Mining belt conveyor belt anomaly detection method based on KAN-YOLO

Abstract

The invention discloses a mining belt conveyor belt anomaly detection method based on KAN-YOLO, and the method comprises the steps: carrying out the shooting through an industrial CCD camera, inputting an image to a PC terminal, and forming an original image data set; and forming a processing data set by adopting binarization, SAPCA denoising, standardization and normalization algorithms. The method has the beneficial effects that the DCGAN is used for expanding the processing data set, so that the problem of sample imbalance in industrial production is solved; and through the trained sample data set, realizing belt anomaly detection of the mining belt conveyor based on YOLOv11. By establishing a KAN network layer model to replace a traditional CNN layer, effective parameter reduction and efficiency improvement are realized; a preset anchor frame is optimized by applying a differential evolution algorithm, so that quick optimization of multiple anchor frames is realized; relative and absolute positions are coded through Transform, and structural blocks are optimized, so that the efficiency is improved.

Description

Technical Field

[0001] The invention relates to a belt anomaly detection method, in particular to a belt anomaly detection method for a mining belt conveyor based on KAN-YOLO, and belongs to the technical field of intelligent detection. Background Art

[0002] Computer Vision is also known as Machine Vision. In this technical field, abnormality detection and recognition is a crucial research topic.

[0003] With the development and progress of the coal industry, underground coal transportation and testing equipment tends to be intelligent and automated. Among them, belt transportation has become the main mode of transportation. During coal transportation, abnormal conditions such as excessive coal accumulation, anchor insertion, and belt tearing can cause damage to equipment and pose a safety hazard.

[0004] Due to the limitations of science and technology, in the traditional industrial production and manufacturing process, manual detection is still the main method for detecting product abnormal information. Due to the limitations of workers, this method not only has large fluctuations in detection accuracy, but also has slow detection speed and low efficiency, resulting in inaccurate detection results; at the same time, in the prior art, such as a method, device and computer-readable storage medium for detecting abnormalities in a conveyor belt disclosed in announcement number CN116758400A, a neural network model is pre-trained, and the pre-trained neural network model is fine-tuned for the model of conveyor belt abnormality detection, and a feature library of conveyor belt abnormalities is constructed, and an adaptive threshold search for abnormality scores is performed according to the feature library to obtain the optimal F1 threshold, so that the abnormal score of the belt image to be tested is compared with the optimal F1 threshold to determine whether the belt image to be tested is abnormal. Combining deep learning and transfer learning and applying them to conveyor belt abnormality detection improves the pertinence and robustness of the model in extracting belt abnormality features, but the existing visual detection method has poor real-time performance and cannot quickly and accurately identify abnormal conditions of mining belt conveyors. Therefore, new technical methods are urgently needed to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a mining belt conveyor belt anomaly detection method based on KAN-YOLO in order to solve at least one of the above technical problems.

[0006] The present invention achieves the above-mentioned purpose through the following technical scheme: a method for detecting belt anomalies of a mining belt conveyor based on KAN-YOLO, the method for detecting belt anomalies of a mining belt conveyor comprising the following steps:

[0007] S1. Data input, obtaining the surface image information of the mining belt conveyor, shooting it with an industrial CCD camera, inputting the image to the PC to form an original image data set;

[0008] S2. Data preprocessing, including:

[0009] S21, image binarization, using an industrial CCD camera to capture the original image information of the mining belt conveyor; selecting a threshold and performing binarization processing on the original image;

[0010] S22, filter noise, use SAPCA to reduce noise, obtain the basic estimated image through forward and inverse transformation and calculation of similar blocks, use Euclidean distance to measure the similarity of the basic estimated blocks, and fully restore the details of the original image;

[0011] S23, standardization and normalization, standardizing and normalizing the denoised binary image;

[0012] S24, data enhancement strategy, using supervised data enhancement to expand on the existing data; by performing corresponding geometric transformations on the denoised binary image, including flipping, rotation, shifting, cropping, deformation and scaling operations, the content of the image itself is changed;

[0013] S3. Network training, including:

[0014] S31, sample expansion, using the DCGAN network architecture to train the model and generate a sample-balanced mining belt conveyor data set;

[0015] S32, model training, using KAN-YOLO neural network to train belt anomaly recognition model;

[0016] S4, abnormality recognition, use the model to identify abnormal conditions of image data acquired in real time; according to the identified abnormal conditions, the mining belt conveyor is adjusted in real time to complete the decision output.

[0017] As a further solution of the present invention: in step S2, the noise filtering process for image preprocessing specifically includes the following steps:

[0018] S221, using 8-direction LPA-ICI to obtain an adaptive shape neighborhood centered on the current pixel; the neighborhood is enclosed in a non-adaptive square block, called a reference block, and the number of pixels in the neighborhood is represented by N el ;

[0019] S222, using block matching to find a block with a high degree of similarity to the reference block, and using the shape information obtained in step S221 to extract an adaptive shape neighborhood from the matching block; the number of matching blocks is N gr express;

[0020] S223, determine the transformation to be applied to the adaptive shape neighborhood, depending on N gr / N el Is it greater than or less than a fixed threshold τ?

[0021] If N gr / N el ≥τ, it is considered that a sufficient number of mutually similar neighborhoods have been found to reliably estimate the second-order matrix; the eigenvectors of this matrix constitute the shape-adaptive PCA; then, the eigenvectors whose corresponding eigenvalues ​​are greater than a predetermined threshold are retained, thereby obtaining a pruned shape-adaptive PCA transform;

[0022] If N gr / N el <τ, it is considered that there are not enough similar neighbors to be used as training data;

[0023] S224, the min(N gr ,N2) Adaptive shape neighborhoods are stacked together to form a three-dimensional array;

[0024] S225, applying the transformation obtained in step S223 to each grouped adaptive shape neighborhood; then, performing a one-dimensional orthogonal transformation along the third dimension of the three-dimensional group;

[0025] S226, shrinking the three-dimensional spectrum;

[0026] S227, inverting the three-dimensional transformation of step S225 to obtain estimated values ​​of all grouped adaptive shape neighborhoods;

[0027] S228. In case of overlap, use the weighted average method to return the obtained estimated value to the original position.

[0028] As a further solution of the present invention: in step S223, the input of the matching block and the reference block is a set of N similar gr Adaptive shape domain; each block in the 2-D domain is represented as a length N el Column vector of Calculate the sample second-order matrix by matrix multiplication;

[0029]

[0030] Subsequently, its eigenvalue decomposition yields:

[0031]

[0032] Where U is an orthogonal matrix and S is a diagonal matrix containing eigenvalues ​​sorted by size: s1>s2>···>s Nel,Finally, the PCs used for adaptive shape neighborhood decomposition are the first N PCs of U trim Column, where N trim is greater than λσ 2 The number of eigenvalues, λ is a fixed threshold;

[0033] As a further solution of the present invention: the standardization and normalization processing specifically includes:

[0034] By using standardization, that is, centering the data according to the mean and then scaling it according to the standard deviation, the data will obey the standard normal distribution with a mean of 0 and a variance of 1, and its range satisfies [-1,1]. The standardization formula is as follows:

[0035]

[0036] Normalization is used, that is, the data x is centered according to the minimum value, and then scaled according to the range, and the final data will converge to [0,1]; the normalization formula is as follows:

[0037]

[0038] As a further solution of the present invention: in step S3, using the DCGAN network architecture to train the model specifically includes:

[0039] YOLOv11 is optimized in the network backbone;

[0040] The initial anchor frame of YOLOv11 is optimized according to the characteristics of belt anomaly recognition of mining belt conveyor;

[0041] In the improved structural block design S3 of the fusion Transformer, YOLOv11 is optimized in the network backbone to build the KAN neural network layer:

[0042] Among them, KAN principle:

[0043] According to the Kolmogorov-Arnold representation theorem: if f is a multivariate continuous function on a bounded domain, it can be written as a composite of finite single variable continuous functions;

[0044] Specifically: For a smooth function:

[0045] f:[0,1] n →R

[0046]

[0047] Among them, x pThat is, it represents the pth element of the vector x, so the range of p is from 1 to n (n is the dimension of the input vector); the index q is used to traverse each component of the external function Φ, so there is a unary function φ p,q (or univariate function) processes the pth component of the input vector x and contributes a term to the sum of the qth external function;

[0048] The theorem shows that every function can be expressed as the sum of univariate functions, and by splicing multiple piecewise polynomials, the univariate function can be made smooth everywhere;

[0049] KANs layer establishment:

[0050] Based on the above characteristics, a one-dimensional function matrix is ​​used to expand the network depth; therefore, the KAN layer is defined as:

[0051] Φ={φ q,p},p=1,2,...,n in ,q=1,2,...,n out

[0052] Among them, the function φ p,q Has trainable parameters;

[0053] In the Kolmogorov-Arnold theorem, the inner function forms a KAN layer with input n and output 2n+1, while the outer function forms a KAN layer with input 2n+1 and output 1;

[0054] Use an integer array to define the KAN layer structure:

[0055] [n0,n1,...,n L ]

[0056] Among them, n i is the number of nodes in the i-th layer of the computational graph; (l,i) represents the i-th neuron in the l-th layer, and x l,i represents the activation value of neuron (l,i); between layer l and layer l+1, there are n l n l+1 Activation function: The activation function connecting (l,i) and (l+1,i) is expressed as:

[0057] φ l,j,i ,l=0,...,L-1,i=1,...,n L ,l=1,...,n l+1

[0058] φ l,j,i The pre-activation is simply represented by x l,i ,φ l,j,i The post-activation is represented by x l,j,i ≡φl,j,i (x l,i ); the activation value of the (l+1,j) neuron is the sum of all incoming post-activations:

[0059]

[0060] In matrix form, this is represented as:

[0061]

[0062] Among them, Φ l Corresponding to the function matrix of the lth KAN layer; the general KAN network contains L layers: For a given input vector x0, the output of KAN is:

[0063]

[0064] Transform the above formula, assume that the output dimension is 1, and define f(x)≡KAN(x):

[0065]

[0066] As a further solution of the present invention: YOLOv11 is optimized in the network backbone part, the traditional CNN layer is replaced by the KAN layer, and the construction of complex nonlinear mapping relationships is realized by adjusting the granularity size in the KAN network. This structure replaces the CNN network and effectively reduces the network parameters and computational loss.

[0067] As a further solution of the present invention: the initial anchor frame of YOLOv11 is optimized, and the preset anchor frame of the mining belt conveyor data set is optimized by using the differential evolution algorithm; specifically including:

[0068] (1) Randomly initialize the population size N, and the evolutionary generation number t = 1;

[0069] (2) Calculate the objective function value of each individual in the initial population;

[0070] (3) Determine whether the termination condition is met or the maximum number of evolution generations is reached; if so, the evolution stops and the best individual at this time is output as the solution; otherwise, proceed to the next step;

[0071] (4) Perform mutation and crossover operations and process boundary conditions to obtain a temporary population;

[0072] (5) Evaluate the temporary population and calculate the fitness value of each individual in the temporary population;

[0073] (6) Perform selection operations to obtain a new population;

[0074] (7) Evolutionary generation t = t + 1, and the new population replaces the old population, and it is determined again whether the termination condition is reached or the maximum evolutionary generation is reached.

[0075] As a further solution of the present invention: the improved structural block design of the fusion Transformer specifically includes:

[0076] The Transformer model includes encoder / decoder, attention mechanism, feedforward network and position encoding module;

[0077] Both encoder and decoder stacks consist of the same stack of layers; each layer contains two sub-layers, the first layer is a multi-head attention mechanism layer, and the second layer is a fully connected feed-forward network layer; residual (ResNet) connections are used in both sub-layers, followed by normalization;

[0078] The attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors;

[0079] The output is calculated using a weighted sum of the values, where the weight assigned to each value is obtained by looking up a compatibility function with the corresponding key; the calculated output matrix is:

[0080]

[0081] The input consists of dimension d k The query and key and dimension are d v The value of is composed of, Softmax is the activation function;

[0082] The definition of the feedforward network is:

[0083] FFN(x)=max(0,xW1+b1)W2+b2

[0084] At the same time, sine and cosine functions of different frequencies are used to describe relative or absolute position information:

[0085]

[0086] Among them, pos represents the position, i is the dimension, d model is the position encoding dimension. That is, each dimension of the position encoding corresponds to a sine wave. The wavelength forms a geometric series from 2π to 10000·2π.

[0087] The beneficial effects of the present invention are as follows: the present invention adopts binarization, SAPCA denoising, standardization and normalization algorithms to form a processing data set; uses a DCGAN network to expand the processing data set to solve the sample imbalance problem in industrial production; realizes belt anomaly detection of a mining belt conveyor based on YOLOv11 through the trained sample data set; replaces the traditional CNN layer by establishing a KAN network layer model to achieve effective parameter reduction and efficiency improvement; realizes rapid optimization of multiple anchor frames by applying a differential evolution algorithm to optimize the preset anchor frame; encodes the relative and absolute positions through Transformer, optimizes the structural blocks to achieve efficiency improvement, and realizes rapid and accurate identification of abnormal conditions of mining belt conveyors. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 is a system flow chart of the present invention;

[0089] Figure 2 This is a schematic diagram of the DCGAN network architecture used in the present invention;

[0090] Figure 3 This is a schematic diagram of the KAN layer structure used in the present invention;

[0091] Figure 4 Schematic diagram of the improved YOLOv11 backbone used in the present invention;

[0092] Figure 5 This is a schematic diagram of the flow of the differential evolution anchor frame optimization algorithm used in the present invention;

[0093] Figure 6 This is a schematic diagram of the improved Transformer structure block used in the present invention. DETAILED DESCRIPTION

[0094] 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.

[0095] Embodiment 1, as Figures 1 to 6 As shown, this embodiment provides a method for detecting belt anomalies of a mining belt conveyor based on KAN-YOLO, and the method for detecting belt anomalies of a mining belt conveyor includes the following steps:

[0096] S1. Data input, obtaining the surface image information of the mining belt conveyor, shooting it with an industrial CCD camera, inputting the image to the PC to form an original image data set;

[0097] S2. Data preprocessing, including:

[0098] S21, image binarization, using an industrial CCD camera to capture the original image information of the mining belt conveyor; selecting a certain threshold to perform binarization processing on the original image;

[0099] S22, filter noise, use SAPCA to reduce noise, obtain the basic estimated image through forward and inverse transformation and calculation of similar blocks, use Euclidean distance to measure the similarity of the basic estimated blocks, and fully restore the details of the original image;

[0100] S23, standardization and normalization, standardizing and normalizing the denoised binary image;

[0101] S24, data enhancement strategy, using supervised data enhancement to expand on the existing data; by performing corresponding geometric transformations on the denoised binary image, including flipping, rotation, shifting, cropping, deformation and scaling operations, the content of the image itself is changed;

[0102] S3. Network training, including:

[0103] S31, sample expansion, using the DCGAN network architecture to train the model and generate a sample-balanced mining belt conveyor data set;

[0104] S32, model training, using KAN-YOLO neural network to train belt anomaly recognition model;

[0105] S4, abnormality recognition, use the model to identify abnormal conditions of image data acquired in real time; according to the identified abnormal conditions, the mining belt conveyor is adjusted in real time to complete the decision output.

[0106] Embodiment 2, in addition to all the technical features of Embodiment 1, this embodiment further includes: In step S2, the noise filtering process for image preprocessing specifically includes the following steps:

[0107] S221, using 8-direction LPA-ICI to obtain an adaptive shape neighborhood centered on the current pixel; the neighborhood is enclosed in a non-adaptive square block, called a reference block, and the number of pixels in the neighborhood is represented by N el ;

[0108] S222, using block matching to find a block with a high degree of similarity to the reference block, and using the shape information obtained in step S221 to extract an adaptive shape neighborhood from the matching block; the number of matching blocks is N gr express;

[0109] S223, determine the transformation to be applied to the adaptive shape neighborhood, depending on Ngr / N el Is it greater than or less than a fixed threshold τ?

[0110] If N gr / N el ≥τ, it is considered that a sufficient number of mutually similar neighborhoods have been found to reliably estimate the second-order matrix; the eigenvectors of this matrix constitute the shape-adaptive PCA; then, the eigenvectors whose corresponding eigenvalues ​​are greater than a predetermined threshold are retained, thereby obtaining a pruned shape-adaptive PCA transform;

[0111] If N gr / N el <τ, it is considered that there are not enough similar neighbors to be used as training data;

[0112] S224, the min(N gr ,N2) Adaptive shape neighborhoods are stacked together to form a three-dimensional array;

[0113] S225, applying the transformation obtained in step S223 to each grouped adaptive shape neighborhood; then, performing a one-dimensional orthogonal transformation along the third dimension of the three-dimensional group;

[0114] S226, shrinking the three-dimensional spectrum;

[0115] S227, inverting the three-dimensional transformation of step S225 to obtain estimated values ​​of all grouped adaptive shape neighborhoods;

[0116] S228. In case of overlap, use the weighted average method to return the obtained estimated value to the original position.

[0117] In step S223, the input of the matching block and the reference block is a set of N similar gr Adaptive shape domain; each block in the 2-D domain is represented as a length N el Column vector of Calculate the sample second-order matrix by matrix multiplication;

[0118]

[0119] Subsequently, its eigenvalue decomposition yields:

[0120]

[0121] Where U is an orthogonal matrix and S is a diagonal matrix containing eigenvalues ​​sorted by size: s1>s2>···>s Nel ,Finally, the PCs used for adaptive shape neighborhood decomposition are the first N PCs of U trim Column, where N trim is greater than λσ2 The number of eigenvalues, λ is a fixed threshold;

[0122] Standardization and normalization processing specifically includes:

[0123] By using standardization, that is, centering the data according to the mean and then scaling it according to the standard deviation, the data will obey the standard normal distribution with a mean of 0 and a variance of 1, and its range satisfies [-1,1]. The standardization formula is as follows:

[0124]

[0125] Normalization is used, that is, the data x is centered according to the minimum value, and then scaled according to the range, and the final data will converge to [0,1]; the normalization formula is as follows:

[0126]

[0127] Embodiment 3: In addition to all the technical features of Embodiment 1, this embodiment also includes:

[0128] In step S3, the DCGAN network architecture is used to train the model, specifically including:

[0129] YOLOv11 is optimized in the network backbone;

[0130] The initial anchor frame of YOLOv11 is optimized according to the characteristics of belt anomaly recognition of mining belt conveyor;

[0131] In the improved structural block design S3 of the fusion Transformer, YOLOv11 is optimized in the network backbone to build the KAN neural network layer:

[0132] Among them, KAN principle:

[0133] According to the Kolmogorov-Arnold representation theorem: if f is a multivariate continuous function on a bounded domain, it can be written as a composite of finite single variable continuous functions;

[0134] Specifically: For a smooth function:

[0135] f:[0,1] n →R

[0136]

[0137] Among them, x p That is, it represents the pth element of the vector x, so the range of p is from 1 to n (n is the dimension of the input vector); the index q is used to traverse each component of the external function Φ, so there is a unary function φ p,q(or univariate function) processes the pth component of the input vector x and contributes a term to the sum of the qth external function;

[0138] The theorem shows that every function can be represented by the sum of one-dimensional functions; however, in practical applications, one-dimensional functions may be non-smooth or even fractal, so the iteration of weight parameters may not be completed in practice. To solve the above problem, the B-spline interpolation function is used. By splicing multiple piecewise polynomials, the one-dimensional function is made smooth everywhere;

[0139] KANs layer establishment:

[0140] Based on the above characteristics, a one-dimensional function matrix is ​​used to expand the network depth; therefore, the KAN layer is defined as:

[0141] Φ={φ q,p},p=1,2,...,n in ,q=1,2,...,n out

[0142] Among them, the function φ p,q Has trainable parameters.

[0143] In the Kolmogorov-Arnold theorem, the inner function forms a KAN layer with input n and output 2n+1, while the outer function forms a KAN layer with input 2n+1 and output 1;

[0144] Strictly, use integer arrays to define the KAN layer structure:

[0145] [n0,n1,...,n L ]

[0146] Among them, n i is the number of nodes in the i-th layer of the computational graph; (l,i) represents the i-th neuron in the l-th layer, and x l,i represents the activation value of neuron (l,i); between layer l and layer l+1, there are n l n l+1 Activation function: The activation function connecting (l,i) and (l+1,i) is expressed as:

[0147] φ l,j,i ,l=0,...,L-1,i=1,...,n L ,l=1,...,n l+1

[0148] φ l,j,i The pre-activation is simply represented by x l,i ,φ l,j,i The post-activation is represented by x l,j,i ≡φ l,j,i (xl,i ); the activation value of the (l+1,j) neuron is the sum of all incoming post-activations:

[0149]

[0150] In matrix form, this is represented as:

[0151]

[0152] Among them, Φ l Corresponding to the function matrix of the lth KAN layer; the general KAN network contains L layers: For a given input vector x0, the output of KAN is:

[0153]

[0154] Transform the above formula, assume that the output dimension is 1, and define f(x)≡KAN(x):

[0155]

[0156] The network backbone of YOLOv11 is optimized, and the KAN network is used to replace the traditional CNN network.

[0157] The network backbone of YOLOv11 is optimized by replacing the traditional CNN layer with the KAN layer. The construction of complex nonlinear mapping relationships is achieved by adjusting the granularity size in the KAN network. This structure replaces the CNN network and effectively reduces the amount of network parameters and computational loss.

[0158] The initial anchor frame of YOLOv11 was optimized, and the preset anchor frame of the mining belt conveyor data set was optimized using the differential evolution algorithm; specifically, the following were included:

[0159] (1) Randomly initialize the population size N, and the evolutionary generation number t = 1;

[0160] (2) Calculate the objective function value of each individual in the initial population;

[0161] (3) Determine whether the termination condition is met or the maximum number of evolution generations is reached; if so, the evolution stops and the best individual at this time is output as the solution; otherwise, proceed to the next step;

[0162] (4) Perform mutation and crossover operations and process boundary conditions to obtain a temporary population;

[0163] (5) Evaluate the temporary population and calculate the fitness value of each individual in the temporary population;

[0164] (6) Perform selection operations to obtain a new population;

[0165] (7) Evolutionary generation t = t + 1, and the new population replaces the old population, and it is determined again whether the termination condition is reached or the maximum evolutionary generation is reached.

[0166] The improved structural block design of the fusion Transformer specifically includes:

[0167] The Transformer model includes encoder / decoder, attention mechanism, feedforward network and position encoding module;

[0168] The encoder and decoder stacks are composed of the same stack of layers; each layer contains two sub-layers. The first layer is a multi-head attention mechanism layer, and the second layer is a fully connected feed-forward network layer; residual (ResNet) connections are used in both sub-layers, followed by normalization;

[0169] The attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors; the output is calculated using the weighted sum of the values, where the weight assigned to each value is obtained by calculating a compatible function of the query and the corresponding key; the calculated output matrix is:

[0170]

[0171] The input consists of dimension d k The query and key and dimension are d v The value of is composed of, Softmax is the activation function;

[0172] The definition of the feedforward network is:

[0173] FFN(x)=max(0,xW1+b1)W2+b2

[0174] At the same time, sine and cosine functions of different frequencies are used to describe relative or absolute position information:

[0175]

[0176] Among them, pos represents the position, i is the dimension, d model is the position encoding dimension. That is, each dimension of the position encoding corresponds to a sine wave. The wavelength forms a geometric series from 2π to 10000·2π.

[0177] The images are taken with an industrial CCD camera and input to the PC to form an original image data set. Binarization, SAPCA denoising, standardization and normalization algorithms are used to form a processed data set. The DCGAN network is used to expand the processed data set, and the belt anomaly detection of the mining belt conveyor is realized based on YOLOv11 through the trained sample data set. The KAN network layer model is established and the differential evolution algorithm is applied to the preset anchor frame to achieve rapid optimization of multiple anchor frames. The relative and absolute positions are encoded through Transformer, and the structural blocks are optimized to improve efficiency. The mining belt conveyor is controlled in real time according to the above-mentioned identified abnormal conditions.

[0178] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0179] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for detecting belt anomalies of a mining belt conveyor based on KAN-YOLO, characterized in that: The method for detecting belt abnormality of a mining belt conveyor comprises the following steps: S1. Data input, obtaining the surface image information of the mining belt conveyor, shooting it with an industrial CCD camera, inputting the image to the PC to form an original image data set; S2. Data preprocessing, including: S21, image binarization, using an industrial CCD camera to capture the original image information of the mining belt conveyor; selecting a threshold and performing binarization processing on the original image; S22, filter noise, use SAPCA to reduce noise, obtain the basic estimated image through forward and inverse transformation and calculation of similar blocks, use Euclidean distance to measure the similarity of the basic estimated blocks, and fully restore the details of the original image; S23, standardization and normalization, standardizing and normalizing the denoised binary image; S24, data enhancement strategy, using supervised data enhancement to expand on the existing data; by performing corresponding geometric transformations on the denoised binary image, including flipping, rotation, shifting, cropping, deformation and scaling operations, the content of the image itself is changed; S3. Network training, including: S31, sample expansion, using the DCGAN network architecture to train the model and generate a sample-balanced mining belt conveyor data set; S32, model training, using KAN-YOLO neural network to train belt anomaly recognition model; S4, abnormality recognition, use the model to identify abnormal conditions of image data acquired in real time; according to the identified abnormal conditions, the mining belt conveyor is adjusted in real time to complete the decision output.

2. The method for detecting belt abnormality of a mining belt conveyor according to claim 1, characterized in that: In step S2, the image preprocessing filtering noise processing specifically includes the following steps: S221, using 8-direction LPA-ICI to obtain an adaptive shape neighborhood centered on the current pixel; the neighborhood is enclosed in a non-adaptive square block, called a reference block, and the number of pixels in the neighborhood is represented by N el ; S222, using block matching to find a block with a high degree of similarity to the reference block, and using the shape information obtained in step S221 to extract an adaptive shape neighborhood from the matching block; the number of matching blocks is N gr express; S223, determine the transformation to be applied to the adaptive shape neighborhood, depending on N gr / N el Is it greater than or less than a fixed threshold τ? If N gr / N el ≥τ, it is considered that a sufficient number of mutually similar neighborhoods have been found to reliably estimate the second-order matrix; the eigenvectors of this matrix constitute the shape-adaptive PCA; then, the eigenvectors whose corresponding eigenvalues ​​are greater than a predetermined threshold are retained, thereby obtaining a pruned shape-adaptive PCA transform; If N gr / N el <τ, it is considered that there are not enough similar neighbors to be used as training data; S224, the min(N gr ,N2) Adaptive shape neighborhoods are stacked together to form a three-dimensional array; S225, applying the transformation obtained in step S223 to each grouped adaptive shape neighborhood, and then performing a one-dimensional orthogonal transformation along the third dimension of the three-dimensional group; S226, shrinking the three-dimensional spectrum; S227, inverting the three-dimensional transformation of step S225 to obtain estimated values ​​of all grouped adaptive shape neighborhoods; S228. In case of overlap, use the weighted average method to return the obtained estimated value to the original position.

3. The method for detecting belt abnormality of a mining belt conveyor according to claim 2, characterized in that: In step S223, the input of the matching block and the reference block is a set of N similar gr Adaptive shape domain; each block in the 2-D domain is represented as a length N el Column vector of Calculate the sample second-order matrix by matrix multiplication; Subsequently, its eigenvalue decomposition yields: Where U is an orthogonal matrix and S is a diagonal matrix containing eigenvalues ​​sorted by size: s1>s2>···>s Nel ,Finally, the PCs used for adaptive shape neighborhood decomposition are the first N PCs of U trim Column, where N trim is greater than λσ 2 is the number of eigenvalues, and λ is a fixed threshold.

4. The method for detecting belt anomalies of a mining belt conveyor based on KAN-YOLO according to claim 3 is characterized in that: The standardization and normalization processing specifically includes: Standardization is used, that is, the data is centered according to the mean and then scaled according to the standard deviation. The data follows a standard normal distribution with a mean of 0 and a variance of 1, and its range satisfies [-1,1]. The standardization formula is as follows: Normalization is used, that is, the data x is centered according to the minimum value, and then scaled according to the range, and the final data will converge to [0,1]; the normalization formula is as follows:

5. The method for detecting belt abnormality of a mining belt conveyor according to claim 1, characterized in that: In step S3, using the DCGAN network architecture to train the model specifically includes: YOLOv11 is optimized in the network backbone; The initial anchor frame of YOLOv11 is optimized according to the characteristics of belt anomaly recognition of mining belt conveyor; In the improved structural block design S3 of the fusion Transformer, YOLOv11 is optimized in the network backbone to build the KAN neural network layer: Among them, KAN principle: According to the Kolmogorov-Arnold representation theorem: if f is a multivariate continuous function on a bounded domain, then it can be written as a composition of finite single variable continuous functions; For smooth functions: f:[0,1] n →R Among them, x p That is, it represents the pth element of the vector x, so the range of p is from 1 to n, where n is the dimension of the input vector; the index q is used to traverse each component of the external function Φ, so there is a unary function φ p,q Process the pth component of the input vector x and contribute a term to the sum of the qth external function; The theorem shows that every function is represented by the sum of univariate functions, and by concatenating multiple piecewise polynomials, the univariate function is made smooth everywhere; KANs layer establishment: Based on the above characteristics, a one-dimensional function matrix is ​​used to expand the network depth; Define the KAN layer as: Φ={φ q,p },p=1,2,...,n in ,q=1,2,...,n out Among them, the function φ p,q Has trainable parameters; In the Kolmogorov-Arnold theorem, the inner function forms a KAN layer with input n and output 2n+1, while the outer function forms a KAN layer with input 2n+1 and output 1; Use an integer array to define the KAN layer structure: [n0,n1,...,n L ] Among them, n i is the number of nodes in the i-th layer of the computational graph; (l,i) represents the i-th neuron in the l-th layer, and x l,i represents the activation value of neuron (l,i); between layer l and layer l+1, there are n l n l+1 Activation function: The activation function connecting (l,i) and (l+1,i) is expressed as: φ l,j,i ,l=0,...,L-1,i=1,...,n L ,l=1,...,n l+1 φ l,j,i The pre-activation is simply represented by x l,i ,φ l,j,i The post-activation is represented by x l,j,i ≡φ l,j,i (x l,i ); the activation value of the (l+1,j) neuron is the sum of all incoming post-activations: In matrix form, this is represented as: Among them, Φ l The function matrix corresponding to the lth KAN layer; the KAN network contains L layers: For a given input vector x0, the output of KAN is: Transform the above formula, assume that the output dimension is 1, and define f(x)≡KAN(x):

6. The method for detecting belt abnormality of a mining belt conveyor according to claim 5, characterized in that: The network backbone of YOLOv11 is optimized by replacing the traditional CNN layer with the KAN layer. The construction of complex nonlinear mapping relationships is achieved by adjusting the granularity size in the KAN network. This structure replaces the CNN network and effectively reduces the amount of network parameters and computational loss.

7. The method for detecting belt abnormality of a mining belt conveyor according to claim 6, characterized in that: The initial anchor frame of YOLOv11 was optimized, and the preset anchor frame of the mining belt conveyor data set was optimized using the differential evolution algorithm; specifically, the following were included: (1) Randomly initialize the population size N, and the evolutionary generation number t = 1; (2) Calculate the objective function value of each individual in the initial population; (3) Determine whether the termination condition is met or the maximum number of evolution generations is reached; if so, the evolution stops and the best individual at this time is output as the solution; otherwise, proceed to the next step; (4) Perform mutation and crossover operations and process boundary conditions to obtain a temporary population; (5) Evaluate the temporary population and calculate the fitness value of each individual in the temporary population; (6) Perform selection operations to obtain a new population; (7) Evolutionary generation t = t + 1, and the new population replaces the old population, and it is determined again whether the termination condition is reached or the maximum evolutionary generation is reached.

8. The method for detecting belt abnormality of a mining belt conveyor according to claim 7, characterized in that: The improved structural block design of the fusion Transformer specifically includes: The Transformer model includes encoder / decoder, attention mechanism, feedforward network and position encoding module; Both encoder and decoder stacks consist of the same stack of layers; each layer contains two sub-layers, the first layer is a multi-head attention mechanism layer, and the second layer is a fully connected feed-forward network layer; Residual connections are used in both sub-layers, followed by normalization. The attention function is described as mapping a query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors; The output is calculated using the weighted sum of the values, where the weight assigned to each value is obtained by looking up a compatible function with the corresponding key; the calculated output matrix is: The input consists of dimension d k The query and key and dimension are d v The value of is composed of, Softmax is the activation function; The definition of the feedforward network is: FFN(x)=max(0,xW1+b1)W2+b2 At the same time, sine and cosine functions of different frequencies are used to describe relative or absolute position information: Among them, pos represents the position, i is the dimension, d model is the position encoding dimension, that is, each dimension of the position encoding corresponds to a sine wave, and the wavelength forms a geometric series from 2π to 10000·2π.

Images