针对BA-FRVM的研究及用于汽车典型故障件的数量预测

发布时间：2018-05-29 19:30

本文选题：汽车故障件 + 预测　；参考：《西南交通大学》2017年硕士论文

【摘要】：近年来,随着汽车企业的快速发展,该行业数据量激增,汽车企业迫切需要在数据中去寻找规律,根据历史数据预测配件需求数量、诊断汽车故障等。在这样的背景下,许多人工智能算法开始被应用于汽车行业,其中,运用最多的是SVM和BP神经网络。鉴于BP神经网络和SVM自身的一些局限性和缺点,本文采用BA-RVM优化后的BA-FRVM算法作为汽车典型故障件数量的预测。本文建立一种基于蝙蝠算法快速相关向量机的汽车典型故障件数量预测模型,在Matlab R2014a软件上实现全部仿真实验。首先,研究干扰相关向量机预测准确率的影响因子-核参数,通过蝙蝠算法选择适合当前数据的核参数,达到核参数的自适应;在汽车典型故障件数据下验证BA-RVM算法,通过仿真实验选择合适的特征归一化方法;其次,通过BA-RVM算法在不同的数据下实验,选出合适的核函数,接着对BA-RVM算法训练效率进行优化,得到BA-FRVM算法,将UCI网站上三种不同类型的数据作为实验数据,以此来验证BA-FRVM算法的可信性和可靠性,然后再在汽车典型故障件数据下进行实验,并与BA-RVM算法对比训练时间、相关向量数以及错误率;再次,对BA-FRVM算法进行多方面的研究,研究不同迭代次数与错误率的关系,不同蝙蝠数量对预测准确率的影响,训练模型错误率与核参数宽度的关系,不同训练样本量对预测准确率的影响,与相似算法BA-SVR对比训练时间、相关(支持)向量数以及错误率,与广泛应用的BA-BP算法对比训练时间和错误率。最后,将BA-FRVM算法用Java编程语言实现,用jblas矩阵库实现矩阵运算,并将BA-FRVM算法试用于实际的故障件数量预测系统中。最终,通过实验结果可知:相比BA-SVR和BA-BP算法,BA-FRVM算法训练速度更快,预测准确率更高,能够更好的适用于汽车典型故障件数量预测。
[Abstract]:In recent years, with the rapid development of automobile enterprises, the volume of data in this industry has increased rapidly. Automobile enterprises urgently need to find the rules in the data, predict the number of spare parts according to historical data, diagnose automobile faults and so on. In this context, many artificial intelligence algorithms are beginning to be applied to the automotive industry, among which, the most widely used are SVM and BP neural networks. In view of the limitations and shortcomings of BP neural network and SVM, this paper uses the optimized BA-FRVM algorithm of BA-RVM as the prediction of the number of vehicle typical fault parts. In this paper, a fast correlation vector machine based on bat algorithm is established to predict the number of typical fault parts. The simulation experiments are implemented on Matlab R2014a software. Firstly, the kernel parameters, which affect the prediction accuracy of interference correlation vector machines, are studied, the kernel parameters suitable for current data are selected by bat algorithm, and the adaptive kernel parameters are achieved. The BA-RVM algorithm is verified under the vehicle typical fault data. The proper feature normalization method is selected through the simulation experiment. Secondly, the appropriate kernel function is selected through the experiment of BA-RVM algorithm under different data, then the training efficiency of BA-RVM algorithm is optimized, and the BA-FRVM algorithm is obtained. Three different types of data on the UCI website are taken as experimental data to verify the credibility and reliability of the BA-FRVM algorithm, and then the experiment is carried out under the typical fault data of the vehicle, and the training time is compared with that of the BA-RVM algorithm. Thirdly, the relationship between different iterations and error rates, the effects of different bat numbers on prediction accuracy, the relationship between the training model error rate and the width of kernel parameters, and the relationship between the training model error rate and the kernel parameter width are studied. The effect of different training samples on prediction accuracy is compared with similar algorithm BA-SVR, correlation (support) vector number and error rate, training time and error rate compared with widely used BA-BP algorithm. Finally, BA-FRVM algorithm is realized by Java programming language, matrix operation is realized by jblas matrix library, and BA-FRVM algorithm is used in actual fault prediction system. Finally, the experimental results show that: compared with BA-SVR and BA-BP algorithm BA-FRVM algorithm training speed is faster, prediction accuracy is higher, can be better applied to the number of vehicle typical fault prediction.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U472;TP18

【相似文献】