基于深度学习的机房设备监控

发布时间：2018-03-06 07:33

本文选题：卷积神经网络　切入点：目标检测　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着电子信息技术不断发展,机房的设备在不断更新和增加。如果机房设备出现故障,就将会直接影响整个系统的正常运行。本文针对无人值守机房的情况,设计了机房设备监控平台,该平台通过实时分析机房现场的图片来实时监控机房设备的运行状况。本文任务的核心是目标检测,检测出机房现场图片里的指定目标。目标检测在人工智能领域已经得到了深入地研究,并且在工业上得到了广大的应用。随着目标检测地被广泛应用,目标检测也存在很多的挑战和问题。传统的目标检测算法采用形态学、统计学的算法,由于需要根据环境设定算法参数,导致算法适应性不好。如果使用选择搜索来提取图片感兴趣区域,这需要对图片进行全局搜索,提取速度非常慢。因此本文采用卷积神经网络来进行目标检测的任务。为了解决上面这些问题,本文设计合适的卷积神经网络来提取感兴趣区域,然后设计卷积神经网络来对感兴趣区域内的目标物体进行识别。本文设计类别误差函数来更新整个网络权重,提高分类的准确率;此外设计位置坐标误差函数,在训练的时候不断更新网络权重,进行线性位置回归,从而提高位置预测的准确率。通过实验可以验证这种方法准确率和检测速度都能满足本文的要求。此外,本文初步讨论了卷积神经网络的改进方式,发现卷积神经网络的前几层卷积层具有相位对称特性,因此可以将前几层卷积层的输出特征图减少一半,另一半由这个镜像得到。通过上面的方式,可以减少网络的权重数量,这样可以提高训练、测试速度。
[Abstract]:With the development of electronic information technology, the equipment of the computer room is constantly updated and increased. If the equipment of the computer room fails, it will directly affect the normal operation of the whole system. The monitoring platform of computer room equipment is designed. The platform can real-time monitor the running condition of the equipment by analyzing the pictures of the computer room in real time. The core of the task of this paper is the target detection. Detection of designated targets in field pictures of computer rooms. Target detection has been deeply studied in artificial intelligence field and has been widely used in industry. With the wide application of target detection, There are also many challenges and problems in target detection. The traditional target detection algorithms use morphological and statistical algorithms, because of the need to set the algorithm parameters according to the environment. This result in poor adaptability of the algorithm. If you use selective search to extract the region of interest of the picture, this requires a global search of the image. The speed of extraction is very slow. Therefore, the task of target detection is based on convolution neural network. In order to solve these problems, this paper designs an appropriate convolutional neural network to extract the region of interest. Then the convolutional neural network is designed to identify the object in the region of interest. In this paper, the class error function is designed to update the weight of the whole network to improve the classification accuracy; in addition, the position coordinate error function is designed. During training, network weights are updated and linear position regression is carried out to improve the accuracy of location prediction. The experimental results show that the accuracy of this method and the speed of detection can meet the requirements of this paper. In this paper, the improved method of convolution neural network is discussed. It is found that the first several layers of the convolution neural network have phase symmetry, so the output characteristic map of the first several layers can be reduced by half. The other half is obtained by this mirror image. In this way, the weight of the network can be reduced, which can improve the training and testing speed.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP308;TP277

【参考文献】