The research team, led by Daniel F. Goncalves together with Aldemir A. Cavallini Jr. and Valder Steffen Jr., focused on a class of machines that is notoriously difficult to monitor: horizontal rotors supported by hydrodynamic bearings. Unlike conventional rolling-element bearings, hydrodynamic bearings support the rotating shaft on a thin, pressurized film of oil, and this fluid-film behavior is inherently nonlinear. That nonlinearity distorts vibration signatures in ways that complicate diagnosis, which is precisely why most existing AI-based fault detection studies, which typically concentrate on rolling bearings, do not translate easily to this class of machinery.

The specific faults the team targeted are among the most common and most confounding in rotating equipment: unbalance, misalignment, and shaft cracks. Each of these defects produces vibration symptoms that overlap substantially, meaning that a maintenance engineer looking at raw vibration data may struggle to distinguish a developing crack from a simple misalignment. The stakes of getting this wrong are high, because a crack that goes undetected can propagate until the shaft fails outright, while an incorrect diagnosis can trigger unnecessary and costly maintenance interventions.

To build their diagnostic system, the researchers combined physical experimentation with high-fidelity simulation. They constructed a test rig consisting of a horizontal steel shaft, 19 millimeters in diameter and 950 millimeters long, driven by an electric motor at a fixed speed of 1,200 revolutions per minute and supported by two hydrodynamic bearings lubricated with ISO VG 68 oil. A corresponding finite element model of the rotor, built from 44 Timoshenko beam elements, was carefully calibrated against the experimental rig using a differential evolution optimization procedure, ensuring that the virtual model faithfully reproduced the machine’s dynamic behavior. This calibrated digital model, implemented with the open-source Rotordynamic Software ROSS, then allowed the team to generate large volumes of synthetic training data covering fault conditions that would be impractical or dangerous to create physically.

From the vibration signals, gathered at two measurement planes with orthogonally positioned proximity sensors, the team extracted features spanning three domains: statistical measures such as skewness, kurtosis, and root-mean-square values in the time domain; harmonic amplitudes in the frequency domain; and wavelet packet sub-band energies in the time-frequency domain. The raw feature set was then distilled using a hybrid strategy that pairs the Relief algorithm, an unsupervised feature ranking method, with SHAP, a model-dependent explainability technique rooted in cooperative game theory. This dual ranking produced a remarkably compact set of just 12 features, a reduction that slashed computational cost while actually improving diagnostic performance compared with using the full feature set or principal component analysis.

The architecture itself operates in two tiers. First, an ensemble of three clustering algorithms, hierarchical agglomerative clustering, DBSCAN, and Gaussian mixture models, votes by majority to flag whether a new vibration sample represents a novelty or deviation from healthy behavior. This unsupervised layer is crucial because real industrial machines rarely come with labeled databases covering every possible operating condition. Then, once an anomaly is detected, supervised classifiers, specifically k-nearest neighbors and support vector machines, take over to identify which specific fault is present. Hyperparameters for every model were tuned through grid search with cross-validation, and the feature selection was performed strictly on training data to prevent any information leakage that could inflate the results.

The performance figures are striking. In numerical tests, both classifiers achieved precision, recall, and F1-scores exceeding 98.9 percent, with the support vector machine reaching a recall of 99.64 percent and the k-nearest neighbors model posting a weighted F1-score of 99.13 percent. In experimental validation on the physical test rig, using 400 samples that included healthy, unbalanced, misaligned, and cracked conditions, the support vector machine achieved a weighted F1-score of 99.30 percent, while the crack class alone reached an F1-score of 98.95 percent despite being the smallest category in the dataset. The few misclassifications that did occur almost exclusively involved confusion between misalignment and cracks, the two faults with the most similar vibration signatures, and even these errors were statistically marginal.

Equally important for industrial deployment is speed. The researchers measured the complete computational pipeline for a single data window, including sliding-window processing, feature extraction via the discrete Fourier transform and wavelet decomposition, feature scaling, and final prediction. The full cycle took roughly 0.17 to 0.18 seconds for classification and about 0.34 seconds for the clustering-based detection stage, comfortably fast enough for real-time monitoring of continuously operating machinery without computational bottlenecks.

What truly sets this work apart, however, is its insistence on transparency. Using SHAP values, the team opened the black box and verified that the classifiers were reasoning in physically meaningful ways. The analysis revealed that the amplitude of the second harmonic of the rotational speed, extracted from measurement plane S1, was the single most influential feature for the support vector machine, contributing a total SHAP value of approximately 0.35, with the largest share attributable to crack conditions. This aligns perfectly with rotor dynamics theory, since misalignment and cracks are known to induce periodic disturbances at twice the rotational frequency. Even more compelling, the skewness of the vibration signal emerged as a critical discriminator for cracks, a result the researchers could explain directly: a propagating crack introduces asymmetry into the rotor system, which skews the distribution of the vibration response. The AI, in other words, had independently rediscovered a genuine physical phenomenon.

The Brazilian team is now developing an application programming interface to bring the methodology into real industrial environments, and they acknowledge that the current results were obtained at a single fixed rotational speed, a deliberate choice that isolated fault signatures from the complications of variable operating conditions. Future work will test the framework across different speeds, imbalanced datasets, and bearing-specific failure modes such as journal ovalization and lubrication problems. Still, the implications are considerable. By pairing robust, lightweight classifiers with explanations that a maintenance engineer can verify against physical intuition, the study offers a template for the kind of trustworthy, interpretable AI that Industry 5.0 demands, where intelligent machines and human experts collaborate rather than compete, and where a computer’s diagnosis of a cracked shaft comes with evidence a seasoned engineer can actually check.

Subject of Research: Explainable AI-based vibration fault diagnosis in rotating machines supported by hydrodynamic bearings

Article Title: Fault recognition approach applied to a rotating machine supported by hydrodynamic bearings using explainable artificial intelligence

Article References: Goncalves, D. F., Cavallini, A. A., Jr., & Steffen, V., Jr. (2026). Fault recognition approach applied to a rotating machine supported by hydrodynamic bearings using explainable artificial intelligence. Applied Intelligence, 56(15), Article 461. https://doi.org/10.1007/s10489-026-07491-9

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07491-9

Keywords: rotating machinery, hydrodynamic bearings, fault diagnosis, explainable artificial intelligence, SHAP, vibration analysis, predictive maintenance, machine learning, support vector machine, k-nearest neighbors, ensemble clustering, rotor dynamics

The core challenge the method confronts is known in the field as sparse reconstruction. In a typical monitoring campaign, accelerometers or strain gauges are placed at a limited number of accessible locations, often far fewer than the number needed to describe the structure’s motion completely. Between the measured points lies a vast space of unobserved behavior, and traditional interpolation or modal expansion techniques struggle to fill it accurately, particularly when the structure is excited by unpredictable forces such as wind, traffic, or seismic shaking. Classical approaches usually assume that the response can be described by a fixed set of vibration modes with slowly varying amplitudes, an assumption that breaks down under nonlinear behavior, transient impacts, or rapidly changing operating conditions.

The innovation at the heart of the new work lies in the way the model processes signals. Rather than treating a vibration record as a simple sequence of numbers sampled in time, the time-frequency gated transformer first transforms the data into a representation that captures both temporal and spectral information simultaneously. This matters because structural responses are inherently multiscale: a bridge deck may carry slow, low-frequency sways driven by wind alongside rapid, high-frequency ripples generated by passing vehicles or local impacts. A model that looks only at the raw time series can miss the spectral signatures that distinguish one excitation mechanism from another, while a purely frequency-domain view loses the precise timing of transient events. By operating in the joint time-frequency domain, the network gains access to a richer description of the physics encoded in the measurements.

The transformer architecture, which has transformed natural language processing and computer vision over the past several years, provides the computational backbone. Transformers rely on attention mechanisms, mathematical operations that allow the model to weigh the relevance of every part of an input sequence when interpreting any given part. Applied to structural data, attention lets the network learn long-range spatial and temporal correlations: for example, how a vibration pattern measured at the base of a tower relates to the response at its top, or how an impact at one point in time influences the oscillations observed seconds later. This ability to capture dependencies across long distances in the data is precisely what sparse reconstruction demands, since unmeasured locations must be inferred from patterns observed elsewhere on the structure.

What distinguishes this implementation from a vanilla transformer is the gating mechanism applied in the time-frequency domain. Gates are learned functions that selectively amplify or suppress components of the representation, allowing the network to decide, on a case-by-case basis, which frequency bands and time windows carry the most reliable information for the reconstruction task at hand. During training, the model adjusts these gates so that noise-dominated or physically irrelevant portions of the signal are downweighted, while informative features pass through to the reconstruction layers. The result is a form of learned signal filtering that adapts dynamically to the data rather than relying on hand-designed filters, a significant advantage when dealing with real-world measurements contaminated by sensor noise, environmental variability, and electromagnetic interference.

The practical implications of such a system are considerable. Structural health monitoring has become a global priority as aging bridges, dams, and buildings face increasing loads from climate extremes, heavier traffic, and material degradation. Full-field reconstruction, the ability to estimate the displacement, velocity, or acceleration at every point of a structure from limited measurements, enables engineers to detect damage earlier, locate it more precisely, and assess remaining service life with greater confidence. In aerospace, where instrumenting every square meter of a wing or fuselage is impractical, a model that infers the complete dynamic response from a sparse sensor array could improve flutter prediction, fatigue tracking, and certification testing. In mechanical engineering, rotating machinery such as wind turbines and jet engines could be monitored with fewer sensors, reducing cost and downtime.

Deep learning approaches to this problem have been attempted before, but they have often stumbled on generalization. A network trained on one structure, one sensor layout, or one class of excitation frequently fails when conditions change, which is a serious limitation for infrastructure that must endure decades of varying environments. The time-frequency gating strategy is aimed squarely at this weakness. By explicitly separating information across time and frequency scales and learning which combinations matter, the model builds a more physically grounded internal representation, one that is less tied to the specific idiosyncrasies of its training data. The authors report that the architecture reconstructs dynamic responses with high accuracy across a range of scenarios, outperforming conventional baseline methods that lack the joint time-frequency treatment.

Like any data-driven method, the approach depends on the quality and diversity of its training data, and the researchers acknowledge that deploying such models on real structures requires careful validation against measured ground truth. Questions of uncertainty quantification also remain active areas of research: engineers need to know not only what the model predicts but how confident it is in each estimate, particularly when the reconstruction informs safety-critical decisions. The field is moving rapidly toward hybrid frameworks that combine the flexibility of neural networks with the guarantees of physics-based models, and gated time-frequency transformers of the kind presented here are likely to serve as powerful components within such systems. The study’s publication in a Nature Portfolio engineering journal signals growing mainstream recognition that machine learning and structural dynamics have converged into a productive research frontier.

The broader significance of the work extends beyond any single application. It exemplifies a wider trend in which attention-based architectures, originally developed for language, are being repurposed to solve problems in physical science and engineering, from weather forecasting to materials discovery. Structural dynamics, with its rich multiscale character and its sparse, noisy measurements, turns out to be a natural fit for these tools. As sensor hardware becomes cheaper and wireless networks make dense instrumentation feasible, the bottleneck will shift from data acquisition to data interpretation, and models like the time-frequency gated transformer will define how effectively that bottleneck is cleared. For the engineers responsible for keeping the world’s infrastructure safe, the study offers a glimpse of monitoring systems that see far more than the sensors they are built on, reconstructing the hidden motion of structures with a fidelity that was recently the province of simulation alone.

Subject of Research: Deep learning-based reconstruction of structural dynamic responses from sparse vibration measurements

Article Title: Time-frequency gated transformer for structural dynamic response reconstruction

Article References: Song, X., Yang, F., Li, R., Ma, X., Wang, Y., Wang, S., Deng, Q., & Zheng, S. (2026). Time-frequency gated transformer for structural dynamic response reconstruction. Communications Engineering. https://doi.org/10.1038/s44172-026-00788-0

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00788-0

Keywords: structural health monitoring, transformer, deep learning, time-frequency analysis, vibration reconstruction, sparse sensors, structural dynamics, attention mechanism, civil infrastructure, aerospace engineering, machine learning, signal processing