Deep transfer learning for failure prediction across failure types

With the increasing development of artificial intelligence (AI) technologies, deep learning-driven approaches have been widely applied to predicate different machinery failures. One key challenge of failure prediction is to collect sufficient data, especially data of various failure types, to train the data-driven models. Existing studies focus on using transfer learning to transfer knowledge across machines or domains, but not across failure types. In this study, we hypothesise that knowledge about failure among similar failure types is transferable. Should the hypothesis hold, companies may no longer require a large amount of all types of failure data for predictive maintenance. This will increase the companies' overall implementation feasibility and productivity gains. We tested our hypothesis on knowledge transferability for failure prediction in an experiment performed on rotating machinery with vibration signals. During the experiment, we first calibrated the performance of the trained deep neural network in each impending failure type. Then, we leveraged the architecture and hyperparameters of the neural network model trained from one type of failure as the pre-trained model for knowledge transfer. The pre-trained model is fine-tuned with data from another type of failure of the same machine. After that, we compared the performance of the neural network model to predict the second type of failure before and after knowledge transfer. Results showed that transferring knowledge obtained from one type of failure could vastly improve the performance of predicting another type of failure, which may not have sufficient data to train a good prediction model. This result implies that predictive analytics can apply parameter-based deep transfer learning (TL) to address the challenge of insufficient data on all types of machine failures for failure prediction.