MS-SSPCANet: A powerful deep learning framework for tool wear prediction

Abstract

Reliable tool condition monitoring (TCM) system becomes increasingly critical to machinery efficiency, workpiece quality, and production cost in modern manufacturing industry. However, existing monitoring models have to face model generalization capacity or other problems, which significantly hampers their further application. Principal component analysis network (PCANet) is a promising interpretable deep learning model, and has been applied in image recognition. Nevertheless, the architecture still needs optimizing. In this paper, a novel multi-scale stacked sparse principal component analysis network (MS-SSPCANet) is proposed to predict tool wear. At first, time–frequency images are generated by short-time Fourier transform (STFT). Then, sparsity, multi-scale filters and multi-scale pooling layer are introduced to optimize PCANet structure. A hybrid pooling layer named GA-SPP, combining global averaging and spatial pyramid pooling layers, is applied to select features. And the selected ones from different filter scales are fused. Finally, support vector regression (SVR) is trained for tool wear prediction. Two milling experiments are conducted, and the vibration signal and corresponding tool wear data are acquired simultaneously. The following performance comparison experiments illustrate the better prediction capacity of the proposed model. Additionally, sparsity and two multi-scale operation (convoluting and pooling) are proven effective, and MS-SSPCANet also outperforms LeNet even after adding −6 dB noise.