Intelligent conveyor belt damage monitoring system and multi-dimensional data analysis method thereof

Abstract

The invention discloses a conveyor belt damage intelligent monitoring system and a multi-dimensional data analysis method thereof, relates to the technical field of industrial equipment maintenance, and solves the problems that an existing conveyor belt damage monitoring system is low in image definition, poor in damage detection accuracy and insufficient in multi-dimensional data fusion. The system comprises a data acquisition module, an image enhancement module, a multi-scale feature extraction module, a damage type classification and positioning module and a multi-dimensional data fusion and analysis module. The working state, the image and the environment parameters of the conveying belt are collected through a sensor network; performing dynamic range compression, noise suppression and detail enhancement on the image through an image enhancement module; capturing detail features of the conveyor belt damage through a multi-scale feature extraction module; the damage types are accurately classified through the damage type classification and positioning module; comprehensive fusion analysis is carried out through a multi-dimensional data fusion and analysis module; the accuracy and reliability of damage detection of the conveying belt are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment maintenance, and more specifically to an intelligent monitoring system for conveyor belt damage and a multi-dimensional data analysis method thereof. Background Art

[0002] With the advancement of Industry 4.0, smart manufacturing and industrial Internet of Things technologies have been widely used. As one of the key equipment in industrial production, the operating status of the conveyor belt is directly related to the stability and production efficiency of the production line. In recent years, with the development of sensor technology, big data analysis and artificial intelligence, the intelligent monitoring system for conveyor belt damage has gradually become the focus of industry attention. By real-time monitoring of the operating status of the conveyor belt, timely detection and warning of potential damage can effectively reduce downtime and maintenance costs and improve production efficiency.

[0003] The existing intelligent monitoring system for conveyor belt damage mainly relies on a variety of sensors and data analysis technologies; common sensors include visual sensors, vibration sensors, temperature sensors and acoustic sensors, which can collect real-time operating data of conveyor belts. Subsequently, the collected data is processed and analyzed through big data analysis and machine learning algorithms to identify and predict conveyor belt damage. For example, Chinese patent CN117274733A proposes to input conveyor belt damage images into a well-trained conveyor belt damage detection model to perform conveyor belt damage detection, thereby improving the detection accuracy of the conveyor belt damage detection model; Chinese patent CN108510488B proposes a conveyor belt four damage detection method based on a residual network, which trains image samples extracted from a conveyor belt running video through a residual network to achieve real-time detection of four types of conveyor belt damage, namely, "tear", "scratch", "edge cutting" and "pothole". However, in practical applications, for example, there are generally weak and uneven lighting, dust and high humidity in coal mines, and the image clarity of the collected images is low or even blurred, which makes it difficult to segment tear damage and affects the accuracy of damage recognition. At the same time, the existing damage detection algorithms may have different recognition effects for different types of damage (such as cracks, wear, tear, etc.), especially when the damage pattern is more complex, it is impossible to integrate multiple dimensions of data such as time series data, spatial distribution information, and physical properties to form valuable insights, and it is difficult to accurately distinguish normal wear from potential faults. Therefore, when dealing with complex damage patterns, the false alarm rate is high. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention discloses an intelligent monitoring system for conveyor belt damage and a multi-dimensional data analysis method thereof, aiming to solve the problems of low image clarity, inaccurate damage type identification, and high false alarm rate caused by insufficient multi-dimensional data fusion in the existing intelligent monitoring system for conveyor belt damage.

[0005] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0006] A conveyor belt damage intelligent monitoring system, comprising:

[0007] The data acquisition module is used to collect various types of data of the conveyor belt through the sensor network, including the working status of the conveyor belt, the conveyor belt image, and the environmental parameters;

[0008] An image enhancement module is used to judge the image clarity through standard deviation and entropy based on the conveyor belt image data collected by the data acquisition module. If the image clarity is lower than a preset threshold, the image is subjected to dynamic range compression, noise suppression and detail enhancement through an adaptive Retinex algorithm and an adaptive threshold histogram equalization method, and the enhanced image is output to a multi-scale feature extraction module; if the image clarity is higher than or equal to the preset threshold, the image is directly output to the multi-scale feature extraction module;

[0009] A multi-scale feature extraction module, which is used to capture the detailed features of conveyor belt damage based on the output image of the image enhancement module through a pyramid pooling convolution structure, where the damage includes cracks, wear and tear; the pyramid pooling convolution structure includes an input layer, a depth-separable convolution layer, a multi-scale pooling layer, a feature fusion layer and an output layer;

[0010] A damage type classification and positioning module is used for classifying and positioning damage based on the output data of the multi-scale feature extraction module through an attention-based target detection algorithm; the attention-based target detection algorithm calculates the weight of each area in the image through the attention mechanism, and classifies the extracted feature vectors through a multi-classification Softmax layer to determine the damage type;

[0011] The multi-dimensional data fusion and analysis module is used to integrate the damage development rate, damage location distribution, material stress and material temperature data based on the collected data of the data acquisition module and the output data of the multi-scale feature extraction module through the graph neural network and the long short-term memory network, perform fusion analysis, and output a damage risk assessment report, including damage type, degree, and development trend.

[0012] As a further technical solution of the present invention, the working method of the image enhancement module for dynamic range compression, noise suppression and detail enhancement of the image is as follows: first, the brightness perception characteristics of the human visual system are simulated by an adaptive Retinex algorithm to perform global and local contrast enhancement on the image, while suppressing the influence of uneven lighting; based on the enhanced image data, the histogram distribution of each small area is dynamically adjusted by an adaptive threshold histogram equalization method to adjust the local dark or bright areas; during the image processing process, the adaptive Retinex algorithm and the adaptive threshold histogram equalization method automatically adjust the parameters according to the actual lighting and noise conditions of the image to ensure the enhancement effect of the image under different lighting conditions.

[0013] As a further technical solution of the present invention, the working principle of the adaptive Retinex algorithm is as follows: based on the conveyor belt image data of the data acquisition module, the standard deviation and entropy of the image are calculated through the image processing library and mathematical statistics methods to quantify the clarity of the image; if the image clarity is lower than the preset threshold, the processing flow of the adaptive Retinex algorithm is triggered; if the image clarity meets or exceeds the preset threshold, the processing flow of the adaptive Retinex algorithm is skipped; when the image clarity is lower than the preset threshold, the image is decomposed into brightness components at multiple scales through the color restoration multi-scale method; the color restoration multi-scale method is based on the Gaussian pyramid model, and the image is smoothed by Gaussian filters of different scales to obtain brightness information at different scales; then, each scale is subjected to the color restoration multi-scale method, and the image is smoothed by Gaussian filters of different scales to obtain brightness information at different scales; then, the image ... The brightness components at different scales are logarithmically transformed. After the transformation, the adaptive Retinex algorithm fuses the logarithmic brightness components at each scale through a weighted average mechanism. During the fusion process, the contribution of different scales to image details is considered, and different weight values ​​are given to the brightness components at different scales through an adaptive gain adjustment mechanism to enhance the global and local contrast of the image. The weight value is based on the local statistical characteristics of the image to achieve adaptive gain adjustment. The weight gain coefficient is automatically adjusted according to the local brightness and contrast information of each pixel in the image based on the local statistical characteristics of the image. After the gain adjustment, the adaptive Retinex algorithm recombines the enhanced brightness component with the color information of the original image through color space conversion to generate an enhanced image.

[0014] As a further technical solution of the present invention, the adjustment method of the adaptive threshold histogram equalization method is:

[0015] Step 401: Divide the image into a number of overlapping or non-overlapping small windows by sliding the window, and each window serves as an independent processing unit;

[0016] Step 402: Calculate the frequency distribution of the pixel grayscale values ​​in each small area by a histogram statistical method to generate a local histogram;

[0017] Step 403: Calculate the brightness mean and variance of each small area according to the local histogram of each small area. If the gray value distribution in the small area is relatively concentrated, lower the preset threshold value. If the gray value distribution in the small area is relatively dispersed, increase the preset threshold value.

[0018] Step 404: based on the dynamically modified threshold, the grayscale histogram is divided into a low grayscale area below the threshold and a high grayscale area above the threshold, and the low grayscale area below the threshold and the high grayscale area above the threshold are respectively subjected to histogram equalization processing by a local histogram equalization method; the local histogram equalization method maps the local histogram of each small area by local contrast stretching and grayscale value redistribution to generate a new grayscale value distribution;

[0019] Step 405: During the equalization process, if the grayscale value variation range in a certain area is small, grayscales in the area are merged into one grayscale by grayscale merging; if the grayscale value variation range in a certain area is large, grayscales in the area are further subdivided by grayscale subdivision;

[0020] Step 406: stitch together the equalization results of each small area through image stitching to generate a final enhanced image.

[0021] As a further technical solution of the present invention, the working steps of the pyramid pooling convolution structure include:

[0022] Step 501: receiving the output image of the image enhancement module through the input layer;

[0023] Step 502: Perform preliminary feature extraction on the input image through the depthwise separable convolution layer; the depthwise separable convolution independently processes the spatial information and channel information of the image through depthwise convolution and pointwise convolution operations respectively;

[0024] Step 503: performing a multi-scale pooling operation on the initially extracted feature map through the multi-scale pooling layer; the multi-scale pooling layer uses pooling windows of different sizes to downsample the feature map to obtain feature representations at different scales;

[0025] Step 504: The multi-scale features output by the multi-scale pooling layer are fused through the feature fusion layer; the feature fusion layer integrates the feature maps of different scales into a unified feature representation by element-by-element addition;

[0026] Step 505: output the feature map integrated by the feature fusion layer as the final feature representation through the output layer.

[0027] As a further technical solution of the present invention, the extraction method of the depth-separable convolution layer for preliminary feature extraction of the input image is as follows: first, an independent convolution kernel is assigned to each channel of the input image through spatial convolution to capture the spatial features within the channel and generate an intermediate feature map; then, the intermediate feature map generated by the spatial convolution is subjected to inter-channel feature fusion through a point-by-point convolution mechanism, and the point-by-point convolution uses a 1x1 convolution kernel to perform weighted summation of the feature values ​​of all channels at each position through linear transformation, and adds a bias term; in the process of depth convolution and point-by-point convolution, each channel of the feature map is normalized through a batch normalization mechanism to normalize the distribution of the feature map and reduce the influence of internal covariate shift; on the feature map after point-by-point convolution, each element of the feature map is operated element-by-element through the nonlinear activation function ReLU, and if the element value is greater than 0, it remains unchanged; if the element value is less than or equal to 0, it is set to 0.

[0028] As a further technical solution of the present invention, the working method of the target detection algorithm based on the attention mechanism is:

[0029] Step 701, pre-processing the real-time conveyor belt image data collected by the data collection module through image enhancement, denoising and normalization processing;

[0030] Step 702: Perform multi-layer convolution operations on the pre-processed input image through a convolutional neural network to extract multi-level features of the image;

[0031] Step 703: Calculate the weight of each region in the image through a multi-scale attention mechanism to generate an attention map. The calculation formula is:

[0032]

[0033] In formula (1), n ​​represents the feature map output by the feature extraction network; G i (x) and G i (m) represents the global average pooling and global maximum pooling operations, which are used to capture global information; X i and M i Represent local average pooling and local maximum pooling operations respectively; i represents different pooling window sizes, which are used to capture multi-scale features; θ represents the inter-channel relationship vector, which is used to capture the correlation between channels; represents the Sigmoid activation function, which is used to limit the value of the attention map to between 0 and 1; P i represents the concatenation operation, which is used to integrate information between different scales and channels; μ cRepresents the convolution operation, which is used to further extract features and generate attention maps;

[0034] Step 704: Based on the attention map outputted in step 703, the attention weight vector is multiplied by the feature value at the corresponding position of the feature map by an element-by-element multiplication operation to generate a weighted feature map;

[0035] Step 705: Based on the weighted feature graph, the extracted feature vector is classified through a multi-classification Softmax layer to determine the damage type; the multi-classification Softmax layer maps the feature vector of the weighted feature graph to a vector space of a fixed length through a fully connected layer to generate a classification feature vector; and the classification feature vector is normalized through an exponential function and a normalization operation to generate a probability distribution of each damage category;

[0036] Step 706: During the training process of the algorithm, the difference between the predicted result and the true label is calculated through the cross entropy loss function, and the back propagation algorithm is applied to pass the error back to each parameter in the network layer by layer.

[0037] As a further technical solution of the present invention, the attention mechanism is also used to guide the region proposal network to generate candidate regions, thereby improving the accuracy of damage positioning. The specific working method is as follows: first, the weight of each region in the image is calculated through the attention mechanism, and the image features are weighted; based on the feature map after weighted processing by the attention mechanism, a series of candidate regions are generated on the feature map by sliding a number of predefined anchor boxes on the feature map. In the process of sliding the anchor box, foreground / background classification and bounding box regression are predicted for each anchor box through the convolution layer; in the process of generating candidate regions, a multi-scale feature map is also constructed through a feature pyramid network to strengthen the association of key features across scales.

[0038] As a further technical solution of the present invention, the working method of the multi-dimensional data fusion and analysis module is: using a graph neural network to process damage location distribution information, constructing a topological structure of conveyor belt damage, capturing the spatial correlation between damages, and processing the damage development rate through a long short-term memory network to capture the change trend of damage over time. Finally, the outputs of the graph neural network and the long short-term memory network are fused through a fully connected layer to generate a damage risk assessment report, including damage type, degree, and development trend.

[0039] As a further technical solution of the present invention, a multi-dimensional data analysis method of a conveyor belt damage intelligent monitoring system includes the following steps:

[0040] S1. Process the damage location distribution information through a graph neural network; the graph neural network constructs the topological structure of the conveyor belt damage through nodes and edges, each node represents a damage point, and the edge represents the spatial relationship between the damage points; the graph neural network also aggregates the features of each node through a graph convolution operation, so that the feature representation of each node not only contains its own attributes, but also integrates the influence of adjacent nodes to form a topological structure diagram of the conveyor belt damage; the formula expression for constructing the topological structure diagram of the conveyor belt damage by the graph neural network is:

[0041]

[0042] S2. In formula (2), represents the feature representation of the kth node in the l+1th layer; σ represents the ReLU nonlinear activation function; N(k) represents the neighbor set of node k; c kj represents the normalization coefficient, which is used to balance the contribution of different neighbor nodes; W l Represents the weight matrix of the lth layer, which is used to learn the relationship between node features; represents the feature representation of the kth neighbor node in the lth layer; b (l) is the bias term of the lth layer, used to adjust the baseline of feature representation;

[0043] S3. Processing the damage development rate through a long short-term memory network to capture the change trend of the damage over time; the long short-term memory network captures the change trend of the damage over time through an input gate, a forget gate, an output gate and a cell state, and dynamically adjusts the memory degree of historical information;

[0044] S4. After completing spatial correlation capture and temporal trend analysis, the spatial features output by the graph neural network and the temporal features output by the long short-term memory network are integrated through the weight matrix and nonlinear activation function of the fully connected layer to generate a comprehensive feature vector z.

[0045] S5. Based on the comprehensive feature vector z output by the fully connected layer, the comprehensive feature vector is normalized by the Softmax function to generate the probability distribution of each damage category; the formula expression of the Softmax function is:

[0046]

[0047] In formula (3), TAU k is the probability of the kth category; z k represents the kth element of the comprehensive feature vector; C is the total number of categories; z j Represents the j-th element of the comprehensive feature vector.

[0048] Positive beneficial effects:

[0049] 1. Through the image enhancement module, using the adaptive Retinex algorithm and the adaptive threshold histogram equalization method, the present invention can effectively improve the clarity of conveyor belt images collected in harsh environments such as low light and high dust. The detailed features in the image, such as cracks, wear, and tear, can be clearly displayed in the enhanced image even if they are blurred in the original image. This directly solves the problem of difficulty in tear damage segmentation caused by low image clarity and significantly improves the accuracy of damage identification.

[0050] 2. Through the pyramid pooling convolution structure of the multi-scale feature extraction module, the present invention can capture the detailed features of conveyor belt damage at different scales, which not only improves the recognition accuracy of single-type damage (such as cracks and wear), but also significantly enhances the recognition ability of complex damage patterns (such as the coexistence of multiple damages). Through the combination of the deep separable convolution layer and the multi-scale pooling layer, the system can extract rich feature information from different scales, providing a solid foundation for subsequent damage classification and positioning.

[0051] 3. The damage type classification and positioning module adopts an attention-based target detection algorithm. By calculating the weight of each area in the image, it can accurately locate the damage position and determine its type, solving the problem of inconsistent recognition effects of existing damage detection algorithms under complex damage patterns and significantly reducing the false alarm rate. At the same time, the multi-classification Softmax layer ensures the accuracy of damage type classification, allowing the system to reliably distinguish between normal wear and potential failures.

[0052] 3. The multi-dimensional data fusion and analysis module integrates data from multiple dimensions such as damage development rate, damage location distribution, material stress, and material temperature, and uses graph neural networks and long short-term memory networks for fusion analysis, which not only achieves a comprehensive assessment of conveyor belt damage, but also improves the accuracy of the assessment. By capturing the changing trends of damage in space and time, the system can more accurately predict the development trend of damage and provide timely warning information to operation and maintenance personnel, thereby effectively avoiding the occurrence of potential failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0054] Figure 1 This is a structural diagram of an intelligent monitoring system for conveyor belt damage according to the present invention;

[0055] Figure 2 A step diagram of an adjustment method of a histogram equalization method of an adaptive threshold according to the present invention;

[0056] Figure 3 It is a processing flow principle framework diagram of the adaptive Retinex algorithm of the present invention;

[0057] Figure 4 A working method diagram of the target detection algorithm based on the attention mechanism of the present invention;

[0058] Figure 5 It is a flow chart of the working method of the multi-dimensional data fusion and analysis module of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, an intelligent monitoring system for conveyor belt damage includes: a data acquisition module for collecting various types of conveyor belt data through a sensor network, including the working status of the conveyor belt, conveyor belt images, and environmental parameters; in the specific implementation, the data acquisition module uses a variety of sensors, first of all, there are photoelectric sensors, pulse-modulated active photoelectric detection system-type electronic sensors are selected, and infrared light detection is used. The photoelectric sensors are installed on both sides below the conveyor belt, and their lateral positions are arranged to detect the maximum damage on the belt edge. Through the reflection and reception of infrared light, the photoelectric sensor can quickly detect the edge damage of the conveyor belt without contact and damage;

[0061] There are also visual sensors (cameras) used to collect real-time images of the conveyor belt. The vertical distance from the conveyor belt is determined according to the focal length of the camera to ensure that the photo area covers the conveyor belt to ensure the continuity and integrity of the image. The light source equipment is installed on both sides of the conveyor belt at a deviation of 30 degrees from the vertical direction; the camera collects damage features such as cracks, wear, and tears on the surface of the conveyor belt through timed shooting or continuous video streaming, and transmits the image data to the data processing center in real time through network transmission or wired connection;

[0062] It also includes a longitudinal tear sensor, which is used to detect longitudinal tearing of the conveyor belt; it is installed below the receiving point of the belt conveyor to ensure that foreign objects penetrating the conveyor belt can be detected. When the conveyor belt is penetrated by foreign objects, the material falls on the sensor, and the sensor sends a signal to the control box, issuing an alarm and a stop signal.

[0063] Secondly, it also includes temperature and humidity sensors, pressure sensors and acceleration sensors; temperature and humidity sensors are used to monitor the temperature and humidity of the conveyor belt and the surrounding environment, and evaluate the thermal stress, wet stress and environmental conditions of the material. Installed in key parts of the conveyor belt, such as bearings, motors and conveyor belt surfaces, as well as key locations in the environment, such as both sides and above the conveyor belt; pressure sensors are used to monitor the load and pressure distribution of the conveyor belt and evaluate the force of the conveyor belt. Pressure sensors are installed on the supporting structure of the conveyor belt, such as rollers and drums, as well as key bearing points of the conveyor belt. The pressure sensor regularly reads the pressure value, collects the load and pressure distribution data of the conveyor belt at different positions, and transmits the pressure data to the data processing center in real time by wireless or wired means. Acceleration sensors are used to monitor the vibration and motion state of the conveyor belt and evaluate the running stability of the conveyor belt. Acceleration sensors are installed on the drive device, bearings and support structure of the conveyor belt. The acceleration sensor regularly reads the acceleration value, collects the vibration amplitude and frequency data of the conveyor belt at different positions, and transmits the acceleration data to the data processing center in real time by wireless or wired means.

[0064] All sensors transmit the collected data to the data processing center in real time through wired or wireless networks (such as Wi-Fi, LoRa, Zigbee, etc.). The data processing center stores and preprocesses the received data to provide a basis for subsequent feature extraction and data analysis. The specific contents of the acquired data include: conveyor belt image data, including high-definition images of the conveyor belt surface, which are used to detect damage features such as cracks, wear, and tear; temperature data, including temperature values ​​of key parts such as the conveyor belt surface, bearings, and motors, which are used to evaluate the thermal stress and environmental conditions of materials; humidity data, including humidity values ​​of key parts such as the conveyor belt surface, bearings, and motors, which are used to evaluate the wet stress and environmental conditions of materials; pressure data, including the load and pressure distribution values ​​of the conveyor belt at different positions, which are used to evaluate the force conditions of the conveyor belt; acceleration data, including the vibration amplitude and frequency values ​​of the conveyor belt at different positions, which are used to evaluate the running stability of the conveyor belt.

[0065] The image enhancement module is used to judge the image clarity through standard deviation and entropy based on the conveyor belt image data collected by the data acquisition module. If the image clarity is lower than the preset threshold, the image is compressed in dynamic range, noise is suppressed and details are enhanced through the adaptive Retinex algorithm and the adaptive threshold histogram equalization method, and the enhanced image is output to the multi-scale feature extraction module; if the image clarity is higher than or equal to the preset threshold, it is directly output to the multi-scale feature extraction module; the working method of the image enhancement module for dynamic range compression, noise suppression and detail enhancement of the image is as follows: first, the brightness perception characteristics of the human visual system are simulated by the adaptive Retinex algorithm, and the image is globally and locally contrast enhanced, while suppressing the influence of uneven illumination; based on the enhanced image data, the histogram distribution of each small area is dynamically adjusted through the adaptive threshold histogram equalization method, and the local dark or bright area is adjusted; in the image processing process, the adaptive Retinex algorithm and the adaptive threshold histogram equalization method automatically adjust the parameters according to the actual illumination and noise conditions of the image to ensure the enhancement effect of the image under different illumination conditions;

[0066] In the specific implementation, after starting the processing flow of the adaptive Retinex algorithm, such as Figure 3 As shown: The module uses the adaptive Retinex algorithm to perform global and local contrast enhancement on the image. Then global and local contrast enhancement is performed to improve image quality. Then the illumination and noise are checked, and if there is uneven illumination, suppression processing is performed. If the image needs further adjustment, adaptive threshold histogram equalization is performed. Adaptive threshold histogram equalization adjusts the histogram distribution according to the local area of ​​the image. According to the results of histogram equalization, the histogram distribution of the image is dynamically adjusted. Finally, check again whether the image enhancement has achieved the desired effect. If yes, the process ends. Otherwise, return to the adaptive Retinex algorithm step and reprocess.

[0067] The working principle of the adaptive Retinex algorithm is as follows: based on the conveyor belt image data of the data acquisition module, the standard deviation and entropy of the image are calculated through the image processing library and mathematical statistics methods to quantify the clarity of the image; if the image clarity is lower than the preset threshold, the processing flow of the adaptive Retinex algorithm is triggered; if the image clarity meets or exceeds the preset threshold, the processing flow of the adaptive Retinex algorithm is skipped; when the image clarity is lower than the preset threshold, the image is decomposed into brightness components at multiple scales through the color restoration multi-scale method; the color restoration multi-scale method is based on the Gaussian pyramid model, and the image is smoothed through Gaussian filters of different scales to obtain brightness information at different scales; then the brightness component at each scale is logarithmically transformed, and after the transformation, the adaptive Re The Retinex algorithm fuses the logarithmic brightness components at each scale through a weighted average mechanism. During the fusion process, the contribution of different scales to the image details is considered, and different weight values ​​are assigned to the brightness components at different scales through an adaptive gain adjustment mechanism to enhance the global and local contrast of the image; wherein, the weight value implements adaptive gain adjustment based on the local statistical characteristics of the image; the weight gain coefficient is automatically adjusted based on the local brightness and contrast information of each pixel in the image based on the local statistical characteristics of the image; after the gain adjustment, the adaptive Retinex algorithm recombines the enhanced brightness component with the color information of the original image through color space conversion to generate an enhanced image; in the specific implementation, the image clarity judgment is to quantify the image quality by calculating the standard deviation and entropy of the image. The standard deviation reflects the contrast of the image, and the entropy reflects the amount of information in the image. The calculation formula of the standard deviation is:

[0068]

[0069] In formula (1), S i is the gray value of the i-th pixel in the image; μ represents the average gray value of the image; N represents the total number of pixels in the image; the entropy calculation formula is:

[0070]

[0071] Among them, p i is the probability that a pixel with gray value i appears in the image, and L is the gray level of the image.

[0072] In practical applications, the standard deviation and entropy of the image are first calculated through an image processing library (such as OpenCV). If the standard deviation and entropy are lower than the preset threshold, the image is considered to be insufficiently clear and needs to be enhanced; otherwise, the image is directly output to the multi-scale feature extraction module. This method can quickly determine the image quality, ensure that only images that need to be enhanced enter the subsequent processing flow, and improve the efficiency of the system.

[0073] The color restoration multi-scale method is an image enhancement technology based on multi-scale analysis. It uses the Gaussian pyramid model to decompose the image into brightness components at different scales. The brightness component at each scale reflects the brightness information of the image at different resolutions. In this way, the detailed features in the image can be captured while suppressing noise and interference. In the intelligent monitoring system for conveyor belt damage, the color restoration multi-scale method is used to pre-process the collected conveyor belt images. By decomposing the image into brightness components at multiple scales, the algorithm can more accurately identify damage features such as cracks, wear, etc.

[0074] Logarithmic transformation is a nonlinear transformation that can compress the brightness value of an image so that an image with a wider brightness range can be more evenly distributed in a new brightness range after the transformation. This transformation helps to enhance the contrast of the image and make the details more prominent. In the adaptive Retinex algorithm, the logarithmic transformation is used to process the brightness component at each scale. Through the transformation, the algorithm can better capture the details in the image and suppress the impact of uneven lighting.

[0075] The weighted average mechanism is a data fusion method based on weight assignment. It assigns a weight value to each data point according to its reliability or importance. Then, the data is fused by calculating the weighted average. This method can retain the key information in the data while suppressing noise and interference. In the adaptive Retinex algorithm, the weighted average mechanism is used to fuse the logarithmic brightness components at different scales. By adaptively adjusting the weight gain coefficient according to the local statistical characteristics of the image, the algorithm can enhance the global and local contrast of the image.

[0076] In the specific implementation, the adaptive Retinex algorithm can automatically adjust parameters according to the actual illumination and noise conditions of the image to ensure the enhancement effect under different illumination conditions. This makes the algorithm more applicable in complex environments and can effectively cope with environments with uneven illumination, dust and high humidity such as underground coal mines. Compared with the traditional fixed parameter method, the adaptive Retinex algorithm can better adapt to different application scenarios and improve the effect of image enhancement. In addition, through multi-scale decomposition and logarithmic transformation, the adaptive Retinex algorithm can capture global and local features in the image and enhance the dynamic range and detail performance of the image. This helps to improve the accuracy and reliability of damage detection, especially for the identification of detail features such as cracks, wear and tear. Compared with the single-scale method, multi-scale processing can capture the features of the image more comprehensively and improve the accuracy of damage detection. Secondly, the use of Gaussian filters can effectively remove high-frequency noise in the image, retain the main structural information of the image, and improve the clarity and contrast of the image. This helps to reduce the false alarm rate and improve the accuracy of damage detection. Compared with the traditional noise suppression method, the Gaussian filter can more effectively remove noise and improve image quality while retaining image details. Through color space conversion, the adaptive Retinex algorithm can maintain the color information of the image while enhancing the brightness and contrast of the image, ensuring that the visual effect of the image is natural and realistic. Compared with the traditional grayscale enhancement method, color preservation can better preserve the visual effect of the image and improve user satisfaction. In addition, the adaptive Retinex algorithm can reduce the amount of calculation and memory consumption while ensuring image quality through multi-scale decomposition and weighted average fusion, thereby improving the real-time and applicability of the algorithm. Compared with complex image enhancement algorithms, the adaptive Retinex algorithm has obvious advantages in computing efficiency and resource occupation, and is suitable for real-time monitoring and large-scale data processing.

[0077] like Figure 2As shown: the adjustment method of the adaptive threshold histogram equalization method is as follows: step 401, divide the image into a number of overlapping or non-overlapping small windows through a sliding window, and each window is used as an independent processing unit; step 402, calculate the frequency distribution of the pixel grayscale value in each small area through a histogram statistical method to generate a local histogram; step 403, calculate the brightness mean and variance of each small area according to the local histogram of each small area, if the grayscale value distribution in the small area is relatively concentrated, then lower the preset threshold, if the grayscale value distribution in the small area is relatively dispersed, then increase the preset threshold; step 404, based on the dynamically modified threshold, divide the grayscale histogram into a low grayscale area below the threshold and a high grayscale area above the threshold In step 405, during the equalization process, if the grayscale value variation range in a certain area is small, part of the grayscale levels in the area are merged into one grayscale level by grayscale merging; if the grayscale value variation range in a certain area is large, the grayscale levels in the area are further subdivided by grayscale subdivision; step 406, the equalization results of each small area are spliced ​​together by image splicing to generate a final enhanced image;

[0078] Among them, sliding window division refers to dividing the entire image into multiple small areas, which can be overlapping or non-overlapping. Each small area is used as an independent processing unit, so that different parts of the image can be processed more finely, especially those areas with different lighting conditions. In this way, the processing results of each small area can be ensured to be more accurate, thereby improving the quality of the overall image. For example, in areas with uneven lighting, a smaller window can be selected to obtain a more detailed processing effect. This division method is helpful for subsequent local histogram statistics and equalization processing, making the processing results of each small area more accurate. Local histogram statistics refers to the statistics of the pixel grayscale values ​​in each small window to generate a local histogram of each small window. The histogram shows the frequency of occurrence of each grayscale value and reflects the distribution of pixels in the small window. Through histogram statistics, the brightness and contrast information in each small window can be understood, which provides a basis for subsequent threshold adjustment and histogram equalization. In practical applications, these histograms can help us understand the brightness and contrast in each small window, providing important reference information for subsequent processing steps. For example, if the grayscale value distribution in a small window is relatively concentrated, it means that the contrast of the area is low and needs to be enhanced. In practical applications, the threshold is adjusted dynamically by calculating the brightness mean and variance of each small window. For example, if the grayscale value distribution in a small window is relatively concentrated, it means that the contrast of the area is low and the threshold needs to be lowered to enhance the contrast; conversely, if the grayscale value distribution is relatively dispersed, it means that the contrast of the area is high and the threshold needs to be increased to avoid over-enhancement. Through this dynamic adjustment, it can ensure that the processing effect of each small window is more reasonable and improve the overall quality of the image. Local histogram equalization refers to the equalization of the histogram in each small window to enhance the details of the image by expanding the local contrast. Specifically, it is to redistribute the grayscale values ​​in each small window so that the frequency of each grayscale value is more uniform, thereby improving the local contrast and detail performance. Grayscale merging and subdivision are adjusted according to the grayscale value variation range in each small window during the equalization process. If the grayscale value variation range in a small window is small, it means that the contrast of the area is low, and the amount of calculation can be reduced by merging some gray levels; if the grayscale value variation range is large, it means that the contrast of the area is high, and the detail performance can be improved by subdividing the gray levels.

[0079] Compared with the existing technology, the adaptive threshold histogram equalization method can more accurately capture the detailed features in the image through local processing and dynamic threshold adjustment, while suppressing noise and interference, thereby improving the image enhancement effect. Secondly, this method can automatically adjust the parameters according to the actual illumination and noise conditions of the image, ensuring the image enhancement effect under different illumination conditions and improving the robustness and stability of the system. In addition, this method further optimizes the distribution of grayscale values ​​through grayscale merging or subdivision, making the grayscale value of the image more uniform and continuous, which is helpful for subsequent damage feature extraction and classification.

[0080] In the practical application of the conveyor belt damage intelligent monitoring system, the hardware working environment of the adaptive Retinex algorithm and the adaptive threshold histogram equalization method includes:

[0081] Computing platform: A high-performance computer or server equipped with a multi-core CPU (such as Intel Xeon E5 series) and high-speed memory (at least 16GB DDR4 RAM) for fast processing of large amounts of image data.

[0082] Graphics Processing Unit: A dedicated GPU (such as NVIDIA Tesla V100) to accelerate the execution of image processing algorithms, especially for complex image processing tasks such as adaptive Retinex algorithm and histogram equalization.

[0083] Storage devices: Large-capacity solid-state drives (SSDs) or RAID arrays to store large amounts of raw and processed image data, ensuring fast and reliable data access.

[0084] Image acquisition device: High-resolution industrial camera (such as Basler ace series) is used to capture high-quality images of the conveyor belt surface, ensuring image clarity and rich details.

[0085] Network equipment: Gigabit Ethernet switch, used to connect various components to ensure high speed and stability of data transmission.

[0086] Power supply: Uninterruptible power supply (UPS) ensures that the system can still operate normally in the event of power failure and protects data security.

[0087] Based on the above hardware environment, a comparative experiment was designed to verify the effectiveness of the adaptive Retinex algorithm combined with the adaptive threshold histogram equalization method (Group A) compared with the traditional global histogram equalization (Group B) in the conveyor belt damage intelligent monitoring system; in this experiment, the traditional algorithm used by Group B is Global Histogram Equalization (GHE). GHE is a simple image enhancement technology that changes the brightness value distribution of each pixel in the image to make the brightness value distribution of the entire image more uniform, thereby achieving the purpose of enhancing the image contrast. Description of the experimental method

[0088] The experiment first randomly selected 5 conveyor belt samples with different degrees of damage from the actual working environment, and took 10 images for each sample. Then, these images were processed using the methods of Group A and Group B respectively, and each method was repeated 5 times for each sample to evaluate the stability and consistency of the method. Finally, the effects of the two methods were analyzed by comparing the contrast, signal-to-noise ratio, and damage recognition accuracy of the images before and after processing. The experimental record table is shown in Table 1:

[0089] Table 1 Comparative experimental records

[0090]

[0091]

[0092] The experimental results show that the adaptive Retinex algorithm combined with the adaptive threshold histogram equalization method (Group A) is significantly superior to the traditional global histogram equalization (Group B) in terms of average contrast improvement, average signal-to-noise ratio improvement, and damage recognition accuracy. The Group A method can better adapt to different lighting conditions, effectively improve the local contrast of the image, and reduce noise amplification, thereby improving the accuracy and reliability of damage detection. This shows that the Group A method is more suitable for use in the intelligent monitoring system for conveyor belt damage, and can provide more accurate damage detection support for actual production.

[0093] A multi-scale feature extraction module is used to capture the detailed features of conveyor belt damage based on the output image of the image enhancement module through a pyramid pooling convolution structure, and the damage includes cracks, wear and tear; the pyramid pooling convolution structure includes an input layer, a depth-separable convolution layer, a multi-scale pooling layer, a feature fusion layer and an output layer; the working steps of the pyramid pooling convolution structure include: step 501, receiving the output image of the image enhancement module through the input layer; step 502, performing preliminary feature extraction on the input image through the depth-separable convolution layer; the depth-separable convolution separates the features through depth-convolution and point-by-point convolution operations; The spatial information and channel information of the image are processed independently; step 503, multi-scale pooling operation is performed on the initially extracted feature map through the multi-scale pooling layer; the multi-scale pooling layer uses pooling windows of different sizes to downsample the feature map to obtain feature representations at different scales; step 504, multi-scale features output by the multi-scale pooling layer are fused through the feature fusion layer; the feature fusion layer uses an element-by-element addition method to integrate feature maps of different scales into a unified feature representation; step 505, the feature map integrated by the feature fusion layer is output as the final feature representation through the output layer.

[0094] Among them, the extraction method of the depth separable convolution layer for preliminary feature extraction of the input image is: first, an independent convolution kernel is assigned to each channel of the input image through spatial convolution, the spatial features in the channel are captured, and an intermediate feature map is generated; then, the intermediate feature map generated by the spatial convolution is subjected to inter-channel feature fusion through a point-by-point convolution mechanism, and the point-by-point convolution uses a 1x1 convolution kernel to perform weighted summation of the feature values ​​of all channels at each position through linear transformation, and adds a bias term; in the process of depth convolution and point-by-point convolution, each channel of the feature map is normalized through a batch normalization mechanism to normalize the distribution of the feature map and reduce the influence of internal covariate shift; on the feature map after point-by-point convolution, each element of the feature map is operated element-by-element through the nonlinear activation function ReLU, if the element value is greater than 0, it remains unchanged; if the element value is less than or equal to 0, it is set to 0;

[0095] In the specific implementation, first, the depthwise separable convolution layer uses independent processing of depthwise convolution and pointwise convolution to effectively reduce the amount of calculation and improve the efficiency of feature extraction. The depthwise convolution focuses on the spatial feature extraction of each input channel, while the pointwise convolution is responsible for combining the features of different channels to generate a new feature map. This separation processing method enables the network to capture the spatial and channel information of the image more efficiently. The multi-scale pooling layer downsamples the feature map by using pooling windows of different sizes to obtain feature representations at different scales. This multi-scale feature extraction method can capture the detailed features of conveyor belt damage (such as cracks, wear, and tear) at different scales, improving the richness and robustness of the features. The feature fusion layer uses the element-by-element addition method to integrate feature maps of different scales into a unified feature representation. This fusion method not only retains the feature information at each scale, but also realizes the complementarity and enhancement of features, providing more accurate and comprehensive feature input for subsequent damage identification and analysis.

[0096] In practical applications, the pyramid pooling convolution structure can significantly improve the performance of the intelligent monitoring system for conveyor belt damage. First, through the efficient feature extraction of the deep separable convolution layer, the system can quickly capture the key information in the conveyor belt image. Subsequently, the multi-scale pooling layer performs multi-scale downsampling on the feature map to capture the detailed features of damage at different scales. These features are integrated and enhanced in the feature fusion layer, and finally output as a unified feature representation. This working method enables the system to more accurately identify and analyze the damage on the conveyor belt. Since the pyramid pooling convolution structure can capture feature information at multiple scales, the system has a good recognition ability for damage of different sizes and shapes. In addition, due to the use of the deep separable convolution layer, the computational efficiency of the system has also been significantly improved, making real-time monitoring and rapid response possible.

[0097] Compared with the existing technology, the pyramid pooling convolution structure can capture richer damage detail features through multi-scale feature extraction and fusion, improving the accuracy and robustness of recognition. Secondly, the use of the depth-separable convolution layer significantly improves the computational efficiency of the system and reduces the computational cost and power consumption. In addition, the pyramid pooling convolution structure also has good generalization ability and can adapt to the needs of conveyor belt damage monitoring under different environments and conditions.

[0098] In the practical application of the conveyor belt damage intelligent monitoring system and its multi-dimensional data analysis method, the hardware working environment of the pyramid pooling convolution structure mainly includes:

[0099] Computing platform: NVIDIA Jetson TX2 embedded computing platform with high-performance GPU and CPU, capable of supporting real-time operation of complex deep learning algorithms.

[0100] Camera: Use an industrial-grade high-definition camera with a resolution of 1080p and a frame rate of 30fps, which can capture clear conveyor belt images.

[0101] Storage device: Equipped with high-speed SSD solid-state hard disk to store large amounts of image data and model parameters.

[0102] Power supply: Use a stable DC power adapter to ensure that the system will not be interrupted by power problems during long-term operation.

[0103] Network communication: Data is transmitted with the monitoring center via wired or wireless networks (such as Wi-Fi) to achieve remote monitoring and management.

[0104] Based on the above hardware environment, a set of comparative experiments was designed to verify the effectiveness and superiority of the pyramid pooling convolution structure in the conveyor belt damage intelligent monitoring system. The experiments were divided into two groups: Group A used the pyramid pooling convolution structure, and Group B used the traditional single-scale feature extraction method (such as SIFT). Both groups of experiments were conducted in the same hardware environment to ensure the consistency of experimental conditions. Each group of experiments was conducted five times, and the damage detection accuracy and processing time of each experiment were recorded. The experimental results are shown in Table 2:

[0105] Table 2 Experimental record table

[0106]

[0107] Through comparative experiments, it can be seen that the pyramid pooling convolution structure (Group A) is superior to the traditional SIFT algorithm (Group B) in terms of accuracy and processing time of conveyor belt damage detection. Specifically, the average accuracy of Group A is 92.8%, while the average accuracy of Group B is 85.6%; the average processing time of Group A is 120 milliseconds, while the average processing time of Group B is 249 milliseconds. This shows that the pyramid pooling convolution structure not only significantly improves the detection accuracy, but also shows obvious advantages in processing speed. Therefore, the pyramid pooling convolution structure has a positive and beneficial effect in the conveyor belt damage intelligent monitoring system, and can effectively improve the performance and practicality of the system.

[0108] A damage type classification and positioning module is used for classifying and positioning damage based on the output data of the multi-scale feature extraction module; the damage is classified and positioned by a target detection algorithm based on an attention mechanism; the target detection algorithm based on an attention mechanism calculates the weight of each area in the image through an attention mechanism, and classifies the extracted feature vectors through a multi-classification Softmax layer to determine the damage type; Figure 4As shown: The working method of the target detection algorithm based on the attention mechanism is as follows: Step 701, pre-processing the real-time conveyor belt image data collected by the data acquisition module through image enhancement, denoising and normalization processing; Step 702, performing multi-layer convolution operations on the pre-processed input image through a convolutional neural network to extract multi-level features of the image; Step 703, calculating the weight of each area in the image through a multi-scale attention mechanism to generate an attention map, and the calculation formula is:

[0109]

[0110] In formula (1), n ​​represents the feature map output by the feature extraction network; G i (x) and G i (m) represents the global average pooling and global maximum pooling operations, which are used to capture global information; X i and M i Represent local average pooling and local maximum pooling operations respectively; i represents different pooling window sizes, which are used to capture multi-scale features; θ represents the inter-channel relationship vector, which is used to capture the correlation between channels; represents the Sigmoid activation function, which is used to limit the value of the attention map to between 0 and 1; P i represents the concatenation operation, which is used to integrate information between different scales and channels; μ cRepresents a convolution operation, which is used to further extract features and generate an attention map; step 704, based on the attention map output by step 703, multiply the attention weight vector by the eigenvalue of the corresponding position of the feature map through an element-by-element multiplication operation to generate a weighted feature map; step 705, based on the weighted feature map, classify the extracted feature vector through a multi-classification Softmax layer to determine the damage type; the multi-classification Softmax layer maps the feature vector of the weighted feature map to a vector space of a fixed length through a fully connected layer to generate a classification feature vector; and normalizes the classification feature vector through an exponential function and a normalization operation to generate a probability distribution of each damage category; step 706, during the training process of the algorithm, the cross entropy loss function is used to calculate the probability distribution of each damage category. The difference between the predicted result and the true label is calculated, and the back-propagation algorithm is applied to pass the error back to each parameter in the network layer by layer; the attention mechanism is also used to guide the region proposal network to generate candidate regions to improve the accuracy of damage location. The specific working method is as follows: first, the weight of each region in the image is calculated through the attention mechanism, and the image features are weighted; based on the feature map after the weighted processing of the attention mechanism, a series of candidate regions are generated on the feature map through a number of predefined anchor boxes sliding on the feature map, and in the process of sliding the anchor box, the foreground / background classification and bounding box regression prediction are performed on each anchor box through the convolution layer; in the process of generating candidate regions, a multi-scale feature map is also constructed through the feature pyramid network to strengthen the key feature association across scales;

[0111] In the specific implementation, the attention mechanism captures the global context information of the image through global average pooling and global maximum pooling, and focuses on the local details of the image by combining local average pooling and local maximum pooling. This dual consideration of global and local enables the algorithm to understand the image content more comprehensively. The introduction of the inter-channel relationship vector further enhances the algorithm's ability to identify key features. By calculating the correlation between different channels, the algorithm can automatically adjust the weight of each channel and optimize the feature extraction process. The Sigmoid activation function, as a nonlinear mapping tool for the attention map, limits the attention value between 0 and 1, ensuring the stability and interpretability of the algorithm during the processing. In addition, the feature extraction network, especially the application of convolutional neural network (CNN), extracts multi-level features of the image through multi-layer convolution operations. These features cover rich visual information from edges, textures to shapes, providing a solid foundation for subsequent damage classification and location. The construction of feature pyramid network (FPN) enhances the algorithm's ability to identify damage areas of different scales through cross-scale feature fusion, and improves the robustness and accuracy of recognition.

[0112] In practical applications, the target detection algorithm based on the attention mechanism first preprocesses the real-time conveyor belt image data collected by the data acquisition module through image enhancement, denoising and normalization. Then, the convolutional neural network is used to perform multi-layer convolution operations on the preprocessed image to extract the multi-level features of the image. Next, the weight of each area in the image is calculated through the attention mechanism, an attention map is generated, and the feature map is weighted. Based on the weighted feature map, the algorithm classifies the extracted feature vectors through the multi-classification Softmax layer to determine the damage type. At the same time, the region proposal network generates candidate regions based on the weighted feature map, and constructs a multi-scale feature map through the feature pyramid network to further improve the accuracy of damage location. Finally, the algorithm outputs information such as damage type, location, and probability distribution for subsequent multi-dimensional data analysis;

[0113] Compared with the existing technology, the algorithm can automatically focus on the damaged area and reduce the interference of background information by weighted processing of the image through the attention mechanism. At the same time, the feature extraction and fusion capabilities of CNN and FPN are combined, so that the algorithm can extract richer image features and improve the accuracy of recognition. Secondly, the combination of attention mechanism, CNN and FPN enables the algorithm to flexibly adapt to different types of conveyor belts and damage conditions, and maintain high recognition accuracy and stability even in complex environments. In addition, by optimizing the structure and parameters of the algorithm, the algorithm can improve the processing speed while ensuring recognition accuracy and realize real-time monitoring.

[0114] In practical applications, the method of generating candidate regions based on the attention mechanism region proposal network can significantly improve the accuracy of damage localization. The specific steps are as follows:

[0115] 1. Image preprocessing: First, the collected conveyor belt images are preprocessed, including image enhancement, denoising and normalization, to ensure that the image quality meets the requirements of subsequent processing.

[0116] 2. Feature extraction: The preprocessed image is subjected to feature extraction through a multi-layer convolutional neural network to generate multi-level feature maps. These feature maps contain local and global information of the image, providing a basis for the subsequent attention mechanism and region proposal network.

[0117] 3. Attention map generation: Generate attention maps through multi-scale pooling and convolution operations. The specific steps are:

[0118] Global average pooling and global maximum pooling operations capture global information.

[0119] Local average pooling and local maximum pooling operations capture local information.

[0120] The correlation between channels is captured by modeling the relationship between channels.

[0121] The Sigmoid activation function is used to clamp the values ​​of the attention map between 0 and 1.

[0122] The information between different scales and channels is integrated through the splicing operation.

[0123] The convolution operation is used to further extract features and generate attention maps.

[0124] 4. Feature weighting processing: Multiply the generated attention map and the feature map element by element to generate a weighted feature map. The weighted feature map can highlight the key areas in the image, suppress irrelevant areas, and improve the robustness and distinguishing ability of the features.

[0125] 5. Candidate region generation: Based on the weighted feature map, candidate regions are generated through the region proposal network. The specific method is as follows: First, slide the predefined anchor box on the weighted feature map to generate a series of candidate regions. Then, the convolution layer performs foreground / background classification and bounding box regression prediction on each anchor box. Foreground / background classification is used to determine whether the anchor box contains damage, and bounding box regression is used to accurately locate the location of the damage.

[0126] 6. Multi-scale feature map construction: Multi-scale feature maps are constructed through feature pyramid networks to strengthen the key feature associations across scales. The specific method is: first, the top-down path is used to pass the high-level semantic information to the low-level. Then, a horizontal connection is made to combine the low-level spatial information with the high-level semantic information. Finally, a multi-scale feature map is generated to capture damage characteristics at different scales.

[0127] In the practical application of the intelligent monitoring system for conveyor belt damage and its multi-dimensional data analysis method, the hardware working environment of the target detection algorithm based on the attention mechanism includes: Computing platform: Using the NVIDIA Jetson Xavier NX embedded computing platform, it has a high-performance GPU and multi-core CPU, which can support the real-time operation of complex deep learning algorithms. Camera: Use an industrial-grade high-definition camera with a resolution of 4K and a frame rate of 30fps, which can capture high-definition conveyor belt images. Storage device: Equipped with a high-speed NVMe SSD solid-state drive to store large amounts of image data and model parameters. Power supply: Use a stable DC power adapter to ensure that the system will not be interrupted due to power problems during long-term operation. Network communication: Data transmission with the monitoring center is carried out through wired or wireless networks (such as 5G) to achieve remote monitoring and management.

[0128] A set of comparative experiments was designed to verify the effectiveness and superiority of the target detection algorithm based on the attention mechanism in the intelligent monitoring system for conveyor belt damage. The experiments were divided into two groups: Group A used the target detection algorithm based on the attention mechanism, and Group B used the traditional single-scale feature extraction method (Faster R-CNN). Both groups of experiments were carried out under the same hardware environment to ensure the consistency of the experimental conditions. Each group of experiments was conducted five times, and the damage detection accuracy and processing time of each experiment were recorded. Faster R-CNN is a classic two-stage target detection algorithm that generates candidate regions through a region proposal network (RPN), and then performs feature extraction and classification through a convolutional neural network. Although Faster R-CNN performs well in many scenarios, its performance may be limited when dealing with complex backgrounds and multi-scale features. The experimental data record table is shown in Table 3:

[0129] Table 3 Comparative experimental record table

[0130]

[0131] Through comparative experiments, it can be seen that the target detection algorithm based on the attention mechanism (Group A) is superior to the traditional Faster R-CNN algorithm (Group B) in terms of accuracy and processing time of conveyor belt damage detection. Specifically, the average accuracy of Group A is 94.6%, while the average accuracy of Group B is 88.3%; the average processing time of Group A is 150 milliseconds, while the average processing time of Group B is 202 milliseconds. This shows that the target detection algorithm based on the attention mechanism not only significantly improves the detection accuracy, but also shows obvious advantages in processing speed. Therefore, the target detection algorithm based on the attention mechanism has a positive and beneficial effect in the conveyor belt damage intelligent monitoring system, and can effectively improve the performance and practicality of the system.

[0132] A multi-dimensional data fusion and analysis module is used to integrate the damage development rate, damage location distribution, material stress and material temperature data through a graph neural network and a long short-term memory network based on the collected data of the data collection module and the output data of the multi-scale feature extraction module, perform fusion analysis, and output a damage risk assessment report, including damage type, degree, and development trend; the working method of the multi-dimensional data fusion and analysis module is as follows: using a graph neural network to process damage location distribution information, construct a topological structure of conveyor belt damage, capture the spatial correlation between damages, and process the damage development rate through a long short-term memory network to capture the change trend of damage over time; finally, the outputs of the graph neural network and the long short-term memory network are fused through a fully connected layer to generate a damage risk assessment report, including damage type, degree, and development trend;

[0133] Furthermore, a multi-dimensional data analysis method for a conveyor belt damage intelligent monitoring system includes the following steps: S1, processing damage location distribution information through a graph neural network; the graph neural network constructs a topological structure of conveyor belt damage through nodes and edges, each node represents a damage point, and the edge represents the spatial relationship between the damage points; the graph neural network also aggregates the features of each node through a graph convolution operation, so that the feature representation of each node not only includes its own attributes, but also integrates the influence of adjacent nodes to form a topological structure diagram of conveyor belt damage; the formula expression for constructing the topological structure diagram of conveyor belt damage by the graph neural network is:

[0134]

[0135] S2. In formula (2), represents the feature representation of the kth node in the l+1th layer; σ represents the ReLU nonlinear activation function; N(k) represents the neighbor set of node k; c kj represents the normalization coefficient, which is used to balance the contribution of different neighbor nodes; W l Represents the weight matrix of the lth layer, which is used to learn the relationship between node features; represents the feature representation of the kth neighbor node in the lth layer; b (l) is the bias term of the lth layer, which is used to adjust the baseline of feature representation; S3, the damage development rate is processed through the long short-term memory network to capture the change trend of the damage over time; the long short-term memory network captures the change trend of the damage over time through the input gate, forget gate, output gate and cell state, and dynamically adjusts the memory degree of historical information; S4, after completing the spatial association capture and time trend analysis, the spatial features output by the graph neural network and the time features output by the long short-term memory network are integrated through the weight matrix and nonlinear activation function of the fully connected layer to generate a comprehensive feature vector z, S5, based on the comprehensive feature vector z output by the fully connected layer, the comprehensive feature vector is normalized by the Softmax function to generate the probability distribution of each damage category; the formula expression of the Softmax function is:

[0136]

[0137] In formula (3), TAU k is the probability of the kth category; z k represents the kth element of the comprehensive feature vector; C is the total number of categories; z j Represents the j-th element of the comprehensive feature vector.

[0138] In specific implementation, in practical applications, the multi-dimensional data analysis method works as follows: first, the damage data of the conveyor belt is obtained through the data acquisition module and the multi-scale feature extraction module; then, the graph neural network is used to process the damage location distribution information and construct the topological structure of the conveyor belt damage; then, the damage development rate is processed through the long short-term memory network to capture the change trend of the damage over time; finally, these two types of information are fused through the fully connected layer to generate a damage risk assessment report.

[0139] In the graph neural network, each node not only represents a damage point, but also connects with adjacent nodes through edges to form a topological structure of conveyor belt damage. This structure is not static, but is constantly updated with the input of new data. The feature representation of each node continuously aggregates the information of adjacent nodes through graph convolution operations. This dynamic update mechanism enables the network to capture the complex spatial associations between damages. The core of the graph convolution operation is to update the feature representation of the node itself by aggregating the features of adjacent nodes. This operation not only considers the properties of the node itself, but also incorporates the influence of adjacent nodes, thereby realizing the perception of local topological structures. Through multi-layer graph convolution operations, the network can capture the associations between nodes at a longer distance and form a global topological structure representation. The long short-term memory network finely regulates the flow of information through three gating mechanisms: input gate, forget gate, and output gate. The input gate determines which new information can enter the cell state, the forget gate determines which old information needs to be forgotten, and the output gate determines which information can be output to the next layer. This fine-tuning mechanism enables the network to dynamically adjust the degree of memory of historical information, thereby achieving accurate processing of time series data. Cell state is the core component of the long short-term memory network, which is responsible for storing the long-term memory of the network. By updating and transmitting the cell state, the network can maintain the memory of past information and use it for decision-making in the current task when needed. This long-term memory mechanism enables the network to capture the changing trend of damage over time, providing strong support for damage risk assessment.

[0140] In the application implementation, the multi-dimensional data analysis method adopts a data-driven approach in actual application, and continuously updates the parameters and status of the model according to the newly collected data. This dynamic update mechanism enables the system to adapt to changes in the conveyor belt environment and maintain stable performance. When processing real-time data, the system needs to increase the processing speed as much as possible while ensuring accuracy. To achieve this goal, the system uses efficient algorithms and hardware acceleration technologies, such as GPU parallel computing, to ensure a balance between real-time and accuracy. In addition, in order to facilitate users to view and use damage risk assessment reports, the system provides a user-friendly interactive interface. Users can intuitively understand the damage status, development trends, potential risk points and other information of the conveyor belt through the interface, providing timely decision-making support for industrial safety production.

[0141] The beneficial effects of this working method are mainly reflected in the following aspects: First, the accuracy of damage identification is improved. By integrating multi-dimensional data, the system can have a more comprehensive understanding of the damage status of the conveyor belt; second, the robustness of the system is enhanced. Even in the face of complex and changeable conveyor belt environments, the system can maintain stable performance; third, the timeliness of damage risk assessment is improved. The system can update the damage risk assessment report in real time, providing timely warning and decision-making support for industrial safety production;

[0142] Compared with the existing technology, the multi-dimensional data analysis method has the following advantages and characteristics: First, higher accuracy. By integrating and deeply analyzing multi-dimensional data, the system can more accurately evaluate the damage status of the conveyor belt; second, stronger robustness. The system can adapt to the complex and changing conveyor belt environment and maintain stable performance; third, more comprehensive damage risk assessment. The system can not only provide basic information such as damage type and degree, but also predict the development trend of damage, providing more comprehensive decision support for industrial safety production; fourth, higher timeliness. The system can update the damage risk assessment report in real time, providing timely warning and response for industrial safety production. These advantages and characteristics make the multi-dimensional data analysis method have broad application prospects and important practical value in the intelligent monitoring system of conveyor belt damage.

[0143] In specific implementation, the hardware working environment of the multi-dimensional data analysis method in the conveyor belt damage intelligent monitoring system mainly includes: High-performance computing server: used to run complex graph neural network and long short-term memory network models to ensure real-time processing and analysis of a large amount of collected conveyor belt damage data. Data acquisition equipment: including sensors, cameras, etc., used to monitor the damage of the conveyor belt in real time and transmit the data to the computing server for processing. Data storage devices: such as hard disk arrays or cloud storage services, used to store historical data and model parameters for long-term analysis and model optimization. Visualization terminal: used to display damage risk assessment reports and real-time monitoring results to facilitate user decision-making and response.

[0144] Based on the above hardware environment, the multi-dimensional data analysis method (Group A) and the traditional algorithm (Group B) were used for comparative experiments. The traditional algorithm uses a rule-based method, that is, it judges the damage of the conveyor belt according to preset rules and thresholds. This method is simple and intuitive, but lacks flexibility and accuracy. During the experiment, both groups were conducted under the same experimental conditions, and each group conducted five experiments, recording the damage identification accuracy, processing time, and system stability of each experiment. The experimental data record table is shown in Table 4:

[0145] Table 4. Multi-dimensional data analysis method experimental comparison data record table

[0146]

[0147]

[0148] Through comparative experiments, it can be seen that the multi-dimensional data analysis method (Group A) is superior to the traditional algorithm (Group B) in terms of damage identification accuracy, processing time and system stability. The damage identification accuracy of Group A averaged 94.6%, while that of Group B was only 83.6%; the processing time of Group A averaged only 2.4 seconds, much lower than the 3.7 seconds of Group B; at the same time, the system stability score of Group A was 5 points, while that of Group B fluctuated. Therefore, the multi-dimensional data analysis method has higher accuracy and efficiency in the intelligent monitoring system for conveyor belt damage, and can better meet the needs of industrial safety production.

[0149] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. An intelligent conveyor belt damage monitoring system; characterized in that: The system comprises: The data acquisition module is used to collect various types of data of the conveyor belt through the sensor network, including the working status of the conveyor belt, the conveyor belt image, and the environmental parameters; An image enhancement module is used to judge the image clarity through standard deviation and entropy based on the conveyor belt image data collected by the data acquisition module. If the image clarity is lower than a preset threshold, the image is subjected to dynamic range compression, noise suppression and detail enhancement through an adaptive Retinex algorithm and an adaptive threshold histogram equalization method, and the enhanced image is output to a multi-scale feature extraction module; if the image clarity is higher than or equal to the preset threshold, the image is directly output to the multi-scale feature extraction module; A multi-scale feature extraction module, which is used to capture the detailed features of conveyor belt damage based on the output image of the image enhancement module through a pyramid pooling convolution structure, where the damage includes cracks, wear and tear; the pyramid pooling convolution structure includes an input layer, a depth-separable convolution layer, a multi-scale pooling layer, a feature fusion layer and an output layer; A damage type classification and positioning module is used for classifying and positioning damage based on the output data of the multi-scale feature extraction module through an attention-based target detection algorithm; the attention-based target detection algorithm calculates the weight of each area in the image through the attention mechanism, and classifies the extracted feature vectors through a multi-classification Softmax layer to determine the damage type; The multi-dimensional data fusion and analysis module is used to integrate the damage development rate, damage location distribution, material stress and material temperature data based on the collected data of the data acquisition module and the output data of the multi-scale feature extraction module through the graph neural network and the long short-term memory network, perform fusion analysis, and output a damage risk assessment report, including damage type, degree, and development trend.

2. The intelligent conveyor belt damage monitoring system according to claim 1 is characterized in that: The working method of the image enhancement module for dynamic range compression, noise suppression and detail enhancement of the image is as follows: first, the brightness perception characteristics of the human visual system are simulated by the adaptive Retinex algorithm, and the global and local contrast of the image is enhanced, while the influence of uneven illumination is suppressed; based on the enhanced image data, the histogram distribution of each small area is dynamically adjusted by the adaptive threshold histogram equalization method, and the local dark or bright area is adjusted; during the image processing process, the adaptive Retinex algorithm and the adaptive threshold histogram equalization method automatically adjust the parameters according to the actual illumination and noise conditions of the image, so as to ensure the enhancement effect of the image under different illumination conditions.

3. The intelligent conveyor belt damage monitoring system according to claim 2 is characterized in that: The working principle of the adaptive Retinex algorithm is as follows: based on the conveyor belt image data of the data acquisition module, the standard deviation and entropy of the image are calculated through the image processing library and mathematical statistics methods to quantify the clarity of the image; if the image clarity is lower than the preset threshold, the processing flow of the adaptive Retinex algorithm is triggered; if the image clarity meets or exceeds the preset threshold, the processing flow of the adaptive Retinex algorithm is skipped; when the image clarity is lower than the preset threshold, the image is decomposed into brightness components at multiple scales through the color restoration multi-scale method; the color restoration multi-scale method is based on the Gaussian pyramid model, and the image is smoothed by Gaussian filters of different scales to obtain brightness information at different scales; then the brightness components at each scale are smoothed. A logarithmic transformation is performed. After the transformation, the adaptive Retinex algorithm fuses the logarithmic brightness components at each scale through a weighted average mechanism. During the fusion process, the contribution of different scales to image details is considered, and different weight values ​​are given to the brightness components at different scales through an adaptive gain adjustment mechanism to achieve enhancement of the global and local contrast of the image; wherein the weight value realizes adaptive gain adjustment based on the local statistical characteristics of the image; the weight gain coefficient is automatically adjusted according to the local brightness and contrast information of each pixel in the image based on the local statistical characteristics of the image; after the gain adjustment, the adaptive Retinex algorithm recombines the enhanced brightness component with the color information of the original image through color space conversion to generate an enhanced image.

4. The intelligent conveyor belt damage monitoring system according to claim 2 is characterized in that: The adjustment method of the adaptive threshold histogram equalization method is: Step 401: Divide the image into a number of overlapping or non-overlapping small windows by sliding the window, and each window serves as an independent processing unit; Step 402: Calculate the frequency distribution of the pixel grayscale values ​​in each small area by using a histogram statistical method to generate a local histogram; Step 403: Calculate the brightness mean and variance of each small area according to the local histogram of each small area. If the gray value distribution in the small area is relatively concentrated, lower the preset threshold value. If the gray value distribution in the small area is relatively dispersed, increase the preset threshold value. Step 404: based on the dynamically modified threshold, the grayscale histogram is divided into a low grayscale area below the threshold and a high grayscale area above the threshold, and the low grayscale area below the threshold and the high grayscale area above the threshold are respectively subjected to histogram equalization processing by a local histogram equalization method; the local histogram equalization method maps the local histogram of each small area by local contrast stretching and grayscale value redistribution to generate a new grayscale value distribution; Step 405: During the equalization process, if the grayscale value variation range in a certain area is small, grayscales in the area are merged into one grayscale by grayscale merging; if the grayscale value variation range in a certain area is large, grayscales in the area are further subdivided by grayscale subdivision; Step 406: stitch together the equalization results of each small area through image stitching to generate a final enhanced image.

5. The intelligent conveyor belt damage monitoring system according to claim 1 is characterized in that: The working steps of the pyramid pooling convolution structure include: Step 501: receiving the output image of the image enhancement module through the input layer; Step 502: Perform preliminary feature extraction on the input image through the depthwise separable convolution layer; the depthwise separable convolution independently processes the spatial information and channel information of the image through depthwise convolution and pointwise convolution operations respectively; Step 503: performing a multi-scale pooling operation on the initially extracted feature map through the multi-scale pooling layer; the multi-scale pooling layer uses pooling windows of different sizes to downsample the feature map to obtain feature representations at different scales; Step 504: The multi-scale features output by the multi-scale pooling layer are fused through the feature fusion layer; the feature fusion layer integrates the feature maps of different scales into a unified feature representation by element-by-element addition; Step 505: output the feature map integrated by the feature fusion layer as the final feature representation through the output layer.

6. The intelligent conveyor belt damage monitoring system according to claim 5, characterized in that: The extraction method of the depth-separable convolution layer for preliminary feature extraction of the input image is as follows: first, an independent convolution kernel is assigned to each channel of the input image through spatial convolution to capture the spatial features in the channel and generate an intermediate feature map; then, the intermediate feature map generated by the spatial convolution is subjected to inter-channel feature fusion through a point-by-point convolution mechanism, and the point-by-point convolution uses a 1x1 convolution kernel to perform weighted summation of the feature values ​​of all channels at each position through linear transformation, and adds a bias term; in the process of depth convolution and point-by-point convolution, each channel of the feature map is normalized through a batch normalization mechanism to normalize the distribution of the feature map and reduce the influence of internal covariate shift; on the feature map after point-by-point convolution, each element of the feature map is operated element-by-element through a nonlinear activation function ReLU, and if the element value is greater than 0, it remains unchanged; If the element value is less than or equal to 0, it is set to 0.

7. The intelligent conveyor belt damage monitoring system according to claim 1 is characterized in that: The working method of the target detection algorithm based on the attention mechanism is: Step 701, pre-processing the real-time conveyor belt image data collected by the data collection module through image enhancement, denoising and normalization processing; Step 702: Perform multi-layer convolution operations on the pre-processed input image through a convolutional neural network to extract multi-level features of the image; Step 703: Calculate the weight of each region in the image through a multi-scale attention mechanism to generate an attention map. The calculation formula is: In formula (1), n ​​represents the feature map output by the feature extraction network; G i (x) and G i (m) represents the global average pooling and global maximum pooling operations, which are used to capture global information; X i and M i Represent local average pooling and local maximum pooling operations respectively; i represents different pooling window sizes, which are used to capture multi-scale features; θ represents the inter-channel relationship vector, which is used to capture the correlation between channels; represents the Sigmoid activation function, which is used to limit the value of the attention map to between 0 and 1; P i represents the concatenation operation, which is used to integrate information between different scales and channels; μ c Represents the convolution operation, which is used to further extract features and generate attention maps; Step 704: Based on the attention map outputted in step 703, the attention weight vector is multiplied by the feature value at the corresponding position of the feature map by an element-by-element multiplication operation to generate a weighted feature map; Step 705: Based on the weighted feature graph, the extracted feature vector is classified through a multi-classification Softmax layer to determine the damage type; the multi-classification Softmax layer maps the feature vector of the weighted feature graph to a vector space of a fixed length through a fully connected layer to generate a classification feature vector; and the classification feature vector is normalized through an exponential function and a normalization operation to generate a probability distribution of each damage category; Step 706: During the training process of the algorithm, the difference between the predicted result and the true label is calculated through the cross entropy loss function, and the back propagation algorithm is applied to pass the error back to each parameter in the network layer by layer.

8. The intelligent conveyor belt damage monitoring system according to claim 1 is characterized in that: The attention mechanism is also used to guide the region proposal network to generate candidate regions and improve the accuracy of damage localization. The specific working method is: first, the weight of each region in the image is calculated through the attention mechanism, and the image features are weighted; based on the feature map after weighted processing by the attention mechanism, a series of candidate regions are generated on the feature map by sliding a number of predefined anchor boxes on the feature map. In the process of sliding the anchor box, the foreground / background classification and bounding box regression are predicted for each anchor box through the convolution layer; in the process of generating candidate regions, a multi-scale feature map is also constructed through the feature pyramid network to strengthen the key feature association across scales.

9. The intelligent conveyor belt damage monitoring system according to claim 1, characterized in that: The working method of the multi-dimensional data fusion and analysis module is as follows: using graph neural networks to process damage location distribution information, constructing the topological structure of conveyor belt damage, capturing the spatial correlation between damages, and processing the damage development rate through long short-term memory networks to capture the change trend of damage over time. Finally, the outputs of the graph neural network and the long short-term memory network are fused through a fully connected layer to generate a damage risk assessment report, including damage type, degree, and development trend.

10. A multi-dimensional data analysis method for a conveyor belt damage intelligent monitoring system, characterized in that: An intelligent conveyor belt damage monitoring system as claimed in any one of claims 1 to 9 comprises the following steps: S1. Process the damage location distribution information through a graph neural network; the graph neural network constructs the topological structure of the conveyor belt damage through nodes and edges, each node represents a damage point, and the edge represents the spatial relationship between the damage points; the graph neural network also aggregates the features of each node through a graph convolution operation, so that the feature representation of each node not only contains its own attributes, but also integrates the influence of adjacent nodes to form a topological structure diagram of the conveyor belt damage; the formula expression for constructing the topological structure diagram of the conveyor belt damage by the graph neural network is: In formula (2), represents the feature representation of the kth node in the l+1th layer; σ represents the ReLU nonlinear activation function; N(k) represents the neighbor set of node k; c kj represents the normalization coefficient, which is used to balance the contribution of different neighbor nodes; W l Represents the weight matrix of the lth layer, which is used to learn the relationship between node features; represents the feature representation of the kth neighbor node in the lth layer; b (l) is the bias term of the lth layer, used to adjust the baseline of feature representation; S2. Processing the damage development rate through a long short-term memory network to capture the change trend of the damage over time; the long short-term memory network captures the change trend of the damage over time through an input gate, a forget gate, an output gate and a cell state, and dynamically adjusts the memory degree of historical information; S3. After completing spatial correlation capture and temporal trend analysis, the spatial features output by the graph neural network and the temporal features output by the long short-term memory network are integrated through the weight matrix and nonlinear activation function of the fully connected layer to generate a comprehensive feature vector z. S4. Based on the comprehensive feature vector z output by the fully connected layer, the comprehensive feature vector is normalized by the Softmax function to generate the probability distribution of each damage category; the formula expression of the Softmax function is: In formula (3), TAU k is the probability of the kth category; z k represents the kth element of the comprehensive feature vector; C is the total number of categories; z j Represents the j-th element of the comprehensive feature vector.

Images