Enhancing multimodal fault diagnosis in mechanical systems via mixture of experts

Abstract Mechanical wear occurs during the operating cycle of all types of complex machinery. In this paper, the spectral, ferro-spectral, physical, and chemical analyses, along with onboard particle counting characteristics under laboratory conditions, are taken as small sample datasets. Wasserstein Generative Adversarial Network is used as the regeneration algorithm model for raw data, and the composite dataset with richer semantic information is used as input. A one-dimensional representation of the composite data is transformed into a two-dimensional image containing richer static information using the Markov Transfer Field transformation concept. The Mixture of Experts based meritocracy architecture selects different expert systems for various features in the dataset by categorizing the expert systems according to combinatorial principles and setting corresponding weight assignments. ConvNeXt, Bidirectional Transformer (BiTransformer), and Bidirectional Long Short-Term Memory are then employed to capture the image features and perform fault diagnosis on the composite one-dimensional mechanical wear data, respectively. An attention mechanism is added to optimize the algorithm globally, weighting the feature information across multiple dimensions to ensure the reliability and completeness of the results. The final results show that the accuracy of fault diagnosis exceeds 95%, demonstrating ideal performance.