Abstract: Due to geological conditions, hydraulic power and automatic following system, abnormal state of hydraulic support generally occurs in the process of automatic movement. Relying solely on manual monitoring and relocation will seriously affect the efficiency of automatic machine following at the working face. In order to solve the above problems, the automatic visual detection method for the abnormal state of hydraulic support at the working face is proposed which captures the status of hydraulic support in real time by non-contact visual perception method Firstly, this method utilizes the YOLOv8 semantic segmentation network to achieve dynamic real-time division of the target area of the working face and accurately obtain the contour and positioning information of the hydraulic support base and push rod. By analyzing the positioning information of different hydraulic support bases and push rods, it can realize the automatic identification of hydraulic support number and extract the local image of the adjacent hydraulic support base. Finally, the improved ResNet50 convolutional classification network which fuses multi-scale feature information is used to extract features from local images. Then, the reliable automatic detection of the abnormal state is realized by combining the support number information The experimental results show that the detection accuracy of hydraulic support abnormal status is 99.17%, and the processing speed is 27.8 frames/s. At the same time, the proposed automatic detection algorithm of the abnormal state of hydraulic support is applied to the AI video surveillance system of coal face ,and meets the real-time and reliability requirements of engineering.