Efficient Tool Wear Prediction in Manufacturing: BiLPReS Hybrid Model with Performer Encoder

Abstract

Predictive maintenance in industrial settings, especially tool wear prediction, remains crucial for operational efficiency and cost reduction. This paper proposes BiLPReS, a novel predictive model leveraging a hybrid architecture integrating bidirectional long short-term memory, Performer encoder, and residual-skip connections. Compared to convolutional and recurrent neural networks, the proposed model achieves long-range dependent global sensing and parallel computing. The Performer encoder reduces the computational complexity by the FAVOR + approach compared to the Transformer encoder. Moreover, the proposed model includes residual-skip connections to enhance information flow efficiency and minimize the risk of information loss during training. The final use of the fully connected layer reduces dimensionality and generates the predicted values. Experiments on the PHM2010 dataset involve the analysis of multichannel sensor signals, including force, acceleration, and acoustic emission. The model undergoes training and validation through k-fold cross-validation. Results unequivocally demonstrate the model’s high accuracy. Furthermore, conducting comparative experiments by selectively reducing modules validates the effectiveness of the utilized modules in enhancing the model’s performance. This study provides a viable solution for optimizing maintenance schedules, reducing downtime, and real-time monitoring of tool machining.