往复机械振动的双树复小波包分析与在线异常检测方法研究

Because of lack of prior knowledge, fault samples and fault mode of machinery condition monitoring, the precision of fault diagnosis is hard to improve. According to the characteristics of multi interference, non-stationary, coupling and periodicity of large and complex reciprocating machine vibration signal, accurate diagnosis is hard to realize. So a method of anomaly detection which based on data-driven of reciprocating machinery is proposed in this project. The new type one-class support vector machine anomaly detection is established to find the fault and alarm in time before the problem of sample distribution of high dimensional kernel feature space and optimum classification hyper plane is analysized; The machinery vibration signal process method of dual-tree complex wavelet packet transform which has the advantage of translation invariance and sparse decomposition is used to improve the signal de-noising and multivariate feature extraction effect. Then a combined multi-objective optimization algorithm will be studied to solve the coupling problem of feature extraction, feature selection and classifier parameters; After analysis new added samples effect of anomaly detection model, a on-line incremental train method is proposed to made the anomaly detection model has the ability of self-learning and generality;? The research results of project will not only expand the application field of fault diagnosis but also provide an new effective, practical detection method for reciprocating machinery early failure and complex fault recognition.

机械状态监测面临的先验知识不足、故障样本稀缺、故障模式不完备，是制约在线监测与故障诊断精度提高的根本原因。针对大型复杂往复机械所监测的振动信号具有干扰多、非平稳、耦合性、强周期性等特点，难以直接实现准确诊断。本项目拟研究一种基于数据驱动的往复机械在线异常检测新方法，实现及时发现并进行故障报警，课题重点分析高维核特征空间样本分布问题和最优分类超平面思想，建立新型单类支持向量机异常检测模型；围绕异常检测模型，研究双树复小波包平移不变性和稀疏分解能力，利用其提高信号降噪和多元故障特征的综合表征能力；针对异常检测面临的特征提取、特征选择和分类器参数耦合难题，提出一种联合多目标优化方法；分析新增样本对异常检测模型的影响，提出一种在线增量式训练算法，实现异常检测的自学习和通用性。本项目研究成果可拓展到其他状态监测与故障诊断领域，为机械早期和复合等故障检测识别难题提供一套有效、实用的新方法。

针对复杂往复机械状态监测信号存在干扰多、非平稳、耦合性、强周期性等特点，测试数据存在正常状态充足、故障状态稀缺的不平衡问题和故障模式不完备等问题，状态监测与故障诊断困难。本项目针对上述问题开展了往复机械异常检测研究：1、利用高频计数器开发了瞬时转速精确测试装置，针对远程状态监测的需求和复杂信号特点，研究了自适应压缩感知的数据压缩方法，提高了压缩率。提出应用品质因子可调小波变换和正交变分模式分解的两种复杂机械信号稀疏分离方法，突出了机械信号的瞬态冲击特征。2、结合双树复小波的优点，提出了重叠块阈值和空域邻域联合降噪两种方法和融合时域、频域和时频域的复杂信号特征提取方法，有效地预制噪声干扰，实现特征参数的有效提取。3、针对机械故障异常检测问题，建立了双控制比例因子的最大间隔超球体分类器异常检测模型和基于Bayes的超球体异常检测模型，UCI数据和实测信号表明能显著提高异常检测的效果。利用改进粒子群的多目标优化方法解决了机械故障异常检测中存在的特征选择、分类器参数优化等耦合难题。4、分析了该异常检测模型的增量学习法则，提出了一种基于快速密度核聚类的大规模训练数据约简方法，利用集成协同半监督学习，获取和利用大量未标记数据数据信息，并构建了集成多个单分类支持向量机，实现了对多故障类型的准确识别。5、为了更深层地挖掘复杂状态监测信号的故障信息，提出了一种基于深度置信网络的发动机故障诊断模型，通过对原始振动数据逐层的学习，自动提取数据中的内在特征，并完成发动机健康状态的识别。本项目取得的研究成果为往复机械提供一套新的监测诊断方法，可以扩展应用于指导相关机械装备的异常监测与诊断中，具有较好的理论和工程应用价值。