A new model with variable correlation feature and local feature fusion for rotating machinery fault diagnosis

Data-driven deep learning techniques have been applied in the field of fault diagnosis. However, most existing deep learning-based fault diagnosis methods require the specification of particular signal data and fixed signal lengths. When dealing with signals of varying variables and sequence lengths, the adjustments of model are often necessary for adaptation. To address this challenge, this study proposes variable correlation feature and local feature fusion fault diagnosis method (VLFFD) for mechanical fault diagnosis under conditions of varying variables and signal lengths. First, a feature fusion network based on multivariable self-attention and grouped query attention is introduced to integrate variable-related features and local fault features. Second, a one-dimensional convolutional feed-forward network is designed to enhance the extraction of local vibration features by leveraging the local feature extraction capability of convolution. Third, a cross-attention mechanism and similarity-based classification head network is developed for dynamic fault identification. The experimental results on three public fault diagnosis datasets demonstrate the effectiveness of the VLFFD method in the fault diagnosis domain.