Selection refines diagnosis: Mamba for acoustic weak fault diagnosis combining feature mode decomposition and selection

Weak fault diagnosis is of great significance in modern industrial production, but have been facing numerous challenges due to the weak nature of the fault feature, noise interference and variations in operating conditions. This paper proposes a Mamba model combining feature mode decomposition (FMD) and selection, aimed at enhancing the performance of acoustic weak fault diagnosis. First, the fault diagnosis network is designed with improved Mamba structure as backbone. Secondly, by combining Mamba's parameterization of input signals with FMD, a Joint Temporal and Component Selection State Space (JTCS3) is proposed, which allows the model to selectively and specifically process different components of the input. Finally, a loss function focusing on weak feature is designed to adjust the feature distribution in the middle part of the network in order to avoid the degradation and annihilation of weak feature components in the training process. The proposed method is validated on an engine dataset and an induction motor dataset. The results show that compared to other advanced methods, the proposed model demonstrates significant advantages in the accuracy, crossing operating conditions and anti-noise performance of weak fault diagnosis.