The invention relates to a fault diagnosis method, in particular to an isolation switch fault diagnosis method based on matrix neural networks. According to a novel isolation switch infrared diagnosis based on image identification, isolation switch infrared images are directly inputted, so that neural network training learning is achieved, traditional neural networks are widened to the matrix neural networks, noise suppression preprocessing of the isolation switch infrared images are firstly performed, singular value decomposition of infrared image matrixes is performed, left weight matrixes and right weight matrixes of each layer are trained and learned, isolation switch fault diagnosis based on the infrared images is achieved, and the isolation switch fault diagnosis method based on the matrix neural networks is finally achieved.