【目的】直流电机系统中的电力电子器件在长期高频开关工作下易发生软故障。针对软故障诊断中存在的时频域特征融合不足、识别准确率低等问题，本文将长短期记忆(LSTM)网络和多尺度时频域交叉注意力机制(CAM)相结合，提出了一种基于LSTM-CAM-Transformer模型的故障诊断方法。【方法】首先，对采集到的故障信号进行预处理，采用霜冰优化算法(RIME)对变分模态分解(VMD)参数进行寻优，精准得到最优分解模态数K和惩罚因子α的组合，有效去除信号中的噪声与干扰成分。然后，提取各个本征模态函数的5维时域参数和5维频域参数，将其作为故障诊断的特征向量。最后，利用多尺度时频域CAM增强特征向量在时域和频域之间的信息交互，使模型进一步挖掘信号的时频域特征。【结果】将本文所提模型和其他四种模型进行对比以验证其优越性。试验结果表明，相较于其他四种模型，本文所提LSTM-CAM-Transformer模型收敛速度最快，稳定性和泛化性更强，且在诊断准确率、F1分数、损失值和召回率上的表现均优于其他四种模型。【结论】本文所提LSTM-CAM-Transformer模型通过融合基于RIME改进的VMD信号预处理策略与CAM时频特征增强机制，有效解决了传统方法中时频域特征融合不足的难题，为直流电机系统中电力电子设备的软故障诊断提供了一种高效可靠的新方法。

[Objective] Power electronic components in DC motor systems are prone to soft faults under prolonged high-frequency switching operation. To address the problems in soft fault diagnosis, including insufficient fusion of time-frequency domain features and low recognition accuracy, this paper proposes a fault diagnosis method based on the LSTM-CAM-Transformer model by combining the long short-term memory (LSTM) network and the multi-scale time-frequency domain cross-attention mechanism (CAM). [Methods] Firstly, the collected fault signal was preprocessed, and the parameters of the variational mode decomposition (VMD) algorithm were optimized using the Rime optimization algorithm (RIME), which accurately obtained the optimal combination of the decomposition modals K and the penalty factor α, and effectively removed the noise and interference components in the signal. Then, the 5-dimensional time domain parameters and 5-dimensional frequency domain parameters of each intrinsic modal function were extracted, which were used as the feature vectors for fault diagnosis. Finally, a multi-scale time-frequency domain CAM was utilized to strengthen the information interaction between the time domain and frequency domain of the feature vectors, allowing the model to further mine the time-frequency domain features of the signals. [Results] The model proposed in this paper was compared with the other four models to verify its superiority. The experimental results showed that compared with the other four models, the LSTM-CAM-Transformer model proposed in this paper has the fastest convergence speed, better stability and generalization, and outperforms the other four models in diagnostic accuracy, F1 score, loss value and recall. [Conclusion] The LSTM-CAM-Transformer model proposed in this paper effectively solves the problem of insufficient fusion of time-frequency domain features in the traditional method by integrating the signal preprocessing strategy of RIME-based improved VMD signal preprocessing strategy with the CAM time-frequency feature enhancement mechanism, which provides an efficient and reliable new method for the soft fault diagnosis of the power electronic equipment in the DC motor system.