Intelligent visual perception multivariable adaptive belt deviation control method and system

Abstract

The present invention discloses a multivariable adaptive belt deviation correction control method and system of intelligent visual perception, the method comprising: S1, detecting the actual position of the belt through an intelligent visual perception total module; S2, comparing the detected actual position of the belt with the belt position setting value to obtain the position deviation, and inputting the position deviation into a PI controller to obtain an accurate position deviation control output; S3, comparing the position deviation control output with the belt offset obtained by observing the current of the deviation correction motor through a state observer to obtain the actual deviation correction amount, and then obtaining an input control law; S4, inputting the input control law into a state space model to obtain a deviation correction signal of the deviation correction motor; S5, outputting the motor current according to the deviation correction signal, and calculating the torque of the deviation correction motor through a torque coefficient to drive a belt deviation correction device to correct the belt. The present invention can correct the belt efficiently and accurately.

Description

Technical Field

[0001] The invention relates to the technical field of belt deviation correction control, and in particular to a multivariable adaptive belt deviation correction control method and system with intelligent visual perception. Background Art

[0002] Industrial belt conveyor systems are widely used in heavy industries such as mining, metallurgy, electricity, and chemicals. Their stability is crucial to production efficiency and safety. However, due to factors such as complex environment, belt load changes, vibration, etc., belts are prone to deviation, which affects equipment life and production efficiency. Belt deviation correction in industrial production is not only related to production efficiency and equipment safety, but also affects product quality and corporate economic benefits. For this reason, it is necessary to pay attention to belt status monitoring and deviation correction control.

[0003] The core goal of the belt deviation correction system is to ensure that the belt always stays on the center track during operation, thereby effectively preventing deviation, friction and damage, and ensuring the safety and efficiency of material transportation. The traditional deviation correction system realizes automatic adjustment through the linkage of sensors and programmable logic controllers (PLCs). Specifically, the deviation sensor detects whether the belt deviates from the center line in real time and transmits the deviation signal to the controller. The controller accurately calculates the direction and amplitude of the deviation based on the preset control algorithm, and then drives the electric push rod to adjust the position of the roller to correct the running track of the belt. However, to build an automatic deviation correction system with excellent performance, a high level of balance must be achieved in terms of real-time performance, accuracy, stability, adaptability and robustness. Limited by the inherent characteristics of the belt deviation correction system - the hysteresis of signal feedback, the strong coupling of variables within the system, and complex and changeable working conditions - the existing technical solutions are often difficult to achieve ideal control effects in practical applications. Summary of the invention

[0004] The present invention provides a multivariable adaptive belt deviation correction control method and system with intelligent visual perception to solve the problems existing in the above-mentioned prior art. The technical solution is as follows:

[0005] On the one hand, a multivariable adaptive belt deviation control method with intelligent visual perception is provided, comprising:

[0006] S1, detect the actual position of the belt through the intelligent visual perception module;

[0007] S2, comparing the detected actual belt position with the belt position setting value to obtain a position deviation, and inputting the position deviation into a PI controller to obtain an accurate position deviation control output;

[0008] S3, comparing the position deviation control output with the belt deviation obtained by observing the current of the deviation correction motor through a state observer, obtaining an actual deviation correction amount, and then obtaining an input control law;

[0009] S4, inputting the input control law into a state space model to obtain a correction signal of the correction motor;

[0010] S5, outputting a motor current according to the deviation correction signal, and obtaining the torque of the deviation correction motor by calculating the torque coefficient, so as to drive a belt deviation correction device to correct the belt;

[0011] The S1 specifically includes:

[0012] S11, obtaining a multi-resolution, multi-view, and multi-state belt RGB image sequence;

[0013] S12, extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature graph;

[0014] S13, inputting the feature map after adding Gaussian noise into the enhanced belt segmentation module, wherein the enhanced belt segmentation module focuses on the key area of ​​the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask;

[0015] S14, inputting the belt segmentation mask into a multi-domain collaborative edge perception module, the multi-domain collaborative edge perception module collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map;

[0016] S15, inputting the final belt edge map into an intelligent linear regression module to generate a belt edge line map;

[0017] S16, inputting the belt edge line diagram into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

[0018] Optionally, the S12 specifically includes:

[0019] S121, organizing the belt RGB image sequence into an image sequence matrix according to state changes;

[0020] S122, creating a weight matrix of the same size as the image sequence matrix , using standard Gaussian distribution to generate the weight matrix Elements in And normalize it;

[0021] S123, use CNN convolutional neural network to extract the feature matrix of each frame image , according to the feature matrix Elements , adjust the weight matrix Elements in , the formula is as follows:

[0022]

[0023] in, is the adjustment coefficient, which controls the influence of the feature on the weight. Perform normalization;

[0024] S124, each frame of the image generates a Gaussian noise matrix with the same size as the image ,Will Elements in After normalization Do the inner product to get the weighted noise matrix for each frame of the image , the formula is as follows:

[0025]

[0026] S125, the weighted noise matrix Superimposed onto the feature matrix The final dynamic weighted feature matrix after adding noise is obtained , the formula is as follows:

[0027] .

[0028] Optionally, the S13 specifically includes:

[0029] S131, using a CNN convolutional neural network as a feature extractor, extracting multi-scale features from the feature map after adding noise, and obtaining a multi-scale feature map;

[0030] S132, dividing the multi-scale feature map into a series of feature blocks, and fusing the feature blocks to form a multi-scale fused feature map;

[0031] S133, input the multi-scale fusion feature map into the feature encoder, the feature encoder introduces position coding on the basis of the multi-scale fusion feature map, identifies the spatial position of the feature, and combines it into a feature vector containing position information. The feature vector passes through the Linear layer to generate three vectors Q, K, and V. Q, K, and V are the core feature vectors used for the attention mechanism in deep learning, which are used to measure the correlation between features. Then, the correlation between Q and K is calculated through the self-attention mechanism to obtain a correlation matrix. Then, the correlation matrix is ​​used to perform weighted summation on the V vector. The weight coefficient of each V vector is the correlation calculated by Q and K, and a new set of Q', K', and V' vectors containing the context information of all input features is obtained, wherein Q' is not passed to the cross-attention mechanism in the decoder, and only K' and V' are passed to the cross-attention mechanism in the decoder. The decoder has a total of N layers. The input of the current layer decoder comes from the Q'' vector generated in the previous N-1 layer decoder. The Q'' vector of the first layer decoder comes from an initialization vector, which represents the initial state of the decoder. The current layer decoder will be based on the input Q'' Combined with K' and V' passed from the encoder of the current layer, the cross-attention mechanism is completed to generate higher-quality feature vectors. These higher-quality feature vectors are used to interact with the decoder of the next layer, and then pass through the feedforward network and normalization. The feedforward network can introduce more complex transformations to the features of each position, enhancing the model's ability to express input data. The purpose of normalization is to normalize the output of each layer. Through normalization, the output vector of each layer is rescaled and offset to make its mean 0 and variance 1, and a new set of Q''', K''', and V''' is output. The obtained Q''', K''', and V''' are calculated through the self-attention mechanism to obtain a new feature vector Imformation;

[0032] S134, passing the new feature vector Imformation through the multi-layer perceptron and segmentation head of the head network to generate the belt segmentation mask.

[0033] Optionally, the S14 specifically includes:

[0034] S141, extracting spatial domain features from the binarized belt segmentation mask, extracting multi-scale edge features through a backbone network, and outputting feature maps of each layer with sizes of [224, 224, 64], [112, 112, 128], [56, 56, 256], and [28, 28, 512], generating edge maps for each layer of features through a sigmoid activation function, and restoring the original image size through upsampling to obtain four edge feature maps of the same size;

[0035] S142, performing fast Fourier transform on the binarized belt segmentation mask to generate a frequency domain feature map, using a flexible high-pass filter in the frequency domain to extract high-frequency components to enhance edge features, and regressing to the spatial domain through inverse Fourier transform to obtain a frequency domain enhanced edge feature map;

[0036] S143, fusing the four edge feature maps with the frequency domain enhanced edge feature map to obtain a final belt edge map.

[0037] Optionally, the S15 specifically includes:

[0038] S151, the final belt edge image is downsampled, the resolution is gradually reduced, the edge features of the belt are extracted, and the data dimension is reduced to improve the calculation efficiency;

[0039] S512, entering the morphological processing stage, performing dilation and corrosion operations on the edge features to remove isolated noise points and enhance the continuity of the edge lines to ensure that the edge features are clear;

[0040] S152, upsampling the feature map after morphological processing to the original image size, and fusing the original images obtained by upsampling at different scales through a fusion layer;

[0041] S153, smoothing the fused feature map to obtain a smoothed feature map, then traversing the extracted edge mutation pixel points through gradient calculation, regressing these edge mutation pixel points using a straight line fitting algorithm, and fitting a straight line representing the edge of the belt;

[0042] S154, upsampling the low-resolution straight line fitting result and gradually restoring it to a size consistent with the original input, thereby generating a clear and continuous line diagram of the belt edge.

[0043] Optionally, the S16 specifically includes:

[0044] Adaptively select n reference point coordinates of the belt edge and combine them with weighted calculation. The weight design fully considers the importance of the reference point and its local characteristics, making the algorithm more adaptable and anti-interference to subtle changes in the edge, and obtaining the final reference point coordinates , the formula is as follows:

[0045]

[0046] in is the weight of the i-th reference point, indicating its importance, ranging from (0, 1), is the coordinate of the i-th reference point, and n is the number of selected reference points.

[0047] Optionally, the state observer is a state observer based on a Kalman filter, which is used to estimate and predict the system state. The gain matrix of the state observer is The design adopts the pole configuration method. Through the pole configuration method, the closed-loop pole distribution of the system is adjusted to a specific position to optimize the dynamic performance of the observer. The introduction of the Kalman filter can improve the stability and accuracy of the signal by smoothing the current signal of the correction motor.

[0048] Optionally, the state space model captures various complex disturbance factors in the belt operation process in real time by dynamically adjusting key parameters, and then accurately characterizes the belt state. Each state space model takes physical drive as the core, uses adaptive parameter optimization strategy, and adjusts model parameters according to real-time feedback of belt operation data to make it more suitable for actual working conditions, which reflects the significant advantages of adaptive parameters in dealing with complex dynamic changes. The dynamic modeling process is as follows:

[0049] (1) The following state variables are defined:

[0050]

[0051] in, Represents the state vector of the belt, which is used to fully characterize the dynamic operation characteristics of the belt; It indicates the belt deviation obtained by observing the current of the correction motor through the state observer. The belt deviation not only provides an important basis for closed-loop control and state tracking in system modeling, but also constitutes the core feedback mechanism of correction control by performing difference calculation with the position deviation detected by vision. Through this mechanism, the system can more accurately judge the correction effect and dynamically adjust the driving amount of the correction motor, thereby effectively avoiding lag or overshoot in the correction process, greatly improving the real-time and stability of the correction control, and ensuring the efficient and accurate operation of the belt; Indicates the current offset speed of the belt, which is the derivative of the belt offset with respect to time. ,This variable describes the dynamic trend of belt offset and is a key indicator for measuring the real-time performance of system offset adjustment; It is the output current of the deviation correction motor. As the core execution variable of the drive system, it directly affects the actual movement response of the belt.

[0052] (2) From the perspective of system dynamics, the model is constructed as follows:

[0053]

[0054] in, is the equivalent mass of the belt system, is the damping coefficient between the belt and the roller, describing the friction characteristics of the belt, and the torque generated by the drive motor is , , is the torque constant of the motor, in units of , motor current Directly converted into driving torque through torque constant; external interference force of belt system The belt vibration caused by the uneven distribution of materials will generate additional disturbance force during the operation of the belt, further affecting the dynamic performance and stability of the system. The external disturbance force of the belt can be described as = The real-time measurement of the vibration acceleration s of the belt is realized by the vibration acceleration sensor, which is used to capture the dynamic disturbance caused by the uneven distribution of materials. It not only reflects the physical relationship between the belt vibration and the material distribution, but also provides accurate dynamic input support for the subsequent control strategy, ensuring that the system can achieve efficient disturbance suppression and stable control under complex working conditions. Therefore, the state equation of the belt system is:

[0055]

[0056] The output equation of the belt system is:

[0057]

[0058] in, represents the output of the system, A is the state matrix, B is the input matrix, is the interference matrix, Torque force of the drive motor Directly related, , so the output matrix is , reflects the mapping relationship between system state and output, In order to directly transfer the matrix and reflect the direct impact of the control signal on the output, the state space model of the system is further constructed by unifying the state equation with the dynamic equation, which reflects the direct impact of the control signal on the system output:

[0059] Based on the definition of the dynamic equations and state variables, the state equations for the offset velocity and motor current dynamics are derived:

[0060]

[0061] This equation reflects the direct impact of the driving signal and the intrinsic system parameters on the excursion dynamics;

[0062] Furthermore, the dynamic behavior of the motor current is expressed as:

[0063]

[0064] This equation reveals the coupling relationship between the motor current change and the system friction, belt offset speed and external disturbance force;

[0065] Convert the above equation into state space form:

[0066]

[0067] State Matrix = , the input matrix = , interference matrix = , the output matrix ,matrix The dynamic coupling relationship of the system is characterized, and adaptive parameter adjustment can be performed according to the on-site working conditions. The construction of the system state equation further reveals the direct regulatory effect of the control signal on the dynamic behavior of the system. This unified modeling helps to analyze the response characteristics of the belt system.

[0068] Optionally, the goal of the input control law is to convert the actual deviation correction amount Gradually decrease and tend to zero, so that the state of the system can gradually converge to a balanced state. Even if the system is disturbed, it can quickly achieve self-stabilization. The actual correction amount The input control law is designed by combining the state of the belt and the deviation information measured by vision. as follows:

[0069]

[0070] in is the speed feedback gain, which is used to adjust the response speed;

[0071] Input control law Substitute into the state equation:

[0072] ;

[0073] Introducing closed-loop energy function , the actual deviation correction amount Substituting the speed feedback into , energy function It decreases monotonically over time, and the system eventually converges to a stable state.

[0074] On the other hand, a multivariable adaptive belt deviation correction control system with intelligent visual perception is provided, the system comprising:

[0075] Intelligent visual perception module, used to detect the actual position of the belt;

[0076] A first comparison module is used to compare the detected actual belt position with the belt position setting value to obtain a position deviation, and input the position deviation into a PI controller to obtain an accurate position deviation control output;

[0077] A second comparison module is used to compare the position deviation control output with the belt deviation obtained by observing the current of the deviation correction motor through the state observer to obtain the actual deviation correction amount, and then obtain the input control law;

[0078] A correction signal acquisition module, used for inputting the input control law into a state space model to obtain a correction signal of the correction motor;

[0079] A deviation correction module, used for outputting a motor current according to the deviation correction signal, and calculating the torque of the deviation correction motor by a torque coefficient, so as to drive a belt deviation correction device to correct the belt;

[0080] The intelligent visual perception module is specifically used for:

[0081] Obtain multi-resolution, multi-view, and multi-state belt RGB image sequences;

[0082] Extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature map;

[0083] The feature map after adding Gaussian noise is input into the enhanced belt segmentation module, which focuses on the key areas of the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask;

[0084] The belt segmentation mask is input into a multi-domain collaborative edge perception module, which collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map;

[0085] Inputting the final belt edge graph into an intelligent linear regression module to generate a belt edge line graph;

[0086] The belt edge line diagram is input into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

[0087] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned intelligent visual perception multivariable adaptive belt deviation correction control method.

[0088] On the other hand, a computer-readable storage medium is provided, in which at least one instruction is stored. The at least one instruction is loaded and executed by a processor to implement the above-mentioned intelligent visual perception multivariable adaptive belt deviation correction control method.

[0089] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0090] 1. The introduction of random weight allocation algorithm can adjust the intensity and influence of noise according to different image features through dynamic weighting of weights, so as to adapt to the influence of multiple factors on image data in complex industrial scenes, flexibly respond to different feature areas, and avoid introducing unnecessary noise in places where features are not obvious, thereby improving the accuracy of image processing and the adaptability of the system.

[0091] 2. An enhanced belt segmentation module is designed. Through multi-scale feature extraction and fusion, and optimization and innovation of the information interaction mechanism between encoder and decoder, high robustness and high precision belt segmentation in complex industrial scenarios is achieved. Specifically, feature block division and fusion technology is used to generate multi-scale fusion feature maps, retaining fine-grained and global features; the cross-attention mechanism is used to enable the decoder to capture the deep correlation between key features in a global scope and improve its ability to understand the global context; combined with the self-attention mechanism, it focuses on the key feature area and effectively suppresses irrelevant noise interference; the processing order of the decoder is optimized, and the fusion effect of global and local features is further improved by prioritizing the execution of the cross-attention mechanism and then the self-attention mechanism.

[0092] 3. Design a multi-domain collaborative edge perception module, which can provide rich edge information through multi-level feature map output. By combining global context information, local edge features and frequency domain enhancement features, the ability to detect belt edges is further enhanced through the collaborative processing of global, local and frequency domain information. The hierarchical structure of the classification network is used to collect spatial features of different scales, and the frequency domain amplitude and phase angle information is extracted through fast Fourier transform. The efficient combination of spatial domain and frequency domain features is achieved in the fusion layer, which improves the robustness and accuracy of edge detection.

[0093] 4. An intelligent linear regression module for the belt edge is designed. By introducing a multi-scale morphological processing method, the intelligent linear regression module successfully achieves edge extraction with continuity, consistency, robustness and efficiency. Even in complex environments (such as irregular edges and uneven lighting), the belt edge features can be accurately extracted.

[0094] 5. The system innovatively realizes the deep integration of image data and multivariable coupling factors (such as belt quality, friction, vibration, etc.), builds a highly intelligent belt state perception and control mechanism, and introduces the Kalman filter algorithm. By filtering out high-frequency noise and retaining low-frequency components of the motor current signal, it significantly optimizes the smoothness and reliability of the motor output signal. It can dynamically capture the mutual influence of the belt running state and external disturbances, accurately control the dynamic characteristics of the system, and through adaptive parameter adjustment and rapid response mechanism, the system shows excellent robustness and adaptability in complex industrial environments, effectively ensuring the high precision, high stability and real-time performance of belt deviation correction control.

[0095] 6. A self-stabilizing feedback control mechanism is adopted to process the deviation between the belt position deviation detected by vision and the belt offset observed by the state observer. The system dynamically adjusts the control parameters according to complex working conditions in real time to achieve adaptive control capability. Through real-time monitoring of the belt status, the system can automatically optimize the correction strategy and quickly reach a stable state according to the input control law, effectively avoiding correction errors or delays caused by incomplete image signals or complex working conditions, thereby ensuring that the system performs correction control efficiently and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0097] Figure 1 This is a flow chart of a multivariable adaptive belt deviation correction control method of intelligent visual perception provided by an embodiment of the present invention;

[0098] Figure 2 It is an overall block diagram of a multivariable adaptive belt deviation correction control method of intelligent visual perception provided by an embodiment of the present invention;

[0099] Figure 3 It is a structural block diagram of the overall module of intelligent visual perception provided by an embodiment of the present invention;

[0100] Figure 4 is a structural block diagram of an enhanced belt segmentation module provided by an embodiment of the present invention;

[0101] Figure 5 It is a structural block diagram of a multi-domain collaborative edge perception module provided by an embodiment of the present invention;

[0102] Figure 6 It is a structural block diagram of an intelligent linear regression module provided by an embodiment of the present invention;

[0103] Figure 7 It is a structural block diagram of a belt edge reference point determination module provided by an embodiment of the present invention;

[0104] Figure 8 is a structural block diagram of a self-stabilizing system composed of a state observer and a state space model provided by an embodiment of the present invention;

[0105] Fig. 9 This is a block diagram of a multivariable adaptive belt deviation correction control system with intelligent visual perception provided by an embodiment of the present invention;

[0106] Fig.10 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0107] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0108] An embodiment of the present invention provides a multivariable adaptive belt deviation correction control method based on intelligent visual perception. The method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 A flowchart of the method is shown. Figure 2 The overall block diagram of the method is shown, and the method may include the following steps:

[0109] S1, detect the actual position of the belt through the intelligent visual perception module;

[0110] S2, comparing the detected actual belt position with the belt position setting value to obtain a position deviation, and inputting the position deviation into a PI controller to obtain an accurate position deviation control output;

[0111] S3, comparing the position deviation control output with the belt deviation obtained by observing the current of the deviation correction motor through a state observer, obtaining an actual deviation correction amount, and then obtaining an input control law;

[0112] S4, inputting the input control law into a state space model to obtain a correction signal of the correction motor;

[0113] S5, outputting a motor current according to the deviation correction signal, and obtaining the torque of the deviation correction motor by calculating the torque coefficient, so as to drive a belt deviation correction device to correct the belt;

[0114] The S1 specifically includes:

[0115] S11, obtaining a multi-resolution, multi-view, and multi-state belt RGB image sequence;

[0116] The embodiment of the present invention obtains a multi-resolution, multi-view, and multi-state RGB image sequence, captures the macroscopic perspective and microscopic details of the belt state, provides detailed image information from the global to the local, and realizes multi-view capture from a wide view to a detailed focus by adjusting the zoom step of the camera. In addition, special image acquisition is performed for different running states of the belt, including normal, offset, and severe offset states.

[0117] S12, extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature graph, such as Figure 3 As shown;

[0118] The embodiment of the present invention uses a random weight allocation algorithm to dynamically adjust the weights to adaptively adjust the intensity and impact of noise according to different image features, rather than directly adding noise to the image itself. This has the advantage of being able to more accurately process belt image features and avoid introducing unnecessary noise in places where features are not obvious, thereby improving the accuracy of feature recognition and the overall performance of the system.

[0119] Optionally, the S12 specifically includes:

[0120] S121, organizing the belt RGB image sequence into an image sequence matrix according to state changes;

[0121] S122, creating a weight matrix of the same size as the image sequence matrix , using standard Gaussian distribution to generate the weight matrix Elements in And normalize it;

[0122] S123, use CNN convolutional neural network to extract the feature matrix of each frame image , according to the feature matrix Elements , adjust the weight matrix Elements in , the formula is as follows:

[0123]

[0124] in, is the adjustment coefficient, which controls the influence of the feature on the weight. Perform normalization;

[0125] S124, each frame of the image generates a Gaussian noise matrix with the same size as the image ,Will Elements in After normalization Do the inner product to get the weighted noise matrix for each frame of the image , the formula is as follows:

[0126]

[0127] S125, the weighted noise matrix Superimposed onto the feature matrix The final dynamic weighted feature matrix after adding noise is obtained , the formula is as follows:

[0128] .

[0129] S13, inputting the feature map after adding Gaussian noise into the enhanced belt segmentation module, wherein the enhanced belt segmentation module focuses on the key area of ​​the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask;

[0130] The enhanced belt segmentation module of the embodiment of the present invention is designed to adapt to the influence of complex environments such as excessive dust concentration and insufficient lighting conditions on the image acquisition quality in industrial sites, so as to effectively focus on the key areas of the image, suppress irrelevant noise interference, and improve the robustness and efficiency of feature extraction. The output result is belt mask segmentation, wherein multi-scale feature extraction and fusion, by extracting multi-scale and multi-level features from the image matrix sequence, enhances the model's ability to capture key features, and the decoder effectively constructs a rich multi-dimensional feature representation by adaptively processing image features at different resolutions and fields of view. To further improve the segmentation accuracy, the module divides features of different scales into independent patches for feature fusion to generate a comprehensive feature map. This method not only retains the feature information of the image at each resolution, but also reveals the implicit relationship between multi-scale features, and improves the model's sensitivity to local and global features. In addition, by constructing a spatial context-aware segmentation network and introducing a multi-level feature decoder, the cross-attention mechanism and adaptive attention mechanism can deeply understand the spatial context relationship of the image based on the feature representation. Finally, the network achieves the segmentation of the belt area with high accuracy.

[0131] Alternatively, if Figure 4 As shown, the S13 specifically includes:

[0132] S131, using a CNN convolutional neural network as a feature extractor, extracting multi-scale features from the feature map after adding noise, and obtaining a multi-scale feature map;

[0133] S132, dividing the multi-scale feature map into a series of feature blocks (instead of directly dividing the original image as in the traditional method, which makes full use of the rich feature information extracted by the convolutional neural network, thereby capturing more fine-grained and global feature representations on belt images of different scales), and fusing these feature blocks to form a multi-scale fused feature map;

[0134] S133, input the multi-scale fusion feature map into the feature encoder, the feature encoder introduces position coding on the basis of the multi-scale fusion feature map, identifies the spatial position of the feature, and combines it into a feature vector containing position information. The feature vector passes through the Linear layer to generate three vectors Q, K, and V. Q, K, and V are the core feature vectors used for the attention mechanism in deep learning, which are used to measure the correlation between features. Then, the correlation between Q and K is calculated through the self-attention mechanism to obtain a correlation matrix. Then, the correlation matrix is ​​used to perform weighted summation on the V vector. The weight coefficient of each V vector is the correlation calculated by Q and K, and a new set of Q', K', and V' vectors containing the context information of all input features is obtained, wherein Q' is not passed to the cross-attention mechanism in the decoder, and only K' and V' are passed to the cross-attention mechanism in the decoder. The decoder has a total of N layers. The input of the current layer decoder comes from the Q'' vector generated in the previous N-1 layer decoder. The Q'' vector of the first layer decoder comes from an initialization vector, which represents the initial state of the decoder. The current layer decoder will be based on the input Q'' Combined with K' and V' passed from the encoder of the current layer, the cross-attention mechanism is completed to generate higher-quality feature vectors. These higher-quality feature vectors are used to interact with the decoder of the next layer, and then pass through the feedforward network and normalization. The feedforward network can introduce more complex transformations to the features of each position, enhancing the model's ability to express input data. The purpose of normalization is to normalize the output of each layer. Through normalization, the output vector of each layer is rescaled and offset to make its mean 0 and variance 1, and a new set of Q''', K''', and V''' is output. The obtained Q''', K''', and V''' are calculated through the self-attention mechanism to obtain a new feature vector Imformation;

[0135] In this way, the system can capture the relationship between features in a global and local range, which enables the system to accurately focus on different areas of the belt and establish deep associations between the feature areas in the image, so as to more accurately identify and locate the state and position of the belt. In addition, in the enhanced belt segmentation of the embodiment of the present invention, the decoder design adopts an innovative processing order: first apply the cross-attention mechanism, and then introduce the self-attention mechanism. Different from the order of using the self-attention mechanism first and then the cross-attention mechanism in the traditional decoder, this adjustment enables the decoder to first extract richer cross information from the encoder output and the target query, thereby providing a more comprehensive global contextual understanding for the belt correction system. This method helps to finely extract and accurately represent the belt features, significantly enhancing the adaptability of the system in complex industrial environments, especially in the face of working conditions such as lighting changes, dust pollution and material occlusion. Through this innovative processing sequence, the improved decoder significantly improves the perception and response capabilities of multivariate interference states, strengthens the real-time monitoring and analysis of the belt running status, and ensures that the system can still maintain efficient and accurate correction performance under harsh working conditions.

[0136] S134, passing the new feature vector Imformation through the multi-layer perceptron and segmentation head of the head network to generate the belt segmentation mask.

[0137] Through this processing step, the hidden layer features are effectively mapped to the belt segmentation mask, generating classification information for each pixel, thereby ensuring that the system can accurately identify and segment the belt edge and its state. This process ensures high-quality segmentation results and provides accurate visual input for subsequent correction control.

[0138] S14, inputting the belt segmentation mask into a multi-domain collaborative edge perception module, the multi-domain collaborative edge perception module collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map;

[0139] Alternatively, if Figure 5 As shown, the S14 specifically includes:

[0140] S141, extracting spatial domain features from the binarized belt segmentation mask. The spatial domain feature extraction extracts multi-scale edge features through a backbone network. The sizes of each layer of feature maps output by the network are: [224, 224, 64], [112, 112, 128], [56, 56, 256], and [28, 28, 512], respectively. Each layer of features generates an edge map through a sigmoid activation function, and is restored to the original image size of [224, 224] by upsampling, so as to obtain four edge feature maps of the same size.

[0141] S142, performing fast Fourier transform on the binarized belt segmentation mask to generate a frequency domain feature map (amplitude map and phase angle map), using a flexible high-pass filter in the frequency domain to extract high-frequency components to enhance edge features, and regressing to the spatial domain through inverse Fourier transform to obtain a frequency domain enhanced edge feature map;

[0142] S143, fusing the four edge feature maps with the frequency domain enhanced edge feature map to obtain a final belt edge map.

[0143] S15, inputting the final belt edge map into an intelligent linear regression module to generate a belt edge line map;

[0144] Alternatively, if Figure 6 As shown, the S15 specifically includes:

[0145] S151, the final belt edge image is downsampled, the resolution is gradually reduced, the edge features of the belt are extracted, and the data dimension is reduced to improve the calculation efficiency;

[0146] S512, entering the morphological processing stage, performing dilation and corrosion operations on the edge features to remove isolated noise points and enhance the continuity of the edge lines to ensure that the edge features are clear;

[0147] S152, upsampling the feature map after morphological processing to the original image size, and fusing the original images obtained by upsampling at different scales through a fusion layer;

[0148] S153, smoothing the fused feature map to obtain a smoothed feature map, then traversing the extracted edge mutation pixel points through gradient calculation, regressing these edge mutation pixel points using a straight line fitting algorithm, and fitting a straight line representing the edge of the belt;

[0149] S154, upsampling the low-resolution straight line fitting result and gradually restoring it to a size consistent with the original input, thereby generating a clear and continuous line diagram of the belt edge.

[0150] S16, inputting the belt edge line diagram into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

[0151] Alternatively, if Figure 7 As shown, the S16 specifically includes:

[0152] Adaptively select n reference point coordinates of the belt edge and combine them with weighted calculation. The weight design fully considers the importance of the reference point and its local characteristics, making the algorithm more adaptable and anti-interference to subtle changes in the edge, and obtaining the final reference point coordinates , the formula is as follows:

[0153]

[0154] in is the weight of the i-th reference point, indicating its importance, ranging from (0, 1), is the coordinate of the i-th reference point, and n is the number of selected reference points.

[0155] The embodiment of the present invention calculates the position deviation of the belt by comparing these final reference points with the set reference reference points (belt position setting value), and inputs the position deviation into the PI controller to obtain a smoother and more accurate belt position deviation control output after filtering (because the calculated position deviation is affected by complex working conditions and is often inaccurate, it needs to be filtered by the PI controller to obtain the accurate position deviation) for subsequent belt deviation correction control.

[0156] Alternatively, if Figure 8 As shown, the state observer is a state observer based on a Kalman filter, which is used to estimate and predict the system state (the observer is designed to combine the measured output of the system and the model prediction value, effectively suppress noise interference, and provide high-precision state feedback information for the implementation of the control strategy). The gain matrix of the state observer is The design adopts the pole configuration method. Through the pole configuration method, the closed-loop pole distribution of the system is adjusted to a specific position to optimize the dynamic performance of the observer. The introduction of the Kalman filter can improve the stability and accuracy of the signal by smoothing the current signal of the correction motor (through this smoothing process, the Kalman filter can provide a more accurate basis for the prediction of the belt state and reduce the uncertainty caused by the current signal fluctuation, thereby ensuring the smoothness and reliability of the state estimation).

[0157] Optionally, the state space model captures various complex disturbance factors in the belt operation process in real time by dynamically adjusting key parameters, and then accurately characterizes the belt state. Each state space model takes physical drive as the core, uses adaptive parameter optimization strategy, and adjusts model parameters according to real-time feedback of belt operation data to make it more suitable for actual working conditions, which reflects the significant advantages of adaptive parameters in dealing with complex dynamic changes. The dynamic modeling process is as follows:

[0158] (1) The following state variables are defined:

[0159] ;

[0160] in, Represents the state vector of the belt, which is used to fully characterize the dynamic operation characteristics of the belt; It indicates the belt deviation obtained by observing the current of the correction motor through the state observer. The belt deviation not only provides an important basis for closed-loop control and state tracking in system modeling, but also constitutes the core feedback mechanism of correction control by performing difference calculation with the position deviation detected by vision. Through this mechanism, the system can more accurately judge the correction effect and dynamically adjust the driving amount of the correction motor, thereby effectively avoiding lag or overshoot in the correction process, greatly improving the real-time and stability of the correction control, and ensuring the efficient and accurate operation of the belt; Indicates the current offset speed of the belt, which is the derivative of the belt offset with respect to time. ,This variable describes the dynamic trend of belt offset and is a key indicator for measuring the real-time performance of system offset adjustment; It is the output current of the deviation correction motor. As the core execution variable of the drive system, it directly affects the actual movement response of the belt.

[0161] (2) From the perspective of system dynamics, the model is constructed as follows:

[0162]

[0163] in, is the equivalent mass of the belt system, is the damping coefficient between the belt and the roller, describing the friction characteristics of the belt, and the torque generated by the drive motor is , , is the torque constant of the motor, in units of , motor current Directly converted into driving torque through torque constant; external interference force of belt system The belt vibration caused by the uneven distribution of materials will generate additional disturbance force during the operation of the belt, further affecting the dynamic performance and stability of the system. The external disturbance force of the belt can be described as = The real-time measurement of the vibration acceleration s of the belt is realized by the vibration acceleration sensor, which is used to capture the dynamic disturbance caused by the uneven distribution of materials. It not only reflects the physical relationship between the belt vibration and the material distribution, but also provides accurate dynamic input support for the subsequent control strategy, ensuring that the system can achieve efficient disturbance suppression and stable control under complex working conditions. Therefore, the state equation of the belt system is:

[0164]

[0165] The output equation of the belt system is:

[0166]

[0167] in, represents the output of the system, A is the state matrix, B is the input matrix, is the interference matrix, Torque force of the drive motor Directly related, , so the output matrix is , reflects the mapping relationship between system state and output, In order to directly transfer the matrix and reflect the direct impact of the control signal on the output, the state space model of the system is further constructed by unifying the state equation with the dynamic equation, which reflects the direct impact of the control signal on the system output:

[0168] Based on the definition of the dynamic equations and state variables, the state equations for the offset velocity and motor current dynamics are derived:

[0169]

[0170] This equation reflects the direct impact of the driving signal and the intrinsic system parameters on the excursion dynamics;

[0171] Furthermore, the dynamic behavior of the motor current is expressed as:

[0172]

[0173] This equation reveals the coupling relationship between the motor current change and the system friction, belt offset speed and external disturbance force;

[0174] Convert the above equation into state space form:

[0175] ;

[0176] Among them, the state matrix = , the input matrix = , interference matrix = , the output matrix ,matrix The dynamic coupling relationship of the system is characterized, and adaptive parameter adjustment can be performed according to the on-site working conditions. The construction of the system state equation further reveals the direct regulatory effect of the control signal on the dynamic behavior of the system. This unified modeling helps to analyze the response characteristics of the belt system.

[0177] Optionally, the goal of the input control law is to convert the actual deviation correction amount Gradually decrease and tend to zero, so that the state of the system can gradually converge to a balanced state. Even if the system is disturbed, it can quickly achieve self-stabilization. The actual correction amount The input control law is designed by combining the state of the belt and the deviation information measured by vision. as follows:

[0178]

[0179] in is the speed feedback gain, which is used to adjust the response speed;

[0180] Input control law Substitute into the state equation:

[0181]

[0182] Introducing closed-loop energy function , the actual deviation correction amount Substituting the speed feedback into , energy function It decreases monotonically over time, and the system eventually converges to a stable state.

[0183] Through the above steps, the state space model and state observer are combined, and the optimized input control law is designed to construct a highly intelligent self-stabilizing control system, such as Figure 8 As shown in the figure, the system takes physical drive and adaptive parameter adjustment as the core, successfully realizes high-precision modeling and real-time stability guarantee of the multi-variable coupled belt correction system, and effectively improves the adaptability and control accuracy of the system under complex working conditions.

[0184] like Fig. 9 As shown, an embodiment of the present invention further provides a multivariable adaptive belt deviation correction control system with intelligent visual perception, the system comprising:

[0185] The intelligent visual perception module 910 is used to detect the actual position of the belt;

[0186] A first comparison module 920 is used to compare the detected actual belt position with the belt position setting value to obtain a position deviation, and input the position deviation into a PI controller to obtain an accurate position deviation control output;

[0187] A second comparison module 930 is used to compare the position deviation control output with the belt deviation obtained by observing the current of the correction motor through the state observer to obtain the actual correction amount, and then obtain the input control law;

[0188] A correction signal acquisition module 940 is used to input the input control law into a state space model to obtain a correction signal of the correction motor;

[0189] The deviation correction module 950 is used to output the motor current according to the deviation correction signal, calculate the torque of the deviation correction motor by the torque coefficient, and drive the belt deviation correction device to correct the belt;

[0190] The intelligent visual perception module is specifically used for:

[0191] Obtain multi-resolution, multi-view, and multi-state belt RGB image sequences;

[0192] Extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature map;

[0193] The feature map after adding Gaussian noise is input into the enhanced belt segmentation module, which focuses on the key areas of the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask;

[0194] The belt segmentation mask is input into a multi-domain collaborative edge perception module, which collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map;

[0195] Inputting the final belt edge graph into an intelligent linear regression module to generate a belt edge line graph;

[0196] The belt edge line diagram is input into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

[0197] An embodiment of the present invention provides a multivariable adaptive belt deviation correction control system with intelligent visual perception, and its functional structure corresponds to a multivariable adaptive belt deviation correction control method with intelligent visual perception provided by an embodiment of the present invention, which will not be repeated here.

[0198] Fig.10 It is a structural diagram of an electronic device 1000 provided in an embodiment of the present invention. The electronic device 1000 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 1001 and one or more memories 1002, wherein at least one instruction is stored in the memory 1002, and the at least one instruction is loaded and executed by the processor 1001 to implement the steps of the above-mentioned intelligent visual perception multivariable adaptive belt deviation correction control method.

[0199] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, which can be executed by a processor in a terminal to complete the multivariable adaptive belt deviation control method of intelligent visual perception. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0200] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0201] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multivariable adaptive belt deviation control method based on intelligent visual perception, characterized in that: The method comprises: S1, detect the actual position of the belt through the intelligent visual perception module; S2, comparing the detected actual belt position with the belt position setting value to obtain a position deviation, and inputting the position deviation into a PI controller to obtain an accurate position deviation control output; S3, comparing the position deviation control output with the belt deviation obtained by observing the current of the deviation correction motor through a state observer, obtaining an actual deviation correction amount, and then obtaining an input control law; S4, inputting the input control law into a state space model to obtain a correction signal of the correction motor; S5, outputting a motor current according to the deviation correction signal, and obtaining the torque of the deviation correction motor by calculating the torque coefficient, so as to drive a belt deviation correction device to correct the belt; The S1 specifically includes: S11, obtaining a multi-resolution, multi-view, and multi-state belt RGB image sequence; S12, extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature graph; S13, inputting the feature map after adding Gaussian noise into the enhanced belt segmentation module, wherein the enhanced belt segmentation module focuses on the key area of ​​the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask; S14, inputting the belt segmentation mask into a multi-domain collaborative edge perception module, the multi-domain collaborative edge perception module collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map; S15, inputting the final belt edge map into an intelligent linear regression module to generate a belt edge line map; S16, inputting the belt edge line diagram into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

2. The method according to claim 1, characterized in that: The S12 specifically includes: S121, organizing the belt RGB image sequence into an image sequence matrix according to state changes; S122, creating a weight matrix of the same size as the image sequence matrix , using standard Gaussian distribution to generate the weight matrix Elements in And normalize it; S123, use CNN convolutional neural network to extract the feature matrix of each frame image , according to the feature matrix Elements , adjust the weight matrix Elements in , the formula is as follows: ； in, is the adjustment coefficient, which controls the influence of the feature on the weight. Perform normalization; S124, each frame of the image generates a Gaussian noise matrix with the same size as the image ,Will Elements in After normalization Do the inner product to get the weighted noise matrix for each frame of the image , the formula is as follows: ； S125, the weighted noise matrix Superimposed onto the feature matrix The final dynamic weighted feature matrix after adding noise is obtained , the formula is as follows: 。 3. The method according to claim 1, characterized in that The S13 specifically includes: S131, using a CNN convolutional neural network as a feature extractor, extracting multi-scale features from the feature map after adding noise, and obtaining a multi-scale feature map; S132, dividing the multi-scale feature map into a series of feature blocks, and fusing the feature blocks to form a multi-scale fused feature map; S133, input the multi-scale fusion feature map into the feature encoder, the feature encoder introduces position coding on the basis of the multi-scale fusion feature map, identifies the spatial position of the feature, and combines it into a feature vector containing position information. The feature vector passes through the Linear layer to generate three vectors Q, K, and V. Q, K, and V are the core feature vectors used for the attention mechanism in deep learning, which are used to measure the correlation between features. Then, the correlation between Q and K is calculated through the self-attention mechanism to obtain a correlation matrix. Then, the correlation matrix is ​​used to perform weighted summation on the V vector. The weight coefficient of each V vector is the correlation calculated by Q and K, and a new set of Q', K', and V' vectors containing the context information of all input features is obtained, wherein Q' is not passed to the cross-attention mechanism in the decoder, and only K' and V' are passed to the cross-attention mechanism in the decoder. The decoder has a total of N layers. The input of the current layer decoder comes from the Q'' vector generated in the previous N-1 layer decoder. The Q'' vector of the first layer decoder comes from an initialization vector, which represents the initial state of the decoder. The current layer decoder will be based on the input Q'' Combined with K' and V' passed from the encoder of the current layer, the cross-attention mechanism is completed to generate higher-quality feature vectors. These higher-quality feature vectors are used to interact with the decoder of the next layer, and then pass through the feedforward network and normalization. The feedforward network can introduce more complex transformations to the features of each position, enhancing the model's ability to express input data. The purpose of normalization is to normalize the output of each layer. Through normalization, the output vector of each layer is rescaled and offset to make its mean 0 and variance 1, and a new set of Q''', K''', and V''' is output. The obtained Q''', K''', and V''' are calculated through the self-attention mechanism to obtain a new feature vector Imformation; S134, passing the new feature vector Imformation through the multi-layer perceptron and segmentation head of the head network to generate the belt segmentation mask.

4. The method according to claim 1, characterized in that The S14 specifically includes: S141, extracting spatial domain features from the binarized belt segmentation mask, extracting multi-scale edge features through a backbone network, and outputting feature maps of each layer with sizes of [224, 224, 64], [112, 112, 128], [56, 56, 256], and [28, 28, 512], generating edge maps for each layer of features through a sigmoid activation function, and restoring the original image size through upsampling to obtain four edge feature maps of the same size; S142, performing fast Fourier transform on the binarized belt segmentation mask to generate a frequency domain feature map, using a flexible high-pass filter in the frequency domain to extract high-frequency components to enhance edge features, and regressing to the spatial domain through inverse Fourier transform to obtain a frequency domain enhanced edge feature map; S143, fusing the four edge feature maps with the frequency domain enhanced edge feature map to obtain a final belt edge map.

5. The method according to claim 1, characterized in that The S15 specifically includes: S151, the final belt edge image is downsampled, the resolution is gradually reduced, the edge features of the belt are extracted, and the data dimension is reduced to improve the calculation efficiency; S512, entering the morphological processing stage, performing dilation and corrosion operations on the edge features to remove isolated noise points and enhance the continuity of the edge lines to ensure that the edge features are clear; S152, upsampling the feature map after morphological processing to the original image size, and fusing the original images obtained by upsampling at different scales through a fusion layer; S153, smoothing the fused feature map to obtain a smoothed feature map, then traversing the extracted edge mutation pixel points through gradient calculation, regressing these edge mutation pixel points using a straight line fitting algorithm, and fitting a straight line representing the edge of the belt; S154, upsampling the low-resolution straight line fitting result and gradually restoring it to a size consistent with the original input, thereby generating a clear and continuous line diagram of the belt edge.

6. The method according to claim 1, characterized in that The S16 specifically includes: Adaptively select n reference point coordinates of the belt edge and combine them with weighted calculation. The weight design fully considers the importance of the reference point and its local characteristics, making the algorithm more adaptable and anti-interference to subtle changes in the edge, and obtaining the final reference point coordinates , the formula is as follows: ； in is the weight of the i-th reference point, indicating its importance, ranging from (0, 1), is the coordinate of the i-th reference point, and n is the number of selected reference points.

7. The method according to claim 1, characterized in that The state observer is a state observer based on a Kalman filter, which is used to estimate and predict the system state. The gain matrix of the state observer is The design adopts the pole configuration method. Through the pole configuration method, the closed-loop pole distribution of the system is adjusted to a specific position to optimize the dynamic performance of the observer. The introduction of the Kalman filter can improve the stability and accuracy of the signal by smoothing the current signal of the correction motor.

8. The method according to claim 1, characterized in that: The state space model captures various complex disturbance factors in the belt operation process in real time by dynamically adjusting key parameters, and then accurately characterizes the belt state. Each state space model takes physical drive as the core, uses adaptive parameter optimization strategy, and adjusts model parameters according to real-time feedback of belt operation data to make it more suitable for actual working conditions, which reflects the significant advantages of adaptive parameters in dealing with complex dynamic changes. The dynamic modeling process is as follows: (1) The following state variables are defined: ； in, Represents the state vector of the belt, which is used to fully characterize the dynamic operation characteristics of the belt; It indicates the belt deviation obtained by observing the current of the correction motor through the state observer. The belt deviation not only provides an important basis for closed-loop control and state tracking in system modeling, but also constitutes the core feedback mechanism of correction control by performing difference calculation with the position deviation detected by vision. Through this mechanism, the system can more accurately judge the correction effect and dynamically adjust the driving amount of the correction motor, thereby effectively avoiding lag or overshoot in the correction process, greatly improving the real-time and stability of the correction control, and ensuring the efficient and accurate operation of the belt; Indicates the current offset speed of the belt, which is the derivative of the belt offset with respect to time. ,This variable describes the dynamic trend of belt offset and is a key indicator for measuring the real-time performance of system offset adjustment; It is the output current of the deviation correction motor. As the core execution variable of the drive system, it directly affects the actual movement response of the belt. (2) From the perspective of system dynamics, the model is constructed as follows: ； in, is the equivalent mass of the belt system, is the damping coefficient between the belt and the roller, describing the friction characteristics of the belt, and the torque generated by the drive motor is , , is the torque constant of the motor, in units of , motor current Directly converted into driving torque through torque constant; external interference force of belt system The belt vibration caused by the uneven distribution of materials will generate additional disturbance force during the operation of the belt, further affecting the dynamic performance and stability of the system. The external disturbance force of the belt can be described as = The real-time measurement of the vibration acceleration s of the belt is realized by the vibration acceleration sensor, which is used to capture the dynamic disturbance caused by the uneven distribution of materials. It not only reflects the physical relationship between the belt vibration and the material distribution, but also provides accurate dynamic input support for the subsequent control strategy, ensuring that the system can achieve efficient disturbance suppression and stable control under complex working conditions. Therefore, the state equation of the belt system is: ； The output equation of the belt system is: ； in, represents the output of the system, A is the state matrix, B is the input matrix, is the interference matrix, is the input control law, Torque force of the drive motor Directly related, , so the output matrix is , reflects the mapping relationship between system state and output, In order to directly transfer the matrix and reflect the direct impact of the control signal on the output, the state space model of the system is further constructed by unifying the state equation with the dynamic equation, which reflects the direct impact of the control signal on the system output: Based on the definition of the dynamic equations and state variables, the state equations for the offset velocity and motor current dynamics are derived: ； This equation reflects the direct impact of the driving signal and the intrinsic system parameters on the excursion dynamics; Furthermore, the dynamic behavior of the motor current is expressed as: ； This equation reveals the coupling relationship between the motor current change and the system friction, belt offset speed and external disturbance force; Convert the above equation into state space form: ； Among them, the state matrix = , the input matrix = , interference matrix = , the output matrix ,matrix The dynamic coupling relationship of the system is characterized, and adaptive parameter adjustment can be performed according to the on-site working conditions. The construction of the system state equation further reveals the direct regulatory effect of the control signal on the dynamic behavior of the system. This unified modeling helps to analyze the response characteristics of the belt system.

9. The method according to claim 8, characterized in that The input control law The goal is to convert the actual deviation amount Gradually decrease and tend to zero, so that the state of the system can gradually converge to a balanced state. Even if the system is disturbed, it can quickly achieve self-stabilization. The actual correction amount The input control law is obtained by the difference between the position deviation control output and the belt offset. The design combines the belt status and the deviation information measured by vision to design the input control law as follows: ； in is the speed feedback gain, which is used to adjust the response speed; Input control law Substitute into the state equation: ； Introducing closed-loop energy function , the actual deviation correction amount Substituting the speed feedback into , energy function It decreases monotonically over time, and the system eventually converges to a stable state.

10. An intelligent visual perception multivariable adaptive belt deviation correction control system, characterized in that: The system comprises: Intelligent visual perception module, used to detect the actual position of the belt; A first comparison module is used to compare the detected actual belt position with the belt position setting value to obtain a position deviation, and input the position deviation into a PI controller to obtain an accurate position deviation control output; A second comparison module is used to compare the position deviation control output with the belt deviation obtained by observing the current of the deviation correction motor through the state observer to obtain the actual deviation correction amount, and then obtain the input control law; A correction signal acquisition module, used for inputting the input control law into a state space model to obtain a correction signal of the correction motor; A deviation correction module, used for outputting a motor current according to the deviation correction signal, and calculating the torque of the deviation correction motor by a torque coefficient, so as to drive a belt deviation correction device to correct the belt; The intelligent visual perception module is specifically used for: Obtain multi-resolution, multi-view, and multi-state belt RGB image sequences; Extracting features from the belt RGB image sequence, and dynamically and randomly weighting and adding Gaussian noise to the extracted feature map; The feature map after adding Gaussian noise is input into the enhanced belt segmentation module, which focuses on the key areas of the image, deeply understands the spatial context relationship of the image, and outputs the belt segmentation mask; The belt segmentation mask is input into a multi-domain collaborative edge perception module, which collaboratively extracts multi-level edge features of the image through spatial domain feature extraction and frequency domain enhancement processing, and outputs a final belt edge map; Inputting the final belt edge graph into an intelligent linear regression module to generate a belt edge line graph; The belt edge line diagram is input into a belt edge reference point determination module to obtain final reference point coordinates as the detected actual belt position.

Images