极端条件下基于特征层面信号融合的电励磁双凸极电机匝间短路故障诊断

Inter-Turn Short Circuit Diagnosis of Wound-Field Doubly Salient Machine Based on Multi-Signal Fusion on Feature Level under Extreme Conditions

电励磁双凸极电机(WFDSM)具有结构简单、可靠性高等优点,适用于航空航天等环境恶劣领域。当发生小匝数短路故障时,由于其电流、振动等信号不会产生明显的变化,难以用传统的检测手段区分。因此,该文提出一种基于特征层面多源信号融合和改进神经网络的WFDSM匝间短路故障诊断方法,用于诊断极端环境下WFDSM早期匝间短路故障。首先,对电流信号进行经验模态分解,获得本征模态函数,同时对振动信号进行小波包变换,并对分解后的各个频段提取峭度和裕度特征,同时计算能量占比;然后,将上述特征矩阵处理后输入改进卷积神经网络训练模型中;最后,通过实验表明,采取特征融合的计算方法诊断准确率可达98%,较数据层面和结果层面的融合计算方法准确率有明显的提升,并且对极端运行环境下的噪声,该方法具有很强的抗干扰能力。

The Wound-Field Doubly Salient Machine(WFDSM)has a simple structure and high reliability,suitable for aerospace and other applications in harsh environments.However,traditional diagnosis methods are challenging to diagnose an inter-turn short-circuit fault due to slight changes in currents and vibration.Therefore,a novel WFDSM inter-turn short circuit diagnosis method is proposed in this paper based on multi-signal fusion at the feature level and improved convolutional neural network(CNN).Firstly,current and vibration data are collected from the 8-10 WFDSM.The 75 sets of four-phase current signals are processed by Empirical Mode Decomposition(EMD),and Intrinsic Mode Function(IMF)of length 101 is taken from each phase current.Thus,each condition of the current feature matrix is 300 sets of length 101.At the same time,the wavelet packet transform is applied to the vibration signal,and each sensor gets kurtosis,margin and energy ratio of length 256.Thus,each condition of the vibration feature matrix is 300 sets of length 768.Interval scaling,normalization,missing value processing,and other feature processing are taken after two sets are arranged in parallel.Each group is randomly divided into the training set and test set of 4∶1.After the input matrix is convolved and pooled,the data features are effectively preserved by the batch normalization layer,next the flattened layer is laminated and finally passed to the SoftMax classifier for diagnosis.The dropout layer is also applied in the diagnosis,while the learning rate is optimized to ensure high classification efficiency without overfitting.The diagnosis results of different detection quantities show that when only current signals are used,the accuracy can reach 82.8%on 800r/min and 83.7%on 1000 r/min.when only vibration signals are used,the accuracy can reach 85.7%and 84.3%.In contrast,when the multi-signal fusion at feature level method is adopted,the accuracy can reach up to 98.3%,while the model convergence speed accelerates significantly.In order to simulate extreme conditions,different degrees of white Gaussian noise are added to the experiment.The fusion accuracy at the feature level method barely changed when the SNR>20 dB,which is still above 95%.Meanwhile,when SNR=10 dB,the accuracy remains at 87.8%.Considering the accuracy and solving efficiency,the fusion at the feature level method is superior to the fusion method at the data or result levels.The following conclusions can be drawn.(1)The magnetic field of WFDSM will be distorted with additional harmonics generated when a short-circuit fault occurs,which also indirectly leads to the change of the current and vibration.(2)Compared with the single signal diagnosis method,the accuracy and convergence speed of the proposed model are greatly improved.(3)The proposed model performs better and has a strong anti-interference ability under extreme conditions.(4)The features extracted by EMD and wavelet packet transformation play an essential role in fault identification,and the improved CNN also has a good classification effect.