A collaborative fault diagnosis model based on an attention-weighted multiscale convolutional neural network and an improved multi-grained cascade forest

Abstract The health of rolling bearings is of great significance to ensure the smooth operation of rotating machinery. Failure of rolling bearings will lead to severe accidents, prolonged downtime or sub-stantial economic losses. Therefore, timely monitoring and diagnosing the health status of rolling bearings is essential to ensure the reliable operation of industrial systems. Deep learning methods have widely applied in industry due to their powerful feature extraction capabilities. However, the fault signals of rolling bearings with complex and variable working conditions exhibit high nonlinearity. To extract the nonlinear features of the fault signal, the traditional deep models often have complex structures and large parameter matrices. It is challenging to train models, obtain high efficiency, and achieve satisfactory results. Therefore, a novel collaborative diagnostic model (MsCNN-IgcForest) based on an attention-weighted multiscale convolutional neural network (MsCNN) and an improved multi-grained cascade forest (IgcForest) is proposed to process bearing fault signals. First, inspired by Xception, we design a lightweight attention-weighted MsCNN feature extraction model that uses attention mechanisms to suppress scattered features and improve the running speed by reducing the number of network training parameters. Second, the improved deep forest employs an attention-weighted MsCNN structure as a classifier instead of multi-grained scanning to reduce memory consumption and achieve fault recognition. Finally, the MsCNN-IgcForest model is verified by two cases of bearing diagnosis and the contrastive results show high fault diagnosis accuracy and strong robustness. In conclusion, the improved model shows good fault diagnosis performance and has a potential reference value for industrial fault diagnosis.