Fault diagnosis of three-phase asynchronous motor based on continuous wavelet and TransXNet for multi-source information fusion

Deep Learning (DL) has been applied to promote intelligent fault diagnosis, achieving significant performance improvements. However, most existing methods cannot effectively capture the temporal information and global characteristics of mechanical equipment, failing to collect sufficient fault data. Meanwhile, due to complex and harsh operating environments, single-source fault diagnosis methods struggle to stably and comprehensively extract fault features. Therefore, this paper proposed a new method for fault diagnosis of threephase asynchronous motors based on continuous wavelet and TransXNet based on Multi-Source Information Fusion (MSIF), which improved the stability of diagnostic performance by extracting and integrating rich features. First, a three-phase asynchronous motor fault experimental platform was built, and acceleration sensors and current sensors were used to collect vibration signals and current signals of the motor under various working conditions to obtain multi-source signals. Second, a new hybrid network module called Dual Dynamic Token Mixer (D-Mixer) was proposed, which dynamically utilized global and local information while introducing a large receptive field and powerful inductive bias. This design enhanced feature extraction capability without compromising input dependency. A Multi-scale Feed-forward Network (MS-FFN) was proposed to perform multiscale feature aggregation in the feed-forward networks. By alternately employing D-Mixer and MS-FFN, a new hybrid CNN-Transformer network named TransXNet was constructed. Subsequently, continuous wavelet transform was applied to perform time-frequency transformation on the multi-source signals, followed by a proposed data-level fusion strategy to generate multi-source information maps. These maps were then fed into TransXNet for feature segmentation and aggregation, completing the feature extraction process for model training and validation. The effectiveness of the proposed TransXNet was validated. Finally, multi-source test samples were utilized to evaluate the diagnostic performance of the proposed method. Experimental results demonstrate that due to TransXNet’s powerful feature extraction capabilities, the recognition accuracy reached 100%. By comparing and adjusting the four evaluation indicators of Rand index, normalized mutual information, F1 score and accuracy, as well as noise immunity analysis, it is concluded that the proposed method is superior to the current most advanced method (SOTA) in the field of fault diagnosis. The field has promising prospects. © 2025, Central South University Press. All rights reserved.