A Fault Diagnosis Approach Based on Adaptive Lightweight 1D-XFormer for Bearings and Electric Motors in Complex Working Conditions

Addressing the challenges of low feature discriminability between normal and fault data of motors under real operating conditions, coupled with the scarcity of high-quality fault samples and the corruption by noise through complex nonlinear transmission paths, this article introduces an adaptive lightweight 1D-Xception-Transformer (1D-XFormer) network. This network integrates input, intermediate, and output channels. Initially, raw vibration signals serve as inputs. Subsequently, enhanced nonlinear feature extraction is achieved by an input channel comprising an improved adaptive activation function, Adaptive GELU (AdaptGELU), integrated with three stacked inverted depthwise separable and residual convolutional blocks. As the intermediate channel, 1D-Transformer, leveraging its Encoder module, effectively captures irregular fault features amidst low-resolution fault samples. The output channel then combines the results through the same AdaptGELU and convolutional blocks combination. Furthermore, 1D-XFormer incorporates deformable convolutions (DeformConv) for adaptive structural adjustments, optimizing the capture of irregular local features. Experiments utilizing vibration signals from test benches and real damaged datasets demonstrate the proposed method’s robust motor fault diagnosis capabilities under complex operating conditions.