基于免疫系统的小样本在线学习异常检测与故障诊断方法
发布时间：2019-05-28 19:31
【摘要】：设备故障样本缺乏、状态检测与故障诊断分离、训练与测试过程相互独立是制约现有智能故障诊断方法广泛应用的主要原因。借鉴生物免疫机理，开展对设备适应性强、对故障样本依赖程度低，并且具有连续学习能力的设备在线学习异常检测与故障诊断方法具有重要的科学意义。 为了提高传统实值反面选择算法检测器的覆盖率和减少冗余检测器，提出了固定边界反面选择算法、精细固定边界反面选择算法、基于边界样本的界面检测器和基于约简边界样本的界面检测器。在深入讨论算法的基础上，应用15组2维人造数据集和Iris数据集进行仿真实验，分析了四种检测器的异常检测性能。与其它异常检测算法相比，训练样本参数相同时，多数情况下，此四种算法具有更好的检测性能；另外，当检测率相近时，检测器(边界样本)数量依次减少。 在全面分析固定边界反面选择算法和基于边界样本界面检测器特性的基础上，提出了小样本在线学习固定边界反面选择算法和小样本在线学习界面检测器，并分析了造成两种算法过学习与欠学习的原因。通过仿真实验，，讨论了这两种小样本在线学习异常检测算法相对于传统反面选择算法的优势，分析了小样本在线学习界面检测器优于小样本在线学习固定边界反面选择算法的原因。借鉴免疫系统的疫苗机理，在小样本在线学习界面检测器算法的基础上引入活性疫苗克服了欠学习降低了误报警率，引入惰性疫苗抑制了过学习提高了检测率。 在深入分析界面检测器特性的基础上，引入异常度和异常等级两个概念，结合界面检测器的连续学习特性，提出了具有小样本在线学习异常检测与故障诊断能力的自适应超环检测器。将分布在非己空间内有限的故障样本构建小样本在线学习故障界面检测器，并引入类间隶属度概念，实现了对已知类型故障样本分类、未知类型故障样本聚类的功能。 使用轴承故障数据进行仿真，讨论了各种条件下自适应超环检测器的小样本在线学习异常检测与故障诊断性能，与其它故障诊断方法相比，自适应超环检测器的诊断准确率更高。自适应超环检测器不仅实现了异常检测与故障诊断一体化，而且具备在线学习能力；不仅具备小样本故障诊断能力，还能识别未知类型故障；不仅能随时加入故障样本，还具备数据压缩功能，具有广泛的应用前景。
[Abstract]:The lack of equipment fault samples, the separation of state detection and fault diagnosis, and the independence of training and testing processes are the main reasons that restrict the wide application of the existing intelligent fault diagnosis methods. Based on the biological immune mechanism, it is of great scientific significance to carry out the methods of equipment online learning anomaly detection and fault diagnosis, which have strong adaptability to equipment, low dependence on fault samples and continuous learning ability. In order to improve the coverage of the traditional real value negative side selection algorithm and reduce the redundant detector, a fixed boundary inverse selection algorithm and a fine fixed boundary reverse surface selection algorithm are proposed. The interface detector based on boundary sample and the interface detector based on reduced boundary sample. On the basis of in-depth discussion of the algorithm, 15 groups of 2D artificial data sets and Iris data sets are used to carry out simulation experiments, and the anomaly detection performance of four kinds of detector is analyzed. Compared with other anomaly detection algorithms, when the training sample parameters are the same, the four algorithms have better detection performance in most cases, in addition, when the detection rate is similar, the number of detector (boundary samples) decreases in turn. Based on the comprehensive analysis of the fixed boundary inverse selection algorithm and the characteristics of the interface detector based on the boundary sample, a small sample online learning fixed boundary negative surface selection algorithm and a small sample online learning interface detector are proposed. The causes of overlearning and underlearning of the two algorithms are analyzed. Through simulation experiments, the advantages of these two small sample online learning anomaly detection algorithms over the traditional negative selection algorithm are discussed. The reason why the small sample online learning interface detector is superior to the small sample online learning fixed boundary negative selection algorithm is analyzed. Based on the vaccine mechanism of immune system, the active vaccine is introduced on the basis of small sample online learning interface detector algorithm to overcome underlearning and reduce the false alarm rate, while the introduction of lazy vaccine suppresses overlearning and improves the detection rate. Based on the in-depth analysis of the characteristics of the interface detector, the concepts of anomaly degree and anomaly grade are introduced, and the continuous learning characteristics of the interface detector are combined. An adaptive hyperloop detector with the ability of small sample online learning anomaly detection and fault diagnosis is proposed. The fault samples distributed in non-self space are constructed to construct small sample online learning fault interface detector, and the concept of inter-class membership degree is introduced to realize the function of classification of known types of fault samples and clustering of unknown types of fault samples. The performance of small sample online learning anomaly detection and fault diagnosis of adaptive hyperloop detector under various conditions is discussed by using bearing fault data. Compared with other fault diagnosis methods, The diagnostic accuracy of adaptive hyperloop detector is higher. The adaptive hyperloop detector not only realizes the integration of anomaly detection and fault diagnosis, but also has the ability of online learning, not only the ability of small sample fault diagnosis, but also the ability to identify unknown types of faults. It can not only add fault samples at any time, but also has the function of data compression, and has a wide range of application prospects.
【学位授予单位】：上海大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP274;TH165.3
[Abstract]:The lack of equipment fault samples, the separation of state detection and fault diagnosis, and the independence of training and testing processes are the main reasons that restrict the wide application of the existing intelligent fault diagnosis methods. Based on the biological immune mechanism, it is of great scientific significance to carry out the methods of equipment online learning anomaly detection and fault diagnosis, which have strong adaptability to equipment, low dependence on fault samples and continuous learning ability. In order to improve the coverage of the traditional real value negative side selection algorithm and reduce the redundant detector, a fixed boundary inverse selection algorithm and a fine fixed boundary reverse surface selection algorithm are proposed. The interface detector based on boundary sample and the interface detector based on reduced boundary sample. On the basis of in-depth discussion of the algorithm, 15 groups of 2D artificial data sets and Iris data sets are used to carry out simulation experiments, and the anomaly detection performance of four kinds of detector is analyzed. Compared with other anomaly detection algorithms, when the training sample parameters are the same, the four algorithms have better detection performance in most cases, in addition, when the detection rate is similar, the number of detector (boundary samples) decreases in turn. Based on the comprehensive analysis of the fixed boundary inverse selection algorithm and the characteristics of the interface detector based on the boundary sample, a small sample online learning fixed boundary negative surface selection algorithm and a small sample online learning interface detector are proposed. The causes of overlearning and underlearning of the two algorithms are analyzed. Through simulation experiments, the advantages of these two small sample online learning anomaly detection algorithms over the traditional negative selection algorithm are discussed. The reason why the small sample online learning interface detector is superior to the small sample online learning fixed boundary negative selection algorithm is analyzed. Based on the vaccine mechanism of immune system, the active vaccine is introduced on the basis of small sample online learning interface detector algorithm to overcome underlearning and reduce the false alarm rate, while the introduction of lazy vaccine suppresses overlearning and improves the detection rate. Based on the in-depth analysis of the characteristics of the interface detector, the concepts of anomaly degree and anomaly grade are introduced, and the continuous learning characteristics of the interface detector are combined. An adaptive hyperloop detector with the ability of small sample online learning anomaly detection and fault diagnosis is proposed. The fault samples distributed in non-self space are constructed to construct small sample online learning fault interface detector, and the concept of inter-class membership degree is introduced to realize the function of classification of known types of fault samples and clustering of unknown types of fault samples. The performance of small sample online learning anomaly detection and fault diagnosis of adaptive hyperloop detector under various conditions is discussed by using bearing fault data. Compared with other fault diagnosis methods, The diagnostic accuracy of adaptive hyperloop detector is higher. The adaptive hyperloop detector not only realizes the integration of anomaly detection and fault diagnosis, but also has the ability of online learning, not only the ability of small sample fault diagnosis, but also the ability to identify unknown types of faults. It can not only add fault samples at any time, but also has the function of data compression, and has a wide range of application prospects.
【学位授予单位】：上海大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP274;TH165.3
【参考文献】
相关期刊论文 前10条
1 孟庆华;赵文礼;樊志华;曾复;;基于改进型阴性选择算法的车辆故障检测方法研究[J];兵工学报;2009年12期
2 陈强;郑德玲;李湘萍;;基于人工免疫的故障诊断模型及其应用[J];北京科技大学学报;2007年10期
3 马立玲;张f
本文编号：2487311
本文链接：https://www.wllwen.com/kejilunwen/jixiegongcheng/2487311.html
