Belt surface damage detection method under dust weak light environment based on self-adaptive space state perception

Abstract

The invention discloses a belt surface damage detection method under a dust weak light environment based on adaptive space state perception, which comprises the following steps: in the operation process of a mining underground belt conveyor, synchronously acquiring dynamic scene image data through a camera and transmitting the dynamic scene image data to an information processing module; the information processing module enhances the acquired image; dynamic target feature information of the belt conveyor is extracted, relative or absolute position information and deep features are obtained from a dynamic scene image, multi-scale size crack features are obtained, and calculation complexity is reduced; representing the image as a spatial state model according to the feature distribution condition of the image; performing crack identification on the image data acquired in real time by using a YOLO improved model; and the crack identification relative position information and the confidence coefficient of each image are obtained and compared with a detection threshold value, and the real-time crack condition of the underground belt conveyor during working is determined. The method can effectively enhance weak textures, accurately extract defect features, and adapt to complex dynamic working conditions.

Description

Technical Field

[0001] This invention relates to a method for detecting surface damage on conveyor belts in dusty and low-light environments based on adaptive spatial state perception, belonging to the field of underground coal mine inspection technology. Background Technology

[0002] In underground coal mines, conveyor belts are crucial production equipment, and their surfaces often develop defects due to long-term operation. Traditional inspection methods are limited by underground environmental conditions, such as insufficient light and dust interference, making it difficult to clearly detect and identify surface defects on conveyor belts.

[0003] Currently, vision-based detection methods, such as conventional edge detection operators, image enhancement methods, and general deep learning models, are widely used due to their non-contact and high-efficiency advantages. However, existing technologies still cannot effectively solve the problem of defect detection in the complex environment of underground coal mines, especially under conditions of low illumination, strong reflection, and dust interference. This is mainly manifested in the following ways: (1) The existing Retinex algorithm mistakenly identifies highlights as texture features or coal dust as background, and cannot achieve physical decoupling and noise reduction; (2) Motion texture interference and computational redundancy: the belt has constant unidirectional high-speed motion characteristics and the background texture is simple; (3) The existing algorithm model adopts an omnidirectional or four-directional scanning mechanism, which not only has a large amount of computation, but also ignores the longitudinal continuity and transverse abrupt flow characteristics of the belt, resulting in insufficient capture ability of longitudinal tear; (4) Large aspect ratio defects are missed: longitudinal tears and scratches usually present an extremely slender shape. The existing loss metric has a sharp decrease in sensitivity when dealing with small offsets of slender targets, resulting in inaccurate positioning. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for detecting surface damage on conveyor belts in dusty and low-light environments based on adaptive spatial state perception. This method can effectively enhance subtle textures, accurately extract defect features, and adapt to complex dynamic working conditions, thereby improving the robustness and accuracy of defect detection in complex environments such as low illumination, dust dispersion, and localized strong light reflection in coal mines.

[0005] The technical solution adopted by this invention to solve its technical problem is: a method for detecting surface damage of conveyor belts in a dusty and low-light environment based on adaptive spatial state perception, comprising the following steps: S1. During the operation of the underground belt conveyor in the mine, dynamic scene image data is collected synchronously by the camera and transmitted to the information processing module; S2. The information processing module enhances the acquired image; the information processing module initially decomposes the image into illumination and reflection parts, converts the image to RGB and HSV color spaces, extracts channel information to form priors through adaptive feature selection method; identifies and mitigates the residuals between illumination and reflection parts, uses super-resolution enhancement to restore hidden details; and supervises the estimation of residuals between parts at different scales based on a multi-scale network architecture. S3. Extract dynamic target feature information of belt conveyor from image enhancement results. Obtain relative or absolute position information and deep features from dynamic scene images based on self-attention and cross-attention. Use axis attention to obtain multi-scale crack features. Reduce computational complexity and improve real-time performance of the model through parameter sharing strategy and axis attention. S4. Based on the feature distribution of the image, the image is represented as a spatial state model, forming a set of state transition equations, structure-aware state fusion equations, and observation equations. S5. Based on the spatial state model and structural perception equations of the image, use the improved YOLO model to identify cracks in the real-time acquired image data. S6. Repeat steps S2 to S5 to obtain the relative position information and confidence level of crack identification for each image, compare it with the detection threshold, and finally determine the real-time crack situation of the underground belt conveyor.

[0006] Furthermore, the camera and information processing module in step S1 are integrated into the visual inspection device, and the camera's field of view covers the dynamic scene below the visual inspection device; the information processing module receives image data from the camera in real time and performs image enhancement, feature extraction and crack recognition online.

[0007] Furthermore, the specific process of step S2 is as follows: S2.1 Construct an image decoupling model based on the prior knowledge of coal dust in dark channels and the physical constraints of high light reflection, to convert low-light images of belt conveyors acquired by cameras in industrial environments into images. Decomposed into lighting and reflection Assume the remaining residual in the illumination component is The residual remaining in the reflectivity component is : ; ; in, Represents the ideal component with no residuals; uses the function To enhance the ideal residual-free component, while introducing estimation error. Enhance the reflected light component The expression is: ; ; The enhanced image is as follows: ; in, Indicates an enhanced image. Indicates an image without residual enhancement. This indicates element-wise multiplication; and Represents the reflection component of the unresidual enhanced illumination; the error between the estimated image and the ideal image is expressed as E Through algebraic calculations, it is derived that: ; ; S2.2, Lighting initialization, specifically including: S2.2-1. Based on different industrial belt conveyor scenarios, an adaptive feature selection method is used to calculate the degree of coal dust obstruction using prior knowledge of dark channels: ; ; in, For dark channels, the minimum light intensity value in a local area represents the concentration of coal dust / fog; The weighting parameter for dust and fog perception is adaptively adjusted in feature fusion, with the weight of high dust areas approaching 0 in dense dust areas and approaching 1 in clear areas. The tanh function is an adjustment factor that controls the sensitivity to dust; it maps the function value to... interval; S2.2-2. Using the diffuse reflection assumption, the reflectivity component is optimized, and the specular suppression loss function is: ; ; in, Represents the SiLU function. This represents the theoretical maximum reflectivity threshold of the belt material. It serves as an indicator mask for highlight areas, determined by the brightness of the input image; The grayscale intensity of the input image; The threshold for highlight determination is set. This step is not a general image enhancement, but a physical constraint mechanism designed for the low diffuse reflectance characteristics of rubber material. It forces the network to strip the high-frequency and high-brightness signals generated by water stains to the illumination component, thereby preventing water light from being falsely detected as cracks. Ultimately, lighting a priori Represented as: ; In the formula, This represents the convolution operation. Indicates a connection along the channel dimension; S2.3, Reflection Initialization: Construct an image decoupling model based on the prior knowledge of dark channel coal dust and the physical constraints of high reflectance, and reflectance components. pass Obtained; adaptive separation using convolutional neural networks : ; In the formula, Represents the SiLU function. The constant value ensures that the denominator is non-zero; this reduces the inter-component similarity between the decomposed illumination and reflection images.

[0008] S2.4 The augmentation network adopts a multi-scale architecture based on residual mitigation and component augmentation modules, including mutual residual estimation (MRES) to simultaneously identify and mitigate inter-component residuals between components and super-resolution enhancement (SES) to recover hidden details; S2.4-1, Mutual Residual Estimation (MRES): This method uses feature similarity to capture residuals. Taking the illumination component as an example, it initializes residuals from different components. , Depth features and those from the reflection component : ; ; ; in, They represent Parameter initialization layer; A similarity matrix between reflectance and illumination features is calculated. This matrix extracts similar features from the reflectance features, thereby capturing the light and shadow residuals in the reflectance component; the estimated reflectance residuals... The expression is as follows: ; in, The learnable parameters for adaptive scaling matrix multiplication are obtained; similarly, the illumination residuals are obtained. : ; After reshaping, the residual reflection is used to... and lighting residue Add raw input By estimating the mutual residual, the optical residual is transferred from reflection to illumination, while the texture residual is transferred from illumination to reflection. S2.4-2, Super-resolution Enhanced SES: Illumination and Reflection Features are respectively Enhanced features They are represented as follows: ; ; The feedforward network FFN consists of three convolutional layers and uses the GELU activation function to achieve nonlinearity; it also uses a conventional image normalization method. For reflection characteristics and lighting characteristics Expand feature resolution to initialize high-resolution features : ; ; in, For an upsampling network with deconvolution and convolutional layers, The scaling factor is used for the captured illumination residual. Utilizing upsampling features initialization Original resolution features i initialization The equation can be rewritten as: ; ; ; S2.5. Based on a multi-scale network architecture, residual estimation between components is supervised at different scales, and deep features of illumination and reflection are reconstructed into enhanced images to avoid inappropriate supervision caused by ground truth decomposition. ; in, Indicates network depth. Each layer consists of 3×3 convolutional layers, used to integrate high-dimensional deep features to synthesize enhanced images. ; S2.6, In decomposing the illumination components At that time, the gradient change in the image was caused by dust, not by changes in illumination, so it was used Error in weighted calculation of dust weight parameters : ; in, The weighting parameters are used to upsample different enhanced images, reducing the loss of detail caused by downsampling of the ground reality, and to calculate pixel-level errors. ; in, Indicates the first An adaptive upsampling network with layers of depth [number] layers. It includes convolutional layers and upsampling layers; Combined with reconstruction loss The overall loss function is: ; ; in, These are parameters for highlight suppression and dust weighting.

[0009] Furthermore, the specific process of step S3 is as follows: S3.1 For the low-light enhanced image, a Transformer-based method is used for feature extraction. The Transformer encoder consists of sequentially connected encoder layers. The encoder input has three channels: query, input, and output. The encoder output is a weighted sum of values, and the query vector is... The dot product and each value Corresponding key vector The attention weights are calculated, and thus the attention function is defined as follows: ; S3.2 Processing and encoding the input sequence to obtain relative or absolute position information: By processing the position encoding operation, the transformed feature map contains relative or absolute position features. The position encoding provides unique position information for each element in sine and cosine format. ; ; in, This refers to the relative position information of image elements. For model dimensions; S3.3 For self-attention layers and cross-attention layers, RGB and HSV color space images As input features, the Transformer framework based on self-attention layers and cross-attention layers is repeatedly computed. Next, its detailed location features are obtained; S3.4, Convolutional network layers are used to implement image processing. Dimensionality reduction, kernel size Depth-wise convolutions are used to obtain high-dimensional information from the image; max-pooling layers are used to realize image... Dimensionality reduction is achieved by extracting the largest feature data from a specified window, thus reducing the size of the feature map. Depthwise convolution and max pooling are calculated as follows: ; S3.5. A parameter sharing strategy is adopted, where the weight parameters of shallow depthwise convolutions are inherited by the weight parameters of deep depthwise convolutions; the number of parameters in the deep convolutional network layer becomes the original. ; S3.6. Employing an axial attention mechanism, the image is divided into preset sizes. Each image patch has an attention value calculated with all other image patches, and the computational complexity is approximately [missing information]. The axis attention mechanism treats a single row or column of image patches as the smallest computational unit to calculate the attention value, and its computational scale is... .

[0010] Furthermore, the specific process of step S4 is as follows: S4.1 Based on state-space model theory, low-light images acquired by cameras in industrial environments are defined as: input sequence ∈ℝ through state variables Convert to output sequence ,in, As a time indicator, The dimension of the state variable; In the underground conveyor belt working environment of coal mines, the processing of data images is not a continuous process. The continuous-time model of the camera video stream is discretized. Zero-order preserved discretization approximates the continuous-time system by keeping the input constant within each discrete time interval, given a time scale. Represents the interval between discrete time steps. For discrete parameters, the discretization rule is applied as follows: ; The discretized recursive formula is: ; ; in, Here is the state transition matrix. Dynamic parameters of the control system; S4.2 The baseline model uses multi-scanning directions to capture the spatial dependencies of neighboring features in the latent state space, and introduces a new structure-aware state fusion equation into the original formula; the entire process of spatial Mamba is described by three equations: the state transition equation, the structure-aware state fusion equation, and the observation equation, which are as follows: ; ; ; in, These are the original state variables. It is a structure-aware state variable. It is a neighborhood group. These are learnable weights. It is a location The Indexes for each domain; original state variables Directly affected by its previous state The influence of structure-aware state variables, while Additional neighboring state variables are included through the fusion mechanism. ,in Indicates the size of the domain set; S4.3. By using an anisotropic flow direction scanning operator, diagonal scanning is abandoned, and a long-range longitudinal flow along the belt movement direction and a local transverse flow in the vertical direction are constructed to decouple the input feature map into two independent directional flows. For the belt movement direction, a longitudinal flow path is defined, and the image is unfolded longitudinally, letting the input sequence be... Define longitudinal system parameters Calculate its longitudinal component: ; ; To capture longitudinal tears, a longer memory region is needed, which is limited during initialization. The characteristic value distribution of the eigenvalues ​​results in a lower corresponding attenuation rate. For the direction perpendicular to the motion, unfold the image row by row, and let the input sequence be... Calculate its horizontal component: ; ; Lateral features are primarily used to determine the crack boundary width and do not require a short memory interval. During initialization, this is achieved by setting... Its high attenuation rate allows it to focus on local features; S4.4 Reshape the output features from both directions back to their original size and fuse them using a gating mechanism: ; in, Here are the gating parameters, Sigmoid is the sigmoid function, and Linear is the linear operation function; ; in, This is the output sequence.

[0011] Furthermore, the specific process of step S5 is as follows: S5.1 The detection framework principle is similar to the YOLO baseline model. Enhanced image information is obtained by acquiring feature information maps through the Transformer module and the state space equation system. The feature pyramid network performs multi-scale processing through image color space, position information and structure-aware features. In this stage, the feature maps achieve spatial resolution adjustment and refinement of features by combining upsampling and downsampling operations, as well as additional convolution processing. The model simultaneously processes high-resolution large-scale features and low-resolution small-scale features, and can more accurately identify and locate cracks of different sizes that may appear in the image at the same time. S5.2. Model the bounding box as a two-dimensional Gaussian distribution. For two two-dimensional Gaussian distributions: ; ; Its second-order Wasserstein distance is defined as: ; Simplify it to: ; in, It is the Frobenius norm; For from the bounding box Modeling Gaussian distribution The above formula can be further expressed as: ; in, These represent the center coordinates, width, and height, respectively. We obtain a new metric for Wasserstein distance (NWD) by normalizing it in exponential form: ; in, C It is a constant closely related to the dataset; S5.3, The standard Wasserstein distance is insensitive to center point offset when dealing with longitudinal tears with extreme aspect ratios; therefore, an aspect ratio sensitivity factor is defined. : ; Using adaptive parameters: ; When the object being tested is a slender crack, the increased denominator leads to a more stringent metric, forcing the model to fit the crack centerline more accurately. The improved NWD metric is as follows: ; Therefore, the loss function is defined as: .

[0012] This invention combines retinal-based image enhancement techniques under low-light and other interference conditions with adaptive feature selection and super-resolution enhancement to effectively restore defect details in images. By decomposing illumination and reflection components and utilizing a multi-scale network architecture and residual estimation methods, it significantly improves defect features. Employing a state-space model and structure-aware dynamic adjustment mechanism, it adaptively adjusts the detection algorithm according to environmental changes. By combining self-attention mechanisms and multi-scale image feature extraction, it can respond in real-time to changes in the mine environment, effectively addressing the challenges of complex and dynamically changing working conditions. By using a parameter-sharing strategy and axial attention mechanism, it optimizes the complexity of model computation, improving the real-time performance and processing efficiency of the detection process. Simultaneously, by utilizing an improved YOLO model combined with real-time image data for crack recognition, it solves the problem of traditional image processing techniques failing to accurately detect defects in complex environments, ensuring stable and rapid defect detection in dynamically changing mine environments, reducing false detection rates, and improving the robustness and accuracy of defect detection. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the illumination and reflection prior feature enhancement method in this invention; Figure 3 This is a structural diagram of the Transformer model with improved feature extraction according to the present invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings.

[0015] like Figure 1 As shown, this invention discloses a method for detecting surface damage on conveyor belts in a dusty, low-light environment based on adaptive spatial state perception, comprising the following steps: S1. During the operation of the underground belt conveyor in the mine, dynamic scene image data is collected synchronously by the camera and transmitted to the information processing module; S2. The information processing module enhances the acquired image; the information processing module initially decomposes the image into illumination and reflection parts, converts the image to RGB and HSV color spaces, extracts channel information to form priors through adaptive feature selection method; identifies and mitigates the residuals between illumination and reflection parts, uses super-resolution enhancement to restore hidden details; and supervises the estimation of residuals between parts at different scales based on a multi-scale network architecture. S3. Extract dynamic target feature information of belt conveyor from image enhancement results. Obtain relative or absolute position information and deep features from dynamic scene images based on self-attention and cross-attention. Use axis attention to obtain multi-scale crack features. Reduce computational complexity and improve real-time performance of the model through parameter sharing strategy and axis attention. S4. Based on the feature distribution of the image, the image is represented as a spatial state model, forming a set of state transition equations, structure-aware state fusion equations, and observation equations. S5. Based on the spatial state model and structural perception equations of the image, use the improved YOLO model to identify cracks in the real-time acquired image data. S6. Repeat steps S2 to S5 to obtain the relative position information and confidence level of crack identification for each image, compare it with the detection threshold, and finally determine the real-time crack situation of the underground belt conveyor.

[0016] This invention provides a priori image decomposition method based on illumination and reflection, utilizing the illumination-reflection mechanism to pre-decompose the image mirror. By decomposing the illumination and reflection components and employing a multi-scale network architecture and residual estimation methods, the feature characteristics of defects are significantly improved. Adaptive feature selection and super-resolution enhancement are used to effectively recover defect details in the image. The image enhancement method includes the following steps: S2.1, such as Figure 2 As shown, an image decoupling model is constructed based on the prior knowledge of coal dust in dark channels and the physical constraints of high light reflection. This model is used to decouple low-light images of belt conveyors captured by cameras in industrial environments. Decomposed into lighting and reflection Assume the remaining residual in the illumination component is The residual remaining in the reflectivity component is : ; ; in, Represents the ideal component with no residuals; uses the function To enhance the ideal residual-free component, while introducing estimation error. Enhance the reflected light component The expression is: ; ; The enhanced image is as follows: ; in, Indicates an enhanced image. Indicates an image without residual enhancement. This indicates element-wise multiplication; and Represents the reflection component of the unresidual enhanced illumination; the error between the estimated image and the ideal image is expressed as E Through algebraic calculations, it is derived that: ; ; S2.2, Lighting initialization, specifically including: S2.2-1. Based on different industrial belt conveyor scenarios, an adaptive feature selection method is used to calculate the degree of coal dust obstruction using prior knowledge of dark channels: ; ; in, For dark channels, the minimum light intensity value in a local area represents the concentration of coal dust / fog; The weighting parameter for dust and fog perception is adaptively adjusted in feature fusion, with the weight of high dust areas approaching 0 in dense dust areas and approaching 1 in clear areas. The tanh function is an adjustment factor that controls the sensitivity to dust; it maps the function value to... interval; S2.2-2. Using the diffuse reflection assumption, the reflectivity component is optimized, and the specular suppression loss function is: ; ; in, Represents the SiLU function. This represents the theoretical maximum reflectivity threshold of the belt material. It serves as an indicator mask for highlight areas, determined by the brightness of the input image; The grayscale intensity of the input image; The threshold for highlight determination is set. This step is not a general image enhancement, but a physical constraint mechanism designed for the low diffuse reflectance characteristics of rubber material. It forces the network to strip the high-frequency and high-brightness signals generated by water stains to the illumination component, thereby preventing water light from being falsely detected as cracks. Ultimately, lighting a priori Represented as: ; In the formula, This represents the convolution operation. Indicates a connection along the channel dimension; S2.3, Reflection Initialization: Construct an image decoupling model based on the prior knowledge of dark channel coal dust and the physical constraints of high reflectance, and reflectance components. pass Obtained; adaptive separation using convolutional neural networks : ; In the formula, Represents the SiLU function. The constant value ensures that the denominator is non-zero; this reduces the inter-component similarity between the decomposed illumination and reflection images.

[0017] S2.4 The augmentation network adopts a multi-scale architecture based on residual mitigation and component augmentation modules, including mutual residual estimation (MRES) to simultaneously identify and mitigate inter-component residuals between components and super-resolution enhancement (SES) to recover hidden details; S2.4-1, Mutual Residual Estimation (MRES): This method uses feature similarity to capture residuals. Taking the illumination component as an example, it initializes residuals from different components. Depth features and those from the reflection component : ; ; ; in, They represent Parameter initialization layer; A similarity matrix between reflectance and illumination features is calculated. This matrix extracts similar features from the reflectance features, thereby capturing the light and shadow residuals in the reflectance component; the estimated reflectance residuals... The expression is as follows: ; in, The learnable parameters for adaptive scaling matrix multiplication are obtained; similarly, the illumination residuals are obtained. : ; After reshaping, the residual reflection is used to... and lighting residue Add raw input By estimating the mutual residual, the optical residual is transferred from reflection to illumination, while the texture residual is transferred from illumination to reflection. S2.4-2, Super-resolution Enhanced SES: Illumination and Reflection Features are respectively Enhanced features They are represented as follows: ; ; The feedforward network FFN consists of three convolutional layers and uses the GELU activation function to achieve nonlinearity; it also uses a conventional image normalization method. For reflection characteristics and lighting characteristics Expand feature resolution to initialize high-resolution features : ; ; in, For an upsampling network with deconvolution and convolutional layers, The scaling factor is used for the captured illumination residual. Utilizing upsampling features initialization Original resolution features The equation can be rewritten as: ; ; ; S2.5. Based on a multi-scale network architecture, residual estimation between components is supervised at different scales, and deep features of illumination and reflection are reconstructed into enhanced images to avoid inappropriate supervision caused by ground truth decomposition. ; in, Indicates network depth. All by It consists of several convolutional layers, used to integrate high-dimensional deep features to synthesize enhanced images. ; S2.6, In decomposing the illumination components At that time, the gradient change in the image was caused by dust, not by changes in illumination, so it was used Error in weighted calculation of dust weight parameters : ; in, The weighting parameters are used to upsample different enhanced images, reducing the loss of detail caused by downsampling of the ground reality, and to calculate pixel-level errors. ; in, Indicates the first An adaptive upsampling network with layers of depth [number] layers. It includes convolutional layers and upsampling layers; Combined with reconstruction loss The overall loss function is: ; ; in, These are parameters for highlight suppression and dust weighting.

[0018] As a preferred implementation method, such as Figure 3 As shown, the specific process of step S3 is as follows: S3.1 For the low-light enhanced image, a Transformer-based method is used for feature extraction. The Transformer encoder consists of sequentially connected encoder layers. The encoder input has three channels: query, input, and output. The encoder output is a weighted sum of values, and the query vector is... The dot product and each value Corresponding key vector The attention weights are calculated, and thus the attention function is defined as follows: ; S3.2 Processing and encoding the input sequence to obtain relative or absolute position information: By processing the position encoding operation, the transformed feature map contains relative or absolute position features. The position encoding provides unique position information for each element in sine and cosine format. ; ; in, This refers to the relative position information of image elements. For model dimensions.

[0019] S3.3 For self-attention layers and cross-attention layers, RGB and HSV color space images As input features, the Transformer framework based on self-attention layers and cross-attention layers is repeatedly computed. Next, its detailed location features are obtained; S3.4, Convolutional network layers are used to implement image processing. Dimensionality reduction, kernel size Depth-wise convolutions are used to obtain high-dimensional information from the image; max-pooling layers are used to realize image... Dimensionality reduction is achieved by extracting the largest feature data from a specified window, thus reducing the size of the feature map. Depthwise convolution and max pooling are calculated as follows: ; S3.5. A parameter sharing strategy is adopted, where the weight parameters of shallow depthwise convolutions are inherited by the weight parameters of deep depthwise convolutions; the number of parameters in the deep convolutional network layer becomes the original. ; S3.6. Employing an axial attention mechanism, the image is divided into preset sizes. Each image patch has an attention value calculated with all other image patches, and the computational complexity is approximately [missing information]. The axis attention mechanism treats a single row or column of image patches as the smallest computational unit to calculate the attention value, and its computational scale is... .

[0020] In a preferred embodiment, the specific process of step S4 is as follows: S4.1 Based on state-space model theory, low-light images acquired by cameras in industrial environments are defined as: input sequence Through state variables Convert to output sequence ,in, As a time indicator, The dimension of the state variable; In the underground conveyor belt working environment of coal mines, the processing of data images is not a continuous process. The continuous-time model of the camera video stream is discretized. Zero-order preserved discretization approximates the continuous-time system by keeping the input constant within each discrete time interval, given a time scale. Represents the interval between discrete time steps. For discrete parameters, the discretization rule is applied as follows: ; The discretized recursive formula is: ; ; in, Here is the state transition matrix. Dynamic parameters of the control system; S4.2 The baseline model uses multi-scanning directions to capture the spatial dependencies of neighboring features in the latent state space, and introduces a new structure-aware state fusion equation into the original formula; the entire process of spatial Mamba is described by three equations: the state transition equation, the structure-aware state fusion equation, and the observation equation, which are as follows: ; ; ; in, These are the original state variables. It is a structure-aware state variable. It is a neighborhood group. These are learnable weights. It is a location The Indexes for each domain; original state variables Directly affected by its previous state The influence of structure-aware state variables, while Additional neighboring state variables are included through the fusion mechanism. ,in Indicates the size of the domain set; S4.3. By using an anisotropic flow direction scanning operator, diagonal scanning is abandoned, and a long-range longitudinal flow along the belt movement direction and a local transverse flow in the vertical direction are constructed to decouple the input feature map into two independent directional flows. For the belt movement direction, a longitudinal flow path is defined, and the image is unfolded longitudinally, letting the input sequence be... Define longitudinal system parameters Calculate its longitudinal component: ; ; To capture longitudinal tears, a longer memory region is needed, which is limited during initialization. The characteristic value distribution of the eigenvalues ​​results in a lower corresponding attenuation rate. For the direction perpendicular to the motion, unfold the image row by row, and let the input sequence be... Calculate its horizontal component: ; ; Lateral features are primarily used to determine the crack boundary width and do not require a short memory interval. During initialization, this is achieved by setting... Its high attenuation rate allows it to focus on local features; S4.4 Reshape the output features from both directions back to their original size and fuse them using a gating mechanism: ; in, Here are the gating parameters, Sigmoid is the sigmoid function, and Linear is the linear operation function; ; in, This is the output sequence.

[0021] In a preferred embodiment, the specific process of step S5 is as follows: S5.1 The detection framework principle is similar to the YOLO baseline model. Enhanced image information is obtained by acquiring feature information maps through the Transformer module and the state space equation system. The feature pyramid network performs multi-scale processing through image color space, position information and structure-aware features. In this stage, the feature maps achieve spatial resolution adjustment and refinement of features by combining upsampling and downsampling operations, as well as additional convolution processing. The model simultaneously processes high-resolution large-scale features and low-resolution small-scale features, and can more accurately identify and locate cracks of different sizes that may appear in the image at the same time. S5.2. Model the bounding box as a two-dimensional Gaussian distribution. For two two-dimensional Gaussian distributions: ; ; Its second-order Wasserstein distance is defined as: ; Simplify it to: ; in, It is the Frobenius norm; For from the bounding box Modeling Gaussian distribution The above formula can be further expressed as: ; in, These represent the center coordinates, width, and height, respectively. We obtain a new metric for Wasserstein distance (NWD) by normalizing it in exponential form: ; in, It is a constant closely related to the dataset; S5.3, The standard Wasserstein distance is insensitive to center point offset when dealing with longitudinal tears with extreme aspect ratios; therefore, an aspect ratio sensitivity factor is defined. : ; Using adaptive parameters: ; When the object being tested is a slender crack, the increased denominator leads to a more stringent metric, forcing the model to fit the crack centerline more accurately. The improved NWD metric is as follows: ; Therefore, the loss function is defined as: .

Claims

1. A method for detecting surface damage of conveyor belts in dusty and low-light environments based on adaptive spatial state perception, characterized in that, Includes the following steps: S1. During the operation of the underground belt conveyor in the mine, dynamic scene image data is collected synchronously by the camera and transmitted to the information processing module; S2, the information processing module enhances the acquired images; S3. Extract dynamic target feature information of belt conveyor from image enhancement results, obtain relative or absolute position information and deep features from dynamic scene images based on self-attention and cross-attention; use axis attention to obtain multi-scale crack features; By using parameter sharing strategies and axis attention, computational complexity is reduced and the real-time performance of the model is improved. S4. Based on the feature distribution of the image, the image is represented as a spatial state model, forming a set of state transition equations, structure-aware state fusion equations, and observation equations. S5. Based on the spatial state model and structural perception equations of the image, use the improved YOLO model to identify cracks in the real-time acquired image data. S6. Repeat steps S2 to S5 to obtain the relative position information and confidence level of crack identification for each image, compare it with the detection threshold, and finally determine the real-time crack situation of the underground belt conveyor.

2. The method for detecting belt surface damage in a dusty, low-light environment based on adaptive spatial state perception according to claim 1, characterized in that, The camera and information processing module in step S1 are integrated in the visual inspection device. The camera's field of view covers the dynamic scene below the visual inspection device. The information processing module receives image data from the camera in real time and performs image enhancement, feature extraction, and crack recognition online.

3. The method for detecting belt surface damage in a dusty, low-light environment based on adaptive spatial state perception according to claim 1, characterized in that, The specific process of step S2 is as follows: S2.1 Construct an image decoupling model based on the prior knowledge of coal dust in dark channels and the physical constraints of high light reflection, to convert low-light images of belt conveyors acquired by cameras in industrial environments into images. I l Decomposed into lighting L l and reflection R l ; Assume the remaining residual in the illumination component is ε R The residual remaining in the reflectivity component is ε L : ； ； in, Represents ideal components with no residuals; uses functions To enhance the ideal residual-free component, while introducing estimation error. Enhance the reflected light component The expression is: ； ； The enhanced image is as follows: ； in, Indicates an enhanced image. Indicates an image without residual enhancement. This indicates element-wise multiplication; Represents the reflection component of the unresidual enhanced illumination; the error between the estimated image and the ideal image is expressed as E Through algebraic calculations, it is derived that: ； ； S2.2, Lighting initialization, specifically including: S2.2-1. Based on different industrial belt conveyor scenarios, an adaptive feature selection method is used to calculate the degree of coal dust obstruction using prior knowledge of dark channels: ； ； in, For dark channels, the minimum light intensity value in a local area represents the concentration of coal dust / fog; The weighting parameter for dust and fog perception is adaptively adjusted in feature fusion, with the weight of high dust areas approaching 0 in dense dust areas and approaching 1 in clear areas. As a regulating factor, it controls the degree of sensitivity to dust; The function maps function values ​​to interval; S2.2-2. Using the diffuse reflection assumption, the reflectivity component is optimized, and the specular suppression loss function is: ； ； in, express function, This represents the theoretical maximum reflectivity threshold of the belt material. It serves as an indicator mask for highlight areas, determined by the brightness of the input image; The grayscale intensity of the input image; The threshold for determining highlights; Ultimately, lighting a priori Represented as: ； In the formula, This represents the convolution operation. Indicates a connection along the channel dimension; S2.3, Reflection Initialization: Construct an image decoupling model based on the prior knowledge of dark channel coal dust and the physical constraints of high reflectance, and reflectance components. pass Obtained; adaptive separation using convolutional neural networks : ； In the formula, express function, The constant is used to ensure that the denominator is non-zero; S2.4 The augmentation network adopts a multi-scale architecture based on residual mitigation and component augmentation modules, including mutual residual estimation and super-resolution augmentation; S2.4-1, Mutual Residual Estimation: Residuals are captured using feature similarity. Taking the illumination component as an example, the residuals from different components are initialized. Depth features and those from the reflection component : ； ； ； in, They represent Parameter initialization layer; A similarity matrix between reflectance and illumination features is calculated. This matrix extracts similar features from the reflectance features, thereby capturing the light and shadow residuals in the reflectance component; the estimated reflectance residuals... The expression is as follows: ； in, The learnable parameters for adaptive scaling matrix multiplication are obtained; similarly, the illumination residuals are obtained. : ； After reshaping, the residual reflection is used to... and lighting residue Add raw input By estimating the mutual residual, the optical residual is transferred from reflection to illumination, while the texture residual is transferred from illumination to reflection. S2.4-2, Super-resolution Enhanced SES: Illumination and Reflection Features are respectively Enhanced features They are represented as follows: ； ； The feedforward network FFN consists of three convolutional layers and uses the GELU activation function to achieve nonlinearity; it also uses a conventional image normalization method. For reflection characteristics and lighting characteristics Expand feature resolution to initialize high-resolution features : ； ； in, For an upsampling network with deconvolution and convolutional layers, The scaling factor is used for the captured illumination residual. Utilizing upsampling features initialization Original resolution features initialization The equation can be rewritten as: ； ； ； S2.

5. Based on a multi-scale network architecture, residual estimation between components is supervised at different scales, and deep features of illumination and reflection are reconstructed into enhanced images to avoid inappropriate supervision caused by ground truth decomposition. ； in, Indicates network depth. All by It consists of several convolutional layers, used to integrate high-dimensional deep features to synthesize multi-scale enhanced images. ; S2.6, In decomposing the illumination components At that time, the gradient change in the image was caused by dust, not by changes in illumination, so it was used Error in weighted calculation of dust weight parameters : ； in, The weighting parameters are used to upsample different enhanced images, reducing the loss of detail caused by downsampling of the ground reality, and to calculate pixel-level errors. : ； in, Indicates the first An adaptive upsampling network with layers of depth [number] layers. It includes convolutional layers and upsampling layers; Combined with reconstruction loss The overall loss function is: ； ； in, These are parameters for highlight suppression and dust weighting.

4. The method for detecting belt surface damage in a dusty, low-light environment based on adaptive spatial state perception according to claim 1, characterized in that, The specific process of step S3 is as follows: S3.1 For the low-light enhanced image, a Transformer-based method is used for feature extraction. The Transformer encoder consists of sequentially connected encoder layers. The encoder input has three channels: query, input, and output. ,key Sum The encoder output is a weighted sum of values, and the query vector is... according to The dot product and each value Corresponding key vector The attention weights are calculated, and thus the attention function is defined as follows: ； S3.2 Processing and encoding the input sequence to obtain relative or absolute position information: By processing the position encoding operation, the transformed feature map contains relative or absolute position features. The position encoding provides unique position information for each element in sine and cosine format. ； ； in, This refers to the relative position information of image elements. For model dimensions; S3.3 For self-attention layers and cross-attention layers, RGB and HSV color space images As input features, the Transformer framework based on self-attention layers and cross-attention layers is repeatedly computed. Next, its detailed location features are obtained; S3.4, Convolutional network layers are used to implement image processing. Dimensionality reduction, kernel size Depth-wise convolutions are used to obtain high-dimensional information from the image; max-pooling layers are used to realize image... Dimensionality reduction is achieved by extracting the largest feature data from a specified window, thus reducing the size of the feature map. Depthwise convolution and max pooling are calculated as follows: ； S3.

5. A parameter sharing strategy is adopted, where the weight parameters of shallow depthwise convolutions are inherited by the weight parameters of deep depthwise convolutions; the number of parameters in the deep convolutional network layer becomes the original. ; S3.

6. Employing an axial attention mechanism, the image is divided into preset sizes. Each image patch has an attention value calculated with all other image patches, and the computational complexity is approximately [missing information]. The axis attention mechanism treats a single row or column of image patches as the smallest computational unit to calculate the attention value, and its computational scale is... .

5. The method for detecting belt surface damage in a dusty, low-light environment based on adaptive spatial state perception according to claim 1, characterized in that, The specific process of step S4 is as follows: S4.1 Based on state-space model theory, low-light images acquired by cameras in industrial environments are defined as: input sequence Through state variables Convert to output sequence ,in, As a time indicator, The dimension of the state variable; The continuous-time model of the camera video stream is discretized. Zero-order preserved discretization approximates the continuous-time system by keeping the input constant in each discrete time interval, given a time scale. , This represents the interval between discrete time steps. For discrete parameters, the discretization rule is applied as follows: The discretized recursive formula is: ； ； in, Here is the state transition matrix. Dynamic parameters of the control system; S4.2 The baseline model uses multi-scanning directions to capture the spatial dependencies of neighboring features in the latent state space, and introduces a new structure-aware state fusion equation into the original formula; the entire process of spatial Mamba is described by three equations: the state transition equation, the structure-aware state fusion equation, and the observation equation, which are as follows: ； ； ； in, These are the original state variables. It is a structure-aware state variable. It is a neighborhood group. These are learnable weights. It is a location The Indexes for each domain, It is the output quantity; the original state variable. Directly affected by its previous state The influence of structure-aware state variables, while Additional neighboring state variables are included through the fusion mechanism. ,in Indicates the size of the domain set; S4.

3. By using an anisotropic flow direction scanning operator, diagonal scanning is abandoned, and a long-range longitudinal flow along the belt movement direction and a local transverse flow in the vertical direction are constructed to decouple the input feature map into two independent directional flows. For the belt movement direction, a longitudinal flow path is defined, and the image is unfolded longitudinally, letting the input sequence be... Define longitudinal system parameters Calculate its longitudinal component: ； ； To capture longitudinal tears, a longer memory interval is required. During initialization, a corresponding decay rate limit is set. The eigenvalue distribution; For the direction perpendicular to the motion, unfold the image row by row, and let the input sequence be... Calculate its lateral component: ； ； Lateral features are primarily used to determine the crack boundary width and do not require a short memory interval. During initialization, this is achieved by setting... The decay rate makes it focus on local features; S4.4 Reshape the output features from both directions back to their original size and fuse them using a gating mechanism: ； in, Here are the gating parameters, Sigmoid is the sigmoid function, and Linear is the linear operation function; ； in, This is the output sequence.

6. The method for detecting belt surface damage in a dusty, low-light environment based on adaptive spatial state perception according to claim 1, characterized in that, The specific process of step S5 is as follows: S5.1 The detection framework principle is similar to the YOLO baseline model. Enhanced image information is obtained by obtaining feature information maps through the Transformer module and the state space equation system. The Feature Pyramid Network performs multi-scale processing using image color space, location information, and structure-aware features. In this stage, the feature maps are refined by combining upsampling and downsampling operations, as well as convolution. The model simultaneously processes high-resolution large-scale features and low-resolution small-scale features, enabling more accurate identification and localization of cracks of varying sizes that may appear in the image at the same time. S5.

2. Model the bounding box as a two-dimensional Gaussian distribution. For two two-dimensional Gaussian distributions: ； ； Its second-order Wasserstein distance is defined as: ； Simplify it to: ； in, It is the Frobenius norm; For from the bounding box Modeling Gaussian distribution The above formula can be further expressed as: ； in, These represent the center coordinates, width, and height, respectively. We obtain a new metric for Wasserstein distance (NWD) by normalizing it in exponential form: ； in, C It is a constant related to the dataset; S5.3, The standard Wasserstein distance is insensitive to center point offset when dealing with longitudinal tears with extreme aspect ratios; therefore, an aspect ratio sensitivity factor is defined. : ； Using adaptive parameters: ； When the object being tested is a slender crack, the increased denominator leads to a more stringent metric, forcing the model to fit the crack centerline more accurately. The improved NWD metric is as follows: ； Therefore, the loss function is defined as: 。

Images