Advanced Mechanical and Mechatronic Systems

Interpretable Spatial Localisation for Visual Anomaly Detection in Industrial Monitoring Systems: A Cross-Dataset Evaluation Using MVTec AD, VisA, and BTAD

Alimuddin Melleng — 2026-10-19

Visual quality inspection requires anomaly detection using limited defect data and accurate spatial localisation. This study evaluates Patch Distribution Modeling (PaDiM), a one-class framework that generates pixel-level anomaly maps using Mahalanobis distance scoring, across three purposively selected categories from each of the MVTec AD and VisA datasets and all three categories of BTAD. Dataset-level results for MVTec AD and VisA are reported as subset means, whereas the BTAD mean represents the complete three-category benchmark. Using ResNet-18, the mean pixel-level AUROC values were 0.9713 for MVTec AD, 0.9835 for VisA, and 0.9764 for BTAD, while the mean image-level AUROC ranged from 0.8194 to 0.9381. For the MVTec AD bottle category, Wide-ResNet-50-2 improved all four evaluation metrics, achieving an image-level AUROC of 1.0000. Counterfactual masking was performed to assess whether regions with high anomaly scores contributed to the image-level anomaly score. Masking the highest-scoring 5% of pixels reduced the score by 10.98%, compared with a 2.08% increase for a size-matched random control and no change (0.00%) for an inverse control. These findings indicate that the highlighted regions contributed more substantially to the anomaly score than either size-matched control, although their removal did not eliminate the remaining anomaly evidence. For the VisA candle category, a high pixel-level AUROC of 0.9827 was accompanied by a low mean intersection-over-union of 0.091, demonstrating that strong pixel-level anomaly ranking does not necessarily translate into accurate spatial overlap with actual defects. Overall, this study demonstrates the potential of PaDiM for anomaly detection while emphasising the limitations of spatial localisation. Its contribution lies in providing empirically evaluated and explicitly bounded spatial localisation rather than a direct explanation of the model's decision-making process.

Adaptive Proportional–Integral–Derivative (PID) Temperature Control for Industrial Heating Systems under Setpoint Variations and Thermal Disturbances

Nguyen Quang Anh — 2026-10-19

This study proposes an adaptive proportional–integral–derivative (PID) controller for an industrial heating system to improve temperature-tracking performance under varying setpoints and thermal load disturbances. Unlike a conventional PID controller with fixed gains, the proposed method updates the proportional, integral, and derivative gains (K_p, K_i, and K_d) online based on the temperature error and its rate of change. The system was simulated for 6,000 s using multiple temperature setpoints ranging from 50 to 100 °C and three thermal load disturbance intervals, with a conventional PID controller used as the benchmark. Control performance was evaluated using the mean absolute error (MAE), root mean square error (RMSE), integral absolute error (IAE), integral squared error (ISE), overshoot, settling time, and steady-state error. The results showed that the adaptive PID controller reduced MAE by 34.90%, RMSE by 23.86%, IAE by 34.90%, and ISE by 42.03% compared with the conventional PID controller. The average steady-state error decreased from 6.251 to 3.081 °C, corresponding to a 50.72% improvement. The adaptive controller also maintained the temperature closer to the setpoint and demonstrated better recovery from thermal load disturbances. However, improvements in overshoot and settling time were not consistent across all setpoint changes because of thermal inertia and heating-power limitations. Overall, the results demonstrate that a simple error-based gain adaptation strategy can enhance temperature-control performance while retaining the practical and straightforward structure of conventional PID control.

Machine Learning for Classification of Locomotor Performance from Electromyographic and Kinetic Gait Parameters

Maria Zofia Goldasz — 2026-10-19

Gait is a complex motor activity that reflects the interaction between neuromuscular control and mechanical loading. This study evaluated the ability of electromyographic (EMG) and kinetic gait parameters to classify observations according to a study-specific index of locomotor performance. A multimodal gait dataset comprising motion-capture, ground reaction force (GRF), and surface EMG measurements was analyzed. EMG features derived from the gluteus maximus and soleus muscles included root mean square (RMS), integrated EMG (iEMG), median frequency (MDF), and a muscle co-contraction ratio. Kinetic features included peak and mean vertical GRF, force impulse, GRF asymmetry, center-of-pressure (COP) path length, and COP velocity. A study-specific Performance Index was constructed from normalized cadence, GRF, COP stability, and asymmetry components and used to categorize observations into Low, Medium, and High performance classes. A total of 1,143 observations were analyzed using a stratified 70:30 training–testing split. Random Forest and XGBoost were applied for EMG-based classification and feature assessment, whereas Random Forest, XGBoost, Support Vector Machine (SVM), and Nearest Mean were evaluated for kinetic-based classification. SHapley Additive exPlanations (SHAP) were used to interpret model predictions, while Spearman correlation analysis was performed to examine EMG–kinetic relationships. The EMG-based Random Forest and XGBoost models both achieved accuracies exceeding 99%. For kinetic data, SVM achieved the highest accuracy (93.59%), followed by Random Forest and XGBoost (90.38% each), whereas Nearest Mean achieved 70.55%. Soleus iEMG was among the most informative EMG features, while GRF magnitude and force impulse were the dominant kinetic features. The strongest EMG–kinetic association was observed between gluteus maximus RMS and right-limb force impulse (r = −0.421). Overall, the findings demonstrate the complementary value of neuromuscular and kinetic information for the study-specific classification of locomotor performance.