Intelligent tool-wear prediction for ball nose tungsten carbide cutters in milling of stainless steel HRC52 using a genetic algorithm-enhanced bi-directional long short-term memory network

Abstract

Tool wear monitoring is crucial in machining, playing a vital role in ensuring quality and controlling costs. Inadequate control over tool wear and life can result in increased expenses or significant damage to both tools and workpieces, making accurate wear prediction essential to avoid failures. While traditional long short-term memory (LSTM) models perform well on time series data, they capture only unidirectional historical information and fail to utilize future data, limiting their effectiveness in complex wear prediction tasks. Moreover, the hyperparameter tuning process for LSTM models is complex and computationally expensive. Manual tuning methods often struggle to find a global optimum in high-dimensional spaces, leading to local optima and restricting the model's generalization capabilities. To address these limitations, this paper introduces a genetic algorithm-optimized bidirectional long short-term memory (GA-BiLSTM) model. Unlike traditional LSTM, BiLSTM captures both forward and backward time series data, enabling comprehensive utilization of sequence features. The genetic algorithm (GA) performs a global search of the hyperparameter space, automatically optimizing key parameters, thus avoiding the inefficiencies of manual tuning and significantly improving the model’s robustness and performance. Experimental results show that GA-BiLSTM reduces mean absolute error (MAE) by up to 72.0% and root mean square error (RMSE) by 64.3% on the PHM2010 dataset, demonstrating its superior predictive accuracy and practical applicability.

Article Highlights

The Bi-LSTM model is synergistically combined with a genetic optimization algorithm for the first time to predict the wear of Ball Nose Tungsten Carbide Cutters. Experimental results demonstrate superior fitting capabilities compared to alternative models.

The sensor signals are precisely utilized for feature extraction in the time domain, frequency domain, and timefrequency domain. These features are filtered using the Pearson correlation coefficient, and a correlation analysis is conducted on the remaining features.

A global optimization strategy utilizing genetic algorithms is employed to fine-tune the learning rate, number of hidden layers, and training batch size of the Bi-LSTM layer.