Few-shot fault diagnosis of rolling bearings based on MAML–CNN–GRU model

In recent years, many diagnosis methods in few-shot fault diagnosis have achieved remarkable results on known faults with limited samples. But in actual industrial conditions, new fault types often emerge during long-term equipment operation. These methods require retraining with new samples, thus failing to meet rapid diagnosis needs. To address this, this paper proposes MAML–CNN–GRU, a novel few-shot model based on the Model-Agnostic Meta-Learning (MAML) framework. First, the CNN–GRU hybrid model can effectively extract the spatiotemporal features from original time-series fault signals. Specifically, it uses convolutional neural network (CNN) to extract local spatial features and gated recurrent unit (GRU) to capture long-term sequence dependencies. Second, leveraging MAML’s meta-training mechanism, the initial parameters of CNN–GRU can be optimized through multiple fault diagnosis tasks. As a result, it gains the ability to learn cross-task general features. Then, a reasonable meta-task generation strategy enables rapid identification of novel fault types or variable condition faults even with limited samples. Ultimately, results from CWRU and XJTU bearing datasets illustrate this method’s outstanding performance in diagnosing novel faults under diverse operating conditions. For unknown or compound faults, this method achieves 98.91% and 96.96% diagnosis accuracy in 3-way 5-shot and 5-way 5-shot tasks, respectively. Evidently, these results validate the method’s effectiveness.