基于卷积神经网络的故障指示器状态识别研究

发布时间：2018-07-22 13:35

【摘要】：随着机器视觉技术的发展,越来越多的产品质量检测采用数字图像处理技术进行分析与识别,能够极大地提高生产的自动化程度。为实现工厂生产的故障指示器产品质量智能检测,本文基于卷积神经网络的故障指示器状态识别进行研究,能够有效地解决产品生产过程中的产品质量检测问题,实现工业生产自动化、智能化、绿色化和高效化。本文针对故障指示器状态智能识别任务,分别从构建识别系统、改进卷积神经网络(Convolutional Neural Networks,CNN)模型、实验验证等方面进行了研究。通过分析故障指示器产品检测场景,构建了图像采集系统并采集了故障指示器的原始视频图像。设计了识别算法流程,并实验验证了将CNN直接用到故障指示器原始图片上进行状态识别的可行性,同时分析实验结果,找出了传统CNN在此任务中存在的问题,启发了后续对原始图片的处理和CNN的改进工作。而后,针对现实场景中模糊、光照不均匀、色偏的故障指示器图片,本文对图片进行滤波、增强和高光消除等预处理,减少了各种因素对识别的影响,进一步采用基于阈值、边缘检测和聚类的方式对图像进行分割实验,接着对图片进行平移、旋转、缩放等数据扩充方式增大数据量,提升小样本对训练卷积神经网络的识别性能。针对传统CNN模型鲁棒性问题,本文改进网络结构,对网络加入尺度估计,提出了多尺度卷积神经网络模型,通过实验验证了该方法的鲁棒性;针对传统的CNN的收敛时间长,识别率低的问题,分析已收敛的CNN各核函数之间存在很大的相关性,提出了小波变换初始化第一层核函数的方法,实验表明该方法既缩短了网络收敛时间,又提高了识别率;将上述两种改进方法的结合起来发挥了各自优势,与传统的卷积神经网络相比,识别率提高7.28%,最终达到96.32%。最后,作为CNN算法的前沿技术,集检测与识别于一体的Faster R-CNN模型也被应用到了本文的故障指示器状态识别任务中,实验结果表明,Faster R-CNN技术能够有效解决故障指示器状态识别问题。
[Abstract]:With the development of machine vision technology, more and more product quality detection uses digital image processing technology to analyze and identify, which can greatly improve the degree of automation of production. In order to realize the intelligent detection of the product quality of the fault indicator in the factory, this paper studies the status recognition of the fault indicator based on the convolution neural network, which can effectively solve the problem of the product quality detection in the process of product production. Realize industrial production automation, intelligence, green and high efficiency. In this paper, the task of intelligent recognition of fault indicator states is studied from the aspects of constructing recognition system, improving the Convolutional Neural Network (CNN) model, and experimental verification. By analyzing the detection scene of the fault indicator, the image acquisition system is constructed and the original video image of the fault indicator is collected. The recognition algorithm flow is designed, and the feasibility of using CNN directly on the original image of the fault indicator is verified. At the same time, the experimental results are analyzed, and the problems existing in the task of traditional CNN are found out. Inspired the subsequent processing of the original images and CNN improvements. Then, aiming at the fault indicator picture of fuzzy, uneven illumination and color deviation in the real scene, this paper preprocesses the image, such as filtering, enhancement and highlight elimination, which reduces the influence of various factors on the recognition, and further adopts the threshold based method. The methods of edge detection and clustering are used to segment the image, and then the image is expanded by translation, rotation, zoom and so on, which increases the amount of data and improves the recognition performance of the small sample to the training convolutional neural network. Aiming at the robustness problem of traditional CNN model, this paper improves the network structure, proposes a multi-scale convolution neural network model for network scale estimation, and verifies the robustness of this method through experiments, aiming at the long convergence time of traditional CNN. The problem of low recognition rate is analyzed. The correlation between the convergent CNN kernel functions is analyzed. A wavelet transform method is proposed to initialize the first layer kernel function. The experiments show that this method not only shortens the convergence time of the network, but also improves the recognition rate. Compared with the traditional convolution neural network, the recognition rate is increased by 7.28%, and finally reaches 96.3232%. Finally, as the cutting-edge technology of CNN algorithm, the Faster R-CNN model, which integrates detection and recognition, is also applied to the task of fault indicator status recognition in this paper. The experimental results show that the Faster R-CNN technique can effectively solve the problem of fault indicator state identification.
【学位授予单位】：华北电力大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】