WO2023082418A1 - 基于人工智能技术的电力综合管廊沉降裂缝识别方法 - Google Patents

基于人工智能技术的电力综合管廊沉降裂缝识别方法 Download PDF

Info

Publication number
WO2023082418A1
WO2023082418A1 PCT/CN2021/138598 CN2021138598W WO2023082418A1 WO 2023082418 A1 WO2023082418 A1 WO 2023082418A1 CN 2021138598 W CN2021138598 W CN 2021138598W WO 2023082418 A1 WO2023082418 A1 WO 2023082418A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
artificial intelligence
output
intelligence technology
power utility
Prior art date
Application number
PCT/CN2021/138598
Other languages
English (en)
French (fr)
Inventor
顾建军
吴留闯
陈杰
蔡人立
万浩
Original Assignee
国网江苏省电力有限公司南通供电分公司
南通送变电工程有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国网江苏省电力有限公司南通供电分公司, 南通送变电工程有限公司 filed Critical 国网江苏省电力有限公司南通供电分公司
Publication of WO2023082418A1 publication Critical patent/WO2023082418A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/40Image enhancement or restoration by the use of histogram techniques
    • G06T5/90
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/13Edge detection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20081Training; Learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20084Artificial neural networks [ANN]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30181Earth observation
    • G06T2207/30184Infrastructure
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04SSYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Definitions

  • the invention relates to the technical field of crack identification of electric power comprehensive pipe gallery, in particular to a method for identifying settlement cracks of electric power comprehensive pipe gallery based on artificial intelligence technology.
  • the purpose of the present invention is to provide a method for identifying settlement cracks of electric utility tunnels based on artificial intelligence technology, using the backpropagation algorithm to carry out BP neural network training, and using the trained BP neural network
  • the network detects and recognizes the image edge features of the acquired pipe gallery wall image, detects the cracks in the pipe gallery wall, and realizes the real-time monitoring of the internal structure settlement of the power comprehensive pipe gallery.
  • the present invention adopts the following technical solutions.
  • a method for identifying settlement cracks in an integrated power utility gallery based on artificial intelligence technology comprising the steps of:
  • the image gray value analysis method includes: converting the acquired color image into a binary gray image, and then performing row-by-row and column-by-column projection on the pixels of the gray image to calculate the gray value The sum of all pixels in each row and each column in the image, through the step change of the gray projection value in the horizontal direction and the vertical direction, obtains the uneven distribution of illumination.
  • step (2) an image histogram equalization method is used to correct the uneven distribution of illumination on the grayscale image
  • n is the sum of image pixels
  • n i is the sum of pixels of the i-th gray level
  • r i is the i-th gray level
  • i 0,1,2,...,L-1.
  • the image edge feature is extracted from the wall image of the pipe corridor by extracting the Canny edge feature of the image.
  • the historically collected crack pictures and normal pictures including the internal structure of the pipe gallery are formed into training samples after image preprocessing and edge feature extraction; the sample data is used through the back propagation algorithm Train the BP neural network.
  • the backpropagation algorithm starts from the output unit, and propagates the weight correction caused by the total error to the hidden layer unit; using the partial derivative ⁇ o (k) of each neuron in the output layer and the output h of each neuron in the hidden layer ho (k) calculates the partial derivative ⁇ w ho (k) of the error function to each neuron in the hidden layer, using the output ⁇ h (k) of each neuron in the hidden layer and the input parameters of each neuron in the input layer Correct the connection weight.
  • D is the total number of training samples
  • outputs is the set of output units
  • w i is the weighting coefficient
  • t kd and O kd are the output values related to the training sample d and the kth output unit, and are determined by the corresponding input value, hidden function as well as the activation function.
  • the gradient descent algorithm is used in the backpropagation algorithm to minimize the square error between the output value of the output unit of the BP neural network and the target value.
  • the extracted edge image is converted into an initial neuron input value, wherein each neuron unit has a certain number of real-valued inputs, and produces a single real-valued output, and the neuron's
  • the final output is as follows:
  • f is the activation function in the neural network
  • the input units in the neuron are x i
  • the beneficial effect of the present invention is that, compared with the prior art, the present invention proposes an improved backpropagation algorithm suitable for detecting cracks on the walls of utility corridors in complex environments. Including crack pictures and normal pictures of the wall in the internal structure of the pipe gallery, for learning and training, and finally realizing the detection and identification of the settlement cracks of the pipe gallery.
  • the backpropagation algorithm uses the gradient descent method to reduce the square error between the network output value and the target value, and comprehensively weights multiple output units to recalculate the reverse error E, so that all The error-weighted sum of the network outputs.
  • Fig. 1 is the flow chart of the identification method for the settlement cracks of the electric utility utility gallery based on the artificial intelligence technology of the present invention
  • Fig. 2 is a schematic diagram of a three-layer BP neural network
  • Fig. 3 is the neuron structure schematic diagram in BP neural network
  • Figure 4 is a picture of cracks in the pipe gallery
  • Fig. 5 is a schematic diagram of feature extraction of the crack edge of the pipe gallery
  • Figure 6 is a schematic diagram of the detection effect of settlement cracks in the power utility tunnel.
  • the artificial intelligence technology-based method for identifying settlement cracks in the integrated power utility gallery of the present invention includes the following steps:
  • the intelligent inspection robot takes real-time pictures of the wall of the pipe gallery in a complex and dangerous environment, and obtains the image of the wall of the pipe gallery, as shown in Figure 4.
  • the acquired wall image of the pipe corridor is analyzed and corrected by using the gray value of the image.
  • Use the gray value of the image to analyze the light distribution of the wall image of the pipe corridor obtain the area with uneven illumination, including areas that are too dark or too bright, and then use the gradient histogram method to correct the uneven light of the image to make it too dark or too bright
  • the area corrected to its original color is used.
  • the environment around the wall of the pipe gallery is complex, and various electrical equipment are installed inside the pipe gallery.
  • the light of the supplementary light source is blocked by the electric equipment installed in the pipe gallery during the inspection of the robot, resulting in light blocking edges on the wall and affecting the real edge of the wall. extraction.
  • the acquired color image is first converted into a binary grayscale image, and then the image pixels are projected row by row and column by column respectively, and the sum of all pixels in each row and column in the image is calculated.
  • the image histogram method is used to equalize the histogram of the image to eliminate the interference edge caused by the occlusion of the light as much as possible.
  • the image histogram is a statistical relationship representing the frequency of occurrence of each gray level in a digital image.
  • the histogram can give a general description of the gray range of the image, the frequency and distribution of each gray level, the average brightness and contrast of the entire image, etc. Assuming that the grayscale range of the grayscale image is [0,L-1], define its histogram as:
  • n is the sum of image pixels
  • nk is the sum of pixels of the kth gray level
  • r k is the kth gray level
  • k 0,1,2,...,L-1.
  • the gray value of the image can be equalized, and the influence of the interference edge caused by light occlusion on the extraction of the real crack edge of the pipe gallery wall can be reduced.
  • the image edge features are extracted from the preprocessed pipe gallery wall image, and the edge features of the pipe gallery wall structure are analyzed, and the extracted edge features are used as the input of the BP neural network.
  • the image of the pipe gallery wall is extracted by extracting the Canny edge features of the image to extract the image edge features, and the extracted edge features are used as the input of the BP neural network input unit.
  • Canny edge extraction algorithm is as follows:
  • Gaussian smoothing is performed on the input image, and the Gaussian kernel function is used to smooth the noise point, which reduces the interference of various noises on image feature extraction, and reduces the edge error rate caused by image noise.
  • the Gaussian kernel function is defined as follows:
  • is the standard deviation
  • the size is set to 0.6 for the tunnel image
  • the Gaussian smoothing window size is set to 7 ⁇ 7 pixels.
  • the gradient magnitude and direction are calculated to estimate the edge strength and direction at each point in the image.
  • the Sobel operator to convolve the input image, assuming that the original image is I(x,y), and the Sobel gradient descriptors in the horizontal and vertical directions are:
  • * is the convolution symbol.
  • non-maximum suppression is performed on the gradient magnitude according to the gradient direction.
  • its gradient direction is approximated as one of the following values (0, 45, 90, 135, 180, 225, 270, 315), and the gradient strength of the pixel is compared with that of the pixel in the positive and negative direction of the gradient direction. If the gradient strength of the pixel is the largest If it is reserved, otherwise it is suppressed (deleted, that is, set to 0).
  • edges are processed and concatenated with double thresholding. There are still many noise points in the image after non-extreme large suppression.
  • Use double thresholds for processing that is, set an upper threshold and a lower threshold. If the pixel in the image is greater than the upper threshold, it must be a boundary, and if it is smaller than the lower threshold, it must not be a boundary, and if it is between the two, it is considered to be a boundary. Candidates. Weak boundaries connected by strong boundaries are considered as boundaries, and other weak boundaries are suppressed.
  • the edge is linked into a contour. When the end point of the contour is reached, the algorithm will find a point that meets the low threshold in the neighborhood of the breakpoint, and then collect new edges based on this point until the edge of the entire image is closed. .
  • a large amount of image data collected in the history of integrated power utility tunnels, including crack pictures and normal pictures in the internal structure of the utility tunnel, is composed of training samples D after image preprocessing and edge feature extraction;
  • the data trains the BP neural network.
  • BP neural network is composed of several neuron units, forming a huge multi-layer BP neural network system.
  • the backpropagation algorithm starts from the output unit and propagates the weight corrections caused by the total error backwards to the hidden layer units.
  • is the corresponding proportional coefficient.
  • is the corresponding proportionality coefficient.
  • D is the total number of training samples
  • outputs is the set of output units
  • w i is the weighting coefficient
  • t kd and O kd are the output values related to the training sample d and the kth output unit, and are determined by the corresponding input value, hidden function as well as the activation function.
  • the hidden function and activation function used in the present invention are respectively defined as follows:
  • the error function is set as:
  • y is the actual result
  • o is the predicted result
  • W i is the weighting coefficient of the corresponding input channel
  • xi is the neuron input value.
  • the process of the gradient descent algorithm is the process of minimizing the function J and solving the corresponding W i .
  • the rate of change as ⁇ , then:
  • the BP neural network will convert the extracted edge image into the initial value of the neuron input according to the edge feature extracted from the image, where each neuron unit has a certain number of real-valued inputs and generates a single real value.
  • Numerical output, the final output of the neuron is shown in the following formula:
  • f is the activation function in the neural network
  • the input units in the neuron are x i
  • the beneficial effect of the present invention is that, compared with the prior art, the present invention proposes an improved backpropagation algorithm suitable for detecting cracks on the walls of utility corridors in complex environments. Including crack pictures and normal pictures of the wall in the internal structure of the pipe gallery, for learning and training, and finally realizing the detection and identification of the settlement cracks of the pipe gallery.
  • the backpropagation algorithm uses the gradient descent method to reduce the square error between the network output value and the target value, and comprehensively weights multiple output units to recalculate the reverse error E, so that all The error-weighted sum of the network outputs.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Quality & Reliability (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

一种基于人工智能技术的电力综合管廊沉降裂缝识别方法,包括步骤:(1)获取管廊墙壁图像;(2)对管廊墙壁图像进行预处理,利用图像灰度值分析光照不均匀情况并进行直方图均衡化校正处理;(3)对预处理后的图像进行图像边缘特征提取;(4)采用反向传播算法进行BP神经网络训练;(5)将提取的图像边缘特征输入训练好的BP神经网络并输出检测识别结果。还提出一种改进的适应于复杂环境下管廊墙壁裂缝检测的反向传播算法,对电力综合管廊中采集到的大量图像数据,包含管廊内部结构中墙壁的裂缝图片和正常图片,进行学习和训练,最终实现管廊沉降裂缝的检测和识别。

Description

基于人工智能技术的电力综合管廊沉降裂缝识别方法 技术领域
本发明涉及电力综合管廊裂缝识别技术领域,具体涉及一种基于人工智能技术的电力综合管廊沉降裂缝识别方法。
背景技术
地下电力综合管廊或者地下电力综合管道在长期服役过程中,由于周围地面环境复杂,受不同地质环境、地面建筑物、河流、高速公路、铁路、管廊结构材料性能、以及地面荷载的长期效应、疲劳效应与突变效应等因素的综合作用,管廊或者管壁将不可避免地出现结构损伤积累,一定程度上会造成管廊沉降或者出现裂缝等情况,甚至在极端情况下导致管廊结构失效和管廊坍塌,造成重大安全事故。因此,对地下电力综合管廊沉降以及发生管廊沉降后出现的管壁裂缝进行智能监测,对地下电力综合管廊安全运营以及减少管廊沉降事故的发生具有重大意义。
根据经验,电力综合管廊发生沉降时,一般管廊内部墙体结构会出现裂缝。随着工业4.0以及智能制造产业的快速发展,人工智能技术越来越多的应用于工业安全生产、机器人智能巡检以及其他重要领域。其中,智能巡检机器人对于地下电力综合管廊安全监测,是人工智能技术在电力行业中的一项重要应用,但现有的裂缝识别方法难以更加准确的对裂缝进行观察,监测效果不能够得到更好的保证,使用存在弊端,且无法适应于复杂环境下管廊墙壁裂缝检测。
发明内容
为解决现有技术中存在的不足,本发明的目的在于,提供一种基于人工智能技术的电力综合管廊沉降裂缝识别方法,采用反向传播算法进行BP神经网络训练,利用训练好的BP神经网络对获取的管廊墙壁图像的图像边缘特征进行检测识别,检测出管廊墙壁的裂缝,实现电力综合管廊内部结构沉降的实时监测。
本发明采用如下的技术方案。
一种基于人工智能技术的电力综合管廊沉降裂缝识别方法,所述方法包括步骤:
(1)获取管廊墙壁图像;
(2)对管廊墙壁图像进行预处理,利用图像灰度值分析光照不均匀情况并进行图像直方图均衡化校正处理;
(3)对预处理后的图像进行图像边缘特征提取;
(4)采用反向传播算法进行BP神经网络训练;
(5)将提取的图像边缘特征输入训练好的BP神经网络并输出检测识别结果。
进一步地,所述步骤(2)中,图像灰度值分析法包括:将获取的彩色图像转化为二值灰度图像,再对灰度图像像素分别进行逐行和逐列投影,计算灰度图像中每一行和每一列中所有像素的和,通过水平方向和竖直方向的灰度投影值的阶跃变化得到光照不均匀分布。
进一步地,所述步骤(2)中,采用图像直方图均衡化方法,对灰度图像进行光照不均匀分布的校正处理;
图像直方图均衡化的数学表达式为:
Figure PCTCN2021138598-appb-000001
其中,n是图像像素的总和,n i是第i个灰度级的像素总和,r i是第i个灰度级,i=0,1,2,…,L-1。
进一步地,所述步骤(3)中,管廊墙壁图像通过提取图像的Canny边缘特征来进行图像边缘特征的提取。
进一步地,Canny边缘特征提取步骤如下:
(3.1)利用高斯核函数对输入图像进行高斯平滑;
(3.2)利用Sobel算子计算图像的梯度幅值和方向;
(3.3)根据梯度方向,对梯度幅值进行非极大值抑制;
(3.4)对非极值大抑制后的图像进行双阈值处理,在高阈值图像中把边缘链接成轮廓,当到达轮廓的端点时,在断点的邻域点中寻找满足低阈值的点,再根据此点收集新的边缘,直到整个图像边缘闭合。
进一步地,所述步骤(4)中,将历史采集的包含管廊内部结构中的裂缝图片和正常图片,经图像预处理与边缘特征提取后,组成训练样本;通过反向传播算法采用样本数据对BP神经网络进行训练。
进一步地,反向传播算法从输出单元开始,将由总误差引起的权重修正向后传播到隐藏层单元;利用输出层各神经元的偏导数δ o(k)和隐藏层各神经元的输出h ho(k)计算误差函数对隐含层各神经元的偏导数Δw ho(k),利用隐藏层各神经元的输出δ h(k)和输入层各神经元的输入参数
Figure PCTCN2021138598-appb-000002
修正连接权值。
进一步地,利用反向传播算法学习多层BP神经网络的权值,并综合加权多个输出单元,重新计算反向误差E,从而将所有网络输出的误差加权相加,公式如下:
Figure PCTCN2021138598-appb-000003
其中,D为训练样本总数,outputs是输出单元的集合,w i是加权系数,t kd和O kd是与训练样例d和第k个输出单元的相关输出值,由对应的输入值、隐藏函数以及激励函数决定。
进一步地,反向传播算法中采用梯度下降算法,最小化BP神经网络输出单元的输出值与目标值之间的平方误差。
进一步地,所述步骤(5)中,将提取后的边缘图像转化为神经元输入初始值,其中每一个神经元单元有一定数量的实值输入,并产生单一的实数值输出,神经元的最终输出结果如下式所示:
Figure PCTCN2021138598-appb-000004
其中,f为该神经网络中的激发函数,神经元中的输入单元分别为x i,与其对应的加权系数为w i,i=1,2,3,…,n-1。
本发明的有益效果在于,与现有技术相比,本发明提出一种改进的适应于复杂环境下管廊墙壁裂缝检测的反向传播算法,对电力综合管廊中采集到的大量图像数据,包含管廊内部结构中墙壁的裂缝图片和正常图片,进行学习和训练,最终实现管廊沉降裂缝的检测和识别。
考虑各单元之间的关联性,反向传播算法采用梯度下降法,减小网络输出值 与目标值之间的平方误差,并综合加权多个输出单元,重新计算反向误差E,从而将所有网络输出的误差加权相加。
利用图像灰度值分析管廊墙壁图像光照不均匀处,将光照不均通过图像光线校正减弱,可以将过暗或者过亮的区域校正到原来本色。
附图说明
图1为本发明的基于人工智能技术的电力综合管廊沉降裂缝识别方法流程图;
图2为三层BP神经网络示意图;
图3为BP神经网络中的神经元结构示意图;
图4为管廊裂缝图片;
图5为管廊裂缝边缘特征提取示意图;
图6为电力综合管廊沉降裂缝检测效果示意图。
具体实施方式
下面结合附图对本申请作进一步描述。以下实施例仅用于更加清楚地说明本发明的技术方案,而不能以此来限制本申请的保护范围。
如图1所示,本发明所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,包括以下步骤:
(1)获取管廊墙壁图像,作为图像输入;
通过智能巡检机器人对复杂危险环境下管廊墙壁进行实时拍摄,获取管廊墙壁图像,如图4所示。
(2)对管廊墙壁图像进行预处理;
将获取到的管廊墙壁图像利用图像灰度值分析并进行校正处理。利用图像灰度值分析管廊墙壁图像光照分布情况,获取光照不均匀区域,包括过暗或者过亮的区域,然后通过梯度直方图方法将图像光线不均匀情况进行校正,将过暗或者过亮的区域校正到原来本色。
管廊墙壁周围环境复杂,各种电力设备安装在管廊内部,造成机器人在巡检时因为补充光源的光线被管廊内安装的电力设备遮挡,导致墙壁上出现光线遮挡 边缘,影响墙壁真实边缘的提取。
利用图像灰度值分析方法,通过分析感兴趣区域边缘的灰度投影值,判别感兴趣区域是否存在灰度遮挡的情况。
先将获取的彩色图像转化为二值灰度图像,然后,再对图像像素分别进行逐行和逐列投影,计算图像中每一行和每一列中所有像素的和。对于m×n大小的图像I(x,y)的感兴趣区域,计算第i行水平方向的像素灰度投影值:
Figure PCTCN2021138598-appb-000005
第j列竖直方向的像素灰度投影值:
Figure PCTCN2021138598-appb-000006
通过分析水平方向和竖直方向的灰度投影值,如果在二者任意方向上出现像素灰度投影值出现明显的阶跃变化,就认为此处可能存在由于投影光线遮挡造成的图像灰度值发生阶跃变化。
基于上述分析,再利用图像直方图方法,对图像进行直方图均衡化,尽量消除上述由于光线被遮挡造成的干扰边缘。
图像直方图是表示数字图像中每一灰度出现频率的统计关系。直方图能给出图像灰度范围、每个灰度的频度和灰度的分布、整幅图像的平均明暗和对比度等概貌性描述。假设灰度图像的灰度级范围为[0,L-1],定义其直方图为:
Figure PCTCN2021138598-appb-000007
其中,n是图像像素的总和,n k是第k个灰度级的像素总和,r k是第k个灰度级,k=0,1,2,…,L-1。
图像进行直方图均衡化的数学表达式为:
Figure PCTCN2021138598-appb-000008
利用上式对图像进行直方图均衡化,即可以实现图像灰度值的均衡化,减小由于光线遮挡造成的干扰边缘对管廊墙壁真实裂缝边缘提取的影响。
(3)对预处理后的图像进行图像边缘特征提取;
如图5所示,将预处理后的管廊墙壁图像进行图像边缘特征提取,通过分析 管廊墙壁结构的边缘特征,以提取的边缘特征作为BP神经网络的输入。
管廊墙壁图像通过提取图像的Canny边缘特征来进行图像边缘特征的提取,以提取的边缘特征作为BP神经网络输入单元的输入。
Canny边缘提取算法如下:
首先,对输入图像进行高斯平滑,利用高斯核函数实现噪声点的平滑,减小各种噪声对图像特征提取的干扰,降低图像噪声产生的边缘错误率。高斯核函数定义如下:
Figure PCTCN2021138598-appb-000009
其中,σ为标准偏差,大小针对管廊图像设为0.6,并且高斯平滑窗口大小设为7×7像素。
然后,计算梯度幅度和方向,来估计图像每一点处的边缘强度与方向。利用Sobel算子对输入图像进行卷积,假设原始图像为I(x,y),水平方向和竖直方向的Sobel梯度描述子分别为:
Figure PCTCN2021138598-appb-000010
Figure PCTCN2021138598-appb-000011
然后,利用对图像I(x,y)进行卷积操作:
Figure PCTCN2021138598-appb-000012
其中,*为卷积符号。
则图像I(x,y)的Sobel梯度大小为:
G=|G x|+|G y|
其次,根据梯度方向,对梯度幅值进行非极大值抑制。对于每个像素点,将其梯度方向近似为以下值中的一个(0,45,90,135,180,225,270,315),比较该像素点和其梯度方向正负方向的像素点的梯度强度,如果该像素点梯度强度最大则保留,否则抑制(删除,即置为0)。
最后,用双阈值处理和连接边缘。经过非极值大抑制后图像中仍然有很多噪声点。利用双阈值进行处理,即设定一个阈值上界和阈值下界,图像中的像素点 如果大于阈值上界则认为必然是边界,小于阈值下界则认为必然不是边界,两者之间的则认为是候选项。通过和强边界相连的弱边界被认为是边界,则其他的弱边界则被抑制。在高阈值图像中把边缘链接成轮廓,当到达轮廓的端点时,该算法会在断点的邻域点中寻找满足低阈值的点,再根据此点收集新的边缘,直到整个图像边缘闭合。
(4)采用反向传播算法进行BP神经网络训练;
将历史采集的大量的电力综合管廊中的图像数据,包含管廊内部结构中的裂缝图片和正常图片,经图像预处理与边缘特征提取后,组成训练样本D;通过反向传播算法采用样本数据对BP神经网络进行训练。
BP神经网络由若干个神经元单元组成,形成一个庞大的多层BP神经网络系统。神经元单元包括输入单元、输出单元,还包含一个或多个隐藏层单元。以三层BP神经网络为例,其结构如图2所示,x i为输入值,y i为输出值,i=1,2,3。
反向传播算法从输出单元开始,将由总误差引起的权重修正向后传播到隐藏层单元。
首先,利用输出层各神经元的偏导数δ o(k)和隐藏层各神经元的输出h ho(k)计算误差函数对隐含层各神经元的偏导数Δw ho(k):
Δw ho(k)=-μδ o(k)h ho(k)
Figure PCTCN2021138598-appb-000013
其中,μ为对应比例系数。
利用隐藏层各神经元的输出δ h(k)和输入层各神经元的输入参数
Figure PCTCN2021138598-appb-000014
修正连接权值:
Figure PCTCN2021138598-appb-000015
其中,η为对应比例系数。
考虑BP神经网络各单元之间的关联性,利用反向传播算法学习多层BP神经网络的权值,并综合加权多个输出单元,重新计算反向误差E,从而将所有网络输出的误差加权相加,公式如下:
Figure PCTCN2021138598-appb-000016
其中,D为训练样本总数,outputs是输出单元的集合,w i是加权系数,t kd和 O kd是与训练样例d和第k个输出单元的相关输出值,由对应的输入值、隐藏函数以及激励函数决定。
本发明中使用的隐藏函数和激励函数分别定义为如下:
假设我们要对第k层隐藏层的参数W (k)和求偏导数b (k)。假设z (k)代表第k层神经元的输入,即定义此时的隐藏函数为:
z (k)=W (k)*n (k-1)+b (k)
Sigmoid激励函数:
Figure PCTCN2021138598-appb-000017
且反向传播算法中采用梯度下降法,减小BP神经网络输出单元的输出值与目标值之间的平方误差。
在梯度下降算法中,设误差函数为:
J=0.5(y-o) 2=0.5(y-f(∑W ix i)) 2
其中,y为实际结果,o为预测结果,o=f(∑W ix i)为神经网络中各种输入函数进行加权输出的结果,W i为对应输入通道加权系数,x i为神经元输入值。
梯度下降算法的过程就是最小化函数J,求解对应的W i的过程。定义变化速率为α,则有:
Figure PCTCN2021138598-appb-000018
Figure PCTCN2021138598-appb-000019
利用上式经过多次运算,可以求出W i,利用此梯度下降算法实现减小误差的目的。
(5)将提取的图像边缘特征输入训练好的BP神经网络并输出检测识别结果;
利用训练好的BP神经网络对获取的图像边缘特征进行检测识别,判断墙壁是否出现裂缝,最后输出检测结果和预警信息,如图6所示。
如图3所示,BP神经网络会根据图像提取的边缘特征,将提取后的边缘图像转化为神经元输入初始值,其中每一个神经元单元有一定数量的实值输入,并 产生单一的实数值输出,神经元的最终输出结果如下式所示:
Figure PCTCN2021138598-appb-000020
其中,f为该神经网络中的激发函数,神经元中的输入单元分别为x i,与其对应的加权系数为w i,i=1,2,3,…,n-1。
本发明的有益效果在于,与现有技术相比,本发明提出一种改进的适应于复杂环境下管廊墙壁裂缝检测的反向传播算法,对电力综合管廊中采集到的大量图像数据,包含管廊内部结构中墙壁的裂缝图片和正常图片,进行学习和训练,最终实现管廊沉降裂缝的检测和识别。
考虑各单元之间的关联性,反向传播算法采用梯度下降法,减小网络输出值与目标值之间的平方误差,并综合加权多个输出单元,重新计算反向误差E,从而将所有网络输出的误差加权相加。
利用图像灰度值分析管廊墙壁图像光照不均匀处,将光照不均通过图像光线校正减弱,可以将过暗或者过亮的区域校正到原来本色。
本发明申请人结合说明书附图对本发明的实施示例做了详细的说明与描述,但是本领域技术人员应该理解,以上实施示例仅为本发明的优选实施方案,详尽的说明只是为了帮助读者更好地理解本发明精神,而并非对本发明保护范围的限制,相反,任何基于本发明的发明精神所作的任何改进或修饰都应当落在本发明的保护范围之内。

Claims (10)

  1. 一种基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,所述方法包括步骤:
    (1)获取管廊墙壁图像;
    (2)对管廊墙壁图像进行预处理,利用图像灰度值分析光照不均匀情况并进行图像直方图均衡化校正处理;
    (3)对预处理后的图像进行图像边缘特征提取;
    (4)采用反向传播算法进行BP神经网络训练;
    (5)将提取的图像边缘特征输入训练好的BP神经网络并输出检测识别结果。
  2. 根据权利要求1所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    所述步骤(2)中,图像灰度值分析法包括:将获取的彩色图像转化为二值灰度图像,再对灰度图像像素分别进行逐行和逐列投影,计算灰度图像中每一行和每一列中所有像素的和,通过水平方向和竖直方向的灰度投影值的阶跃变化得到光照不均匀分布。
  3. 根据权利要求2所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    所述步骤(2)中,采用图像直方图均衡化方法,对灰度图像进行光照不均匀分布的校正处理;
    图像直方图均衡化的数学表达式为:
    Figure PCTCN2021138598-appb-100001
    其中,n是图像像素的总和,n i是第i个灰度级的像素总和,r i是第i个灰度级,i=0,1,2,…,L-1。
  4. 根据权利要求1所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    所述步骤(3)中,管廊墙壁图像通过提取图像的Canny边缘特征来进行图像边缘特征的提取。
  5. 根据权利要求4所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    Canny边缘特征提取步骤如下:
    (3.1)利用高斯核函数对输入图像进行高斯平滑;
    (3.2)利用Sobel算子计算图像的梯度幅值和方向;
    (3.3)根据梯度方向,对梯度幅值进行非极大值抑制;
    (3.4)对非极值大抑制后的图像进行双阈值处理,在高阈值图像中把边缘链接成轮廓,当到达轮廓的端点时,在断点的邻域点中寻找满足低阈值的点,再根据此点收集新的边缘,直到整个图像边缘闭合。
  6. 根据权利要求1所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    所述步骤(4)中,将历史采集的包含管廊内部结构中的裂缝图片和正常图片,经图像预处理与边缘特征提取后,组成训练样本;通过反向传播算法采用样本数据对BP神经网络进行训练。
  7. 根据权利要求6所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    反向传播算法从输出单元开始,将由总误差引起的权重修正向后传播到隐藏层单元;利用输出层各神经元的偏导数δ o(k)和隐藏层各神经元的输出h ho(k)计算误差函数对隐含层各神经元的偏导数Δw ho(k),利用隐藏层各神经元的输出δ h(k)和输入层各神经元的输入参数
    Figure PCTCN2021138598-appb-100002
    修正连接权值。
  8. 根据权利要求7所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    利用反向传播算法学习多层BP神经网络的权值,并综合加权多个输出单元,重新计算反向误差E,从而将所有网络输出的误差加权相加,公式如下:
    Figure PCTCN2021138598-appb-100003
    其中,D为训练样本总数,outputs是输出单元的集合,w i是加权系数,t kd和O kd是与训练样例d和第k个输出单元的相关输出值,由对应的输入值、隐藏函数以及激励函数决定。
  9. 根据权利要求7所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    反向传播算法中采用梯度下降算法,最小化BP神经网络输出单元的输出值与目标值之间的平方误差。
  10. 根据权利要求1所述的基于人工智能技术的电力综合管廊沉降裂缝识别方法,其特征在于,
    所述步骤(5)中,将提取后的边缘图像转化为神经元输入初始值,其中每一个神经元单元有一定数量的实值输入,并产生单一的实数值输出,神经元的最终输出结果如下式所示:
    Figure PCTCN2021138598-appb-100004
    其中,f为该神经网络中的激发函数,神经元中的输入单元分别为x i,与其对应的加权系数为w i,i=1,2,3,…,n-1。
PCT/CN2021/138598 2021-11-09 2021-12-16 基于人工智能技术的电力综合管廊沉降裂缝识别方法 WO2023082418A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111319884.3 2021-11-09
CN202111319884.3A CN113989257A (zh) 2021-11-09 2021-11-09 基于人工智能技术的电力综合管廊沉降裂缝识别方法

Publications (1)

Publication Number Publication Date
WO2023082418A1 true WO2023082418A1 (zh) 2023-05-19

Family

ID=79747328

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/138598 WO2023082418A1 (zh) 2021-11-09 2021-12-16 基于人工智能技术的电力综合管廊沉降裂缝识别方法

Country Status (2)

Country Link
CN (1) CN113989257A (zh)
WO (1) WO2023082418A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116630321A (zh) * 2023-07-24 2023-08-22 铁正检测科技有限公司 基于人工智能的桥梁健康智能监测系统
CN116630813A (zh) * 2023-07-24 2023-08-22 青岛奥维特智能科技有限公司 一种公路路面施工质量智能检测系统
CN117037006A (zh) * 2023-10-09 2023-11-10 山东中宇航空科技发展有限公司 一种高续航能力的无人机跟踪方法
CN117237368A (zh) * 2023-11-16 2023-12-15 临沂市公路事业发展中心兰陵县中心 一种桥梁裂缝检测方法及系统
CN117593651A (zh) * 2024-01-18 2024-02-23 四川交通职业技术学院 一种隧道裂纹分割识别方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034530A (zh) * 2018-06-19 2018-12-18 天津大学 一种基于神经网络算法的管廊规划方法
CN110044268A (zh) * 2019-03-19 2019-07-23 天津大学前沿技术研究院有限公司 基于光纤反射原理的盾构隧道接缝张开与错台监测系统
CN111401131A (zh) * 2020-02-13 2020-07-10 深圳供电局有限公司 隧道管廊的图像处理方法、装置、计算机设备和存储介质
CN112861681A (zh) * 2021-01-29 2021-05-28 长兴云尚科技有限公司 一种基于云端处理的管廊视频智能分析方法及系统

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108629775B (zh) * 2018-05-14 2021-08-03 华中科技大学 一种热态高速线材表面图像处理方法
CN109685786A (zh) * 2018-12-20 2019-04-26 龙口味美思环保科技有限公司 一种禽蛋表面裂纹的无损检测及自动分级方法
CN110119687A (zh) * 2019-04-17 2019-08-13 浙江工业大学 基于图像处理和卷积神经网络相结合的道路表面裂痕缺陷的检测方法
US20230014823A1 (en) * 2019-12-09 2023-01-19 Agency For Science, Technology And Research Defect detection in image space
CN112233067A (zh) * 2020-09-21 2021-01-15 武汉钢铁有限公司 一种热轧钢卷端面质量检测方法及系统
CN112766460A (zh) * 2021-01-12 2021-05-07 上海电机学院 一种基于子区域和bp神经网络的焊缝识别方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034530A (zh) * 2018-06-19 2018-12-18 天津大学 一种基于神经网络算法的管廊规划方法
CN110044268A (zh) * 2019-03-19 2019-07-23 天津大学前沿技术研究院有限公司 基于光纤反射原理的盾构隧道接缝张开与错台监测系统
CN111401131A (zh) * 2020-02-13 2020-07-10 深圳供电局有限公司 隧道管廊的图像处理方法、装置、计算机设备和存储介质
CN112861681A (zh) * 2021-01-29 2021-05-28 长兴云尚科技有限公司 一种基于云端处理的管廊视频智能分析方法及系统

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116630321A (zh) * 2023-07-24 2023-08-22 铁正检测科技有限公司 基于人工智能的桥梁健康智能监测系统
CN116630813A (zh) * 2023-07-24 2023-08-22 青岛奥维特智能科技有限公司 一种公路路面施工质量智能检测系统
CN116630813B (zh) * 2023-07-24 2023-09-26 青岛奥维特智能科技有限公司 一种公路路面施工质量智能检测系统
CN116630321B (zh) * 2023-07-24 2023-10-03 铁正检测科技有限公司 基于人工智能的桥梁健康智能监测系统
CN117037006A (zh) * 2023-10-09 2023-11-10 山东中宇航空科技发展有限公司 一种高续航能力的无人机跟踪方法
CN117037006B (zh) * 2023-10-09 2023-12-15 山东中宇航空科技发展有限公司 一种高续航能力的无人机跟踪方法
CN117237368A (zh) * 2023-11-16 2023-12-15 临沂市公路事业发展中心兰陵县中心 一种桥梁裂缝检测方法及系统
CN117237368B (zh) * 2023-11-16 2024-01-26 临沂市公路事业发展中心兰陵县中心 一种桥梁裂缝检测方法及系统
CN117593651A (zh) * 2024-01-18 2024-02-23 四川交通职业技术学院 一种隧道裂纹分割识别方法
CN117593651B (zh) * 2024-01-18 2024-04-05 四川交通职业技术学院 一种隧道裂纹分割识别方法

Also Published As

Publication number Publication date
CN113989257A (zh) 2022-01-28

Similar Documents

Publication Publication Date Title
WO2023082418A1 (zh) 基于人工智能技术的电力综合管廊沉降裂缝识别方法
Chen et al. A self organizing map optimization based image recognition and processing model for bridge crack inspection
CN107782733B (zh) 金属表面缺陷的图像识别无损检测装置及方法
CN113469177B (zh) 基于深度学习的排水管道缺陷检测方法及系统
CN112967243A (zh) 一种基于yolo的深度学习芯片封装裂纹缺陷检测方法
Yoo et al. Development of a crack recognition algorithm from non-routed pavement images using artificial neural network and binary logistic regression
CN108492291B (zh) 一种基于cnn分割的太阳能光伏硅片瑕疵检测系统及方法
CN111652295B (zh) 一种铁路货车钩尾销托梁脱落故障识别方法
CN111079821A (zh) 一种脱轨自动制动拉环脱落故障图像识别方法
CN111860143B (zh) 一种面向巡检机器人的火焰实时检测方法
CN110726725A (zh) 一种输电线路金具锈蚀检测方法及装置
CN113313107B (zh) 一种斜拉桥缆索表面多类型病害智能检测和识别方法
CN115147418B (zh) 缺陷检测模型的压缩训练方法和装置
CN111402249B (zh) 基于深度学习的图像演化分析方法
CN115995056A (zh) 一种基于深度学习的桥梁病害自动识别方法
CN116309292A (zh) 一种基于视觉转换层与实例分割的焊缝缺陷智能识别方法
CN111667473A (zh) 基于改进Canny算法的绝缘子憎水性等级判断方法
CN109859199B (zh) 一种sd-oct图像的淡水无核珍珠质量检测的方法
CN111310899B (zh) 基于共生关系和小样本学习的电力缺陷识别方法
CN111091554B (zh) 铁路货车摇枕断裂故障图像识别方法
CN116777898B (zh) 基于AFFormer实现3D打印挡墙施工过程裂缝测量的方法
CN112669269A (zh) 一种基于图像识别的管道缺陷分类分级方法及系统
CN110349119B (zh) 基于边缘检测神经网络的路面病害检测方法和装置
CN114140731B (zh) 一种牵引变电所异常检测方法
CN113793327B (zh) 一种基于token的高铁异物检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21963856

Country of ref document: EP

Kind code of ref document: A1