Convolutional neural network model fault tolerance method based on key layer rollback mechanism

The invention discloses a convolutional neural network model fault tolerance method based on a key layer rollback mechanism, and the method comprises the following steps: S1, searching a key layer, searching a key node which has the greatest influence on classification through a mode of adding a parameter lambda to an original model for training, and determining the key layer distribution of a network; S2, inputting a classification picture, recognizing the input picture through a convolutional neural network, obtaining an output feature of a key layer, and obtaining a detection threshold according to the output feature; and S3, modeling the hardware constraint, and calculating the number of detection points under the condition that the hardware constraint is satisfied by using a simulated annealing algorithm. Key node distribution of the convolutional neural network is obtained through analysis experiments, and key layers of the network and output distribution of the key layers are obtained on the basis of t