【作者】 刘曼兰；

【导师】 崔淑梅；

【作者基本信息】 哈尔滨工业大学 ， 电机与电器， 2007， 博士

【摘要】 小功率永磁直流电机因其具有结构简单、运行可靠、体积小、重量轻、效率高等优点,被广泛地应用于汽车工业以及航空、航天领域。为了这些电机的安全运行,本文研究了一种简便的永磁直流电机故障在线监测与智能诊断方法。针对在线故障监测的实时性与简便性的要求,提出了一种基于电流信号分析与处理的永磁直流电动机常见故障的诊断方法。以易于检测的永磁直流电机电枢电流为故障诊断的信号源,综合利用基于信号处理的方法来实现故障特征的提取,基于人工智能方法实现故障模式识别。在分析了永磁直流电动机的故障机理和电机动、稳态电流在正常与故障状态下的变化特性的基础上,建立了基于电枢电流分析的永磁直流电机故障诊断的数学模型。确立稳态电流i的均值ia v、稳态电枢电流频谱谱峰对应的脉动频率f及其脉动幅度i std、电机起动电流的峰值im和峰值点附近的下降速率k这五个量确立为永磁直流电机故障诊断的特征参数,并将它们相对于正常状态的变化量组成故障特征向量: Te = [ΔfΔi_avΔi_stdΔi_mΔk ]。以数台永磁直流电机为实验样本,对其常见的元件开路、匝间短路、电刷磨损,绕组脱焊四种故障进行空、负载电枢电流故障特征提取及故障机理分析,实验结果与故障机理分析的一致性表明:本文提出的基于电流分析的永磁直流电机故障诊断方法对永磁直流电机的常见故障的诊断具有通用性。故障模式识别是实现故障智能诊断的关键环节。专家系统在故障诊断领域中的应用,实现了基于人类专家经验知识的故障模式智能识别技术。但由于永磁直流电机故障特征参数的分散性、随机性和模糊性较大,获取有效的永磁直流电机的诊断知识成为建造故障模式识别专家系统的关键问题之一。机器学习是解决知识获取问题的主要途径。目前应用较为广泛的基于机器学习的故障模式识别方法主要有统计模式识别方法和人工神经网络方法。但统计模式识别方法和人工神经网络方法都是建立在传统的统计学基础之上的分类算法,即只有在足够多的样本前提下,其算法才是合理的,而当训练的是小样本数据集时,就不宜再沿用该分类算法。工程应用中的永磁直流电机尤其航空航天应用领域中的电机不可能带故障长期运行,故难以获得大量典型故障样本。这就在一定程度上制约了统计模式识别和人工神经网络方法在该领域中的应用。本文以支持向量机算法的数学模型为理论依据,构建了一种基于支持向量机的永磁直流电机故障模式识别分类器,对永磁直流电机的数种故障模式进行识别,并对其和应用BP神经网络识别方法在小样本故障模式识别中的应用进行了对比实验研究。结果表明:基于支持向量机的故障模式识别方法在小样本情况下的诊断精度要高于同样样本情况下的BP神经网络,且避免了基于BP神经网络方法中存在的过学习问题和陷入局部极小值问题。针对永磁直流电机故障在线诊断中存在类样本数目不平衡、误判损失不等、在线样本数据缺少类别标识以及复杂环境中样本数据存在大量噪声野点等问题,本文对基于支持向量机的永磁直流电机故障模式识别算法作了如下两方面的改进:首先,通过对支持向量机数学模型中的误差惩罚因子进行加权,构建了一种基于加权支持向量机的永磁直流电机故障模式识别算法。理论分析和实验结果表明:该算法可以提高小样本类（故障样本类）诊断精度,降低误判损失。其次,采用了一种基于模糊C均值的聚类算法,对无类别标识的在线数据进行模糊聚类,并根据模糊聚类的隶属度,来判断每个样本数据的所属类,由此同时定位数据中的野点,消除野点后,再对消噪后的数据运用基于支持向量机的模式识别算法进行训练与测试。提高了永磁直流电机复杂运行工况下在线样本数据的故障识别精度,扩充了基于支持向量机的故障模式识别算法的应用领域。本方法只需要监测电机电枢电压与电流,不需采集更多的信号,对传感器等硬件的要求小,实现简单,对系统影响小,成本低。在电机生产线的质量控制、汽车与航空与航天器中电机的在线状态监测方面具有实用性,会有广泛的应用前景。更多还原

【Abstract】 Permanent-magnetic DC motors are widely used in automobile and military or civil airplanes due to their lots of merits such as simple configuration, reliable running, small size, light weight, high efficiency, etc. To ensure these motors work safely, this paper studies a simple and efficient method to monitor the condition of permanent-magnetic DC motor and diagnose the possible online faults intellectualized.In order to meet the requirement of real-time and simplicity in online faults detection,a universal diagnosing method was given to detect the typical faults of permanent-magnetic DC motor based on current signal analyzing and processing. Several signal analysis methods were synthetically used to extract the characteristic of failure from the armature current which is easy to be detected. Then the fault pattern recognization was made by artificial intelligence method.After analyzing faults mechanisms and the changing characteristic of dynamic or stable current in cases of both failure and no failure, mathematical model of faults diagnosing based on armature current was created. Parameters: average ia v of stable currenti , pulsant frequency f , pulsant range is td, peak amplitude im of starting current and its gradient k near the peak, were defined as characteristics for faults diagnosis of permanent-magnetic DC motor. Differences of the five partners in case of failure and no failure made up the characteristic vector Te = [ΔfΔi_avΔi_stdΔi_mΔk ].Several failed experiments in no loading or loading condition are done on motors with four common faults in production or application: open circuit of components, brush wear, short circuit between coils and loose weld coils. And their fault mechanism was also analyzed in theory. The consistency between experimental results and the failed mechanism analysis shows that this method is feasible and universal to diagnose the faults of permanent-magnetic DC motor.Pattern recognition is the key to realize fault intelectual diagnosis. The application of Expert System in field of fault diagnosis realized fault diagnosis intellectually. Due to the decentralization, randomicity and fuzziness of the fault character of permanent-magnetic DC motor, it is hard to get complete and efficient knowledge for creating the expert system for fault diagnosis of permanent-magnetic DC motor. The main approach to solve problem of knowledge acquisition is Machine Learning.The Statistical Pattern Recognition and the Artificial Neural Network are the basic means which are most widely used in fault diagnosis based on machine learning at present. But the theoretical basis of Statistical Pattern Recognition and Artificial Neural Network is traditional statistics. This kind of sort algorithm is valid only in the case of the number of trained sample being infinity and the results from this kind of sort algorithm is invalid when the number of trained sample is small. However, it is difficulty to gain a great deal of representative fault samples because it is impossible to let the motor long running with failure in engineering application especially used in airplanes.Fault diagnosis method based on SVM （Support Vector Machine） is developed for permanent-magnetic DC motor according to the mathematical model of SVM. The fault diagnosis results of this method in cases where only limited training samples are available is compared with that of another classification algorithm BP ANN. It shows that SVM have better performance than ANN both in training speed and recognition rate. SVM can also avoid over-fitting and trapping in local extreme which often happened in the neural networks algorithm, especially in case of limited trained samples.To overcome the problems existing in the online fault diagnosis of permanent-magnetic DC motor, such as non-symmetry of dataset, different loss by misjudgments and interference of noisy or outliers, the recognition algorithms of SVM is improved in following two ways. Firstly, a weighted support vector machine algorithm is developed through weighting error punishing factor of SVM. Both results of several experiments and analysis in theory show that this weighted support vector machines improve classification accuracy for class with small size, and reduce the different loss by misjudgments in fault diagnosis. Secondly, an improved support vector machine algorithm based on fuzzy C-means is proposed. The online data are clustered by the fuzzy C-means and the outliers are recognized according to the membership grade calculated from the fuzzy C-means. And then the data which removed the outliers are trained and tested by the support vector machine algorithm above mentioned. The results from experiments shows that support vector machine algorithm based on fuzzy C-means have better tolerance of noise and anti-noise performance. It enhances fault recognition precision in the complex case and extends the application field of fault diagnosis method based on SVM.This method needs no more messages than measured voltage and current of armature. Therefore, the advantages are lower request for hardware like sensor, more simple realization, less influence on system and lower cost. It is practical and of good perspective in online monitoring online the condition of a working permanent-magnetic DC motor in an automobile or airplane or space probe, and in controlling the quality of motor manufacture line.更多还原