IPSA-DSB-ResNet: a deep learning fault diagnosis method of circuit boards using infrared thermographic sequence

Abstract The fault diagnosis of circuit boards using single-moment infrared thermal images is prone to underdiagnosis and misdiagnosis of components with small temperature differences between normal and faulty states. To address this issue, this paper proposes a novel deep learning-based fault diagnosis method using infrared thermographic sequence: IPSA-DSB-ResNet. Firstly, a set of infrared thermographic sequences of circuit boards from power-up to stable operation period are collected by an infrared thermographic camera, preprocessed, and constructed as a dataset. Secondly, the ResNet50 residual network model is modified by the strategies of embedding two improved pyramid split attention modules (IPSA) into the shallow and deeper layers of the network and introducing a maximum pooling operation to realize the down sampling shortcut branch (DSB) in Layer 3, which selectively emphasize multi-scale key detail features, target contour features, and cross-channel interaction features, and reduce feature loss. Finally, ablation and comparison experiments are conducted on the same dataset. The results show that the method can effectively identify component undercurrent, overcurrent, and normal states and increase the accuracy of diagnosis.