基于CNN-SVM融合算法的矿用充电机功率模块早期故障识别

Early Fault Identification of Power Modules in Mining Battery Chargers Based on CNN-SVM Fusion Algorithm

摘要: 针对矿用充电机功率模块早期故障辨识存在检测灵敏度偏低、故障特征易受环境噪声干扰、小样本数据集难以充分挖掘有效信息等现实痛点，本文提出融合改进变分模态分解(IVMD)、多尺度注意力卷积神经网络(MS-CNN-Attention)与粒子群优化支持向量机(PSO-SVM)的一体化智能故障诊断模型。首先依托IVMD算法完成原始电气信号自适应分解，结合排列熵(PE)筛选含有微弱故障信息的模态分量并完成信号重构，实现故障特征信噪比优化；其次搭建嵌入SE挤压激励注意力机制的多尺度卷积神经网络，通过多尺寸卷积核并行提取多感受野故障特征，借助通道动态加权机制强化关键故障特征表征效果；最后将网络输出的高维特征输入经粒子群算法完成超参数寻优的SVM分类器，完成功率模块早期故障精细化分类识别。仿真测试结果表明，所提混合模型对四类典型早期故障的平均识别精度可达98.8%，在高强度噪声工况下具备优良抗干扰性能，模型单次推理耗时满足设备在线实时监测指标。研究成果可为矿山电力电子装备的预测性维保工作提供新的技术实现路径。

Abstract: To solve the prominent problems including insufficient detection sensitivity, severe noise interference on fault features and inefficient utilization of limited small-scale datasets during early fault identification of power modules for mining chargers, an integrated intelligent diagnosis model combining Improved Variational Mode Decomposition (IVMD), Multi-scale Attention Convolutional Neural Network (MS-CNN-Attention) and PSO optimized Support Vector Machine (PSO-SVM) is proposed in this paper. Firstly, IVMD is adopted to adaptively decompose original electrical signals, and Permutation Entropy (PE) is used to screen sensitive modal components containing faint fault information for signal reconstruction, which improves the signal-to-noise ratio of fault features effectively. Then, an MS-CNN-Attention network embedded with Squeeze-and-Excitation (SE) attention is constructed. Multi-size convolution kernels are applied to extract multi-receptive-field features in parallel, and dynamic channel weighting is realized to enhance the representation capability of critical fault characteristics. Ultimately, high-dimensional features extracted by the neural network are imported into the SVM classifier whose hyperparameters are optimized via particle swarm optimization, realizing accurate classification of incipient faults in power modules. Simulation experiments verify that the proposed hybrid model achieves an average diagnostic accuracy of 98.8% for four categories of typical early faults, presents outstanding anti-noise robustness under harsh noisy conditions, and its single inference latency satisfies the requirement of real-time online monitoring. The research provides an innovative technical reference for predictive maintenance of mining power-electronic equipment.