基于机器学习的故障预测与健康管理（PHM）方法研究

发布时间：2018-08-04 08:16

【摘要】：在一些重要场合,设备的失效会造成巨大的经济损失。而科技的发展使得设备的内部结构日益复杂,人工维护的难度和开销越来越高。故障预测与健康管理技术在几十年的发展后已经成为了降低设备维护开销的重要手段,在多处场合有了广泛的应用。而对于数字电路的故障预测尚处于探索阶段。介于数字电路是人们生活之中不可或缺的一部分,其故障预测与健康管理方案的研究具有很高的价值。本文通过国内外的故障预测与健康管理研究动向,总结了故障预测与健康管理技术的研究实现方法,对针对数字电路的故障预测与健康管理方案进行了开拓性研究。结合了目前热门的机器学习方法,本文以74LS280为案例,分析了其芯片的特征与相关的性能指标,在老化测试数据中挖掘其性能退化的数据特征,提出预测故障发生的方法。该方法具有容易实现,泛用性较强的特点。为了完成本文的实验内容,本文设计了用于测试74LS280的特征参量的平台。平台设计包含选取传感器、模拟使用环境的硬件平台,以及用于读取测试数据并传输至PC机的软件平台。基于支持向量机方法,本文利用74LS280的测试数据建立一个故障模型,利用该模型诊断其它芯片的故障与否。作为延伸,本文给出了两种数据预处理方式用于优化训练模型,以达到更快的训练结果或者更低的模型误差。模型建立之后,本文提出了利用最小二乘法估计74LS280的健康值走势,预测剩余寿命,并提出了惯性存储数据存储方法,以改善寿命预测值。由于数据随时间的变化并非呈现出单调的趋势,虽然本文的剩余寿命预测值无法保证数值的准确性,但其足够起到对于维护人员的提示性作用,具有一定的实用性。
[Abstract]:In some important occasions, equipment failure will cause huge economic losses. With the development of science and technology, the internal structure of equipment is becoming more and more complex, and the difficulty and expense of manual maintenance are becoming higher and higher. After decades of development, the technology of fault prediction and health management has become an important means to reduce the cost of equipment maintenance, and has been widely used in many situations. The fault prediction of digital circuit is still in the exploration stage. Digital circuit is an indispensable part of people's life, and the research of its fault prediction and health management scheme is of great value. Through the research trends of fault prediction and health management at home and abroad, this paper summarizes the research and implementation methods of fault prediction and health management technology, and makes a pioneering study on fault prediction and health management schemes for digital circuits. Combined with the popular machine learning methods, this paper takes 74LS280 as an example, analyzes the chip features and related performance indexes, mining the data features of its performance degradation in the aging test data, and puts forward a method to predict the occurrence of faults. This method is easy to implement and has strong universality. In order to complete the experiment, a platform is designed to test the characteristic parameters of 74LS280. The platform design includes selecting sensors, simulating the hardware platform of using environment, and software platform for reading test data and transmitting to PC. Based on the support vector machine (SVM) method, a fault model is established by using the test data of 74LS280, which is used to diagnose the faults of other chips. As an extension, two data preprocessing methods are presented to optimize the training model to achieve faster training results or lower model errors. After the model is established, this paper proposes to estimate the trend of 74LS280 health value by least square method and to predict the residual life, and to improve the life prediction value by using the method of inertial storage data storage. Because the change of data with time does not show a monotonous trend, although the prediction value of residual life in this paper can not guarantee the accuracy of the numerical value, but it has enough to serve as an indication to the maintenance personnel, it has certain practicability.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN79

【相似文献】