# An Explainable Intelligent Fault Diagnosis for Rotating Machinery via Multi-Source Information Fusion Under Noisy Environments and Small Sample Conditions

## TL;DR

This paper introduces a new method for diagnosing faults in rotating machinery using multi-source data fusion, which improves accuracy even with noisy data and limited samples.

## Contribution

The novel approach combines multi-sensor fusion, adaptive thresholding, and a multi-scale transformer to enhance fault diagnosis in challenging conditions.

## Key findings

- The proposed method achieves high diagnostic accuracy on the CWRU and PT890 datasets under noisy environments.
- The fusion module effectively highlights informative sensor channels, improving feature extraction.
- The method demonstrates robustness with small sample sizes and outperforms existing techniques.

## Abstract

In modern industrial systems, the fault diagnosis of rotating machinery is crucial for ensuring safe equipment operation. However, practical fault data are often contaminated by noise, and the scarcity of samples across fault conditions makes effective feature extraction challenging. Moreover, single-sensor measurements provide limited and incomplete information, further degrading the accuracy and reliability of diagnostic models. To address these challenges, this paper proposes an explainable intelligent fault diagnosis for rotating machinery via multi-source information fusion under noisy environments and small sample conditions. Firstly, a multi-sensor data intelligent fusion module (MSDIFM) is developed. It converts multi-sensor vibration signals into time–frequency maps via continuous wavelet transform (CWT). Pixel-level cross-channel fusion is then performed using a variance-driven dynamic weighting strategy to generate a unified fusion map, adaptively highlighting high information channels. Secondly, a multi-dimensional adaptive asymmetric soft-threshold residual shrinkage block (MASRSB) is proposed to implement differentiated and dynamic threshold control for positive and negative features, enhancing representation and discrimination capabilities. Thirdly, the multi-scale Swin Transformer (MSSwin-T) is designed. This module significantly enhances the model’s feature extraction capability by expanding multi-level receptive fields, strengthening key channel representations, and reinforcing cross-window feature interactions. Finally, to validate the effectiveness of the proposed method, experiments are conducted on both the Case Western Reserve University (CWRU) dataset and the self-created PT890 dataset. Results demonstrate that the proposed method exhibits outstanding diagnostic performance and robustness under noisy conditions and with small sample sizes.