Vibration data fusion and sparsity-driven convolutional network for enhanced tool wear monitoring

This research presents a novel approach for tool condition monitoring that addresses the challenges of overfitting in deep learning and limited accuracy in traditional methods. It employs a weighted multi-sensor fusion technique, which dynamically adjusts the significance of each sensor based on its contribution. Raw vibration data from five sensors is processed to compute a Fast Fourier Transform (FFT). The FFT is input to the convolutional neural network (CNN) and passed to feature extraction layers. Thereafter, a softmax activation layer generates importance scores for each sensor, and the weighted outputs are passed through fully connected layers for final classification. The key innovation of the work is introducing a mechanism to compute the Shannon entropy of weights and regularization to combat overfitting, promoting the sparse activation of filters. The fusion method achieves 99.90% accuracy, significantly outperforming the individual sensors, Sensor 1 (83.03%) and Sensor 2 (60.51%). The multi-sensor fusion approach improves accuracy by 3.4%, surpassing existing state-of-arts methods.