Track gap abnormal state detection system based on improved symmetric point variational mode capsule network

Abstract

The invention relates to the field of conveying, in particular to a track gap abnormal state detection system based on an improved symmetric point variational mode capsule network, which comprises an inspection dynamic data acquisition module, a fixed point data acquisition module, a conveyor abnormal state detection module and a conveyor intelligent operation and maintenance module, the abnormal state of the conveyor track gap is detected through the system, manual control of an operator is not needed by adopting the system and the method, 24-hour multi-data type signal real-time collection and state display are achieved, conveyor abnormal information can be timely and accurately provided, unmanned monitoring of unsafe areas is achieved, and the safety of the conveyor is improved. Through the intelligent operation and maintenance platform, comprehensive state information of operation of the conveyor in a complex and dangerous environment is provided, and a safety guarantee is provided for stable operation of the conveyor.

Description

Technical Field

[0001] The present invention relates to the field of conveyors, and particularly to an abnormal state detection system for track gaps based on an improved symmetric point variational mode capsule network. Background Art

[0002] The trailer-type conveyor is a new type of belt conveyor. This conveyor uses trailers to support the conveyor belt, and the trailers run on the track. There is a rigid contact between the railway track and the wheel set, and the trailer wheel set runs with a small rolling friction force; the trailer and the conveyor belt are relatively stationary, eliminating the material wear resistance caused by the indentation resistance of the rollers and the wave motion of the conveyor belt in the traditional roller-type conveyor, saving energy during operation. Compared with the traditional roller-type belt conveyor, its energy consumption is saved by 60% and the carrying capacity is increased by 40%. The rail-type belt conveyor has a large single-load capacity and a long single-machine transportation distance, and is more suitable for long-distance continuous mineral resource transportation in the wild environment.

[0003] When the trailer-type conveyor is installed in an environment with poor conditions and complex terrain, in order to ensure safe, reliable and efficient operation, the conveyor line needs to be maintained. Most of the existing intelligent operation and maintenance systems for conveyors rely on sensors installed at fixed points for data collection, and can only monitor the status of a certain area, unable to reflect the overall operation of the equipment. For example, in the existing method with fixed intervals, its real-time performance is poor, and it may miss the real-time data of key abnormal points or sudden problems in the interval, and the method has great limitations.

[0004] The track is an important component of the trailer-type conveyor. If the track gap appears abnormally, it will seriously affect the operation of the trailers on the track and cause accidents. Therefore, the detection of the track gap is extremely important. Since it is a new model, there is currently no safe and reliable detection method. In the existing detection methods, most of the vibration data is analyzed in one-dimensional signals. Due to the existence of many noises during the on-site operation process, the signals are chaotic, resulting in low accuracy. The commonly used one-dimensional to two-dimensional analysis method is used, and there are various two-dimensional image analysis methods. Among them, in the symmetric point variational mode method, the feature information it contains is not comprehensive enough, and the petal left-right symmetric structure features are repeated, and the feature expression rate is low.

[0005] Therefore, in the detection of abnormal states of the track gaps of the trailer-type conveyor, aiming at the problems of complexity of abnormal features, incomplete feature recognition, and low abnormal recognition rate, the present invention adopts an improved symmetric point variational mode and an improved CBAM attention mechanism to improve the efficiency of feature expression of the track gaps and the accuracy and robustness of model abnormal recognition. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides an abnormal state detection system for conveyor track gaps based on an improved symmetric point variational mode capsule network, including an inspection dynamic data acquisition module, a fixed-point data acquisition module, a conveyor abnormal state detection module, and a conveyor intelligent operation and maintenance module; the system is used to detect the abnormal state of the conveyor track gaps.

[0007] Further, the inspection dynamic data acquisition module includes an inspection robot, on which a vibration sensor, a laser three-dimensional profile scanner, and an industrial camera are arranged, and the inspection robot runs synchronously with the conveyor to acquire inspection dynamic data; The inspection dynamic data includes track gap vibration data collected by the vibration sensor, three-dimensional data of track surface wear collected by the laser three-dimensional profile scanner, and track gap image data collected by the industrial camera.

[0008] Further, the fixed-point data acquisition module includes an industrial camera arranged at a fixed location, and the industrial camera acquires fixed-point data at the fixed position. The fixed-point data includes perimeter environment video data of the non-safe area of the conveyor collected by the industrial camera, which is used to monitor whether there are personnel intrusions in the dangerous area.

[0009] Further, the conveyor abnormal state detection module is internally provided with a variety of intelligent algorithm models. The conveyor abnormal state detection module analyzes and processes the dynamic data and fixed-point data obtained by collecting the data from the inspection data acquisition module and the fixed-point data acquisition module through the intelligent algorithm models internally provided.

[0010] Among them, the intelligent algorithm model includes: an improved symmetric point variational mode model, and its steps are: S1: Perform autocorrelation analysis preprocessing on the track gap vibration data collected by the vibration sensor to eliminate random noise interference, and use the uniform sliding step segmentation method for sample expansion to increase the number of available vibration data samples; S2: For the track gap vibration data under different working conditions, after the denoising process in S1, perform VMD variational mode decomposition to obtain the intrinsic mode functions in different frequency bands, decompose the complex frequency spectrum into multiple simple frequency spectra, and perform symmetric point graph transformation on each IMF so that all the obtained components are fused together to obtain a new symmetric point variational mode image; S3: Introduce the image generated in S2 into the capsule network with an attention mechanism, deploy and train the model for the four different abnormal types that may occur in the track gap, generate a final detection model, and through this model, realize the detection of the abnormal state of the track gap.

[0011] Furthermore, the intelligent algorithm model further includes: a PCA-NICP three-dimensional contour registration method model, and its steps are as follows: S1: Preprocess the three-dimensional data of the track surface wear through the CCDZ data preprocessing algorithm to eliminate invalid data and outliers; S2: Perform two-step registration, namely rough registration and fine registration, on the preprocessed three-dimensional contour data of the track to finally obtain the three-dimensional contour data of the track with high matching accuracy; S3: Visualize the track surface by using the heat map method for the finally obtained three-dimensional contour data of the track with high matching accuracy, and intuitively show the track surface wear condition by dividing with different colors.

[0012] Furthermore, the intelligent algorithm model further includes: a generative adversarial YOLOv8 network model for abnormal detection of personnel intrusion, and its steps are as follows: S1: Preprocess the video data of the perimeter environment of the non-safe area of the conveyor collected by the industrial camera, that is, the video data of personnel intrusion; S2: Put the data preprocessed in S1 into the generative adversarial YOLOv8 network model for training; S3: Verify the personnel intrusion video with the model generated in S2 to complete the abnormal detection of personnel intrusion in the environment around the conveyor.

[0013] Furthermore, the intelligent algorithm model further includes: a machine vision TCH track gap adaptive recognition method model for detecting the abnormal state of the track gap, and its steps are as follows: S1: Compare the image data of the track gap collected by the industrial camera, specifically compare the pixel values of the track gap image with a preset threshold, extract the feature area of the track gap, and obtain the feature value; S2: After threshold comparison and segmentation of the image, perform morphological processing to enhance the image features, including dilation and erosion operation processing of the image, so that the edges of the track fan image are obvious. S3: After morphological processing, use the Canny edge detection algorithm to extract the edges of the processed image. Through multi-level screening and calculation, the edges of the track gap in the image can be accurately detected; S4: Further process the extracted image edges by using the Kirchhoff transform to accurately identify the width of the track gap.

[0014] Furthermore, the detection of the abnormal state of the track gap also includes a neural network model. By adopting an improved CBAM attention mechanism, the channel attention module and the spatial attention module are connected in parallel, and then bilinear pooling feature fusion is inserted to improve the recognition ability of the abnormal state of the track gap and the robustness of the system.

[0015] Furthermore, the intelligent operation and maintenance module of the conveyor includes real-time monitoring of the conveyor state, display of faults, alarms and risk control.

[0016] The present invention adopts a data acquisition method that takes inspection data acquisition as the main body and fixed-point data acquisition as the auxiliary, as well as a real-time monitoring intelligent operation and maintenance system. Compared with the existing single fixed-point method, it can collect the areas passed by the conveyor during operation more comprehensively. It can not only collect and display the information of dangerous sections in real time, reduce manual participation, but also effectively avoid the blindness and personnel danger of manual inspection, timely discover abnormal information or emergencies of the conveyor and handle them, increasing the safety and stability of the conveyor operation.

[0017] The present invention adopts an improved abnormal state detection method based on symmetric point variational mode graph transformation to detect the abnormal state of the conveyor track gap. Compared with the unimproved method, this method breaks the petal symmetry structure of the original symmetric point transformation graph, reduces duplicate features and introduces frequency domain features, enabling a comprehensive expression of time-frequency domain information, which has more advantages and higher accuracy for subsequent abnormal state detection of the track gap.

[0018] The present invention adopts an improved CBAM attention mechanism. By connecting the channel attention module and the spatial attention module in parallel and then inserting bilinear pooling feature fusion, it improves the high-order interaction ability, can more finely mine feature relationships, enhances the expression ability of information fusion, establishes a stronger correlation in the channel and spatial dimensions, and improves the abnormal recognition accuracy and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] From the following detailed description in conjunction with the drawings, the above and other objects and advantages of the present invention will become more completely clear, where the same or similar elements are denoted by the same reference numerals.

[0020] Figure 1 : Flowchart of the working principle of the present invention; Figure 2 : Diagram of the CBAM-Capsnet anomaly detection model; Figure 3 : Improved symmetric point variational mode fusion image; Figure 4 : Relationship diagram between the number of iterations and the accuracy rate during the training process of the network model for detecting the abnormal state of the conveyor track gap; Figure 5 : Relationship diagram between the number of iterations and the loss value during the training process of the network model for detecting the abnormal state of the conveyor track gap; Figure 6 : Confusion matrix diagram during the training process of the network model for detecting the abnormal state of the conveyor track gap; Figure 7 : Relationship diagram between each model and the accuracy rate in the ablation experiment design; Figure 8 : Screen capture display diagram of the intelligent operation and maintenance display module; Label description: 0 represents the normal track gap, 1 represents the wide gap without height difference, 2 represents the track gap with height difference, and 3 represents the track gap with low height difference. Specific implementation mode

[0021] The present invention will be further described in detail below in conjunction with embodiments. The advantages and features of the present invention will become clearer with the description. However, these embodiments are exemplary only and do not constitute any limitation to the scope of the present invention. Those skilled in the art should understand that the details and forms of the technical solutions of the present invention can be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the protection scope of the present invention.

[0022] As Figures 1 to 8 shown, the present invention provides an abnormal state detection system for track gaps based on an improved symmetric point variational mode capsule network, including an inspection dynamic data acquisition module, a fixed-point data acquisition module, a conveyor abnormal state detection module, and a conveyor intelligent operation and maintenance module; the abnormal state of the conveyor track gap is detected through this system.

[0023] After performing autocorrelation analysis and noise reduction processing on the collected vibration data, variational mode decomposition is used to obtain the intrinsic mode functions in different frequency bands. An improved symmetric point variational mode transformation is performed on each component so that all the obtained components are fused together to obtain a new symmetric point variational mode image. Finally, it is introduced into the capsule network with an attention mechanism to detect the abnormal state of the track gap. It can not only collect and display the information of dangerous sections in real time, reduce manual participation, but also effectively avoid the blindness and personnel danger of manual inspection, timely discover abnormal information or emergencies of the conveyor and handle them, and increase the safety and stability of the conveyor operation.

[0024] Among them, the intelligent algorithm model includes: an improved symmetric point variational mode model, and its steps are: S1: Perform autocorrelation analysis preprocessing on the track gap vibration data collected by the vibration sensor to eliminate random noise interference, and use the uniform sliding step segmentation method for sample expansion to increase the number of available vibration data samples; S2: For the track gap vibration data under different working conditions, after the denoising process of S1, perform VMD variational mode decomposition to obtain the intrinsic mode functions in different frequency bands, decompose the complex frequency spectrum into multiple simple frequency spectra, and perform symmetric point graph transformation on each IMF so that all the obtained components are fused together to obtain a new symmetric point variational mode image; S3: Introduce the image generated in S2 into the capsule network with an attention mechanism, deploy and train the model for the four different abnormal types that the track gap may generate, generate the final detection model, and through this model, realize the detection of the abnormal state of the track gap.

[0025] Specifically: When performing autocorrelation analysis on the vibration data of the track gap, since the vibration signal of the conveyor track gap has periodic characteristics and is often disturbed by external noise, the method of the variance of the autocorrelation function is used to screen the IMF (Intrinsic Mode Function). The autocorrelation function can reveal the correlation between different time points of a time series, that is, measure the similarity of a signal to itself at a delayed time. Assume the signal is , the similarity of the signal after the delay point :

[0026] Let the modulation signal be:

[0027] where A is the amplitude, f1 and f2 are the modulation and carrier frequencies respectively, and n(t) is the noise signal,

[0028] Therefore, it can be observed that as time goes by, the autocorrelation function value of the noise signal will rapidly decay to zero, while the autocorrelation function of the modulation signal remains stable, reflecting that the modulation frequency and carrier frequency of the modulation signal do not change.

[0029] The denoised vibration data obtained under different working conditions above is decomposed by the VMD (Variational Mode Decomposition) variational mode function to obtain the intrinsic mode functions in different frequency bands, decompose the complex spectrum into multiple simple spectra, and perform a symmetric point map transformation on each IMF so that all the obtained components are fused together to obtain a new symmetric point variational mode image. The decomposition process of VMD for the signal mainly includes two major steps: First, construct a variational problem, and second, solve the variational problem. The signal x(t) is decomposed into K IMF components u(t) through the two steps, and it is ensured that the sum of each u(t) is the signal x(t), that is:

[0030] In this embodiment, an improved symmetric point variational mode map transformation method is adopted. The traditional symmetric point variational mode map transformation method only simply performs a polar coordinate transformation on the IMF modal time-domain signal of the variational mode decomposition under the condition of left-right symmetry, while the improved symmetric point variational mode map transformation method changes the original symmetry, reduces duplicate features, changes the left-right symmetric petals into asymmetric petals of the time-domain signals of different IMF modal abnormal working conditions, and at the same time adds the frequency-domain signal on this basis to form an image with a time-frequency domain interval distribution pattern, which is more conducive to the comprehensive expression of the characteristics of the non-linear conveyor track gap abnormal state signal. This method uses the discrete Fourier transform to obtain the frequency-domain signal and the original time-domain signal as inputs. Assume the track gap vibration data signal is x(t), and the time-domain point x of the track gap vibration signal is obtained through the following formulai and the frequency domain point z i transformed into the polar coordinate space where the polar coordinate radius r i is mapped by the point x i l i is mapped by the point z i and can be expressed as: , where x max and x min respectively represent the maximum and minimum values of the time domain signal x of the orbit gap vibration, and z max and z min respectively represent the maximum and minimum values of the time domain signal z of the orbit gap vibration.

[0031] The counterclockwise and clockwise rotation angles corresponding to the initial line are mapped by adjacent time points and its mathematical expression is: , where is the angle of the m-th mirror symmetry plane, m = 0, 1......, N, and N is the number of mirror symmetry planes, represents the gain of the plotting angle , is the time interval factor .

[0032] In S3: The generated image is introduced into the capsule network with an attention mechanism, and the model of four different abnormal types that may occur in the orbit gap is deployed and trained to generate the final detection model. Through this model, the detection of orbit gap anomalies is achieved; the capsule network described in S3 above has spatiality in feature extraction compared with the traditional CNN convolutional neural network. Because the basic unit of the capsule network is a capsule, representing a vector, compared with the basic unit scalar neuron of the CNN network, it can not only extract features, but also extract spatial information such as the position, state, and direction of the features, enriching the expression of features. It is a four-layer capsule network, including an input layer, a convolutional layer, a primary capsule layer, and a fully connected capsule layer.

[0033] The input layer generates a new fused image through the symmetric point variational mode decomposition algorithm for vibration data and adjusts the size of the input picture to 128×128. This layer inputs the data for which we need feature recognition. In this paper, the symmetric point variational mode decomposition diagram of the conveyor is received.

[0034] The convolutional layer extracts features from the input signal through convolutional kernels for different-scale convolutional windows. The convolutional layer connects the convolutional window with the temporal local region of the feature representation layer, performs weighted summation on the temporal local information, and transmits it to the activation function to generate corresponding features and input them into the next layer. One convolutional kernel of size 3*3 and one of size 5*5 are selected in this layer to extract low-level features. Suppose the convolutional layer has k convolutional kernels with a stride of 1, where the i th convolutional kernel is , c is the width of the convolutional kernel window for identifying local temporal features in the graph; 2 d is the dimension of the vector matrix input to the convolutional layer. Each convolutional kernel slides from top to bottom on the text matrix to perform convolution operations. Then, the feature map i generated by the th convolutional kernel is: , where is a continuous c time sampling points; is the bias term; f is the ReLU function, which can alleviate the gradient explosion and gradient vanishing phenomena; Arranging the obtained in sequence can obtain the output feature matrix of the convolutional layer: .

[0035] The capsule layer is the core of the capsule network, including the primary capsule layer and the fully connected capsule layer, i.e., the digital capsule layer. Each capsule outputs a vector representing the entity features existing at the current layer. Capsules combine the features extracted at the same position by abstracting from low-level features to high-level features, and can effectively capture the correlations within the variational mode graph of data symmetry points. The relationships between capsules are learned through the dynamic routing algorithm, and information is passed to the next layer of capsules. The primary capsule layer encapsulates the vector matrix M i = (i = 1, 2…, n - c + 1) of length 1 in the convolutional layer with different attributes. M i is the i-th row vector of M. The primary capsule layer can convert M into a capsule matrix through the transformation matrix. Suppose the dimension of the primary capsule is l i , and the transformation matrix of the i-th primary capsule is . Each transformation matrix is operated on M i , then the feature maps P i can be generated one by one:

[0036] where e iis the bias term; g is the non-linear activation function.

[0037] Since each capsule includes l i conversion matrices, the output of each capsule . Rearranging the outputs of all capsules gives the output feature matrix of the primary capsule layer: U = \left [ {{u}_{1}, {u}_{2}...} \right ]\in {R}^{(n - c + 1)\times {l}_{i}\times q} Each capsule in the digital capsule layer is connected to the capsules in the previous layer. Using the capsule matrix U obtained from the previous-level capsules, multiplying it by the conversion matrix gives the prediction vector, and by using the dynamic routing mechanism, a distinction is made for the abnormal categories of the gap of the pair-type conveyor track.

[0038] , where, represents the capsule of the j th class.

[0039] In this embodiment, in the improved symmetric point variational mode model, an improved CBAM attention mechanism module is added to assign weights to different regions of the image. When the model detects abnormal working conditions of the device, it can focus on the regions that reflect key abnormal features and suppress irrelevant information at the same time. In this way, through more reasonable resource allocation, the extraction and detection accuracy of abnormal features can be improved. The process of the improved CBAM attention mechanism is as follows: where ⊗ describes element-wise multiplication.

[0040] After the tandem output of the two, a bilinear pooling fusion module is added. The process is as follows: , where, is the final output.

[0041] In this embodiment, compared with traditional one-dimensional signals, using two-dimensional signals for anomaly detection has higher accuracy. Using the improved symmetric point variational mode graph transformation anomaly state detection method to detect the abnormal state of the conveyor track gap, this method, compared with before improvement, breaks the petal symmetry structure of the original symmetric point transformation graph, reduces duplicate features and introduces frequency domain features, can achieve a comprehensive expression of time-frequency domain information, and has more advantages and higher accuracy for subsequent detection of the abnormal state of the track gap.

[0042] In this embodiment, a PCA-NICP three-dimensional contour registration method model is also provided, and its steps are: S1: preprocess the three-dimensional data of track surface wear through the CCDZ data preprocessing algorithm to eliminate invalid data and outliers; S2: perform coarse registration and fine registration on the preprocessed three-dimensional track contour data. The coarse registration uses the principal component analysis method to construct the covariance matrix of the gradient vector in the contour curve to calculate the average gradient value of the contour curve and determine the rotation and translation parameters. The fine registration uses the normal vector iterative closest point algorithm to utilize the characteristics of the actual surface to filter out erroneous point matches, and finally obtain the track three-dimensional contour data with high matching accuracy; S3: visualize the track surface with the heat map method using the high matching accuracy track three-dimensional contour data obtained in the end, and intuitively show the track surface wear through different color divisions.

[0043] In this embodiment, the generative adversarial YOLOv8 network model for personnel intrusion anomaly detection comprises the following steps: S1: preprocessing the video data of the perimeter environment of the non-safe area of the conveyor, i.e., the personnel intrusion video data, collected by the industrial camera; S2: putting the preprocessed data of S1 into the generative adversarial YOLOv8 network model for training; S3: verifying the personnel intrusion video with the model generated by the S2 training to complete the personnel intrusion anomaly detection in the surrounding environment of the conveyor.

[0044] In this embodiment, for the machine vision TCH track gap adaptive recognition method model for track gap anomalies, the steps are as follows: S1: Compare the track gap image data collected by the industrial camera. Specifically, compare the pixel values of the track gap image with a preset threshold, and classify the pixels into two categories, one is the track gap area, and the other is the non-gap area, effectively extracting the characteristic area of the track gap and obtaining the characteristic value; S2: After threshold segmentation of the image, perform morphological processing to enhance the image features. This includes dilation and erosion operations on the image. The dilation operation makes the edges of the track gap more obvious by expanding the boundaries of the target area in the image; the erosion operation is to shrink the target area to remove small noise points and discontinuous gap areas. The combined operation of dilation and erosion helps to highlight the main features of the gap and remove unnecessary interference information; S3: After morphological processing, use the Canny edge detection algorithm to extract the edges of the processed image. Through multi-level screening and calculation, it can accurately detect the edges of the track gap in the image. The edge information is crucial for the subsequent calculation of the track gap width because the accurate extraction of the edges directly affects the accuracy of gap recognition; S4: Use the Kirchhoff transform to further process the extracted image edges to accurately identify the width of the track gap. The Kirchhoff transform is a mathematical transformation method used to process geometric shape information in images. Through this transformation, the system can extract the precise position, shape, and width information of the track gap, thereby realizing the recognition of track gap anomalies.

[0045] The detection of the abnormal state of the track gap also includes a neural network model that uses an improved CBAM attention mechanism. This mechanism combines the channel attention module and the spatial attention module in parallel, further enhancing the model's sensitivity and discriminative ability to track gap anomalies. The channel attention module first extracts the global feature information of each channel through global pooling operations and assigns weights to each channel based on this information, thus effectively highlighting the key feature channels where the abnormal gaps are located; at the same time, the spatial attention module obtains spatial features through the fusion of global average pooling and maximum pooling operations and assigns different attention weights according to the relative importance of spatial positions, thereby focusing on the key areas of track gap anomalies. After the parallel output of the channel and spatial attention modules, the system introduces bilinear pooling feature fusion to perform high-order fusion of the feature information of the two modules. By capturing the complex correlations between channel and spatial features, it further improves the richness and discriminative ability of feature expression. Through this improved CBAM attention mechanism, the system can adaptively adjust its focus in the detection of track gap abnormal states, greatly improving the model's robustness, adaptability to different gap morphology factors, and higher recognition accuracy.

[0046] The following is an example of specific parameter design in this embodiment. In this embodiment, the vibration sensor used is of the IEPE type, and the sampling frequency is 25,600 Hz. The abnormal data of the track gap collected includes four types: high-low difference track gap, low-high difference track gap, wide gap without high-low difference, and normal track gap.

[0047] Perform autocorrelation analysis on the different abnormal data of the track gap collected in S1 above, perform denoising processing, and obtain the frequency domain signal using discrete Fourier transform; apply the improved symmetric point variational mode method to the vibration signal after one-dimensional data processing in S2 above, decompose it into multiple IMF mode functions and perform improved symmetric point polar coordinate transformation to obtain a two-dimensional point-to-point variational mode fusion image. Generate one picture for every 2,048 data points, with a size of 128×128. Divide the samples of each type into 500 groups, and a total of 2,000 groups of samples are used. Divide them according to the ratio of the training set to the test set of 7:3. The improved symmetric point variational mode fusion image is as Figure 3 shown. Subsequently, introduce the image generated in S2 into the network model, which includes a CBAM attention mechanism module, a bilinear pooling fusion module, an input layer, a convolutional layer, a primary capsule layer, and a fully connected capsule layer. Design and train a deep learning model, and input the fused image features into the network to achieve accurate discrimination of abnormal types.

[0048] Specifically, as Figure 4 , Figure 5 and Figure 6 shown, during the training process of the conveyor track gap abnormal state detection network model, the batch size is set to 64, and the number of iterations is 100. According to the accuracy and loss curve results of training and testing, it can be seen that the accuracy of the training set basically reaches convergence after about 20 iterations, and the accuracy is 99.3%. The training loss tends to be stable at about 25 iterations, approaching 0.001. The accuracy of the test set also tends to be stable after about 20 iterations, reaching 99.67%. At the same time, the test loss curve drops rapidly during the iteration process and completely converges after the 25th iteration, approaching 0.003. According to the confusion matrix, the classification results of the rail abnormality are as follows: 0 represents a normal track gap, 1 represents a wide gap without high-low difference, 2 represents a high-low difference track gap, and 3 represents a low-high difference track gap.

[0049] In this embodiment, an ablation experiment is also set up to prove the effectiveness. As Figure 7As shown below, the ablation test design is as follows: Model A: Symmetric Point Variational Mode Method, using a standard attention mechanism capsule network model without any improvements. Model B: Improved Symmetric Point Variational Mode Method, which improves the symmetric petal structure in the traditional method to an asymmetric structure, reduces duplicate features, and adds frequency domain information. It uses a standard attention mechanism capsule network model without any improvements. Model C: Improved Symmetric Point Variational Mode Method, using a standard capsule network model with the attention mechanism structure changed to a parallel structure. Model D: Improved Symmetric Point Variational Mode Method, using a standard capsule network model. The improved attention mechanism adds a bilinear pooling feature fusion module on the basis of Model C. The training and optimization of the models are consistent with the above training steps.

[0050] In this embodiment, the conveyor intelligent operation and maintenance module includes real-time monitoring of the conveyor status, display of faults, alarms, and risk control. The main function of this module is to comprehensively display the operating status of the conveyor through means such as real-time data collection, status monitoring, and fault alarms, helping operation and maintenance personnel to timely discover potential problems and handle them. The display module needs to have a friendly user interface, clear alarm prompts, and intuitive status information to ensure efficient and safe operation and maintenance work. Its monitoring center can display videos of multiple areas such as the head and tail of the conveyor in real time, and vibration waveform diagrams of the tracks at each monitoring point, etc. The operation and maintenance status area displays information such as monitoring stations, abnormal stations, discovered hidden dangers, handled hidden dangers, and operation and maintenance days. The risk control area displays low, medium, and high risk sector charts for different faults to more intuitively predict the potential hidden danger trends that the conveyor may generate in the future. Among them, the abnormal station is used to display information such as the type, location, and time of the fault occurrence. Handling hidden dangers is used to display fault reports and maintenance plans so that staff can handle them immediately. The construction of the conveyor intelligent operation and maintenance system platform is as Figure 8 shown.

[0051] The present invention has the following beneficial effects compared with the prior art: 1) The present invention is an intelligent operation and maintenance system for data collection and real-time monitoring with inspection as the main body and fixed points as the auxiliary. Compared with the existing single fixed-point method, it can collect more comprehensively the areas passed through during the operation of the conveyor, avoiding the limitations of a single fixed-point area. Since the conveyor transportation environment is mostly unmanned areas with complex and dangerous terrains, through the inspection method, not only can the information of dangerous sections be collected and displayed in real time, reducing manual participation, but also the blindness and personnel danger of manual inspection can be effectively avoided, and abnormal information or emergencies of the conveyor can be timely discovered and handled, increasing the safety and stability of the conveyor operation.

[0052] 2) Compared with traditional one-dimensional signals, the two-dimensional signals used in the present invention have higher accuracy in anomaly detection. This patent adopts an improved symmetric point variational mode graph transformation anomaly state detection method to detect the anomaly state of the conveyor track gap. Compared with the unimproved method, this method breaks the petal symmetry structure of the original symmetric point transformation graph, reduces duplicate features and introduces frequency domain features, which can achieve a comprehensive expression of time-frequency domain information and has more advantages and higher accuracy in subsequent detection of track gap anomaly states.

[0053] 3) In the detection of the anomaly state of the conveyor track gap in this patent, the anomaly features are complex and the feature information is difficult to identify. This patent introduces an improved CBAM attention mechanism. By connecting the channel attention module and the spatial attention module in parallel and then inserting bilinear pooling feature fusion, the high-order interaction ability is improved, the feature relationship can be mined more finely, the expression ability of information fusion is improved, a stronger correlation is established in the channel and spatial dimensions, and the anomaly recognition accuracy and robustness of the model are improved.

[0054] 4) The present invention does not require manual operation by the operator, realizes real-time acquisition and status display of multi-data type signals for 24 hours, can provide conveyor anomaly information in a timely and accurate manner, realizes unmanned monitoring in non-safe areas, and provides comprehensive status information of the conveyor operating in complex and dangerous environments through an intelligent operation and maintenance platform, providing a safety guarantee for the stable operation of the conveyor.

[0055] The implementation schemes in the above embodiments can be further combined or replaced, and the embodiments only describe the preferred embodiments of the present invention, rather than limiting the concept and scope of the present invention. Without departing from the design idea of the present invention, various changes and improvements made by those skilled in the art to the technical solutions of the present invention all belong to the protection scope of the present invention.

Claims

1. An abnormal state detection system for track gaps based on an improved symmetric point variational mode capsule network, characterized in that: The system includes a patrol dynamic data acquisition module, a fixed-point data acquisition module, a conveyor abnormal state detection module and a conveyor intelligent operation and maintenance module.

2. The detection system according to claim 1, wherein: The inspection dynamic data acquisition module includes an inspection robot, which is equipped with a vibration sensor, a laser three-dimensional profile scanner and an industrial camera. The inspection robot runs synchronously with the conveyor and collects inspection dynamic data; the inspection dynamic data includes track gap vibration data collected by the vibration sensor, track surface wear three-dimensional data collected by the laser three-dimensional profile scanner and track gap image data collected by the industrial camera.

3. The detection system according to claim 2, characterized in that: The fixed-point data acquisition module includes an industrial camera set at a fixed location. The industrial camera collects fixed-point data at a fixed position. The fixed-point data includes video data of the perimeter environment of a non-safe area of a conveyor collected by the industrial camera, which is used to monitor whether there is human intrusion in a dangerous area.

4. The detection system according to claim 3, characterized in that: The conveyor abnormal state detection module has multiple intelligent algorithm models built in it. The conveyor abnormal state detection module uses the intelligent algorithm model built in it to obtain dynamic data and fixed-point data collected by the inspection data acquisition module and the fixed-point data acquisition module for analysis and processing.

5. The detection system according to claim 4, characterized in that: The intelligent algorithm model includes: an improved symmetric point variational modal model, and its steps are: S1: performing autocorrelation analysis preprocessing on the track gap vibration data collected by the vibration sensor to eliminate random noise interference, and using a uniform sliding step size segmentation method to expand the sample to increase the number of available vibration data samples; S2: for the track gap vibration data under different working conditions, after the denoising process of S1, VMD variational modal decomposition is performed to obtain the intrinsic mode functions in different frequency bands, the complex spectrum is decomposed into multiple simple spectrums, and each IMF is subjected to symmetric point graph transformation, so that all the components obtained are fused together to obtain a new symmetric point variational modal image; S3: the image generated by S2 is introduced into the capsule network with the attention mechanism added, and the four different abnormal types that may be generated by the track gap are modeled and trained to generate a final detection model, through which the abnormal state of the track gap is detected.

6. The detection system according to claim 4, characterized in that: The intelligent algorithm model also includes: a PCA-NICP three-dimensional contour registration method model, and its steps are: S1: preprocessing the three-dimensional data of track surface wear through the CCDZ data preprocessing algorithm to eliminate invalid data and outliers; S2: performing two-step registration of coarse registration and fine registration on the preprocessed track three-dimensional contour data, and finally obtaining track three-dimensional contour data with high matching accuracy; S3: using the heat map method to visualize the track surface with the final high matching accuracy track three-dimensional contour data, and intuitively showing the track surface wear through different color divisions.

7. The detection system according to claim 4, wherein: The intelligent algorithm model further includes: a YOLOv8 network model for generating and opposing for abnormal detection of personnel intrusion, and its steps are as follows: S1: Preprocess the video data of the perimeter environment of the non-safe area of the conveyor collected by the industrial camera, that is, the video data of personnel intrusion; S2: Put the data preprocessed in S1 into the YOLOv8 network model for generating and opposing for training; S3: Use the model generated in S2 to verify the video of personnel intrusion to complete the abnormal detection of personnel intrusion in the environment around the conveyor.

8. The detection system according to claim 4, wherein: The intelligent algorithm model further includes: a machine vision TCH track gap adaptive recognition method model for detecting the abnormal state of track gaps, and its steps are as follows: S1: Compare the image data of the track gaps collected by the industrial camera, specifically compare the pixel values of the track gap images with a preset threshold, extract the feature regions of the track gaps, and obtain the feature values; S2: After threshold comparison and segmentation of the images, perform morphological processing to enhance the image features, including dilation and erosion operations on the images, so that the edges of the track fan images are obvious. S3: After morphological processing, use the Canny edge detection algorithm to extract the edges of the processed images. Through multi-level screening and calculation, the edges of the track gaps in the images can be accurately detected; S4: Use the Kirchhoff transform to further process the extracted image edges to accurately identify the width of the track gaps.

9. The detection system according to claim 8, characterized in that: The detection of the abnormal state of the track gap also includes a neural network model. An improved CBAM attention mechanism is adopted. By connecting the channel attention module and the spatial attention module in parallel and then inserting bilinear pooling feature fusion, the recognition ability of the abnormal state of the track gap and the robustness of the system are improved.

10. The detection system according to claim 3, wherein: The intelligent operation and maintenance module of the conveyor includes real-time monitoring of the conveyor state, display of faults, alarms and risk control.

Images