An intelligent cross-domain fault diagnosis framework based on XGBoost and CNN-KLD for multi-source data

Abstract

To address the limitations in model generalization caused by discrepancies in the distribution of multi-source data and insufficient labeling in the fault diagnosis of high-speed train bearings, this paper constructs an intelligent diagnostic computing pipeline designed for cross-operating-condition scenarios. This issue manifests as inconsistencies between the distribution of training data and actual operational data, making it difficult for traditional models to maintain stable performance. Methodologically, we propose an integrated algorithmic system comprising signal processing, feature modeling, and cross-domain learning. First, signal quality is enhanced through multi-level processing involving anomaly correction, wavelet denoising, and adaptive filtering. Second, a multi-channel feature fusion strategy based on signal-to-noise ratio (SNR) weighting is designed to achieve a unified representation of multi-source data. Building on this foundation, XGBoost is introduced for source-domain modeling, and a CNN-KLD model—combining a convolutional neural network with KL divergence constraints—is constructed. Joint optimization is employed to achieve feature distribution alignment and cross-domain transfer. Experimental results show that this method achieves a classification accuracy close to 100% in the source domain and an accuracy of 83.50% in the target domain, with a micro-average AUC of 0.9899, significantly outperforming models without distribution alignment. In summary, this computational pipeline achieves multi-source fusion and improved cross-domain generalization capabilities while maintaining low computational complexity. It demonstrates good robustness and versatility and can be extended to signal diagnosis tasks under complex operating conditions.