Abstract: Disc brake is an important part of high-speed railway trains and rail vehicles. The friction between the brake pad friction block and the brake disc is used to ensure the deceleration and normal stop of the train. The braking interface between the friction block and the brake disc is the working hub of the friction braking system, and maintaining good friction behavior is the guarantee of the reliability of friction braking. However, due to the influence of the slope of the upper and lower lines, high temperature, and environmental conditions during the train running, the friction block on the brake pad inevitably appears partial wear. Friction block eccentric wear will not only make the train braking process noise, affect the comfort of passengers, but also affect the braking performance of the brake itself. Therefore, it is urgent to monitor the status of brake friction blocks for high-speed trains. In order to better realize the state monitoring of the brake friction block, this study first analyzed the causes of the partial wear of the friction block, and then designed the friction block according to the possible types of partial wear of the friction block, and built the state monitoring model of the friction block through the deep learning framework tensorflow for the tangential vibration acceleration generated by the friction block during braking. In order to improve the feature extraction ability of the model, this study combined the convolutional block attention module (CBAM) with the multi-scale convolution module to design the model. CBAM considered not only the importance of pixels in different channels, but also the importance of pixels in different positions in the same channel. By giving different weights to the channels and spaces of the data, more critical and important information was extracted, so that the model can make accurate judgments. By adding two different multi-scale convolution modules designed in this study to the monitoring model, not only the feature extraction performance of the model was improved, but also the width of the model was increased, and the robustness of the model was improved, so that the model can extract more abundant features. In order to prevent the phenomenon of gradient explosion and gradient disappearance in the model, this study added residual connections at different positions of the model. In order to verify the feature extraction ability of the proposed method, the CBAM-CNN model was applied to the brake pad eccentric wear data set obtained from the self-made test bench of high-speed trains. In the training process of the model, the parameters (such as training accuracy, training accuracy and training recall rate) can achieve good results, indicating that the designed model was reasonable. By comparing the existing depth algorithm with the traditional machine learning algorithm through ten experiments, the model proposed in this study can not only achieve higher test accuracy, but also had better stability. In addition, the t-SNE visualization analysis of the features obtained from the last hidden layer of the model showed that the four different state friction block features extracted by CBAM-CNN model not only had obvious separability in space, but also were better than other deep learning models in spatial clustering. This showed that the proposed model had a strong feature extraction ability, and the characteristics of each eccentric wear state had obvious separability in space, which verified the effectiveness of CBAM-CNN model in monitoring the eccentric wear state of brake friction blocks of high-speed trains.