Acoustic Emission-Driven Explainable Temporal Convolutional Network for Continuous Abrasive Belt Condition Monitoring

Abrasive belt wear condition influences machining quality and process stability in grinding operations. Acoustic emission (AE) provides informative time-series data reflecting the dynamic evolution of wear. This paper proposes an explainable and statistically validated temporal convolutional network (TCN)-based approach for abrasive belt condition monitoring using AE. Multiple features are extracted, including root mean square, counts, energy and average power for cumulative energy levels, and centroid frequency for frequency-domain properties. Statistical correlation analysis combined with SHAP-based interpretability identifies counts and RMS as dominant predictive indicators, with threshold, centroid frequency and amplitude providing secondary contributions. The designed TCN captures temporal dependencies in wear processes through dilated convolutions and residual connections, while maintaining low computational cost and short inference latency. Experimental evaluation with cross-validation demonstrates that the proposed method achieves an R² of approximately 0.91, with performance variability reported as mean ± standard deviation. These results highlight both the predictive performance and interpretability of the approach, while its computational efficiency suggests promising potential for practical deployment in online belt wear monitoring.