Few-shot fault diagnosis of mechanical transmission system based on multi-image fusion and relation network

Abstract In practical engineering, mechanical transmission systems exhibit variable operating conditions, complex fault types, and scarce fault samples. As a result, traditional diagnostic models that rely on large-scale historical data often perform poorly under varying working conditions and small-sample scenarios. Although meta-learning has the ability to rapidly adapt to new tasks with only a few samples, existing meta-learning methods still show limitations in signal processing, feature extraction, and model lightweight. To address these issues, this paper proposes a small-sample intelligent fault diagnosis method based on multi-channel image fusion and an improved relation network (RN). First, 1D vibration signals are converted into 2D images using short-time Fourier transform, continuous wavelet transform, and grayscale texture encoding. These images are then aligned and concatenated along the channel dimension to form fused images that effectively integrate time-domain, frequency-domain, and texture information. Second, a squeeze-and-excitation module is incorporated into a lightweight feature extractor to adaptively adjust feature weights through optimized loss, thereby enhancing the representation of salient information. Subsequently, a local residual block is introduced to fuse low-level and high-level features, reducing the risk of overfitting. Finally, dilated convolutions are employed to replace standard convolutions, expanding the receptive field and improving multi-scale feature capture. Experimental results show that the proposed method achieves an accuracy of 99.56% in the 12-way 20-shot task, representing 9.78% improvement over the original RN. Unlike existing studies that predominantly assume constant operating conditions, this paper further conducts comprehensive evaluations through experiments involving types increasing, samples decreasing, speed transfer and load transfer experiments.