阅读说明：本技术 一种皮带纵向撕裂检测方法及装置 (Belt longitudinal tearing detection method and device ) 是由 胡友民 郭志恒 张鑫 于 2020-05-28 设计创作，主要内容包括：本发明属于皮带损伤识别相关技术领域,并具体公开了一种皮带纵向撕裂检测方法及装置。该方法包括如下步骤：采集待测皮带的实时图像并进行传输和保存；对实时图像依次进行灰度转化、二位图转化、背景过滤、局部扫描和区域提取处理,进而得到该实时图像的核心区域；利用预先训练的卷积神经网络对核心区域进行识别,以此得到待测皮带有无纵向撕裂的检测结果。本发明通过无接触测量的方式采集待测皮带的实时图像,将对工作现场的影响降到最小；然后通过一系列的预处理以提取核心区域,进而有效提高识别效率和识别精度；同时采用卷积神经网络对图像信息进行快速准确的识别和分类,有效提高了皮带纵向撕裂的检测精度,实现了皮带纵向撕裂的实时监测。(The invention belongs to the technical field of belt damage identification and particularly discloses a method and a device for detecting longitudinal tearing of a belt. The method comprises the following steps: acquiring a real-time image of a belt to be detected, and transmitting and storing the real-time image; sequentially carrying out gray level conversion, bitmap conversion, background filtering, local scanning and region extraction processing on the real-time image so as to obtain a core region of the real-time image; and identifying the core area by utilizing a pre-trained convolutional neural network so as to obtain a detection result of whether the belt to be detected is longitudinally torn or not. The invention collects the real-time image of the belt to be measured in a non-contact measuring mode, and the influence on the working site is reduced to the minimum; then, a core area is extracted through a series of preprocessing, and therefore the identification efficiency and the identification precision are effectively improved; meanwhile, the convolutional neural network is adopted to rapidly and accurately identify and classify the image information, so that the detection precision of the longitudinal tearing of the belt is effectively improved, and the real-time monitoring of the longitudinal tearing of the belt is realized.)

1. A method of detecting longitudinal tearing in a belt, the method comprising the steps of:

s1, acquiring a real-time image of the belt to be measured, and transmitting and storing the image;

s2, sequentially carrying out gray level conversion, bipartite graph conversion, background filtering, local scanning and region extraction processing on the real-time image so as to obtain a core region of the real-time image;

s3, recognizing the core area by using a pre-trained convolutional neural network, so as to obtain a detection result of whether the belt to be detected has longitudinal tear.

2. The method for detecting longitudinal tear of a belt as claimed in claim 1, wherein in step S1, the real-time image is obtained by real-time capturing the image of the lower surface of the belt to be detected from bottom to top with an industrial camera.

3. The method for detecting longitudinal tear of a belt as claimed in claim 1, wherein the step S2 includes the sub-steps of:

s21, carrying out gray level conversion on the real-time image to obtain a gray level image;

s22, converting the gray level image into a binary bitmap for facilitating subsequent operation;

s23, carrying out local processing on the bipartite graph, further screening out background information and reserving a belt area image;

s24, training a convolution kernel based on the normal belt image, and traversing the belt region image by using the convolution kernel so as to identify the abnormal region belt image;

s25, extracting and storing the area with the maximum difference from the normal image in the abnormal area belt image, so as to obtain the core area of the real-time image.

4. The method of claim 3, wherein in step S24, the aspect ratio of the convolution kernel is not greater than 1: 4.

5. The method for detecting longitudinal tear of a belt as claimed in claim 1, wherein the step S3 includes the sub-steps of:

s31, inputting the core area of the real-time image into the pre-trained convolutional neural network, and extracting the features by using a convolutional layer to obtain the image features;

s32, performing feature selection and information filtering on the image features by using a pooling layer, so as to obtain a detection result of whether the belt to be detected is longitudinally torn or not.

6. The method for detecting the longitudinal tear of the belt as claimed in claim 1, wherein in step S3, a Soft-max classifier is used to obtain the detection result of whether the belt to be detected has the longitudinal tear.

7. A device for detecting the longitudinal tear of a belt by using the method as claimed in any one of claims 1 to 6, which is characterized by comprising an image acquisition unit, an image processing unit and a control unit, wherein the image acquisition unit is used for acquiring a real-time image of the belt to be detected, transmitting and storing the image; the image processing unit is used for acquiring a core area of the real-time image and identifying the core area based on a convolutional neural network so as to obtain a detection result of whether the belt to be detected has longitudinal tearing or not; the control unit is used for controlling the running state of the belt according to the detection result.

8. The belt longitudinal tear detection device of claim 7, wherein said image capturing unit comprises an image capturing sensor for capturing a real-time image of the belt under test and an image capturing card for transmitting and storing said real-time image.

Technical Field

The invention belongs to the technical field of belt damage identification, and particularly relates to a belt longitudinal tearing detection method and device.

Background

The belt conveyor is common transportation equipment in industrial production, and the stable operation and the safety management of the belt conveyor are very important for guaranteeing the transportation safety. With the continuous expansion of the transportation industry, the requirements on the carrying scale, speed and distance of the belt conveyor are continuously increased. Due to the large working load and long operation, various faults are easy to occur to the belt conveyer, and great threats are caused to safe transportation.

Belt tearing is a common and highly damaging type of failure, and often manifests itself as longitudinal tearing of the belt (surveys have shown that longitudinal tearing accounts for 90% of belt tearing incidents). Because the belt is expensive in manufacturing cost, if the tearing of the belt is not found in time, the whole belt is torn, the materials are splashed, the speed reducer is damaged, even the frame structure is damaged, great economic loss is caused, and the transportation safety is seriously threatened.

Detecting and preventing longitudinal tearing of a belt has become a major issue for health management of belt conveyors. At present, the monitoring of the running state of the belt conveyor by utilizing an information technology, a sensor technology and an image recognition technology becomes a reliable mode. The fault detection based on machine vision is a brand-new detection theory, and can be used for carrying out high-efficiency and automatic detection on longitudinal tearing of the belt.

Disclosure of Invention

In view of the above-mentioned drawbacks and/or needs of the prior art, the present invention provides a method and apparatus for detecting longitudinal tear of a belt, wherein a real-time image of the belt to be measured is acquired by means of a non-contact measurement, so as to minimize the impact on the work site; then, a core area is extracted through a series of preprocessing, and therefore the identification efficiency and the identification precision are effectively improved; meanwhile, the convolutional neural network is adopted to quickly and accurately identify and classify the image information, so that the detection precision of the longitudinal tearing of the belt is effectively improved, and the real-time monitoring of the longitudinal tearing of the belt is realized.

To achieve the above object, according to an aspect of the present invention, there is provided a belt longitudinal tear detecting method including the steps of:

s1, acquiring a real-time image of the belt to be measured, and transmitting and storing the image;

s2, sequentially carrying out gray level conversion, bipartite graph conversion, background filtering, local scanning and region extraction processing on the real-time image so as to obtain a core region of the real-time image;

s3, recognizing the core area by using a pre-trained convolutional neural network, so as to obtain a detection result of whether the belt to be detected has longitudinal tear.

Preferably, in step S1, the real-time image is obtained by capturing the image of the lower surface of the belt to be measured in real time from bottom to top with an industrial camera.

As a further preference, step S2 includes the following sub-steps:

s21, carrying out gray level conversion on the real-time image to obtain a gray level image;

s22, converting the gray level image into a binary bitmap for facilitating subsequent operation;

s23, carrying out local processing on the bipartite graph, further screening out background information and reserving a belt area image;

s24, training a convolution kernel based on the normal belt image, and traversing the belt region image by using the convolution kernel so as to identify the abnormal region belt image;

s25, extracting and storing the area with the maximum difference from the normal image in the abnormal area belt image, so as to obtain the core area of the real-time image.

Further preferably, in step S24, the aspect ratio of the convolution kernel is not greater than 1: 4.

As a further preference, step S3 includes the following sub-steps:

s31, inputting the core area of the real-time image into the pre-trained convolutional neural network, and extracting the features by using a convolutional layer to obtain the image features;

s32, performing feature selection and information filtering on the image features by using a pooling layer, so as to obtain a detection result of whether the belt to be detected is longitudinally torn or not.

Preferably, in step S3, a Soft-max classifier is used to obtain a detection result of whether the belt to be detected has a longitudinal tear.

According to another aspect of the invention, a device for detecting longitudinal tearing of a belt by using the method is provided, and the device comprises an image acquisition unit, an image processing unit and a control unit, wherein the image acquisition unit is used for acquiring a real-time image of the belt to be detected, transmitting and storing the image; the image processing unit is used for acquiring a core area of the real-time image and identifying the core area based on a convolutional neural network so as to obtain a detection result of whether the belt to be detected has longitudinal tearing or not; the control unit is used for controlling the running state of the belt according to the detection result.

Preferably, the image acquisition unit comprises an image acquisition sensor and an image acquisition card, the image acquisition sensor is used for acquiring a real-time image of the belt to be detected, and the image acquisition card is used for transmitting and storing the real-time image.

Generally, compared with the prior art, the above technical solution conceived by the present invention mainly has the following technical advantages:

1. the belt longitudinal tearing detection method provided by the invention collects the real-time image of the belt to be detected in a non-contact measurement mode, so that the influence on the working site is reduced to the minimum; then, a core area is extracted through a series of preprocessing, and therefore the identification efficiency and the identification precision are effectively improved; meanwhile, the convolutional neural network is adopted to quickly and accurately identify and classify the image information, so that the detection precision of the longitudinal tearing of the belt is effectively improved, the real-time monitoring of the longitudinal tearing of the belt is realized based on the machine vision technology, the economic loss caused by the longitudinal tearing accident of the belt is reduced, the running reliability of the belt conveyor is improved, and the production efficiency and the economic benefit of enterprises are improved;

2. particularly, the method can effectively improve the stability and robustness of the longitudinal tearing detection of the belt by optimizing the processing process of the real-time image;

3. in addition, the invention also provides a belt longitudinal tearing detection device, which utilizes a machine vision technology to detect the belt, can ensure that a detected object is not contacted and changed, has stable detection characteristics, is easy to maintain, and has the characteristics of high automation degree, high information integration degree, and good stability and robustness.

Drawings

FIG. 1 is a flow chart of a method of detecting longitudinal tearing of a belt in accordance with the present invention;

FIG. 2 is a schematic view of a longitudinal tear detection device for a belt provided in accordance with the present invention;

FIG. 3 is a schematic flow chart of the acquisition of real-time images in the preferred embodiment of the present invention;

FIG. 4 is a real-time image of the belt under test in different states in the preferred embodiment of the present invention, wherein (a) is normal, (b) is longitudinal tear, (c) is point loss, and (d) is transverse tear;

FIG. 5 is a flow chart of the extraction of the core region in the preferred embodiment of the present invention, wherein (a) is a longitudinal tear specimen, (b) is a point loss specimen, (c) is a transverse tear specimen, and (d) is a normal specimen;

FIG. 6 is a schematic diagram of the recognition using a convolutional neural network in the preferred embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

As shown in fig. 1, an embodiment of the present invention provides a method for detecting a longitudinal tear of a belt, including the following steps:

s1, shooting the lower surface image of the belt to be measured in real time from bottom to top by using an industrial camera to obtain a real-time image, and transmitting and storing the real-time image;

s2 sequentially performs gray level conversion, binary image conversion, background filtering, local scanning, and region extraction on the real-time image, so as to obtain a core region of the real-time image, which specifically includes the following sub-steps:

s21, carrying out gray level conversion on the colored real-time image to obtain a gray level image;

s22, converting the gray level image into a binary bitmap for facilitating subsequent operation;

s23, carrying out local processing on the bipartite graph, further screening out background information and reserving a belt area image;

s24 training a convolution kernel based on the normal belt image, and performing local scanning operation by using the convolution, wherein the convolution kernel can traverse the image at the pixel level, perform convolution and accumulation operation, and compare the prior range of the scanning result of the normal belt image so as to identify the belt image in the abnormal area, namely the image of the belt breakage position;

s25, extracting and storing the area with the maximum difference with the normal image in the belt image of the abnormal area, so as to obtain the core area of the real-time image;

s3, recognizing the core area by using a pre-trained two-layer convolutional neural network to obtain a detection result of whether the belt to be detected has longitudinal tear, specifically comprising the following substeps:

s31, inputting the core area of the real-time image into a pre-trained Convolutional Neural Network (CNN), and extracting the features by using a convolutional layer to obtain the image features;

s32, feature selection and information filtering are carried out on image features by using a pooling layer, a Soft-max classifier is constructed at the end of the convolutional neural network, and finally (1,0) or (0,1) is output to respectively represent two results of 'existence of longitudinal tearing' or 'absence of longitudinal tearing', and the detection result is fed back to a control center to determine whether shutdown processing is carried out or not.

In the above process, the image acquisition and core area extraction part is mainly applied to the related technology of machine vision, and the non-contact measurement mode can minimize the influence on the work site. In the crack identification part, a convolution neural network is mainly adopted, so that the image information is quickly and accurately identified and classified, and the method is suitable for the application background of belt longitudinal tearing detection. In the process of feature recognition, a knowledge base is an important recognition basis and mainly comprises belt acquisition samples accumulated previously, and a convolutional neural network model is trained by using the belt acquisition samples. After the recognition is finished, the correct recognition result is also put into a knowledge base so as to train a more perfect and accurate convolutional neural network model.

Further, the aspect ratio of the convolution kernel used in the local scanning operation in step S24 is generally not greater than 1:4, so that it can better identify the longitudinal tear feature of the belt, and extract and store the feature map, i.e., the image of the core region.

According to another aspect of the present invention, as shown in fig. 2, there is provided an apparatus for longitudinal tear detection of a belt using the above method, the apparatus comprising an image acquisition unit, an image processing unit and a control unit, wherein: the image acquisition unit is used for acquiring a real-time image of the belt to be detected and transmitting and storing the real-time image, and comprises an image acquisition sensor and an image acquisition card, wherein the image acquisition sensor is used for acquiring the real-time image of the belt to be detected, and the image acquisition card is used for transmitting and storing the real-time image; the image processing unit is used for acquiring a core area of a real-time image and identifying the core area based on a convolutional neural network so as to obtain a detection result of whether the belt to be detected has longitudinal tearing or not; the control unit is used for controlling the running condition of the belt according to the detection result.

The present invention will be further specifically described below with reference to specific examples.

Step 1: image acquisition

In order to acquire high-quality image materials, an industrial camera with the HIKVISION model of MV-CE100-30GM is adopted. As shown in FIG. 3, the industrial camera is erected under the belt and is used for shooting and sampling from bottom to top, more belt images are collected by the camera and background information is collected by the camera less by adjusting the focal length and the installation position, and the detection accuracy and the detection speed are guaranteed. The collected image materials have four types of normal, longitudinal crack, point loss and transverse crack, as shown in fig. 4, and the images are transmitted to an image processing unit through a field high-speed network for subsequent processing. And after the result that whether the belt is longitudinally torn is obtained, a signal is sent to a control unit of the belt conveyor, and then a stop instruction is sent to the belt conveyor by the control unit.

Step 2: image region extraction

(1) Grayscale image conversion

As shown in fig. 5, the image region extraction step first converts the image gradation. Since the image collected by the industrial camera is a three-channel color image, the analysis is inconvenient. Therefore, the color image is subjected to gray scale conversion to obtain a gray scale image.

(2) Two-bit map translation

After the gray scale conversion, the gray scale image is converted into a two-bit (0,1) image, so that the subsequent background filtering is facilitated.

(3) Background filtering

After the conversion of the binary image, the image is locally processed based on the prior knowledge, background information is screened out, and a belt image area is reserved.

(4) Local scanning

Based on the normal belt image, a rectangular convolution kernel is trained, so that the rectangular convolution kernel can output a determined range after pixel-level translation convolution and result accumulation are carried out on the normal belt image. When the damaged image is processed, the output is different. And the conditions of illumination and regional reflection are considered in the convolution kernel training process, and training samples are taken from different illumination conditions. Therefore, after the rectangular convolution kernel is subjected to image traversal once, the belt image of the abnormal area can be identified. And extracting and storing the area with the largest difference of the results to obtain the core area part of the image to be detected. The size of the convolution kernel is confirmed based on training experience, but the length-width ratio of the convolution kernel is generally not more than 1:4, namely the rectangular kernel is in a vertical strip shape, so that the longitudinal tearing characteristics of the belt can be better extracted.

(5) Core region extraction

And (4) performing omnibearing scanning operation on the image to be detected by using the convolution kernel trained in the local scanning. When the convolution kernel is convolved on a normal belt image, the output thereof can be stabilized within a certain range θ. When the convolution kernel is convolved on an image in which a tear state exists, the output value is not in the range θ. And after the convolution kernel finishes one-time scanning on the image to be detected, selecting the area with the maximum deviation between the convolution output value and the range theta, namely the core area of the image to be detected. The region is extracted and stored, and is applied to the subsequent crack identification step, so that the calculation complexity of subsequent image identification can be greatly reduced, and the identification precision of longitudinal tearing of the belt is improved.

And step 3: crack recognition (CNN-based belt longitudinal tear feature recognition)

After the image is subjected to the region extraction operation, the image processing unit invokes the CNN algorithm in the host to complete the feature recognition, and the recognition flow is shown in fig. 6. And (3) inputting the image core area extracted in the step (2) into a convolutional neural network for feature extraction, wherein the feature extraction is mainly divided into two steps of convolution and pooling. The convolution layer mainly plays a role in feature extraction, and regularly scans an image by setting a convolution kernel with a fixed size, wherein a single scanning area is also called a receptive field. And performing matrix element multiplication summation on the input features in the receptive field and superposing deviation values, thereby completing one-time extraction of the image features. The output feature map is then fed into a pooling layer containing a pre-set pooling function that replaces the results of a single point in the feature map with the feature map statistics of its neighboring regions to enable feature selection and information filtering. The parameters of the convolutional layer and the pooling layer can be set artificially, so that the optimal characteristic identification effect is achieved.

To complete the above feature recognition process, a knowledge base is first used to perform model training on the convolutional neural network, so that the model can better fit the feature distribution on the image. The model training of the convolutional neural network mainly depends on supervised training, namely, pictures containing labels (whether longitudinal tearing exists or not, if the longitudinal tearing exists, the labels are 1, if the longitudinal tearing does not exist, the labels are 0, and samples with the labels of 0 comprise common tearing, transverse tearing and point loss conditions) are used for network training, and the weight and the bias of neurons in each layer are adjusted layer by layer. Through the training of a large number of picture materials, the model is gradually perfected, and the method can be applied to the real-time detection of the longitudinal tearing of the belt. In addition, because the construction process of the knowledge base needs a large number of pictures containing labels, the pictures need to be marked artificially in the early stage of the construction of the knowledge base. And determining whether the image is torn or not through manual inspection, and training the convolutional neural network model after marking to a certain number. And then, the model can obtain rich and labeled picture materials to further improve the knowledge base in the real-time detection process. Along with the continuous perfection of the knowledge base, the result of model training can be more in accordance with the actual situation, so that the detection accuracy rate of longitudinal belt tearing is improved.

The training process of the convolutional neural network is divided into a forward transmission process and a back propagation process. The forward transmission is that data is input from an input layer, passes through a plurality of convolution layers and pooling layers, and is transmitted to a full connection layer and an output layer to obtain an output result. The result of the forward pass is compared to the desired output (artificially labeled result) and the mean square error is usually used to represent this difference σ. The back propagation process passes the error σ to the output of each layer, and the gradient of the parameter is found by the derivative of the function of each layer to the parameter. With the method for calculating the gradient, the optimal value can be found through optimization based on the gradient, and the updating process of parameters such as the weight, the offset, the convolution kernel and the like of the convolution neural network is completed.

The input of the CNN network is a two-dimensional graph of 10 multiplied by 50, and the result is mapped into a Soft-max classifier at the tail end through two layers of convolution kernel pooling operation. The classifier then finally outputs (1,0), (0,1) two results, representing that the image was identified as having longitudinal tear and not having longitudinal tear. For images without longitudinal tears, there may be instances of point loss and transverse tears, but the present invention is primarily useful for the detection of longitudinal tears because they occur more frequently and are more hazardous.

And 4, step 4: feedback control

If the longitudinal tearing phenomenon of the belt is identified, the control console immediately sends an instruction to a belt conveyor control system to stop the belt conveyor so as to avoid causing greater loss.

Experiments prove that the belt longitudinal tearing detection scheme provided by the invention has the detection precision of 90.4% of longitudinal tearing, can play a good role in online detection and identification, and has certain practical value.

It will be understood by those skilled in the art that the foregoing is merely a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included within the scope of the present invention.