The research, published in Frontiers in Forests and Global Change, examines Austria, Brazil, Canada, China, Finland, France, Germany, Sweden, and the United States over the period from 1961 to 2022, using 62 annual observations per country drawn from the Global Footprint Network. The team, led by İlkay Güler of Ankara Haci Bayram Veli University together with colleagues from Bingöl University, Arkin University of Creative Arts and Design, Tarsus University, and Artvin Coruh University, focused on a measure called forest ecological balance, or FEB. This indicator is calculated as forest biocapacity minus the forest product footprint, capturing both the supply side, meaning the regenerative capacity of a nation’s forests, and the demand side, meaning the pressure that annual consumption of forest products places on those forests.

The choice of indicator matters. Most previous work on the stochastic properties of environmental indicators has concentrated on carbon dioxide emissions or aggregate ecological footprint, leaving forest-specific conditions comparatively underexplored. Forests underpin the United Nations’ Sustainable Development Goal 15, Life on Land, but they also feed into goals on food security, clean water, and climate action. Roughly 1.6 billion people depend directly or indirectly on forests for their livelihoods, and forest resources can account for around 22 percent of household income in some tropical developing countries. Understanding whether damage to these systems fades away or lingers is therefore far more than an academic exercise.

The statistical logic at the heart of the study is deceptively simple. If an environmental indicator is stationary, it fluctuates around a constant mean, and the effects of any shock, whether a policy change, a drought, or a market disruption, eventually dissipate as the series returns toward its long-term equilibrium. If the series is non-stationary, containing what statisticians call a unit root, shocks can have permanent consequences, and the system may drift to a new state rather than recover. For policymakers, this distinction determines whether an intervention produces a temporary correction or a lasting transformation, and whether uniform international frameworks can plausibly deliver comparable outcomes across different national contexts.

To probe these properties, the researchers applied two complementary tests. The conventional KPSS test, introduced by Kwiatkowski and colleagues in 1992, treats stationarity as the null hypothesis and decomposes a series into a deterministic trend, a random walk, and a stationary error term. Its weakness is that it ignores structural breaks, which can bias results toward rejecting stationarity. The Fourier KPSS test addresses this by incorporating sine and cosine functions that flexibly capture smooth, unknown, and multiple structural changes in the data. When the trigonometric terms prove statistically significant, the Fourier version provides the more reliable reading. Lag lengths in both tests were selected automatically using the Newey-West bandwidth with a Bartlett kernel to handle serial correlation and heteroskedasticity in the residuals.

The second pillar of the methodology is the maximal overlap discrete wavelet transform, or MODWT, which splits each country’s forest ecological balance series into short-term components spanning two to eight years, medium-term components spanning eight to thirty-two years, and long-term components beyond thirty-two years. The team used the Daubechies Least Asymmetric filter of length eight, known as LA8, together with a reflection boundary condition to preserve sample size and minimize information loss at the edges of the data. The authors preferred MODWT over the classical discrete wavelet transform because it is phase-independent, accommodates arbitrary data lengths, and retains outlying observations, making it better suited to relatively short annual samples.

The headline results reveal striking heterogeneity. For the original, undecomposed series, the conventional KPSS test found non-stationarity in every country, but the Fourier KPSS test, after accounting for smooth structural changes, identified stationary behavior in China, Germany, and the United States. The remaining six countries, Austria, Brazil, Canada, Finland, France, and Sweden, showed non-stationary dynamics, meaning shocks there may leave more persistent marks. The authors attribute China’s apparent stability to roughly three decades of large-scale reforestation and ecological restoration programs, Germany’s to long-standing institutionalized sustainable forest management aligned with European Union environmental goals, and the United States’ to pioneering anti-pollution policies, advanced forest monitoring systems including the Forest Inventory and Analysis program, and the ecosystem services approach.

The frequency-domain analysis adds a crucial layer of nuance. In the short run, most countries displayed non-stationary behavior, with only Finland and France showing mean-reverting dynamics, suggesting that short-term disturbances in those two countries dissipate more rapidly. In the medium run, the picture changed dramatically: both KPSS and Fourier KPSS tests supported stationarity for all nine countries once structural breaks were considered, indicating that medium-term forest ecological balance is generally resistant to persistent shocks regardless of national governance differences. The long run was the most heterogeneous of all. Fourier KPSS results flagged persistent, non-stationary behavior in China and Finland, while the other seven countries, Austria, Brazil, Canada, France, Germany, Sweden, and the United States, exhibited stationary long-run components where policy shocks would prove only temporary.

These findings carry a pointed message for forest policy. Because persistence varies by both country and time horizon, uniform forest-policy frameworks are unlikely to generate comparable outcomes across major forest-producing economies. The authors recommend short-term measures such as enforcing existing forest laws, cracking down on illegal logging, deploying rangers, drones, and satellite monitoring, and rapid reforestation, alongside long-term strategies including restoration of degraded areas, institutional capacity building, and conservation incentives. Country-specific recommendations range from Brazil’s need for stricter enforcement against illegal logging and protection of Indigenous land rights, to Canada’s focus on wildfire prevention and Indigenous co-governance, to Sweden’s push to reduce intensive clear-cutting in sensitive areas. The authors caution that their evidence concerns temporal persistence rather than causal effects, and that stationarity alone should not be read as proof of ecological resilience. Still, by showing that the same forest indicator can behave differently at different frequencies, the study offers an empirical foundation for forest management strategies tailored to both the nation and the timescale, a combination that one-size-fits-all policy has long overlooked.

Subject of Research: Frequency-dependent persistence of forest ecological balance in major forest product-producing countries

Article Title: Assessing forest ecological sustainability through wavelet decomposition and stationarity tests: a comparative analysis of major forest product producers

Article References: Güler, İ., Naimoğlu, M., Özbek, S., Adalı, Z., & Şimşek, O. (2026). Assessing forest ecological sustainability through wavelet decomposition and stationarity tests: a comparative analysis of major forest product producers. Frontiers in Forests and Global Change, 9, Article 1622811. https://doi.org/10.3389/ffgc.2026.1622811

Image Credits: AI Generated

DOI: 10.3389/ffgc.2026.1622811

Keywords: forest ecological balance, wavelet decomposition, stationarity tests, Fourier KPSS, forest biocapacity, ecological footprint, forest sustainability, MODWT, forest policy, structural breaks, timber production, mean reversion

The question matters because the geniculostriate pathway, the main conduit carrying visual information from the eye through the thalamus to the cortex, is built from parallel neuronal projections that each originate at a particular retinal location. If the wiring between anatomical areas were perfectly precise, positional information would be preserved from stage to stage. But biological wiring is never perfect, and any imprecision in these projections would place hard constraints on the visual system’s ability to perform tasks that require knowing where things are, such as identifying the letters of a word. Zhu and Baldwin set out to quantify exactly how much such spatial scrambling would hurt performance, and where along the visual pathway the scrambling would be most damaging.

To do this, the researchers built their analysis around a canonical model of the early visual cortex, focusing on the first orientation-tuned mechanisms in the processing hierarchy, which are analogous to the simple cells found in primary visual cortex, or V1. These cells respond to oriented edges, such as the vertical stroke of a letter T or the diagonal of an N, and they sit at a crucial junction in the visual pipeline. The authors considered two distinct points at which scrambling could occur. The first is the input to these orientation-tuned mechanisms, before orientation is even extracted; the second is the output from those mechanisms, after orientation information has already been computed. They refer to the first case as subcortical scrambling and the second as cortical scrambling, and the distinction turns out to be far from academic.

The methodological heart of the study is a wavelet decomposition and resynthesis algorithm that the researchers developed to mimic these two kinds of distortion. Wavelets are localized mathematical building blocks that can represent an image as a collection of small, oriented patches at different positions and scales, which makes them a natural stand-in for the receptive fields of orientation-tuned neurons. By decomposing a letter image into wavelets and then resynthesizing it after perturbing the spatial positions of those components, the team could generate distorted letters that simulated either subcortical or cortical scrambling. Crucially, the two manipulations produce visibly different results: scrambling the input to orientation mechanisms injects orientation noise into the image, while scrambling the outputs preserves the orientation content but jumbles where each oriented fragment ends up.

With their stimulus-generation machinery in place, Zhu and Baldwin measured how human observers performed on letter identification when the letters had been distorted by the two types of scrambling. Participants viewed the scrambled letters and attempted to identify them, allowing the researchers to derive identification thresholds, the level of distortion at which observers could still perform the task at a criterion level of accuracy. The results showed that subcortical and cortical scrambling have distinguishable effects on both the perceived noisiness of the letters and on the identification thresholds themselves. In other words, the two kinds of distortion are not perceptually interchangeable: the human visual system treats them differently, and the difference is measurable both in how noisy the letters look and in how hard they are to read.

To understand why, the researchers turned to convolutional neural networks, the workhorses of modern computational vision. They assembled a suite of pre-trained and custom-built CNNs and trained them on the same scrambled stimuli that the humans had seen. By comparing the distortion level at which humans could just identify a letter with the level at which a network could do the same, the team computed a relative efficiency, expressed as the ratio of human to CNN thresholds. This efficiency metric is a standard tool in vision science because it separates two sources of performance limits: the information actually available in the stimulus, and the observer’s ability to extract that information. A human who needs twice the distortion tolerance of an optimal machine observer is said to be only half as efficient.

The comparison produced a striking and somewhat counterintuitive pattern. Relative efficiency, calculated against CNNs trained on the scrambled stimuli, was higher for cortical scrambling than for subcortical scrambling. On the surface, that suggests humans cope relatively well when the distortion happens after orientation has been extracted. But when the researchers modeled human inefficiency in a different way, by reducing the proportion of wavelets available to the CNNs and asking how much of the neural representation had to be removed to match human performance, the picture flipped: humans were less efficient under cortical scrambling than under subcortical scrambling. The two analyses, taken together, reveal that the direction of the human advantage depends on which kind of distortion is in play, and that the flip is informative about the underlying neural computation.

The resolution of this apparent paradox, the authors argue, lies in what each type of scrambling does to the information content of the stimulus. Cortical scrambling produces stimuli that are orientation-redundant: the oriented fragments are still present and consistent, just displaced, so the visual system can pool over many redundant orientation measurements to recover the letter’s identity. Subcortical scrambling, by contrast, produces orientation-noisy stimuli, in which the orientation signals themselves are corrupted and cannot be recovered by pooling. The efficiency differences show that humans are better at processing orientation-redundant stimuli than orientation-noisy ones. Human vision, in other words, excels at integrating consistent orientation evidence across space but struggles when that evidence is internally inconsistent, a pattern that no amount of clever pooling can fully repair.

From this pattern, Zhu and Baldwin draw a hypothesis about cortical architecture. The differences in efficiency between the two scrambling conditions, they propose, reflect differences in the integration properties at the input and output stages of simple cells in the cortex. At the input stage, where signals from the thalamus converge onto orientation-tuned neurons, the integration may be relatively tolerant of positional imprecision, because the cell’s receptive field pools over a spatial neighborhood and can average out small displacements. At the output stage, where the orientation signals are passed onward to downstream areas, the integration may operate differently, and positional jumbling of already-extracted orientation features may be harder for the system to undo. The asymmetry in human performance thus becomes a probe into how information is combined at successive stages of the early visual hierarchy.

Beyond its implications for basic neuroscience, the study offers a template for how computational models and human psychophysics can be combined to test hypotheses about neural coding that cannot be directly observed. By constructing stimuli that isolate a specific computational manipulation, measuring human behavior on those stimuli, and benchmarking that behavior against machine observers with known properties, researchers can infer properties of the biological system that would otherwise remain hidden. The work also speaks to a broader question in vision science: the human visual system is not a perfect measuring device, and its inefficiencies are not random. They are structured, task-dependent, and revealing. When letters are warped by physiologically inspired spatial scrambling, the way our performance degrades tells us something fundamental about how the brain wires itself to read the world, one oriented edge at a time.

Subject of Research: The effects of physiologically inspired spatial scrambling at early stages of the human visual pathway on letter identification

Article Title: Identification of letters distorted by physiologically-inspired spatial scrambling

Article References: Zhu, R. X., & Baldwin, A. S. (2026). Identification of letters distorted by physiologically-inspired spatial scrambling. PLOS Computational Biology, 22(10), e1014856. https://doi.org/10.1371/journal.pcbi.1014856

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014856

Keywords: visual neuroscience, spatial scrambling, letter identification, V1 simple cells, wavelet decomposition, convolutional neural networks, relative efficiency, geniculostriate pathway, orientation tuning, PLOS Computational Biology, psychophysics, cortical processing

The heart of the contribution is ISL-MDN, a publicly released dataset of inertial recordings capturing continuous multi-digit number signing. Forty-one participants wore a glove instrumented with six IMU sensors sampling at 100 hertz, producing streams of acceleration and angular velocity data as they signed sequences of two, three, and four digits. Crucially, the dataset includes sign boundary annotations, meaning researchers know exactly where one digit sign ends and the next begins. This annotation scheme allows the systematic evaluation of two fundamentally different recognition paradigms within a single benchmark: segment-based approaches that classify each pre-segmented sign individually, and sequence-based approaches that must infer the digit string from the unbroken sensor stream. The researchers also reserved unseen six-digit sequences specifically to test whether models trained on shorter sequences could generalize to longer, previously unencountered combinations—a rigorous test of true sequence understanding rather than memorization.

The choice of inertial sensing is deliberate and consequential. Unlike video, IMU data contains no images of the signer, so privacy is preserved by design; there is nothing to identify a face, a location, or a bystander. The data volume is also dramatically smaller. The authors calculate that, normalized by the number of represented classes, ISL-MDN requires more than fifty times less storage per class than comparable video-based datasets. That compactness matters enormously for edge AI, where memory, bandwidth, and battery budgets are tight. A glove with six tiny motion sensors is cheap, unobtrusive, and works identically in a dark room or bright sunlight—conditions that routinely defeat camera systems.

On the algorithmic side, the team benchmarked architectures spanning both recognition paradigms. For segment-based recognition, they built a ResNet convolutional baseline and a more sophisticated model they call the Transformer Core Method, or TCM. TCM exploits the transformer architecture’s celebrated attention mechanism in two directions at once: it models dependencies between the six sensor channels within each IMU frame, capturing how the motion of one part of the hand relates to another at a single instant, and it models temporal relationships across the segmented sign units, capturing how the hand’s configuration evolves from the start to the end of a digit. This dual attention lets the network weigh which sensor channels and which moments in time carry the most discriminative information for distinguishing, say, a signed four from a signed five.

Segment-based recognition, however, sidesteps the hardest part of the problem: in the real world, nobody tells the system where one sign ends and the next begins. For sequence-based recognition, the authors proposed their centerpiece, the Hierarchical Temporal Compression Transformer, or HTCT. Continuous sign language recognition from raw sensor streams is notoriously difficult because the model must implicitly align an input sequence of arbitrary length with an output label sequence of unknown length, without frame-level supervision. HTCT tackles this with a training objective based on Connectionist Temporal Classification, or CTC, a technique originally developed for speech recognition that allows the network to learn the alignment itself by summing over all possible alignments between inputs and labels.

The truly novel ingredient in HTCT is its hierarchical temporal compression. Raw IMU streams at 100 hertz are highly redundant: consecutive samples carry largely overlapping information, and the discriminative structure of a signed digit unfolds over hundreds of milliseconds, not milliseconds. HTCT combines a fixed Haar wavelet decomposition—a classical signal processing tool that decomposes a signal into coarse trends and fine details at multiple scales—with a learnable, causal temporal compression module. The wavelet stage provides a principled multi-scale representation, capturing both the rapid transients of a hand flick and the slower posture changes that define a digit, while the learnable compression stage is trained to discard redundant temporal information and retain only the features that help the CTC objective. The result is a much shorter, information-dense sequence that the transformer can attend over efficiently, improving both accuracy and speed.

The numbers tell a compelling story. The ResNet baseline achieved a Word Error Rate of 16.3 percent, where a word corresponds to a signed digit and the error rate measures how often the recognized digit sequence differs from the ground truth. TCM improved on that with 14.6 percent, confirming the value of modeling cross-channel dependencies with attention. But HTCT left both far behind, reaching a Word Error Rate of just 5.2 percent—roughly a threefold reduction in errors relative to the baseline. Even more remarkable is what that accuracy costs computationally: only 1.73 million floating-point operations per inference. For context, that is orders of magnitude below what typical video-based sign language transformers require, placing HTCT squarely in the territory of microcontrollers and low-power wearable processors. The authors note that the model achieves an effective accuracy–efficiency trade-off through low computational cost and inference latency, while candidly acknowledging that power consumption was not directly measured in the study.

The generalization test adds further weight. When confronted with six-digit sequences—longer than anything seen during training—the system’s ability to recognize previously unseen combinations demonstrates that HTCT is learning compositional structure rather than memorizing fixed-length patterns. This matters because real-world usage is inherently open-ended: a signer might need to communicate a bank account number, a serial code, or a date of arbitrary length. A recognizer that only works for the sequence lengths in its training set would be of limited practical value; one that composes digit signs flexibly can scale to the task at hand.

The release of ISL-MDN as an open resource, hosted on Zenodo with source files and documentation on GitHub, may prove as influential as the model itself. Inertial sensing remains relatively underexplored for continuous sign language recognition compared with vision, and one persistent obstacle has been the scarcity of annotated, signer-diverse datasets. By providing boundary annotations, multiple sequence lengths, held-out generalization data, and forty-one participants, ISL-MDN gives the community a unified benchmark on which segment-based and sequence-based methods can be compared fairly. It also extends coverage to Iranian Sign Language, a language underrepresented in the predominantly English- and Chinese-centric sign language recognition literature, and the data collection followed institutional ethical guidelines with informed consent and full anonymization of participant data.

The broader significance lies in what this work suggests about the future of assistive wearables. Sign language recognition has long been framed as a computer vision problem, but the wrist and fingers are where the signal actually lives, and miniature motion sensors can capture it directly, privately, and cheaply. Combined with aggressively compressed transformer architectures like HTCT, the vision of a glove that translates continuous signing into text entirely on-device—no cloud, no camera, no privacy trade-off—moves from aspiration toward engineering reality. Challenges remain, including scaling beyond digits to full vocabularies, testing across more sign languages, and validating battery life in the field. But with a 5.2 percent error rate at 1.73 million FLOPs, this study makes a persuasive case that the next breakthrough in accessible communication may come not from bigger models watching us, but from smaller ones riding along on our hands.

Subject of Research: Inertial-sensor-based continuous sign language number recognition using transformer models for edge AI

Article Title: Inertial-sensor–driven continuous multi-digit sign language number recognition using hierarchical temporal compression and transformer-based methods for edge AI

Article References: Parvizi, A., Jamshidi, K., & Shahbazi, H. (2026). Inertial-sensor–driven continuous multi-digit sign language number recognition using hierarchical temporal compression and transformer-based methods for edge AI. Multimedia Tools and Applications, 85(10), Article 792. https://doi.org/10.1007/s11042-026-21947-7

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21947-7

Keywords: sign language recognition, inertial sensors, wearable technology, transformer, edge AI, IMU, CTC, wavelet decomposition, Iranian Sign Language, dataset, deep learning, accessibility

The study, published in the International Journal of Aeronautical and Space Sciences, addresses one of the most stubborn measurement challenges in hypersonic testing. Strain-gauge balances and other force sensors inside a wind tunnel do not simply record the steady lift, drag, and moments acting on a model. They record a mixture: the true aerodynamic load, plus the dynamic response of the entire model-balance-support structure as it is excited by the tunnel’s start-up shock. In impulse and shock tunnels, where useful test times can be extremely short, engineers traditionally had to wait for vibrations to decay before trusting the data. For large-scale heavy-duty models, that waiting period may consume the entire test window, and the low-frequency vibration components overlap with the aerodynamic signal itself, making conventional filtering useless because a filter that removes the vibration would also remove the force.

The core idea behind the new framework is deceptively simple: instead of asking a neural network to learn everything from a raw, tangled time series, first decompose the signal into physically meaningful frequency bands using wavelet analysis, and then let the deep learning model work on that structured representation. Wavelet decomposition acts like a prism for signals, splitting a measurement into components that each occupy a limited frequency range while preserving information about when events occur in time. This frequency-selective representation gives the network a head start. Rather than discovering on its own that a slow oscillation and a fast ripple carry different kinds of information, the model receives those features already separated, which the researchers found improves convergence speed and reduces the number of training samples needed to reach a given accuracy.

That last point matters enormously in practice. Deep learning models are notoriously data-hungry, and wind-tunnel experiments are among the most expensive data sources in engineering. Every run of a hypersonic facility consumes compressed gas, high-enthalpy heating power, instrumented model time, and hours of skilled operator attention. A method that achieves accurate force identification with fewer training runs directly lowers the cost and increases the throughput of hypersonic testing campaigns. The team reports that integrating wavelet decomposition into the learning architecture maintained identification accuracy while cutting the required sample count, a combination that could change how force-measurement campaigns are planned at facilities testing large aircraft and hypersonic vehicle models.

Equally important is the framework’s interpretability, a quality often missing from black-box neural approaches. Because the input to the learning model is a set of wavelet bands rather than an undifferentiated signal, researchers can compare which bands the model relies on and how the input-output relationship behaves band by band. The authors describe this as feature-level interpretability through wavelet-band representation and input-output comparison. In an engineering discipline where results feed directly into flight-vehicle design decisions, being able to inspect why a model produced a particular force estimate is not a luxury. It allows test engineers to verify that the network is learning genuine aerodynamic content rather than memorizing artifacts of a particular tunnel, balance, or model configuration.

The validation strategy is notable for its breadth. The framework was tested first in numerical simulation, where the true aerodynamic force is known exactly and identification error can be quantified without ambiguity. It then progressed to bench-scale experiments, which isolate the structural dynamics of the measurement system under controlled conditions. Finally, it was demonstrated in full-scale hypersonic wind-tunnel tests, the environment where all the complicating factors, start-up shock transients, structural ringing, and heavy-model low-frequency vibration, coexist exactly as they do in real campaigns. Passing through all three stages gives the results a credibility that simulation-only studies cannot claim, and it demonstrates that the method survives contact with the messy reality of operational test facilities.

The work builds on a growing body of research into intelligent force measurement in shock tunnels. Earlier efforts explored empirical mode decomposition with soft thresholding and transfer learning, wave-system fitting for long-duration shock tunnels, and calibration methods based on transient response analysis. Deep networks have been applied to strain-gauge balance measurements, and convolutional architectures have been used to predict aerodynamic flow fields more broadly. What distinguishes the present approach is the deliberate marriage of wavelet physics with learning architecture: the decomposition is not a preprocessing afterthought but a structural component of the model’s input representation, chosen specifically because it mirrors how vibration contamination and aerodynamic content are separated in frequency.

The implications extend beyond hypersonics. Any measurement scenario in which a sensor records a mixture of a desired quasi-static quantity and structural vibration, from aircraft model testing in pulse tunnels to dynamic force identification in mechanical systems, faces the same frequency-overlap dilemma. A framework that separates signals by physically meaningful bands, learns the mapping from contaminated measurement to true force, and explains its reasoning at the band level offers a template that other instrumentation communities can adapt. The authors also note that the method is designed for short-duration quasi-static identification under start-up conditions, meaning it targets precisely the window of data that engineers previously had to discard.

Funding for the research came from the National Natural Science Foundation of China and the Open Fund of the National Key Laboratory of Aerospace Physics in Fluids, reflecting the strategic priority that hypersonic flight research now occupies in aerospace programs worldwide. As vehicles flying at Mach 5 and beyond move from concept to hardware, every percentage point of uncertainty removed from ground-test force measurements translates into lighter structures, more efficient trajectories, and safer flight envelopes. Methods like this one, which squeeze reliable data out of the noisiest moments of a tunnel run, may prove to be quiet enablers of the hypersonic era, turning vibration-contaminated signals that once had to be thrown away into trustworthy aerodynamic knowledge.

Subject of Research: A wavelet-based deep learning method for identifying aerodynamic forces from vibration-contaminated measurements in hypersonic wind tunnel tests of large-scale heavy-duty models

Article Title: A Wavelet-Based Short-Time Aerodynamic Identification Method with Band-Level Interpretability for Large-Scale Heavy-Duty Models in Hypersonic Wind Tunnels

Article References: Xu, H., Li, S., Gao, H., Zhu, Z., & Qian, H. (2026). A Wavelet-Based Short-Time Aerodynamic Identification Method with Band-Level Interpretability for Large-Scale Heavy-Duty Models in Hypersonic Wind Tunnels. International Journal of Aeronautical and Space Sciences. https://doi.org/10.1007/s42405-026-01293-6

Image Credits: AI Generated

DOI: 10.1007/s42405-026-01293-6

Keywords: hypersonic wind tunnel, aerodynamic-force identification, wavelet decomposition, deep learning, force measurement, signal processing, large-scale models, structural vibration, interpretability, strain-gauge balance, shock tunnel, machine learning