A novel method for signal labeling and precise location in a variable parameter milling process based on the stacked-BiLSTM-CRF and FLOSS

Abstract

Unlabeled time series signals collected during manufacturing typically have low value density and must be labeled and intercepted according to the specific application scenario. During variable-parameter milling, particularly high-precision machining, machining parameters vary, and associated discrepancies in vibration signals are small. In this scenario, signal features that are extracted by hand or via deep learning methods cannot typically distinguish machining states via classification models. To solve this problem, a sequence labeling model developed using a stacked bidirectional long short-term memory network with a conditional random field layer (stacked-BiLSTM-CRF) is proposed in this study to automatically label and intercept vibration signals. The stacked BiLSTM receives the shallow features obtained by the short-time Fourier transform of the vibration signals and then outputs the extracted deep features to capture the before and after dependence of the signals. The stacked BiLSTM is then extended by stacking a CRF layer to explicitly model the dependence of signal labels. In a more accurate labeling scenario, the fast low-cost online semantic segmentation algorithm (FLOSS) is used to acquire more fine-grained signal boundary locations after obtaining the frame-level signal label using the stacked BiLSTM-CRF model. In addition, to evaluate model performance, a novel evaluation index for signal labeling is proposed. The feasibility and effectiveness of the proposed method are verified using the vibration signals collected from variable parameter cutting experiments, and results show that the proposed model achieves the best labeling performance of tested methods in nearly all scenarios.