Intelligent process monitoring and feedback method and system for resource processing of construction waste

Abstract

The present invention relates to the technical field of intelligent process monitoring, and discloses an intelligent process monitoring and feedback method and system for resource processing of construction waste. The method comprises: collecting and standardizing process parameters of multiple monitoring points of a construction waste processing production line to obtain a preprocessing data set; constructing a graph data structure, calculating the correlation coefficient between nodes to determine the node connection relationship, and dynamically optimizing through a capability-aware graph neural network to obtain a process data association graph; inputting the process data association graph into a process parameter association model to perform graph convolution operations and attention feature extraction to obtain a process parameter feature vector; inputting the process parameter feature vector into a multi-constrained non-convex optimizer to calculate the joint optimization of processing efficiency, resource utilization and product quality to obtain a process parameter regulation strategy. The present invention realizes intelligent process monitoring of resource processing of construction waste and effectively integrates the optimal control strategies under different working conditions.

Description

Technical Field

[0001] The present invention relates to the field of intelligent process monitoring, and in particular to an intelligent process monitoring and feedback method and system for resource processing of construction waste. Background Art

[0002] Traditional methods of construction waste treatment mainly rely on manual experience to adjust process parameters and lack scientific monitoring and feedback mechanisms.

[0003] In the current construction waste treatment process, the treatment effect fluctuates greatly due to the complex material composition and unstable moisture content. The process parameter monitoring data has noise and outliers, and the correlation between the monitoring points is complex, making it challenging to accurately grasp the process status and adjust the treatment parameters in time. In addition, the regulation of process parameters needs to take into account multiple goals such as treatment efficiency, resource utilization and product quality. The traditional single optimization method is difficult to achieve multi-objective coordinated optimization. Summary of the invention

[0004] The present invention provides an intelligent process monitoring and feedback method and system for resource processing of construction waste. The present invention realizes intelligent process monitoring of resource processing of construction waste and effectively integrates optimal control strategies under different working conditions.

[0005] In a first aspect, the present invention provides an intelligent process monitoring and feedback method for resource processing of construction waste, the intelligent process monitoring and feedback method for resource processing of construction waste includes:

[0006] Collecting process parameters at multiple monitoring points of a construction waste treatment production line to obtain an original monitoring data set, and removing outliers and standardizing the original monitoring data set to obtain a preprocessed data set;

[0007] Constructing a graph data structure based on the preprocessed data set and the plurality of monitoring points, calculating the correlation coefficient between nodes to determine the node connection relationship, and performing dynamic optimization through a capability-aware graph neural network to obtain a process data association graph;

[0008] Inputting the process data association graph into the process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector;

[0009] The process parameter feature vector is input into a multi-constrained non-convex optimizer to calculate the joint optimization of processing efficiency, resource utilization and product quality to obtain a process parameter control strategy.

[0010] In a second aspect, the present invention provides an intelligent process monitoring and feedback system for resource processing of construction waste, the intelligent process monitoring and feedback system for resource processing of construction waste includes:

[0011] The acquisition module is used to collect process parameters at multiple monitoring points of the construction waste treatment production line to obtain an original monitoring data set, and to remove outliers and perform standardization on the original monitoring data set to obtain a pre-processed data set;

[0012] A construction module, used to construct a graph data structure based on the preprocessed data set and the multiple monitoring points, calculate the correlation coefficient between nodes to determine the node connection relationship, and perform dynamic optimization through a capability-aware graph neural network to obtain a process data association graph;

[0013] An extraction module, used for inputting the process data association graph into a process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector;

[0014] The optimization module is used to input the process parameter feature vector into a multi-constrained non-convex optimizer, calculate the joint optimization of processing efficiency, resource utilization and product quality, and obtain a process parameter control strategy.

[0015] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned intelligent process monitoring and feedback method for resource processing of construction waste.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored, which, when executed on a computer, enable the computer to execute the above-mentioned intelligent process monitoring and feedback method for resource processing of construction waste.

[0017] In the technical solution provided by the present invention, the process parameters are dynamically optimized through the capability-aware graph neural network, unreliable monitoring data are effectively identified and filtered, and data quality and reliability are improved; a multi-level feature extraction structure is adopted, including graph convolution operations and multiple attention mechanisms, to deeply explore the correlation between process parameters and enhance the characterization capability of complex process states; through multi-objective optimization, the coordinated optimization of processing efficiency, resource utilization and product quality is achieved, and the scientific nature of the optimization results is ensured; a complete process parameter monitoring and feedback adjustment closed loop is constructed, and the real-time and accuracy of process parameter regulation are ensured through continuous data collection, analysis and optimization; a hierarchical and progressive processing architecture is adopted, from data preprocessing, graph structure construction, feature extraction to parameter optimization, to form a systematic process monitoring scheme, which improves the reliability and robustness of the present invention; a strategy integration method based on cluster division and importance weighting effectively integrates the optimal control strategies under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 A schematic diagram of the steps of the intelligent process monitoring and feedback method for resource processing of construction waste in an embodiment of the present invention;

[0020] Figure 2 It is a structural schematic diagram of an intelligent process monitoring and feedback system for resource processing of construction waste in an embodiment of the present invention;

[0021] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The embodiment of the present invention provides an intelligent process monitoring and feedback method and system for resource processing of construction waste. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 An embodiment of the intelligent process monitoring and feedback method for resource processing of construction waste in the embodiment of the present invention includes:

[0024] Step S1, collecting process parameters at multiple monitoring points of the construction waste treatment production line to obtain an original monitoring data set, and performing outlier removal and standardization processing on the original monitoring data set to obtain a preprocessed data set;

[0025] It is understandable that the execution subject of the present invention may be an intelligent process monitoring and feedback system for resource processing of construction waste, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0026] Specifically, monitoring sensors are set up at each key link of the construction waste treatment production line, including the feed end, crushing end, screening end, conveying end and finished product end. Through these sensors, data on a series of process parameters such as material particle size, moisture content, impurity content, equipment load, processing volume and energy consumption are collected in real time to obtain the original collected data. The original collected data are grouped according to the time series, and each group of data generates a unique identifier according to its collection time to ensure that the data can be effectively managed and analyzed according to the time sequence and logical relationship to obtain a labeled data set. The labeled data set is segmented to form time series segmented data, which represents the dynamic changes of process parameters within a certain time range. Through statistical analysis of the time series segmented data, the statistical characteristics of each process parameter in each time period, such as mean, variance, kurtosis and skewness, are obtained to reflect the distribution characteristics, fluctuations and abnormal trends of the data. The feature data is detected for outliers according to the set threshold interval. By comparing whether the statistical feature value exceeds the preset reasonable range, if the statistical characteristics of a data point exceed these threshold intervals, the data point is marked as an outlier, and these outliers are organized into an outlier tag set. Through the marker set, data points that do not conform to normal process fluctuations are identified and eliminated in the data cleaning stage to ensure the quality of the final data set. On this basis, the abnormal data points in the original collected data are eliminated to obtain a cleaned data set. The maximum and minimum values ​​of each parameter in the cleaned data set are normalized. By converting each parameter data into a unified standard range, usually [0,1] or [-1,1], the influence of different data scales on subsequent analysis models is eliminated, making data processing more efficient and consistent. The normalized data can avoid analysis deviations caused by large differences in data ranges and provide standardized input for subsequent process parameter optimization. The normalized data is organized and stored according to the correspondence between the monitoring points and the process links to form a complete preprocessing data set.

[0027] Step S2: construct a graph data structure based on the preprocessed data set and multiple monitoring points, calculate the correlation coefficient between nodes to determine the node connection relationship, and dynamically optimize through the capability-aware graph neural network to obtain a process data association graph;

[0028] Specifically, each monitoring point in the preprocessed data set is constructed according to its process position in the production line to construct a node feature matrix. Each node represents a monitoring point, and the node features are various process parameters related to the monitoring point, such as material particle size, moisture content, etc. Based on the time series data, these node features are extracted. The extraction process captures the temporal trend of the process parameters, and the initial node representation of each node is obtained, reflecting the dynamic characteristics of the monitoring point in the time dimension. The Pearson correlation coefficient is calculated between any two monitoring points in the initial node representation to measure the similarity between them. The Pearson correlation coefficient is an indicator for measuring linear correlation, which effectively reveals the relationship between different monitoring points. A correlation coefficient threshold λ is set. If the correlation coefficient of two nodes is greater than the threshold, it means that the relationship between the two monitoring points is relatively close, and connecting edges should be established in the graph, thereby obtaining a preliminary graph structure, in which nodes represent monitoring points and edges represent the correlation between monitoring points. The preliminary graph structure is input into the capability-aware graph neural network for optimization. The capability-aware graph neural network includes three layers of graph convolution layers, two layers of attention layers, and two layers of fully connected layers. The function of the graph convolution layer is to update the features of the nodes by aggregating the information of the adjacent nodes, so as to learn a richer node representation. The initial graph structure is input to the graph convolution layer through the adjacency matrix. The graph convolution layer weights and sums the features of the adjacent nodes according to the connection relationship between the nodes, and then uses the ReLU activation function to perform a nonlinear transformation to obtain the updated node embedding features. The node embedding features contain the attributes of the node itself and the influence of its neighboring nodes, which can better reflect the process characteristics of the node. The node embedding features are input to the attention layer. The attention mechanism adaptively assigns weights according to the relationship between nodes. By calculating the attention weight matrix between nodes, it identifies which nodes are more important to the current task. The attention weight matrix is ​​normalized by the softmax function to ensure that the sum of the attention weights of all nodes is 1. The node features are weighted and aggregated using the normalized attention weights to obtain the attention enhancement features. The attention enhancement features are input to the fully connected layer. The fully connected layer calculates the reliability score of each node through the sigmoid activation function, which reflects the importance and credibility of the node in the entire process system. Nodes with low node reliability scores are considered unreliable nodes. According to the reliability score, the importance of nodes in the initial graph structure is re-evaluated, and nodes with scores below the set threshold β are marked as unreliable nodes. For unreliable nodes, their connecting edges are weighted attenuated to reduce their influence in the graph, and the features of unreliable nodes are replaced with the weighted average of their reliable neighboring nodes to ensure that the impact of unreliable nodes in the graph structure on the entire process is minimized, and an optimized process data association graph is obtained.

[0029] Step S3: input the process data association graph into the process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector;

[0030] Specifically, the process data association graph is input into the process parameter association model for graph convolution operation and attention feature extraction. The nodes and edges in the process data association graph represent the relationship between the monitoring points of the production line. The model extracts the deep features of the nodes through the graph convolution neural network. The process parameter association model uses a three-layer graph convolution network with residual connection for feature aggregation. This design helps to avoid the gradient vanishing problem in deep network training. Each layer of the graph convolution network contains 32 convolution kernels with a convolution kernel size of 3×3. The neighborhood features of the nodes are aggregated through graph convolution operation, and the ReLU activation function is used to perform nonlinear transformation on the output of each layer of convolution to obtain the initial feature map. The initial feature map is input into the two-layer spatial attention network in the process parameter association model for modeling the relationship between nodes. Each layer of the spatial attention network contains 8 attention heads, and the feature dimension of each attention head is 64. By calculating the attention weights between nodes, the spatial attention network can automatically focus on the nodes and connections that are most important to the process and filter out key information. In order to ensure the stability and efficiency of training, the network uses LayerNorm for normalization at each layer to ensure that the scales of features of different nodes are consistent and avoid the problem of non-convergence of training due to large feature differences. In this way, the spatial attention network can effectively extract the spatial correlation between nodes and generate spatial attention features. The target area is screened for the spatial attention features. According to the specific location information of the node in the process flow, the regional mask matrix is ​​set. The mask matrix filters out regional features that are irrelevant to the current target according to the physical layout of the process flow and the logical relationship of the production process to obtain regional attention features. The regional attention features are input into the two-layer temporal attention network in the process parameter association model. The role of the temporal attention network is to model the changes of process parameters at different time scales and capture the dynamic changes in time. Each layer of the temporal attention network contains 4 attention heads and the feature dimension is 32. Through the temporal attention mechanism, the model can identify the dependency between process parameters at different time points and provide temporal related features for subsequent process prediction and regulation. The temporal related features are input into the two-layer graph pooling layer in the process parameter association model for feature dimensionality reduction. The graph pooling layer aggregates the features of adjacent nodes in the graph structure to reduce the dimension of the features, thereby reducing the computational burden and retaining important information. The compressed features after pooling are input into the three-layer fully connected network for feature fusion. The fully connected network fuses the features of each layer by weighted summation to produce a more compact and efficient fused feature vector. The model performs feature selection on the fused feature vector. The most important features are selected through a certain algorithm, and redundant and irrelevant features are eliminated to obtain the core feature vector. The core feature vector is mapped to finally obtain the process parameter feature vector.

[0031] Step S4: input the process parameter feature vector into a multi-constrained non-convex optimizer, calculate the joint optimization of processing efficiency, resource utilization and product quality, and obtain the process parameter control strategy.

[0032] Specifically, the process parameter feature vector is input into the Weighted-K-Means clusterer in the multi-constrained non-convex optimizer to partition the parameter space. The clusterer calculates the similarity between samples by weighted Euclidean distance and divides the process parameter space into K process parameter clusters. In this process, the center of each cluster represents the characteristics of a class of similar process parameters. The centers of the clustered K process parameter clusters provide important initial information for subsequent optimization and decision-making. The centers of the K process parameter clusters are state-encoded. In the state encoding stage, the three key performance indicators of processing efficiency, resource utilization and product quality are encoded into state vectors, and these state vectors are integrated together to form a multi-dimensional state representation. The multi-dimensional state representation is input into the multi-constrained non-convex optimizer and optimized using the Actor network of the DDPG (Deep Deterministic Policy Gradient) algorithm. The Actor network generates a set of reasonable process parameter control actions based on the current state information. In this process, the Actor network consists of three fully connected layers with feature dimensions of [256, 128, 64], and each layer is nonlinearly transformed by the tanh activation function. After these processes, the Actor network can generate a set of effective control actions, which represent the optimal process parameter adjustment direction under a given state. The control action and the multi-dimensional state representation are combined and input into the Critic network of the DDPG algorithm. The role of the Critic network is to evaluate the value of a given control action in the current state, that is, to evaluate the long-term effect of the action. The Critic network consists of four fully connected layers with feature dimensions of [512, 256, 128, 64]. Through these layers of processing, the Critic network can accurately evaluate the value of the action. After evaluating the action value, the temporal difference error is calculated based on this, and the parameters of the Actor network and the Critic network are updated through the back propagation algorithm to gradually optimize the process control strategy. During the optimization process, the reward calculation is performed for the optimized control strategy. The design of the reward function considers three key factors: processing efficiency, resource utilization, and product quality. By setting the weight of each factor, the weight of the processing efficiency is set to w1, the weight of the resource utilization is set to w2, and the weight of the product quality is set to w3, and the weighted reward value is obtained. According to the obtained weighted reward value, the gradient ascent optimization control parameter calculation is performed for each process parameter cluster to obtain the local optimal strategy. The gradient ascent method can effectively adjust the process parameters according to the reward value so that the production process reaches the optimal state in the local range. The local optimal strategies of each cluster are integrated and weighted according to the importance weight of the cluster to form a comprehensive process parameter control strategy.

[0033] The process parameter feature vector is scored for parameter importance to determine the degree of influence of each process parameter in the entire process. By evaluating the contribution of each parameter to the process result, weight coefficients are assigned to different parameters to obtain the parameter weight vector. A weighted distance measurement function is constructed based on the parameter weight vector. This function is used to calculate the weighted distance between samples, so as to adjust the distance measurement method according to the weight difference of each sample in each parameter dimension. K samples are randomly selected as the initial clustering center for preliminary clustering. The remaining samples are assigned to the cluster with the closest distance according to the weighted Euclidean distance calculation to obtain the initial clustering division. After the initial division is completed, the weighted centroid is calculated for each process parameter cluster. The centroid is obtained based on the weighted average of all samples in the cluster, reflecting the central tendency of the cluster. According to the updated cluster center, the weighted distance between each sample and the cluster center is recalculated, and then the sample is redistributed to the cluster with the closest distance to complete the optimized clustering division. In order to evaluate the current clustering results, the intra-class sum of squares of the optimized clustering division is calculated as the evaluation index of clustering. The intra-class sum of squares reflects the compactness of samples within the cluster. The smaller the value, the more concentrated the samples and the better the clustering effect. Based on the calculation results of the intra-class sum of squares, it is judged whether the cluster division has converged. If the change in the cluster center is less than the set change threshold, or the maximum number of iterations is reached, the clustering process is considered to be completed, the iteration is stopped, and the final cluster division is obtained. The weighted center point coordinates are calculated for each process parameter cluster in the final cluster division. The weighted center point coordinates represent the typical process parameter combination in the cluster, which can effectively summarize and express the overall characteristics of the cluster, and finally obtain K process parameter cluster centers.

[0034] The gradient ascent method is applied to calculate the parameter gradient of the weighted reward value to determine the optimization direction. The weighted reward value comprehensively considers factors such as processing efficiency, resource utilization and product quality, and can provide clear feedback for the optimization goal. In this process, the gradient ascent method obtains the direction of parameter update by calculating the gradient of the reward value to the process parameters. The process parameters are iteratively updated based on the optimization direction to adjust the key parameters in the process. By continuously updating the process parameters, an updated parameter sequence is obtained, which can guide the process to adjust to the optimal state and gradually optimize the various performances of the construction waste treatment process. The updated parameter sequence is input into the target optimization function for optimization, and the parameter change between two adjacent iterations is calculated, and it is checked whether the change is less than the set parameter threshold. If the parameter change is less than the threshold or the maximum number of iterations is reached, the optimization process is considered to have converged, the iteration is stopped, and the local optimal strategy is obtained. In this process, the target optimization function quantifies the changes in each iteration to ensure that the optimization process can be stopped at the right time, thereby avoiding unnecessary calculations caused by excessive iterations and ensuring the efficiency and practicality of process parameter regulation. According to the local optimal strategy, the equipment parameters of each process link are constrained and checked to ensure that the updated process parameters are feasible in actual production and meet the physical and operational limitations of the process flow. The process parameters involved in the test include feed rate, crushing gap, screening amplitude and conveyor belt speed, etc. These parameters have their own value ranges. By constraining these process parameters, it is ensured that there will be no over-adjustment or non-conformity with process requirements during the optimization process, and a parameter combination that meets the actual production is obtained. On the basis of satisfying the constraints, the weight coefficient of each process parameter cluster is calculated. These weight coefficients reflect the contribution of each cluster to the overall process optimization. By calculating the weight coefficient of the cluster, the influence of different clusters is quantified. The calculated weight coefficient is normalized to ensure that the weight of each cluster is between [0,1] to avoid excessive or too small weight affecting the final control strategy. The global optimal parameter combination is obtained by linearly combining the standardized cluster weight coefficient with the local optimal strategy of each cluster. The linear combination weights the optimization results of each cluster according to their importance to obtain a fully optimized process parameter set. Based on the global optimal parameter combination, a control instruction sequence is generated to specifically guide the operation of equipment such as feeders, crushers, vibrating screens and conveyor belts. The control instruction sequence includes the control values ​​of process parameters such as feeder speed, crusher gap, vibrating screen amplitude and conveyor belt speed. Through these control instructions, the equipment operation status is dynamically adjusted to achieve precise control of process parameters.

[0035] In the embodiment of the present invention, the process parameters are dynamically optimized through the capability-aware graph neural network, and unreliable monitoring data is effectively identified and filtered, thereby improving data quality and reliability. A multi-level feature extraction structure is adopted, including graph convolution operations and multiple attention mechanisms, to deeply explore the correlation between process parameters and enhance the characterization capability of complex process states. Through multi-objective optimization, the coordinated optimization of processing efficiency, resource utilization and product quality is achieved, and the scientific nature of the optimization results is ensured. A complete process parameter monitoring and feedback adjustment closed loop is constructed, and the real-time and accuracy of process parameter regulation are ensured through continuous data collection, analysis and optimization. A hierarchical and progressive processing architecture is adopted, from data preprocessing, graph structure construction, feature extraction to parameter optimization, to form a systematic process monitoring scheme, thereby improving the reliability and robustness of the present invention. The strategy integration method based on cluster division and importance weighting effectively integrates the optimal control strategies under different working conditions.

[0036] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0037] Monitoring sensors are installed at the feeding end, crushing end, screening end, conveying end and finished product end of the production line to collect data on material particle size, moisture content, impurity content, equipment load, processing capacity and energy consumption indicators to obtain original collected data;

[0038] The original collected data is grouped according to the time series, and a unique identifier is generated for each group of data to obtain a labeled data set;

[0039] Segment the labeled data set to obtain time series segmented data, and calculate the mean, variance, kurtosis and skewness of each parameter based on the time series segmented data to obtain a statistical feature set;

[0040] A threshold interval is set for the statistical feature set, and data exceeding the threshold interval is marked as an outlier to obtain an outlier mark set, and outlier data points in the original collected data are removed according to the outlier mark set to obtain a cleaned data set;

[0041] The parameters in the cleaned data set are normalized to their maximum and minimum values ​​to obtain normalized data, and the normalized data are organized and stored according to the corresponding relationship between the monitoring points and the process links to obtain the preprocessed data set.

[0042] Specifically, all process links of the production line are comprehensively monitored and data collected. Corresponding monitoring sensors need to be set up at the feeding end, crushing end, screening end, conveying end and finished product end. These sensors can collect key indicator data in the production process, including material particle size, moisture content, impurity content, equipment load, processing volume and energy consumption, etc. These data constitute the original collection data set, reflecting the operating status and processing efficiency of the production line. Group the data according to the time series. Each group of data is classified according to its collection time to ensure the timeliness and sequence of the data. For each group of data, a unique identifier is generated to identify the position of this group of data in the entire time series. The obtained labeled data set contains the production line process parameters and time information corresponding to each time point. Data segmentation is performed on the labeled data set. The time series data is divided into multiple continuous sub-segments, each of which represents the data change trend within a certain time range. Based on each time series segment data, the statistical characteristics of each parameter are calculated, including mean, variance, kurtosis and skewness, etc., to reveal the basic distribution law of the data. The mean, variance, kurtosis and skewness of each parameter are used as the statistical feature set representing the parameter. Statistical features can help identify the normal fluctuation range of data and whether there are obvious abnormal fluctuations. Based on the statistical feature set, a threshold interval is set, and the part of the data that exceeds the threshold interval is marked as an outlier. The detection of outliers is based on the threshold range set based on experience or prior knowledge, or automatically determined according to the distribution of data. By setting the threshold interval, outliers are automatically identified to form an outlier tag set. The detection of outliers uses the following formula:

[0043] Outliers ;

[0044] in, represents the data points in the original data, is the mean value of the parameter, is the standard deviation of the parameter, is a constant, usually 2 or 3, used to control the range of anomalies. Points outside this range are marked as outliers. After marking the outliers, the abnormal data points in the original data set are removed to obtain the cleaned data set. The cleaned data set is normalized to ensure that the data scale is consistent between different process parameters. The normalization method uses maximum and minimum value normalization to map the data values ​​to the interval [0,1] according to their maximum and minimum values. The normalized material particle size is calculated by the following formula:

[0045] ;

[0046] in, is the original data, and are the minimum and maximum values ​​in the data set, respectively. is the normalized value. After normalization, all data are converted into a unified standard, which is convenient for subsequent model training and optimization. The normalized data are organized and stored according to the correspondence between the monitoring points and the process links to form a preprocessing data set. The normalized data of each monitoring point will be combined with its corresponding process link to form a complete data set containing the information of each monitoring point.

[0047] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0048] The preprocessed data set is used to construct a node feature matrix according to the process position of the monitoring point, and the node feature matrix is ​​subjected to time series feature extraction to obtain the initial node representation;

[0049] Calculate the Pearson correlation coefficient between any two monitoring points in the initial node representation, set the correlation coefficient threshold λ, establish connecting edges for node pairs with correlation coefficients greater than the correlation coefficient threshold λ, and obtain the initial graph structure;

[0050] The initial graph structure is input into the capability-aware graph neural network, which includes three graph convolutional layers, two attention layers, and two fully connected layers.

[0051] The initial graph structure is input into the three-layer graph convolution layer, the node features are aggregated through the adjacency matrix, and the ReLU activation function is used for nonlinear transformation to obtain the node embedding features;

[0052] The node embedding features are input into two attention layers, the attention weight matrix between nodes is calculated, the attention scores are normalized by the softmax function, and the node features are weighted and aggregated to obtain the attention enhancement features;

[0053] The attention enhancement features are input into two fully connected layers, and the reliability score of each node is calculated through the sigmoid function to obtain the node reliability vector;

[0054] Re-evaluate the importance of nodes in the initial graph structure according to the node reliability vector, mark nodes with reliability scores lower than the score threshold β as unreliable nodes, and perform weight decay on the connecting edges of unreliable nodes to obtain the target edge weight;

[0055] The adjacency matrix of the initial graph structure is updated based on the target edge weights, and the features of unreliable nodes are replaced with the weighted average of their reliable neighbor nodes to obtain the process data association graph.

[0056] Specifically, each monitoring point in the preprocessed data set is used to construct a node feature matrix according to its position in the process flow. The various process parameters collected at each monitoring point (such as material particle size, moisture content, equipment load, etc.) constitute the feature vector of the node. Each row in the feature matrix represents the feature vector of a monitoring point, and each column corresponds to a specific process parameter. Time series feature extraction is performed on the node features to capture the regularity of changes in process parameters over time. More useful features are mined from the original time series data to help improve the performance of subsequent graph neural networks. Assuming that the feature data of each monitoring point is a time series, the time series features representing the characteristics of the node are extracted by performing statistical calculations (such as mean, variance, etc.) on the data in each time period. For a monitoring point, the time window width is set to , and calculate its mean and variance using the following formula:

[0057] ;

[0058] in, Indicates The data at the moment, is the mean value in the time series window, The time series feature extraction process generates a more stable time representation for each monitoring point, making the original time series data easier to analyze and calculate in the subsequent processing. Calculate the correlation coefficient between any two monitoring points in the initial graph structure, and determine whether there is a connection between the nodes based on the size of the correlation coefficient. Use the Pearson correlation coefficient to measure the similarity between the features of two monitoring points. The calculation formula of the Pearson correlation coefficient is as follows:

[0059] ;

[0060] in, Indicates and The Pearson correlation coefficient between the monitoring points is and Monitoring points and At the moment The characteristic value of and is their mean. If the correlation coefficient between the nodes is greater than the set threshold value, a connection edge will be established between the two nodes to obtain the initial graph structure. This process is judged by the following conditions:

[0061] ;

[0062] The initial graph structure after the connection is established contains multiple nodes and edges. Each node represents a monitoring point, and the edge represents the correlation between monitoring points. The initial graph structure is input into the capability-aware graph neural network for processing. The capability-aware graph neural network consists of three graph convolution layers, two attention layers, and two fully connected layers. In the graph convolution layer, the network aggregates node features through the adjacency matrix and uses the ReLU activation function for nonlinear transformation to obtain the embedded features of the nodes. The calculation of the graph convolution layer is expressed as:

[0063] ;

[0064] in, Indicates The node characteristics of the layer, is the adjacency matrix, is the weight matrix of this layer. In this way, the graph neural network can effectively aggregate and update node features and extract high-order features of nodes. The node embedding features are input into the two-layer attention layer to calculate the attention weight matrix between nodes. In this process, the importance of the node in the aggregation process is determined by calculating the attention weight between nodes. The attention weight matrix is ​​normalized by the Softmax function to obtain the weighted aggregation feature of each node. Assume that the attention weight between nodes is , the calculation of the Softmax function is as follows:

[0065] ;

[0066] in, For Node and nodes The normalized attention weights between Is a node The neighbor node set of . In this way, the network can give higher weights to important nodes and enhance the aggregation effect of important features. The attention enhancement features are input into two fully connected layers, and the Sigmoid activation function is used to calculate the reliability score of each node. The calculation formula of the reliability score is:

[0067] ;

[0068] in, Is a node The reliability score, represents the Sigmoid activation function, and is the weight matrix of the fully connected layer, Is a node Based on the reliability scores of the nodes, the nodes in the initial graph structure are re-evaluated and the nodes with reliability scores below the threshold are marked. The node is an unreliable node. The weight of the connecting edge of the unreliable node is decayed to obtain the target edge weight. The calculation of the target edge weight is expressed by the following formula:

[0069] ;

[0070] in, Is a node and nodes The original edge weights between is an indicator function, indicating that when the node The reliability score is below the threshold The weight of the edge will decay when . The adjacency matrix of the initial graph structure is updated based on the target edge weight, and the features of the unreliable nodes are replaced by the weighted average of their reliable neighbor nodes to obtain the process data association graph, in which the features of each node represent the status of each process link in the production line, and the weight of the edge reflects the similarity and mutual influence between nodes.

[0071] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0072] The process data association graph is input into the three-layer graph convolution network with residual connection in the process parameter association model for feature aggregation to obtain the initial feature map. Each layer of the graph convolution network contains 32 convolution kernels with a convolution kernel size of 3×3 and uses the ReLU activation function.

[0073] The initial feature map is input into the two-layer spatial attention network in the process parameter association model to model the relationship between nodes and obtain the spatial attention features. Each layer of the spatial attention network contains 8 attention heads, and the feature dimension of each attention head is 64, which is normalized using LayerNorm.

[0074] The target area is screened for the spatial attention features, and the regional mask matrix is ​​set according to the position information of the node in the process flow. The features of irrelevant areas are filtered through the mask matrix to obtain the regional attention features;

[0075] The regional attention features are input into the two-layer temporal attention network in the process parameter association model to model the changes of process parameters at different time scales and obtain the temporal related features. Each layer of the temporal attention network contains 4 attention heads and the feature dimension is 32.

[0076] Input the time series related features into the two-layer graph pooling layer in the process parameter association model to reduce the feature dimension and obtain the compressed features. Then input the compressed features into the three-layer fully connected network in the process parameter association model to perform feature fusion and obtain the fused feature vector.

[0077] Feature selection is performed on the fused feature vector to obtain a core feature vector, and feature mapping is performed on the core feature vector to obtain a process parameter feature vector.

[0078] Specifically, the process data association graph is input into the graph convolution network in the process parameter association model. The graph convolution network consists of three layers of graph convolution layers, each layer contains 32 convolution kernels, the size of the convolution kernel is 3×3, and the ReLU activation function is used. Graph convolution aggregates node features through the adjacency relationship between nodes to obtain a higher-order feature representation. Assume that the input node feature matrix is , the calculation process of each layer of graph convolutional network is expressed as:

[0079] ;

[0080] in, Indicates The node feature matrix of the layer, is the adjacency matrix of the graph, It is The weight matrix of the layer, ReLU is the activation function. The purpose of graph convolution is to propagate and aggregate the features of the nodes through the structure of the graph (i.e., the adjacency matrix), so that the features of each node can contain more information about neighboring nodes. 32 convolution kernels are used to enhance the expressiveness of the model. The size of the convolution kernel is 3×3, so each convolution operation will examine the 3×3 area in the neighborhood to update the features of the node. In this way, the initial feature map gradually obtains a rich feature representation under the action of each layer of graph convolution. The initial feature map is input into the spatial attention network in the process parameter association model to model the relationship between nodes. The spatial attention network gives higher attention to certain nodes by calculating the attention weights between different nodes, so that the model can better capture the mutual influence between nodes. At this stage, the spatial attention network consists of two layers, each layer contains 8 attention heads, and the feature dimension of each attention head is 64. The method for calculating the attention weight matrix between nodes is: for each pair of nodes and , calculate the correlation score between them , and then normalize the score through the Softmax function to get the attention weight:

[0081] ;

[0082] in, Representation Node and nodes The attention weights between is the calculated node relevance score, Is a node The normalized attention weight determines the contribution of each node in feature aggregation. In this way, the relationships between nodes can be modeled more accurately, allowing the network to focus on more important nodes and relationships. The spatial attention features are screened in the target area to filter out features of irrelevant areas. According to the position information of the node in the process flow, a region mask matrix is ​​set to select node features related to the target task. The region mask matrix is ​​a binary matrix in which each element indicates whether the corresponding region should be retained. If a node is not in the target area, the feature of the node will be masked. Assume that the region mask matrix is , then the process of screening through the mask matrix is ​​expressed as:

[0083] ;

[0084] in, represents element-wise multiplication, is the current node feature matrix, is the region mask matrix, is the feature matrix after the target area is filtered. In this way, the network can focus only on the task-related areas, thereby improving efficiency and reducing computational burden. The regional attention features are input into the two-layer temporal attention network in the process parameter association model. The temporal attention network captures key information in the time dimension by analyzing the changing rules of process parameters at different times. Temporal feature modeling helps the network understand the dynamic changes in the production process, especially dealing with parameter changes at different time scales. Assume that the temporal feature of the node is , then the calculation process of the temporal attention network is similar to that of the spatial attention network. By calculating the attention weights between nodes in the time series, the time series related features are obtained. Each layer of the temporal attention network contains 4 attention heads and the feature dimension is 32. The calculation of temporal attention is similar to that of spatial attention. The temporal attention score is calculated, and then the weight is obtained by Softmax normalization to aggregate the node features. After completing the temporal feature modeling, the temporal related features are input into the two-layer graph pooling layer in the process parameter association model. The graph pooling layer reduces the dimension of the node features, retains key information and reduces the amount of calculation. The pooling operation is implemented by selecting the maximum value or the average value. Assume that the feature after pooling is , then the calculation of the graph pooling layer is expressed as:

[0085] ;

[0086] Among them, Pool represents the pooling operation, and Pooling Method is a method such as maximum pooling or average pooling. is the input time series feature, is the pooled feature. The compressed feature is input into the three-layer fully connected layer in the process parameter association model for feature fusion to form a comprehensive feature vector. Assume that the pooled feature is , then the calculation through the fully connected layer is expressed as:

[0087] ;

[0088] in, and is the weight matrix and bias term of the fully connected layer, is the feature vector after the fully connected layer. Feature selection is performed on the fused feature vector to determine which features are most important for process parameter control. Feature selection is achieved through a variety of methods, such as importance-based scoring, L1 regularization, etc. Through feature selection, the core feature vector is obtained , these features are the key factors that are finally input into the process parameter control strategy. The core feature vector is mapped to obtain the final process parameter feature vector, which is expressed as:

[0089] ;

[0090] The final output of this step is It is the input of process parameter control strategy, which helps the decision-making system optimize process parameters and realize intelligent feedback and control.

[0091] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0092] The process parameter feature vector is input into the Weighted-K-Means clusterer in the multi-constrained non-convex optimizer to partition the parameter space. The similarity between samples is calculated by weighted Euclidean distance to obtain K process parameter cluster centers.

[0093] The K process parameter cluster centers are state-encoded, and the processing efficiency, resource utilization and product quality are encoded as state vectors to obtain a multi-dimensional state representation.

[0094] The multi-dimensional state representation is input into the Actor network of the DDPG algorithm in the multi-constrained non-convex optimizer, and the parameter control action is obtained after passing through three fully connected layers with feature dimensions of [256, 128, 64] and tanh activation function;

[0095] The parameter control action and multi-dimensional state representation are combined and input into the Critic network of the DDPG algorithm in the multi-constrained non-convex optimizer, and the action value evaluation is obtained through four fully connected layers with feature dimensions of [512, 256, 128, 64].

[0096] The temporal difference error is calculated based on the action value evaluation, and the parameters of the Actor network and the Critic network are updated through back propagation to obtain the optimized control strategy;

[0097] The reward calculation is performed on the optimized control strategy, and the weight of processing efficiency is set to w1, the weight of resource utilization is set to w2, and the weight of product quality is set to w3 to obtain the weighted reward value;

[0098] According to the weighted reward value, the gradient ascent optimization control parameter calculation is performed for each process parameter cluster to obtain the local optimal strategy, and the local optimal strategy is integrated and combined based on the cluster importance weight to obtain the process parameter control strategy.

[0099] Specifically, the process parameter feature vector is input into the weighted K-means clusterer in the multi-constrained non-convex optimizer to partition the parameter space. K-means clustering divides the data set into clusters, each of which contains similar data points. In weighted K-means, the similarity between samples is calculated by weighted Euclidean distance, and the weights are adjusted according to the importance of different parameters to obtain the center of each cluster. Suppose the feature vector of the sample is The center of the cluster is , the weight is , then the calculation formula of weighted Euclidean distance is:

[0100] ;

[0101] in, and are samples and cluster centers in The value of the dimension, is the weight of the corresponding dimension. By weighted Euclidean distance, the importance of certain dimensions is ensured to receive higher attention in the clustering process. The K-means algorithm iteratively updates the cluster center until the clustering converges. The obtained K process parameter cluster centers represent the division of the parameter space, which is convenient for subsequent optimization operations. The K process parameter cluster centers are state-encoded, and the key indicators of the production process - processing efficiency, resource utilization and product quality - are encoded into a multi-dimensional state vector to form a state representation. Assume that processing efficiency, resource utilization and product quality are , , , then the state vector s is expressed as:

[0102] ;

[0103] The state vector contains the core performance indicators of the process and is used as input in the multi-constrained non-convex optimizer to guide subsequent control strategy optimization. The multi-dimensional state representation is input into the Acton network of the deep deterministic policy gradient (DDPG) algorithm in the multi-constrained non-convex optimizer for processing. The DDPG algorithm is a reinforcement learning algorithm based on deep learning that maximizes cumulative rewards by continuously optimizing strategies. In the Actor network, the state vector s is input into a neural network with three fully connected layers, each with feature dimensions of 256, 128, and 64, and nonlinear transformation is performed using the tanh activation function. The output of the network is the parameter control action , indicating the process parameter control measures that should be taken under the current state. Assume that the output of the Actor network is , then after three layers of fully connected layers and the calculation of activation function, it is expressed as:

[0104] ;

[0105] in, , , are the weight matrices of each fully connected layer, , , is the bias term, s is the input state vector, is the output parameter control action. The parameter control action a and the state representation s are input into the Critic network in the DDPG algorithm. The Critic network evaluates the value of an action in a given state, that is, calculates the Q value of the action. The Q value reflects the cumulative reward that can be obtained by taking an action starting from the current state. The Critic network consists of four fully connected layers, and the feature dimensions of each layer are 512, 256, 128 and 64 respectively. Through the forward propagation of the network, the Critic network outputs the action value evaluation , that is, for a given state and action, the Q value evaluation output by the Critic network:

[0106] ;

[0107] in, It means to concatenate the state and action together as input. arrive is the weight matrix of the Critic network, arrive is the bias term, is the output action value evaluation. By calculating the Q value, backpropagating through the temporal difference (TD) error, the parameters of the Actor network and the Critic network are updated to continuously improve the performance of the strategy. The calculation formula of the TD error is:

[0108] ;

[0109] in, is the immediate reward of the current step, is the discount factor, is the Q value of the next state and the next action is the Q value of the current state and action. By minimizing the TD error, the network can gradually optimize the control strategy. Based on the optimized control strategy, the reward calculation is performed, and the weights of processing efficiency, resource utilization, and product quality are set as , and , and get the weighted reward value:

[0110] ;

[0111] Through the weighted reward value, the DDPG algorithm performs gradient ascent optimization control between multiple process parameter clusters and calculates the local optimal strategy for each process parameter cluster. The local optimal strategy is the optimal control solution obtained by optimizing the parameters in each cluster. In order to improve the global optimization effect, the local optimal strategies are integrated and combined based on the importance weights of the clusters to obtain the final process parameter control strategy. Assume that the importance weight of each cluster is The final process parameter control strategy It is expressed as:

[0112] ;

[0113] in, is the local optimal strategy for each process parameter cluster, is the weight of this type of cluster, and the final process parameter control strategy is the weighted average of each local optimal strategy.

[0114] In a specific embodiment, the execution step inputs the process parameter feature vector into the Weighted-K-Means clusterer in the multi-constrained non-convex optimizer to partition the parameter space, and the similarity between samples is calculated by weighted Euclidean distance to obtain K process parameter cluster centers, which may specifically include the following steps:

[0115] Score the importance of the process parameter feature vector, assign a weight coefficient based on the degree of influence of each parameter on the process, and obtain a parameter weight vector;

[0116] A weighted distance measurement function is constructed based on the parameter weight vector to obtain the distance calculation formula between samples;

[0117] Randomly select K samples from the process parameter feature vector as the initial clustering centers, and assign the remaining samples to the closest cluster to obtain the initial cluster division;

[0118] Calculate the weighted centroid of each process parameter cluster to obtain the updated cluster center, and recalculate the weighted distance between each sample and the cluster center based on the updated cluster center, redistribute the sample to the cluster with the closest distance, and obtain the optimized cluster division;

[0119] Calculate the intra-class sum of squares for the optimized cluster division to obtain the clustering evaluation index;

[0120] The clustering evaluation index is used to determine whether the cluster division has converged. If the change in the cluster center is less than the change threshold or reaches the maximum number of iterations, the iteration is stopped to obtain the final cluster division.

[0121] The weighted center point coordinates are calculated for each process parameter cluster in the final cluster division as the representative process parameter combination of the process parameter cluster, and K process parameter cluster centers are obtained.

[0122] Specifically, the importance of the process parameter feature vector is scored, and a weight coefficient is assigned according to the degree of influence of each parameter on the process. The scoring is based on the degree of influence of the process parameter on the production process. For example, a higher weight is assigned to a parameter that affects production efficiency, while a lower weight is assigned to a parameter that has a smaller impact on product quality or resource consumption. Assume that process parameters, and the importance score of each parameter is , then the weight vector w is expressed as:

[0123] ;

[0124] These weight coefficients reflect the relative importance of each parameter in the clustering process. Parameters with larger weights will have a greater impact on the clustering results. Based on the weight vector of the parameter, a weighted distance metric function is constructed to calculate the distance between samples. Assume that each process parameter feature vector is , then the weighted Euclidean distance calculation formula is:

[0125] ;

[0126] in, and The samples are and In the The values ​​of the process parameters, is the weight of the parameter. By weighted Euclidean distance, the influence of different dimensions in calculating distance is adjusted according to the importance of each parameter, so that the change of important parameters has a greater impact on clustering. After the weighted distance metric function is determined, the initial clustering is performed. By randomly selecting The samples are used as the initial cluster centers, and the remaining samples are assigned to the clusters with the closest distance. The initial cluster centers are selected by random selection or by some heuristic method. Assume that the selected initial cluster centers are , then each sample It will be assigned to the nearest cluster according to its weighted distance to the center of each cluster:

[0127] ;

[0128] in, It is a sample The cluster index to which it belongs. Through this process, the initial cluster division is obtained. Based on the initial cluster division, the center of each process parameter cluster is updated. In the weighted K-means algorithm, the update of the cluster center not only considers the mean of the sample, but also needs to weight the importance of each parameter. Assume that the cluster The sample set in is , then the weighted centroid of the cluster The calculation is done by the following formula:

[0129] ;

[0130] in, Belongs to the cluster A sample of is the weight of each process parameter. The calculation of the weighted centroid takes into account the weight coefficient of each parameter, making it more sensitive to the adjustment of important parameters. The updated cluster center better represents the characteristics of the samples in the cluster. After the cluster center is updated, the weighted distance from each sample to the center of each cluster is recalculated, and the samples are redistributed according to the distance. This process is iterated until the cluster division no longer changes significantly. In each iteration, the redistribution of samples is based on the weighted distance metric function:

[0131] ;

[0132] By continuously updating the cluster center and sample allocation, an optimized cluster division is obtained. In order to evaluate the effect of cluster division, the within-cluster sum of squares (WCSS) is calculated for each cluster, that is, the sum of the squares of the weighted distances between each sample in the cluster and the cluster center. As an indicator of clustering quality, the smaller the WCSS, the more reasonable the cluster division. The specific calculation formula is:

[0133] ;

[0134] The total WCSS of a cluster is the sum of the WCSS of all clusters:

[0135] ;

[0136] By calculating WCSS, the effect of cluster division is judged. After each iteration, the change of cluster center is checked. If the change is less than the set change threshold, , or the maximum number of iterations is reached, the iteration is stopped and the cluster division is considered to have converged. The judgment formula for cluster convergence is expressed as:

[0137] ;

[0138] in, and Represents clusters In the and The center in the iteration, is the set convergence threshold. After optimization and convergence judgment, the final process parameter clustering is obtained. For each cluster, the coordinates of its weighted center point are calculated as the representative process parameter combination of the cluster, and The representative process parameter combination of each cluster is calculated by the following weighted mean:

[0139] ;

[0140] These cluster centers represent different combinations of process parameters, which are typical parameter combinations under different conditions in the production process and are used for further optimization and decision-making.

[0141] In a specific embodiment, the execution step performs gradient ascending optimization control parameter calculation for each process parameter cluster according to the weighted reward value to obtain a local optimal strategy, and integrates the local optimal strategy, performs weighted combination based on the importance weight of the cluster, and obtains the process parameter control strategy. The process may specifically include the following steps:

[0142] Apply the gradient ascent method to calculate the parameter gradient of the weighted reward value to obtain the optimization direction, and iteratively update the process parameters based on the optimization direction to obtain the updated parameter sequence;

[0143] Input the updated parameter sequence into the target optimization function, calculate the parameter change between two adjacent iterations, and stop the iteration if it is less than the parameter threshold or reaches the maximum number of iterations to obtain the local optimal strategy;

[0144] According to the local optimal strategy, the equipment parameters of each process link are constrained and tested to obtain the parameter combination that meets the constraints. The constraint test includes the value range of the process parameters of feeding speed, crushing gap, screening amplitude and conveyor belt speed.

[0145] The cluster weight coefficient is calculated for the parameter combination that meets the constraints to obtain the cluster importance weight, and the cluster importance weight is normalized to obtain the standardized weight coefficient;

[0146] The standardized weight coefficients are linearly combined with the local optimal strategies of various clusters to obtain the global optimal parameter combination;

[0147] Based on the global optimal parameter combination, a control instruction sequence is generated to obtain a process parameter regulation strategy, which includes control parameter values ​​of feeder speed, crusher gap, vibrating screen amplitude and conveyor belt speed.

[0148] Specifically, in order to optimize the process parameters, the gradient ascent method is applied to calculate the parameter gradient of the weighted reward value to obtain the optimization direction. Assume that the reward function is a parameter vector

[0149] A function where each parameter Represents different process parameters (such as feeder speed, crusher gap, etc.). In order to optimize this function, its gradient with respect to the parameters is calculated, that is, the reward function About parameters The partial derivative of :

[0150] ;

[0151] Based on the gradient information, the parameters are updated, that is, the process parameters are iteratively updated along the gradient direction. The update rule is:

[0152] ;

[0153] in, is the learning rate, which controls the step size of each iteration. In this way, the process parameters are gradually optimized to increase the reward value and obtain a better process parameter combination. In the process of updating the parameters, the parameter change between two adjacent iterations is checked. If the change is less than the set parameter threshold Or when the maximum number of iterations is reached, the algorithm is considered to have converged and the iteration stops. Assume that the parameter change is:

[0154] ;

[0155] if , or the number of iterations exceeds the maximum limit , the iteration is terminated and the current parameter combination is returned as the local optimal strategy. Constraint checks are performed on the equipment parameters of each process link. Equipment parameters include feed rate, crushing gap, screening amplitude, and conveyor belt speed. Each parameter has a valid range of values, called the constraint interval. For example, feed rate The valid range is , breaking gap The range is . Make sure the optimized parameter combinations are within these ranges:

[0156] ;

[0157] By performing constraint checks on each parameter, we ensure that the obtained process parameter combination is not only optimal in theory but also feasible in actual production. We calculate the weight coefficient of each process parameter cluster to evaluate the importance of the cluster. Assume that process parameter clusters, each cluster The weight coefficient is , indicating the importance of the cluster in the overall optimization. The weight coefficient of the cluster is calculated based on the distribution density of samples in the cluster, the compactness of the cluster or the priority of the cluster. The calculation of is as follows:

[0158] ;

[0159] in, It is The number of samples in the cluster, It is The number of samples in the cluster. The cluster weight coefficient indicates the contribution of each cluster to the final process parameter control strategy. Clusters with larger weights will occupy a larger proportion in the global optimization. The cluster weight coefficient is normalized to obtain the standardized weight coefficient , which is used to ensure that the sum of the weight coefficients of all clusters is 1. The calculation formula for the standardized weight coefficient is:

[0160] ;

[0161] Standardization ensures that each cluster contributes fairly to the final control strategy. The local optimal strategies of various clusters are linearly combined to obtain the global optimal parameter combination. The local optimal strategy of the cluster is , then the global optimal parameter combination Calculated by the following linear combination formula:

[0162] ;

[0163] The weighted combination method ensures that the global optimal parameter combination can fully consider the importance of each cluster, thereby balancing the contribution of each cluster and obtaining a global optimal process parameter setting. , generate a control instruction sequence to guide the process parameter control in the actual production process. Assume that the process parameters include the feeder speed , Crusher gap , vibration screen amplitude and conveyor belt speed The process parameter control strategy contains the specific values ​​of these control parameters to ensure that the parameters of each device in the production process can reach the optimal state. For example, if the final global optimal parameter combination is , the control instruction sequence will instruct the corresponding speed or gap adjustment of the feeder, crusher, screening equipment and conveyor belt.

[0164] The above describes the intelligent process monitoring and feedback method for resource processing of construction waste in the embodiment of the present invention. The following describes the intelligent process monitoring and feedback system for resource processing of construction waste in the embodiment of the present invention. Figure 2 An embodiment of the intelligent process monitoring and feedback system for resource processing of construction waste in the embodiment of the present invention includes:

[0165] The acquisition module is used to collect process parameters at multiple monitoring points of the construction waste treatment production line to obtain an original monitoring data set, and to remove outliers and perform standardization on the original monitoring data set to obtain a pre-processed data set;

[0166] A construction module is used to construct a graph data structure based on a preprocessed data set and multiple monitoring points, calculate the correlation coefficient between nodes to determine the node connection relationship, and dynamically optimize through a capability-aware graph neural network to obtain a process data association graph;

[0167] An extraction module is used to input the process data association graph into the process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector;

[0168] The optimization module is used to input the process parameter feature vector into the multi-constrained non-convex optimizer, calculate the joint optimization of processing efficiency, resource utilization and product quality, and obtain the process parameter control strategy.

[0169] Through the collaborative cooperation of the above-mentioned components, the process parameters are dynamically optimized through the capability-aware graph neural network, and unreliable monitoring data are effectively identified and filtered, thereby improving data quality and reliability. A multi-level feature extraction structure is adopted, including graph convolution operations and multiple attention mechanisms, to deeply explore the correlation between process parameters and enhance the ability to characterize complex process states. Through multi-objective optimization, the coordinated optimization of processing efficiency, resource utilization and product quality is achieved, ensuring the scientific nature of the optimization results. A complete process parameter monitoring and feedback adjustment closed loop is constructed, and the real-time and accuracy of process parameter regulation are ensured through continuous data collection, analysis and optimization. A hierarchical and progressive processing architecture is adopted, from data preprocessing, graph structure construction, feature extraction to parameter optimization, to form a systematic process monitoring solution, which improves the reliability and robustness of the present invention. The strategy integration method based on cluster division and importance weighting effectively integrates the optimal control strategies under different working conditions.

[0170] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0171] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0172] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0173] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0174] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0175] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0176] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent process monitoring and feedback method for resource processing of construction waste, characterized in that: The method comprises: Collecting process parameters at multiple monitoring points of a construction waste treatment production line to obtain an original monitoring data set, and removing outliers and standardizing the original monitoring data set to obtain a preprocessed data set; A graph data structure is constructed based on the preprocessed data set and the multiple monitoring points, the correlation coefficient between nodes is calculated to determine the node connection relationship, and dynamic optimization is performed through the capability-aware graph neural network to obtain a process data association graph; specifically comprising: constructing a node feature matrix of the preprocessed data set according to the process position of the monitoring point, and extracting time series features from the node feature matrix to obtain an initial node representation; calculating the Pearson correlation coefficient between any two monitoring points in the initial node representation, setting a correlation coefficient threshold λ, establishing connecting edges for node pairs whose correlation coefficients are greater than the correlation coefficient threshold λ, and obtaining an initial graph structure; inputting the initial graph structure into the capability-aware graph neural network, the capability-aware graph neural network comprising three graph convolution layers, two attention layers and two fully connected layers; inputting the initial graph structure into the three-layer graph convolution layer, aggregating node features through an adjacency matrix, and using Re LU activation function is used to perform nonlinear transformation to obtain node embedding features; the node embedding features are input into the two attention layers, the attention weight matrix between nodes is calculated, the attention score is normalized by the softmax function, and the node features are weighted aggregated to obtain attention enhancement features; the attention enhancement features are input into two fully connected layers, and the reliability score of each node is calculated by the sigmoid function to obtain a node reliability vector; the importance of the nodes in the initial graph structure is re-evaluated according to the node reliability vector, and the nodes with reliability scores lower than the score threshold β are marked as unreliable nodes, and the connecting edges of the unreliable nodes are weighted attenuated to obtain the target edge weight; the adjacency matrix of the initial graph structure is updated based on the target edge weight, and the features of the unreliable nodes are replaced by the weighted average of their reliable neighbor nodes to obtain a process data association graph; Inputting the process data association graph into the process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector; The process parameter feature vector is input into a multi-constrained non-convex optimizer to calculate the joint optimization of processing efficiency, resource utilization and product quality to obtain a process parameter control strategy.

2. The intelligent process monitoring and feedback method for resource processing of construction waste according to claim 1 is characterized in that: The process parameters of multiple monitoring points of the construction waste treatment production line are collected to obtain an original monitoring data set, and the original monitoring data set is subjected to outlier elimination and standardization processing to obtain a preprocessed data set, including: Monitoring sensors are installed at the feeding end, crushing end, screening end, conveying end and finished product end of the production line to collect data on material particle size, moisture content, impurity content, equipment load, processing capacity and energy consumption indicators to obtain original collected data; The original collected data are grouped according to time series, and a unique identifier is generated for each group of data to obtain a labeled data set; Segmenting the labeled data set to obtain time series segmented data, and calculating the mean, variance, kurtosis and skewness of each parameter based on the time series segmented data to obtain a statistical feature set; Setting a threshold interval for the statistical feature set, marking data exceeding the threshold interval as an outlier, obtaining an outlier marking set, and removing outlier data points in the original collected data according to the outlier marking set, obtaining a cleaned data set; The parameters in the cleaned data set are normalized to their maximum and minimum values ​​to obtain normalized data, and the normalized data are organized and stored according to the corresponding relationship between the monitoring points and the process links to obtain a preprocessed data set.

3. The intelligent process monitoring and feedback method for resource processing of construction waste according to claim 1 is characterized in that: The step of inputting the process data association graph into the process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector includes: Input the process data association graph into a three-layer graph convolution network with residual connection in the process parameter association model for feature aggregation to obtain an initial feature map, where each layer of the graph convolution network contains 32 convolution kernels with a convolution kernel size of 3×3 and uses a ReLU activation function; Input the initial feature map into the two-layer spatial attention network in the process parameter association model to model the relationship between nodes, and obtain spatial attention features. Each layer of the spatial attention network contains 8 attention heads, and the feature dimension of each attention head is 64, which is normalized using LayerNorm; The spatial attention feature is subjected to target area screening, a regional mask matrix is ​​set according to the position information of the node in the process flow, and features of irrelevant regions are filtered through the mask matrix to obtain regional attention features; Input the regional attention features into the two-layer temporal attention network in the process parameter association model, model the changes of process parameters at different time scales, and obtain temporal related features. Each layer of the temporal attention network contains 4 attention heads and the feature dimension is 32; Input the time series related features into the two-layer graph pooling layer in the process parameter association model to perform feature dimension reduction to obtain compressed features, and input the compressed features into the three-layer fully connected network in the process parameter association model to perform feature fusion to obtain a fused feature vector; Feature selection is performed on the fused feature vector to obtain a core feature vector, and feature mapping is performed on the core feature vector to obtain a process parameter feature vector.

4. The intelligent process monitoring and feedback method for resource processing of construction waste according to claim 3 is characterized in that: The process parameter feature vector is input into a multi-constrained non-convex optimizer to calculate the joint optimization of processing efficiency, resource utilization and product quality to obtain a process parameter control strategy, including: The process parameter feature vector is input into the Weighted-K-Means clusterer in the multi-constrained non-convex optimizer to partition the parameter space, and the similarity between samples is calculated by weighted Euclidean distance to obtain K process parameter cluster centers; Performing state encoding on the K process parameter cluster centers, encoding processing efficiency, resource utilization and product quality into state vectors, and obtaining a multi-dimensional state representation; The multi-dimensional state representation is input into the Actor network of the DDPG algorithm in the multi-constrained non-convex optimizer, and a parameter control action is obtained through three fully connected layers with feature dimensions of [256, 128, 64] and a tanh activation function; The parameter control action and the multi-dimensional state representation are combined and input into the Critic network of the DDPG algorithm in the multi-constrained non-convex optimizer, and the action value evaluation is obtained through four fully connected layers with feature dimensions of [512, 256, 128, 64]; Based on the action value evaluation, the temporal difference error is calculated, and the parameters of the Actor network and the Critic network are updated through back propagation to obtain an optimized control strategy; Perform reward calculation on the optimized control strategy, set the processing efficiency weight to w1, the resource utilization weight to w2, and the product quality weight to w3, and obtain a weighted reward value; According to the weighted reward value, a gradient-ascending optimization control parameter calculation is performed for each process parameter cluster to obtain a local optimal strategy, and the local optimal strategy is integrated and combined based on the cluster importance weights to obtain a process parameter regulation strategy.

5. The intelligent process monitoring and feedback method for resource processing of construction waste according to claim 4 is characterized in that: The process parameter feature vector is input into the Weighted-K-Means clusterer in the multi-constrained non-convex optimizer to partition the parameter space, and the similarity between samples is calculated by weighted Euclidean distance to obtain K process parameter cluster centers, including: Score the importance of the process parameter feature vector, assign a weight coefficient according to the influence of each parameter on the process, and obtain a parameter weight vector; A weighted distance measurement function is constructed based on the parameter weight vector to obtain a distance calculation formula between samples; Randomly select K samples from the process parameter feature vector as initial cluster centers, and assign the remaining samples to the closest clusters to obtain initial cluster divisions; Calculate the weighted centroid for each process parameter cluster to obtain an updated cluster center, and recalculate the weighted distance between each sample and the cluster center based on the updated cluster center, and redistribute the sample to the cluster with the closest distance to obtain an optimized cluster division; Calculating the intra-class sum of squares for the optimized cluster division to obtain a clustering evaluation index; Judging whether the cluster division has converged according to the cluster evaluation index, if the change of the cluster center is less than the change threshold or reaches the maximum number of iterations, the iteration is stopped to obtain the final cluster division; The weighted center point coordinates are calculated for each process parameter cluster in the final cluster division as a representative process parameter combination of the process parameter cluster, and K process parameter cluster centers are obtained.

6. The intelligent process monitoring and feedback method for resource processing of construction waste according to claim 5 is characterized in that: The step of performing gradient ascending optimization control parameter calculation on each process parameter cluster according to the weighted reward value to obtain a local optimal strategy, integrating the local optimal strategy, and performing weighted combination based on the importance weight of the cluster to obtain a process parameter control strategy includes: Applying the gradient ascent method to perform parameter gradient calculation on the weighted reward value to obtain an optimization direction, and iteratively updating the process parameters based on the optimization direction to obtain an updated parameter sequence; The update parameter sequence is input into the target optimization function, and the parameter change between two adjacent iterations is calculated. If the change is less than the parameter threshold or the maximum number of iterations is reached, the iteration is stopped to obtain a local optimal strategy; According to the local optimal strategy, the equipment parameters of each process link are constrained and tested to obtain a parameter combination that meets the constraints, wherein the constraint test includes the value range of the process parameters of the feeding speed, the crushing gap, the screening amplitude and the conveyor belt speed; Calculating cluster weight coefficients for the parameter combinations that satisfy the constraints to obtain cluster importance weights, and normalizing the cluster importance weights to obtain standardized weight coefficients; Linearly combining the standardized weight coefficients with the local optimal strategies of various clusters to obtain a global optimal parameter combination; A control instruction sequence is generated based on the global optimal parameter combination to obtain a process parameter regulation strategy, wherein the process parameter regulation strategy includes control parameter values ​​of feeder speed, crusher gap, vibrating screen amplitude and conveyor belt speed.

7. An intelligent process monitoring and feedback system for resource processing of construction waste, characterized in that: The intelligent process monitoring and feedback method for resource-based treatment of construction waste according to any one of claims 1 to 6, wherein the intelligent process monitoring and feedback system for resource-based treatment of construction waste comprises: The acquisition module is used to collect process parameters at multiple monitoring points of the construction waste treatment production line to obtain an original monitoring data set, and to remove outliers and perform standardization on the original monitoring data set to obtain a pre-processed data set; A construction module, used to construct a graph data structure based on the preprocessed data set and the multiple monitoring points, calculate the correlation coefficient between nodes to determine the node connection relationship, and perform dynamic optimization through a capability-aware graph neural network to obtain a process data association graph; An extraction module, used for inputting the process data association graph into a process parameter association model to perform graph convolution operation and attention feature extraction to obtain a process parameter feature vector; The optimization module is used to input the process parameter feature vector into a multi-constrained non-convex optimizer, calculate the joint optimization of processing efficiency, resource utilization and product quality, and obtain a process parameter control strategy.

8. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the intelligent process monitoring and feedback method for resource processing of construction waste as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the intelligent process monitoring and feedback method for resource recovery of construction waste as described in any one of claims 1 to 6.

Images