基于隐马尔科夫模型抗体的自组织免疫网络多通道故障诊断技术研究

With the continuous progress of China's industrial upgrading, fault diagnosis is faced with the problem of multiple fault diagnosis under the conditions of the large amount of data and multi-channel. This grant funds use clustering and other data mining techniques to cluster one-dimensional time sequence of the feature value and utilize hidden Markov model (HMM) to learn the typical dynamic samples. Thus, fault dynamic trend vector is composed of the likelihood outputs of the HMMs. Then, structural property is introduced into artificial immune network by the innovative extension of the affinity definition. Fault is characterized by HMM antibody chain which means the specific response of specific antibody combinations. At last, fault is diagnosed by the topology and response vector of HMM antibody chain. So, hybrid intelligent diagnostic model is constructed in this application, by combining dynamic modeling capabilities of HMM and self-organizing ability of AIS. And fault dynamic trend vector is applied to establish contact between multi-fault with the multi-channel signals. The success of this project will advance the immune theories of AIS, and is expected to build new theories and new methods for dynamic fault diagnosis, such as intelligent manufacturing equipment fault diagnosis, with structure of multi-channel data signals / fault dynamic trend vector/ multiple fault diagnosis.

随着我国产业升级的不断推进，故障诊断领域面临着在多通道、大数据条件下进行多故障诊断的问题。申请首先通过对数据聚类等数据挖掘技术对特征值的一维时间序列进行聚类；再采用隐马尔科夫模型对聚类后的典型动态过程建模，从而通过模型似然率输出得到故障动态趋势矢量。然后通过创新性地对人工免疫网络的亲和力定义进行扩展，将结构特性引入人工免疫网络，采用隐马尔科夫模型抗体链这一特定抗体组合的特定响应表征故障；最后通过隐马尔科夫模型抗体链的拓扑结构以及响应值编码矢量对故障进行诊断。本项研究结合了隐马尔科夫模型的动态建模能力以及人工免疫系统的自组织能力等优点，构建了混合智能诊断模型，通过故障动态趋势矢量建立了多通道信号与多故障之间的联系。项目的成功实施有望进一步拓展人工免疫系统的免疫响应方式与原理，构建针对智能制造装备等强调动态诊断的"多通道大数据信号--故障动态趋势矢量--多故障诊断"的故障诊断新理论与新方法。

设备的安全可靠是设备运行的首要条件，故障诊断系统的作用十分重要。以数据学习为基础的智能诊断系统存在结果可解释性较低以及可靠性受限制于原始数据。本基金资助项目提出HMM抗体的概念，通过最优状态序列联系AIS系统，分别对基于HMM 抗体的人工免疫网络构建技术、基于HMM 聚类技术、故障动态过程表征技术、以及HMM 抗体免疫网络的训练与识别技术进行了研究。最终提出了HMM逆Viterbi评估算法、HMM模型逆Viterbi聚类AIS训练算法、HMM模型自聚类最优组合选择算法、柔性HMM模型及其训练识别算法、采用柔性HMM模型的故障动态趋势矢量检索技术以及形态分量分析与双谱分析的故障动态过程特征表征技术。项目组建了可实现耦合故障模拟的转子故障实验平台，基于NI-Labview采集系统、Matlab与C++开发了对应的HMM故障诊断系统。最终通过轴承特征提取实验验证了形态分量分析与双谱分析技术的有效性，再通过对实验台31种设备状态(耦合故障)的故障信号采用HMM 抗体人工免疫网络诊断系统进行诊断。诊断结果表明系统的多类故障诊断性能达到要求，同时模型训练自动化水平高，模型特征可在大范围内自动筛选，系统适应性好，具备良好的应用潜力。