【作者】 柳长源；

【导师】 毕晓君；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2013， 博士

【摘要】 相关向量机(Relevance Vector Machine, RVM)是贝叶斯统计学习理论(StatisticalLearing Theory,SLT)发展的产物，是一种有监督机器学习的模式识别新方法。该方法由支持向量机(Support Vector Machine, SVM)理论演变而来，相比后者，具有解更稀疏、核函数选择更自由、泛化能力更强、鲁棒性更好等优点，在小样本的统计学习问题中的表现尤其突出，近几年已经在应用领域得到了快速发展，在模式分类、故障诊断、智能预测、语音及图像信息处理等方面均有很好的表现。但是，相关向量机在解决多类模式识别问题时，由于计算过程比较复杂，仍存在分类精度与训练识别时间无法兼顾的矛盾。本课题针对RVM算法存在的不足，对算法的结构和关键步骤，如核函数的选择、分类器的设计以及控制参数的调整等进行了深入研究和大量实验仿真工作，并对多分类问题中应用最广泛且分类精度最高的“一对一”分类器进行了改进。改进后的多分类方法，在基本保持原有的分类精度的基础上，大幅度提升了算法在类别数较多的模式识别问题上的分类时间，使RVM算法应用的实时性有了明显的提高。此外，从应用角度出发，将改进后的RVM算法应用于人脸识别及汽车发动机失火故障检测等问题的模式识别中，均取得了良好的效果。首先，详细论述相关向量机的研究现况和基本理论,并且提出相关向量机中仍需解决的关键问题。为了提高相关向量机学习算法在多模式识别中的分类速度，对相关向量机多分类方法进行了分析和研究，发现比较次数过多是该方法计算量大的主要原因。提出了一种在每轮比较中，排除最差类别的新方法。该方法使比较次数逐级减少，并且当类别数较多时，总计算量减少尤其明显。通过仿真实验说明了该方法的有效性,对数据分类的实验结果表明，新方法与传统分类器相比，在基本不影响分类正确率的前提下，机器训练与识别次数显著减少，算法运行速度明显提高。其次，为了解决人脸识别问题中对准确性、实时性、稳定性的要求，对传统的人脸识别方法进行了研究，提出一种基于改进相关向量机的人脸识别方法。文章利用小波变换对人脸图像进行预处理；根据PCA方法对处理后的人脸图像进行特征提取；利用相关向量机多分类模型进行人脸识别。与基于SVM的人脸识别方法进行比较，结果表明RVM具有高于SVM的鲁棒性，人脸识别的正确率更高、实时性好、可靠性更强。再次，当人脸图像含有较多噪声时，识别正确率会有很明显的下降。目前的人脸识别技术对此问题尚无较好的解决办法。本文提出一种采用相关向量机的人脸识别方法，利用机器学习对小波分解和PCA变换后的人脸数据库样本进行训练，得到的相关向量构成“超平面”作为差异样本的分类面，并利用改进的“一对一”方法实现多类别模式识别。对加噪声的识别对象进行了大量的仿真实验结果表明，与传统方法相比，新方法对图像噪声不敏感，具有更高的识别率和很强的鲁棒性。另外，对拍照光线、角度变化、物体遮挡、分辨率不足等条件下的人脸图像识别，也采用新方法进行了实验分析和讨论。最后，应用RVM算法研究汽车发动机故障诊断问题。研究发现算法中的惩罚因子和径向基核函数参数对分类准确率有着很大的影响，本文结合粒子群（PSO）算法对参数进行优化，并把该优化算法用于汽车发动机故障诊断中。针对样本的特征参数会随发动机转速变化的问题，提出了一种超参数自适应拟合的增量学习方法。在发动机失火故障诊断中，建立汽车尾气中各气体的体积分数与失火故障原因的映射关系，并对不同档位归一化处理的数据进行增量机器训练，对得到的超参数进行非线性拟合，并将训练好的RVM模型应用于故障分类诊断。仿真实验表明新方法不仅诊断结果准确可靠，而且解决了传统方法实现变速动态检测的困难。在论文的结尾，对课题的研究工作进行了总结，并对进一步研究工作进行了展望。更多还原

【Abstract】 Relevance vector machine (RVM) technique is a novel pattern recognition method ofsupervision machine learning which is based on Bayesian learning theory. It was developedon the basis of Support vector machine(SVM) learning theory, compared with the SVM, it hasthe benefits of sparser model、the facility to utilize arbitrary kernel functions、more accurate、strong robusticity、 intensity generalization ability， and so on. RVM algorithm has beendeveloped rapidly in the fields of application, and proved better in pattern classification、faultdiagnosis、intelligent forecast、information processing of voice and image, etc. Whereas，tosolve the questions of multi-mode pattern recognition，RVM algorithm still suffer from theproblems of looking after both sides of accuracy and real-time because of the complexity ofcomputational process.According to the insufficiency of RVM, the structure and key steps of the algorithm,including kernel functions selection, classifier construct, and control parameter setting, aredeeply investigated in this paper and improved on “one against one” classifier which has beenthe highest accuracy and broad application are proposed to improve. The improvedclassification method proposed is utilized successfully to reduce the time of classification andincrease in real-time without cutting down the accuracy. In addition, improved algorithm inthis paper is applied in face recognition&automobile engine fault diagnosis, and behavedwell.Firstly, the research situation and the fundamental theory about RVM is detail discussedin this paper, and the solved key problem of RVM is presented. In order to improveclassification speed of multiclass pattern recognition based on relevance vector machinelearning algorithm， investigated the method of relevance vector machine algorithm inmulti-mode classification, and found that the comparison too many times was the main reasonfor large amount of calculation. Proposed a new waythat eliminated the most dissimilar classin each round of comparison. Comparison times were reduced step by step per cycle, and theclassification was more, the decrease in the total calculation amount was more obvious. Thevalidity of this method has been proved by some simulated examples, and the experimentalresults of data classification show that compared with traditional classifier, the training timesand the recognition times of the novel method are greatly reduced under the premise of hardly influence classification accuracy, the algorithm running speed is improved obviously.Secondly, to solve the problems existed in face recognition, such as lack of accuracy、real-time and stability，a new face recognition approach based on improved relevance vectormachine is presented in this paper. Firstly, the wavelet transform is applied to preprocess faceimage to reduce the impact of expression change. Then, in order to extract key features of theprocessed face image, use the principal component analysis (PCA) method. Finally, the RVMclassification model is adopted for identifying. In comparison with the support vectormachine(SVM) method, the RVM approach performs well and can obtain more satisfactoryresults in terms of recognition rates、real-time and reliability.Thirdly, the face images recognition accuracy will be obvious decline when the objectscontain more noise. At present, face recognition technology to solve this problem is no betterway. In this paper, a new method of face recognition based on relevance vector machine waspresented. After the wavelet decomposition and PCA transform, relevance vectors fromsample training constituted a "hyperplane" as the differences in the classification of thesamples by machine learning algorithm and used the improved “one against one” method toachieve multi-class pattern recognition. Compared with the former method, a large number ofsimulation results show that the new method used in noisy objects being recognized is notsensitive to image noise, with a more accurate and strong robusticity. In addition, photo light、angle change、occlusion、low resolution ratio and so on, also be discussed and analysed inexperiment with new method in this paper.Finally, application of RVM in the automobile engine fault diagnosis is investigated.From the study we know that the parameter of penalty factor and kernel paraeter play a veryimportant roal on the diagnosis model, so the Particle Swarm Optimization(PSO) is used tooptimize the parameters, this algorithms practical applied to automotive engine fault diagnosis.Considering the problem of the variation of characteristic parameters are followed by enginerotational speed, puts forward a adaptive fitting of super-parameters on incremental learning.To the problem of engine misfire, mapping relations established between gas volume fractionand the cause of the misfire, used normalized data with different gears in machine training,ajusted super-parameters by curve fitting, and the trained RVM model applied in faultclassification and diagnosis.The simulation experiments shows the results of new method isnot only accurate and reliable but also resolve the problem of dynamic detection with variable speed in traditional methods.In conclusion, a simple summary is made and some research aspects are presented in thefuture.更多还原