Diagnosis of Inter-turn Short Circuit of Permanent Magnet Synchronous Motor Based on Deep learning and Small Fault Samples

Abstract

An efficient and accurate method based on a conditional generative adversarial net (CGAN) and an optimized sparse auto encoder (OSAE) is proposed to detect the inter-turn short circuit (ITSC) problem for permanent magnet synchronous motors (PMSMs). In order to achieve an accurate detection of the ITSC, the CGAN is adopted to augment the few fault samples, and a noise injection strategy is applied to enhance the generalization ability of the network in the framework of the OSAE. Specifically, we made a combination of two types of signals to create a training set that is augmented by the CGAN, and the parameters of the OSAE are determined by the training process of networks. The experimental results indicate that the proposed method for the fault diagnosis of this fault achieves high accuracy 98.9%.

Introduction

The permanent magnet synchronous motor (PMSM) has some special characteristics in many rotating machinery such as high efficiency, power density and wide speed range [1]. But it is vulnerable to various types of faults. Especially the inter-turn short circuit (ITSC) is a root of all apparent faults of PMSMs which can result in other serious faults such as an inter-phase short circuit, a single-phase ground and so on. These coupling faults may lead to serious demagnetization of a permanent magnet, which causes the PMSM to stop running. Traditionally, the ITSC is detected by regular disassembly. Wu et al. [2] proposed a new detection coil which is installed inside the generator for online detection of this fault. However, the above methods are destructive and are not suitable for many application scenarios. So the nondestructive diagnosis is the hot spot of detecting the ITSC.

The nondestructive diagnosis relies on the physical changes of the PMSM to detect the ITSC faults. For example, the temperature is the most intuitive manifestation of this kind of faults, if the number of short circuit coils increases unregularly, the temperature inside the PMSM may exceed 100 degrees. The deviation of a three-phase current fluctuates greatly more than normal with the increase of eddy loss, and a magnetomotive force is opposite to the direction of a resultant magnetic, which causes a demagnetization of the PMSM easily. Besides, negative sequence characteristics, a harmonic current and vibration are also fit for detecting the ITSC. These parameters are easily interfered by several objective conditions such as humidity, dust and so on. So the combination of these fault samples are benefit for reducing the influence of environmental factors, which enhancing the diagnosis reliability of the short circuit fault effectively.

However, the probability of detecting such faults in rotating machinery is only 1%, which makes the diagnosis of such faults become a difficult few-shot problem due to the lack of fault samples. A series of mathematical analysis methods such as principal component analysis(PCA) [3] empirical mode decomposition (EMD) [4] BP network [5] support vector machines (SVM) [6] and other machine learning methods can be used to identify the fault based on these few samples. But there are several limitations in the application of fault diagnosis.

• The fault samples, despite few, are crucial to the classification quality of the above-mentioned methods. Hence, the issue of lack of fault samples significantly impairs the classification performance of those methods.

• The distribution characteristics of these samples are not compatible with the analysis capability provided by these traditional methods. For example, the PCA method is not the best method for identifing the few sample with the non-Gaussian distribution.

• The SVM and artificial neural networks are unable to learn complex non-linear relationships in the task of fault diagnosis due to their shallow architectures. So deeper architectures are conducive automatically extract valid fault samples from big data.

Fault diagnosis methods based on deep learning are very meaningful for us to discover early fault of motors timely, improve system reliability and reduce maintenance cost [7]. Deep learning methods which includes convolutional neural network (CNN) [8] , deep neural networks [9] , sparse auto encoder (SAE) [10], recursive neural network (RNN) [11], deep belief network (DBN) [12], and generative adversarial net (GAN) [13] can provide effective solution to overcome the above limitations. Deep architectures have more multiple hidden layers than traditional networks. They are useful for learning representative features from the raw data directly, and automatically selecting discriminative representations of fault samples.

A large number of short circuit fault samples and deep learning based methods are the core to improve the stability of fault diagnosis based on professional knowledge. However, the structure of the PMSM is becoming more and more complex, which causes the short circuit to be invisible and random. Meanwhile the correlated interference usually leads to the homogenization of these samples and covers up the fault performance of the PMSM. So it is necessary to modify the diversity of the training samples and augment these limited features. Li et al. [14] adopted the GAN to generate fault samples and enhanced the diagnosis accuracy obviously. But the GAN has some defects such as the problem of convergence between a generative model and a discriminative model. The conditional generative adversarial net (CGAN) transforms the unsupervised GAN into a supervised model which makes the output data strongly correlate with the input samples [15]. In addition to the problem of the small samples and networks’ parameters, many types of optimization methods are proposed to accelerate the training speed of models, such as the adaptive sub-gradient method [16] and adaptive moment estimation (Adam) [17].

The denoising auto encoder (DAE) is an effective improvement of the auto encoder[18]. In order to make the hidden features more robust, a certain probability of noise is added to the training samples, where the input data of each hidden layer is randomly set to zero. Because the training mechanism of the auto encoder requires that the output data is as much as possible close to the input data without noise pollution, it means that the DAE learn how to remove the noise from the input data. The DAE enhances the generalization ability of the auto encoder in the condition of practical application. So we draw on this idea of the DAE for finding the way of noise injection in the background of detecting the ITSC.

The main contributions of this paper are summarized as follows.

• We proposed a fault diagnosis method based on the small sample of the PMSM, and completed an optimized solution with two types of fault features. The CGAN is used to augment the number of fault samples to increase the diversity of data, and a certain probability of noise is added to the training set to enhance the generalization ability of the SAE. The entire architecture of the optimized SAE (OSAE) was finally adjusted by these task-specific samples, and several hierarchical representations about the ITSC were learned from the training samples.

• The performance of the proposed method based on the CGAN and the OSAE is compared with the performance of deep learning algorithms such as the CNN, the RNN and the BP methods. The former leads to faster training and better accuracy in terms of fault classification of the PMSM.

The rest of this paper is organized as follows. We presented a diagnosis method based on the CGAN and the OSAE for the ITSC fault of the PMSM. Firstly, an electromagnetic torque and a negative sequence current are acquired to describe the fault status of the PMSM. Secondly, the CGAN is used to augment the sample data-set based on the previously combined characteristics. Finally, a concise expression of fault characteristics is created and optimized by the OSAE model, and then a softmax classifier is trained to detect the ITSC of the PMSM. Sections 2 and 3 discuss the principles and fault characteristics of deep learning-based models, respectively. The experimental analysis and conclusions are introduced in Section 4 and Section 5, respectively

The research using machine learning techniques is very fruitful in the field of fault diagnosis. Wang et al.[18] adopted the DAE combined with softmax classifier to accomplish 6 types of fault diagnosis of asynchronous motor with an accuracy over 95%. Tang et al. [19] proposed an adaptive DBN with Nesterov momentum for the diagnosis of the rotating machinery. Janssens et al. [20] developed a learning model based on the CNN for bearing fault detection, and Chen et al. [21] used the CNN to identify faults in gearboxes. Piotr et al. [22] adopted the RNN to realize the behavior of a chaos engineering system for fault detection. Sun et al. [23] presented the SAE to achieve samples’ learning for the identification of induction motor faults. If the number of hidden layer nodes is larger than the number of input nodes, a sparse restriction is added to the auto encoder for activating some nodes. Chang et al. [24] proposed a mutual-channel loss for improving the classification ability of deep networks, which is also fit for the SAE algorithm. These above models can be easily deployed in industrial Intelligence IoT [25].