基于DCNN和图像识别的电动汽车用PMSM故障诊断技术研究

In order to avoid security incident from the failure of PMSM in electric vehicle, it is necessary to monitor the operation status of the motor regularly or in real time. Under considering the influences of the motor speed, load, inverter power supply and closed loop control, it is very difficult to acquire the high accuracy of fault diagnosis, and this is now facing a big problem. In order to solve the problem, this project will realize the fault classification and fault severity identification of the motor by recognizing image way based on the deep convolutional neural network (DCNN) and the deep learning method. First of all, a simulated driving system of electric vehicle is constructed, and the current data of PMSM is sampled by running the system, the current data are converted into images classified by the fault types of motor, they can formed the labeled image database for fault diagnosis of PMSM. Second, determining the number of convolution kernel, convolution activation layer and pooling layer, and selecting a classifier and determining the corresponding the neuron number, we design a DCNN which is suitable for fault diagnosis for the PMSM. It can extract more fault characteristic information from the motor current, so can achieve high accuracy of fault diagnosis; Finally, the fault features extracted from the motor current by the deep neural network are studied, the new fault features in the motor current will be find out. The above research explores the application of deep learning and image recognition in PMSM fault diagnosis of electric vehicles, which provides a new ideas for motor fault diagnosis.

为了避免电动汽车用永磁同步电机(PMSM)出现故障而造成安全事故，需定期或实时监测电机运行状态。在考虑电机转速、负载、逆变器供电和闭环控制的影响下，要高精度地诊断电机故障非常困难，这是目前面临的一大难题。为了解决此难题，本项目拟利用深度卷积神经网络（DCNN）和深度学习方法以识别图像的方式，实现电机故障分类和故障严重程度识别。首先，构建并运行电动汽车用模拟驱动系统，采集电机电流数据，将其转化为能按电机故障类型分类的图像，形成用于故障诊断的带标签的图像数据库；其次，确定卷积核、卷积激活层和池化层的数量，选择分类器并确定其神经元个数，设计适用于电机故障诊断的DCNN，从电机电流中提取更多的故障特征信息，实现高精度的故障诊断；最后研究DCNN从电流中提取的故障特征，找出电流中蕴含的新故障特征。上述研究探索了深度学习和图像识别在电动汽车用PMSM故障诊断中的应用，这为电机故障诊断提供了新思路。

将电动汽车用电驱动系统的电机运行电流转化为灰度图像，利用深度卷积神经网络（DCNN）分类图像具有的独特优势，以此对电动汽车电机进行故障诊断。这可有效地抑制信号采集时噪声、电机闭环控制和电机负载变化对电机故障诊断带来的影响，对提高电机故障诊断精度，确保电动汽车安全运行具有十分重要的意义，该技术有很好的应用前景。课题组围绕项目计划书的研究内容展开了较为全面深入的研究工作，取得了较好的研究成果。在综合分析电动汽车用永磁同步电动机（PMSM）故障机理的基础上，用有限元法建立含故障的PMSM的仿真模型，建立了电动汽车电驱动系统的仿真系统；利用自相关矩阵方法将电机运行电流转化为二维灰度图像，并理论证明了该转化的有效性，构建了用于故障诊断的图像数据集；深入研究深度卷积神经网络以及进行图像分类的原理，设计了用于电机故诊断的DCNN,实现了电机的故障诊断；从DCNN的某些卷积层取出输出图像，并用特征图和热图标注其特征，找出特征与电机故障类型和强度的对应关系；定制带各类故障的PMSM，搭建了电动汽车模拟电驱动控制系统。根据项目任务书，课题组精心安排研究成员，创造性地开展了工作，取得了较好的研究成果和人才培养工作。四年来，发表相关科研文章共10篇,sci检索7篇，EI检索1篇，主要发表在International Journal of Control, Automation and Systems，Control Enginee-ring and Applied Informatics, Applied science, Identification and Control等国内外重要核心期刊和国际会议上；授权发明专利5项;培养青年教师1名，硕士研究生5名，其中柳波海获国家励志奖学金。按照项目计划书的研究计划要点，本项目的研究内容均已完成，基本达到预期目标，后续研究工作仍将进行。

内容获取失败，请点击重试