【作者】 岑健；

【导师】 胥布工；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 随着旋转机械设备的大型化、高速化、复杂化,复合故障发生的概率也在不断的增大。由于多重故障并发时,不同故障特征相互混杂呈现出多耦合、模糊性等复杂征兆,并非多个单故障的简单线性叠加,很难用准确的数学模型加以描述,也难以完全依靠确定性判据进行故障诊断,复合故障具有复杂性、相关性和不确定性的特点,给诊断带来了极大的困难。本文以化工机组复合故障为主要研究对象,以人工免疫与其他智能方法集成来提高复合故障的准确率为研究内容,分析、研究机组复合故障的特性,优化特征分类能力,形成核心算法,研究依托免疫系统分布性、多样性原理的多类无量纲免疫检测器集成诊断技术,通过智能整合思想为问题的解决提供一条可行的途径,深入探索复合故障精确诊断的机理,研究出有效诊断方法,突破复合故障诊断难的瓶颈,为人工免疫系统在机组复合故障诊断技术的工业应用打下理论基础及提供技术支撑,在工业机组复合故障诊断领域实现智能化、高诊断准确率的目标。具体研究如下:考察和分析现有的故障检测器生成算法,发现基于人工免疫的算法中一般都针对一定的匹配规则,目前都没有很好地解决阈值的问题,预设常数阈值对不同的故障存在局限性;应用单一匹配规则,生成检测器的质量不高,会生成错误的检测器,从而在检测时产生比较大的误报率;同时还会产生检测器集的黑洞问题,在检测时造成比较大的漏报率和误报率;而且现有算法需要比较大的检测器集合来保证较高的检测率,导致每次检测器生成过程和检测过程花费较多时间。为此,研究一种故障检测器生成算法,基于形态学空间的分析,提出利用检测器串覆盖问题的思想来生成检测器集合,该算法在进化过程中实现匹配阈值的自适应变化,有效减少黑洞和边界不清晰的问题,防止故障误检和漏检。构造一种符合人工免疫系统多样性、分布性原理的高效的、快捷的检测器训练方式和算法,生成高效的免疫检测器,提高故障检测准确率。针对多重故障并发时,不同故障特征相互混杂呈现出多耦合、模糊性等复杂征兆,提出免疫编程的特征构造方法。通过抗体集的进化来获得最优抗体,得到具有最佳识别能力的新特征指标,解决分类能力不足的问题,通过算法训练后获得最佳分类能力的成熟免疫检测器,进一步提高免疫检测器对复合故障的分类能力。针对目前采用神经网络、模糊聚类理论、专家系统等集成的算法存在计算复杂、在线性差,难于满足诊断快速性要求。把多重复合故障看成单独一类新出现故障参与到整个故障空间中进行训练,定义5种时域指标生成的无量纲免疫检测器同时进行交叉检测,利用证据理论的决策系统作进一步决策,通过多信息融合技术,推导出能诊断复合故障的优秀免疫检测器,研究出一种更为简单、高效、快速、实用的集成诊断算法,形成基于多类无量纲免疫检测器集成的机组复合故障诊断技术。利用人工免疫系统的特性,根据缓变故障的特点,采用阴性选择算法,选取合适的编码位数,对缓变故障特征进行提取和约简,得到独有故障特征,再集成诊断,通过实验验证了这种方法对缓变的复合故障诊断的有效性。把研究成果应用于工业机组的故障诊断中,“橡胶装置二线挤压脱水机GY6204及膨胀干燥机GY6205智能故障诊断系统”的投入使用,为人工免疫系统在故障诊断的工业应用上打下了基础,初步实现了工业机组复合故障领域智能化、高诊断准确率的目标。通过本文的研究,尤其是人工免疫系统在复合故障诊断问题的深入探讨,为今后进一步的研究打下了坚实的基础,同时也发掘出了一些新的研究问题和研究思路。更多还原

【Abstract】 Due to the high speed and complexity of large-scale rotating machinery, complex fault probability is also constantly increasing. When multiple concurrent faults happen, those different faults are mixed with each other and feature multiple-coupling and ambiguity, they can’t be reduplicated by simple linear fault, so it’s hard to accurately describe the mathematical model of those faults and it’s very difficult to diagnose them through deterministic criteria. Complex fault has the features of complexity, correlation and uncertainty; therefore it is very difficult to diagnose. This paper aims to discuss the problem of accurate complex fault diagnosis of machine unit, by means of artificial immune method combining with other intelligence methods, the accuracy ratio of complex fault diagnosis can be improved, and a feasible way to complex fault diagnosis may be provided through the idea of intelligent integrated. This article analyzes the characteristics of unit complex faults, optimizes feature classification, forms the core algorithm, and based on principle of immune system distribution and diversity, studies the integrated diagnostic technology to more types of non-dimensional immune detectors. Precise diagnosis mechanisms of complex fault are studied further in this paper. In order to break the bottleneck of complex fault diagnosis, the effective diagnosis method will be developed. Those methods can form a solid theoretical foundation and provide strong technical support for artificial immune system application to unit complex faults diagnose, and those methods can help to carry out intelligence and high diagnostic accuracy ratio of industrial units complex fault diagnosis. The main contents of this dissertation are summarized as followings:Investigation and analysis of the existed generating algorithms to fault detector, we find that they have fixed matching rules based on artificial immune algorithm; the threshold problem is currently not well resolved. The default constant threshold has limitations for different faults. The quality of detectors generated by using single matching rule is not very satisfactory, sometime, error detectors will be generated, and it can result in larger detection false. In the same time it will produce a detector set black hole, and lead to larger false rate and omitting rate. In order to guarantee the higher detector rate, the existed algorithm needed larger detector set, but it take long time to generate detector and detect. In this paper, detector-generating algorithm is studied. Based on the form analysis, the detector set can be generated by the technology of detector covering; the algorithm can adaptively match on the threshold in the course of evolution. According to the principle of the diversity and distribution of artificial immune system, the detectors set can be optimized, efficient and fast detector training method and algorithm can be proposed through the principle of the diversity and distribution of artificial immune system. The more efficient detector can reduce black holes and boundary uncertainty, and prevent misdiagnosis and omitting diagnosis.As concurrent multiple faults happen, different fault features mix with each other and feature multiple coupling and ambiguity, the algorithm of characteristics construction of immune programming is proposed. By through of evolution of antibodies set, the optimal antibodies can be obtained and the new features indicator of best recognition ability can be formed, it can solve the problem of insufficient classification capacity. By algorithm training, mature immune detector with best classification ability can be obtained such that the detector further enhances the immune detector classification capability to the complex fault.Considering that the use of neural networks, fuzzy clustering theory, expert systems integration algorithms has complexity and on-line diagnosis can not meet the fast requirements in recent years, through the multiple complex fault participating in the non-ego training space as a new type fault, the dimensionless immune detectors have been generated by defining five kinds of time domain performance index, those detectors can crossly test. Using decision-making systems of the evidence theory, the excellent detector can be derived through multi-information fusion technology, the detector can diagnose complex fault. A kind of integrated diagnosis algorithm with more simplicity, efficiency, rapidity and practice has been proposed, and basing on integration of some class of dimensionless immune detectors, the unit complex fault diagnosis technology has been formed.Using characteristics of the artificial immune system, negative selection algorithm is adopted and the appropriate coding bits are selected based on characteristics of slowly-varying fault, slowly-varying fault feature can be extracted and reduced, unique fault feature can be obtained, and then the integrated diagnosis can be proceed. Experiment has shown the effectiveness of the method to slowly varying complex fault diagnosis.The research results have been applied to fault diagnosis of industrial units. The intelligent diagnosis systems such as the rubber device GY6204 and GY6205 have been used in the industrial area. Those use of fault diagnosis technology in the industrial sector has made a solid foundation for artificial immune system, and the intelligence and high accuracy ratio of complex fault diagnosis can be realized in the field of industrial unit complex fault diagnosis.Through some study in our paper, the especially further research about the use of the artificial immune system to complex fault can not only make a solid foundation in this field, but also help us to discover some new research problems and possible solutions.更多还原