Data imbalance fault diagnosis method based on an ensemble multi-scale convolutional attention network

In recent years, mechanical intelligence fault diagnosis methods based on deep learning are in the ascendant. However, the problem of noise interference and data imbalance is often faced in practical applications, so it is still a challenge to achieve high precision and reliable fault diagnosis. To solve the problem that traditional convolutional neural networks have poor anti-noise performance and are easy to ignore the minority class samples, this paper proposes a mechanical intelligence fault diagnosis method based on an ensemble multi-scale convolutional attention network (EMCAN). First, a multi-scale convolutional attention network is constructed as the base classifier, which is mainly composed of the multi-scale convolutional denoising module (MCDM) and the cooperative attention module (CAM). MCDM suppresses high-frequency noise and extracts multi-scale discriminant features. Differentiated CAMs adaptively focus on important features and increase the diversity of base classifiers. Second, an ensemble strategy based on improved weighted voting is proposed, and balanced training subsets are constructed for each base classifier by sampling with replacement to improve the robustness and generalization of the ensemble model. The proposed EMCAN is validated on a bearing open dataset and a gearbox experimental dataset. Compared with the state-of-the-art comparison method, the Gmean of the proposed EMCAN is 4.60% and 12.11% higher under the most imbalanced conditions, respectively, which proves the validity and superiority of EMCAN.