Pneumatic control valves act as typical terminal actuators in the process industry, which suffer from difficulties in fault identification and severe consequences of faults due to the high incidence of faults and diverse fault types. Therefore, intelligent fault detection and diagnosis of pneumatic control valves have crucial practical significance. In this paper, an adaptive multi-scale features fusion network is proposed for the pneumatic control valve fault diagnosis. Firstly, a multi-scale feature extraction network with fusion self-attention mechanism is constructed to automatically extract spatial and detail features of signals. Then a weighted adaptive feature fusion network is designed to perform the weighted fusion of multi-scale features to improve the fault feature characterization capability of model. Finally, the feature identification and fault classification are performed by the Long short-term memory neural network with SoftMax function. The experimental results show that the model achieves an average accuracy of 96. 82% on the DAMADICS valve benchmark experimental platform, which are higher than other comparative models. Comparison with the detection results in the latest literature reveals that the model developed in this paper also has certain advantages such as the number of detectable faults and detection accuracy, and the detection performance of model is experimentally verified.