Equipment such as pipelines and valves in key areas of thermal power plants often experience leakage faults of substances such as oil and steam. In order to improve the segmentation and detection accuracy of pipelines and valves leakage faults in thermal power plants and accelerate model inference speed, a pipeline and valve leakage segmentation and detection algorithm in thermal power plants based on the improved YOLOv7 is proposed. By adding relevant instance segmentation modules to the YOLOv7 network, the parallelization of instance segmentation and target detection tasks are achieved. By integrating polarized self-attention mechanism and separated vision transformer modules, the interference of complex backgrounds is weakened, the edge extraction of the leakage area is strengthened. Then, in the post-processing stage, the CP-cluster(confidence propagation cluster) mechanism is applied to improve the accuracy of model localization; Finally, Ghostconv is used in the neck network to reduce feature redundancy, and the model is compressed by channel pruning technology to achieve lightweight model. The experimental results show that after adding various improvements, the mAP@0.5:0.95 of the algorithm's segmentation task and detection task reach 75.7% and 82.2%, respectively, which increase by 11.9% and 7.1% compared to the basic model, and the model parameter decreases by 30.3%. The model can be effectively applied in the actual production environment of power plants.