As the key component of pulse power system in particle accelerator, the anomaly detection of modulator is very important to ensure the reliable operation of the equipment. To solve the problem that existing detection methods based on single domain feature are not sensitive enough to transient anomalies, a multi-domain collaborative autoencoder (MDC-AE) anomaly detection framework is proposed in this paper. First, the time domain waveform of the original pulse voltage and current is converted into frequency domain matrix by short-time Fourier transform (STFT). On this basis, a dual-path autoencoder network architecture is designed, and one-dimensional (1D) convolutional neural network (CNN) is used to extract feature representations in time domain and frequency domain respectively. In order to enhance the reconstruction ability of the model, a cross-domain consistency constraint loss function is introduced to force the reconstructed time domain signal to maintain spectral energy alignment with the directly reconstructed frequency domain features after STFT conversion. In addition, the attention mechanism is embedded in the feature fusion stage to realize the weight adjustment of time-frequency domain features. The experimental results on the solid-state modulator (SSM) show that the proposed method achieves 99.67% F1-score index, which is 12.66% higher than that of the single time domain autoencoder, and is significantly better than other common anomaly detection methods.