Adaptive fault diagnosis of machining processes enabled by hybrid deep learning and incremental transfer learning

Abstract

Faults occurring during machining processes can severely impact productivity and product quality. Deep learning models have been actively used to develop fault diagnosis approaches. However, it is challenging for industries to adopt the approaches due to their inability to adapt to varying machining conditions. To address the issue, a novel diagnostic approach is designed based on a hybrid convolutional neural network (CNN)-long short-term memory (LSTM) model and an incremental transfer learning strategy. Based on the incremental transfer learning, the CNN-LSTM model can acquire knowledge from previous machining conditions (source domain) and effectively apply it to new conditions (target domain). In the diagnostic approach, instance-based transfer learning, knowledge-based transfer learning, and incremental transfer learning are combined to improve the training efficiency and overcome the issue of forgetting previously learned knowledge. The CNN-LSTM-attention model is designed as a supplementary model when the data complexity is high. Experimental results show that the approach increased the average training accuracy from 88.63 % to 97.10 %, and required training datasets were reduced by 96.97 %. In addition, the incremental transfer learning reduced false detections for 71.24 %.