Remaining Useful Life Prediction of End Mills Using DCNN-McBiLSTM-LRSA with Multi-Source Sensory Signals

In precision mold manufacturing, the machining of HRC52 hardened steel causes severe tool wear and high noise in multi-source sensor signals, making accurate remaining useful life (RUL) prognostics challenging. To address this, we propose a hybrid model based on a two-stage VB-to-RUL estimation strategy. The network first performs high-fidelity flank wear (VB) trajectory tracking; the RUL is then deduced via threshold mapping. The model integrates three components: a one-dimensional deep convolutional neural network (DCNN), a low-resolution self-attention (LRSA) module with 1D-to-2D spatiotemporal reconstruction, and a multi-channel bidirectional long short-term memory network (McBiLSTM). A Gaussian smoothing filter is first applied to denoise the 50 kHz signals, followed by physical-period sliding windows for feature extraction. A multi-strategy fusion pooling layer (mean, max, and last-quarter features) further improves prediction accuracy. Using the PHM 2010 milling cutter dataset under leave-one-out cross-validation, the proposed model achieves a mean absolute percentage error (MAPE) of 1.45% and a root mean square error (RMSE) of 2.76 μm, reducing prediction error by up to 75.6% compared to Transformer, LSTM, and GRU baselines. These results demonstrate that the model effectively extracts degradation features even during the accelerated wear stage, providing a potential solution for tool health monitoring and predictive maintenance under complex cutting conditions.