Meta-learning-based hybrid LSTM-Transformer framework for cross-condition small sample fault diagnosis

Abstract

Recently, intelligent fault diagnosis has emerged as a widely used method for identifying equipment failures. It leverages advanced feature learning abilities to ensure both efficient and highly accurate fault detection. However, in practical applications, it encounters a significant challenge due to the insufficient availability of fault samples under varying operating conditions, which greatly undermines the accuracy of fault diagnosis. Therefore, a meta-learning-based fault diagnosis method for small sample learning under variable operating conditions is proposed in this article. First, long short-term memory (LSTM) and Transformer are fused into a meta-learner to efficiently capture long-term dependencies within sequential signals and to strengthen the connection between local and global features of temporal signals. Then, the model-agnostic meta-learning (MAML) algorithm is combined to improve the ability of the model to extract small sample fault features under multiple working conditions. Finally, the LSTM transformer with MAML algorithm is validated for effectiveness on our established dataset of road machinery hydraulic system and two publicly available datasets. The results show that in cross-condition small-sample fault diagnosis, the proposed method achieves 98.84% diagnostic accuracy in only 50 cross-condition fault samples. This provides an effective solution for investigating cross-case small-sample diagnostics in real engineering.