基于梅尔倒谱系数特征集的储能变流器开路故障诊断方法

Open-Circuit Fault Diagnosis Method of Energy Storage Converter Based on MFCC Feature Set

电池储能电站功率转换系统(power conversion system,PCS)故障诊断在储能电站智能运维中发挥着重要作用。现有方法在非侵入式识别PCS内部IGBT开路故障时,易出现信号特征提取困难、数据维度爆炸以及阈值判定区间不稳定等问题。提出一种基于梅尔倒谱系数(Mel-scale frequency cepstral coefficients,MFCC)特征集的储能变流器开路故障诊断方法。首先,以交流侧三相电流为输入信号,通过分析不同频率区间的信号频谱能量分布情况和包络特征,构建MFCC故障特征数据集。然后,结合核主成分分析(kernel principal components analysis,KPCA),实现充放电工况下非线性故障特征的降维筛选;其次,以低维故障特征集为输入,构建基于贝叶斯优化算法(bayesian optimization algorithm,BOA)与一维卷积神经网络(1d-convolutional neural network,CNN-1D)的故障状态诊断模型;最后,通过并网储能变流器的故障仿真实验,与现有方法进行比较,结果表明:所提方法在复杂的噪声环境下的鲁棒性和准确性更优。

Power conversion system(PCS)fault diagnosis plays an important role in the intelligent operation and maintenance of battery energy storage power stations.Considering the difficulties in signal feature extraction,data dimension explosion,and instability of threshold determination interval in the non-invasive identification of IGBT open-circuit fault in PCS by the existing methods,this paper proposes an open-circuit fault diagnosis method of energy storage converter based on Mel-scale frequency cepstral coefficients(MFCC)feature set,so as to support the normal operation and maintenance of PCS.Firstly,the three-phase current on the alternating current(AC)side is taken as the input signal,and an MFCC fault feature data set is constructed by analyzing the signal spectrum energy distribution and envelope characteristics in different frequency intervals.Then,through kernel principal component analysis(KPCA),the dimension reduction screening of nonlinear fault features under charge and discharge conditions is realized.Secondly,a fault state diagnosis model based on the Bayesian optimization algorithm(BOA)and onedimensional convolutional neural network(CNN-1D)is constructed with a low-dimensional fault feature set as an input.Finally,through the fault simulation experiment of a grid-connected energy storage converter,the proposed method is compared with existing methods,and the results show that the proposed method has better robustness and accuracy in complex noise environments.