Detection of anomalies in piezoelectric signals for structural health monitoring using machine learning

Piezoelectric sensors can be used on the smart structure for structure health monitoring in real time. Piezoelectric sensor captures the low frequency vibration signals of the structure. In active vibration control, this signal is sent to the processor to generate anti vibration signal to be sent to piezoelectric actuator. Generally, control law is unable to handle any unknown frequency signal called as anomaly. In this paper, a synthetic sinusoidal damping anomaly signal of frequency 15 Hz is added to actual experimentally acquired sensor signal of frequency 17 Hz. Binary labled dataset is created with actual piezoelectric sensor signal and synthetic anomalous signal. Machine learning based techniques (RF, DT, SVM, NN) are used for classification of signals. To ensure reliable evaluation and avoid data leakage, a group-wise cross-validation strategy is adopted, and performance is reported as mean values across folds. The results demonstrate that the DT model achieves the highest performance with an accuracy of 100% and ROC–AUC of 1.00, reflecting the high separability of the controlled anomaly. The RF model achieves an accuracy of 98–99% with an ROC–AUC of approximately 0.99, while the SVM model yields an accuracy of 95–97% and ROC–AUC of 0.96–0.98. The NN model achieves an accuracy of 94–96% with an ROC–AUC of 0.95–0.97. Additional robustness analysis under noise, clipping, and transient disturbances shows gradual degradation in performance, confirming the stability of the proposed framework. The proposed approach is a proof-of-concept using synthetic anomaly signal with small dataset, with future work focusing on more realistic anomalies and larger datasets for improved generalization.