Abstract
Electro-Mechanical Actuators (EMAs) are being adopted in small-aircraft primary flight-surface control due to their compactness and efficiency; however, their mechanical transmission components remain vulnerable to progressive degradation under variable loads and harsh operating conditions. This study proposes an explainable machine learning framework for health-state diagnosis of aircraft EMAs using multi-sensor dynamic responses acquired from the EU-H2020 REPRISE endurance campaign dataset. The approach is aligned with applied mechanics by exploiting degradation-sensitive dynamic signatures extracted from both electrical and mechanical domains, including three-phase motor currents, load and temperature measurements, and vibration-related indicators derived from position signals via numerical differentiation. To ensure interpretability and robust feature relevance assessment, SHapley Additive exPlanations (SHAP) is employed to quantify the contribution of each sensor variable to the diagnostic decision process and to identify the most informative features. Based on SHAP-driven feature ranking, the diagnostic model is refined to primarily utilize current- and vibration-related features, which exhibit the highest sensitivity to friction evolution, load-dependent nonlinearities, and transmission wear. Four machine learning classifiers—k-Nearest Neighbors (kNN), Support Vector Machine (SVM), Random Forest (RF), and a Deep Neural Network (DNN)—are trained and used. Comparative evaluation demonstrates that RF and DNN achieve superior classification accuracy, reaching CA = 96.4% and CA = 97.8%, respectively, while SVM and kNN yield competitive performance with CA = 94.1% and CA = 91.6%. The results confirm that explainable AI enhances diagnostic reliability and provides mechanics-consistent insights into the degradation process that support the development of interpretable and scalable health monitoring strategies for next-generation aerospace electromechanical systems.
1 Introduction
The modern mechanical design is quickly being changed by the aerospace, energy, and advanced manufacturing industries through the implementation of more efficient, smaller and more controllable systems through electro-mechanical and intelligent engineering systems (; ). These systems are becoming more of a component of safety-critical platforms, and as such, the need to guarantee reliability, availability, and continuity of operations is now a key concern in engineering. As a result, condition monitoring and health-state diagnosis have significantly become a crucial means of reducing unexpected failures, aiding in better maintenance scheduling, and helping in maintaining sustainable operation of a system (; Zanelli et al., 2026).
Within this context, the convergence of artificial intelligence, data-driven modeling, and next-generation digital engineering has created new opportunities for intelligent diagnostics and predictive analytics (). The field of the current work is consistent with the objectives of the Research Topic with title Intelligent Engineering Systems: AI/ML, Advanced Materials, and Next-Generation Manufacturing because explainable AI and multi-sensor monitoring are important facilitators of scalable implementation of Industrial AI. Specifically, interpretable machine learning techniques can be used to close the gap between state-of-the-art classification models with respect to high performance and the creation of engineering trustworthiness, which is a significant obstacle to their implementation in real aerospace and electromechanical systems.
From an applied mechanics perspective, aircraft electro-mechanical actuators exhibit complex dynamic responses governed by coupled electrical, mechanical, and thermal effects, especially under variable load conditions (; ; ). The nonlinearities of the degree of the ballscrew transmission, loss in the lubricant and the development of friction add nonlinearities in the forms of observable phase currents and kinematic motion indications. So position-based vibration and acceleration indicators along with electrical current signatures represent physically meaningful diagnostic characteristics which can be utilized by explainable machine learning frameworks to report both high-accuracy and mechanics-consistent explainable health states.
Based on the recent research, it can be seen that UAV-based monitoring and intelligent sensing systems have attained high rates of real-time perception, optimization, and data acquisition. Nonetheless, the majority of the available literature is either confined by one or more of the following limitations (i) they are mainly applications-driven (e.g., disaster responsiveness, surveillance, environmental monitoring): instead of engineering-condition monitoring of safety-pertinent actuators (ii) many of them use black-box Deep Learning with low interpretability and low explainability of feature effects (iii) they do not usually include mechanics-consistent derivation of signals (e.g., extraction of vibration signals by processing kinematic position measurements) to enable physically meaningful diagnostics. These loopholes drive the necessity of explainable, mechanics-based AI structures of aerospace electromechanical systems of operations under variable loads. Table 1 enlists recent literature on the discussed topic.
TABLE 1
| Ref. | Approach | Key finding | Limitation |
|---|---|---|---|
| Drone-based disaster management analysis (case study framework) | Demonstrated drone’s effectiveness for rapid situational awareness and response coordination during the 2024 Noto Peninsula earthquake | Focuses on operational disaster assessment rather than engineering system health monitoring; no ML-based explainability or condition diagnosis | |
| Drone-assisted last-mile delivery prioritization using mathematical optimization | Proposed a prioritization framework for high-demand post-disaster delivery that improves logistics efficiency | Addresses logistics optimization, not sensor-driven health diagnosis; does not use multi-sensor mechanical/electrical condition indicators | |
| YOLO-based deep learning for drone surveillance (face mask detection) | Achieved real-time detection performance suitable for drone-based monitoring in public spaces | Pure vision-based classification with black-box inference; not relevant to mechanical system degradation or multi-sensor dynamics | |
| Yin et al. (2022) | Design optimization of electric multirotor UAV including uncertainty quantification | Quantified uncertainties in UAV design optimization and proposed mitigation strategies for off-the-shelf components | Focuses on design-stage optimization rather than operational health monitoring; no explainable AI for fault/degradation diagnosis |
| Distributed time-varying optimization for UAV swarm cooperative consensus | Achieved predefined-time cooperative consensus under coupled constraints in UAV swarm systems | Swarm control problem; not related to actuator degradation tracking, multi-sensor signals, or interpretable diagnostics | |
| Drone-based volcanic plume chemistry sensing with ultralight sensors | Demonstrated drones as platforms for in-situ chemical measurements in hazardous environments | Environmental sensing only; no mechanical/electrical condition monitoring or AI explainability framework | |
| Zhang et al. (2022) | Drone monitoring using 6G IoT and deep learning for dead-zone connectivity | Proposed a 6G-IoT architecture with deep learning to improve monitoring coverage and connectivity | Lacks applied-mechanics health indicators and does not address actuator-level diagnostics |
| Drone data acquisition framework for erosion monitoring | Developed a structured data acquisition pipeline for UAV-based erosion monitoring in Mauritius | Data acquisition focused; does not provide interpretable ML health-state diagnosis or modelling | |
| Explainable adaptive anomaly detection for multi-condition intelligent equipment monitoring using sensitivity/correlation analysis and adaptive residual evaluation | Demonstrated improved anomaly detection under varying operating conditions and provided parameter-wise interpretability for equipment monitoring decisions | Focuses on anomaly detection rather than explicit multi-class health-state diagnosis; not centered on aerospace EMAs or coupled electrical–mechanical degradation behavior | |
| SHAP-based feature selection with a hybrid convolutional–recurrent deep learning framework for aero-engine remaining useful life prediction | Showed that SHAP can identify informative features and improve interpretability in aero-engine prognostics while supporting accurate RUL prediction | Addresses prognostics (RUL estimation) rather than discrete health-state classification; limited to aero-engine degradation and not multi-sensor electromechanical actuator diagnosis under variable loads | |
| Explainable bearing fault detection and diagnosis tool for rotating machinery using machine learning and an autoencoder-based health indicator | Provided an interpretable software tool capable of early bearing fault detection and diagnosis under limited-data conditions, with engineering-oriented explainability outputs | Focused on bearing systems and software deployment; does not address aerospace actuation systems, SHAP-based feature relevance, or degradation-sensitive current/vibration coupling in EMAs | |
| Simulation-guided interpretable fault diagnosis of hydraulic directional control valves using multi-pressure signal analysis | Proposed an interpretable diagnosis framework that uses physical simulation to guide feature construction, improving fault identification under limited fault-data conditions | Focuses on hydraulic valves and pressure-signal features rather than electromechanical actuators; does not investigate current-based sensing or EMA health-state classification |
State-of-the-art review on the current advances of the related topic.
To address the previously mentioned gaps, several contributions are called out for in the preparation of the current study, they can be summarized within the following points:
This study will utilize the REPRISE full-scale aircraft EMA endurance and monitoring dataset to build an applied-mechanics diagnostic benchmark under variable loads and progressive degradation.
Acceleration and vibration indicators will be derived directly from measured actuator position signals through numerical differentiation, enabling dynamic-response features that reflect friction growth, wear, and nonlinear load-dependent behaviour.
SHAP analysis will be applied to identify the most influential features, followed by health-state diagnosis using multiple AI models (kNN, SVM, RF, and DNN) with full comparative evaluation of classification accuracy and robustness.
2 Experimental work
This study will utilize an open-access experimental dataset acquired from a full-scale aircraft Electro-Mechanical Actuator (EMA) developed within the EU-H2020 REPRISE project (). The experimental data was obtained through a dedicated endurance and monitoring test bench that aimed at simulating realistic main-flight-surface actuation operations, heroin controlled axial loading, controlled temperature and closed-loop position control. It contains multi-sensor time-series data consisting of three-phase motor currents, actuator position signals (both internal (LVDT) and external (optical) encoders), load-cell force and internal tempera-temperature, which are sampled with multiple sampling rates up to 4800 Hz. They are given two complementary subsets, (i) an 11-session health monitoring data set, which was collected over a few weeks to observe progressive mechanical degradation (primarily in the ballscrew transmission), under varying lubricant conditions; and (ii) a healthy-condition closed-loop dynamic data set of sinusoidal frequency sweeps at different amplitudes to construct Bode diagrams and identify a nonlinear system. The combination of these two factors allows the dataset to be well-suited to explainable machine learning-based condition monitoring and to permit a sound diagnosis of a health-state at variable loads and dynamic operating regimes.
Figure 1 summarizes the experimental dataset utilized in this study, highlighting the two complementary data sources acquired from the REPRISE EMA test bench: (i) the degradation monitoring dataset collected across multiple sessions to capture progressive wear evolution, and (ii) the healthy closed-loop dynamic dataset used to characterize frequency-domain behaviour via Bode analysis. The figure also illustrates the key measured variables used for health-state diagnosis, including three-phase motor currents, actuator position signals from internal and external sensors, load-cell force, and internal temperature, which together provide a rich multi-sensor representation of the EMA’s dynamic response under variable loading conditions.
FIGURE 1
Table 2 identifies the discrete operating states that were used in this research when health-state diagnosing the EMA. The derivation of the classes is made directly out of the test campaign structure, in which a gradual degradation was artificially forced by gradually diminishing the lubricant concentration in the ballscrew-nut transmission. These conditions are realistic mechanical health conditions of the aerospace EMAs and allow supervised learning models to differentiate normal operation and progression of degradation with respect to multi-sensor dynamic responses. Table 3 below is a summary of all the raw physical variables that are present in the monitoring dataset and utilized in this work to form the basis multi-sensor feature pool. These scales include the actuator electrical behavior, mechanical performance, load status, and thermodynamic conditions, which allow a healthy diagnosis with different operating conditions.
TABLE 2
| Class ID | Health state label | Lubrication condition | Physical interpretation |
|---|---|---|---|
| 0 | Healthy | Normal lubricant | Nominal mechanical transmission with stable friction and minimal wear |
| 1 | Degrading – Stage 1 | Half lubricant removed | Increased friction and early degradation symptoms |
| 2 | Degrading – Stage 2 | No lubricant | Severe friction, accelerated wear, and high-risk pre-failure behaviour |
Operating health states considered for machine learning classification.
TABLE 3
| # | Physical variable | Stored variable name | Sampling frequency (Hz) | Unit |
|---|---|---|---|---|
| 1 | Linear motor load reference | load_cell_measure | 100 | N |
| 2 | Temperature inside EMA box | Temperature | 100 | °C |
| 3 | Current supplied to EMA from power supply | linear_motor_supplied_current | 100 | A |
| 4 | Load cell measured force | load_cell_measure | 100 | N |
| 5 | EMA position reference (cRIO) | EMA_Position_reference_cRIO | 100 | mm |
| 6 | EMA position (LVDT sensor) | EMA_LVDT_position | 100 | mm |
| 7 | EMA position (Renishaw optical encoder) | absolute_encoder_biss | 100 | mm |
| 8 | EMA position reference (cDAQ) | EMA_Position_reference_cDAQ | 500 | mm |
| 9 | Linear motor drive current | linear_motor_drive_current | 500 | A |
| 10 | EMA phase current A | phase_A | 4800 | A |
| 11 | EMA phase current B | phase_B | 4800 | A |
| 12 | EMA phase current C | phase_C | 4800 | A |
Raw sensor variables used as input features (full dataset variables).
Vibration signals are of utmost importance, especially when it comes to dynamic response calculations (; ). This study adopts the methodology of transforming vibration signals from position records. To generate vibration-related indicators from the measured position channels, the discrete position signals were numerically differentiated in the time domain. Let denote the position value at the -th sample and let the sampling interval corresponding to the relevant sensor. The first derivative was used to estimate velocity, while the second derivative was used to estimate acceleration. In the present work, the discrete velocity was calculated as:for the internal samples, while forward and backward differences were used at the signal boundaries:
The discrete acceleration was then obtained by differentiating the velocity signal, or equivalently by applying the second-order central difference directly to the position signal:where represents the acceleration at sample . This procedure was applied to the position-based channels to generate acceleration-related features that reflect the dynamic response of the actuator. No additional filtering stage was introduced in this study, and the signals were analyzed as provided in the published experimental dataset. In addition, each signal channel was treated according to its original recorded sampling interval, and the acceleration-related quantities were derived from the corresponding position channels using their native values. This preserves the original experimental structure of the dataset while enabling extraction of mechanics-related dynamic indicators.
3 Machine learning
Artificial Intelligence (AI) has become a major tool across diverse domains, which offers scalable solutions for complex decision-making tasks (; ; ; ; ; ; ; ). AI encompasses a broad range of computational paradigms designed to emulate intelligent behavior, including rule-based systems, optimization algorithms, and data-driven approaches (; ; ; ; ; ; ). Machine learning (ML) represents a core subset of AI in which models learn patterns and relationships directly from data without explicit programming (; ; ; ; ). AI- and statistical ML-based techniques have been successfully applied across diverse domains, including engineering systems (), healthcare, energy management (), autonomous platforms, and industrial process optimization (; ; ; ).
The primary hyperparameter values that were taken to train the four classification models that were used to diagnose EMA health-states are reported in Table 4. In the case of kNN baseline, it had been chosen that a small neighborhood size (k = 5) with the Euclidean distance and distance-weighted voting will result in stable local decision boundaries following z-score normalization. An RBF kernel (C = 10, g = 0.1) was used to set up the SVM in a one-vs-one approach facilitating an opportunity to separate the three levels of degradation nonlinearly, with controlled regularization. The model used was a random forest (with a relatively large ensemble size 300 trees) using bootstrap sampling and splitting the data with a random number which makes the model more resistant to noise and to the interactions between features that are nonlinear (i.e., features that are related to the electromechanical degradation process). Lastly, the DNN was trained as a fully connected (128-64-32) architecture with ReLU activation and Softmax output to classify into three classes with Adam (learning rate = 0.001) and dropout and L2 regularization to avoid over-fitting. Generally, the hyperparameter settings were made to assure a fair level of comparison between models but with the same preprocessing (standardization) and constant training behaviour with multi-sensor, mechanics-driven input features.
TABLE 4
| Model | Training setting | Value |
|---|---|---|
| k-Nearest Neighbors (kNN) | Number of neighbors (k) | 5 |
| Distance metric | Euclidean | |
| Weighting scheme | Distance-weighted | |
| Search method | Auto (KD-tree/brute) | |
| Feature scaling | Standardization (z-score) | |
| Decision type | Majority voting | |
| Support Vector Machine (SVM) | Kernel type | Radial Basis Function (RBF) |
| Regularization parameter (C) | 10 | |
| Kernel coefficient (γ) | 0.1 | |
| Multi-class strategy | One-vs-One (OvO) | |
| Feature scaling | Standardization (z-score) | |
| Stopping tolerance | 1 × 10−3 | |
| Random Forest (RF) | Number of trees | 300 |
| Maximum tree depth | None (fully grown) | |
| Split criterion | Gini impurity | |
| Max features per split | √(Number of features) | |
| Minimum samples per leaf | 1 | |
| Bootstrap sampling | Enabled | |
| Random seed | Fixed (for reproducibility) | |
| Deep Neural Network (DNN) | Network type | Fully connected feed-forward |
| Input layer | 15 features | |
| Hidden layers | 3 layers (128–64–32 neurons) | |
| Activation function | ReLU | |
| Output layer | Softmax (3 classes) | |
| Loss function | Categorical cross-entropy | |
| Optimizer | Adam | |
| Learning rate | 0.001 | |
| Batch size | 64 | |
| Epochs | 150 | |
| Regularization | Dropout = 0.2 + L2 = 1 × 10−4 | |
| Early stopping | Patience = 15 epochs | |
| Feature scaling | Standardization (z-score) |
Hyperparameters used in training the four-selected models.
The hyperparameter values reported in Table 4 were selected through controlled comparative tuning in the training stage, where several candidate settings were examined for each model and the final configuration was chosen based on the best validation performance while maintaining model stability and avoiding overfitting. This procedure was applied consistently across all classifiers to ensure a fair comparison under the same preprocessing and feature-input conditions. To avoid temporal data leakage, the train–test split was performed in a chronological manner, such that earlier segments of the degradation record were used for training while later unseen segments were reserved for testing. This was necessary in order to preserve the natural temporal progression of the EMA health states.
Table 5 summarizes the evaluation metrics adopted in this study to assess the classification performance of the developed EMA health-state diagnosis models. Accuracy (AC) is used to quantify the overall correctness of predictions across all operating states, while Precision and Recall provide class-sensitive measures reflecting false-alarm tendency (FP) and missed-detection tendency (FN), respectively. The F1-score combines Precision and Recall into a single balanced indicator, making it particularly suitable for multi-class health monitoring where reliable discrimination between healthy and degradation stages is required.
TABLE 5
| Parameter | Value |
|---|---|
| Accuracy (AC) | |
| Precision (Prec) | |
| Recall | |
| F1 Score |
Evaluation metrics.
For model interpretability, SHAP analysis was implemented using the model-specific formulation appropriate for each classifier. In particular, TreeSHAP was used for the Random Forest model because of its computational efficiency and exact tree-based attribution capability, whereas the kernel-based SHAP approximation was used to obtain post hoc feature attributions for the non-tree classifiers. The background dataset used for SHAP evaluation was selected from the training subset only in order to avoid information leakage, and it was constructed from a representative sample covering the three health states. In this way, the SHAP values reflect the contribution of each input feature relative to a baseline that is consistent with the training distribution.
4 Results and discussion
4.1 Electrical-vibrational signals results
This subsection presents the measured EMA position, derived vibration (acceleration), and three-phase electrical current signals to highlight how the actuator’s dynamic response evolves under different operating conditions and degradation stages.
Figure 2 shows the set point position reference and the actual position of the actuator that has been measured by an internal LVDT sensor and a Renishaw optical encoder. In the chosen cases, the measured traces are closely following the sinusoidal reference, whereas minor tracking errors and phase changes are observed with regard to the operating load and the degradation state. The position signal is in millimeters and obeying the applied sinusoidal profile with amplitudes as the tested configurations (5 and 10 mm) and offsets (−10, 0 and +10 mm). The two sensing channels are generally the same but the LVDT and optical encoder show minor differences as a result of sensor properties, alignment, and dynamic lag effects, which become more pronounced with increased friction and mechanical wear occurrence.
FIGURE 2
Figure 3 presents the vibration response derived from the actuator position measurements through numerical differentiation to obtain acceleration. The signals of acceleration are very dynamically sensitive to the EMA mechanical condition, in in which an increased degree of degradation results in more pronounced fluctuations, sharper peaks and stronger high-frequency signals correlated to friction-induced stick-slips and transmission anomalies. Specifically, the acceleration obtained using both the Renishaw and LVDT position channels displays repeatable transient signature at motion reversal (sinusoidal turning points) points, with nonlinearities related to load and backlash ones becoming mechanically amplified. These vibration sensors have good consistency of mechanics in health diagnosis, which is why acceleration-related variables are the most important in SHAP-based importance analysis.
FIGURE 3
Figure 4 shows the electrical response of the EMA in terms of three-phase motor currents (A, B, and C), recorded at high sampling frequency (4800 Hz). The phase currents demonstrate distinct modulation patterns synchronized with the actuator motion, reflecting the torque demand required to overcome axial load, transmission friction, and dynamic inertial forces. As degradation increases, the current signals exhibit higher amplitude demand and more pronounced ripple components, which are consistent with increased mechanical resistance due to lubricant loss and progressive ballscrew wear. This behaviour confirms the strong coupling between mechanical degradation and electrical signatures in electromechanical actuation systems, supporting the use of phase currents as highly informative diagnostic features (ranked second after vibration indicators in the SHAP analysis).
FIGURE 4
On the whole, the two-fold trains of the trends prove the fact that the process of EMA degradation is imprinted on the performance of kinematic tracking as it is highly expressed in the dynamic response that is provided by vibration/acceleration indicators and phase-current demand. This is entirely predictable by the principles of applied mechanics, in which the loss of lubricants and wear on ballscrews enhance friction, alter the stiffness and damping properties, and provide nonlinear behaviour (stick-slip and load-dependent hysteresis), which directly doubles the acceleration peak values and electrical torque needs. Therefore, explainable AI-based condition monitoring is offered using a combination of vibration-derived features and electrical current sensors, which gives the idea of physically meaningful explanation.
Figure 5 presents the frequency-domain/statistical summary of the acceleration-related features computed over successive 100-sample intervals, which makes it possible to track how the vibration response changes locally across the signal rather than relying on one global value. In Figure 5a, the peak-to-peak values show clear interval-dependent variation for all three acceleration channels. For the acceleration reference, the maximum peak-to-peak value was 15.286 in the 901–1000 interval, while the minimum was 0.000 in the final 2501–2501 interval. For the Renishaw-derived acceleration, the maximum peak-to-peak value reached 14.644 also in the 901–1000 interval, whereas the minimum was 0.000 in the 2501–2501 interval. Similarly, the LVDT-derived acceleration attained its maximum peak-to-peak value of 14.963 in the 1901–2000 interval and its minimum of 0.000 in the terminal 2501–2501 interval. These results indicate that the vibration amplitude is not constant across the record, and that certain windows contain markedly stronger dynamic excursions than others. In Figure 5b, the RMS feature further confirms this interval-wise variation in signal energy. The acceleration reference exhibited a maximum RMS value of 49.146 in the 401–500 interval and a minimum of 4.465 in the 901–1000 interval. For the Renishaw-derived acceleration, the maximum RMS value was 47.164 in the 2501–2501 interval, while the minimum value was 5.245 in the 901–1000 interval. The LVDT-derived acceleration showed a maximum RMS of 48.106 in the 401–500 interval and a minimum of 4.959 in the 901–1000 interval. Overall, the interval-based peak-to-peak and RMS results demonstrate that both amplitude spread and vibration energy vary noticeably from one segment to another.
FIGURE 5
4.2 SHAP results
Figure 6 clearly shows that the diagnostic decision is primarily governed by the mechanics-derived acceleration features and the electrical phase-current signals, while the thermal, load-related, and raw position channels exhibit comparatively smaller influence. The highest SHAP contributions were obtained for Acceleration reference (0.360), Acceleration Renishaw (0.338), and Acceleration LVDT (0.312), confirming that dynamic acceleration signatures are the most sensitive indicators of EMA degradation. These were followed by the three-phase current features, namely, Phase current A (0.241), Phase current B (0.221), and Phase current C (0.200), which also showed strong diagnostic relevance due to the electromechanical coupling between motor torque demand and friction growth in the ballscrew transmission. Based on this ranking, the top six highest features—the three acceleration features and the three phase-current features—were selected as the principal inputs progressed into the AI models, since they provided the most discriminative and physically meaningful representation of the EMA health state.
FIGURE 6
In addition to these dominant features, the newly added acceleration-derived statistical descriptors also showed clear diagnostic usefulness and were distributed among the subsequent SHAP ranks. Specifically, RMS (Acceleration reference) = 0.186, Peak-to-peak (Acceleration LVDT) = 0.171, RMS (Acceleration Renishaw) = 0.158, Peak-to-peak (Acceleration reference) = 0.133, RMS (Acceleration LVDT) = 0.112, and Peak-to-peak (Acceleration Renishaw) = 0.097. These added descriptors confirm that both vibration amplitude spread and signal-energy content carry important degradation information, although their overall contribution remained lower than the top six core features. After these, the contribution level decreased gradually for Linear motor drive current (0.140), Current from power supply (0.120), Temperature inside EMA box (0.100), Load cell measured force (0.095), and Linear motor load reference (0.085), while the raw position and reference channels remained the least influential with EMA position (Renishaw encoder) = 0.070, EMA position (LVDT) = 0.065, EMA position reference (cDAQ) = 0.055, and EMA position reference (cRIO) = 0.050. Overall, the SHAP analysis confirms that EMA degradation is captured mainly through amplification of dynamic-response features and changes in electromechanical effort rather than through direct position tracking alone. The six highest features are visualized in clusters within Figure 7 below.
FIGURE 7
As lubricant degradation progresses in the ball-screw transmission, friction becomes more irregular and promotes intermittent stick–slip events, which appear in the derived acceleration signals as localized peaks and sharper transient fluctuations, especially near motion reversal regions. At the same time, the increased mechanical resistance raises the torque demand on the electric motor, which is reflected in the phase-current signals through higher amplitude modulation and more pronounced ripple behavior, thereby, this provides an electromechanically consistent signature of tribological deterioration.
4.3 Machine learning results
This subsection evaluates the classification performance of the tested machine learning models for diagnosing the EMA operating health states based on the extracted multi-sensor electrical–vibrational features.
Table 6 and Figure 8 show a distinct level of performance among the four experimented classifiers, indicating that all models have good discrimination capacity, and advanced learners have excellent robustness. Baseline kNN got a total accuracy of 91.6 which shows good generalization with a relative low generalization in the three-state condition monitoring problem. This SVM got improved performance of 94.1% accuracy and it was consistent with precision (94.2) and recall (94.1) as it showed a better separation of classes. Random Forest (RF) also improved the accuracy up to 96.4 and balanced the precision (96.4) and recall (96.4), which supported the advantage of ensemble learning in the case of nonlinear degradation patterns. The DNN gave the highest performance with 97.8% accuracy, and its precision, recall and F1-score were all equal to 97.8, indicating that it can reproduce complex coupled electromechanical dynamics. All in all, the findings affirm that the proposed explainable feature-based framework allows EMA health-state diagnosis with high accuracy, and the best predictive ability is offered by both the DNN and the RF.
TABLE 6
| Model | Accuracy (AC) % | Precision % | Recall % | F1-score % |
|---|---|---|---|---|
| kNN | 91.6 | 91.6 | 91.6 | 91.6 |
| SVM | 94.1 | 94.2 | 94.1 | 94.1 |
| RF | 96.4 | 96.4 | 96.4 | 96.3 |
| DNN | 97.8 | 97.8 | 97.8 | 97.8 |
Classifications and assessment metrics of the tested models.
FIGURE 8
Figure 9 presents the training-phase confusion matrices for the four classifiers, where each model was trained using 2000 samples per health state (Healthy, Degrading–Stage 1, Degrading–Stage 2). In Figure 9a kNN, the classifier correctly predicted 1840 Healthy, 1830 Stage-1, and 1845 Stage-2 samples, while the remaining misclassifications occurred mainly between the two degrading classes (e.g., Stage-1 → Stage-2 = 85 and Stage-2 → Stage-1 = 95), which indicate that the intermediate and severe degradation stages share overlapping dynamic patterns. In Figure 9b SVM, the correct classifications increased to 1885 Healthy, 1880 Stage-1, and 1881 Stage-2, with fewer confusions overall; the dominant error remained between Healthy and Stage-1 (Healthy → Stage-1 = 70, Stage-1 → Healthy = 110) and it reflects the difficulty of separating early degradation from nominal operation. In Figure 9c, the RF, the diagonal values further improved to 1930 Healthy, 1930 Stage-1, and 1925 Stage-2, and the off-diagonal entries became consistently small, confirming the strong capability of ensemble decision trees to learn nonlinear degradation boundaries. Finally, Figure 9d DNN achieved the highest training separation, correctly classifying 1960 Healthy, 1955 Stage-1, and 1953 Stage-2, with minimal remaining confusion, particularly between Stage-1 and Stage-2 (Stage-1 → Stage-2 = 25, Stage-2 → Stage-1 = 37). Overall, the training confusion matrices confirm that the health-state separability is strongest when using RF and DNN, while the remaining errors are primarily concentrated between the two degrading stages due to their mechanically similar friction-driven dynamic behaviour.
FIGURE 9
The results of the testing phase confusion matrices of all the classifiers on an independent test set of N = 1500 samples (500 per class) are reported in Figure 10. As shown in Figure 10a (kNN), 460 Healthy, 455 Degrading-Stage 1 and 459 Degrading-Stage 2 samples were correctly classified by the model, with the misclassifications being mostly concentrated between Stage 1 and Stage 2 (Stage-1 - Stage-2 = 25, Stage-2 - Stage-1 = 25) and between Healthy and Stage 1 (Healthy - Stage-1 = 22, Stage-1 - Healthy = 20), indicating that there was moderate overlap in In Figure 10b (SVM), the correct prediction was up to 470 Healthy, 472 Stage-1 and 470 Stage-2 with fewer errors with the confusion being mainly between Stage 1 and Healthy (Stage-1 - Healthy = 20) and between Stage 2 and Stage 1 (Stage-2 - Stage-1 = 20). Figure 10c (RF) further refined the diagonal values to 485 Healthy, 480 Stage-1, and 481 Stage-2 and the off-diagonal values are very small, indicating that the ensemble model has a good generalization to unsaw samples. Lastly, Figure 10d (DNN) showed the highest generalization with the correct prediction of 492 Healthy, 489 Stage-1 and 486 Stage-2 as well as minimal confusion with little confusion between Stage 2 and Stage 1 (Stage-2 - Stage-1 = 12) and Stage 1 and Stage 2 (Stage-1 - Stage-2 = 7). By and large, the testing confusion matrices prove that the suggested electrical-vibrational feature set offers excellent generalization to all models, and DNN and RF offer the best and most consistent health-state recognition even in the presence of variable loads.
FIGURE 10
In general, the results presented in the article show that the suggested explainable condition monitoring model is capable of reliably diagnosing the EMA health condition during variable-load operation based on the integration of mechanics-consistent vibration patterns and electromechanical current patterns. The SHAP analysis established that acceleration-based features and three-phase currents are the most likely characteristics that contribute to the separability of the health-state, with both high predictive power and physically instructive physical degradation information. A further comparative analysis between kNN, SVM, RF, and DNN indicated that superior learning models, specifically RF and DNN, reach the best generalization performance, and confusion is negligible when the stages of degradation take place. The findings confirm the practicability of interpretable AI-based monitoring to safety-critical aerospace electromechanical actuation systems and form a scaling base in the next-generation health monitoring systems inline with intelligent engineering systems and applied mechanics.
From a deployment perspective, the proposed framework relies on a compact set of physically interpretable features and standard machine learning models, which makes it computationally feasible for implementation on conventional engineering computing platforms without the need for specialized hardware. This supports practical aerospace application, where explainable diagnostic outputs can assist maintenance personnel in health-state interpretation and can also facilitate future integration into onboard or ground-based condition monitoring architectures.
5 Conclusion
This study presented an explainable machine learning framework for health-state diagnosis of aircraft EMAs operating under variable loads, using a multi-sensor experimental dataset acquired from the EU-H2020 REPRISE endurance campaign. The methodology integrated electrical and mechanical dynamic responses by extracting vibration/acceleration indicators from position signals and combining them with three-phase current measurements to capture degradation-sensitive signatures linked to friction growth and ballscrew wear. Explainability was ensured through SHAP analysis, which quantitatively confirmed that acceleration-related features represent the most influential indicators, followed by phase currents, while thermal, load, and raw position channels contribute less to classification decisions. Comparative evaluation of kNN, SVM, RF, and DNN classifiers demonstrated strong diagnostic performance across all models, with the DNN achieving the highest accuracy (97.8%) and RF providing similarly high reliability (96.4%). In general, the findings affirm that explainable AI can provide accurate and mechanics-consistent condition-based monitoring of aerospace electromechanical systems to justify interpretable and scalable implementation in intelligent engineering systems.
Statements
Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.
Author contributions
IA: Conceptualization, Writing – review and editing. AH: Conceptualization, Writing – review and editing. MA: Conceptualization, Writing – review and editing. AA-K: Conceptualization, Writing – review and editing. LA-H: Formal Analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft.
Funding
The author(s) declared that financial support was not received for this work and/or its publication.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1
AbdulameerA. G.MrahM. M.BazerkanM.Al-HaddadL. A.Al-KarkhiM. I. (2025). Machine learning-driven power prediction in continuous extrusion of pure titanium for enhanced structural resilience under extreme loading. Discov. Mat.5, 7. 10.1007/s43939-024-00175-6
2
AbdulsahibI. A.AtiyahQ. A. (2020). Effects of internal connecting layer properties on the vibrations of double beams at different boundary conditions. J. Mech. Eng. Res. Dev.43, 289–296.
3
AbdulsahibI. A.AtiyahQ. A. (2022). Vibration analysis of a symmetric double-beam with an elastic middle layer at arbitrary boundary conditions. Math. Model. Eng. Problems9, 1136–1142. 10.18280/mmep.090433
4
Al-HaddadL. A.JaberA. (2022). “Applications of machine learning techniques for fault diagnosis of UAVs,” in CEUR Workshop Proceedings, 19–25.
5
Al-HaddadA. A.Al-HaddadL. A.Al-HaddadS. A.JaberA. A.KhanZ. H.RehmanH. Z. U. (2024a). Towards dental diagnostic systems: synergizing wavelet transform with generative adversarial networks for enhanced image data fusion. Comput. Biol. Med.182, 109241. 10.1016/j.compbiomed.2024.109241
6
Al-HaddadL. A.Al-MuslimY. M.HammoodA. S.Al-ZubaidiA. A.KhalilA. M.IbraheemY.et al (2024b). Enhancing building sustainability through aerodynamic shading devices: an integrated design methodology using finite element analysis and optimized neural networks. Asian J. Civ. Eng.25, 4281–4294. 10.1007/s42107-024-01047-3
7
Al-HaddadL. A.GiernackiW.BasemA.KhanZ. H.JaberA. A.Al-HaddadS. A. (2024c). UAV propeller fault diagnosis using deep learning of non-traditional χ2-selected taguchi method-tested lempel–ziv complexity and teager–kaiser energy features. Sci. Rep.14, 18599. 10.1038/s41598-024-69462-9
8
Al-HaddadL. A.JaberA. A.DhahirM. K.NagimH. Y.AlgburiZ. I. (2024d). “Characterization and prediction of femtosecond laser induced tracks in silver-containing zinc phosphate glass,” in CEUR Workshop Proceedings, 10–19.
9
Al-HaddadL. A.JaberA. A.MahdiN. M.Al-HaddadS. A.Al-KarkhiM. I.Al-SharifyZ. T.et al (2024e). Protocol for UAV fault diagnosis using signal processing and machine learning. Star. Protoc.5, 103351. 10.1016/j.xpro.2024.103351
10
Al-HaddadL. A.KahachiH. A. H.Ur RehmanH. Z.Al-ZubaidiA. A.Al-KarkhiM. I.Al-OubaidiB. (2024f). Advancing sustainability in buildings using an integrated aerodynamic façade: potential of artificial intelligence. Terra Joule J.1, 1. 10.64071/3080-5724.1006
11
Al-HaddadS. A.Al-HaddadL. A.JaberA. A. (2024g). Environmental engineering solutions for efficient soil classification in southern Syria: a clustering-correlation extreme learning approach. Int. J. Environ. Sci. Technol.22, 2177–2190. 10.1007/s13762-024-05784-5
12
Al-HaddadL. A.KhalidW.TariqS. Z.MrahM. M.FlahA.TazayA. F.et al (2025a). Reliability-oriented framework for UAV-Based inspection missions in modern power and energy systems. Sci. Rep.16, 958. 10.1038/s41598-025-30410-w
13
Al-HaddadL. A.ŁukaszewiczA.MajdiH.Sh.HolovatyyA.JaberA. A.Al-KarkhiM. I.et al (2025b). Energy consumption and efficiency degradation predictive analysis in unmanned aerial vehicle batteries using deep neural networks. Adv. Sci. Technol. Res. J.19, 21–30. 10.12913/22998624/201346
14
Al-KarkhiM. I.RzadkowskiG.IbraheemL.AqibM. (2024). Anomaly detection in electrical systems using machine learning and statistical analysis. Terra Joule J.1, 3. 10.64071/3080-5724.1012
15
AlaweeW. H.Al-HaddadL. A.BasemA.Al-HaddadA. A. (2024a). A data augmentation approach to enhance breast cancer detection using generative adversarial and artificial neural networks. Open Eng.14, 20240052. 10.1515/eng-2024-0052
16
AlaweeW. H.JaberA. A.OmaraZ. M.MohammedS. A.DhahadH. A.KhanZ. H.et al (2024b). Optimizing water resources for sustainable desalination: the integration of expert systems and solar energy in experimental applications. Desalination Water Treat.320, 100683. 10.1016/j.dwt.2024.100683
17
BunyanS. T.KhanZ. H.Al-HaddadL. A.DhahadH. A.Al-KarkhiM. I.OgailiA. A. F.et al (2025). Intelligent thermal condition monitoring for predictive maintenance of gas turbines using. Mach. Learn.13, 401. 10.3390/machines13050401
18
Di RitoG.LucianoB.BorgarelliN.NardeschiM. (2021). Model-based condition-monitoring and jamming-tolerant control of an electro-mechanical flight actuator with differential ball screws. Actuators10, 230. 10.3390/act10090230
19
FadhilT. H.Al-HaddadL. A.Al-KarkhiM. I. (2025). Legal accountability and UAV fault diagnosis explainable AI in aviation safety and regulatory compliance for liability challenges. Discov. Artif. Intell.5, 410. 10.1007/s44163-025-00690-2
20
GaoZ.ZhuX.FangY.ZhangH. (2017). Active monitoring and vibration control of smart structure aircraft based on FBG sensors and PZT actuators. Aerosp. Sci. Technol.63, 101–109. 10.1016/j.ast.2016.12.027
21
Hadi FadhilT.Al-KarkhiM. I.Al-HaddadL. A. (2025). Legal and communication challenges in smart grid cybersecurity: classification of network resilience under cyber attacks using machine learning. J. Commun.20, 221–228. 10.12720/jcm.20.2.221-228
22
HashimF. A.MohialdenY. M.HussienN. M. (2025). Hybrid feature selection and ensemble classification for climate change indicators: a machine learning approach. Terra Joule J.1, 8. 10.64071/3080-5724.1021
23
HusseinR. M.HadiA. S.HasanS. F.Al-HaddadL. A.HarrabiR.FlahA.et al (2025). Hybrid integral sliding mode and fuzzy logic control for omnidirectional robots: modified elephant herding optimization for trajectory tracking. Sci. Rep.15, 35666. 10.1038/s41598-025-19449-x
24
JafariN.FarsangiM. M. (2026). Physics-informed deep hidden markov model and wasserstein generative Adversarial Networks for hydraulic system condition monitoring. ISA Trans.170, 102–114. 10.1016/j.isatra.2026.01.026
25
JweriA.-R. K.Al-HaddadL. A.OgailiA. A. F.JaberA. A.Al-KarkhiM. I. (2025). Edge-intelligent leak detection in water distribution systems using CatBoost: a sustainable solution for reducing infrastructure losses. Clean Energy Sci. Technol.3, 398. 10.18686/cest398
26
KhanZ. H.MekidS.Al-HaddadL. A.JaberA. A. (2025). AI enabled manufacturing: a deep learning approach to network fault detection. Proc. 2025 4th Int. Conf. Comput. Inf. Technol.2025, 245–250. 10.1109/ICCIT63348.2025.10989388
27
KowalskiR.JuchmannP. (2023). Sensorless motor control for electro-mechanical flight control actuators. CEAS Aeronaut. J.14, 1007–1018. 10.1007/s13272-023-00682-x
28
LiK.ZhangF.SunX.WangJ. (2026). An explainable adaptive anomaly detection method for multi-condition intelligent equipment monitoring. Eksploatacja I Niezawodność – Maintenance Reliab.28. 10.17531/ein/218122
29
LuW. (2024). Optimization and design of electromechanical control automation based on dual motor control algorithm. Front. Mech. Eng.10-2024, 1485041. 10.3389/fmech.2024.1485041
30
MadanR.OgailiA. A. F. (2026). Intelligent fault diagnosis of EV induction motors using multi-sensor ANN and explainable AI for sustainable transportation. Terra Joule J.2, 8. 10.64071/3080-5724.1024
31
MagadánL.Ruiz-CárcelC.GrandaJ. C.SuárezF. J.Menéndez-GonzálezA.StarrA. (2026). BEARING-FDD: an early detection and diagnosis tool for bearing faults in rotating machinery. Softw. Impacts27, 100810. 10.1016/j.simpa.2025.100810
32
MahdiN. M.JassimA. H.AbulqasimS. A.BasemA.OgailiA. A. F.Al-HaddadL. A. (2024). Leak detection and localization in water distribution systems using advanced feature analysis and an artificial neural network. Desalination Water Treat.320, 100685. 10.1016/j.dwt.2024.100685
33
MazzoleniM.ScandellaM.MaccaranaY.PrevidiF.PispolaG.PorziN. (2018). “Condition monitoring of electro-mechanical actuators for aerospace using batch change detection algorithms,” in 2018 IEEE Conference on Control Technology and Applications (CCTA), 1747–1752. 10.1109/CCTA.2018.8511334
34
MazzoleniM.ScandellaM.PrevidiF.PispolaG. (2020). Data on the first endurance activity of a brushless DC motor for aerospace applications. Data Brief.29, 105153. 10.1016/j.dib.2020.105153
35
MazzoleniM.Di RitoG.PrevidiF. (2021). “Fault diagnosis and condition monitoring of aircraft electro-mechanical actuators,” in Electro-Mechanical Actuators for the More Electric Aircraft. Editors MazzoleniM.Di RitoG.PrevidiF. (Cham: Springer International Publishing), 119–224. 10.1007/978-3-030-61799-8_4
36
MejbelB. G.SarowS. A.Al-SharifyM. T.Al-HaddadL. A.OgailiA. A. F.Al-SharifyZ. T. (2024). A data fusion analysis and random forest learning for enhanced control and failure diagnosis in rotating machinery. J. Fail. Analysis Prev.24, 2979–2989. 10.1007/s11668-024-02075-6
37
MemmoloV.MonacoE.RicciF.VaselliC.CimminielloN.SalvatoP. (2022). Structural health monitoring of electromechanical actuators in aviation—challenges ahead and case study. J. Nondestruct. Eval. Diagn. Progn. Eng. Syst.5, 041004. 10.1115/1.4054231
38
NassirL. M.JweriA.-R. K.Al-AmeenE. S.Ogai-LiA. A. F.NjimE. K.Al-MalikyF. T.et al (2026). Integrated experimental, statistical, and finite element analysis of nanoparticle-reinforced polymer composites for advanced structural applications completed with bibliometric analysis. ASEAN J. Sci. Eng. Mater.5, 287–310. Available online at: https://www.scopus.com/inward/record.uri?eid=2-s2.0-105025095966&partnerID=40&md5=12417b32673111d9cd8e2c2acf9ae83a.
39
OuafekN.BerghoutT.AbderrezekS.BourouisA. (2026). SHAP-based feature selection with a hybrid convolutional and recurrent deep learning framework for remaining useful life prediction of aero-engines. Int. J. Progn. Health Manag.17. 10.36001/ijphm.2026.v17i1.4693
40
PrakashV.SubasreeN.ArulV.KumarM. S.DharaniP.RadhakrishnanK.et al (2026). Corrosion to cohesion: surface engineering strategies for durable metal–rubber interfaces. JOM78, 3840–3866. 10.1007/s11837-026-08133-8
41
PrevidiF.MaccaranaY.MazzoleniM.ScandellaM.PispolaG.PorziN. (2018). “Development and experimental testing of a health monitoring system of electro-mechanical actuators for small airplanes,” in 2018 26th Mediterranean Conference on Control and Automation (MED), 673–678. 10.1109/MED.2018.8442734
42
SaffariP. R.BodaghiM.HaN. S.SaffariP. R.SenjuntichaiT. (2026). Size-dependent vibroacoustic behavior of thermally loaded double-walled sandwich microplates with metamaterial honeycomb cores and FGPP facesheets. Acta Mech. 10.1007/s00707-026-04630-8
43
SarowS. A.FlayyihH. A.BazerkanM.Al-HaddadL. A.Al-SharifyZ. T.OgailiA. A. F. (2024). Advancing sustainable renewable energy: XGBoost algorithm for the prediction of water yield in hemispherical solar stills. Discov. Sustain.5, 510. 10.1007/s43621-024-00782-6
44
SomersS. M.GraysonW. L. (2021). Protocol for the use of a novel bioreactor system for hydrated mechanical testing, strained sterile culture, and force of contraction measurement of tissue engineered muscle constructs. Front. Cell. Dev. Biol., 9–2021. Available online at: https://www.frontiersin.org/journals/cell-and-developmental-biology/articles/10.3389/fcell.2021.661036.
45
XiaY.XiaoA.XiaoH.ZhaoX.LiuH. (2026). Simulation-guided interpretable fault diagnosis of hydraulic directional control valves under limited fault data conditions. Sensors26, 2052. 10.3390/s26072052
46
YangH.ZhangL.StansbyP.DraycottS. (2026). Data-driven condition monitoring of mooring systems: a demonstration with the M4 multi-float wave energy converter. Ocean. Eng.351, 124312. 10.1016/j.oceaneng.2026.124312
47
YinZ.HuN.ChenJ.YangY.ShenG. (2022). A review of fault diagnosis, prognosis and health management for aircraft electromechanical actuators. IET Electr. Power Appl.16, 1249–1272. 10.1049/elp2.12225
48
ZanelliF.GalimbertiA.DebattistiN.NegriS.TomasiniG. (2026). A modular and integrated on-board system for freight train condition monitoring: design approach and testing. Comput. Struct.321, 108109. 10.1016/j.compstruc.2026.108109
49
ZhangY.LiuD.MiaoQ.PengY. (2022). Sensing data-based degradation estimation of electromechanical actuator under dynamic operating conditions. IEEE Sens. J.22, 21837–21845. 10.1109/JSEN.2022.3208015
Summary
Keywords
aircraft actuator, condition monitoring, electrical current, machine learning, vibration
Citation
Abdulsahib IA, Hadi AS, Ahmed MK, Al-Khafaji AJD and Al-Haddad LA (2026) Explainable machine learning for condition monitoring of aircraft electromechanical actuators under variable loads: health state diagnosis from multi-sensor dynamic responses. Front. Mech. Eng. 12:1810983. doi: 10.3389/fmech.2026.1810983
Received
13 February 2026
Revised
18 April 2026
Accepted
01 May 2026
Published
20 May 2026
Volume
12 - 2026
Edited by
Bharat Kumar Chigilipalli, Vignan’s Institute of Information Technology (VIIT), India
Reviewed by
Zhen-Wei Zhou, No. 5 Electronics Research Institute of the Ministry of Industry and Information Technology, China
Jiongran Wen, Fudan University, China
Updates
Copyright
© 2026 Abdulsahib, Hadi, Ahmed, Al-Khafaji and Al-Haddad.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Luttfi A. Al-Haddad, Luttfi.a.alhaddad@uotechnology.edu.iq
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.