A smart sand and gravel screening efficiency optimization control system

Abstract

This invention discloses an intelligent sand and gravel screening efficiency optimization control system, belonging to the field of optimization control technology. The system includes: a screening control data integration module, a sand and gravel screening efficiency prediction module, a sand and gravel screening control parameter optimization module, and an efficiency optimization control module. Specifically, this invention determines the working condition based on the comprehensive variance, generates an enhanced feature vector, performs dynamic calibration based on criticality scores, constructs a time-series feature interaction graph, generates two enhanced views, constructs positive and negative sample pairs, and fuses time-series and spatial losses to construct a comprehensive loss function, obtaining a sand and gravel screening efficiency prediction model, providing data support for subsequent dynamic optimization of control parameters; it sets up a state space and action space, constructs a multi-objective weighted reward function, optimizes network parameters with the objective of maximizing the expected cumulative discount reward, and outputs the optimal operating control parameters for sand and gravel screening, significantly improving the adaptability and accuracy of sand and gravel screening efficiency optimization control.

Description

Technical Field

[0001] This invention belongs to the field of optimization control technology, specifically referring to an intelligent sand and gravel screening efficiency optimization control system. Background Technology

[0002] The sand and gravel screening efficiency optimization control system is an intelligent system that utilizes artificial intelligence technology to optimize screening efficiency by collecting real-time data on material characteristics, equipment operating status, and environmental parameters. This helps formulate precise equipment control strategies, thereby improving sand and gravel screening efficiency and promoting the intelligent upgrading of the sand and gravel production process. However, existing sand and gravel screening efficiency optimization control systems suffer from several problems. Firstly, they fail to differentiate between operating conditions, resulting in insufficient targeted feature extraction. Secondly, they neglect complex interactions between features, focusing only on temporal features while lacking spatial correlation mining. This leads to large prediction errors and low accuracy in sand and gravel screening efficiency, failing to provide a reliable basis for optimization control. Furthermore, existing systems suffer from incomplete state descriptions, failing to consider the operating conditions of the screening process. They also rely on single reward indicators and control parameters that depend heavily on experience, making dynamic adjustment difficult. This results in an inability to accurately perceive the real-time screening scenario, leading to low adaptability and accuracy in sand and gravel screening efficiency optimization control. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent sand and gravel screening efficiency optimization control system. Addressing the problems of existing systems that fail to differentiate between working conditions, resulting in insufficient feature extraction, neglecting complex interactions between features, focusing only on temporal features while lacking spatial correlation mining, leading to large prediction errors and low accuracy in sand and gravel screening efficiency, thus failing to provide a reliable basis for optimization control, this solution determines the working condition based on comprehensive variance, generates enhanced feature vectors, performs dynamic calibration based on keyness scoring, generates initial screening efficiency prediction values, constructs a temporal feature interaction graph based on feature type and corresponding working condition, generates random sampling disturbance coefficients to obtain two enhanced views, constructs positive and negative sample pairs, fuses temporal and spatial losses to construct a comprehensive loss function, obtains a sand and gravel screening efficiency prediction model, calibrates the initial screening efficiency prediction values, and outputs a model that simultaneously conforms to temporal changes and spatial correlations. The final predicted value of the regularity significantly improves the prediction accuracy, providing accurate and reliable data support for the subsequent dynamic optimization of control parameters. Addressing the problems in existing sand and gravel screening efficiency optimization control systems, such as one-sided state descriptions that fail to consider the working conditions of the screening process, single reward indicators, and control parameters that rely heavily on experience and are difficult to dynamically adjust, resulting in an inability to accurately perceive the real-time screening scenario and low adaptability and accuracy of sand and gravel screening efficiency optimization control, this solution sets up a state space and an action space, constructs an efficiency-gradient composite reward item, an equipment health-stability reward item, and an energy consumption-adjustment cost reward item. Based on working condition adaptive weight vector weighted fusion, a multi-objective weighted reward function is obtained. A reinforcement learning agent based on a near-end policy optimization algorithm is constructed to optimize network parameters with the objective of maximizing the expected cumulative discount reward, outputting the optimal operating control parameters for sand and gravel screening, significantly improving the adaptability and accuracy of sand and gravel screening efficiency optimization control.

[0004] The present invention provides an intelligent sand and gravel screening efficiency optimization control system, including a screening control data integration module, a sand and gravel screening efficiency prediction module, a sand and gravel screening control parameter optimization module, and an efficiency optimization control module;

[0005] The screening control data integration module collects historical screening control time series data and preprocesses it to construct training datasets and test datasets.

[0006] The sand and gravel screening efficiency prediction module determines the working condition based on the comprehensive variance, generates an enhanced feature vector, performs dynamic calibration based on the criticality score, generates an initial screening efficiency prediction value, constructs a time-series feature interaction graph based on the feature type and corresponding working condition, generates random sampling perturbation coefficients to obtain two enhanced views, constructs positive and negative sample pairs, integrates time-series and spatial losses to construct a comprehensive loss function, obtains the sand and gravel screening efficiency prediction model, and calibrates the initial screening efficiency prediction value.

[0007] The sand and gravel screening control parameter optimization module sets the state space and action space, constructs an efficiency-gradient composite reward item, an equipment health-stability reward item, and an energy consumption-adjustment cost reward item, obtains a multi-objective weighted reward function based on the working condition adaptive weight vector weighted fusion, constructs a reinforcement learning agent based on the near-end policy optimization algorithm, optimizes the network parameters with the goal of maximizing the expected cumulative discount reward, and outputs the optimal operating control parameters for sand and gravel screening.

[0008] The efficiency optimization control module sends the optimal operating control parameters for sand and gravel screening to the programmable logic controller of the sand and gravel screening equipment, thereby realizing real-time adaptive optimization control of sand and gravel screening efficiency.

[0009] Furthermore, the screening control data integration module collects historical screening control time series data and performs preprocessing, dividing the preprocessed data into training datasets and test datasets;

[0010] The historical screening control time series data includes timestamps, material characteristic data, equipment operating status data, environmental parameter data, and screening efficiency, with screening efficiency used as the data label;

[0011] The preprocessing includes data cleaning, data encoding, and data normalization.

[0012] Furthermore, the sand and gravel screening efficiency prediction module predicts sand and gravel screening efficiency by constructing a sand and gravel screening efficiency prediction model. It includes a working condition feature fusion unit, a dynamic attention calibration unit, a feedforward prediction unit, a time-series feature interaction graph construction unit, a graph enhancement comparison unit, and a model optimization convergence unit; specifically, it includes the following:

[0013] The working condition feature fusion unit sets a threshold for the comprehensive variance, calculates the comprehensive variance within the window corresponding to each time step of the training dataset according to the sliding time window, and determines the working condition status of the data corresponding to each time step; extracts corresponding dynamic features according to different working condition statuses, extracts the linear trend features within the window for stable working condition data, and extracts the abrupt change features between adjacent time steps for transitional working condition data, fuses them to generate an enhanced feature vector, slices the enhanced sequence data in a sliding window manner, and performs standardization processing;

[0014] The dynamic attention calibration unit performs linear projection on the standardized enhanced data vector to generate a query matrix, a key matrix, and a value matrix. It calculates the basic self-attention, calculates the criticality score based on the working condition and feature fluctuation information at each time step, dynamically calibrates the attention weights based on the criticality score, and generates the calibrated attention output.

[0015] Feedforward prediction unit: The calibrated attention output is nonlinearly mapped through a feedforward neural network to obtain the initial screening efficiency prediction value. The temporal attention loss function is constructed by combining the mean square error and the attention weight regularization term.

[0016] The temporal feature interaction graph construction unit constructs a node set using each feature dimension as a graph node. Based on the feature type and corresponding working condition, it initializes an embedding vector for each node, calculates the interaction strength between any two nodes, selects the top C nodes with the highest interaction strength as adjacent nodes for each node, generates an edge set and an adjacency matrix, and constructs a temporal feature interaction graph based on the node set and edge set.

[0017] The graph enhancement contrast unit generates two independent random sampling perturbation coefficients for the temporal feature interaction graph, calculates the perturbation interaction strength, and updates the adjacency matrix to obtain two enhanced views. The graph convolutional network is used to aggregate features of the two enhanced views respectively, and the adjacency node information is fused to update the node embedding. Based on the interaction strength and node type, positive sample pairs and negative sample pairs are constructed from each enhanced view. The spatial contrast loss function is constructed by combining the view contrast loss and the dynamic interaction consistency loss.

[0018] The model is optimized and converged. By combining the temporal attention loss function and the spatial contrast loss function, a comprehensive loss function is constructed. With the goal of minimizing the comprehensive loss, the Adam optimizer is used to update all learnable parameters synchronously to obtain the trained sand and gravel screening efficiency prediction model. The initial screening efficiency prediction value is calibrated and used as the final output of the sand and gravel screening efficiency prediction model.

[0019] Furthermore, the sand and gravel screening control parameter optimization module, by constructing a sand and gravel screening control parameter optimization model, dynamically optimizes the adjustable operating parameters of the screening equipment based on reinforcement learning to generate the optimal operating control parameters for sand and gravel screening. It includes a state-action space setting unit, a reward function setting unit, and an optimal control generation unit; specifically, it includes the following:

[0020] State and Action Space Setting Unit: Define the fusion vector containing material characteristic data, environmental parameter data, equipment operating status data, and working condition status markers during the sand and gravel screening process as the state space to obtain the state vector; Define the set of values ​​of the adjustable operating parameters of the screening equipment within its feasible range as the action space to obtain the action vector.

[0021] The reward function setting unit integrates the current screening efficiency achievement rate and efficiency rise gradient to construct an efficiency-gradient composite reward item; integrates the stability of adjustable parameter fluctuations and equipment health status to construct an equipment health-stability reward item; integrates motor energy consumption cost and parameter adjustment cost to construct an energy consumption-adjustment cost reward item; based on the current operating condition status label, an adaptive weight vector of operating conditions is generated, and the three reward items are weighted and integrated to finally construct a multi-objective weighted reward function.

[0022] The optimal control generation unit is constructed based on state vectors, action vectors, and a multi-objective weighted reward function. A reinforcement learning agent based on a proximal policy optimization algorithm is built. The agent's policy network takes the current state vector as input and outputs the probability distribution of each adjustable operating parameter value in the action vector. With the goal of maximizing the expected cumulative discount reward, the policy network parameters are iteratively optimized, and the current parameters are taken as the optimal policy network parameters. The corresponding network is a sand and gravel screening control parameter optimization model. The state vector is input into the trained policy network, and the recommended optimal operating control parameters for sand and gravel screening under the current state are output.

[0023] Furthermore, the efficiency optimization control module collects real-time screening control timing data, which includes timestamps, material characteristic data, equipment operating status data, and environmental parameter data. After data cleaning, encoding, and normalization, the real-time screening control timing data is sequentially input into the trained sand and gravel screening efficiency prediction model and sand and gravel screening control parameter optimization model to obtain the optimal operating control parameters for sand and gravel screening. These parameters are then sent to the programmable logic controller of the sand and gravel screening equipment to achieve real-time adaptive optimization control of sand and gravel screening efficiency.

[0024] The beneficial effects achieved by the present invention using the above solution are as follows:

[0025] (1) To address the problems in existing sand and gravel screening efficiency optimization control systems, such as insufficient feature extraction due to lack of differentiation between working conditions, neglect of complex interactions between features, focus only on temporal features while lacking spatial correlation mining, resulting in large prediction errors and low accuracy of sand and gravel screening efficiency, and failing to provide a reliable basis for optimization control, this scheme determines the working condition based on the comprehensive variance, generates enhanced feature vectors, achieves accurate matching between features and working conditions, and retains key information under different working conditions; performs dynamic calibration based on criticality scoring to enhance sensitivity to key temporal information affecting efficiency; generates initial screening efficiency prediction values, and outputs data that closely matches actual efficiency changes. The initial predicted value of the law is obtained; based on the feature type and corresponding working condition, a time-series feature interaction graph is constructed to improve the comprehensiveness and effectiveness of feature representation; random sampling perturbation coefficients are generated to obtain two enhanced views, and positive and negative sample pairs are constructed to improve the reliability of spatial feature representation; a comprehensive loss function is constructed by fusing time-series and spatial losses to guide the model to simultaneously optimize the learning effect of time-series and spatial features; a sand and gravel screening efficiency prediction model is obtained, and the initial screening efficiency prediction value is calibrated to output the final prediction value that simultaneously conforms to the time-series change and spatial correlation law, significantly improving the prediction accuracy and providing accurate and reliable data support for subsequent dynamic optimization of control parameters.

[0026] (2) To address the problems in existing sand and gravel screening efficiency optimization control systems, such as one-sided state descriptions, failure to consider the working conditions of the screening process, single reward indicators, and reliance on experience for control parameters, which makes dynamic adjustment difficult and results in an inability to accurately perceive the real-time screening scenario, leading to low adaptability and accuracy of sand and gravel screening efficiency optimization control, this solution sets up a state space and an action space to comprehensively depict the screening scenario state and ensure accurate perception of real-time working conditions; constructs an efficiency-gradient composite reward item, an equipment health-stability reward item, and an energy consumption-adjustment cost reward item, covering key objectives of the entire screening production process, providing comprehensive optimization guidance for agent learning; obtains a multi-objective weighted reward function based on working condition adaptive weight vector weighted fusion, enhancing the working condition adaptability of the reward mechanism and guiding the output of parameters that are more in line with actual production needs; constructs a reinforcement learning agent based on a near-end strategy optimization algorithm to get rid of experience dependence and realize intelligent generation of control parameters; optimizes network parameters with the goal of maximizing expected cumulative discount rewards, outputs optimal operating control parameters for sand and gravel screening, and significantly improves the adaptability and accuracy of sand and gravel screening efficiency optimization control. Attached Figure Description

[0027] Figure 1 A schematic diagram of an intelligent sand and gravel screening efficiency optimization control system provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the sand and gravel screening efficiency prediction module;

[0029] Figure 3This is a schematic diagram of the sand and gravel screening control parameter optimization module.

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] Example 1, see Figure 1 The present invention provides an intelligent sand and gravel screening efficiency optimization control system, including a screening control data integration module, a sand and gravel screening efficiency prediction module, a sand and gravel screening control parameter optimization module, and an efficiency optimization control module.

[0034] The screening control data integration module collects historical screening control time series data and preprocesses it to construct training datasets and test datasets, and then sends the data to the sand and gravel screening efficiency prediction module.

[0035] The sand and gravel screening efficiency prediction module receives data sent by the screening control data integration module, determines the working condition based on the comprehensive variance, generates an enhanced feature vector, performs dynamic calibration based on the criticality score, generates an initial screening efficiency prediction value, constructs a time-series feature interaction graph based on the feature type and corresponding working condition, generates random sampling perturbation coefficients to obtain two enhanced views, constructs positive and negative sample pairs, fuses time-series and spatial losses to construct a comprehensive loss function, obtains the sand and gravel screening efficiency prediction model, calibrates the initial screening efficiency prediction value, and sends the data to the sand and gravel screening control parameter optimization module.

[0036] The sand and gravel screening control parameter optimization module receives data sent by the sand and gravel screening efficiency prediction module, sets the state space and action space, constructs an efficiency-gradient composite reward item, an equipment health-stability reward item, and an energy consumption-adjustment cost reward item, obtains a multi-objective weighted reward function based on working condition adaptive weight vector weighted fusion, constructs a reinforcement learning agent based on the near-end policy optimization algorithm, optimizes network parameters with the goal of maximizing the expected cumulative discount reward, outputs the optimal operating control parameters for sand and gravel screening, and sends the data to the efficiency optimization control module;

[0037] The efficiency optimization control module receives data sent by the sand and gravel screening control parameter optimization module, and sends the optimal operating control parameters for sand and gravel screening to the programmable logic controller of the sand and gravel screening equipment to realize real-time adaptive optimization control of sand and gravel screening efficiency.

[0038] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the screening control data integration module, historical screening control time series data is collected and preprocessed. The preprocessed data is divided into training dataset and test dataset.

[0039] The historical screening control time series data includes timestamps, material characteristic data, equipment operating status data, environmental parameter data, and screening efficiency. Screening efficiency is used as the data label. Screening efficiency = (actual mass of undersize product × percentage of particles smaller than the screen aperture in the feed) / total feed mass of the screening equipment × 100%;

[0040] The material characteristic data includes feed rate, particle size distribution, material moisture content, material density, and impurity rate;

[0041] The equipment operating status data includes screen rotation speed, screen amplitude, motor current, motor power, screen tilt angle, screen pressure, and screen temperature;

[0042] The environmental parameter data includes ambient temperature and ambient humidity;

[0043] The preprocessing includes data cleaning, data encoding, and data normalization;

[0044] The data cleaning process involves identifying and processing invalid data, filling missing values ​​with the mean, and detecting and processing outliers using the 3σ principle.

[0045] The data encoding uses One-Hot encoding to convert categorical data into numerical data.

[0046] The data normalization method uses the max-min scaling method to unify numerical data to the range [0, 1].

[0047] Example 3, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, predicts sand and gravel screening efficiency by constructing a sand and gravel screening efficiency prediction model in the sand and gravel screening efficiency prediction module. It includes a working condition feature fusion unit, a dynamic attention calibration unit, a feedforward prediction unit, a time-series feature interaction graph construction unit, a graph enhancement comparison unit, and a model optimization convergence unit; specifically, it includes the following:

[0048] The working condition feature fusion unit addresses the issue of stable and transitional working conditions in sand and gravel screening. Traditional models fail to differentiate between these conditions, leading to insufficient feature extraction and confusion between trend features of stable conditions and abrupt changes in transitional conditions, resulting in large prediction errors. This unit sets a threshold for the comprehensive variance and calculates the comprehensive variance within each time window of length ω for the data in the training dataset, determining the working condition status of the data at each time point. If the comprehensive variance is less than or equal to the variance threshold, the data is classified as stable working condition data, and its working condition status value is marked as 0. If the comprehensive variance is greater than the variance threshold, the data is classified as... Transitional operating condition data is used, and its operating condition state value is marked as 1. Corresponding dynamic features are extracted based on different operating conditions. For stable operating condition data, linear trend features within a window are extracted, while for transitional operating condition data, abrupt change features between adjacent time points are extracted. These are fused to generate an enhanced feature vector. The enhanced sequence data is then sliced ​​using a sliding window method and standardized. This achieves feature matching with operating conditions. The fused enhanced feature vector retains key information under different operating conditions, enabling the model to adapt to fluctuations in the screening process and reduce prediction bias during operating condition switching. The formulas used are as follows:

[0049] ;

[0050] ;

[0051] ;

[0052] In the formula, is the overall variance of the data corresponding to the t-th timestamp, where t is the timestamp index, D is the total number of features, d is the feature index, and ω is the sliding window length, ω=5. It is the value of the d-th feature at the i-th time step, where i is the time step index. It is the mean of the d-th feature within the window corresponding to the t-th timestamp. and S t These are the enhanced feature vector and the original feature vector corresponding to the data at the t-th timestamp, respectively, where T is the vector transpose operation. It is the linear trend feature vector of features within the window under stable operating conditions, with a variance threshold ∈ [0.01, 0.05]. It is the abrupt change feature vector of the t-th timestamp and the (t-1)-th timestamp under the transition condition. St-1 L is the original feature vector of the data corresponding to the (t-1)th timestamp. t It is the operating status value corresponding to the t-th timestamp. It is a standardized augmented data vector. , and These are the standardized enhanced feature vectors corresponding to the t-th, t-1-th, and t-ω+1-th timestamps, respectively. , σ x and μ x These are the variance vector and mean vector of each dimension of the enhanced feature vector within the sliding window, respectively. It is a smoothing term. , It is a Hadama pile;

[0053] Dynamic Attention Calibration Unit: Traditional self-attention mechanisms distribute feature weights equally across all time steps of the screened data, failing to highlight critical time points or important information with drastic feature fluctuations during transitional operating conditions. This leads the model to focus on irrelevant data, limiting prediction accuracy. This unit performs linear projection on the standardized augmented data vectors to generate query, key, and value matrices. It then calculates the basic self-attention and, based on the operating conditions and feature fluctuations at each time step, calculates a criticality score. The attention weights are dynamically calibrated based on this score, generating a calibrated attention output. This allows the model to focus on core data with large feature fluctuations during transitional operating conditions, improving the relevance of the attention output, reducing interference from redundant data in stable operating conditions, ensuring the model's sensitivity to key time-series information affecting efficiency, and improving the real-time performance and accuracy of efficiency prediction. The formulas used are as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] In the formula, Q, K, and V are respectively The query matrix, key matrix, and value matrix obtained by linear projection, d k It is the attention head dimension. It is a normalization function. and These are the basic self-attention output and the calibrated attention output, respectively, and the score. i It represents the key score at the i-th time step within the sliding window, where α is the score weight coefficient, α∈[0.7, 0.9]. and These are the standard deviations of the enhanced feature vectors at the i-th and j-th time steps, respectively, where j is the time step index, and L...i W is the operating condition value corresponding to the i-th time step. calib It is the calibrated attention weight matrix. , It is the criticality score vector [Score1, Score2, ..., Score... ω ] T The matrix after broadcast expansion, Score1, Score2, and Score ω These are the criticality scores for the 1st, 2nd, and ωth time steps within the sliding window, respectively. It is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0058] Feedforward prediction unit: Training solely with mean squared error can easily lead to the model overemphasizing short-term efficiency errors, neglecting the rationality of attention weights, resulting in overfitting to some outliers and reducing the model's applicability in actual screening scenarios. A feedforward neural network is used to perform a nonlinear mapping on the calibrated attention output to obtain the initial screening efficiency prediction value. Combining the mean squared error and the attention weight regularization term, a temporal attention loss function is constructed. This optimizes prediction accuracy while constraining the rationality of weights, avoiding model overfitting, and making the initial prediction value more closely reflect the actual variation of screening efficiency, providing a reliable foundation for subsequent calibration. The formulas used are as follows:

[0059] ;

[0060] ;

[0061] In the formula, This is the initial predicted screening efficiency. It is a feedforward neural network. It is the temporal attention loss function, T train It is the total number of timestamps in the training dataset. and These are the initial screening efficiency prediction value and the actual screening efficiency value corresponding to the t-th timestamp, respectively, where λ is the regularization coefficient, λ∈[0.005, 0.015]. It is the calibrated attention weight of the i-th time step in the sliding window corresponding to the t-th timestamp;

[0062] A temporal feature interaction graph construction unit is used. Complex interaction relationships exist among various features in sand and gravel screening. Traditional models fail to explore the spatial correlations between features, resulting in insufficient modeling of the mechanisms by which multiple parameters synergistically affect efficiency. This unit uses each feature dimension as a graph node, constructs a node set, initializes an embedding vector for each node based on the feature type and corresponding operating condition, calculates the interaction strength between any two nodes, and selects the top C nodes with the highest interaction strength as adjacent nodes for each node, generating an edge set and adjacency matrix. A temporal feature interaction graph is then constructed based on the node set and edge set. This accurately captures the synergistic effects between features such as material properties and equipment status, enabling the model to learn the combined impact of multiple parameters on efficiency, compensating for the shortcomings of focusing only on temporal features and improving the comprehensiveness of feature representation. The formulas used are as follows:

[0063] ;

[0064] ;

[0065] In the formula, and Let be the initial embedding vectors of node u and node v at the t-th timestamp, where u and v are node indices. The feature type corresponding to node u is: material characteristic data, equipment operating status data, and environmental parameter data. Their features correspond to material type, equipment type, and environment type, respectively. yes The weights for the types are as follows: material type, equipment type, and environment type are 1.0, 0.9, and 0.7, respectively. and These are the standardized enhanced feature vectors corresponding to node u and node v at the t-th timestamp, respectively. δ is the embedding layer function, where δ is the transition condition embedding enhancement coefficient, δ∈[0.1, 0.3], and C is the number of adjacent nodes, C∈[5, 12]. γ is the interaction strength between node u and node v at the t-th timestamp, and γ is the eigenvalue difference decay coefficient, γ∈[0.1, 0.5]. It is the maximum L2 norm of the feature vectors of all nodes within timestamp t in the training dataset. It is the L2 norm;

[0066] Graph-enhanced contrast unit: Feature interaction graphs are susceptible to data noise, and the model lacks robustness in learning feature associations, failing to stably capture the dynamic changes in feature interactions under different working conditions, resulting in insufficient spatial feature learning accuracy. For temporal feature interaction graphs, two independent random sampling perturbation coefficients are generated, the perturbation interaction strength is calculated, and the adjacency matrix is ​​updated to obtain two enhanced views. A graph convolutional network is used to aggregate features from the two enhanced views, fusing adjacency node information to update node embeddings. Based on interaction strength and node type, positive and negative sample pairs are constructed from each enhanced view. A spatial contrast loss function is constructed by combining view contrast loss and dynamic interaction consistency loss. This enhances the model's ability to learn feature associations, reduces noise interference, makes spatial feature representation more reliable, and provides accurate spatial dimensional support for comprehensive prediction. The formulas used are as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] In the formula, It represents the perturbation-induced interaction strength between node u and node v in the s-th augmented view at the t-th timestamp, where s is the augmented view index. These are the random sampling perturbation coefficients corresponding to node u and node v in the s-th augmented view. , and These are the output feature vectors of node u in the s-th augmented view at the t-th time stamp, located at the (l+1)-th and 1-th layers of the graph convolutional network. It is the output feature vector of node v in the s-th augmented view at the t-th timestamp, located at the l-th layer of the graph convolutional network, where l is the layer index. It is a linear rectified function. and These are the sets of adjacent nodes of node u and node v in the s-th augmented view at the t-th timestamp. and They are and The number of nodes in and These are the edge type weight matrix and bias weight matrix of the l-th layer of the graph convolutional network, respectively. and These are the positive and negative sample pairs corresponding to node u in the s-th augmented view at the t-th timestamp, respectively. and These are the output feature vectors of node u and node v in the s-th augmented view at the t-th time stamp, respectively, at the L-th layer of the graph convolutional network, where L is the number of layers in the graph convolutional network. It is the type of the feature corresponding to node v. and These are the logical AND operator and the logical OR operator, respectively. , and These are the view contrast loss values ​​corresponding to the s-th, 1st, and 2nd enhanced views, respectively, and B is the number of positive sample pairs sampled in a single image. This is the dynamic interaction consistency loss value, where M is the total number of nodes in the temporal feature interaction graph. It represents the model-predicted interaction strength between nodes u and v at the t-th timestamp. It is a spatial contrast loss function;

[0074] Model optimization convergence unit: Relying solely on a single loss function can easily lead to model bias towards a particular temporal or spatial dimension, resulting in an ambiguous convergence criterion and failing to ensure the model reaches global optimum in the screening efficiency prediction task. Furthermore, the predicted values ​​are not calibrated using feature interactions. By combining the temporal attention loss function and the spatial contrast loss function, a comprehensive loss function is constructed. With the goal of minimizing the comprehensive loss, the Adam optimizer is used to synchronously update all learnable parameters. When the comprehensive loss of the test dataset decreases by less than [a certain value] for five consecutive training iterations, the model is optimized. When the number of training iterations reaches the preset maximum, the model is considered converged and training is stopped. During training, the model parameters corresponding to the minimum overall loss in the test dataset are saved to obtain the trained sand and gravel screening efficiency prediction model. The initial screening efficiency prediction value is calibrated and used as the final output of the sand and gravel screening efficiency prediction model. This ensures that the final prediction value conforms to both temporal variation and spatial correlation patterns, improving prediction accuracy and providing a precise basis for subsequent control parameter optimization. The formula used is as follows:

[0075] ;

[0076] ;

[0077] in, It is a comprehensive loss function. It is the calibrated predicted screening efficiency value at the t-th timestamp.

[0078] By performing the above operations, this solution addresses the problems in existing sand and gravel screening efficiency optimization control systems. These problems include insufficient feature extraction due to a lack of differentiation between working conditions, neglect of complex interactions between features, and a focus solely on temporal features without spatial correlation mining. This results in large prediction errors and low accuracy in sand and gravel screening efficiency, failing to provide a reliable basis for optimization control. This solution determines the working condition based on comprehensive variance, generates enhanced feature vectors to achieve precise matching between features and working conditions, and retains key information under different working conditions. Dynamic calibration is performed based on criticality scoring to enhance sensitivity to key temporal information affecting efficiency. Finally, an initial predicted screening efficiency value is generated, outputting an efficiency that closely matches the actual performance. The initial predicted value of the change pattern is obtained; based on the feature type and corresponding working condition, a time-series feature interaction graph is constructed to improve the comprehensiveness and effectiveness of feature representation; two enhanced views are obtained by generating random sampling perturbation coefficients and constructing positive and negative sample pairs to improve the reliability of spatial feature representation; a comprehensive loss function is constructed by fusing time-series and spatial losses to guide the model to simultaneously optimize the learning effect of time-series and spatial features; a sand and gravel screening efficiency prediction model is obtained, and the initial screening efficiency prediction value is calibrated to output the final prediction value that simultaneously fits the time-series change and spatial correlation pattern, significantly improving the prediction accuracy and providing accurate and reliable data support for subsequent dynamic optimization of control parameters.

[0079] Example 4, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, in the sand and gravel screening control parameter optimization module, constructs a sand and gravel screening control parameter optimization model, dynamically optimizes the adjustable operating parameters of the screening equipment based on reinforcement learning, and generates the optimal operating control parameters for sand and gravel screening. It includes a state-action space setting unit, a reward function setting unit, and an optimal control generation unit; specifically, it includes the following:

[0080] State-Action Space Setting Unit: In traditional control, the state description is one-sided, the range of adjustable parameters is vague, and it does not take into account the characteristics of the screening process, which makes it impossible for the reinforcement learning agent to accurately perceive the scene state and output reasonable control actions. The fusion vector containing material characteristic data, environmental parameter data, equipment operating status data, and operating status markers in the sand and gravel screening process is defined as the state space, and the state vector is obtained. The set of adjustable operating parameters of the screening equipment within its feasible range is defined as the action space, thus obtaining the action vector. The system comprehensively depicts the screening scenario, ensuring accurate perception and compliant actions by the intelligent agent. This guarantees that optimization aligns with the actual operating capabilities of the equipment, preventing equipment malfunctions caused by parameters exceeding limits. The adjustable operating parameters of the screening equipment include screen rotation speed, screen amplitude, motor power, and screen tilt angle. t and a t These are the state vector and action vector corresponding to the t-th timestamp, respectively, F t Et G t and L t These are the material characteristic data vector, equipment operating status data vector, environmental parameter data vector, and operating condition value corresponding to the t-th timestamp, respectively, n t A t P t and These are the screen rotation speed, screen amplitude, motor power, and screen tilt angle corresponding to the t-th timestamp, respectively, and n t ∈[500, 1500] rpm, A t ∈[2, 10] mm, P t ∈[15, 55] kW, ∈[15,30] degree;

[0081] The reward function setting unit addresses the issue that a single reward indicator can easily lead to large fluctuations in equipment parameters, excessive energy consumption, or health problems, failing to meet the multi-objective requirements of "high efficiency, low consumption, and safety" in the screening process. This unit integrates the current screening efficiency achievement rate with the efficiency rise gradient to construct an efficiency-gradient composite reward item; it also integrates the stability of adjustable parameter fluctuations with the equipment health status to construct an equipment health-stability reward item; and it integrates motor energy consumption cost with parameter adjustment cost to construct an energy consumption-adjustment cost reward item. Based on the current operating condition status label, an adaptive weight vector is generated, and the three reward items are weighted and integrated to finally construct a multi-objective weighted reward function. This ensures the reward function aligns with actual production needs, guiding the agent to output control parameters that consider multiple objectives. The formulas used are as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] In the formula, , , and r t These are the efficiency-gradient composite reward term, the equipment health-stability reward term, the energy consumption-adjustment cost reward term, and the multi-objective weighted reward function value at the t-th timestamp, respectively. η is the calibrated predicted screening efficiency value at the (t-1)th timestamp. min and η max These are the minimum and maximum screening efficiencies in historical data, respectively. It is a hyperbolic tangent function, where H is the number of adjustable operating parameters and h is the parameter index. and These are the adjustment values ​​of the h-th adjustable operating parameter at the t-th and t-1-th timestamps, respectively. and These are the maximum and minimum allowable values ​​for the h-th adjustable parameter, respectively. This is the k-th equipment health indicator at the t-th timestamp. The equipment health indicators include motor current and screen temperature. P is the safety threshold for the k-th device health indicator. t P is the motor power at the t-th timestamp. max and P min These are the maximum and minimum allowable values ​​of the motor power, ω. base and ω trans These are the weighted base values ​​for steady-state and transient operating conditions, respectively. , ω t It is the adaptive weight vector of the working condition at the t-th timestamp. , and They are ω t Chinese correspondence , and The weights;

[0088] Optimal control generation unit; traditional control parameter setting relies on experience, cannot be dynamically adjusted according to real-time operating conditions, and is difficult to balance long-term cumulative benefits and short-term efficiency, resulting in poor adaptability of control parameters; based on state vectors, action vectors, and multi-objective weighted reward functions, a reinforcement learning agent based on a proximal policy optimization algorithm is constructed, and the policy network π of this agent... θ With the current state vector q t As input, output action vector a t The probability distribution of the values ​​of each adjustable operating parameter is used to iteratively optimize the policy network parameter θ with the goal of maximizing the expected cumulative discount reward. The optimization is performed when the change in the cumulative discount reward over five consecutive iterations on the test dataset is less than... Training is complete when the preset maximum number of iterations is reached, and the parameters at this point are used as the optimal policy network parameters θ. best The corresponding network is a sand and gravel screening control parameter optimization model, which inputs the state vector into the trained policy network. This function outputs the optimal operating control parameters for sand and gravel screening under the current conditions. The output parameters aim to improve current efficiency while ensuring long-term stable operation of the equipment, thus enhancing the dynamic adaptability and economy of the control parameters. The formula used is as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] Among them, U t It represents the cumulative discount reward starting from time t, where b is the index of a future time step. It is a discount factor. , It is the multi-objective weighted reward function value at the (t+b)th timestamp. It is the objective function for expected cumulative discount rewards. These are the learnable parameters of the policy network. In strategy Sampling trajectory Expectations These are the optimal operating control parameters for sand and gravel screening at the t-th timestamp.

[0094] By performing the above operations, this solution addresses the problems in existing sand and gravel screening efficiency optimization control systems, such as one-sided state descriptions that fail to consider the working conditions of the screening process, single reward indicators, and control parameters that rely heavily on experience and are difficult to dynamically adjust. These issues result in an inability to accurately perceive the real-time screening scenario, leading to low adaptability and accuracy in sand and gravel screening efficiency optimization control. This solution establishes a state space and action space to comprehensively characterize the screening scenario state, ensuring accurate perception of real-time working conditions. It constructs efficiency-gradient composite reward items, equipment health-stability reward items, and energy consumption-adjustment cost reward items, covering key objectives throughout the entire screening production process and providing comprehensive optimization guidance for the agent's learning. A multi-objective weighted reward function is obtained through weighted fusion of working condition adaptive weight vectors, enhancing the working condition adaptability of the reward mechanism and guiding the output of parameters that better meet actual production needs. A reinforcement learning agent based on a near-end strategy optimization algorithm is constructed to overcome experience dependence and achieve intelligent generation of control parameters. The network parameters are optimized with the goal of maximizing the expected cumulative discount reward, outputting the optimal operating control parameters for sand and gravel screening, significantly improving the adaptability and accuracy of sand and gravel screening efficiency optimization control.

[0095] Example 5, see Figure 1This embodiment is based on the above embodiment. In the efficiency optimization control module, real-time screening control timing data is collected. The real-time screening control timing data includes timestamps, material characteristic data, equipment operating status data, and environmental parameter data. After data cleaning, data encoding, and data normalization processing, the real-time screening control timing data is sequentially input into the trained sand and gravel screening efficiency prediction model and sand and gravel screening control parameter optimization model to obtain the optimal operating control parameters for sand and gravel screening. These parameters are then sent to the programmable logic controller of the sand and gravel screening equipment to realize real-time adaptive optimization control of sand and gravel screening efficiency.

[0096] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0098] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent sand screening efficiency optimization control system, characterized in that: The sandstone screening efficiency prediction model is constructed, the sandstone screening efficiency is predicted, and a working condition feature fusion unit, a dynamic attention calibration unit, a feedforward prediction unit, a time sequence feature interaction graph construction unit, a graph enhancement comparison unit and a model optimization convergence unit are arranged. The screening control data integration module collects historical screening control time series data and pre-processes the time series data, constructs a training data set and a test data set, and determines a working condition state according to a comprehensive variance. The sandstone screening efficiency prediction module generates an enhanced feature vector according to the comprehensive variance, the working condition state includes a transition working condition and a stable working condition, performs dynamic calibration based on a key degree score, generates an initial screening efficiency prediction value, constructs a time sequence feature interaction graph according to a feature type and a corresponding working condition state, material characteristic data, equipment operation state data and environmental parameter data are respectively corresponding to a material type, an equipment type and an environmental type, generates a random sampling disturbance coefficient to obtain two enhanced views, constructs a positive and negative sample pair, fuses a time sequence and a space loss to construct a comprehensive loss function, obtains a sandstone screening efficiency prediction model, and calibrates the initial screening efficiency prediction value. The sandstone screening control parameter optimization module sets a state space and an action space, constructs an efficiency-gradient compound reward item, an equipment health-stability reward item and an energy consumption-regulation cost reward item, fuses the items based on a working condition adaptive weight vector to obtain a multi-objective weighted reward function, constructs a reinforcement learning intelligent agent based on a proximal policy optimization algorithm, optimizes network parameters to maximize an expected cumulative discount reward, and outputs sandstone screening optimal operation control parameters. The efficiency optimization control module sends the sandstone screening optimal operation control parameters to a programmable logic controller of a sandstone screening device, and realizes real-time self-adaptive optimization control of the sandstone screening efficiency.

2. The intelligent grit screening efficiency optimization control system of claim 1, wherein: The sandstone screening efficiency prediction module, by constructing a sandstone screening efficiency prediction model, predicts the sandstone screening efficiency, and is provided with a working condition feature fusion unit, a dynamic attention calibration unit, a feedforward prediction unit, a time sequence feature interaction graph construction unit, a graph enhancement comparison unit and a model optimization convergence unit. The working condition feature fusion unit; A threshold of the comprehensive variance is set, the comprehensive variance in a window corresponding to each time is calculated according to a sliding time window, and the working condition state corresponding to the data at each time is determined; according to different working condition states, corresponding dynamic features are extracted, linear trend features in the window of the stable working condition data are extracted, and mutation features between adjacent time steps of the transition working condition data are extracted, and an enhanced feature vector is generated by fusion; the enhanced sequence data is sliced in a sliding window manner, and standardized processing is performed; The dynamic attention calibration unit; The standardized enhanced data vector is linearly projected to generate a query matrix, a key matrix and a value matrix, a basic self-attention is calculated, a key degree score is calculated according to the working condition state and the feature fluctuation information of each time step, the attention weight is dynamically calibrated based on the key degree score, and a calibrated attention output is generated; The feedforward prediction unit; The time sequence feature interaction graph construction unit; The graph enhancement comparison unit; The model optimization convergence unit; by combining the temporal attention loss function and the spatial contrast loss function, a comprehensive loss function is constructed, the Adam optimizer is used to update all learnable parameters synchronously to minimize the comprehensive loss, and the trained sand and stone screening efficiency prediction model is obtained. The initial screening efficiency prediction value is calibrated as the final output of the sand and stone screening efficiency prediction model.

3. The intelligent grit screening efficiency optimization control system of claim 2, wherein: The feedforward prediction unit performs nonlinear mapping on the calibrated attention output through a feedforward neural network to obtain an initial screening efficiency prediction value, and constructs a temporal attention loss function by combining mean square error and attention weight regularization terms.

4. The intelligent grit screening efficiency optimization control system of claim 3, wherein: The time series feature interaction graph construction unit constructs a node set with each feature dimension as a graph node, initializes an embedding vector for each node according to the feature type and the corresponding working condition state, calculates the interaction strength between any two nodes, and selects the node with the optimal interaction strength for each node as the adjacent node to generate an edge set and an adjacency matrix. The time series feature interaction graph is constructed based on the node set and the edge set.

5. The intelligent grit screening efficiency optimization control system of claim 4, wherein: The graph enhancement contrast unit generates two independent random sampling perturbation coefficients for the time series feature interaction graph, calculates the perturbed interaction strength, and updates the adjacency matrix to obtain two enhanced views. Feature aggregation is performed on the two enhanced views through a graph convolution network, the node embedding is updated by fusing the adjacent node information, and the positive sample pairs and negative sample pairs are constructed from each enhanced view based on the interaction strength and the node type. The spatial contrast loss function is constructed by combining the view contrast loss and the dynamic interaction consistency loss.

6. The intelligent grit screening efficiency optimization control system of claim 5, wherein: The sand and stone screening control parameter optimization module optimizes the controllable operating parameters of the screening equipment based on reinforcement learning by constructing a sand and stone screening control parameter optimization model to generate optimal sand and stone screening operation control parameters. It is provided with a state action space setting unit, a reward function setting unit and an optimal control generation unit; specifically including the following contents: The state action space setting unit defines the fusion vector containing material property data, environmental parameter data, equipment operating state data and working condition state markers in the sand and stone screening process as the state space to obtain the state vector; and defines the value set of the controllable operating parameters of the screening equipment in its feasible interval as the action space to obtain the action vector. The reward function setting unit; The optimal control generation unit.

7. The intelligent grit screening efficiency optimization control system of claim 6, wherein: The reward function setting unit fuses the current screening efficiency compliance degree and the efficiency rising gradient to construct an efficiency-gradient composite reward term; fuses the controllable parameter fluctuation stability and the equipment health state to construct a device health-stability reward term; and fuses the motor energy consumption cost and the parameter adjustment cost to construct an energy consumption-adjustment cost reward term. Based on the current working condition state marker, a working condition adaptive weight vector is generated, and the three reward terms are weighted and fused to finally construct a multi-objective weighted reward function.

8. The intelligent grit screening efficiency optimization control system of claim 7, wherein: The optimal control generation unit constructs a reinforcement learning agent based on a proximal policy optimization algorithm based on the state vector, the action vector and the multi-objective weighted reward function. The policy network of the agent takes the current state vector as input and outputs the probability distribution of the value of each adjustable operating parameter in the action vector, so as to maximize the expected cumulative discounted reward and iteratively optimize the policy network parameters to obtain optimal policy network parameters. The corresponding network is a sandstone screening control parameter optimization model. The state vector is input into the trained policy network, and the recommended optimal operation control parameters of sandstone screening under the current state are output.

9. The intelligent aggregate screening efficiency optimization control system of claim 8, wherein: The screening control data integration module collects and pre-processes historical screening control time series data, divides the pre-processed data into training data set and test data set; the historical screening control time series data includes timestamp, material characteristic data, equipment operating state data, environmental parameter data and screening efficiency, and the screening efficiency is taken as data label; The preprocessing includes data cleaning, data encoding and data normalization.

10. The intelligent grit screening efficiency optimization control system of claim 9, wherein: The efficiency optimization control module collects real-time screening control time series data, which includes timestamp, material characteristic data, equipment operating state data and environmental parameter data. After data cleaning, data encoding and data normalization processing, the real-time screening control time series data is input into the trained sandstone screening efficiency prediction model and sandstone screening control parameter optimization model in turn, to obtain the optimal operation control parameters of sandstone screening, and is issued to the programmable logic controller of the sandstone screening equipment, to realize real-time self-adaptive optimization control of sandstone screening efficiency.

Images