Belt tearing laser scanning detection device and belt longitudinal tearing laser detection method

Abstract

The invention relates to the technical field of laser image detection, and particularly discloses a belt tearing laser scanning detection device and a belt longitudinal tearing laser detection method. The belt tearing laser scanning detection device comprises an ellipsoidal carrier roller, laser equipment and visual equipment, the outer surface of the ellipsoidal carrier roller is an arc-shaped surface, a belt is tensioned and wound around the arc-shaped surface, and the laser equipment is fixedly arranged in the mode that laser faces the ellipsoidal carrier roller and can irradiate the belt. The visual equipment is fixedly arranged in the mode that the visual field faces the ellipsoid carrier roller and the belt is in the visual field, the structural arrangement can reconstruct the carrying working state of the belt, the tearing characteristic of the belt in the carrying state is highlighted, and the belt longitudinal tearing laser detection method is more suitable for long-distance conveying belt longitudinal tearing detection; and the method has higher detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of laser image detection. Background Art

[0002] Long-distance belt conveyors are indispensable equipment in mines, ports, and power enterprises, used for efficiently transporting bulk materials such as lumps and powders, capable of achieving uninterrupted transportation of large volumes of materials at high speeds over long distances at low cost. However, during long-term operation, with the continuous increase in the material carrying capacity and operating cycle of the belt, as well as other accidental factors causing serious wear and damage to the belt, accidents such as longitudinal tearing of the belt occur from time to time. Such accidents have a huge impact on the normal operation of enterprise production, the stability of equipment, and the safety of personnel.

[0003] Currently, existing related solutions and technologies for detecting longitudinal tearing of belts, such as a smart belt tearing detection method and device disclosed in Chinese Patent Application No. 202010405231.6. The implementation of this solution is to use a light emitter to irradiate one side of the belt, and a light intensity sensor on the other side receives the light intensity measurement value. By calculating the Mahalanobis distance of the light intensity measurement value, it is determined whether there is belt tearing; and based on the image of the belt, a deep learning image segmentation algorithm is used to extract the tearing features of the belt and obtain the feature parameters to achieve the detection function of belt tearing. In this solution, the detection is mainly carried out when the belt conveyor is in the no-load state, and the laser energy passing through the belt is detected by a photosensitive component. Since this solution relies on the light intensity received by the photosensitive component, it is greatly affected by changes in the external light environment.

[0004] Another example is a belt tearing detection method disclosed in Chinese Patent Application No. 202010965954.1. In this solution, a deep learning model is used to detect the belt tearing image. By obtaining pictures of the normal and torn conditions of the belt through a belt monitoring system, a belt tearing detection data set is formed; the belt tearing detection data set is labeled and randomly divided into a training set and a test set; a convolutional neural network model is trained using the training set to obtain a belt tearing detection model; and then based on the belt tearing detection model, the belt tearing situation in the monitoring area is detected in real time. This solution realizes the detection of belt tearing images by training a deep learning model based on a convolutional neural network. This solution requires monitoring cameras to be arranged at multiple positions of the belt conveyor to collect belt images, and the entire belt tearing monitoring system will consume a large amount of hardware costs.

[0005] In the above two solutions, a deep learning model is used to process the belt image to detect the belt tearing feature. Using the deep learning model can improve the robustness of defect detection to a certain extent. However, since the surface material of the belt is generally black rubber, the visible light camera imaging has low image contrast and cannot well image some belt tearing features. Secondly, due to the interference of the belt surface texture or scratches, there may be missed detections and false detections during the detection process. Summary of the Invention

[0006] The object of the present invention is to provide a belt tearing laser scanning detection device that can reconstruct the working state of the belt when carrying objects and highlight the tearing features existing in the belt when carrying objects. Another object of the present invention is to provide a belt longitudinal tearing laser detection method that is more suitable for detecting longitudinal tearing of long-distance conveyor belts and has a higher detection accuracy.

[0007] To achieve the above object, the technical solution of the present invention is: a belt tearing laser scanning detection device, including an ellipsoidal idler, a laser device, and a vision device. The outer surface of the axial cross-section of the ellipsoidal idler is an outwardly convex arc surface for the belt to be tensioned and wound around the arc surface of the ellipsoidal idler. The laser device is fixedly arranged in such a way that its laser is directed towards the ellipsoidal idler and the laser can irradiate on the belt tensioned and wound around the arc surface of the ellipsoidal idler. The vision device is fixedly arranged in such a way that its field of view is directed towards the ellipsoidal idler and the belt tensioned and wound around the arc surface of the ellipsoidal idler is within its field of view.

[0008] The vision device adopts an industrial area array camera with global exposure, and / or, the laser device adopts a red light laser device.

[0009] The vision device is arranged in such a way that its optical axis extends radially relative to the ellipsoidal idler, and / or, the laser device is arranged in such a way that its laser light surface irradiates radially relative to the ellipsoidal idler and the laser light surface corresponds to the roller axis direction of the ellipsoidal idler in the width direction.

[0010] A belt longitudinal tearing laser detection method, the steps of the detection method are as follows:

[0011] S1. The laser device of the pre-built belt tearing laser scanning detection device emits laser towards the belt of the belt conveyor, and the vision device of the belt tearing laser scanning detection device obtains the belt laser stripe image of the belt conveyor in real time;

[0012] The belt tearing laser detection device further includes an ellipsoidal idler. The outer surface of the axial section of the ellipsoidal idler is an outwardly convex arc surface. The belt is tensioned and wound around the arc surface of the ellipsoidal idler. The laser device emits laser light that irradiates on the belt wound around the ellipsoidal idler in a tensioned manner, and the width of the laser surface of the laser corresponds to the width direction of the belt. The visual device obtains in real time the section of the belt that is wound around the ellipsoidal idler and irradiated with laser light, thereby obtaining the belt laser stripe image;

[0013] S2. Extract the laser stripes of the belt laser stripe image obtained in real time to obtain real-time laser stripes and calculate their real-time three-dimensional coordinates in three-dimensional space;

[0014] S3. Perform three-dimensional point cloud reconstruction through the real-time three-dimensional coordinates of the real-time laser stripes to obtain the reconstructed three-dimensional point cloud data of the belt surface features;

[0015] S4. Input the three-dimensional point cloud data of the belt surface features into the pre-constructed and trained network recognition and detection model for belt tearing recognition and detection.

[0016] In step S2, after step S2, the real-time stripe optical center is further extracted from the real-time laser stripes. According to the calibration results of the internal parameters of the processed camera, the external parameters of the camera in different state postures, and the calibration of the laser light knife plane of the laser, the real-time three-dimensional coordinates of the real-time laser stripes in three-dimensional space are calculated and processed through the laser light plane equation.

[0017] The visual device uses an industrial area array camera with global exposure, and / or the laser device uses a red light laser device, and / or the visual device is arranged in the direction in which its optical axis extends radially relative to the ellipsoidal idler, and / or the laser device is arranged in the direction in which its laser light surface irradiates radially relative to the ellipsoidal idler.

[0018] Before extracting the real-time laser stripes in step S2, the Gaussian noise in the belt laser stripe image is smoothed using the Gaussian filtering algorithm.

[0019] The pre-constructed and trained network recognition and detection model in step S4 is a CNN-LSTM recognition and detection model, which includes a CNN module, an LSTM module, and a classification module. The CNN module part includes three one-dimensional convolution operations, adaptively extracts the spatial features of the point cloud sequence using the wide convolution kernel of the convolutional neural network, learns the spatial features of the point cloud sequence by using the convolutional feature information as the input feature of the LSTM network, and classifies the output feature information based on the fully connected layer function to realize the abnormal classification of the line laser point cloud sequence.

[0020] Step S4 includes preprocessing the three-dimensional point cloud data of the belt surface by the network recognition and detection model and extracting multi-scale geometric features;

[0021] The multi-scale geometric feature extraction determines the multi-scale local neighborhood space range according to the sampling density of the point cloud data, sets the scale radius r, where the minimum space radius is 3-5 times the average spacing d of the point cloud; calculates the geometric features of the three-dimensional point cloud data on the belt surface, including curvature, normal vector change rate, local density, and anisotropy;

[0022] The curvature calculation is expressed as where λ1, λ2, and λ3 are the eigenvalues of the covariance matrix calculated for the neighborhood point set within the radius r of point i;

[0023] The calculation expression for the normal vector change rate is where n i , nj represent the normal vectors of point i and its neighborhood point j;

[0024] The calculation expression for the local density is

[0025] The calculation expression for anisotropy is where λ1 and λ3 represent the maximum and minimum eigenvalues respectively.

[0026] After preprocessing and multi-scale geometric feature extraction of the three-dimensional point cloud data on the belt surface, the network recognition and detection model detects the tear features of the three-dimensional point cloud data of the reconstructed belt surface for each frame as a data sample unit of each frame.

[0027] By adopting the above technical solutions, the beneficial effects of the present invention are as follows: When the belt tear laser scanning detection device is applied, by setting an ellipsoidal idler in the idlers on the slack side of the long-distance belt conveyor to improve the state of the slack side of the belt, it can achieve the reconstruction of the long-distance belt in the load-carrying working state, highlight the tear features existing in the belt in the load-carrying state, facilitate the laser scanning of the laser device and the acquisition of laser images by the vision device, and reconstruct the tear features of the belt with high precision for identification and detection in the belt longitudinal tear laser detection method. The above belt longitudinal tear laser detection method further adopts a CNN-LSTM recognition and detection network model. By extracting multi-scale geometric feature information of the point cloud, constructing a line laser point cloud sequence structure data sample, and designing a CNN-LSTM model that fuses CNN and LSTM networks to process the long-distance belt point cloud data features and the relationship between point cloud sequence data, it is used to detect whether there are tear features in the point cloud sequence of the line laser scanning long-distance belt, and finally realizes the detection function of the long-distance belt tear, and can issue abnormal warning information in real time according to the tear feature recognition result, maximizing the safety production and stable operation of the conveyor belt. Description of the Drawings

[0028] Figure 1The present invention relates to a flow chart of a method for detecting longitudinal tearing of a belt by laser.

[0029] Figure 2 The present invention relates to a calibration flow chart of a laser detection method for longitudinal tearing of a belt.

[0030] Figure 3 This is a state reference diagram of a belt tear laser scanning detection device involved in the present invention when used on a belt conveyor.

[0031] Figure 4 This is a schematic diagram of the use of a belt tear laser scanning detection device involved in the invention.

[0032] Figure 5 This is a schematic diagram of the belt tearing on an ordinary roller.

[0033] Figure 6 This is a schematic diagram of the belt tearing on the ellipsoid roller involved in this invention.

[0034] Figure 7 The present invention relates to a schematic diagram of the framework structure of a CNN-LSTM recognition and detection network model in a belt longitudinal tear laser detection method. DETAILED DESCRIPTION

[0035] In order to further explain the technical solution of the present invention, the present invention is described in detail below through specific embodiments.

[0036] The present invention uses a belt longitudinal tear laser detection method in this embodiment to clearly describe the specific technical solutions of the belt tear laser scanning detection device and belt longitudinal tear laser detection method of the present invention. This method can be applied to the detection of belt longitudinal tears on long-distance belt conveyors.

[0037] This embodiment provides a laser detection method for longitudinal tearing of a belt. Figure 1 As shown, the detection method steps are as follows:

[0038] S1. Use the laser equipment of the pre-built belt tear laser scanning detection device to emit laser light to the belt of the belt conveyor, and use the visual equipment of the belt tear laser scanning detection device to obtain the laser stripe image of the belt of the belt conveyor in real time.

[0039] The belt tearing laser scanning detection device includes an ellipsoidal idler, a tensioning idler, a laser device and a vision device. The outer surface of the axial section of the ellipsoidal idler is an arc surface protruding outwards for the belt to bypass. The tensioning idler is arranged on both sides of the ellipsoidal idler to tension the belt on the arc surface of the ellipsoidal idler. The laser device is fixedly arranged in such a way that its laser is directed towards the ellipsoidal idler and the laser can irradiate on the belt tensioned and bypassing the arc surface of the ellipsoidal idler. The vision device is fixedly arranged in such a way that its field of view is directed towards the ellipsoidal idler and the belt tensioned and bypassing the arc surface of the ellipsoidal idler is within its field of view.

[0040] In the present invention, the load-bearing state of the belt is reconstructed by improving the loose-side idler of the long-distance belt conveyor, and the hardware settings and its installation structure for three-dimensional laser of the belt are designed to improve the design of the equipment. The schematic diagram of the improved design installation is as Figure 3 shown. Specifically, the ellipsoidal idler is arranged in the idlers of the entire belt rotary drive of the long-distance belt conveyor, and the belt runs over the ellipsoidal idler in a tensioned state on the surface of the ellipsoidal idler, as Figure 4 shown. The camera is fixedly arranged below the ellipsoidal idler at a distance where the lens field of view can completely cover the belt running over the ellipsoidal idler, and the laser device is fixedly arranged at a set angle and with the laser capable of covering the width of the belt. The effect of reflecting the longitudinal tear characteristics on the belt after the improvement of the present invention is as Figure 6 shown. Figure 5 For the effect of reflecting the longitudinal tear characteristics on the belt when using ordinary idlers, it can be seen by comparison that the improved structure with the ellipsoidal idler can expand the longitudinal tear on the belt, making the longitudinal tear characteristics obvious. That is, the application of the ellipsoidal idler can enable the belt to better reconstruct (or simulate) the load-bearing state of the belt when passing through the ellipsoidal idler, and can highlight the belt tearing characteristics.

[0041] In this embodiment, to maximize the reconstruction of the working condition of the belt in the load-bearing state, the ellipsoidal idler is processed according to the standard ellipsoidal surface, and its ellipsoidal shape expression is: Among them, x, y, and z are three-dimensional coordinates in a coordinate system. a and c respectively represent the lengths of the minor semi-axes of the ellipsoid along the x and z axes, and b represents the length of the major semi-axis of the ellipsoid along the y axis. The ellipsoidal idler uses an ellipsoidal idler with an axial length of 1000 mm, that is, the major semi-axis b = 500 mm, and the length of the minor semi-axis of the ellipsoidal idler is 200 mm, that is, the minor semi-axes a = c = 200 mm. In the embodiment, the camera is fixedly arranged at a position with the closest distance of 500 mm from the belt. Since the object moves relative to the camera during the scanning process, the camera preferably uses an industrial area array camera with global exposure. The camera selects a suitable lens according to the working scene requirements of the belt and is installed with its optical axis arranged in the direction radially extending relative to the ellipsoidal idler. Since the surface of the long-distance belt is usually made of black material and will absorb the light energy of some bands, the laser device preferably uses a red laser with a longer wavelength. This laser can adjust the laser width according to the working scene. The laser device is arranged under the ellipsoidal idler and is arranged with its laser light surface in the direction radially irradiating the ellipsoidal idler, and the laser light surface corresponds to the width direction of the belt in the width direction. The included angle θ between the laser light surface of the laser and the optical axis of the camera is 25°.

[0042] S2. Extract the real-time laser stripe from the real-time belt laser stripe image. Before extraction, the Gaussian noise in the belt laser stripe image can be smoothed by using the Gaussian filtering algorithm, which can reduce the influence of Gaussian noise on the accuracy of this step and the next extraction step.

[0043] S3. Extract the real-time stripe optical center from the real-time laser stripe. This extraction is to obtain the stripe optical center by calculating the sub-pixel center point of the real-time laser stripe, and the calculation is realized by using the Steger algorithm to extract the sub-pixel center point of the laser stripe.

[0044] S4. According to the preprocessed camera internal parameter calibration, the camera external parameter calibration in different state postures, and the calibration result of the laser light knife plane of the laser device, and through the laser light plane equation, the real-time three-dimensional coordinates of the real-time laser stripe in the three-dimensional space are calculated and processed for the extracted real-time stripe optical center.

[0045] The pre - processed camera intrinsic parameter calibration, camera extrinsic parameter calibration for different states and poses, and laser knife - edge plane calibration are carried out after the above - mentioned belt - tearing laser scanning detection device is built, installed and fixed. A calibration board with one - third of the short - side size of the camera's field of view is used to collect calibration images at different positions in the camera's field of view. The number of collected images should cover the entire camera's field of view at least; at the same time, pose images of the calibration board at different heights are collected at different positions in the camera's field of view, and laser stripe images corresponding to the poses are collected simultaneously. A calibration board data set is established, and based on the calibration board data set and the Zhang's calibration algorithm, the camera intrinsic parameter calibration and the camera calibration board pose extrinsic parameter calibration are realized based on the OpenCV algorithm library. The laser stripes are extracted from the collected laser stripe images corresponding to the poses, and the sub - pixel center points of the laser stripes are calculated to obtain the laser optical center of the calibration board. Based on the camera intrinsic parameter calibration, the camera calibration board pose extrinsic parameter calibration results, and the laser optical center extraction results of the calibration board, the three - dimensional point cloud reconstruction of the laser optical center of the calibration board is carried out. According to the three - dimensional point clouds of the laser optical centers of the calibration boards with different poses, the calculation and fitting of the laser knife - edge plane are realized to complete the laser knife - edge plane calibration, as Figure 2 shown.

[0046] S5. The three - dimensional point cloud data of the reconstructed belt surface is obtained through three - dimensional point cloud reconstruction using the real - time three - dimensional coordinates of the real - time laser stripes.

[0047] S6. The belt - tearing recognition and detection are carried out on the three - dimensional point cloud data of the reconstructed belt surface obtained through reconstruction by using the pre - constructed and trained network recognition and detection model.

[0048] In this embodiment, the pre - constructed and trained network recognition and detection model is a CNN - LSTM recognition and detection model, and its model structure is as Figure 7 shown, including a CNN module, an LSTM module and a classification module. By designing three one - dimensional convolution operations in the CNN part, the extraction of sample feature information is realized, and at the same time, the sample is compressed from the original step size of 2048 to 32, greatly reducing the subsequent calculation amount of the model.

[0049] During the model training process, the data flow and data dimension changes of the input data among various parts of the model are as follows: The initial dimension of the input sample data is (2048, 13). After transposing the input data, it is used as the input of the CNN module. After being processed by multi-layer and multi-scale convolutional kernels, the data dimension of the output of the CNN module is (256, 32). By transposing the output data of the CNN module, it is input into the LSTM part with a data dimension of (32, 256). After being processed by the LSTM module, the output data dimension is (32, 256). The LSTM output result is flattened along the time step dimension and input into the classification module. In the classification module, the classification output of the sample is obtained through the fully connected layer and the activation function. During model training, a mixed-precision training strategy and a progressive learning rate adjustment strategy are adopted. FP16 is used for forward propagation and storage, and FP32 is used for weight update and gradient accumulation processes. This method can accelerate the training process and reduce memory occupancy while maintaining the accuracy of the model as much as possible. The single-cycle learning rate adjustment strategy is adopted to help the model converge faster and improve the overall performance. Here, it should be noted that FP16 is an index for data operation in the network model training process, which refers to 16-bit floating-point numbers, also known as half-precision; FP32 refers to 32-bit floating-point numbers, known as single-precision. The memory occupied by FP32 in the operation will be twice that of FP16. Therefore, FP16 is used for forward propagation and data storage during network training to improve the calculation speed and save video memory space; while in the gradient update process, since the gradient values are all very small, to ensure calculation accuracy and avoid gradient disappearance, FP32 is adopted. After completing the model training based on the above training strategy, the optimal model file is statistically screened and saved as the final model weight file for the belt tear detection deployment of the real-time processed reconstructed belt surface three-dimensional point cloud data.

[0050] This step includes the following steps:

[0051] S6.1. The network recognition and detection model preprocesses the belt surface three-dimensional point cloud data and extracts multi-scale geometric features. The main purpose of the preprocessing here is to normalize the point cloud data coordinates, and the implementation method can adopt the zero-mean normalization algorithm. The multi-scale geometric feature extraction determines the multi-scale local neighborhood space range according to the sampling density of the point cloud data, and sets the scale radius r, where the minimum spatial radius is 3-5 times the average spacing d of the point cloud. In this solution, the multi-scale radius is determined as r ∈ {3d, 6d, 12d}; according to the geometric features of the point cloud data, the geometric features that need to be calculated in this solution include curvature, normal vector change rate, local density, and anisotropy.

[0052] Among them, the curvature calculation is expressed as In the formula, λ1, λ2, λ3 are the eigenvalues of the covariance matrix calculated from the neighborhood point set of point i within the radius r, and the curvature The closer the value is to 1, the greater the degree of convexity or concavity of the surface; conversely, the closer the value is to 0, the flatter the point cloud surface is.

[0053] Among the geometric features of the point cloud, the normal vector change rate can better reflect the consistency of the local surface direction, and its calculation expression is Where n i ,nj represents the normal vector of point i and its neighboring point j. The change rate value range is [0,2]. The larger it is, the more drastic the surface mutation is. The closer it is to 0, the smoother the point cloud surface is.

[0054] Geometric characteristics of point clouds Among them, local density represents the sampling density of point clouds in local space, that is, the number of point clouds in the local neighborhood space. Higher density indicates that the corresponding area may be a solid area or a repeated structure, and low density may be a hole or a sparse abnormal area. The local density calculation expression is Logarithmic transformation was used to reduce the magnitude difference and keep the data characteristics stable.

[0055] Geometric characteristics of point clouds Among them, anisotropy represents the directionality of the spatial distribution of local point clouds, and its calculation expression is Where λ1 and λ3 represent the maximum and minimum eigenvalues, respectively, and anisotropy The value of is in the range [0,1]. The larger the value, the more anisotropic it is, and the smaller the value, the more isotropic it is.

[0056] The CNN-LSTM recognition and detection model uses the CNN-LSTM algorithm to identify belt cracks based on line laser scanning. Based on the geometric features of the point cloud, the convolutional neural network uses a wide convolution kernel to adaptively extract spatial features from the point cloud sequence. This convolutional feature information is used as input to the LSTM network to learn the spatial features of the point cloud sequence. The output feature information is then classified using a fully connected layer function to achieve anomaly classification of the line laser point cloud sequence.

[0057] Training data preparation and dataset construction for the CNN-LSTM recognition and detection model. According to the parameters of the belt-line laser scanning camera and the reconstruction accuracy of the laser stripe point cloud, taking the embodiments of this solution as an example, for the belt contour reconstruction based on the pixel points of the laser line, it can be determined that the number of single-frame line laser reconstruction point cloud is 2048. Taking one-frame point cloud data as a sequence data unit, the normal sample and torn sample data of the belt are obtained through reconstruction by the laser detection device. By calculating 12 feature dimensions including the curvature, normal vector change rate, local density, and anisotropy of the multi-scale point cloud and the Z-axis coordinate data in the depth direction of the point cloud, a total of 13-dimensional feature data is obtained. At the same time, the data of the 13 dimensions are respectively normalized to eliminate the influence of the data dimension on the model performance. The data unit in the constructed dataset is a 13×2048-dimensional data sample, and classification labels are assigned to each frame of the collected and reconstructed data sample. For example, in the embodiments of this solution, only whether the belt has a torn feature is detected, so the data labels are two categories: "normal" and "torn".

[0058] S6.2. During the detection of the CNN-LSTM recognition and detection model, the three-dimensional point cloud data of the reconstructed belt surface is classified and processed. After the preprocessing and extraction in step S6.1, the data samples in units of each frame are detected to realize the detection of the torn feature of the three-dimensional point cloud data of the reconstructed belt surface per frame, and to realize the detection of whether the belt has a torn feature.

[0059] S7. According to the real-time detection results of the CNN-LSTM recognition and detection model, the abnormal state information of the point cloud is statistically analyzed and fed back. The cumulative number of frames continuously classified as the "torn" category in the reconstructed belt point cloud is statistically analyzed, and an abnormal warning is sent to the detection system in real time to realize the detection of the torn feature and the abnormal warning function.

[0060] The above detection method realizes the reconstruction of the longitudinal tear feature of the belt by designing a belt laser point cloud reconstruction algorithm. At the same time, based on the deep learning method, a CNN-LSTM network is designed to process the point cloud data of the belt reconstruction, realize the detection of the longitudinal tear feature of the belt, and perform real-time abnormal warning. It is particularly suitable for the intelligent monitoring and warning of the operation status of long-distance conveyor belts, such as the monitoring and operation and maintenance of the operation status of long-distance belt conveyors in the construction of intelligent mines.

[0061] In this technical solution, a belt laser scanning device structure is designed for the longitudinal tear detection of a long-distance belt. At the same time, according to the existing loose-side idler structure of the long-distance belt conveyor, an ellipsoidal idler is designed to address the possible impacts on belt tear detection, which is used to reconstruct the belt's load-bearing state on the loose side of the belt and highlight the belt tear characteristics. A belt laser scanning detection device is used to perform laser scanning on the surface of the long-distance belt for three-dimensional reconstruction of the belt surface features. The feature information of the belt surface is constructed using the three-dimensional point cloud information, and a point cloud multi-scale geometric feature sequence data set is constructed based on the three-dimensional point cloud reconstruction of the belt surface. The CNN-LSTM network model is built by integrating CNN and LSTM to extract the features and relationships of the point cloud sequence data. Finally, by inputting the serialized data features into a binary classifier, the discrimination of whether there are tear features in the point cloud sequence data is carried out, and ultimately the detection of tear features in the long-distance belt is achieved. An abnormal alarm is sent to the device management platform according to the detection result. Compared with the existing technical solutions, since this solution uses laser scanning to obtain the surface information of the long-distance belt, the influence of general external environmental light changes on the detection system is reduced. At the same time, since the implementation of this solution is based on the anomaly detection of three-dimensional point cloud information, misdetection caused by image texture features can be better avoided. Moreover, compared with the existing two-dimensional image-based anomaly detection methods, the method of designing a CNN-LSTM network based on point cloud multi-scale geometric features to detect the tear features of a long-distance belt in this invention uses the point cloud multi-scale geometric feature information to extract the line laser point cloud sequence features and identify the tear features of the point cloud sequence through the CNN-LSTM network, realizing the identification of the tear features of the long-distance belt.

[0062] The above embodiments and diagrams do not limit the product form and style of the present invention. Any appropriate changes or modifications made by those of ordinary skill in the relevant technical field shall be regarded as not departing from the patent scope of the present invention.

Claims

1. A laser scanning detection device for belt tearing, characterized in that, It includes an ellipsoidal idler, a laser device, and a vision device. The outer surface of the axial section of the ellipsoidal idler is an outwardly convex arc surface, around which the belt is tensioned and wound. The laser device is fixedly arranged in such a way that its laser is directed towards the ellipsoidal idler and the laser can irradiate on the belt tensioned and wound around the arc surface of the ellipsoidal idler. The vision device is fixedly arranged in such a way that its field of view is directed towards the ellipsoidal idler and the belt tensioned and wound around the arc surface of the ellipsoidal idler is within its field of view.

2. The belt tearing laser scanning detection device according to claim 1, characterized in that, The vision device uses an industrial area array camera with global exposure, and / or the laser device uses a red light laser device.

3. The belt tearing laser scanning detection device according to claim 1 or 2, characterized in that, The vision device is arranged in a direction where its optical axis extends radially relative to the ellipsoidal idler, and / or the laser device is arranged in a direction where its laser light surface irradiates radially relative to the ellipsoidal idler and the laser light surface corresponds to the roller axis direction of the ellipsoidal idler in the width direction.

4. A laser detection method for longitudinal tearing of a belt, characterized in that, The detection method steps are as follows: S1. The laser device of the pre-built belt tear laser scanning detection device emits laser towards the belt of the belt conveyor, and the vision device of the belt tear laser scanning detection device obtains the belt laser stripe image in real time. The belt tear laser detection device further includes an ellipsoidal idler. The outer surface of the axial section of the ellipsoidal idler is an outwardly convex arc surface, and the belt is tensioned and wound around the arc surface of the ellipsoidal idler. The laser device emits laser and irradiates on the belt tensioned and wound around the ellipsoidal idler, and the width of the laser light surface corresponds to the belt width direction. What the vision device obtains in real time is the section tensioned and wound around the ellipsoidal idler and irradiated with laser, so as to obtain the belt laser stripe image. S2. Extract the laser stripes of the belt laser stripe image obtained in real time to obtain the real-time laser stripes and calculate their real-time three-dimensional coordinates in the three-dimensional space. S3. Perform three-dimensional point cloud reconstruction through the real-time three-dimensional coordinates of the real-time laser stripes to obtain the reconstructed three-dimensional point cloud data of the belt surface features. S4. Input the three-dimensional point cloud data of the belt surface features into the pre-constructed and trained network recognition detection model for belt tear recognition detection.

5. The laser detection method for longitudinal tearing of a belt according to claim 4, characterized in that, In step S2, after step S2, the real-time stripe optical center is further extracted from the real-time laser stripes. According to the pre-processed camera internal parameter calibration, the camera external parameter calibration in different state poses, and the calibration result of the laser light knife plane of the laser, the real-time three-dimensional coordinates of the real-time laser stripes in the three-dimensional space are calculated and processed through the laser light plane equation.

6. The laser detection method for longitudinal tearing of a belt according to claim 4, characterized in that, The vision device uses an industrial area array camera with global exposure, and / or the laser device uses a red light laser device, and / or the vision device is arranged in a direction where its optical axis extends radially relative to the ellipsoidal idler, and / or the laser device is arranged in a direction where its laser light surface irradiates radially relative to the ellipsoidal idler.

7. A laser detection method for longitudinal tearing of a belt according to claim 5, characterized in that, Before extracting the real-time laser stripes in step S2, the Gaussian filtering algorithm is used to smooth the Gaussian noise in the belt laser stripe image.

8. A laser detection method for longitudinal tearing of a belt according to claim 4, 5, 6 or 7, characterized in that In the step S4, the pre-constructed and trained network recognition and detection model is a CNN-LSTM recognition and detection model, which includes a CNN module, an LSTM module, and a classification module. The CNN module part includes three one-dimensional convolution operations, which adaptively extract the spatial features of the point cloud sequence by using the wide convolution kernel of the convolutional neural network. The convolutional feature information is used as the input feature of the LSTM network to learn the spatial features of the point cloud sequence, and the output feature information is classified based on the fully connected layer function to realize the abnormal classification of the line laser point cloud sequence.

9. The laser detection method for longitudinal tearing of a belt according to claim 8, characterized in that In the step S4, the network recognition and detection model preprocesses the three-dimensional point cloud data on the belt surface and extracts multi-scale geometric features; The multi-scale geometric feature extraction determines the multi-scale local neighborhood space range according to the sampling density of the point cloud data, and sets the scale radius r, where the minimum spatial radius is 3-5 times the average point cloud spacing d; calculates the geometric features of the three-dimensional point cloud data on the belt surface, including curvature, normal vector change rate, local density, and anisotropy; The curvature calculation is expressed as where λ1, λ2, and λ3 are the eigenvalues of the covariance matrix calculated for the neighborhood point set of point i within radius r; The calculation expression for the change rate of the normal vector is where n i , and nj represent the normal vectors of point i and its neighboring point j; The local density calculation expression is The anisotropy calculation expression is where λ1 and λ3 represent the maximum and minimum eigenvalues respectively.

10. The laser detection method for longitudinal tearing of a belt according to claim 9, characterized in that, After preprocessing the three-dimensional point cloud data on the belt surface and extracting multi-scale geometric features, the network recognition and detection model detects the tearing features of the three-dimensional point cloud data on the belt surface reconstructed frame by frame with the data after preprocessing and multi-scale geometric feature extraction as the data samples per frame.

Images