Control valves serve as critical actuators in industrial process control systems, and their operational status directly impacts production safety and product quality. Addressing the limitations of existing control valve fault diagnosis methods—such as delayed pressure and flow signal responses, susceptibility of vibration signals to interference, and inadequate extraction of characteristic information—this study proposes a fault diagnosis method for pneumatic control valves based on microwave vibration measurement and time-frequency domain feature fusion. First, microwave vibration measurement technology is employed to achieve non-contact, high-precision acquisition of control valve stem vibration signals, overcoming the application limitations of traditional contact-based sensors. Stem vibration can more directly reflect the status of critical components such as the valve core, spring, and seals. Second, a network structure with multi-scale time-frequency domain dual-channel feature fusion is constructed. In the time-domain branch, multi-scale one-dimensional convolutions combined with bidirectional gated recurrent units are designed to fully extract the temporal dynamic features of the signal. In the frequency-domain branch, short-time Fourier transforms are used to convert one-dimensional signals into two-dimensional time-frequency spectrograms, and multi-scale two-dimensional convolutional networks are employed to extract spectral texture features. A channel attention mechanism is introduced to adaptively learn feature importance weights, and a cross-attention mechanism is employed to achieve deep fusion of time-frequency domain features, fully leveraging complementary information across different modalities. Experiments were conducted on a pneumatic control valve fault simulation test bench equipped with a microwave vibration measurement system. The experimental results show that the proposed method achieves a classification accuracy of 96.25% for six operating states on the fault simulation test bench, demonstrating superior diagnostic performance compared to common deep learning models. In the validation of 11 fault modes on the DAMADICS platform, the method achieved an average classification accuracy of 99.24%, demonstrating the model′s excellent generalization capability and providing a new technical approach for control valve fault diagnosis.