A task-oriented theil index-based meta-learning network with gradient calibration strategy for rotating machinery fault diagnosis with limited samples

In industrial scenarios, rotating machinery operates in harsh environments under complex and variable conditions, which leads to a scarcity of available data. This brings challenges to intelligent model-based rotating machinery fault diagnosis. For this issue, a task-oriented theil index-based meta-learning network with gradient calibration strategy (TTIMN-GCS) is proposed for rotating machinery fault diagnosis with limited samples. Firstly, a fine-grained feature learner (FGFL) is designed to extract high-dimensional fine-grained fault information from limited samples. The FGFL is modeled after the human recognition process of fine-grained objects, enhancing distinguishing between fault categories with subtle differences. Secondly, a task inequality metric named task-oriented theil index is developed to acquire more competitive update rules from limited samples, which creatively frees the initial performance of the meta-FGFL from being overly tied to specific tasks. Finally, a gradient calibration strategy is proposed to adjust the optimization trajectory of TTIMN-GCS, which facilitates the diagnostic model evolution toward robust generalization performance. Four diagnostic cases on several datasets are designed, and the diagnostic accuracies under the 5-shot setting reach 98.18 %, 96.68 %, 94.60 %, and 93.90 %, respectively, which are better than other state-of-the-art methods. Experimental results exhibit that the TTIMN-GCS has a remarkable capability to identify new fault categories from a few samples and is potentially promising for engineering applications.