The invention discloses a CNN (convolutional neural network)-based method for monitoring internal component surface of a nuclear pile. The method comprises the following steps: S1, acquiring video data of the internal component surface of the nuclear pile; S2, obtaining image data; S3, dividing the image data into a training dataset and a testing dataset; S4, inputting the training dataset into aCNN for training to obtain a feature recognition model and output primary recognition features; S5, classifying the primary recognition features according to known corrosion features to obtain a corrosion type; S6, inputting the testing dataset into the feature recognition model for testing to output secondary recognition features; S7, judging whether the secondary recognition features accord withthe corrosion type; S8, displaying the corrosion type on a human-computer interaction interface of a monitoring and early warning module. The method solves the problems that labor input is large andefficiency is low due to manual monitoring and requirements for timeliness and continuity of material surface corrosion monitoring cannot be met in the prior art.