Material dynamic regulation system and control method

Abstract

The application relates to the technical field of industrial intelligent control, and provides a material dynamic regulation and control system and a control method. Through construction of an intelligent regulation and control system from data sensing, strategy generation to precise control, the regulation and control system realizes multidimensional real-time monitoring and dynamic optimization of the material state, can quickly respond to material state changes and process requirement differences, significantly improves the precision, stability and production efficiency of material processing, and simultaneously forms self-learning capability through continuous accumulation of historical regulation and control data and optimization of regulation and control strategies, effectively reduces the artificial intervention cost and production risk, and provides an innovative solution for automatic and intelligent regulation and control of materials in industrial production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial intelligent control, in particular to a material dynamic regulation system and a control method. BACKGROUND

[0002] In modern industrial production, material regulation technology is widely used in food processing, chemical industry, pharmaceutical industry and other fields. Its regulation effect directly affects product quality, production efficiency and energy consumption level. With the development of industrial automation and intelligentization, traditional material regulation technology is also evolving.

[0003] Early material regulation mainly relies on manual experience. The operator manually adjusts the equipment parameters according to the subjective judgment of the equipment running state and the appearance of the material. This method is not only low in efficiency, but also greatly affected by the experience and state of the operator, and it is difficult to ensure the consistency and accuracy of the regulation. Especially when dealing with complex materials or large-scale production, the disadvantages are more obvious.

[0004] With the progress of technology, automatic regulation systems based on single control model have emerged, such as PID control system. The PID system sets fixed parameters to control the temperature and flow of the material in a closed loop, which improves the stability and efficiency of the regulation to a certain extent. However, the single control model is often based on idealized assumptions, and has poor adaptability to changes in material characteristics and complex working conditions. When the density, viscosity and structure of the material fluctuate, the fixed parameter control model cannot be adjusted in time, which easily leads to regulation lag or over-regulation, affecting production quality and efficiency. Most of them do not solve how to perceive multi-dimensional data under complex working conditions and connect the material state to realize intelligent regulation of the material. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a material dynamic regulation system and a control method.

[0006] In a first aspect, the present application provides a material dynamic regulation system, which comprises: acquiring modal data of a material in a target device, and extracting material characteristics of the modal data as material density distribution and material structure characteristics; fusing the material density distribution and the material structure characteristics to generate a feature tensor, and determining the material state of different spatial positions in the target device through spatiotemporal coding;

[0007] Based on the attention mechanism, a weight matrix of the process parameter set and the feature tensor is determined, the process parameter set and the feature tensor are subjected to tensor product operation according to the weight matrix, an associated tensor is generated, and the associated tensor is fused and processed with a process knowledge graph to output associated features, while the material state of different spatial positions is combined to generate a regulation strategy;

[0008] The material in the target device is controlled according to the regulation strategy by the actuator, the control effect is monitored, the impedance of the actuator is matched based on the material structure characteristics, the execution parameters of the actuator are adjusted, and the effect deviation is obtained by comparing the control effect with the target effect, and whether to update the regulation strategy and the material state in different positions of the target device is judged according to the effect deviation.

[0009] As an optional implementation, the generation logic of the feature tensor comprises:

[0010] Modal data of the material in the target device is acquired, and the modal data is preprocessed by an adaptive denoising algorithm;

[0011] The preprocessed modal data is respectively mapped to a density feature space and a structure feature space by a variational autoencoder, the density change of the modal data is extracted by a convolutional neural network to obtain a material density distribution, and the topological relationship between the particles of the material is analyzed by a graph neural network to obtain a material structure feature;

[0012] Feature weights of the material density distribution and the material structure feature are generated by an attention mechanism, and the feature tensor is generated by weighted summation.

[0013] As an optional implementation, the determination logic of the material state comprises:

[0014] The feature tensor is given spatiotemporal coordinate information, the spatial position and time sequence of the material in the target device are marked, and the feature tensor is interacted by a multi-head attention mechanism to extract the change feature of the material in the spatial position and time sequence;

[0015] The material state of the target device at different spatial positions is predicted based on the change feature by a Bayesian neural network, and the prediction uncertainty is output;

[0016] The local feature and the global feature of the feature tensor are extracted, the fusion weight is generated according to the predicted material state and the prediction uncertainty, the local feature and the global feature are fused according to the fusion weight to obtain a fusion feature, and the fusion feature is mapped to determine the material state of the target device at different spatial positions.

[0017] As an optional implementation, the output logic of the associated feature comprises:

[0018] The feature tensor is received and a set of process parameters is acquired, the associated weight of the set of process parameters and the feature tensor in the spatial dimension, the time dimension and the feature dimension is calculated respectively by an attention mechanism, and the associated weight is multiplied to obtain a weight matrix of the set of process parameters and the feature tensor;

[0019] The causal relationship between the process parameter set and the material feature is identified through a causal reasoning network, and the weight matrix of the process parameter set and the feature tensor is corrected according to the causal relationship;

[0020] The process parameter set and the feature tensor are subjected to tensor product operation based on the corrected weight matrix, an associated tensor is generated, the associated tensor is fused with the process knowledge graph, the fused process knowledge graph is reasoned through a graph neural network, and the associated feature is output.

[0021] As an optional implementation, the generation logic of the regulation strategy includes:

[0022] A reinforcement learning is constructed, in which the state space is a global material state formed by splicing the material state and the associated feature of different spatial positions in the target device, the action space includes a parameter configuration set of the actuator, and the regulation efficiency, energy consumption and stability are taken as the reward function;

[0023] Based on the global material state and the associated feature, a global regulation strategy is output, and at the same time, the global regulation strategy is taken as a constraint condition to output a local regulation strategy for the spatial position where the material state is abnormal, and a time sequence regulation strategy is output according to the change of the material state, so as to generate the regulation strategy;

[0024] The prediction uncertainty increment is added to the state space, the regulation strategy is screened and optimized through a genetic algorithm, an optimized regulation strategy is obtained, and the optimized regulation strategy is subjected to feasibility verification to correct the regulation strategy.

[0025] As an optional implementation, the adjustment logic of the execution parameter includes:

[0026] The execution parameter of the actuator in the regulation strategy is extracted, the execution parameter is disassembled into a basic parameter and a modulation parameter, the adjustment range of the execution parameter is determined in combination with the prediction uncertainty, and the confidence interval of the adjusted execution parameter is calculated through a Monte Carlo algorithm;

[0027] According to the process requirement, the historical regulation data and the material state, the basic parameter and the modulation parameter are prioritized, the execution parameter is adjusted according to the priority order, the adjustment effect feedback of the execution parameter is monitored, and the execution parameter is re-adjusted in combination with the confidence interval of the adjusted execution parameter;

[0028] After each adjustment of the execution parameter, the coordination between the basic parameter and the modulation parameter is checked to trigger reverse adjustment based on the priority order, and the adjustment effect feedback of each time is recorded to the historical regulation data.

[0029] As an optional implementation, the matching sub-logic of the actuator impedance includes:

[0030] The material structure features are decomposed into component features by wavelet transform, and the component features include microscopic features, macroscopic features and mesoscopic features;

[0031] The actuator impedance includes mass, damping and stiffness, and a mapping relationship between the component features and the actuator impedance is established based on historical control data to determine an adjustment direction and an adjustment amplitude of the actuator impedance;

[0032] During the material processing, the change of the material state and the operation feedback of the actuator are monitored in real time to dynamically adjust the actuator impedance and the mapping relationship between the component features and the actuator impedance through a double-time-scale control mechanism.

[0033] As an optional implementation, the update logic of the control strategy includes:

[0034] The effect deviation between the control effect and the target effect is calculated by the Mahalanobis distance, and the correlation between the effect deviations is measured, and the effect deviations include state deviation, energy consumption deviation and stability deviation;

[0035] A decision tree is constructed based on historical control data, the adjustment direction of the control strategy is generated according to the combination of the effect deviations, and the decision tree is regularly optimized according to the update success rate of the control strategy;

[0036] The constraint conditions for adjusting the execution parameters are determined according to the physical limitations of the actuator and the material properties, and the execution parameters are iteratively adjusted by adaptive step length;

[0037] The control effect of the adjusted execution parameters on the material in the target device is monitored in real time, and the effect deviation is determined to optimize the control strategy in a closed loop.

[0038] As an optional implementation, the update mechanism of the material state includes:

[0039] A deviation threshold is configured, when any effect deviation is greater than the deviation threshold, a causal probability that the effect deviation is caused by the change of the material state is calculated according to the causal relationship between the process parameter set and the material features;

[0040] The parameter explanatory power of the effect deviation is verified by adjusting the execution parameters, and the update state of the material state is comprehensively judged according to the causal probability and the parameter explanatory power;

[0041] When the material state needs to be updated, the update step length of the material state is dynamically adjusted by fuzzy logic control, the effect deviation is recalculated after the material state is updated, and the material state is iteratively optimized.

[0042] In a second aspect, the application provides a control method of a material dynamic control system, the method comprising: obtaining modal data of material in a target device, and extracting material features of the modal data as material density distribution and material structure features;

[0043] The material density distribution is fused with the material structure features to generate a feature tensor, and the material states of different spatial positions in the target equipment are determined through space-time coding;

[0044] A weight matrix of the process parameter set and the feature tensor is determined based on an attention mechanism, and the process parameter set and the feature tensor are subjected to a tensor product operation according to the weight matrix to generate a correlation tensor;

[0045] The correlation tensor is fused and processed with a process knowledge graph to output correlation features, and a regulation and control strategy is generated in combination with the material states of different spatial positions;

[0046] The material in the target equipment is controlled according to the regulation and control strategy through an actuator, the control effect is monitored, the actuator impedance is matched based on the material structure features, and the execution parameters of the actuator are adjusted;

[0047] An effect deviation is obtained by comparing the control effect with a target effect, and it is determined whether to update the regulation and control strategy and the material states of different spatial positions in the target equipment according to the effect deviation.

[0048] Compared with the prior art, the beneficial effects of the present application are: by constructing an intelligent regulation and control system from data perception, strategy generation to precise control, which is different from the traditional material processing mode relying on artificial experience or single control, the regulation and control system realizes multi-dimensional real-time monitoring and dynamic optimization of the material state, can quickly respond to the change of the material state and the difference of the process demand, significantly improves the precision, stability and production efficiency of the material processing, and at the same time, the regulation and control system forms a self-learning ability by continuously accumulating historical regulation and control data and optimizing the regulation and control strategy, effectively reduces the artificial intervention cost and production risk, and provides an innovative solution for the automatic and intelligent regulation and control of materials in industrial production.

[0049] By acquiring the modal data of the material and extracting the material density distribution and the material structure features, accurate material state basic information is provided for the regulation and control system, the generated feature tensor fuses multi-dimensional information, and the state change of the material in the target equipment is located through space-time coding, so that the regulation and control system can master the characteristics of the material under different spatial positions and time sequences in real time and accurately, and provides reliable data support for subsequent process decision and control, avoiding the regulation and control deviation caused by inaccurate information.

[0050] Based on the attention mechanism and the causal reasoning network, the correlation between the process parameters and the material characteristics can be deeply analyzed, the potential causal logic can be mined, and the data-driven analysis and industry experience knowledge can be combined by fusing with the process knowledge graph to generate the associated features containing rich semantic information. On this basis, the regulation and control strategy is generated by combining the material state, which effectively solves the problems of low matching degree of parameters and material characteristics and single strategy in traditional regulation and control, improves the pertinence and optimization effect of the regulation and control strategy, and helps to realize the fine control of the production process.

[0051] Under the guidance of the regulation and control strategy, the precise control of the material is realized through the actuator. In the control process, the actuator parameters can be dynamically adjusted based on the matching of the material structure characteristics and the actuator impedance to avoid equipment wear or poor material processing effect caused by impedance mismatch. At the same time, by comparing the control effect with the target effect, it is judged whether to update the regulation and control strategy and the material state in real time to form a closed-loop optimization mechanism. This makes the regulation and control system quickly adapt to the changes in the production process, continuously optimize the regulation and control strategy, improve the material processing quality and production efficiency, reduce energy consumption, and enhance the operation stability of the regulation and control system. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0053] Figure 1 The system flowchart of the material dynamic regulation and control system provided by the embodiments of the present application;

[0054] Figure 2 The determination logic diagram of the material state of the material dynamic regulation and control system provided by the embodiments of the present application;

[0055] Figure 3 The matching sub-logic diagram of the actuator impedance of the material dynamic regulation and control system provided by the embodiments of the present application;

[0056] Figure 4 The method flowchart of the control method of the material dynamic regulation and control system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings of the specification. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments.

[0058] Embodiment 1

[0059] As Figure 1 shown, a system flowchart of the material dynamic regulation system is provided for the embodiments of the present application, and the system includes a modal perception module, a process correlation module and a dynamic regulation module.

[0060] The actual application scenario here takes automatic sugar and water adding as an example, so the target device is a water tank, a sugar tank and a mixing tank, and the materials are water and sugar.

[0061] The modal perception module is used to obtain modal data of the materials in the target device, and extract material features of the modal data as material density distribution and material structure features, fuse the material density distribution and the material structure features to generate a feature tensor, and determine the material state of different spatial positions in the target device through spatiotemporal coding.

[0062] Specifically, the generation logic of the feature tensor includes:

[0063] Obtain the modal data of the materials in the target device, and pre-process the modal data through an adaptive denoising algorithm;

[0064] Map the pre-processed modal data to a density feature space and a structure feature space through a variational autoencoder, extract the density change of the modal data through a convolutional neural network to obtain the material density distribution, and analyze the topological relationship between the particles of the materials through a graph neural network to obtain the material structure features;

[0065] Generate feature weights of the material density distribution and the material structure features through an attention mechanism, and generate the feature tensor through weighted summation.

[0066] The original modal data of the materials in the target device is easily disturbed by environmental noise, and direct use for feature extraction will lead to inaccurate results, so the modal data needs to be pre-processed to improve data quality; the physical properties of sugar and water are quite different, so customized sensor configuration and denoising strategies are needed, that is, the sugar tank needs to monitor caking and bridging phenomena, the water tank needs to focus on liquid level fluctuations and turbulent state, a vibration sensor is deployed outside the sugar tank to detect particle flow state, a laser particle size analyzer is installed at the top to monitor sugar particle distribution in real time, a bottom pressure sensor detects the risk of discharge port blockage, and an ultrasonic level meter is used to measure the water level of the water tank, an acoustic sensor is installed on the wall of the water tank to capture liquid flow noise, and a water quality sensor monitors indicators such as dissolved oxygen.

[0067] The acquired modal data is processed by an adaptive noise reduction algorithm. For sugar particle flow signals, the filtering of high-frequency vibration noise needs to be enhanced. For water flow signals, Kalman filtering is used to eliminate noise caused by liquid level fluctuations. The sensor configuration is optimized according to the characteristics of the material, improving the perception of key states such as sugar clumping and water turbulence. The adaptive noise reduction processing ensures the reliability of the modal data, and the precise sugar particle distribution data and water level information provide high-quality input for subsequent density and structural feature extraction.

[0068] The original modal data cannot be directly used for material state analysis and needs to be converted into density distribution and structural features that can reflect the essential characteristics of the material. The subsequent feature tensor is generated by extracting features of sugar particle characteristics and water fluid characteristics through different algorithms. Sugar particle characteristics include particle size distribution and bulk density, and water fluid characteristics include flow rate and liquid level. The preprocessed modal data is mapped to the density feature space and the structural feature space through a variational autoencoder. In terms of density feature extraction, the laser particle size analyzer data is combined with the sugar tank pressure distribution to identify the changes in bulk density at different heights in the sugar tank through a convolutional neural network. When local pressure abnormally rises, it is determined that there is clumping. The convolutional neural network automatically extracts density change features of different scales in the data through multiple convolutional layers and pooling layers.

[0069] In terms of structural feature extraction, based on the water flow noise captured by the acoustic sensor, a graph neural network is used to analyze the turbulence and vortex structure of the water flow, including identifying specific frequency acoustic wave patterns to determine whether there is abnormal turbulence caused by pipe blockage. The topological relationship between sugar particles is constructed using vibration sensor data, and the agglomeration state of sugar particles is analyzed through a graph neural network. When the vibration signal decays, it is determined that there is a bridging phenomenon. The feature extraction algorithm is customized according to the characteristics of sugar and water, which can accurately capture the clumping risk of granular materials and the flow abnormalities of liquid materials. The extracted material density distribution and material structural features provide key inputs for subsequent feature tensor generation, ensuring that the feature tensor can fully reflect the material state.

[0070] To more comprehensively and efficiently represent the material characteristics, it is necessary to fuse the material density distribution and structural characteristics to generate a unified feature tensor, which is convenient for subsequent spatio-temporal coding and material state determination, that is, the particle characteristics of sugar and the fluid characteristics of water need to be fused to form a unified representation for subsequent spatio-temporal coding analysis; the attention mechanism is introduced to analyze the material density distribution and material structure characteristics to generate their feature weights. The attention mechanism automatically assigns weights according to the importance of different characteristics in reflecting the material state. When the sugar tank is at risk of caking, the weight of the sugar structure feature is automatically increased. When the water level in the water tank is close to the warning line, the weight of the water structure feature is increased. Then, the material density distribution and the material structure characteristics are fused by weighted summation to generate a feature tensor that integrates the density and structure information of the material. The feature tensor is stored in the form of a multi-dimensional array and can clearly reflect the feature combination of the material in different dimensions. The feature weights are dynamically adjusted so that the feature tensor can highlight key information according to the working conditions and improve the representation ability of the material state. The fused feature tensor contains the spatio-temporal change information of sugar and water, realizes efficient fusion and unified representation of material characteristics, and can more comprehensively describe the material state. Compared with using a single feature, the representation ability of the material state is greatly improved, providing a rich data basis for subsequent spatio-temporal coding.

[0071] Specifically, as shown in Figure 2 the determination logic of the material state includes:

[0072] The feature tensor is assigned with spatio-temporal coordinate information to mark the spatial position and time sequence of the material in the target device, and the feature tensor is interacted through the multi-head attention mechanism to extract the change features of the material in the spatial position and time sequence;

[0073] The Bayesian neural network is used to predict the material state of the target device at different spatial positions based on the change features and output the prediction uncertainty;

[0074] The local features and global features of the feature tensor are extracted, the fusion weights are generated according to the predicted material state and prediction uncertainty, the local features and global features are fused according to the fusion weights to obtain the fused features, and the fused features are mapped to determine the material state of the target device at different spatial positions.

[0075] The state of the material in the target device is not only related to its own characteristics, but also affected by the spatial position and time sequence. To accurately determine the material state, the feature tensor needs to be spatio-temporally coded and its change features in the spatio-temporal dimension need to be extracted. First, the feature tensor is assigned with spatio-temporal coordinate information to clearly mark the spatial position and time sequence of the material in the target device, that is, three-dimensional coordinate systems are established for the sugar tank and the water tank to mark the spatial positions of sugar and water at different time points, including the discharge port position of the sugar tank and the water inlet position of the water tank.

[0076] Then the feature tensor is interacted by the multi-head attention mechanism, which analyzes the feature tensor from multiple angles, captures the correlation and changes of material features between different spatial positions and time points, pays attention to the relationship between sugar dissolution rate and water flow speed during the sugar-water mixing stage, and captures the spatio-temporal features of the uneven mixing area, thereby extracting the change features of the material in the spatial position and time sequence; the spatio-temporal encoding and feature interaction enable the information of the material features in the spatio-temporal dimension to be fully mined and integrated, which can more accurately describe the dynamic change process of the material in the target device and provide more rich spatio-temporal information for material state prediction, and the extracted change features provide key inputs for the Bayesian neural network to predict the material state.

[0077] After obtaining the spatio-temporal change information of the material features, the material state at different positions in the target device needs to be predicted by the Bayesian neural network to provide decision basis for subsequent material regulation, which is to predict the states such as the dissolution degree of sugar and the liquid level change of water in actual application to provide decision support for automatic sugar and water addition; the Bayesian neural network processes the extracted change features to predict the material state at different spatial positions of the target device, which can not only output the predicted value of the material state but also output the prediction uncertainty. The Bayesian neural network learns a large amount of historical data to establish the probability relationship between the material features and the material state, and in the prediction process, according to the current input change features, the probability of different material states is calculated by combining the learned probability relationship, and the state with the maximum probability is selected as the predicted value, and the uncertainty of the prediction result is output to reflect the reliability of the prediction, i.e. inputting the features such as sugar particle distribution and water flow speed, and outputting the predicted value and uncertainty of the sugar dissolution state, i.e. predicting the dissolution percentage of sugar in water, and giving the confidence interval. When there is uncertainty in the dissolution process, the dependence on the sugar concentration sensor data is increased and the prediction weight of the Bayesian neural network is reduced.

[0078] The application of the Bayesian neural network enables the material state prediction to not only give the prediction result, i.e. the sugar dissolution state, but also to evaluate the prediction uncertainty, which provides more comprehensive information for decision-making, helps users to reasonably develop regulation strategies according to the reliability of the prediction, reduces the decision risk, and specifically takes a conservative strategy when the uncertainty is high, i.e. prolongs the stirring time. The material state prediction value and prediction uncertainty provide an important reference for subsequent local and global feature fusion, which is used to generate fusion weights to guide the feature fusion process to more accurately determine the final state of the material.

[0079] In order to comprehensively consider the local details and global trends of the material, accurately determine the final state of the material at different positions in the target device, the local features and global features of the feature tensor need to be fused, and the predicted material state and uncertainty are combined for decision-making. In practice, it is to comprehensively consider the local details and global trends of sugar dissolution, and accurately judge the mixing state. First, the local features and global features of the feature tensor are extracted. The local features are obtained by analyzing the local area of the feature tensor, focusing on the subtle changes and local characteristics of the material, that is, the arrangement of the material particles in a small area. The global features are obtained by comprehensive analysis of the entire feature tensor, reflecting the overall distribution and macro trend of the material, that is, the overall density distribution of the material in the device.

[0080] Then, according to the predicted material state and the prediction uncertainty, a fusion weight is generated. If the prediction uncertainty is high, the fusion weight of the local feature will be appropriately increased to pay more attention to the local details of the material and ensure accurate judgment of the material state. Conversely, the fusion weight of the global feature is increased. Then, the local feature and the global feature are weighted and fused according to the fusion weight to obtain a fused feature. Finally, the fused feature is input into the mapping rule to determine the final state of the material at different positions in the target device. In practical applications, the local feature focuses on the dissolution state of sugar particles near the tank discharge port to determine whether there are undissolved sugar blocks. The global feature analyzes the sugar concentration distribution in the entire mixing tank to evaluate the mixing uniformity. When the prediction uncertainty is high, the fusion weight of the local feature is increased to focus on the existing agglomeration. When the scheduling system is stable, the fusion weight of the global feature is increased to focus on the overall mixing effect.

[0081] The mapping rule includes dividing the material state into three levels of judgment results, including normal, warning and abnormal. Relevant rules are developed for different material types and device areas, that is, when the fused feature shows that the local density is greater than the density threshold and the global density gradient is abnormal at the bottom area of the sugar tank, it is judged as a lumping warning, but accompanied by a vibration sensor signal attenuation, it is upgraded to a lumping warning. When the fused feature shows that the local turbulence intensity is less than the standard intensity and the global flow rate is uneven at the mixing area of the water tank, it is judged as a mixing abnormality.

[0082] By fusing local features and global features and making decisions combined with the predicted material state, the material state can be determined comprehensively and accurately, the characteristics of the material under different scales and different reliability levels are fully considered, the sugar dissolution state and mixing uniformity can be accurately judged under different working conditions, compared with single feature analysis, the accuracy and reliability of the material state determination are significantly improved, the misjudgment caused by local abnormalities is avoided, the accurately determined material state provides accurate basis for the process correlation module to generate a control strategy, so that the control strategy can more targetedly control the material, and also provides a clear target for the dynamic control module to execute the control operation, and improves the control effect and operation efficiency of the whole material dynamic control system.

[0083] The process correlation module is configured to determine a weight matrix of the process parameter set and the feature tensor based on an attention mechanism, perform a tensor product operation on the process parameter set and the feature tensor according to the weight matrix to generate a correlation tensor, and fuse and process the correlation tensor with a process knowledge graph to output a correlation feature, while generating a control strategy combined with the material state of different spatial positions.

[0084] Specifically, the output logic of the correlation feature includes:

[0085] The feature tensor is received and the process parameter set is obtained, the correlation weights of the process parameter set and the feature tensor in the spatial dimension, the time dimension and the feature dimension are calculated respectively through the attention mechanism, the correlation weights are multiplied to obtain the weight matrix of the process parameter set and the feature tensor;

[0086] The causal relationship between the process parameter set and the material feature is identified through the causal reasoning network, and the weight matrix of the process parameter set and the feature tensor is corrected according to the causal relationship;

[0087] The process parameter set and the feature tensor are subjected to a tensor product operation based on the corrected weight matrix to generate a correlation tensor, the correlation tensor is fused with the process knowledge graph, the fused process knowledge graph is reasoned through a graph neural network, and a correlation feature is output.

[0088] The distribution of sugar particles and the turbulent state of water during the mixing process are significantly affected by the space-time of process parameters. The coupling relationship between parameters and features needs to be analyzed by multi-dimensional weight to avoid control lag or excessive adjustment. Here, the attention mechanism is used to calculate the correlation weight of process parameter set and feature tensor in spatial dimension, time dimension and feature dimension. The correlation weight in spatial dimension is calculated by deploying sensor nodes at the discharge port of the sugar tank, the mixing area of the water tank and the pipe elbow, including setting pressure, vibration and liquid level sensors. For each process parameter, including sugar adding speed, stirring speed and sugar tank valve opening, the influence of the process parameter on the material features at different spatial positions is calculated by scanning the spatial grid with a two-dimensional convolution kernel. For example, when the valve opening at the bottom of the sugar tank increases, the convolution kernel outputs the weight of the action on the increase of sugar density at the bottom of the mixing tank. The darker the color, the higher the weight, which intuitively shows the spatial influence range.

[0089] The correlation weight in time dimension is calculated by bidirectional gate recurrent unit processing historical control data, with a time window of 2 hours to capture the time sequence dependence of process parameter adjustment and material state change. When the analysis finds that the lag time of the process in which the increase of stirring speed increases the water turbulence intensity is about 30 seconds, the bidirectional gate recurrent unit will automatically assign a higher weight to the influence of this process parameter on the turbulence features in the next 30 seconds. The correlation weight in feature dimension is calculated by constructing a semantic correlation matrix of process parameters and material features. The rows of the matrix represent process parameters, and the columns represent material features, including sugar clumping rate and water turbulence intensity. The element value of the semantic correlation matrix is calculated by self-attention mechanism, i.e. the semantic similarity between stirring speed and water turbulence intensity is 0.8, while the semantic similarity between stirring speed and sugar clumping rate is only 0.3, so as to determine the correlation weight in feature dimension. The three-dimensional weight matrix realizes the space-time fine analysis of the influence of process parameters on material features, and can accurately locate the problem of insufficient valve opening at the left side of the sugar tank leading to low sugar concentration at the lower right of the mixing tank, avoiding the global average adjustment mode of traditional PID control. The generated weight matrix provides correlation data with space-time labels for causal reasoning, and the spatial dimension weight can help determine whether there is a real causal relationship between the sugar adding speed and the local clumping of the mixing tank.

[0090] There are parameters with statistical correlation but no causal relationship in the mixing process of sugar and water, which need to be removed through causal reasoning to ensure that the weight matrix reflects the real action path; the causal reasoning network is constructed by inputting process parameters and material characteristics, and the influence of process parameter changes on material state is simulated through intervention experiments. Preferably, the water temperature is fixed at 25℃, and the stirring speed is forced to increase from 500 rpm to 800 rpm, and the change of sugar dissolution time is observed. If the dissolution time is shortened, it is confirmed that there is a causal effect of stirring speed on dissolution rate. If not, it is determined to be a false correlation, and the value of its weight matrix is reduced. At the same time, the causal effect intensity is determined, that is, the influence of stirring speed on dissolution rate accounts for 70%, and the weight matrix is corrected. If the causal relationship between stirring speed and sugar clumping rate is found to be weak, the weight is reduced to avoid excessive adjustment of stirring speed. Causal reasoning effectively removes false correlations in data, making the weight matrix more physically interpretable, improving the reliability of subsequent feature fusion, and the corrected weight matrix provides more accurate weighting basis for tensor product operation, ensuring that the correlation tensor can truly reflect the interaction between process parameters and feature tensors.

[0091] A single process parameter cannot fully reflect the complexity of the mixing process, that is, the sugar addition speed needs to work with stirring speed and water temperature, and high-dimensional correlation features need to be constructed through tensor product operation, and the semantic interpretability of feature tensors needs to be improved combined with process knowledge; based on the corrected weight matrix, the process parameters and feature tensors are subjected to tensor product operation to generate correlation tensors containing time and space and feature dimensions. Specifically, the sugar flow parameter is multiplied by the sugar density feature at different heights of the mixing tank to highlight the influence of flow on local density. The process knowledge graph is called, which stores rules such as too fast sugar addition leading to clumping and low water temperature requiring longer stirring time. Each rule is a knowledge node, and the matching degree between the correlation tensor and the knowledge node is inferred through the graph neural network. When the correlation tensor shows that the sugar addition speed is too fast and the water temperature is too low, the clumping risk node in the process knowledge graph is automatically matched to enhance the ability to identify abnormal states. Tensor product operation realizes the multi-dimensional fusion of process parameters and feature tensors, and process knowledge graph introduces the experience of domain experts, so that the correlation features can not only reflect real-time data, but also conform to process common sense. The output correlation features integrate data-driven quantitative analysis and knowledge-driven qualitative judgment, providing a comprehensive decision basis for generating control strategies.

[0092] Specifically, the generation logic of the control strategy includes:

[0093] A reinforcement learning is constructed, in which the state space is the global material state obtained by splicing the material state at different spatial positions in the target device and the correlation features, the action space includes the parameter configuration set of the actuator, and the control efficiency, energy consumption and stability are taken as the reward function;

[0094] Based on the global material state and associated features, output the global regulation strategy, and at the same time, output the local regulation strategy for the spatial position where the material state is abnormal, and output the time sequence regulation strategy according to the change of the material state, to generate the regulation strategy;

[0095] Add the prediction uncertainty increment to the state space, screen and optimize the regulation strategy through the genetic algorithm, obtain the optimized regulation strategy, and verify the feasibility of the optimized regulation strategy to correct the regulation strategy.

[0096] The mixing of sugar and water needs to optimize the regulation efficiency, energy consumption and stability, and the traditional control method is difficult to dynamically balance, so it is necessary to search for the optimal strategy combination in the state space through reinforcement learning; the state space is spliced by the material state and associated features of each position of the mixing tank, when the sugar concentration at the bottom of the mixing tank is lower than the target value and the associated features show that the sugar adding speed weight is higher, the state space will highlight the abnormal state of this area; the action space includes the discrete combination of execution parameters such as sugar tank valve opening and stirring speed, the stirring speed is divided into low, medium and high three grades, and combined with the fast, medium and slow three grades of sugar flow to form 9 action options; the reward function is designed as the weighted sum of regulation efficiency, energy consumption and stability; reinforcement learning can automatically explore the optimal parameter combination, adapt to the dynamic changes of execution parameters such as sugar tank valve opening and stirring speed in the mixing process of sugar and water, and construct reinforcement learning to provide an optimization target for hierarchical regulation strategy generation, ensuring the synergy of global regulation strategy, local regulation strategy and time sequence regulation strategy.

[0097] During the mixing process, global uniformity, local abnormality handling and time sequence lag need the cooperation of different levels of strategies, among which the global strategy ensures the overall concentration to meet the standard, the local strategy eliminates the bottom block, and the time sequence strategy compensates for the delay effect of execution parameter adjustment; based on the overall sugar concentration distribution and associated features, the global regulation strategy is generated, when the global average sugar concentration is less than 20% of the target value, the global instruction of increasing the sugar flow to 120% of the rated value is output to quickly improve the overall concentration; for the bottom area where the block is detected in the mixing tank, the local regulation strategy is generated as a constraint of the global regulation strategy, which increases the pulse vibration of the bottom area to eliminate the block while increasing the sugar flow globally; considering the lag effect of execution parameter adjustment, it takes 5 minutes to completely dissolve after adding sugar, a strategy sequence with a time window is generated, that is, first add sugar quickly at high flow, and then start high-speed stirring after 5 minutes to ensure dissolution, ensuring the time alignment of action and effect; the hierarchical regulation strategy realizes the three-dimensional control of global optimization to local repair and then to time sequence smoothing, avoids the global imbalance or local handling not in time caused by a single strategy, and generates a rich candidate solution set for the optimization of the genetic algorithm, improving the diversity and robustness of the regulation strategy.

[0098] Sugar particle size fluctuation and sensor noise, etc. will cause uncertainty in the state space, including the difference in dissolution rate of coarse and fine sugar, which needs to be quantified in the optimization of the control strategy to avoid aggressive decisions in high uncertainty scenarios; the predicted uncertainty output by the modal perception module is added to the state space as the input value of reinforcement learning, and when the sugar particle size detection data is missing, the dissolution time prediction uncertainty is high, the control system automatically switches to a conservative strategy, that is, the stirring time is extended by 10%; the control strategy generated by reinforcement learning is optimized by genetic algorithm, and the control strategy with strong stability in high uncertainty scenarios is preferentially retained, that is, in the working condition of large water temperature fluctuation, the control strategy of low-speed stirring first and then gradually increasing the speed is selected, and the feasibility of the optimized control strategy is verified to check whether the motor power is out of limit and whether the pipeline is blocked; the consideration of uncertainty enables the control system to dynamically adjust the aggressiveness of the control strategy according to the data reliability, avoiding blind optimization in high-risk scenarios, and the optimized control strategy is output to the dynamic control module for execution after the feasibility verification, ensuring the safety and operability of the control strategy.

[0099] The dynamic control module is used to control the material in the target device according to the control strategy through the actuator, monitor the control effect, match the actuator impedance based on the material structure characteristics, adjust the execution parameters of the actuator, and compare the control effect with the target effect to obtain the effect deviation, and determine whether to update the control strategy and the material state in different positions of the target device according to the effect deviation.

[0100] Further, as shown in Figure 3 The matching sub-logic of the actuator impedance includes:

[0101] The material structure characteristics are received, and the material structure characteristics are decomposed into component characteristics by wavelet transform, including microscopic characteristics, macroscopic characteristics and mesoscopic characteristics;

[0102] The actuator impedance includes mass, damping and stiffness, and the mapping relationship between the component characteristics and the actuator impedance is established based on the historical control data to determine the adjustment direction and adjustment amplitude of the actuator impedance;

[0103] During the material processing process, the change of the material state and the running feedback of the actuator are monitored in real time to dynamically adjust the actuator impedance and the mapping relationship between the component characteristics and the actuator impedance through the double-time scale control mechanism.

[0104] The structural characteristics of different materials vary greatly, and a single control method cannot be adapted. By decomposing the material structure characteristics according to the micro, meso and macro dimensions, the impedance of the actuator can be matched more accurately. After obtaining the material structure characteristics through the modal perception module, the wavelet transform technology is used for multi-scale decomposition to obtain the component characteristics, including micro, macro and meso characteristics. For granular materials such as sugar, the micro level can analyze the particle surface roughness and particle size distribution, the meso level studies the size and shape of the particle agglomerates, and the macro level focuses on the bulk density and flowability of the material. For water, the characteristics are decomposed from the molecular structure (micro), the water flow vortex shape (meso) and the overall liquid level (macro). Thus, the material structure characteristics are analyzed in detail, providing multi-dimensional basis for subsequent actuator impedance adjustment, avoiding a single control method, and the decomposed component characteristics provide a basis for establishing a mapping relationship with the actuator impedance, making the impedance adjustment more targeted.

[0105] The mass, damping and stiffness parameters of the actuator have a significant impact on the material processing effect, and a corresponding relationship needs to be established according to the material structure characteristics to ensure that the actuator impedance matches the material structure characteristics. Based on historical control data, a mapping relationship between component characteristics and actuator impedance is established. When sugar clumping is detected (meso characteristic anomaly), the actuator stiffness needs to be increased to break the clumps. When the water flow is unstable (macro characteristic anomaly), the damping parameter needs to be adjusted to stabilize the water flow. The control system determines the adjustment direction and approximate amplitude of the actuator impedance through these mapping relationships, thereby quickly finding an actuator impedance that matches the current material state, improving control efficiency, reducing equipment damage or poor material processing effect caused by mismatched actuator impedance, and providing a reference standard for real-time dynamic adjustment of actuator impedance, making the adjustment process more efficient and accurate.

[0106] The material state changes in real time during production, and the change speed varies, so a double-time-scale control mechanism needs to be used to balance fast response and long-term optimization. On the fast time scale, the control system monitors the material state changes and actuator operation feedback in real time, including motor current and vibration frequency. Once an anomaly is detected, the motor current increases sharply due to sugar tank discharge port blockage, and a small amplitude impedance parameter fine-tuning is performed immediately to increase the stiffness and quickly solve the immediate problem. On the slow time scale, the mapping relationship is optimized through deep reinforcement learning algorithm based on long-term operation data. If it is found that the actuator impedance adjustment effect is not good under a certain working condition, the control system automatically learns and corrects the corresponding relationship to make the actuator impedance more adaptive to complex working conditions. The fast time scale ensures the timely response of the control system to sudden conditions, and the slow time scale optimizes long-term operating conditions, improving the stability and adaptability of the control system. The dynamically adjusted actuator impedance provides a stable basis for actuator parameter adjustment, and the data accumulated during the adjustment process also provides a reference for updating the control strategy.

[0107] Specifically, the execution parameter adjustment logic comprises:

[0108] extracting the execution parameters of the actuators in the control strategy, decomposing the execution parameters into basic parameters and modulation parameters, determining the adjustment range of the execution parameters in combination with the prediction uncertainty, and calculating the confidence interval of the adjusted execution parameters through the Monte Carlo algorithm;

[0109] According to the process requirements, historical control data and material state, the basic parameters and modulation parameters are prioritized, the execution parameters are adjusted according to the priority order, the adjustment effect of the execution parameters is monitored and fed back, and the execution parameters are re-adjusted in combination with the confidence interval of the adjusted execution parameters;

[0110] After adjusting the execution parameters each time, the coordination between the basic parameters and the modulation parameters is checked to trigger the reverse adjustment based on the priority order, and the adjustment effect feedback of each time is recorded to the historical control data.

[0111] The types of execution parameters are various, and the effects are different. After decomposition, precise adjustment can be made according to the characteristics of different parameters, while considering the prediction uncertainty to avoid excessive adjustment due to inaccurate data. The execution parameters in the control strategy are decomposed into basic parameters and modulation parameters, wherein the basic parameters refer to the parameters that directly affect the material conveying and processing, i.e. the valve opening, and the modulation parameters refer to the parameters used for fine-tuning the control effect, i.e. the fluctuation frequency of the stirring motor speed. At the same time, in combination with the prediction uncertainty output by the modal perception module, the adjustment sensitivity is set for different execution parameters. For the area with high prediction uncertainty, the adjustment range of the basic parameters is small, and the modulation parameters are carefully fine-tuned. For the area with low uncertainty, the adjustment range can be appropriately increased, and the confidence interval of the adjusted execution parameters is calculated through the Monte Carlo algorithm to quantify the reliability of parameter adjustment. Thus, the classification and fine adjustment of the execution parameters are realized, the control risk caused by data uncertainty is reduced, the control accuracy is improved, and the determined adjustment range and confidence interval provide a quantitative basis for subsequent parameter priority sorting and adjustment.

[0112] The influence degree of each execution parameter on the control effect is different at different stages and working conditions in the production process, and preferentially adjusting the key parameters can improve the control efficiency; according to the process requirements, historical control data and real-time material state, the priority of the basic parameters and the modulation parameters is sorted, when the material is blocked, the priority of the sugar tank valve opening and the vibrator vibration frequency is improved, when the mixing uniformity needs to be improved, the priority of the stirring motor speed is higher, the execution parameters are adjusted according to the priority order, and the effect feedback of the adjusted parameters is monitored in real time, if the effect is not as expected, the second adjustment is made combined with the confidence interval of the adjusted parameters; so as to ensure that the key problems are preferentially handled under complex working conditions, avoid blind adjustment of parameters to cause instability of the control system, improve the control efficiency and effect, and the effect feedback of the adjusted execution parameters is used to check the coordination between parameters, and data is accumulated for the control strategy update.

[0113] The execution parameters are related to each other, and single parameter adjustment will affect the effect of other parameters, checking the coordination can avoid the conflict between parameters and ensure the stable operation of the control system; after adjusting the execution parameters each time, the control system automatically checks the coordination between the basic parameters and the modulation parameters, when the sugar tank valve opening is increased to improve the sugar adding speed, whether the stirring motor speed matches it is checked, if the speed is too low, the material will be accumulated, and whether the vibration frequency of the vibrator needs to be adjusted to prevent blockage is checked, if the execution parameters are not coordinated, the reverse adjustment based on the priority order is triggered, if the stirring motor load is too large, the valve opening is appropriately reduced or the speed is increased, to ensure the stable operation of the control system, after the adjustment is completed, the adjustment effect feedback is recorded to the historical control data, to provide a reference for subsequent control; so as to timely find and solve the conflict between execution parameters, ensure the stable operation of the control system, and optimize the control strategy by recording the feedback data, the coordinated execution parameters ensure the effective execution of the current control strategy, and the accumulated data provides a basis for the control strategy update and material state update.

[0114] Specifically, the update logic of the control strategy includes:

[0115] The effect deviation between the control effect and the target effect is calculated by the Mahalanobis distance, and the correlation between the effect deviations is measured, the effect deviation includes state deviation, energy consumption deviation and stability deviation;

[0116] A decision tree is constructed based on the historical control data, the adjustment direction of the control strategy is generated according to the combination of the effect deviations, and the decision tree is optimized regularly according to the success rate of the control strategy update;

[0117] The constraint conditions for adjusting the execution parameters are determined according to the physical limitations of the actuators and the material properties, and the execution parameters are adjusted by adaptive step iteration;

[0118] The control effect of the adjusted execution parameters on the material in the target equipment is monitored in real time to determine the effect deviation for closed-loop optimization of the regulation and control strategy.

[0119] To determine whether the current regulation and control strategy is effective, the actual control effect is compared with the target effect, the deviation is calculated, and the correlation is analyzed to provide a direction for adjustment of the regulation and control strategy. The effect deviation between the control effect and the target effect is calculated by Mahalanobis distance, covering state deviation, energy consumption deviation, and stability deviation. The state deviation includes the difference between the actual density of the material and the target density. The energy consumption deviation includes the comparison between the actual energy consumption and the expected energy consumption. The stability deviation includes the vibration amplitude of the equipment operation and the parameter fluctuation. The correlation between the effect deviations is also measured. If the material density deviation is large and the energy consumption is also high, and the correlation between them is strong, it indicates that there is a double problem caused by unreasonable regulation and control strategy. Thus, the execution effect of the regulation and control strategy is quantified comprehensively to identify the root cause of the problem and avoid misjudgment caused by single index analysis. The calculated effect deviation and correlation analysis results provide data support for building a decision tree and determining the adjustment direction of the regulation and control strategy.

[0120] A large amount of experience is contained in historical regulation and control data, which can be converted into rules through the construction of a decision tree to quickly determine the adjustment direction of the regulation and control strategy. Based on historical regulation and control data, a decision tree is constructed with the combination of effect deviations as the decision node. If the state deviation is greater than the state deviation threshold and the energy consumption deviation is also high, the decision tree points to adjusting the basic parameters to optimize the material processing efficiency. If the stability deviation is outstanding, the decision tree points to checking the equipment operation parameters and adjusting the regulation and control parameters. According to the currently calculated effect deviation, the decision tree automatically generates the adjustment direction of the regulation and control strategy. At the same time, the regulation and control system optimizes the decision tree regularly based on the update success rate of the regulation and control strategy, deletes invalid branches, and adds new effective rules to make it more suitable for actual production needs. The historical experience is converted into operable rules to improve the pertinence and efficiency of the regulation and control strategy adjustment, reduce the trial and error cost, and provide guidance for the determination of execution parameter adjustment constraints and strategy iteration optimization.

[0121] According to the adjustment direction of the regulation strategy, the execution parameters are adjusted under the physical limitations of the actuators and the constraints of the material properties, and closed-loop monitoring is used to ensure continuous optimization of the regulation strategy; the constraint conditions for adjusting the execution parameters are determined according to the physical limitations of the actuators and the material properties, the physical limitations of the actuators include the maximum power of the motor and the maximum opening of the valve, and the material properties include the maximum pressure and temperature range that the material can withstand, and an adaptive step iterative adjustment is used to adjust the execution parameters, a larger step is used to quickly approach the target at the first adjustment, and then the step is gradually reduced for fine adjustment according to the effect feedback, the control effect of the adjusted execution parameters on the material in the target device is monitored in real time, the effect deviation is recalculated, if the effect deviation still does not reach the target range, the execution parameters are continuously adjusted, a closed-loop optimization is formed, and a satisfactory control effect is achieved; thereby ensuring the safety of the equipment and the quality of the material, realizing the continuous optimization of the regulation strategy, improving the stability and production efficiency of the regulation system, and providing protection for the continuous operation of the dynamic regulation module, and the data in the adjustment process also provides a reference for the material state update.

[0122] Specifically, the material state update mechanism includes:

[0123] A deviation threshold is configured, when any effect deviation is greater than the deviation threshold, the causal probability of the effect deviation caused by the change of the material state is calculated according to the causal relationship between the process parameter set and the material characteristics;

[0124] The parameter interpretation of the effect deviation is verified by adjusting the execution parameters, and the update state of the material state is comprehensively judged according to the causal probability and the parameter interpretation;

[0125] When the material state needs to be updated, the update step of the material state is dynamically adjusted through fuzzy logic control, the effect deviation is recalculated after the material state is updated, and the material state is iteratively optimized.

[0126] Not all effect deviations need to update the material state, significant deviations are selected by setting a threshold, and the causal relationship between the significant deviations and the change of the material state is analyzed to avoid unnecessary update operations; a deviation threshold is configured in advance, when any effect deviation is greater than the corresponding deviation threshold, the subsequent process is triggered, the causal probability of the effect deviation caused by the change of the material state is calculated according to the causal relationship between the process parameter set and the material characteristics established by the process correlation module, if the material density deviation is detected to be out of tolerance, the regulation system analyzes whether it is caused by process parameter adjustment or material state change after querying the causal relationship, and calculates the corresponding causal probability; thereby accurately identifying the deviation conditions that need attention, reducing invalid calculation and misoperation, improving the operation efficiency of the regulation system, and the calculated causal probability provides an important basis for judging whether to update the material state.

[0127] Part of the effect bias can be solved by adjusting the execution parameters, and the parameter interpretation ability needs to be verified to comprehensively judge whether the material state really needs to be updated; by fine-tuning the execution parameters, the change of the effect bias is observed, the parameter interpretation ability is verified, and when the sugar adding speed is appropriately reduced, if the material density deviation is significantly reduced, it means that the deviation can be solved by parameter adjustment, and the material state does not need to be updated, if the deviation does not change significantly, the material state needs to be updated according to the causal probability, when the causal probability is high and the parameter interpretation ability is low, it is determined that the material state needs to be updated; otherwise, the current state is maintained; thereby avoiding the error update of the material state caused by misjudgment, ensuring the stable operation of the regulation system, reducing unnecessary regulation operation, and determining the material state update decision to start the subsequent state update operation or maintain the existing regulation strategy.

[0128] In the material state update process, the update step under different working conditions is difficult to fix, which can be dynamically adjusted according to the actual situation through fuzzy logic control to ensure the stability and effectiveness of the update process; when it is determined that the material state needs to be updated, the update step is dynamically adjusted through the fuzzy logic controller, the fuzzy logic controller takes the effect deviation and the causal probability as input variables, if the effect deviation is large and the causal probability is high, a larger update step is output to speed up the material state adjustment speed; if the effect deviation is small, a small step is used for fine-tuning, the material state is updated, the effect deviation is recalculated, if it still does not reach the target range, the update step is continuously adjusted through the fuzzy logic control for iterative optimization until the material state meets the production requirements; thereby realizing the adaptive adjustment of the material state update step, avoiding the system fluctuation caused by improper step, improving the accuracy and stability of the material state adjustment, and feeding back the updated material state to the modal perception module and the process correlation module to provide the latest data for subsequent regulation, and also as a reference basis for further optimization of the regulation strategy.

[0129] Embodiment 2

[0130] As shown in Figure 4 , a method flow chart of a control method of a material dynamic regulation system is provided for the embodiments of the present application, the method comprises:

[0131] Modal data of the material in the target device is acquired, and material features of the modal data, i.e. material density distribution and material structure features, are extracted;

[0132] The material density distribution and the material structure features are fused to generate a feature tensor, and the material state of different spatial positions in the target device is determined through space-time coding;

[0133] The weight matrix of the feature tensor and the process parameter set is determined based on the attention mechanism, the feature tensor and the process parameter set are subjected to tensor product operation according to the weight matrix, and the correlation tensor is generated;

[0134] The correlation tensor is fused and processed with the process knowledge graph to output correlation features, and a regulation and control strategy is generated in combination with the material states at different spatial positions;

[0135] The material in the target device is controlled according to the regulation and control strategy by the actuator, the control effect is monitored, the actuator impedance is matched based on the material structure features, and the execution parameters of the actuator are adjusted;

[0136] The effect deviation is obtained by comparing the control effect with the target effect, and whether to update the regulation and control strategy and the material states at different spatial positions in the target device is determined according to the effect deviation.

[0137] Since the principle of the method in the embodiments of the present application solves the problem is similar to the system described above, the implementation of the method refers to the implementation of the system, and the repeated parts will not be repeated.

Claims

1. A material dynamic control system, characterized in that, include: Modal data of materials in the target equipment are acquired, and the material features of the modal data are extracted as material density distribution and material structure features. The material density distribution and material structure features are fused to generate a feature tensor, and the material state at different spatial locations in the target equipment is determined by spatiotemporal coding. The weight matrix of the process parameter set and feature tensor is determined based on the attention mechanism. The process parameter set and feature tensor are multiplied by the weight matrix to generate the association tensor. The association tensor is then fused and processed with the process knowledge graph to output the association features. At the same time, the control strategy is generated by combining the material state at different spatial locations. The output logic for the associated features includes: Receive the feature tensor and obtain the process parameter set. Calculate the correlation weights between the process parameter set and the feature tensor in the spatial dimension, temporal dimension, and feature dimension through the attention mechanism. Multiply the correlation weights by the weight matrix of the process parameter set and the feature tensor. The causal relationship between the process parameter set and material characteristics is identified by the causal reasoning network, and the weight matrix of the process parameter set and feature tensor is corrected according to the causal relationship. Tensor product operation is performed on the process parameter set and feature tensor based on the corrected weight matrix to generate the association tensor. The association tensor is then fused with the process knowledge graph. The fused process knowledge graph is then inferred through a graph neural network to output the association features. The actuator controls the material in the target equipment according to the control strategy, monitors the control effect, matches the actuator impedance based on the material structure characteristics, and adjusts the actuator's execution parameters. At the same time, the control effect is compared with the target effect to obtain the effect deviation. Based on the effect deviation, it is determined whether to update the control strategy and the material status at different locations in the target equipment.

2. The material dynamic control system as described in claim 1, characterized in that, The logic for generating the feature tensor includes: Modal data of materials in the target equipment are acquired, and the modal data are preprocessed using an adaptive noise reduction algorithm; The preprocessed modal data is mapped to the density feature space and the structural feature space by a variational autoencoder. The density changes of the modal data are extracted by a convolutional neural network to obtain the material density distribution. The topological relationship between the particles of the material is analyzed by a graph neural network to obtain the material structural features. Feature weights for material density distribution and material structure characteristics are generated through an attention mechanism, and feature tensors are generated through weighted summation.

3. The material dynamic control system as described in claim 2, characterized in that, The logic for determining the state of the material includes: Assign spatiotemporal coordinate information to the feature tensor, mark the spatial location and time series of the material in the target equipment, and perform feature interaction on the feature tensor through a multi-head attention mechanism to extract the change features of the material in spatial location and time series; The material state of the target device at different spatial locations is predicted based on the changing characteristics using a Bayesian neural network, and the prediction uncertainty is output. Local and global features are extracted from the feature tensor. Fusion weights are generated based on the predicted material state and prediction uncertainty. The local and global features are then fused according to the fusion weights to obtain fused features. The fused features are then mapped to determine the material state of the target device at different spatial locations.

4. The material dynamic control system as described in claim 3, characterized in that, The generation logic of the regulation strategy includes: Reinforcement learning is constructed, where the state space is the global material state formed by splicing the material states and associated features at different spatial locations in the target device, and the action space includes the parameter configuration set of the actuator, and the control efficiency, energy consumption and stability are used as reward functions; Based on the global material state and related characteristics, a global control strategy is output. At the same time, with the global control strategy as a constraint, a local control strategy is output for the spatial location where the material state is abnormal. Furthermore, a time-series control strategy is output based on the changes in the material state to generate a control strategy. The prediction uncertainty increment is added to the state space, and the regulation strategy is screened and optimized by a genetic algorithm to obtain the optimized regulation strategy. The feasibility of the optimized regulation strategy is then verified to correct the regulation strategy.

5. The material dynamic control system as described in claim 4, characterized in that, The logic for adjusting the execution parameters includes: The execution parameters of the actuator in the control strategy are extracted, and the execution parameters are decomposed into basic parameters and modulation parameters. The adjustment range of the execution parameters is determined by combining the prediction uncertainty, and the confidence interval of the adjusted execution parameters is calculated by Monte Carlo algorithm. Based on process requirements, historical control data, and material status, the basic parameters and modulation parameters are prioritized, and the execution parameters are adjusted according to the priority order. The feedback on the adjustment effect of the execution parameters is monitored, and the execution parameters are readjusted based on the confidence interval of the adjusted execution parameters. After each adjustment of the execution parameters, the coordination between the basic parameters and the modulation parameters is checked to trigger a reverse adjustment based on priority order, and the effect of each adjustment is updated and recorded in the historical control data.

6. The material dynamic control system as described in claim 5, characterized in that, The matching sub-logic for the actuator impedance includes: The material structure features are received and decomposed into component features through wavelet transform. The component features include micro-features, macro-features and meso-features. Actuator impedance includes mass, damping, and stiffness. A mapping relationship between component characteristics and actuator impedance is established based on historical control data to determine the adjustment direction and adjustment range of actuator impedance. During the material handling process, the changes in the material state and the operation feedback of the actuator are monitored in real time, so as to dynamically adjust the mapping relationship between the actuator impedance and component characteristics and the actuator impedance through a dual time scale control mechanism.

7. The material dynamic control system as described in claim 6, characterized in that, The update logic of the control strategy includes: The deviation between the control effect and the target effect is calculated by Mahalanobis distance, and the correlation between the deviations is measured. The deviations include state deviation, energy consumption deviation and stability deviation. A decision tree is constructed based on historical control data. The direction of adjustment of the control strategy is generated based on the combination of effect deviations. The decision tree is then optimized periodically based on the success rate of updating the control strategy. The constraints for adjusting the execution parameters are determined based on the physical limitations of the actuator and the properties of the material, and the execution parameters are adjusted iteratively through adaptive step size. Real-time monitoring of the control effect of adjusted execution parameters on materials in the target equipment, determination of effect deviation, and closed-loop optimization of control strategies.

8. The material dynamic control system as described in claim 7, characterized in that, The material status update mechanism includes: Configure a deviation threshold. When any effect deviation exceeds the deviation threshold, calculate the causal probability that the effect deviation is caused by the change in the material state based on the causal relationship between the process parameter set and the material characteristics. The explanatory power of the parameters that verify the effect deviation is verified by adjusting the execution parameters, and the update status of the material status is judged by combining the causal probability and the explanatory power of the parameters. When the material status needs to be updated, the update step size of the material status is dynamically adjusted through fuzzy logic control. After the material status is updated, the effect deviation is recalculated to iteratively optimize the material status.

9. A control method for a material dynamic control system, characterized in that, include: Obtain modal data of materials in the target equipment, and extract material characteristics from the modal data as material density distribution and material structure characteristics; Material density distribution and material structure features are fused to generate a feature tensor, and the material state at different spatial locations in the target equipment is determined by spatiotemporal coding. The weight matrix of the process parameter set and feature tensor is determined based on the attention mechanism. The process parameter set and feature tensor are then multiplied by tensor based on the weight matrix to generate the associated tensor. The correlation tensor is fused and processed with the process knowledge graph to output correlation features, and control strategies are generated by combining the material state at different spatial locations. The output logic for the associated features includes: Receive the feature tensor and obtain the process parameter set. Calculate the correlation weights between the process parameter set and the feature tensor in the spatial dimension, temporal dimension, and feature dimension through the attention mechanism. Multiply the correlation weights by the weight matrix of the process parameter set and the feature tensor. The causal relationship between the process parameter set and material characteristics is identified by the causal reasoning network, and the weight matrix of the process parameter set and feature tensor is corrected according to the causal relationship. Tensor product operation is performed on the process parameter set and feature tensor based on the corrected weight matrix to generate the association tensor. The association tensor is then fused with the process knowledge graph. The fused process knowledge graph is then inferred through a graph neural network to output the association features. The actuator controls the material in the target equipment according to the control strategy, monitors the control effect, matches the actuator impedance based on the material structure characteristics, and adjusts the actuator's execution parameters. By comparing the control effect with the target effect, the effect deviation is obtained. Based on the effect deviation, it is determined whether to update the control strategy and the material state at different spatial locations in the target equipment.

Images