To address the issue of subjective threshold setting and the difficulty in aligning feature signals with fault information in existing inverter open-circuit fault diagnosis methods, this paper proposes a novel diagnostic approach integrating model-driven and signal analysis techniques. First, a first-order differential tracker is constructed, and a wavelet analysis method is applied to denoise and decouple disturbances from the three-phase output voltages of the inverter. The residuals are then obtained by subtracting the nominal three-phase voltages, after filtering out noise disturbances and derived from the generalized state-space averaging method. Second, a voltage residual evaluation function based on short-time Fourier transform is introduced to calculate voltage time-frequency features and determine alarm thresholds, enabling fault detection through comparison. Meanwhile, the algebraic and polar relationships of the discrete Fourier coefficients of the voltage residuals are derived and summarized to identify the specific faulty power switch devices. Finally, a series of simulation results and hardware-in-the-loop experimental results verify the real-time response capability and diagnostic accuracy of the proposed method.