BiMamba-AE: Bidirectional State Space Model based Autoencoder for Industrial Fault Diagnosis

This paper investigates a novel anomaly diagnosis model specifically designed for handling one-dimensional time series. The model explores the feasibility and practical effectiveness of integrating state space equations (Mamba) into an autoencoder (AE) architecture, resulting in MAE-Bimamba and VAE-Bimamba. In existing diagnostic research, most methods focus on classification tasks, aiming to enhance the model’s discriminative ability in high-noise environments or achieve crossdomain transfer. However, in practical industrial applications, the labeled datasets required for classification tasks are often unavailable, especially when the components and mechanisms requiring fault diagnosis are high-value or when the consequences of failure are extremely serious. Under such conditions, it is difficult to obtain labeled samples needed for classification, particularly when there is a severe lack of negative samples in the target domain. In this real-world situation, some research has turned to generating synthetic samples, but there are still logical and reliability issues in the recognition architecture of such mechanisms. To address the above difficulties and challenges, this paper proposes introducing Mamba into the MAE and VAE architectures to achieve unsupervised anomaly detection on time series. Experiments were conducted on the publicly available PHM2009 and HTBF datasets. The experiments revealed that the combination of Mamba and MAE led to unstable training due to gradient vanishing. To overcome this problem, Bi-mamba was introduced to replace the Mamba architecture, effectively solving the gradient vanishing issue and improving model stability. Finally, the improved Bi-Mamba architecture outperformed traditional MAE and VAE models, demonstrating the feasibility and effectiveness of the proposed method.