Prior knowledge-informed multi-task dynamic learning for few-shot machinery fault diagnosis

The lack of machine fault data renders few-shot fault diagnosis important in modern industrial applications. Small sample sizes are sufficient to train a good model if prior diagnosis knowledge is incorporated into the learning process. This study presents a prior knowledge-informed multi-task dynamic learning model for few-shot fault diagnosis, which consists of a main task for fault identification and an auxiliary task for prior knowledge learning. Our framework starts with a selection of signal indicators as the source of prior knowledge based on domain knowledge. After that, a dynamic penalty on knowledge inconsistency, which is adjustable based on the training requirements, is designed for the auxiliary prior knowledge learning task. The knowledge learned in the auxiliary task is then utilized by the main task through a dedicated shared network structure, and thus the auxiliary information reduces the demand for training data in the main task. The proposed framework is applied to two fault diagnosis cases with small sample sizes. We demonstrate that compared with state-of-the-art methods, the proposed solution achieves an identification accuracy of up to 0.9803 with only 20 fault training samples.