As the growing requirements for the stability and safety of process industries, the fault detection and diagnosis of pneumatic control valves have crucial practical significance. Many of the approaches were presented in the literature to diagnose faults through the comparison of residual sequences with thresholds. In this study, a novel hybrid neural network model has been developed to address the issue of pneumatic control valve fault diag-nosis. First, the feature extractor automatically extracts in-depth features of the signals through multi-scale convolutional neural networks with different kernel sizes, which not only adequately explores the local dis-tinguishable features, but also takes into account the global features. The extracted features are then fused by the feature fusion layer to reduce redundant features. Finally, the long short-term memory for fault identification and the dense layer for fault classification. Experimental results demonstrate that the average test accuracy is above 94% and 16 out of the 19 conditions can be successfully detected in the simulated actual industrial en-vironment. The effectiveness and practicability of the proposed method have been verified through a com-parative analysis with existing intelligent fault diagnosis methods, and the results suggest that the developed model has better robustness.