Enhancing machine tool predictive maintenance: A dual-model approach integrating improved deep autoencoders and graph attention network

Along with the increasing amount and complication of machine tools, various machine faults appear and the stability of the inherent manufacturing process may be threatened. On the one hand, high-dimensional data is difficult to be effectively used to accurately predict the timing of a fault. On the other hand, even with a rough estimate of the fault time, it is difficult for maintenance personnel to quickly find the root cause of the fault during the stage of predictive maintenance. The fundamental reason is the lack of cognition and reasoning about the correlation of a large amount of underlying knowledge in fault prediction and maintenance. To address this issue, a dual-model driven predictive maintenance approach is proposed by using deep autoencoder and graph attention neural network to enhance the stability of machine tools. Firstly, a system architecture of the proposed dual-model driven predictive maintenance is designed with fault data acquisition of machine tools, fault prediction model driven by CNN-BiLSTM-Autoencoder, and Graph Attention Network-driven maintenance service recommendation, which can be illustrated minutely as follows. Based on the acquires high-dimensional fault data, a CNN-BiLSTM-Autoencoder-based fault prediction model is proposed for machine tools, which can capture the underlying relationships among fault features to calculate accurate fault prediction results. Based on the accurate fault prediction result, an effective maintenance service recommendation approach is proposed based on the improved Graph Attention Network combined with Neural Tensor Networks, which can capture more complex relationships among fault causes and maintenance services. Based on this, maintenance services that well match maintenance requirements for actual faults can be recommended. Finally, comparative experiments are conducted within a machining workshop featuring a diversity of machine tools, which confirm that the proposed approach can exceed traditional methods in terms of fault prediction and maintenance recommendation performance.