一种改进卷积神经网络的逆变器故障诊断

Improved Inverter Fault Diagnosis Based on Convolutional Neural Network

针对传统二极管钳位式三电平逆变器故障诊断方法存在的诊断效率低且准确率不高的问题,将一种自适应正则化系数引入卷积神经网络CNN(convolutional neural network),对逆变器进行故障诊断。在传统CNN模型引入正则化去拟合中,正则化系数常采用全局统一的常数型参数,训练过程中需不断试错且效果甚微,针对此提出根据目标损失函数梯度变化,自适应调整正则化系数的CNN模型,能够加快其在逆变器故障诊断中的收敛速度,增强模型泛化能力,提高故障识别准确率。实验表明,与传统BP神经网络和原始CNN模型相比,改进的CNN模型能对逆变器复杂故障做出实时准确诊断。

In view of the low diagnostic efficiency and low accuracy of the traditional diode clamping three-level inverter fault diagnosis method,adaptive regularization coefficient is introduced into convolutional neural network(CNN)for the inverter fault diagnosis.In the de-fitting of the traditional CNN model with the introduction of regularization,the regularization coefficient often uses global unified constant parameters;meanwhile,it is necessary to continuously try and fail in the training process but with little effect.To solve this problem,adaptive adjustment of regularization coefficient is proposed according to the gradient change in the target loss function to speed up the convergence speed of the CNN model in the inverter fault diagnosis,enhance the model’s generalization capability,and improve the accuracy of fault identification.Experimental results show that compared with the traditional BP neural network and the original CNN model,the improved CNN model can make a real-time and accurate diagnosis of the complex fault of the inverter.