【目的】四象限脉冲整流器绝缘栅双极型晶体管(IGBT)开路故障具有较强隐蔽性，极易引发网侧电流畸变、功率因数衰减及直流母线电压波动。应用于故障诊断领域的传统卷积神经网络(CNN)模型存在诊断精度低、易陷入局部最优解的缺陷。针对上述问题，本文引入粒子群优化(PSO)算法，提出一种融合小波变换与PSO-CNN的智能故障诊断方法。【方法】首先基于Simulink搭建单相两电平四象限脉冲整流器模型，设置7种梯度电压波动工况与10种不同的IGBT开路故障状态，模拟复杂、真实的设备运行场景。随后采集网侧电流与直流侧电压信号，采用db3小波基对信号进行6层小波分解，并结合频带能量系数构建标准化特征向量。考虑到电流与电压信号的特征差异，构建双通道CNN，分别提取网侧电流与直流侧电压的故障特征。在此基础上，引入PSO算法对CNN关键超参数进行自适应优化。【结果】选取加权F1分数与混淆矩阵作为模型性能的定量评价指标，将所提PSO-CNN模型与传统CNN模型进行对比。试验结果表明，相较于传统CNN模型，PSO-CNN模型可将加权F1分数由0.45提升至0.92，且能有效改善特征重叠故障的误分类问题，在电压波动工况下仍可保持稳定的故障诊断能力，仅在少量高度相似的故障类别中存在轻微识别偏差。【结论】小波变换与PSO算法相结合，显著提升了CNN模型的特征提取能力与分类可靠性。所提智能诊断方法适用于四象限脉冲整流器的实时状态监测，可为各类电力电子变流装置的智能运维与性能优化提供可行、可靠的技术参考。

[Objective] Open-circuit faults of insulated-gate bipolar transistor (IGBT) in four-quadrant pulse rectifiers exhibit strong concealment characteristics, which can inevitably induce grid-side current distortion, power factor degradation, and DC-side voltage fluctuation. Traditional convolutional neural network (CNN) model applied to fault diagnosis is restricted by low diagnosis accuracy and easy trapping in local optimal solutions. To address these limitations, this paper adopts the particle swarm optimization (PSO) algorithm and proposes an intelligent fault diagnosis method integrating wavelet transform and the PSO-CNN. [Methods] Firstly, a single-phase two-level four-quadrant pulse rectifier model was established based on Simulink. Seven graded voltage fluctuation operating conditions and ten different IGBT open-circuit fault states were set to simulate complex and realistic equipment operation scenarios. Subsequently, grid-side current and DC-side voltage signals were collected. A six-layer wavelet decomposition was performed on the signals using the db3 wavelet basis, and normalized feature vectors were constructed by combining the frequency band energy coefficients. Considering the feature differences between current and voltage signals, a dual-channel CNN was built to extract fault features of grid-side current and DC-side voltage respectively. On this basis, the PSO algorithm was introduced to adaptively optimize the key hyperparameters of the CNN. [Results] Weighted F1-score and confusion matrix were taken as quantitative evaluation metrics for model performance, and comparative tests were conducted between the proposed PSO-CNN model and the traditional CNN model. The experimental results showed that compared with the traditional CNN model, the PSO-CNN model increased the weighted F1-score from 0.45 to 0.92. It effectively reduced the misclassification of faults with overlapping features and maintained stable fault diagnosis performance under voltage fluctuation conditions, with only minor recognition deviations in a small number of highly similar fault categories. [Conclusion] The combination of wavelet transform and PSO algorithm significantly enhance the feature extraction ability and classification reliability of the CNN model. The presented intelligent diagnosis method is suitable for real-time condition monitoring of four-quadrant pulse rectifiers, and it provide a feasible and reliable technical reference for intelligent operation and maintenance as well as performance improvement of various power electronic conversion devices.