Research on intelligent diagnosis of multi-modal piston axis wear state based on FMD and resnet-CBAM-TCN spatio-temporal architecture

Abstract In the injection - molding mixing device of PDCPD materials, the piston shaft operates under high pressure and corrosive environments for an extended period. This leads to surface wear, yet it is challenging to detect such wear promptly using conventional methods. This situation not only undermines the reliability of equipment operation but also has a direct impact on product processing quality and production efficiency.To tackle this problem, this study analyzed the wear mechanism of the piston shaft and put forward a multi - modal intelligent diagnostic model for the wear state of the piston shaft based on the FMD and ResNet - CBAM - TCN spatio - temporal architecture. Initially, Feature Mode Decomposition (FMD) was employed to pre - process the original vibration signals. This pre - processing step enhanced the wear - related fault features and mitigated noise interference. Subsequently, the Gramian Angular Difference Field (GADF) algorithm was utilized to transform the time - domain signals into time - frequency images, thereby constructing a multi - modal input.During the feature extraction phase, the model utilized a Residual Network with a Convolutional Block Attention Module (ResNet - CBAM) to extract spatial features from the time - frequency images. Simultaneously, the Time - Domain Convolutional Network (TCN) was used to capture the temporal characteristics of the original signals. Finally, a cross - attention mechanism was introduced to realize the adaptive fusion of the two types of modal features.Experimental results demonstrate that this model overcomes the limitations of single - modal features. Moreover, the strategic integration of multi - level attention mechanisms significantly improves the model’s diagnostic capabilities for piston shaft wear under complex conditions, enabling accurate diagnosis of the piston shaft wear state.