Cross-domain fault detection method based on multi-scale fusion and deformable convolution

The invention provides a cross-domain fault detection method based on multi-scale fusion and deformable convolution, and solves the corresponding problems in train fault image cross-domain target detection. Comprising the following steps: preparing a train fault data set in a normal scene as a source domain data set and a train fault data set in a foggy day scene as a target domain data set; designing a domain adaptive feature pyramid, and adding the domain adaptive feature pyramid into a RetinaNet target detection network to realize feature alignment of a source domain and a target domain; a deformable convolution module is added to a backbone in the target detection network to realize the adaptive capacity to faults with large morphological differences; an ASFF module is added to an FPN in an original detection network, and the problem that learning targets of different scales are inconsistent in the FPN is solved; and a regression sub-network loss function of the RetinaNet network is changed into CIOU loss