融合型深度学习亚健康识别算法的研究

发布时间：2019-04-08 08:44

【摘要】：在现代的工业生产过程中,逐步向生产设备大型化、智能化、复杂化、自动化发展,如果其中某零部件发生故障,会对整个生产过程产生一定的影响。但是若能对当前系统状态进行准确识别,并对相关故障环节的设备进行及时的更换,就能有效的避免故障的发生。对于不可预测的外力作用所导致的突发性系统故障来说一般是没有任何征兆以及不可控的。但是工业过程控制系统发生的故障大多是由于设备磨损或者元件老化导致的延时性故障,表征为设备工作时表现不明显,因此延时性故障研究和设备可靠性研究成为工业设备故障诊断领域的重点,并引起很多的专家学者的关注。本文主要对深度自动编码器及其改进进行研究,并以滚动轴承的“亚健康”状态识别作为应用场景。在阅读大量的深度学习和故障诊断方法后,发现一个设计良好的学习率策略可以显著提高深度学习模型的收敛速度,于是本文提出一种自适应性学习率来提高深度网络模型的收敛速度。同时本文结合稀疏自动编码器以及边缘降噪编码器的优点,对原深度模型的代价函数进行改进,增加网络模型的泛化能力以及鲁棒性。本文采用层叠编码机作为深度学习的网络结构,通过这种结构可以对机械振动信号中的噪声进行过滤,有助于有利特征的提取。实验结果表明在基本保证准确率的情况下加快了模型的收敛速度。本文使用相关向量机取代传统深度学习中的SoftMax层对提取的深度特征进行识别分类,同时,相关向量机的核函数选取以及核参数的选取对于最终的分类结果尤为重要,本文采用混合模式的核函数,并且根据Fisher准则以及最大熵准则提出一种最优映射的相关向量机核参数选取方法。实验结果表明这种方式下选取的核参数有助于提高模型的识别精度。为了进一步提高识别的精确性,本文将由相关向量机分类器所得到的输出归一化处理后作为D-S证据理论的第一证据空间,将由SoftMax分类器得到的结果归一化后作为第二个证据空间,然后根据D-S证据理论的融合规则将两个证据空间进行融合得到最终的识别结果。实验结果表明该方法在提高滚动轴承“亚健康”状态识别精度上是有效的。
[Abstract]:In the process of modern industrial production, the production equipment is gradually developed to large-scale, intelligent, complex and automatic. If one of the parts is broken down, it will have a certain impact on the whole production process. However, if the current system status can be accurately identified, and the equipment of the relevant fault links can be replaced in a timely manner, the failure can be effectively avoided. For unexpected system failures caused by unpredictable external forces, there are generally no signs and uncontrollable. However, most of the industrial process control system failures are caused by equipment wear or aging, which is characterized as the performance of the equipment is not obvious. Therefore, the research on delay fault and equipment reliability has become the focal point in the field of industrial equipment fault diagnosis, and has attracted the attention of many experts and scholars. This paper mainly studies the depth automatic encoder and its improvement, and takes the "sub-health" state recognition of rolling bearings as the application scenario. After reading a large number of in-depth learning and fault diagnosis methods, it is found that a well-designed learning rate strategy can significantly improve the convergence rate of the deep learning model. Therefore, this paper proposes a self-adaptive learning rate to improve the convergence rate of the deep network model. At the same time, combining the advantages of sparse automatic encoder and edge de-noising encoder, the cost function of the original depth model is improved to increase the generalization ability and robustness of the network model. In this paper, the cascade coding machine is used as the network structure of in-depth learning. Through this structure, the noise in mechanical vibration signal can be filtered, which is helpful to the extraction of favorable features. The experimental results show that the convergence speed of the model is speeded up under the condition that the accuracy of the model is basically guaranteed. In this paper, we use the correlation vector machine to replace the SoftMax layer in the traditional depth learning to identify and classify the extracted depth features. At the same time, the kernel function selection and the kernel parameter selection of the correlation vector machine are very important for the final classification results. In this paper, the kernel function of mixed mode is used, and according to the Fisher criterion and the maximum entropy criterion, a method for selecting the kernel parameters of the correlation vector machine is proposed. The experimental results show that the kernel parameters selected in this way are helpful to improve the recognition accuracy of the model. In order to further improve the accuracy of recognition, the output normalization obtained from the correlation vector machine classifier is regarded as the first evidence space of DES evidence theory. The results obtained by SoftMax classifier are normalized as the second evidence space, and then the two evidence spaces are fused according to the fusion rule of DES evidence theory to get the final recognition result. The experimental results show that this method is effective in improving the accuracy of "sub-health" state identification of rolling bearings.
【学位授予单位】：辽宁大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】