Adaptive Diagnostics on Machining Processes Enabled by Transfer Learning
Authors: Liang, Y., Li, W., Wang, S. and Lu, X.
Publisher: Springer Nature
Faults on machines or cutting tooling during machining processes generate negative impacts on productivity, production quality and scrap rate. Effective diagnostics to identify faults throughout the lifecycle of a machining process adaptively is foremost for achieving overall manufacturing sustainability. In recent years, the research of leveraging deep learning algorithms to develop diagnostics approaches has been actively conducted. However, the approaches have not been widely adopted by industries yet due to their inadaptability of addressing the changing working conditions of customized machining processes. Re-collecting a large amount of data and re-training the approaches for new conditions is significantly time-consuming and expensive. To overcome the limitation, this chapter presents a novel deep transfer learning enabled adaptive diagnostics approach. In the approach, firstly, a Long Short-term Memory-Convolutional Neural Network (LSTM-CNN) is designed to perform diagnostics on machining processes. Then, a transfer learning strategy is incorporated into the LSTM-CNN to enhance the adaptability of the approach on different machining conditions via the following steps: (1) The input datasets from different conditions are optimally aligned to facilitate data reuse between the conditions; (2) The weights of the trained LSTM-CNN are regularized using an improved optimization algorithm to minimize the mismatches of feature distributions of the conditions in implementing cross-domain transfer learning. Based on the steps, the LSTM-CNN based diagnosis trained in one condition can be adaptively applied into new conditions efficiently, and thereby the re-training processes of the LSTM-CNN from scratch can be alleviated. Comparative experiment results indicated that the approach achieved 96% in accuracy, which is significantly higher than other approaches without transfer learning mechanisms.