Conveyor belt deviation correction method and system based on embodied perception and bidirectional diffusion prediction

Abstract

The application provides a conveying belt deviation correction method and system based on embodiment perception and bidirectional diffusion prediction, and belongs to the technical field of conveying belt deviation correction and embodiment intelligence. Multi-dimensional data such as deviation, tension, load, vibration and image are collected through a multi-modal sensor, the running state is comprehensively described in combination with numerical and latent space image features, an expert knowledge base is constructed to provide deviation correction decision rules, and embodiment cognition of the conveying belt state is realized comprehensively and accurately. Numerical time series and latent space image diffusion models are used to predict future features respectively, the results are corrected through bidirectional consistency constraints, the deviation trend is predicted in advance and perceived accurately, the historical and predicted states are fused in the time layer, expert correction is introduced to improve the robustness, and multi-carrier roller global collaborative optimization is realized in the space layer, forming a closed-loop decision and execution, effectively avoiding control conflicts and over-correction. The application can realize high-precision forward prediction and adaptive collaborative correction of the conveying belt deviation in a complex industrial environment.

Description

Technical Field

[0001] This invention belongs to the field of conveyor belt correction and embodied intelligence technology, and specifically refers to a conveyor belt correction method and system based on embodied perception and bidirectional diffusion prediction. Background Technology

[0002] Belt conveyors are widely used continuous conveying equipment in industrial fields such as mining, metallurgy, power, and ports. Their operational stability directly affects production efficiency, equipment lifespan, and on-site safety. During actual operation, due to the combined effects of various factors such as uneven material distribution, belt tension fluctuations, equipment wear and aging, and complex environmental disturbances, conveyor belts are prone to deviation. If this deviation continues, it can cause abnormal wear, tearing, and even structural damage to the conveyor belt edges. It can also lead to material spillage, equipment downtime, and cascading failures, resulting in significant economic losses and safety risks.

[0003] With the development of industrial intelligence, traditional conveyor belt correction methods that rely on single-point detection, threshold alarms, and local actuator responses are no longer sufficient to meet the demands for high-precision, adaptive, and continuous stable control under complex operating conditions. Especially in industrial environments with long distances, heavy loads, and frequent dynamic disturbances, the conveyor belt's operating status exhibits significant multi-factor coupling and spatial correlation characteristics. Relying solely on a single sensor makes it difficult to comprehensively characterize the overall system operation mechanism. Furthermore, traditional methods fail to utilize deviation trend prediction results, allowing correction only after deviation occurs, hindering proactive control. In terms of correction decision-making, existing methods often employ independent adjustment of single idlers or local control strategies, lacking multi-idler group control mechanisms. When conveyor belt deviation is detected, only individual idlers are adjusted, easily leading to control conflicts or system oscillations, thus affecting equipment stability and increasing mechanical wear. Summary of the Invention

[0004] To address the technical problems existing in the prior art, the present invention provides a conveyor belt correction method and system based on embodied sensing and bidirectional diffusion prediction, the technical solution of which is as follows: On the one hand, a conveyor belt correction method based on embodied perception and bidirectional diffusion prediction is provided, the method comprising: S1. Collect multi-dimensional data on the operating status of the conveyor belt through a multi-modal sensor. The multi-dimensional data includes various physical parameter data and image data. S2. By performing physical parameter modeling on the various physical parameter data, numerical sensor features are constructed. By performing latent space image modeling on the image data, latent space image features are constructed. The numerical sensor features and latent space image features are then combined to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. S3. The numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework uses the historical numerical sensor feature sequence of the conveyor belt as input to predict the numerical sensor feature sequence in the future period and obtain the numerical trend of conveyor belt deviation. At the same time, combined with the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework, the latent space image feature sequence of the conveyor belt is predicted in the future period using the historical latent space image feature sequence of the conveyor belt as input to obtain the information on the change of conveyor belt image. S4. At the time level, by combining the historical operating state feature sequence perceived by multimodal sensors with the predicted future operating state feature sequence, a single idler roller full-time state feature is constructed. A confidence-driven expert correction mechanism and cross-entropy online optimization are introduced to correct and optimize the full-time state features with low confidence. S5. After completing the full-time state modeling and expert correction of a single idler, the multi-idler spatial group control collaborative decision-making algorithm is used to make collaborative optimization decisions for all correction idlers at the spatial level, so as to realize the overall collaborative correction of the conveyor belt deviation and obtain the optimal correction roller control quantity for this control cycle. S6. Control the movement of the correction roller using the optimal correction roller control quantity, and update the decision for the next control cycle through status feedback to form a complete closed loop.

[0005] On the other hand, a conveyor belt correction system based on embodied perception and bidirectional diffusion prediction is provided, the system comprising: The acquisition module is used to acquire multidimensional data on the operating status of the conveyor belt through a multimodal sensor. The multidimensional data includes various physical parameter data and image data. The modeling module is used to construct numerical sensor features by performing physical parameter modeling on the various physical parameter data, construct latent space image features by performing latent space image modeling on the image data, and stitch the numerical sensor features and latent space image features together to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. The prediction module is used to use the numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the numerical sensor feature sequence of the conveyor belt in the future period, and obtain the numerical trend of the conveyor belt deviation change. At the same time, it combines the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the latent space image feature sequence of the conveyor belt in the future period, and obtain the information on the change of the conveyor belt image. The module is used to construct the all-time state features of a single idler roller by combining the historical operating state feature sequences sensed by multimodal sensors with the predicted future operating state feature sequences at the time level. It also introduces a confidence-driven expert correction mechanism and cross-entropy online optimization to correct and optimize the all-time state features with low confidence. The collaborative optimization module is used to perform collaborative optimization decisions on all correction rollers at the spatial level through a multi-roller spatial group control collaborative decision-making algorithm after completing the full-time state modeling and expert correction of a single roller. This enables the overall collaborative correction of conveyor belt deviation and obtains the optimal correction roller control quantity for the current control cycle. The control module is used to control the movement of the correction roller with the optimal correction roller control quantity, and to update the decision for the next control cycle through status feedback, forming a complete closed loop.

[0006] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described conveyor belt correction method based on embodied perception and bidirectional diffusion prediction.

[0007] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement the above-described conveyor belt correction method based on embodied perception and bidirectional diffusion prediction.

[0008] The beneficial effects of the technical solution provided by this invention include at least the following: This invention significantly improves the monitoring and control level of conveyor belt deviation by constructing a conveyor belt deviation correction method and system that combines multimodal embodied perception, bidirectional diffusion trend prediction, and expert knowledge-constrained spatiotemporal group control. Specifically: (1) A method for describing the operating status of a conveyor belt system based on multimodal intelligent sensing and expert knowledge base: To address the problem that existing conveyor belt correction systems primarily rely on single-sensor information and struggle to comprehensively reflect the conveyor belt's operational status, this invention proposes a method for describing the operational status of a conveyor belt system based on multimodal intelligent sensing and an expert knowledge base. Multimodal sensors (position sensors, tension sensors, weighing sensors, vibration sensors, and conveyor belt image sensors) are deployed at key locations on the conveyor belt as the sensing component of the conveyor belt's embodied intelligent entity. These sensors collect multidimensional information in real time, including conveyor belt offset, tension changes, load distribution, vibration characteristics, and conveyor belt image features. Numerical sensor features are constructed through physical parameter modeling, and latent space image features are constructed through latent space image modeling. The operational status features obtained by connecting the numerical sensor features and latent space image features comprehensively reflect the conveyor belt's operational status. Simultaneously, a knowledge base is constructed from long-term accumulated operational experience and expert knowledge. This knowledge base contains multiple correction decision rules to provide correction adjustments for subsequent decision-making stages. Through this process, a more comprehensive and accurate intelligent understanding of the conveyor belt's operational status is achieved, while providing reliable rule constraints for subsequent correction decisions.

[0009] (2) A method for predicting deviation trends based on bidirectional diffusion: To address the shortcomings of existing conveyor belt misalignment monitoring and control methods, such as a lack of forward-looking prediction capabilities and insufficient robustness and stability of single-modal predictions, this invention proposes a misalignment trend prediction method based on a bidirectional diffusion mechanism. This method serves as an auxiliary sensing tool for the conveyor belt's embodied intelligent system, enabling advanced prediction and accurate perception of conveyor belt misalignment trends. First, a numerical time-series diffusion model is used, taking historical numerical sensor feature sequences of the conveyor belt as input, to predict the numerical sensor feature sequences over a future period, thereby obtaining the numerical trend of conveyor belt misalignment. Simultaneously, a latent space image diffusion model is combined, taking historical latent space image feature sequences of the conveyor belt as input, to predict the latent space image feature sequences over a future period, thus obtaining information on changes in conveyor belt image information. Based on this, a bidirectional consistency constraint mechanism is constructed between numerical prediction and image prediction, allowing the two prediction results to guide each other and perform consistency correction, thereby improving the accuracy and stability of misalignment trend prediction and achieving high-precision forward-looking prediction of conveyor belt misalignment status.

[0010] (3) Spatiotemporal group control and correction method based on expert knowledge constraints driven by embodied intelligence: To address the shortcomings of existing conveyor belt deviation control methods, which often rely on independent adjustment of single idlers and lack global coordination and expert knowledge guidance, this invention proposes an embodied intelligence-driven, expert knowledge-constrained spatiotemporal group control deviation correction method as the decision-making and execution unit of the conveyor belt embodied intelligence system. At the temporal level, this method combines historical operating state feature sequences perceived by multimodal sensors with predicted future operating state feature sequences to construct a single idler's all-time state feature. It also introduces a confidence-driven expert correction mechanism and online cross-entropy optimization to correct and optimize low-confidence all-time state features, improving the system's reliability and robustness in low-confidence scenarios. At the spatial level, a multi-idler spatial group control collaborative decision-making algorithm is used to uniformly plan and globally optimize the adjustment angles of multiple adjustable idlers, achieving overall collaborative deviation correction of the conveyor belt, effectively improving system stability and deviation correction response efficiency. Finally, optimized control commands are used to control the deviation correction roller's movement, and state feedback updates the next cycle's decision, forming a complete closed loop. This ensures stable, adaptive, and forward-looking deviation correction operation of the conveyor belt under complex working conditions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a schematic diagram of the conveyor belt-embedded intelligent system provided in an embodiment of the present invention; Figure 2 This is a general block diagram of a conveyor belt correction method based on embodied perception and bidirectional diffusion prediction provided by an embodiment of the present invention; Figure 3 This is a flowchart of a conveyor belt correction method based on embodied perception and bidirectional diffusion prediction provided by an embodiment of the present invention; Figure 4 This is a flowchart of a method for describing the operating status of a conveyor belt system based on multimodal intelligent sensing and an expert knowledge base, provided in an embodiment of the present invention. Figure 5 This is a flowchart of the deviation trend prediction method based on bidirectional diffusion provided in the embodiments of the present invention; Figure 6 This is a block diagram of the numerical feature noise prediction network structure provided in the embodiments of the present invention; Figure 7 This is a block diagram of the latent space image noise prediction network structure provided in an embodiment of the present invention; Figure 8 This is a flowchart of a multi-roller time-constrained group control and correction method based on multimodal state perception and deviation trend prediction provided in an embodiment of the present invention. Figure 9 This is a block diagram of a conveyor belt correction system based on embodied perception and bidirectional diffusion prediction provided by an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0014] The conveyor belt correction method of this invention uses the physical conveyor belt body and correction rollers (such as...) Figure 1 As shown, it contains a total of Using individual guiding rollers as the carrier, the system acquires its operating status and motion trends in real time through multimodal sensing and bidirectional diffusion mechanisms. It employs a multi-roller collaborative group control strategy for guiding decisions, with a PLC controller as the execution unit. After execution, a feedback loop is formed through sensing operations. The entire system possesses physical embodiment, intelligent sensing, autonomous decision-making, execution, and a feedback loop, constituting a complete embodied intelligent system. Embodied intelligence emphasizes that the system must rely on a specific physical carrier, forming autonomous behavioral capabilities oriented towards task objectives through continuous interaction of sensing, cognition, decision-making, execution, and environmental feedback. In the conveyor belt guiding scenario, the conveyor belt itself, guiding rollers, PLC control unit, and multimodal sensor network together constitute the physical foundation of the embodied intelligent system; multi-dimensional operating information such as offset, tension, load, vibration, and images, along with trend prediction models, constitute the system's sensing input; an expert knowledge base and group control optimization algorithms constitute the system's decision-making core; and the PLC-controlled guiding rollers changing their actions and the resulting feedback constitute the closed-loop interaction between the system and the environment. Therefore, introducing the concept of embodied intelligence into the conveyor belt correction system will help promote the upgrade of the system from traditional mechanical and passive correction to an intelligent correction system with comprehensive perception capabilities, forward prediction capabilities, and collaborative execution capabilities.

[0015] like Figure 2 As shown, the embodied intelligence system comprises three stages: The first stage is the embodied intelligent perception stage of the conveyor belt, which collects multi-dimensional operational data (offset, tension, load, vibration, and image) from multimodal sensors at key locations on the conveyor belt in real time, and improves the characterization ability of the conveyor belt's operating state through physical parameter modeling and latent space image modeling. Simultaneously, an expert knowledge base is constructed to provide rule constraints for subsequent decision-making processes. The second stage is the embodied intelligent assisted perception stage of the conveyor belt, which uses the historical operating state feature sequence output from the first stage as input to perform time-series modeling of numerical sensor features and latent space image features respectively. A numerical time-series diffusion model predicts the changing trend of numerical sensor features, and a latent space image diffusion model predicts future latent space image features, thereby obtaining the evolution information of the conveyor belt's deviation state. The third stage is a spatiotemporal group control and correction method driven by embodied intelligence and based on expert knowledge constraints. This method comprehensively utilizes historical and future operating state feature sequences to fuse historical and predicted states at the time level and introduces expert corrections to improve robustness. At the spatial level, it achieves global collaborative optimization of multiple idlers, ultimately forming a closed-loop decision-making and execution system, effectively avoiding control conflicts and over-correction.

[0016] This invention provides a conveyor belt correction method based on embodied perception and bidirectional diffusion prediction. This method can be implemented by an electronic device, which can be a terminal or a server. Figure 3 The diagram shown is a flowchart of the method. The processing flow may include the following steps: S1. Collect multi-dimensional data on the operating status of the conveyor belt through a multi-modal sensor. The multi-dimensional data includes various physical parameter data and image data. Optionally, such as Figure 4 As shown, S1 specifically includes: Multiple sensors are deployed near each straightening roller on the conveyor belt to collect data on the conveyor belt status near each roller. These sensors include: A laser displacement sensor is used to detect the conveyor belt offset and obtain the lateral offset of the conveyor belt. ,in The order of the alignment rollers, when When the value is positive, the lateral offset is to the right, directly reflecting the belt deviation. Tension sensors are used to monitor changes in conveyor belt tension and obtain the output values ​​from the tension sensors on both sides of the conveyor belt. This is used to determine whether the deviation is caused by tension imbalance; The weighing sensor is used to acquire the load information of the materials on the left and right sides, and obtain the output loads on the left and right sides respectively. This is used to determine whether the deviation is caused by uneven material loading. A piezoelectric accelerometer is used to collect vibration signals, and the output vibration signal is... It is used to monitor for operational anomalies; Industrial cameras are used to acquire images of the conveyor belt in operation, thus obtaining images. .

[0017] S2. By performing physical parameter modeling on the various physical parameter data, numerical sensor features are constructed. By performing latent space image modeling on the image data, latent space image features are constructed. The numerical sensor features and latent space image features are then combined to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. Optionally, such as Figure 4 As shown, S2 specifically includes: The collected multidimensional data is used to model the operational state characteristics, where various physical parameter data are used to construct numerical features. Including offset , offset change rate Tension difference Load distribution index and time window vibration amplitude within The calculation formula is as follows: For images of conveyor belt operation, a VAE encoder is used to extract features from the images: Input image First, the latent variables are mapped to the latent space through multiple convolutional and fully connected layers, and then the mean vector of the latent variables is output. with standard deviation ; Then, the latent space is sampled to obtain the feature vector. : in, Let be a random variable that follows a standard normal distribution; During the training phase, feature vectors The input is fed into the VAE decoder, which consists of multiple deconvolution (convolution transpose) layers used to reconstruct the original image from the latent space, resulting in the reconstructed image. The feature extraction process is optimized by calculating the reconstruction loss through comparing the differences between the input image and the reconstructed image. Obtain the feature vector As latent space image features; Characteristics of numerical sensors After splicing, the first Operating status characteristics near the straightening roller .

[0018] Optionally, such as Figure 4 As shown, the method further includes, based on the operating state characteristics and combined with the conveyor belt operating mechanism and on-site operation and maintenance experience, pre-constructing an expert knowledge base for the conveyor belt correction system. The expert knowledge base includes multiple correction decision rules, each rule including operating state conditions, abnormal cause judgment, and corresponding correction strategy. The correction rules in the expert knowledge base are represented in the following form: in: Representing the Rule 1 The larger the value, the higher the priority. Representing the Near the straightening roller The conditions of the rules Indicates the first The direction of deviation near the correction roller (optional: left deviation, right deviation, no deviation). Indicates the first The deviation intensity near the correction roller can be selected as: weak, medium, or strong. Indicates the first The direction of the deviation trend near the correction roller (optional: left deviation trend, right deviation trend, no deviation trend). Indicates the first The degree of deviation tendency near the correction roller (optional: weak, medium, strong); Indicates the first The status of the conveyor belt near the correction roller (optional: normal, first-level warning, second-level alarm, third-level protection).

[0019] For example:

[0020] The rules above These represent the first threshold for lateral offset, the first threshold for tension difference, the first threshold for offset change rate, and the first threshold for load distribution index, respectively. Similar regulations also include a second and third threshold. The first threshold marks the boundary between no deviation and deviation, while the second threshold is... and The boundary between weak and medium, the third threshold is and The boundary between medium and strong. It is a time window The third threshold of internal vibration amplitude, when When the belt is in a level 3 protection state, similar rules apply to the first threshold and the second threshold. The first threshold is the dividing line between normal and level 1 warning, and the second threshold is the dividing line between level 1 warning and level 2 alarm.

[0021] S3. The numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework uses the historical numerical sensor feature sequence of the conveyor belt as input to predict the numerical sensor feature sequence in the future period and obtain the numerical trend of conveyor belt deviation. At the same time, combined with the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework, the latent space image feature sequence of the conveyor belt is predicted in the future period using the historical latent space image feature sequence of the conveyor belt as input to obtain the information on the change of conveyor belt image. The overall goal of this step is based on time. The following operational status characteristic sequences have been obtained within the historical observation window: and ,(in (The length of the window for both history and the future) to predict the future. Sequence of the corrector roller's operating status within a given time period: and Meanwhile, setting the future Within the time period of the first moment The actual operating state sequence of the straightening roller is as follows and To achieve the above objectives, this embodiment of the invention employs a combination of a numerical temporal diffusion model and a latent space image diffusion model for future state prediction; simultaneously, it introduces a bidirectional consistency constraint to ensure that the representations of the system's future operating state by the two modes are mutually calibrated, thereby significantly improving the stability and reliability of the prediction results.

[0022] Optionally, such as Figure 5 As shown, the numerical time-series diffusion model in S3, which utilizes a bidirectional diffusion-based misalignment trend prediction framework, takes the historical numerical sensor feature sequence of the conveyor belt as input to predict the numerical sensor feature sequence over a future period, thereby obtaining the numerical trend of conveyor belt misalignment. Specifically, this includes: S3-1-1. Employing a one-dimensional temporal convolutional network to process historical numerical sensor feature sequences. Encoding is performed to extract the global dynamic pattern of numerical sensor features over time (using a one-dimensional temporal convolutional network to compress long-series signals into fixed feature vectors for easier subsequent fusion analysis; this can be understood as transforming the matrix into a globally encoded vector that incorporates all historical data), including: Will Write in matrix form , The numerical sensor feature dimension at a single moment is used to obtain a high-dimensional temporal representation of the historical sequence after passing through multiple layers of one-dimensional temporal convolution and ReLU activation function. ,in It is the dimension of the high-dimensional temporal representation, and then the historical numerical global encoding vector is obtained through global average pooling. ; S3-1-2. Constructing a numerical time-series diffusion model to predict numerical sensor characteristics. This process involves building a diffusion framework consisting of forward and backward numerical time-series diffusion, including: Let the future real numerical sensor feature sequence undergoing forward diffusion be... , For the future Within the time period of the first moment The true numerical sensor characteristic sequence of each correction roller; Numerical time series forward diffusion: The role of forward diffusion is to gradually spread towards... Adding Gaussian noise to obtain noisy samples Used during training The network is gradually guided to learn the transformation from Gaussian noise to real samples, including: First to The number of diffusion steps performed is The forward diffusion process, in which the first The forward diffusion process is represented as in: Indicates the first The noise intensity parameters of the step, Represents the identity matrix; The above equation is the form of a forward-diffusion Markov chain. Based on the above equation, ... The process is summarized step by step to obtain noisy samples at any given time. about Closed expression: Among them, cumulative coefficient Defined as , , Noise added at each step; Numerical temporal backdiffusion: Numerical temporal backdiffusion uses historical numerical global encoding vectors As a conditional input, standard Gaussian noise As initial noise, in the diffusion step The system predicts the noise contained in each noisy sample at each step, and gradually restores the numerical sensor feature prediction results for future time periods by removing the noise contained in the noisy samples. like Figure 6 As shown, the numerical temporal noise prediction network: in the first... During the back-diffusion process, the noise prediction network is used to estimate the current noisy sample. The noise contained therein, its input includes: , No. Noisy samples of the step and diffusion step The output is The noise contained In order for the network to be able to perceive the current diffusion stage, the time step Mapping to time-step embedding vectors using fully connected layers:

[0023] in, This represents the mapping of fully connected layers. Indicates the embedding dimension of the time step; Then and Extending along the time dimension, we obtain: ; Next, the current noisy sample, the historical numerical global encoding vector, and the time step embedding vector are concatenated along the feature dimension to construct the first noisy sample. Input to the step-by-step numerical temporal noise prediction network:

[0024] The numerical temporal noise prediction network is constructed using a multi-layer one-dimensional convolutional structure. After being input into the network, the predicted noise is projected onto a noise space of the same dimension as the noisy sample through the output layer, thus obtaining the noise prediction result. ; Numerical time-series progressive denoising generation mechanism: from initial noise Begin the gradual execution of the noise prediction and denoising update process: The denoising formula for each step is as follows: When

[0025] in: , ,

[0026] After the reverse diffusion process is completed, the future is obtained. Predict the numerical sensor feature sequence within a given time period. , that is .

[0027] Optionally, such as Figure 5 As shown, the latent space image diffusion model in S3, which combines a bidirectional diffusion-based deviation trend prediction framework, uses the historical latent space image feature sequence of the conveyor belt as input to predict the latent space image feature sequence over a future period, thereby obtaining conveyor belt image change information. Specifically, it includes: S3-2-1. Employing a one-dimensional temporal convolutional network to process the feature sequences of historical latent space images. Encoding is performed to extract global dynamic patterns that vary over time in the latent space image, including: Will Write in matrix form , The latent space image feature dimension at a single time step is used to obtain a high-dimensional temporal representation of the historical sequence after passing through multiple layers of one-dimensional temporal convolution and ReLU activation function. ,in It is the dimension of the high-dimensional temporal representation, and then the global encoding vector sequence of historical images is obtained through global average pooling. ; S3-2-2, Constructing a latent space image diffusion model to predict latent space image features. This process involves constructing a diffusion framework consisting of forward diffusion and backward diffusion of the latent space image, including: Let the future real latent space image feature sequence undergoing forward diffusion be... , For the future Within the time period of the first moment A sequence of real latent space image features of a single correction roller; Forward diffusion of latent space images: The number of diffusion steps performed is The forward diffusion process of latent space images: Among them, cumulative coefficient Defined as , , Noise added at each step; Latent space image backdiffusion: The backdiffusion process uses the global encoded vector of historical images. As a conditional input, standard Gaussian noise As initial noise, it is progressively predicted and removed during the diffusion step. The noise contained within the image is used to generate latent space image feature prediction results for future time periods. ; like Figure 7 As shown, the latent space image noise prediction network: in the first... During the backdiffusion step, the latent space image noise prediction network is used to estimate the noise contained in the noisy samples of the current step. Its inputs include: , No. Noisy samples of the step and diffusion step The output is The noise contained In order for the network to be able to perceive the current diffusion stage, the time step Mapping to time-step embedding vectors using fully connected layers:

[0028] in, This represents the mapping of fully connected layers. Indicates the embedding dimension of the time step; Then and Extending along the time dimension, we obtain: Next, the current noisy sample, the global encoding vector of historical images, and the embedding vector of time steps are concatenated along the feature dimension to construct the first noisy sample. Input to the latent space image noise prediction network:

[0029] The latent space image noise prediction network is constructed using a multi-layer one-dimensional convolutional structure. After being input into the network, the predicted noise is projected onto a noise space of the same dimension as the noisy sample through the output layer, thus obtaining the noise prediction result. ; Image progressive denoising generation mechanism: from initial noise Begin the gradual execution of the noise prediction and denoising update process: The denoising formula for each step is shown below:

[0030] in: , , ; Once the reverse diffusion process is complete, the future is obtained. Latent space image feature sequence within a time period , that is .

[0031] Optionally, the bidirectional diffusion deviation prediction framework, the training loss function (the design goal of the loss function is to ensure the consistency between image prediction and numerical time series prediction, while guaranteeing the accuracy and stability of the prediction results) includes: Noise prediction loss: During forward and backward diffusion, the noise prediction network needs to accurately predict the noise at each step. The loss function uses the mean squared error, as shown in the following formula:

[0032] Bidirectional consistency loss: A bidirectional consistency constraint loss is introduced to enhance the consistency between latent space image feature prediction and numerical sensor feature prediction.

[0033] in: for The feature mapping neural network from latent space image to numerical sensor is obtained and implemented using a fully connected layer. for The feature mapping neural network from numerical sensor features to latent space image features is obtained and implemented using a fully connected layer. The final total loss function is:

[0034] in, Weights for each part of the loss.

[0035] During training, the parameters of the entire network are optimized using bidirectional consistency loss. In practical use, the numerical-to-image feature mapping neural network and the image-to-numerical feature mapping neural network do not participate in inference; the output... and That's the output for this part.

[0036] S4. At the time level, by combining the historical operating state feature sequence perceived by multimodal sensors with the predicted future operating state feature sequence, a single idler roller full-time state feature is constructed. A confidence-driven expert correction mechanism and cross-entropy online optimization are introduced to correct and optimize the full-time state features with low confidence. Optionally, such as Figure 8 As shown, S4 specifically includes: The input includes information reflecting the historical operating status. and Predictions reflect future operating conditions. and ,set up The output is the full-time state feature. , The three dimensions represent the strength of the deviation, the future trend of the deviation strength, and the confidence index, respectively. The strength of the deviation ranges from -1 to 1 (negative values ​​represent left deviation, positive values ​​represent right deviation, and the larger the absolute value, the greater the deviation) and reflects the current deviation of the belt. The future trend of the deviation strength ranges from -1 to 1 (negative values ​​represent left deviation, positive values ​​represent right deviation, and the larger the absolute value, the greater the deviation). The confidence index ranges from 0 to 1 (a confidence score greater than 0.9 indicates that the current full-state characteristics are reliable, and vice versa). A full-time state feature mapping network is constructed. First, one-dimensional convolution is performed on the fused multimodal time series to extract local dynamic change features at each time step. Then, global average pooling is performed in the time dimension to unify and integrate the temporal features of historical windows and future prediction windows, and to correct the bias. Historical numerical sensor features, latent space image features, and predicted future numerical sensor features, latent space image features are mapped to a method that can comprehensively reflect the correction roller. Uncorrected full-time state characteristics of historical and future states ; A set of rules is obtained according to the pre-built expert knowledge base. ; Then according to To determine if there is an abnormality in the conveyor belt, if If the condition is normal, then perform normal correction. When the alert status is at Level 1, the system will provide audible and visual warnings, prompts from the host computer, and normal error correction. When the alarm is at level two, continuous alarm and normal correction will be performed. Interlock shutdown is initiated when the alarm status is at level three. When the conveyor belt system can perform normal correction, according to The confidence level is used to determine whether to make a correction. If the confidence level is greater than 0.9, no adjustment is needed; simply set it to [the desired level]. As a feature of all-time state Output; if If the value is less than or equal to 0.9, a correction is required. The correction process is as follows: First of all Reconstructing into expert state features , The three dimensions represent the strength of the shift, the future trend of the shift strength, and the confidence level, respectively. The definitions of these three quantities are... The two variables with the same name are defined identically, with a confidence level set to 0.9. The reconstruction method is as follows: left deviation corresponds to negative offset strength; right deviation corresponds to positive offset strength; left deviation trend corresponds to negative future offset strength trend; right deviation trend corresponds to positive future offset strength trend; weak corresponds to 0.3; medium corresponds to 0.5; strong corresponds to 0.7. What will be obtained in the end As a full-time state feature output To further improve system stability, expert state features were also constructed when the confidence level was below 0.9. The cross-entropy loss function between the two is used to optimize the neural network online for low-confidence predictions during the actual inference stage (i.e., the parameters need to be optimized during the training stage and also during the actual use stage, which is online optimization). This makes the output gradually approach the stable decision boundary constructed by expert knowledge rules, thereby effectively reducing the probability of low-confidence predictions and improving the overall robustness of the system.

[0037] S5. After completing the full-time state modeling and expert correction of a single idler, the multi-idler spatial group control collaborative decision-making algorithm is used to make collaborative optimization decisions for all correction idlers at the spatial level, so as to realize the overall collaborative correction of the conveyor belt deviation (to avoid control conflicts, over-correction or system oscillation caused by the independent action of multiple idlers) and obtain the optimal correction roller control quantity for this control cycle. Optionally, such as Figure 8 As shown, S5 specifically includes: The input is the The output is the control quantity of the correction roller.

[0038] First, construct the desired correction control value for a single correction roller:

[0039] Wherein: the first item is the current offset feedback; the second item is the future trend feedforward compensation; To control the gain; Then define the objective function for group control optimization:

[0040] The first term in the group control optimization objective function is a reference tracking term, used to determine the final control quantity of each correction roller. The first term is to approximate the desired correction control value of a single correction roller as closely as possible. This ensures that the control behavior of the single idler roller responds to its current offset strength and future offset trend. The second term is the control energy consumption term, which suppresses drastic angle changes in the idler roller by penalizing excessive control output, thereby reducing mechanical shock and energy consumption of the actuator, while also reducing equipment wear and improving system stability. The third term is the spatial coordination term, in which... , and The first and the The spatial position of the correction roller This term describes the spatial coupling relationship between idlers. The closer the idlers are, the greater their weight. This term penalizes the differences in control values ​​between adjacent idlers, making the adjustment values ​​of spatially adjacent idlers tend to be smooth and consistent, thereby avoiding over-correction or system oscillation problems caused by multiple idlers making large adjustments independently; the fourth term is the time smoothing term: For the previous moment The control quantity of the correction roller is used to suppress drastic changes in the control quantity over time and reduce frequent reverse adjustment behavior. The first constraint is the magnitude constraint, in which... The first constraint represents the lower and upper limits of the physical adjustment of the idler roller actuator, respectively. This constraint ensures that the control quantity is always within the safe operating range allowed by the idler roller structure, preventing it from exceeding mechanical limits. The second constraint is a rate of change constraint, in which... This is the maximum adjustment range per unit of time, which can prevent sudden adjustments. In the collaborative decision-making stage of multi-idler group control, under the condition that the desired correction control quantity, spatial coupling weight, and control quantity at the previous moment are considered as known parameters and the spatial coupling weight is a symmetric positive matrix, the objective function of the group control optimization is expressed as a convex quadratic programming problem with linear constraints. The solution process is as follows: Rewritten in the form of a standard convex quadratic programming problem:

[0041] in, It is a symmetric positive semi-definite matrix, which is composed of the quadratic coefficients corresponding to the reference tracking term, control energy consumption term, spatial coordination term and time smoothing term; It is determined by the desired control quantity of a single idler roller and the control quantity at the previous moment; The linear constraint matrix and vector are constructed from the amplitude constraint and control rate of change constraint of the idler roller actuator, respectively. Since the convex quadratic programming problem has a unique global optimal solution, the optimal solution can be obtained directly using a quadratic programming solver. This serves as the optimal control quantity for the correction roller in this control cycle.

[0042] S6. Control the movement of the correction roller using the optimal correction roller control quantity, and update the decision for the next control cycle through status feedback to form a complete closed loop.

[0043] PLC control and feedback stage: The PLC receives the optimal solution for each correction roller obtained in the previous step. Next, the angle of the correction roller is adjusted and controlled. After execution, the operating status of the conveyor belt is re-acquired by a multi-modal sensor, and the updated status information is fed back to the next control cycle, realizing a closed-loop control process of perception-decision-execution-feedback, thereby ensuring the stable, adaptive, and forward-looking correction operation of the conveyor belt under complex working conditions.

[0044] like Figure 9 As shown, this embodiment of the invention also provides a conveyor belt correction system based on embodied perception and bidirectional diffusion prediction, the system comprising: The acquisition module 910 is used to acquire multidimensional data of the conveyor belt's operating status through a multimodal sensor. The multidimensional data includes various physical parameter data and image data. The modeling module 920 is used to construct numerical sensor features by performing physical parameter modeling on the various physical parameter data, construct latent space image features by performing latent space image modeling on the image data, and stitch the numerical sensor features and latent space image features together to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. The prediction module 930 is used to use the numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the numerical sensor feature sequence of the conveyor belt in the future period, and obtain the numerical trend of the conveyor belt deviation change; at the same time, it combines the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the latent space image feature sequence of the conveyor belt in the future period, and obtain the conveyor belt image change information, using the historical latent space image feature sequence of the conveyor belt as input. Module 940 is used to construct the all-time state features of a single idler roller by combining the historical operating state feature sequences sensed by multimodal sensors with the predicted future operating state feature sequences at the time level. It also introduces a confidence-driven expert correction mechanism and cross-entropy online optimization to correct and optimize the all-time state features with low confidence. The collaborative optimization module 950 is used to perform collaborative optimization decisions on all correction rollers at the spatial level through a multi-roller spatial group control collaborative decision-making algorithm after completing the full-time state modeling and expert correction of a single roller. This achieves overall collaborative correction of conveyor belt deviation and obtains the optimal correction roller control quantity for the current control cycle. The control module 960 is used to control the movement of the correction roller with the optimal correction roller control quantity, and to update the decision of the next control cycle through status feedback to form a complete closed loop.

[0045] The conveyor belt correction system based on embodied perception and bidirectional diffusion prediction provided in this embodiment of the invention has a functional structure that corresponds to the conveyor belt correction method based on embodied perception and bidirectional diffusion prediction provided in this embodiment of the invention, and will not be described again here.

[0046] Figure 10 This is a schematic diagram of the structure of an electronic device 1000 provided in an embodiment of the present invention. The electronic device 1000 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 1001 and one or more memories 1002. The memory 1002 stores at least one instruction, which is loaded and executed by the processor 1001 to implement the steps of the conveyor belt correction method based on embodied perception and bidirectional diffusion prediction described above.

[0047] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the conveyor belt correction method based on embodied perception and bidirectional diffusion prediction. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0048] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0049] 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A conveyor belt deviation correction method based on embodied perception and bidirectional diffusion prediction, characterized in that, The method includes: S1. Collect multi-dimensional data on the operating status of the conveyor belt through a multi-modal sensor. The multi-dimensional data includes various physical parameter data and image data. S2. By performing physical parameter modeling on the various physical parameter data, numerical sensor features are constructed. By performing latent space image modeling on the image data, latent space image features are constructed. The numerical sensor features and latent space image features are then combined to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. S3. The numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework uses the historical numerical sensor feature sequence of the conveyor belt as input to predict the numerical sensor feature sequence in the future period and obtain the numerical trend of conveyor belt deviation. At the same time, combined with the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework, the latent space image feature sequence of the conveyor belt is predicted in the future period using the historical latent space image feature sequence of the conveyor belt as input to obtain the information on the change of conveyor belt image. S4. At the time level, by combining the historical operating state feature sequence perceived by multimodal sensors with the predicted future operating state feature sequence, a single idler roller full-time state feature is constructed. A confidence-driven expert correction mechanism and cross-entropy online optimization are introduced to correct and optimize the full-time state features with low confidence. S5. After completing the full-time state modeling and expert correction of a single idler, the multi-idler spatial group control collaborative decision-making algorithm is used to make collaborative optimization decisions for all correction idlers at the spatial level, so as to realize the overall collaborative correction of the conveyor belt deviation and obtain the optimal correction roller control quantity for this control cycle. S6. Control the movement of the correction roller using the optimal correction roller control quantity, and update the decision for the next control cycle through status feedback to form a complete closed loop.

2. The method according to claim 1, characterized in that, S1 specifically includes: Multiple sensors are deployed near each straightening roller on the conveyor belt to collect data on the conveyor belt status near each roller. These sensors include: A laser displacement sensor is used to detect the conveyor belt offset and obtain the lateral offset of the conveyor belt. ,in The order of the alignment rollers, when When the value is positive, the lateral offset is to the right, directly reflecting the belt deviation. Tension sensors are used to monitor changes in conveyor belt tension and obtain the output values ​​from the tension sensors on both sides of the conveyor belt. This is used to determine whether the deviation is caused by tension imbalance; The weighing sensor is used to acquire the load information of the materials on the left and right sides, and obtain the output loads on the left and right sides respectively. This is used to determine whether the deviation is caused by uneven material loading. A piezoelectric accelerometer is used to collect vibration signals, and the output vibration signal is... It is used to monitor for operational anomalies; Industrial cameras are used to acquire images of the conveyor belt in operation, thus obtaining images. .

3. The method according to claim 2, characterized in that, S2 specifically includes: The collected multidimensional data is used to model the operational state characteristics, where various physical parameter data are used to construct numerical features. Including offset , offset change rate Tension difference Load distribution index and time window vibration amplitude within The calculation formula is as follows: For images of conveyor belt operation, a VAE encoder is used to extract features from the images: Input image First, the latent variables are mapped to the latent space through multiple convolutional and fully connected layers, and then the mean vector of the latent variables is output. with standard deviation ; Then, the latent space is sampled to obtain the feature vector. : in, Let be a random variable that follows a standard normal distribution; During the training phase, feature vectors The input is fed into the VAE decoder, which consists of multiple deconvolutional layers used to reconstruct the original image from the latent space, resulting in the reconstructed image. The feature extraction process is optimized by calculating the reconstruction loss through comparing the differences between the input image and the reconstructed image. Obtain the feature vector As latent space image features; Characteristics of numerical sensors After splicing, the first Operating status characteristics near the straightening roller .

4. The method according to claim 1, characterized in that, The method further includes, based on the operating state characteristics and combined with the conveyor belt operating mechanism and on-site maintenance experience, pre-constructing an expert knowledge base for the conveyor belt correction system. The expert knowledge base includes multiple correction decision rules, each rule including operating state conditions, anomaly cause judgment, and a corresponding correction strategy. The correction rules in the expert knowledge base are represented in the following form: in: Representing the Rule 1 The larger the value, the higher the priority. Representing the Near the straightening roller The conditions of the rules Indicates the first The direction of deviation near the correction roller; Indicates the first Deviation strength near the correction roller; Indicates the first The direction of the deviation trend near the correction roller; Indicates the first The intensity of the deviation tendency near the correction roller; Indicates the first The condition of the conveyor belt near the straightening roller.

5. The method according to claim 3, characterized in that, The numerical time-series diffusion model in S3, which utilizes a bidirectional diffusion-based misalignment trend prediction framework, takes the historical numerical sensor feature sequence of the conveyor belt as input to predict the numerical sensor feature sequence over a future period, thereby obtaining the numerical trend of conveyor belt misalignment. Specifically, this includes: S3-1-1. Employing a one-dimensional temporal convolutional network to process historical numerical sensor feature sequences. Encoding is performed to extract global dynamic patterns of numerical sensor features over time, including: Will Write in matrix form , The numerical sensor feature dimension at a single moment is used to obtain a high-dimensional temporal representation of the historical sequence after passing through multiple layers of one-dimensional temporal convolution and ReLU activation function. ,in It is the dimension of the high-dimensional temporal representation, and then the historical numerical global encoding vector is obtained through global average pooling. ; S3-1-2. Constructing a numerical time-series diffusion model to predict numerical sensor characteristics. This process involves building a diffusion framework consisting of forward and backward numerical time-series diffusion, including: Let the future real numerical sensor feature sequence undergoing forward diffusion be... , For the future Within the time period of the first moment The true numerical sensor characteristic sequence of each correction roller; Numerical time series forward diffusion: The role of forward diffusion is to gradually spread towards... Adding Gaussian noise to obtain noisy samples Used during training The network is gradually guided to learn the transformation from Gaussian noise to real samples, including: First to The number of diffusion steps performed is The forward diffusion process, in which the first The forward diffusion process is represented as in: Indicates the first The noise intensity parameters of the step, Represents the identity matrix; The above equation is the form of a forward-diffusion Markov chain. Based on the above equation, ... The process is summarized step by step to obtain noisy samples at any given time. about Closed expression: Among them, cumulative coefficient Defined as , , Noise added at each step; Numerical temporal backdiffusion: Numerical temporal backdiffusion uses historical numerical global encoding vectors As a conditional input, standard Gaussian noise As initial noise, in the diffusion step The system predicts the noise contained in each noisy sample at each step, and gradually restores the numerical sensor feature prediction results for future time periods by removing the noise contained in the noisy samples. Numerical temporal noise prediction network: in the first During the back-diffusion process, the noise prediction network is used to estimate the current noisy sample. The noise contained therein, its input includes: , No. Noisy samples of the step and diffusion step The output is The noise contained In order for the network to be able to perceive the current diffusion stage, the time step Mapping to time-step embedding vectors using fully connected layers: in, This represents the mapping of fully connected layers. Indicates the embedding dimension of the time step; Then and Extending along the time dimension, we obtain: ; Next, the current noisy sample, the historical numerical global encoding vector, and the time step embedding vector are concatenated along the feature dimension to construct the first noisy sample. Input to the step-by-step numerical temporal noise prediction network: The numerical temporal noise prediction network is constructed using a multi-layer one-dimensional convolutional structure. After being input into the network, the predicted noise is projected onto a noise space of the same dimension as the noisy sample through the output layer, thus obtaining the noise prediction result. ; Numerical time-series progressive denoising generation mechanism: from initial noise Begin the gradual execution of the noise prediction and denoising update process: The denoising formula for each step is as follows: When in: , , After the reverse diffusion process is completed, the future is obtained. Predict the numerical sensor feature sequence within a given time period. , that is .

6. The method according to claim 7, characterized in that, The latent space image diffusion model in S3, which combines a bidirectional diffusion-based deviation trend prediction framework, uses the historical latent space image feature sequence of the conveyor belt as input to predict the latent space image feature sequence over a future period, thereby obtaining information on conveyor belt image changes. Specifically, this includes: S3-2-1. Employing a one-dimensional temporal convolutional network to process the feature sequences of historical latent space images. Encoding is performed to extract global dynamic patterns that vary over time in the latent space image, including: Will Write in matrix form , The latent space image feature dimension at a single time step is used to obtain a high-dimensional temporal representation of the historical sequence after passing through multiple layers of one-dimensional temporal convolution and ReLU activation function. ,in It is the dimension of the high-dimensional temporal representation, and then the global encoding vector sequence of historical images is obtained through global average pooling. ; S3-2-2, Constructing a latent space image diffusion model to predict latent space image features. This process involves constructing a diffusion framework consisting of forward diffusion and backward diffusion of the latent space image, including: Let the future real latent space image feature sequence undergoing forward diffusion be... , For the future Within the time period of the first moment A sequence of real latent space image features of a single correction roller; Forward diffusion of latent space images: The number of diffusion steps performed is The forward diffusion process of latent space images: Among them, cumulative coefficient Defined as , , Noise added at each step; Latent space image backdiffusion: The backdiffusion process uses the global encoded vector of historical images. As a conditional input, standard Gaussian noise As initial noise, it is progressively predicted and removed during the diffusion step. The noise contained within the image is used to generate latent space image feature prediction results for future time periods. ; Latent space image noise prediction network: in the first During the backdiffusion step, the latent space image noise prediction network is used to estimate the noise contained in the noisy samples of the current step. Its inputs include: , No. Noisy samples of the step and diffusion step The output is The noise contained In order for the network to be able to perceive the current diffusion stage, the time step Mapping to time-step embedding vectors using fully connected layers: in, This represents the mapping of fully connected layers. Indicates the embedding dimension of the time step; Then and Extending along the time dimension, we obtain: Next, the current noisy sample, the global encoding vector of historical images, and the embedding vector of time steps are concatenated along the feature dimension to construct the first noisy sample. Input to the latent space image noise prediction network: The latent space image noise prediction network is constructed using a multi-layer one-dimensional convolutional structure. After being input into the network, the predicted noise is projected onto a noise space of the same dimension as the noisy sample through the output layer, thus obtaining the noise prediction result. ; Image progressive denoising generation mechanism: from initial noise Begin the gradual execution of the noise prediction and denoising update process: The denoising formula for each step is shown below: in: , , ; Once the reverse diffusion process is complete, the future is obtained. Latent space image feature sequence within a time period , that is .

7. The method according to claim 6, characterized in that, The loss function trained in the bidirectional diffusion deviation prediction framework includes: Noise prediction loss: During forward and backward diffusion, the noise prediction network needs to accurately predict the noise at each step. The loss function uses the mean squared error, as shown in the following formula: Bidirectional consistency loss: A bidirectional consistency constraint loss is introduced to enhance the consistency between latent space image feature prediction and numerical sensor feature prediction. in: for The feature mapping neural network from latent space image to numerical sensor is obtained and implemented using a fully connected layer. for The feature mapping neural network from numerical sensor features to latent space image features is obtained and implemented using a fully connected layer. The final total loss function is: in, Weights for each part of the loss.

8. The method according to claim 6, characterized in that, S4 specifically includes: The input includes information reflecting the historical operating status. and Predictions reflect future operating conditions. and ,set up The output is the full-time state feature. , The three dimensions represent the strength of the offset, the future trend of the offset strength, and the confidence index, respectively. A full-time state feature mapping network is constructed. First, one-dimensional convolution is performed on the fused multimodal time series to extract local dynamic change features at each time step. Then, global average pooling is performed in the time dimension to unify and integrate the temporal features of historical windows and future prediction windows, and to correct the bias. Historical numerical sensor features, latent space image features, and predicted future numerical sensor features, latent space image features are mapped to a method that can comprehensively reflect the correction roller. Uncorrected full-time state characteristics of historical and future states ; A set of rules is obtained according to the pre-built expert knowledge base. ; Then according to To determine if there is an abnormality in the conveyor belt, if If the condition is normal, then perform normal correction. When the alert status is at Level 1, the system will provide audible and visual warnings, prompts from the host computer, and normal error correction. When the alarm is at level two, continuous alarm and normal correction will be performed. Interlock shutdown is initiated when the alarm status is at level three. When the conveyor belt system can perform normal correction, according to The confidence level is used to determine whether to make a correction. If the confidence level is greater than 0.9, no adjustment is needed; simply set it to [the desired level]. As a feature of all-time state Output; if If the value is less than or equal to 0.9, a correction is required. The correction process is as follows: First of all Reconstructing into expert state features , The three dimensions represent the strength of the shift, the future trend of the shift strength, and the confidence level, respectively. The definitions of these three quantities are... The two variables with the same name are defined identically, with a confidence level set to 0.

9. The reconstruction method is as follows: left deviation corresponds to negative offset strength; right deviation corresponds to positive offset strength; left deviation trend corresponds to negative future offset strength trend; right deviation trend corresponds to positive future offset strength trend; weak corresponds to 0.3; medium corresponds to 0.5; strong corresponds to 0.

7. What will be obtained in the end As a full-time state feature output To further improve system stability, expert state features were also constructed when the confidence level was below 0.

9. The cross-entropy loss function between the two is used to optimize the neural network online for low-confidence predictions during the actual inference stage. This allows the output to gradually approach the stable decision boundary constructed by expert knowledge rules, thereby effectively reducing the probability of low-confidence predictions and improving the overall robustness of the system.

9. The method according to claim 8, characterized in that, S5 specifically includes: The input is the The output is the control quantity of the correction roller. First, construct the desired correction control value for a single correction roller: Wherein: the first item is the current offset feedback; the second item is the future trend feedforward compensation; To control the gain; Then define the objective function for group control optimization: The first term in the group control optimization objective function is a reference tracking term, used to determine the final control quantity of each correction roller. The first term is to approximate the desired correction control value of a single correction roller as closely as possible. This ensures that the control behavior of the single idler roller responds to its current offset strength and future offset trend. The second term is the control energy consumption term, which suppresses drastic angle changes in the idler roller by penalizing excessive control output, thereby reducing mechanical shock and energy consumption of the actuator, while also reducing equipment wear and improving system stability. The third term is the spatial coordination term, in which... , and The first and the The spatial position of the correction roller This term describes the spatial coupling relationship between idlers. The closer the idlers are, the greater their weight. This term penalizes the differences in control values ​​between adjacent idlers, making the adjustment values ​​of spatially adjacent idlers tend to be smooth and consistent, thereby avoiding over-correction or system oscillation problems caused by multiple idlers making large adjustments independently; the fourth term is the time smoothing term: For the previous moment The control quantity of the correction roller is used to suppress drastic changes in the control quantity over time and reduce frequent reverse adjustment behavior. The first constraint is the magnitude constraint, in which... The first constraint represents the lower and upper limits of the physical adjustment of the idler roller actuator, respectively. This constraint ensures that the control quantity is always within the safe operating range allowed by the idler roller structure, preventing it from exceeding mechanical limits. The second constraint is a rate of change constraint, in which... This is the maximum adjustment range per unit of time, which can prevent sudden adjustments. In the collaborative decision-making stage of multi-idler group control, under the condition that the desired correction control quantity, spatial coupling weight, and control quantity at the previous moment are considered as known parameters and the spatial coupling weight is a symmetric positive matrix, the objective function of the group control optimization is expressed as a convex quadratic programming problem with linear constraints. The solution process is as follows: Rewritten in the form of a standard convex quadratic programming problem: in, It is a symmetric positive semi-definite matrix, which is composed of the quadratic coefficients corresponding to the reference tracking term, control energy consumption term, spatial coordination term and time smoothing term; It is determined by the desired control quantity of a single idler roller and the control quantity at the previous moment; The linear constraint matrix and vector are constructed from the amplitude constraint and control rate of change constraint of the idler roller actuator, respectively. Since the convex quadratic programming problem has a unique global optimal solution, the optimal solution can be obtained directly using a quadratic programming solver. This serves as the optimal control quantity for the correction roller in this control cycle.

10. A conveyor belt correction system based on embodied perception and bidirectional diffusion prediction, characterized in that, The system includes: The acquisition module is used to acquire multidimensional data on the operating status of the conveyor belt through a multimodal sensor. The multidimensional data includes various physical parameter data and image data. The modeling module is used to construct numerical sensor features by performing physical parameter modeling on the various physical parameter data, construct latent space image features by performing latent space image modeling on the image data, and stitch the numerical sensor features and latent space image features together to obtain operating status features that can comprehensively reflect the operating status of the conveyor belt. The prediction module is used to use the numerical time-series diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the numerical sensor feature sequence of the conveyor belt in the future period, and obtain the numerical trend of the conveyor belt deviation change. At the same time, it combines the latent space image diffusion model of the bidirectional diffusion deviation trend prediction framework to predict the latent space image feature sequence of the conveyor belt in the future period, and obtain the information on the change of the conveyor belt image. The module is used to construct the all-time state features of a single idler roller by combining the historical operating state feature sequences sensed by multimodal sensors with the predicted future operating state feature sequences at the time level. It also introduces a confidence-driven expert correction mechanism and cross-entropy online optimization to correct and optimize the all-time state features with low confidence. The collaborative optimization module is used to perform collaborative optimization decisions on all correction rollers at the spatial level through a multi-roller spatial group control collaborative decision-making algorithm after completing the full-time state modeling and expert correction of a single roller. This enables the overall collaborative correction of conveyor belt deviation and obtains the optimal correction roller control quantity for the current control cycle. The control module is used to control the movement of the correction roller with the optimal correction roller control quantity, and to update the decision for the next control cycle through status feedback, forming a complete closed loop.

Images