro-cracking of electrodes under mechanical stress, rising internal resistance, thermal stress, and cell-to-cell inconsistency all conspire to shorten a battery’s useful life. In the worst case, localized overheating can trigger thermal runaway—a fire or explosion risk that no manufacturer can afford to ignore.

Traditional monitoring methods, the review explains, struggle to keep pace with this complexity. Capacity-based estimation requires full charge-discharge cycles that vehicles rarely complete in the real world. Electrochemical impedance spectroscopy detects aging sensitively but is costly and largely an offline laboratory technique. Equivalent circuit models run fast but sacrifice physical fidelity, while full electrochemical models are accurate yet computationally punishing. Kalman filter observers track states online but falter when the underlying model drifts from reality. Under the non-linear, multifactorial, and time-varying conditions typical of contemporary lithium-ion packs, these classical approaches often miss hidden degradation dynamics entirely—precisely the dynamics that matter most for safety and longevity.

Enter deep learning. Machine learning frameworks can learn intricate relationships between voltage, current, temperature, impedance, and cycle behavior without requiring explicit physical simplifications. For SOC estimation—the measure of a battery’s remaining charge relative to its rated capacity—the review highlights gated recurrent unit (GRU) encoder-decoder networks equipped with dual-stage attention, which map sequences of voltage, current, and temperature readings onto charge states while taming the vanishing-gradient problems that plague conventional recurrent networks. Time-delayed recurrent neural networks add another twist, using delayed hidden-state feedback to capture electrochemical hysteresis and battery memory effects that simpler models gloss over.

The most compelling SOC results come from optimization-augmented hybrids. A genetic algorithm tunes the architecture of a GRU network—searching across one to three layers and 32 to 256 neurons—using mean squared error as its fitness function, yielding faster training and greater robustness under dynamic operating conditions. In comparative tests across standard drive cycles including US06, FUDS, BJDST, HWFET, and UDDS, the standout performers were CNN-BWGRU and Bi-LSTM models, which posted root mean square errors as low as 0.00659 and mean absolute errors down to 0.0017, with coefficients of determination exceeding 0.999. Those numbers translate to charge tracking accurate to fractions of a percent—precision that directly extends driving range confidence and pack safety.

State of health estimation, which tracks how far a battery’s maximum usable capacity has fallen from its factory-fresh value toward the roughly 80 percent end-of-life threshold, has spawned an equally creative arsenal. Physics-informed neural networks embed electrochemical constraints—such as the monotonic relationship between incremental capacity features and SOH—directly into the training loss, forcing the network to learn degradation physics alongside data patterns. Graph convolutional networks take a radically different approach: each charge-discharge cycle becomes a node in a graph, with discharge voltage profile similarities defining the edges, allowing the model to trace how degradation propagates from cycle to cycle.

Transformer architectures, borrowed from the language-modeling revolution, have proven especially potent here. A regression vision transformer slices one-dimensional battery cycling features into patches and processes them in parallel, capturing long-range temporal dependencies that recurrent models miss. The review’s comparative analysis found that attention-enhanced hybrids dominate the SOH leaderboard: an artificial bee colony-optimized bidirectional GRU achieved an R² of 0.9973, while a CNN-Transformer combination posted the lowest mean absolute error at 0.0027 and the lowest mean absolute percentage error at 0.0037. The IPSO-CNN Bi-LSTM model proved the most reliable tracker of non-linear aging, holding its cycle-wise error band to roughly 0 to 0.02 even through sharp post-cycle-25 declines in capacity.

Remaining useful life prediction—the forecast of how many cycles remain before a battery hits end-of-life—is where the review’s findings become most dramatic. Denoising transformer-based neural networks combine a one-dimensional convolutional front-end that strips sensor noise with stacked multi-head self-attention layers that learn charging and discharging trends across extended horizons. But the single best performer was a decomposition-assisted hybrid: ICEEMDAN signal decomposition feeding an optimal deep belief network tuned by a Bayesian algorithm. Trained on NASA and CALCE benchmark cells, it achieved a mean absolute error of 9.7261 × 10⁻⁵ for SOH and an absolute error of zero cycles for RUL—essentially perfect failure-cycle prediction on the test data. A chaotic particle swarm-optimized particle filter separately recorded the lowest MAPE of 0.0056 for end-of-life forecasting.

Perhaps the most forward-looking section of the review concerns joint estimation: single integrated models that predict SOC, SOH, and RUL simultaneously. Because all three metrics derive from the same underlying electrochemical state, shared feature extraction improves consistency and slashes computational overhead—a critical advantage for embedded battery management systems. One FPGA-accelerated LSTM implementation achieved SOC mean absolute error of just 0.016 percent while simultaneously estimating SOH, demonstrating that real-time multi-state monitoring is no longer theoretical. The review also catalogs a joint AR-ECM and hybrid kernel extreme learning machine approach fusing Kalman filtering with particle swarm optimization, which kept SOH error below 1 percent and SOC error below 2 percent on automotive cells.

Challenges remain, and the authors are candid about them. Model explainability is limited, a serious problem in safety-critical battery management where engineers must understand why a network flags a cell as failing. Laboratory-to-field transferability suffers from domain shift, noise in voltage and current measurements degrades robustness, and uncertainty quantification is still immature. The review’s prescription is a research agenda built on explainable AI, hybrid physics-informed learning that marries electrochemical models with data-driven intelligence, transfer learning across chemistries and operating conditions, and edge-deployment optimization for embedded hardware. If that agenda succeeds, the batteries in tomorrow’s vehicles and grids will not merely store energy—they will understand themselves, predicting their own decline with enough precision to schedule maintenance, prevent failures, and squeeze every last cycle from the multi-terawatt-hour fleet the energy transition demands.

Subject of Research: AI and deep learning techniques for estimating lithium-ion battery state of charge, state of health, and remaining useful life

Article Title: AI-driven lithium-ion battery health prediction: a review of SOC, SOH and RUL estimation techniques

Article References: Tamizharasi, G., & Rajini, G. K. (2026). AI-driven lithium-ion battery health prediction: a review of SOC, SOH and RUL estimation techniques. Frontiers in Artificial Intelligence, 9, Article 1858302. https://doi.org/10.3389/frai.2026.1858302

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1858302

Keywords: lithium-ion batteries, artificial intelligence, deep learning, battery management systems, state of charge, state of health, remaining useful life, transformers, electric vehicles, energy storage, physics-informed neural networks, predictive maintenance

The study behind this framework was carried out by researchers at Mälardalen University in Eskilstuna, Sweden, together with a colleague at Research Institutes of Sweden in Kista, and it appears in the journal Frontiers in Manufacturing Technology as part of the journal’s Digital Manufacturing section. The team set out to address a gap they identified directly in the manufacturing literature: while staged process models for AI development in industrial settings already exist, they tend to describe recommended activities at a level of abstraction that leaves practitioners guessing about what to actually do on Monday morning. Other scholarly efforts have produced detailed methods, but those methods typically cover only narrow slices of the development journey, such as the elicitation of requirements, leaving the rest of the process undocumented.

To build something more complete, the researchers began with a systematic review of development frameworks that had already proven useful in adjacent fields. AI applications in clinical medicine, human-AI interaction design, and software engineering have each generated mature bodies of work on how to structure the creation of intelligent systems, and the authors mined these traditions for transferable insights. They then combined and modified the existing frameworks to fit the particular realities of manufacturing, where production engineers, quality specialists, machine operators, data scientists, interface designers, and software developers must all coordinate around systems whose data pipelines, machine-learning models, and user interactions are deeply interdependent.

The result is a framework of 147 task recommendations organised into six conceptual stages of development progression. Each stage functions as what methodologists call a determinant framework: rather than prescribing a rigid sequence, it lays out factors and tasks that have been found to influence development outcomes, giving project leaders a detailed and actionable map they can adapt to their own circumstances. The granularity matters because the framework’s intended users are development project leaders with manufacturing backgrounds—people who may be experts in production engineering but who must nonetheless orchestrate contributions from specialists whose vocabularies and priorities differ sharply from their own.

The choice of project leaders as the primary audience reflects a deliberate theory about where AI implementation succeeds or fails. When front-line employees such as operators and technicians are the ultimate users of an AI system, the person driving the project must understand those employees’ operational workflows, work contexts, and cognitive processes well enough to communicate the system’s benefits, concerns, and limitations in language the front line actually uses. In the researchers’ conceptualisation, an AI solution system contains both a technical subsystem—the AI-enabled software or hardware itself—and a social subsystem encompassing the human actors and workflows that interact with it. Human actors can augment AI components by providing control and feedback on inference results, while AI components can in turn enhance human situation awareness, data analysis, and decision-making. Neither subsystem can be developed in isolation from the other.

Transparency emerges as a second guiding principle. The authors argue that a major barrier to effective collaboration in cross-functional AI teams is the lack of knowledge overlap: each participant can comprehend only the parts of the solution immediately relevant to their own expertise, so nobody anticipates what colleagues are doing or how their own activities ripple outward. By making the entire development process visible and legible to all participants, the framework aims to let a quality engineer understand why a data scientist needs certain labelling conventions, or let an operator see how their feedback loops shape model retraining. Transparency also pays dividends when manufacturing organisations must rely on external parties to acquire or co-develop AI systems, because an internal team that understands the full development process can set realistic, well-scrutinised requirements for vendors and research partners rather than accepting vague promises.

The empirical grounding of the framework came through two complementary routes. First, external reviewers examined the framework and assessed its coherence and practical relevance. Second, the researchers applied the framework in pilot form to two real-world development projects, observing how its task recommendations played out against the messy contingencies of actual industrial work. These pilots were preliminary rather than definitive validations, but they surfaced concrete patterns of utility. In early-phase project planning, the framework prompted deep multi-domain dialogues in which participants from different disciplines confronted questions they might otherwise have deferred until late in the project, when course corrections are far more expensive. Midway through projects, the framework supported reflective planning, giving teams a structured occasion to step back and assess whether their trajectory still matched operational realities.

The pilots also generated candid remarks about limitations. The researchers note that the framework’s use in iterative development environments requires care: modern AI development rarely proceeds in a clean linear sequence, and teams that work in agile or spiral patterns must adapt the six stages to cycles of prototyping, testing, and revision rather than treating them as a one-way march. The authors also confine the post-deployment stage—what happens after a system goes live in production—outside the detailed scope of the model, meaning organisations will need complementary guidance for maintaining, monitoring, and retraining systems once operational.

One of the study’s sobering findings is its characterisation of the current state of practice: the development of human-integrated AI systems in industrial environments remains immature. This immaturity helps explain the AI graveyard phenomenon. When early prototypes demonstrate inspiring potential but applications rarely cross into real operational environments, the fault often lies not in model accuracy but in the unmanaged seams between data collection, curation, inference-based decision-making, error handling, and system maintenance—activities in which shop-floor personnel are deeply embedded. Operators and technicians who collect data and handle errors are not peripheral users; they are components of the solution system itself, and treating them as afterthoughts undermines the socio-technical harmony the technology requires.

The stakes of getting this right are amplified by the environments in which manufacturing AI operates. These systems are frequently intended for high-stakes operational processes where faulty decisions or controls can severely damage production performance. A mis-classified part, an erroneous maintenance prediction, or an ill-timed control adjustment carries consequences that a failed recommendation engine in a consumer application never would. This asymmetry raises the cost of the knowledge gap: a data scientist who does not understand quality engineering constraints, or a production engineer who cannot articulate what the model needs, each introduce risks that compound silently until deployment.

Application areas for AI in manufacturing span production planning, automation, quality inspection, process control, and predictive maintenance, and each domain brings its own patterns of human involvement. In quality inspection, for instance, image-recognition systems may flag defects that human inspectors must verify, creating a division of labour that must be designed rather than assumed. In predictive maintenance, technicians’ tacit knowledge about machine behaviour may be essential for curating the training data, yet that contribution is often invisible in conventional project plans. The framework’s 147 recommendations give project leaders a vocabulary for surfacing these hidden dependencies early, converting what would otherwise be implicit assumptions into explicit, discussable tasks.

For manufacturing organisations weighing whether to adopt such structured approaches, the study suggests that the investment in process discipline may be what separates systems that create lasting value from those that stall at the prototype stage. The framework does not promise to eliminate the challenges of harmonising human and AI work, but it offers a shared map that lets diverse experts see the whole journey, anticipate one another’s contributions, and catch misalignments while they are still cheap to fix. As manufacturers continue to experiment with machine learning across their operations, the evidence from this research indicates that the organisations most likely to escape the graveyard are those that plan for the human side of the system with the same rigour they apply to the algorithms themselves.

Subject of Research: A development framework for integrating human work into AI systems in manufacturing

Article Title: A development framework for human work integrated AI systems in manufacturing

Article References: Yamamoto, Y., Aranda-Muñoz, Á., & Sandström, K. (2025). A development framework for human work integrated AI systems in manufacturing. Frontiers in Manufacturing Technology, 5, Article 1601903. https://doi.org/10.3389/fmtec.2025.1601903

Image Credits: AI Generated

DOI: 10.3389/fmtec.2025.1601903

Keywords: artificial intelligence, manufacturing, human-AI collaboration, socio-technical systems, machine learning, development framework, Industry 4.0, technology readiness levels, workflow integration, digital manufacturing, project management, predictive maintenance

The model, named MSFF-SATADPN, which stands for Multi-scale Feature Fusion and Self-Attention based Time-Series Anomaly Detection and Prediction Network, was developed by Tengda Chen of Liuzhou Polytechnic University in Guangxi, China. Its design targets a well-documented weakness in existing sequence models. Classic statistical baselines like SARIMA are too rigid to track nonlinear wear trajectories, while hybrid deep architectures such as CNN-LSTM combinations capture local features but lose long-range context. Even powerful attention-based transformers like Informer and PatchTST, which excel at modeling dependencies across long time horizons, tend to overfit when mechanical background noise contaminates the training data. The result is a family of models that either blur sudden structural changes into smoothed-out averages or oscillate wildly when a transient shock hits the sensor waveform. Chen’s approach attacks the problem at two levels simultaneously: how local features are captured, and how the network organizes what it has learned.

The first level relies on progressive dilated convolution. Rather than feeding the raw six-channel sensor stream, which combines three orthogonal cutting force components and three vibration acceleration axes, into a single fixed receptive field, the architecture splits it across three parallel one-dimensional convolution branches with dilation rates of 1, 2, and 4. Small dilation rates give branches tight receptive fields that catch instantaneous, high-frequency impacts, such as the sharp spikes that appear when a cutting edge chips or a workpiece geometry changes abruptly. Larger dilation rates let the kernels skip across time steps, capturing slow, low-frequency structural resonances and gradual degradation trends. After the branches run in parallel, their outputs are concatenated along the channel dimension and compressed through an affine dimensionality-reduction layer, producing a fused local feature sequence that retains information from multiple frequency bands without letting any single band dominate.

On top of this multi-scale backbone sits a self-attention module with absolute positional encoding. Sine and cosine functions inject a deterministic sense of temporal order into the fused features, and the scaled dot-product attention operation then computes similarity scores between every pair of time steps, allowing the network to link an early, subtle micro-fracture at the cutting edge to the severe macro-vibrations that follow hundreds of cycles later. Visualization of the attention weight matrix against the raw waveform shows exactly this behavior: sharp diagonal activation patches align with transient impact events in the physical signal, while a diffuse bright region in the lower corner of the matrix captures the long-range degradation dependency that spans the machining lifecycle. In other words, the attention head learns to watch both the sparks and the slow burn.

The genuinely novel ingredient, however, is what happens after attention. The feature representations are mapped into a higher-dimensional hidden manifold space, and then singular value decomposition is applied to the covariance matrix of the features across the training set. Because steady-state cutting forces dominate the variance of real sensor data while anomalous transients contribute only low-variance residual components, the orthogonal eigenvector basis can be split by singular value magnitude: the top 18 components define a normal steady-state subspace, and the remaining 78 components define an anomalous disturbance subspace. A penalty term, computed as the Frobenius norm of the inner product between the two projected feature sets, is added to the joint training loss, forcing the geometric angle between the normal and anomaly clusters to widen during backpropagation. This orthogonal constraint, frozen between epochs so that the basis vectors are estimated stably over the full training set, effectively filters mechanical background noise at the level of learned representation rather than at the raw signal level.

Downstream, the decoupled features feed two heads. A linear regression layer maps the steady-state matrix to continuous tool wear predictions, optimized with mean squared error, while a fully connected classifier maps the anomaly subspace to a binary fault probability, optimized with cross-entropy. The three loss terms are combined with dynamically balanced hyperparameters into a single joint objective, so regression accuracy, classification sharpness, and geometric decoupling are all optimized in one backpropagation path. The entire model contains only about 24,124 trainable parameters, a striking contrast to the 15.41 million parameters of the Informer baseline and the 7.63 million of PatchTST. That compactness matters: it suggests the performance gains come from architecture rather than brute-force capacity, and it hints at eventual deployment on modest industrial hardware rather than data-center GPUs.

The empirical evaluation used the international PHM 2010 CNC milling benchmark, which records force and vibration data across the full lifecycles of several cutting tools under dry cutting conditions. Three tools were reserved for training and one completely unseen tool served as the blind test set, with tool-level partitioning ensuring that no correlated windows from a single tool’s life leaked between subsets. During the structurally turbulent interval of cycles 215 to 275, where wear accelerates nonlinearly, the Informer model’s residuals oscillated between −0.058 and 0.059 millimeters, while the new model’s residuals converged to a band of −0.015 to 0.018 millimeters, hugging the zero-error baseline. Across the full blind test, the model achieved a root mean square error of 0.073 millimeters and a global mean absolute error of 0.062 millimeters, reductions of 18.0 and 20.5 percent respectively compared with PatchTST, and its coefficient of determination reached 0.901, meaning it explained more than 90 percent of the variance in the tool’s degradation trajectory. Paired t-tests across ten independent random seeds confirmed that all improvements over every baseline were statistically significant at p < 0.001.

Classification performance showed a similar pattern. Under the naturally noisy cutting conditions, the model’s precision and recall converged to 0.941 and 0.936, yielding an overall F1 score of 0.938 with a standard deviation of only 0.008, versus 0.859 for PatchTST and 0.793 for Informer. Its area under the precision-recall curve reached 0.946, outperforming the runner-up by 0.075. Ablation experiments quantified which components carried the weight: removing the orthogonal projection constraint dropped the F1 score from 0.938 to 0.859 and raised the RMSE from 0.073 to 0.089 millimeters; replacing self-attention with simple temporal average pooling was the most damaging change for regression, pushing RMSE to 0.101 millimeters. Sensitivity analyses located clear optima at 18 retained singular directions and an orthogonal-loss weight of 0.08, and t-SNE visualizations showed normal and anomaly feature clusters, which overlapped heavily without the constraint, separating into distinct geometric regions when it was applied.

The author is candid about the limits. The validation remains intra-dataset, conducted on one milling platform under fixed speed and feed, so generalization to other machines, materials, and cutting parameters is unproven, and robustness to noise was inferred from naturally noisy signals rather than controlled injection at defined signal-to-noise ratios. The SVD computation adds computational overhead, and online latency on lightweight edge devices has not been verified. Future work will explore low-rank approximations for lighter decoupling, domain adaptation and meta-learning for cross-machine generalization, and controlled noise experiments at 5 to 20 dB. Still, the core message is compelling: instead of asking a neural network to ignore noise, you can build the geometry of that ignorance directly into its training. For factories betting on predictive maintenance to keep spindles turning and tools from failing mid-cut, that could be the difference between a warning and a surprise.

Subject of Research: Deep learning-based anomaly detection and wear prediction for CNC machine tool time-series sensor data

Article Title: A model for timing anomaly detection and prediction of CNC machine tools based on multi-scale feature fusion and self-attention

Article References: Chen, T. (2026). A model for timing anomaly detection and prediction of CNC machine tools based on multi-scale feature fusion and self-attention. Frontiers in Mechanical Engineering, 12, Article 1967578. https://doi.org/10.3389/fmech.2026.1967578

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1967578

Keywords: CNC machine tools, tool wear prediction, anomaly detection, self-attention, dilated convolution, singular value decomposition, orthogonal projection, predictive maintenance, time-series analysis, PHM 2010 dataset, deep learning, digital manufacturing

That gap is now addressed by a systematic literature review published in Frontiers in the Internet of Things by Nayyar R. Kibria and Javeed Kittur of the University of Oklahoma. Using the PRISMA 2020 framework, the researchers searched seven academic databases, screened 541 records, and retained 45 peer-reviewed publications that explicitly engage with digital twins in computer engineering contexts. They then applied inductive thematic analysis, distilling the literature into six major themes that together map how digital twin systems are designed, deployed, and evaluated from a computing perspective.

The timing of the analysis is telling. The 45 articles span 2017 to early 2026, but 87 percent were published between 2022 and 2025, with a peak of 13 articles in 2025. First authors came from 18 countries, led by the United States and China with six papers each, followed by India and the United Kingdom. Nearly all contributions appeared in archival peer-reviewed journals rather than conference proceedings, a signal that digital twin work in computer engineering is maturing from preliminary demonstrations into sustained, citable scholarship.

Where are digital twins actually being used? Among the 20 studies with clearly specified application domains, power and energy systems dominated at 55 percent, covering smart grids, networked microgrids, substations, lithium-ion batteries, and direct-current microgrids. Industry 4.0, manufacturing, and predictive maintenance accounted for another 20 percent, while networking, extended reality, smart retail, and cross-domain surveys filled out the remainder. The concentration in energy reflects a practical reality: grids are cyber-physical systems where a twin that models both electricity flow and information flow can justify its computational cost.

The review’s first theme, computer engineering foundations, examines the local machinery of twinning: embedded and real-time execution platforms, hardware-software co-design, and synchronization. Standout studies include runtime period adaptation of real-time control tasks under fixed-priority scheduling, a multi-agent-system-based cyber-physical system model spanning nine layers from physical hardware to application software, and a Web of Things architecture called the Relativistic Digital Twin that aims to weave twins into the IoT ecosystem while managing latency and consistency. Collectively, these works show that a digital twin is not a visualization dashboard but an end-to-end systems engineering problem.

The second theme, networked and distributed infrastructure, captured 15 articles and highlights how 5G, 6G, edge computing, and cloud orchestration shape twin performance. Exemplars include Digital Twin Edge Networks toward 6G, digital-twin-aided intelligent offloading with edge selection in mobile edge computing, and secure, latency-aware twin-assisted resource scheduling for 5G-enabled distribution grids. A recurring insight is bidirectional: wireless networks both carry twin data and are themselves twinned for proactive optimization. The reviewers note, however, that most proposed architectures are validated only in simulation or controlled prototypes, not in longitudinal field deployments.

Intelligence is the third and fastest-growing theme. Here, machine learning transforms twins from mirrors into decision engines. One study pairs a semiempirical capacity model with metaheuristically optimized machine learning—using algorithms such as the Improved Grey Wolf Optimizer with AdaBoost—validated by ten-fold cross-validation on NASA battery aging data to predict lithium-ion discharge capacity with very low error. Another trains a reinforcement learning controller entirely inside a digital twin of a DC microgrid, avoiding the safety risks of exploratory actions on live hardware, and then demonstrates improved disturbance response on a real testbed. A third, EcoEdgeTwin, uses deep reinforcement learning to jointly optimize quality of experience, latency, and energy in 6G-era edge networks, treating the twin as an active control system rather than a passive replica.

Security and dependability form the fourth theme, where eight articles position twins as defenders rather than targets. Studies describe IoT-based twins that help networked microgrids withstand coordinated false data injection and denial-of-service attacks, probabilistic twin representations that store object states as probability distributions for risk-sensitive control under uncertainty, and a user-preference-based trust computation model for federated clouds. The reviewers also flag a notable blind spot: distributed ledger technologies could provide tamper-evident data exchange and cross-organization auditability, but they introduce latency and storage costs that must be weighed against real-time requirements.

The fifth theme, modeling, simulation, and testing, is the largest with 25 articles, reflecting how much work goes into making twins trustworthy. Reference architectures like C2PS organize sensing, data handling, and compute services into reusable building blocks, while a systematic review of digital-twin-based testing for cyber-physical systems treats the twin as a test asset, raising questions about test oracles and model visibility. Perhaps the most striking finding is an absence: no single theoretical framework appeared repeatedly across studies, meaning the field has not yet converged on a unified design methodology, and standardized validation metrics for synchronization accuracy, timing jitter, and real-world transfer remain scarce.

The review closes with a research agenda grounded in engineering trade-offs. The best digital twin, the authors conclude, is rarely the highest-fidelity model—it is the model most verifiable under operational constraints of timing, bandwidth, energy, and safety. They call for hardware-aware twin design, safe continual-learning pipelines that can survive distribution shift, interoperability treated as a measurable engineering target rather than a limitation paragraph, and reusable artifacts such as co-simulation testbeds, benchmarks, and open-source reference implementations. As twins edge toward agentic behavior powered by transformers and large language models, those guarantees will only become more critical. For a technology pitched as the nervous system of smart infrastructure, the review’s message is clear: the field must now prove its reliability on real, constrained hardware, not just in slides and simulators.

Subject of Research: Digital twin technologies in computer engineering

Article Title: The emerging role of digital twin technologies in computer engineering research: a systematic literature review

Article References: Kibria, N. R., & Kittur, J. (2026). The emerging role of digital twin technologies in computer engineering research: a systematic literature review. Frontiers in The Internet of Things, 5, Article 1870191. https://doi.org/10.3389/friot.2026.1870191

Image Credits: AI Generated

DOI: 10.3389/friot.2026.1870191

Keywords: digital twins, computer engineering, cyber-physical systems, Internet of Things, edge computing, 6G, on-device AI, systematic literature review, security, interoperability, real-time systems, predictive maintenance

The research team, led by Sizhe Ma with Katherine A. Flanigan and Mario Bergés, benchmarked three architecturally distinct multi-task TSFMs—MOMENT, UniTS, and FFTS—across four downstream tasks within a predictive maintenance context: imputation, anomaly detection, classification, and regression. The testbed was the Commercial Modular Aero-Propulsion System Simulation, or C-MAPSS, a NASA-developed dataset that simulates a large commercial turbofan engine of the 90,000-pound thrust class across varying altitudes, speeds, and throttle settings. By systematically injecting deterioration into the engine’s five rotating components—the fan, low-pressure compressor, high-pressure compressor, high-pressure turbine, and low-pressure turbine—the simulation produces complete run-to-failure trajectories, something real-world operational datasets typically cannot provide because they are truncated to prevent catastrophic failure. The model outputs 21 noisy sensor streams measuring gas-path temperatures, pressures, and shaft speeds, of which 14 sensors responsive to degradation were used for the evaluation.

The choice of predictive maintenance as the case study is deliberate. Under the Open System Architecture for Condition-Based Maintenance standard, predictive maintenance is a hierarchical process whose modules map naturally onto the four benchmarked tasks. Data manipulation requires imputation to recover missing sensor information, simulating sensor failure or transmission loss common in industrial environments. State detection is computationally framed as anomaly detection, identifying deviations from normal behavior. Health assessment diagnoses the specific root cause of degradation through classification. Prognostics assessment projects the future trajectory of the asset to estimate remaining useful life, or RUL, via regression. Because these requirements for interpreting complex, multivariate sensor streams mirror the fundamental data challenges found across digital twin applications, performance on this benchmark serves as a proxy for the analytical layer that any mature digital twin would host.

The experimental design is a controlled test of generalization. All three models were evaluated under identical data preprocessing, splitting, and evaluation frameworks, with ten-fold cross-validation partitioned at the level of the engine unit rather than the window, so that no window spans a group boundary. The test set was held out entirely and never used for training, hyperparameter selection, or epoch selection. Three adaptation regimes were compared: zero-shot inference, in which the pre-trained backbone is applied with no parameter update; lightweight adaptation, which freezes the backbone and updates only a small set of parameters such as task-specific tokens or a linear head; and full fine-tuning, which updates all parameters. Critically, no model was initialized with random weights—every experiment started from a pre-trained backbone derived from official repositories or a reproduction of the original pre-training procedure, as in the case of FFTS, which releases no checkpoint and was pre-trained from scratch using its published federated strategy over 18 datasets from the Monash repository.

The headline finding is a trade-off between generalizability and precision. Although the TSFMs generalized to every evaluated task, they rarely surpassed specialized, domain-specific baselines trained only on C-MAPSS. In imputation, the best TSFM configuration—UniTS under full fine-tuning—reached a mean squared error of 0.123, against 0.105 for the specialized SAITS baseline. In classification, the same configuration reached an accuracy of 0.978, against 0.988 for the DA-Net baseline. In regression, UniTS full fine-tuning achieved an RMSE of 14.21, against 12.90 for the specialized Bi-LSTM Attention baseline. Anomaly detection was the closest contest of the four: the best TSFM configuration—MOMENT in zero-shot mode—reached an adjusted F1 of 0.872 and an AUC of 0.983, against 0.882 and 0.987 for the Anomaly Transformer baseline, meaning the frozen backbone fell within 0.004 AUC of a model trained directly on the target data without any parameter update.

One of the most striking results concerns the task absent from all three models’ pre-training objectives. TSFM pre-training in the time-series domain has centered overwhelmingly on forecasting, driven by the abundance of unlabeled temporal data available for self-supervision. RUL estimation, by contrast, is a regression task that maps an input window to an exogenous scalar target—the number of cycles to failure. The benchmark reveals that the backbone combining channel-mixing with explicit task tokens, UniTS, adapts best to this out-of-pre-training objective, and most visibly when adaptation is restricted to a small set of parameters. From a window width of 64 onward, UniTS prompt learning reached the best lightweight result of the three models on both regression metrics; at a width of 128 it achieved an RMSE of 14.93, against 17.61 for MOMENT linear probing and 19.02 for FFTS linear probing. Remarkably, this frozen configuration also surpassed the fully fine-tuned MOMENT and FFTS and the specialized LSTM baseline, all while leaving the pre-trained weights untouched.

The study also uncovered a counterintuitive pattern in anomaly detection. For all three models, fine-tuning outperformed zero-shot inference at narrow window widths of 16 and 32, but fell below it from a width of 64 onward. The authors offer a plausible explanation rooted in the unsupervised training strategy: because the reconstruction objective is fitted on the entire run-to-failure training data, a longer window exposes the model to more degraded behavior, and a backbone with sufficient capacity can learn to reconstruct that degraded behavior as faithfully as the healthy behavior it is meant to characterize. The share of anomalous training data rises from roughly 9 percent at the narrowest width to approximately 45 percent at the widest, so both the proportion of contaminated samples and the degraded content of each sample increase with window width. The practical implication reverses the pattern observed in the other three tasks: in reconstruction-based monitoring, zero-shot or lightweight use of a TSFM may be preferable to full adaptation.

Window width itself emerged as a decisive constraint. All three models degraded at the shortest tested window width of 16, and the degradation was largest for UniTS in every task. A candidate explanation lies in patch-based tokenization: MOMENT uses a patch size of 8 and can resolve a complete patch at width 16, whereas UniTS and FFTS employ a patch size of 16 and are fed input sequences shorter than a single patch—a scenario unseen during their pre-training. The evidence is mixed, however, since FFTS shares UniTS’s patch size yet degrades least of the three, reaching the best zero-shot adjusted F1 in anomaly detection and the best lightweight RMSE in regression at that width. The authors therefore report patch size as a plausible contributor rather than an established cause, and note that isolating it would require an ablation varying patch size within a single backbone. The practical lesson stands regardless: for monitoring high-frequency assets at short window widths, the alignment between a model’s tokenization and the inference window length should be considered per model rather than assumed.

In the zero-shot regime, no single backbone led across tasks, and the results were only partially consistent with an effect of pre-training data scale. MOMENT, pre-trained on the largest corpus—the 1.23-billion-time-point Time Series Pile—reached the best zero-shot imputation error at every window width and the best zero-shot AUC in anomaly detection at two widths. Yet at other widths, FFTS, whose corpus is less than half the size, reached the higher zero-shot adjusted F1. Because the three models differ simultaneously in architecture, parameter count, pre-training corpus and objective, and patch size, the authors are careful to frame these as patterns observed under one shared benchmarking framework rather than as isolated causes. An appendix examining the pre-training corpora confirms that all three reach only the modality level of familiarity with C-MAPSS: temperatures, pressures, and rate-like quantities occur in all three corpora, but none contains gas-path instrumentation of any engine, and C-MAPSS itself is absent from all three pre-training datasets.

The authors are equally careful about scope: the evaluation was conducted on raw sensor series alone, without the semantic or physics-based components of a mature digital twin, so the findings establish an empirical baseline for the analytical layer such a system would host rather than an evaluation of digital twin integration itself. Looking forward, they argue that closing the gap to specialized models may require moving beyond purely data-driven pre-training toward knowledge-guided TSFMs that are supplied with the physical constraints, system topology, and failure logic a digital twin already holds, rather than expecting the backbone to recover them from raw sensor data. Whether that strategy closes the gap remains an open question—but this benchmark provides the raw-data reference point against which it can finally be asked.

Subject of Research: Artificial Intelligence

Article Title: Benchmarking time-series foundation models for multitask predictive maintenance in digital twins

Article References: Ma, S., Flanigan, K. A., & Bergés, M. (2026). Benchmarking time-series foundation models for multitask predictive maintenance in digital twins. Frontiers in Aerospace Engineering, 5, Article 1937868. https://doi.org/10.3389/fpace.2026.1937868

Image Credits: AI Generated

DOI: 10.3389/fpace.2026.1937868

Keywords: digital twins, time-series foundation models, predictive maintenance, machine learning, aerospace engineering, remaining useful life, C-MAPSS, benchmarking, time-series, foundation, models, multitask