Powder feeding process ai optimization control method and system

Abstract

The present application relates to the technical field of powder feeding, and particularly relates to a powder feeding process AI optimization control method and system, which predicts powder feeding process evolution trend, constructs a multi-target constraint directed dependence graph, and obtains a constraint influence propagation matrix by using a graph neural network, identifies a hierarchical conflict structure based on the matrix, generates adaptive weights by using a graph attention mechanism, degrades according to priority and automatically recovers in conflict, and solves an optimal control action sequence by using reinforcement learning, thereby improving the constraint satisfaction degree and process stability of the powder feeding process.

Description

Technical Field

[0001] This invention relates to the field of powder feeding technology, and in particular to an AI-optimized control method and system for the powder feeding process. Background Technology

[0002] Powder feeding is a common process in industrial production, widely used in chemical, pharmaceutical, and food industries. The core challenge lies in precisely controlling the flow rate, uniformity, and energy consumption of powder materials to meet the requirements of downstream processes for material proportioning and process stability. Current conventional methods rely primarily on manual experience combined with proportional-integral-derivative (PID) controllers for adjustment, with operators manually correcting silo valve openings or screw feeder speeds based on real-time instrument data. Some highly automated production lines employ model-based predictive control methods, establishing linear or weakly nonlinear models of powder flow to predict future feeding conditions and optimize control parameters.

[0003] Existing control methods struggle to effectively coordinate the inherent contradictions between multiple objective constraints during the feeding process. For example, powder uniformity and feeding rate often mutually restrict each other, while the goal of minimizing energy consumption may conflict with these constraints. Current fixed-weight multi-objective optimization or piecewise rule-based processing methods are highly susceptible to constraint violations or control oscillations under dynamic operating conditions, ultimately impacting product quality and production efficiency. Reliance on manual experience and offline debugging mechanisms prevent the system from sensing the time-varying characteristics of powder flowability, such as bridging in the hopper due to moisture absorption and agglomeration or fluidization anomalies caused by electrostatic aggregation. These nonlinear and non-stationary disturbances often lead to a sharp decline in control performance, even triggering safety shutdowns. Therefore, existing technologies lack the ability to predict dynamic constraint conflicts and adaptive adjustment mechanisms, making it difficult to maintain stable and efficient feeding control under complex disturbances. Summary of the Invention

[0004] This invention provides an AI-optimized control method and system for the powder feeding process, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides an AI-optimized control method for a powder feeding process, comprising: Temporal features were extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process; Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network for message passing and aggregation to obtain the constraint influence propagation matrix. Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained; Based on the hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node, and an adaptive weight configuration is dynamically generated by combining real-time operating condition deviations. When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to the process safety priority, and automatically restored when the recovery conditions are met; Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

[0006] Temporal features are extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process, including: Collect multi-dimensional status data of the powder feeding system within the current time window. The multi-dimensional status data includes feeding flow rate, silo pressure, stirring speed and material temperature. Arrange the multi-dimensional status data in chronological order to form a time-series status sequence. Multi-scale temporal feature extraction is performed on the temporal state sequence. State change rate features are extracted on a short time scale, and state trend features are extracted on a long time scale. The state change rate features and the state trend features are fused to obtain a temporal feature vector. The time-series feature vector is input into the time-series prediction neural network. The time-series prediction neural network is trained by the mapping relationship between the historical time-series feature vector and the corresponding actual evolution results, and outputs a multi-dimensional state prediction value sequence within the future time window. The multi-dimensional state prediction value sequence constitutes the evolution trend of the powder feeding process.

[0007] Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. A graph neural network is used to perform message passing and aggregation on the directed constraint dependency graph to obtain a constraint influence propagation matrix, including: Based on the multidimensional state prediction value sequence in the evolution trend of the powder feeding process, uniformity constraints, rate constraints and energy consumption constraints are calculated and used as constraint nodes to construct a multi-objective constraint set; Extract the current working condition feature vector from the multidimensional state prediction value sequence. The current working condition feature vector includes the satisfaction margin and state change rate of each constraint node. Input the current working condition feature vector into the directed edge weight prediction network and output the time-varying coupling strength value between each pair of constraint nodes. When the time-varying coupling strength value exceeds the preset strength threshold, establish a directed edge between the pair of constraint nodes and use the time-varying coupling strength value as the weight to form a time-varying directed constraint dependency graph. A time-varying directed constraint dependency graph is input into a multi-layer graph neural network. In each graph convolutional layer, the constraint nodes perform weighted aggregation of the features of adjacent nodes through directed edge weights and update their own representations to obtain the node representation matrix of each layer. The representation difference between constraint node pairs in the first layer node representation matrix is ​​calculated to obtain the first-order constraint influence matrix. The representation difference between constraint node pairs in the second layer node representation matrix is ​​calculated to obtain the second-order constraint influence matrix. The first-order constraint influence matrix and the second-order constraint influence matrix are combined hierarchically to form a hierarchical constraint influence propagation matrix.

[0008] Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained, including: The constraint influence propagation matrix is ​​subjected to a second power operation to obtain a second-order constraint influence matrix. Elements with influence intensity exceeding a preset influence threshold are extracted from the constraint influence propagation matrix and the second-order constraint influence matrix to form a first-order conflict propagation path and a second-order conflict propagation path, respectively. The two paths are then merged to obtain a set of conflict propagation paths. The propagation attenuation coefficient is calculated based on the path length of each conflict propagation path, and the propagation attenuation coefficient is multiplied by the influence intensity of each node on the path to obtain the cascade effect intensity of each conflict propagation path. The path convergence degree is obtained by counting the number of times each constraint node is traversed by the conflict propagation path. The node influence output strength is obtained by calculating the sum of the cascade effect strength of each constraint node. The node conflict score is obtained by weighted summing of the path convergence degree and the node influence output strength. Constraint nodes whose node conflict scores exceed a preset score threshold are identified as conflict nodes. Calculate the shortest propagation distance from each conflict node to other constraint nodes, and divide the constraint nodes into source conflict layer, directly affected layer and indirectly affected layer according to the shortest propagation distance to form a hierarchical conflict structure.

[0009] Based on the aforementioned hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node. An adaptive weight configuration is then dynamically generated by combining real-time operating condition deviations, including: Each constraint node in the hierarchical conflict structure is treated as a graph node, and the conflict propagation path between constraint nodes is treated as a graph edge to construct a constraint conflict propagation graph. The initial attention weight of the source conflict layer node is higher than that of the directly affected layer node, and the initial attention weight of the directly affected layer node is higher than that of the indirectly affected layer node. Based on the cascading effect strength and path convergence of each graph node in the constraint conflict propagation graph, the attention coefficient of each graph node relative to its neighboring graph nodes is calculated through the graph attention mechanism. The attention coefficient is then weighted and aggregated with the features of the neighboring graph nodes to obtain the structural importance score. Obtain the real-time operating parameters of each constraint node, calculate the target deviation of the real-time operating parameters relative to the target set value of the constraint node and the relative deviation of the real-time operating parameters relative to the adjacent constraint nodes in the conflict propagation path, weight and fuse the two deviations to obtain the comprehensive operating deviation value, and compare the comprehensive operating deviation value with the preset deviation sensitivity threshold to obtain the operating condition sensitivity score. The structural importance score and the operating condition sensitivity score are weighted and fused to obtain a comprehensive importance score. Based on the comprehensive importance score, resource weights are allocated and optimized to obtain an adaptive weight configuration.

[0010] When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to process safety priority, and automatically restored when the recovery conditions are met, including: The adaptive weight configuration is applied to each constraint node, the constraint satisfaction status is detected, and constraint nodes with a constraint satisfaction status of non-satisfaction are identified as residual conflict nodes. The number and distribution of residual conflict nodes are counted to obtain residual conflict features. When the residual conflict characteristics meet the preset conflict triggering conditions, the level of the residual conflict node and the conflict propagation path are extracted from the hierarchical conflict structure. The process safety priority score is calculated by combining the process safety impact range and fault propagation risk of the residual conflict node. The downgrade candidate sequence is obtained by sorting the process safety priority scores from low to high. From the candidate degradation sequence, select residual conflict nodes in sequence to reduce the constraint strength to the preset degradation level, recalculate the constraint satisfaction state after degradation, stop degradation when all constraint nodes satisfy the constraints, and record the set of degradation nodes and the degradation time. The system monitors the real-time operating parameters and constraint satisfaction status of the degraded nodes. When the real-time operating parameters of the degraded node continuously meet the original constraint strength and the duration exceeds the preset stable duration, the constraint strength of the degraded node is restored from the degraded level to the original constraint strength and the set of degraded nodes is updated.

[0011] Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree under the adaptive weight configuration and the process deviation, including: Based on the set of degraded nodes and the non-degraded constraint nodes, a constraint domain with priority management is constructed. The constraint boundary corresponding to the degraded node is set according to the degrade level, and the constraint boundary corresponding to the non-degraded constraint node is set according to the original constraint strength. The constraint domain is used as the action space constraint of the reinforcement learning strategy. The process state parameters at the current moment are obtained as the state input of the reinforcement learning strategy. Based on the adaptive weight configuration, the degree of constraint satisfaction of each constraint node in the current state is calculated. The process target parameters at the current moment are obtained and the deviation from the actual process state parameters is calculated to obtain the process deviation. The degree of constraint satisfaction and the process deviation are weighted and fused to construct the reward function. Based on the reward function and the constraint domain, a reinforcement learning strategy is used to iteratively search for control actions within the action space constraints, calculate the cumulative reward value corresponding to each candidate control action, and select the control action sequence with the largest cumulative reward value as the optimal control action sequence. The optimal control action sequence is converted into execution instructions for the powder feeding equipment. The process feedback parameters after execution are obtained to update the process state parameters. The updated process state parameters are used as the state input of the reinforcement learning strategy at the next moment to form a closed-loop optimization.

[0012] A second aspect of the present invention provides an AI optimization control system for a powder feeding process, comprising: The predictive analysis unit is used to extract and predict the time-series features of the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process. The constraint graph modeling unit is used to construct a set of multi-objective constraints, including uniformity constraints, rate constraints, and energy consumption constraints, based on the evolution trend and model it as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network to perform message passing and aggregation to obtain a constraint influence propagation matrix. The conflict propagation unit is used to calculate the conflict propagation path and cascade effect intensity of each constraint node based on the constraint influence propagation matrix, identify conflict nodes, and obtain a hierarchical conflict structure. The weight adaptive unit is used to calculate the importance score of each constraint node according to the hierarchical conflict structure using a graph attention mechanism, and dynamically generate an adaptive weight configuration in combination with real-time operating condition deviations. The degradation recovery unit is used to temporarily downgrade the conflicting constraints according to the process safety priority when conflicts still exist under the adaptive weight configuration, and to automatically recover them when the recovery conditions are met. The reinforcement learning unit is used to solve the optimal control action sequence within the constraint domain after priority management using a reinforcement learning strategy, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this invention are: By extracting and predicting temporal features, the evolution trend of the powder feeding process can be grasped in advance, laying the foundation for subsequent constraint management. Directed dependency graph modeling of multi-objective constraint sets and graph neural network message passing accurately quantify the mutual influence between constraints, obtain the constraint influence propagation matrix, and thus clearly reveal the coupling relationship between key constraint nodes. This significantly improves the control accuracy and global coordination capability of complex feeding processes, and avoids the local imbalance problem caused by traditional single constraint optimization.

[0016] Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity are calculated, accurately identifying conflict nodes and constructing a hierarchical conflict structure, effectively uncovering the root causes of conflict among multi-objective constraints. A graph attention mechanism dynamically calculates the importance score of each constraint node and generates an adaptive weight configuration based on real-time operating condition deviations. This allows the weights to automatically adjust with fluctuations in operating conditions, significantly enhancing the system's adaptability and robustness to changes in operating conditions and avoiding performance degradation caused by fixed weights when operating conditions shift.

[0017] When conflicts still exist under adaptive weights, conflicting constraints are temporarily downgraded based on process safety priorities and automatically restored when recovery conditions are met, ensuring both core safety and dynamic balance of constraints. Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence. The reward function integrates the constraint satisfaction degree under adaptive weights with process deviations, achieving energy-saving, efficient, and uniform feeding control effects under the global optimization objective. This significantly reduces manual intervention and improves the intelligence level of the feeding process, achieving synergistic optimization of energy consumption, rate, and uniformity. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the AI-optimized control method for powder feeding; Figure 2 Build a flowchart for the hierarchical conflict structure. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the AI ​​optimization control method for powder feeding process according to an embodiment of the present invention.

[0022] AI-optimized control methods for powder feeding processes include: Temporal features were extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process; Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network for message passing and aggregation to obtain the constraint influence propagation matrix. Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained; Based on the hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node, and an adaptive weight configuration is dynamically generated by combining real-time operating condition deviations. When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to the process safety priority, and automatically restored when the recovery conditions are met; Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

[0023] Temporal features are extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process, including: Collect multi-dimensional status data of the powder feeding system within the current time window. The multi-dimensional status data includes feeding flow rate, silo pressure, stirring speed and material temperature. Arrange the multi-dimensional status data in chronological order to form a time-series status sequence. Multi-scale temporal feature extraction is performed on the temporal state sequence. State change rate features are extracted on a short time scale, and state trend features are extracted on a long time scale. The state change rate features and the state trend features are fused to obtain a temporal feature vector. The time-series feature vector is input into the time-series prediction neural network. The time-series prediction neural network is trained by the mapping relationship between the historical time-series feature vector and the corresponding actual evolution results, and outputs a multi-dimensional state prediction value sequence within the future time window. The multi-dimensional state prediction value sequence constitutes the evolution trend of the powder feeding process.

[0024] Accurate real-time data acquisition is fundamental for subsequent analysis and prediction during powder feeding. For the powder feeding system, four core state variables—feed flow rate, silo pressure, stirring speed, and material temperature—are simultaneously collected within the current time window, forming a multi-dimensional state data set. Feed flow rate reflects the change in mass or volume of powder passing through the feed inlet per unit time; silo pressure characterizes the influence of the internal air pressure environment on powder flowability; stirring speed describes the mechanical motion state of the mixing device; and material temperature is related to the physical properties of the powder (such as flowability and agglomeration tendency). These four types of data are acquired synchronously within the same sampling period. The sampling frequency is set according to the dynamic response characteristics of the feeding process, typically selected within the range of 100ms to 500ms to ensure effective capture of rapid changes in operating conditions. The collected multi-dimensional state data are arranged sequentially in chronological order to form a time-series state sequence. ,in The number of sampling points within the time window. The number of state dimensions (here) (corresponding to four types of state variables).

[0025] For multi-scale temporal feature extraction of time-series state sequences, independent feature analysis is performed at both short and long time scales. At the short time scale, the focus is on the rate of change of state between adjacent sampling points, i.e., the first-order difference of each dimension of state variables and its statistical characteristics, which can capture local dynamic information such as sudden changes in feed flow and instantaneous fluctuations in silo pressure. Specifically, the difference operation is performed on the state vectors of adjacent time points in the sequence, and the mean, variance, and other statistics are calculated within a sliding window to obtain the state change rate feature vector. On a long-term scale, the focus is on the overall trend of state variables over a longer period. Moving averages or low-frequency filtering methods are used to smooth the time-series state sequences, extracting the trend components of state variables in each dimension. Trend features are then quantified using indicators such as linear regression slope and trend strength to obtain a state trend feature vector. Short-timescale features focus on capturing instantaneous disturbances in the operating conditions, while long-timescale features focus on revealing the macroscopic evolution direction of the process. The two are complementary in describing the dynamic behavior of powder feeding.

[0026] The feature fusion stage will use the state change rate feature vector With state trend feature vector The features are concatenated to form a fused temporal feature vector. ,in This represents a vector concatenation operation. The fused temporal feature vector contains both short-term dynamic information and long-term trend information, enabling a more comprehensive characterization of the powder feeding system's operational status within the current time window. In practical engineering, different fusion weights can be assigned to short-term and long-term features according to different process stages to adapt to the varying needs for feature emphasis at different stages such as feed start-up, stable operation, and shutdown.

[0027] Temporal prediction neural networks use fused temporal feature vectors The neural network takes as input and outputs a sequence of multidimensional state predictions within a future time window. It employs an encoder-decoder structure. The encoder performs high-dimensional semantic compression on the input temporal feature vector, extracting implicit representations effective for future state predictions. The decoder, based on the implicit representations output by the encoder, progressively generates multidimensional state predictions for each future sampling time. The encoder can utilize a multilayer long short-term memory (LSTM) network or a temporal convolutional network (TCN). The former excels at modeling long-range dependencies in time series, while the latter offers advantages in parallel computation efficiency. The decoder also employs a recursive structure, feeding back the prediction output from the previous step to the input of the current step in each prediction step, achieving multi-step rolling prediction.

[0028] The training process of a time-series prediction neural network relies on the mapping relationship between historical time-series feature vectors and corresponding actual evolution results. Specifically, a large number of time window samples are extracted from historical running data, and each sample is input with a time-series feature vector. and the corresponding future actual multidimensional state sequence Composition, including superscript This is the sample index. The training objective is to minimize the mean squared error between the predicted output and the actual state, and the loss function is defined as... ,in The total number of training samples, For the network to the first The predicted output for each sample, This corresponds to the actual multidimensional state sequence. The network parameters are iteratively updated using the backpropagation algorithm until the loss function converges to a set threshold. To prevent overfitting, a Dropout regularization mechanism is introduced during training, and historical data is divided into training and validation sets in chronological order to ensure the model's generalization ability to unseen conditions.

[0029] After training, the time-series predictive neural network is deployed online in the powder feeding control system. It receives the latest time-series feature vector every sampling period and outputs a real-time sequence of predicted values ​​for feed flow rate, silo pressure, stirring speed, and material temperature within the future time window. This predicted value sequence constitutes the evolution trend of the powder feeding process, providing forward-looking state information support for subsequent constraint modeling and optimized control. When the prediction error exceeds a set confidence threshold, an online model update mechanism is triggered, using the latest accumulated actual operating data to incrementally fine-tune the network parameters, ensuring that the prediction accuracy remains within an effective range as process conditions change. This online adaptive update mechanism allows the evolution trend prediction to follow the long-term drift and seasonal changes in the feeding process, ensuring that control optimization is always based on accurate process state predictions.

[0030] Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. A graph neural network is used to perform message passing and aggregation on the directed constraint dependency graph to obtain a constraint influence propagation matrix, including: Based on the multidimensional state prediction value sequence in the evolution trend of the powder feeding process, uniformity constraints, rate constraints and energy consumption constraints are calculated and used as constraint nodes to construct a multi-objective constraint set; Extract the current working condition feature vector from the multidimensional state prediction value sequence. The current working condition feature vector includes the satisfaction margin and state change rate of each constraint node. Input the current working condition feature vector into the directed edge weight prediction network and output the time-varying coupling strength value between each pair of constraint nodes. When the time-varying coupling strength value exceeds the preset strength threshold, establish a directed edge between the pair of constraint nodes and use the time-varying coupling strength value as the weight to form a time-varying directed constraint dependency graph. A time-varying directed constraint dependency graph is input into a multi-layer graph neural network. In each graph convolutional layer, the constraint nodes perform weighted aggregation of the features of adjacent nodes through directed edge weights and update their own representations to obtain the node representation matrix of each layer. The representation difference between constraint node pairs in the first layer node representation matrix is ​​calculated to obtain the first-order constraint influence matrix. The representation difference between constraint node pairs in the second layer node representation matrix is ​​calculated to obtain the second-order constraint influence matrix. The first-order constraint influence matrix and the second-order constraint influence matrix are combined hierarchically to form a hierarchical constraint influence propagation matrix.

[0031] Based on the multidimensional state prediction sequence output from the powder feeding process evolution trend, three types of constraints are calculated and used as constraint nodes to construct a multi-objective constraint set. The uniformity constraint is obtained by evaluating the spatial variance of the predicted powder concentration distribution or mixing ratio. When the variance of the predicted concentration distribution exceeds the process allowable upper limit, the uniformity constraint is violated. The rate constraint is calculated by the slope of the change in feeding flow rate or screw speed in the predicted state sequence. Excessive rate changes can cause process anomalies such as material blockage or splashing. The energy consumption constraint is estimated by the power integral of the drive motor during the prediction period. When the predicted energy consumption exceeds the set single-batch energy consumption upper limit, the energy consumption constraint is triggered. These three types of constraint nodes together constitute the multi-objective constraint set, providing the node foundation for the subsequent construction of the directed constraint dependency graph.

[0032] The current working condition feature vector is extracted from the multidimensional state prediction value sequence. This feature vector contains two key pieces of information: the satisfaction margin of each constraint node and the rate of change of the corresponding state variable of each constraint node. The satisfaction margin is defined as the normalized distance between the current constraint index value and its upper and lower thresholds; the smaller the margin, the closer the constraint is to the boundary violation. The rate of change of state is calculated by the time difference of the corresponding state variable in the prediction sequence, reflecting the dynamic evolution direction of the constraint pressure. The above working condition feature vector is input into a directed edge weight prediction network. This network is based on a multilayer perceptron and takes the concatenated vector of working condition features of any two constraint nodes as input, outputting a scalar value as the time-varying coupling strength value between the two constraint nodes. ,in and These are the indices of the constraint nodes. When Exceeding the preset intensity threshold At that time, from the node To the node Establish a directed edge, and The weight of this edge constitutes a time-varying directed constrained dependency graph. The orientation of this graph reflects the direction of causal influence between constraints. For example, violations of rate constraints often precede the deterioration of uniformity constraints. Therefore, the coupling strength from rate constraint nodes to uniformity constraint nodes is usually higher.

[0033] Transform the time-varying directed constraint dependency graph Input is processed through a multi-layer graph neural network for message passing and aggregation. In the first graph convolutional layer, each constraint node... Receive all its incoming neighbor nodes Features, and edge weights The neighbor features are weighted and summed, then concatenated with the node's own features, and processed by a linear transformation and activation function to obtain the updated node representation vector. After completing the first-level aggregation for all constraint nodes, the node representation vectors are concatenated in node order to obtain the first-level node representation matrix. Its number of rows is equal to the total number of constraint nodes. The number of columns is equal to the dimension of the first hidden layer. In the second convolutional layer, with... Using the representation vectors of each node as input, the above directed weighted aggregation and nonlinear transformation process is repeated to obtain the second-layer node representation matrix. The stacking of two layers of graph convolutions allows the representation of each node to include not only the influence of its direct neighbors, but also the indirect influence information of constrained nodes within a two-hop range, thereby capturing deeper constraint coupling relationships.

[0034] Based on the first-level node representation matrix Calculate all constraint node pairs The difference in representation between nodes With nodes The first layer represents the difference between vectors. Norm as the influence strength of first-order constraint The first-order constraint influence matrix is ​​composed of the first-order influence strength values ​​of all node pairs. Similarly, based on the second-level node representation matrix... The representation differences between each pair of constraint nodes are calculated to obtain the second-order constraint influence matrix. The first-order constraint influence matrix reflects the degree of direct mutual influence between constraint nodes after single-hop message passing, while the second-order constraint influence matrix further characterizes the degree of indirect influence transmitted through intermediate nodes. and The hierarchical combination method involves stacking two matrices along the newly added dimension to form a hierarchical constraint influence propagation matrix. The first layer of slices corresponds to first-order influence, and the second layer corresponds to second-order influence. This hierarchical structure preserves the hierarchical differences in constraint influence at different propagation depths, providing a fine-grained description of constraint propagation for subsequent conflict node identification and hierarchical conflict structure analysis.

[0035] In practical powder feeding scenarios, complex nonlinear coupling relationships often exist between uniformity constraints, rate constraints, and energy consumption constraints. For example, increasing the feeding rate can shorten batch time and thus reduce energy consumption per unit time, but an excessively high rate can lead to uneven powder mixing, causing uniformity constraints to be violated; while decreasing the rate is beneficial for uniformity, it may prolong the feeding time and increase total energy consumption. The directed edge weight prediction network learns the statistical patterns of these coupling relationships from historical operating data, enabling it to dynamically adjust the directed edge weights between constraint nodes under different operating conditions. This allows the time-varying directed constraint dependency graph to accurately reflect the actual dependence strength and direction between constraints under the current operating condition, rather than using a fixed static graph structure. Building upon this, a multi-layer graph neural network uses a message passing mechanism to propagate local constraint coupling information layer by layer, allowing the layered constraint influence propagation matrix to comprehensively characterize the influence structure of the constraint system at different propagation depths, providing reliable structured input for subsequent conflict identification and priority management.

[0036] Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained, including: The constraint influence propagation matrix is ​​subjected to a second power operation to obtain a second-order constraint influence matrix. Elements with influence intensity exceeding a preset influence threshold are extracted from the constraint influence propagation matrix and the second-order constraint influence matrix to form a first-order conflict propagation path and a second-order conflict propagation path, respectively. The two paths are then merged to obtain a set of conflict propagation paths. The propagation attenuation coefficient is calculated based on the path length of each conflict propagation path, and the propagation attenuation coefficient is multiplied by the influence intensity of each node on the path to obtain the cascade effect intensity of each conflict propagation path. The path convergence degree is obtained by counting the number of times each constraint node is traversed by the conflict propagation path. The node influence output strength is obtained by calculating the sum of the cascade effect strength of each constraint node. The node conflict score is obtained by weighted summing of the path convergence degree and the node influence output strength. Constraint nodes whose node conflict scores exceed a preset score threshold are identified as conflict nodes. Calculate the shortest propagation distance from each conflict node to other constraint nodes, and divide the constraint nodes into source conflict layer, directly affected layer and indirectly affected layer according to the shortest propagation distance to form a hierarchical conflict structure.

[0037] Combination Figure 2 The flowchart for constructing a hierarchical conflict structure is explained below, and the influence propagation matrix of hierarchical constraints is obtained. Next, it is necessary to further explore the conflict propagation structure between the constraint nodes. Performing exponentiation, i.e., calculating This yields the second-order constraint influence matrix. The physical meaning of the quadratic operation is: if the constraint nodes... For nodes There is a direct impact on the nodes. For nodes If there is a direct impact, then The corresponding element in reflects via intermediate nodes right The cumulative strength of the indirect impact. This operation extends the matrix, which originally could only describe one-hop propagation relationships, to the range of two-hop propagation, allowing the conflict risk implicit in the indirect coupling path to be quantitatively expressed.

[0038] from Extract all elements whose values ​​exceed a preset impact threshold. element pairs These elements form a set of first-order conflict propagation paths for their corresponding directed paths. Similarly, from Extracting values ​​exceeding the same threshold The element pairs constitute the set of second-order conflict propagation paths. Taking the union of the two sets yields the complete set of conflict propagation paths. Threshold The setting of the constraint coupling strength needs to be calibrated based on the sensitivity of the actual process to the constraint coupling strength. Too low a value will introduce a large number of weakly correlated paths that interfere with subsequent analysis, while too high a value may miss potential conflict propagation channels. In powder feeding scenarios, there is often a strong direct coupling between uniformity constraints and rate constraints, while the coupling between uniformity constraints and energy consumption constraints is more reflected in second-order indirect paths. Therefore, considering both first-order and second-order paths is crucial for a complete characterization of the conflict structure.

[0039] For the set of conflict propagation paths Each path in the array, based on its path length Calculate the propagation attenuation coefficient (i.e., the number of directed edges traversed by the path). The propagation attenuation coefficient reflects the physical law that the impact of a collision gradually weakens with increasing propagation hops, and is defined as follows: ,in This is a single-hop attenuation factor, which can be adjusted according to specific process characteristics. For a first-order path, The attenuation coefficient is For second-order paths, The attenuation coefficient is Based on this, the propagation attenuation coefficient is multiplied by the influence intensity of each node along the path to obtain the cascade effect intensity of the path. Specifically, for the sequence of nodes... The path, ,in For the propagation matrix The influence strength value of the corresponding node pair. This multiplication mechanism ensures that the strength of the cascade effect is constrained by both the path length and the coupling strength of each segment on the path, avoiding the one-sidedness caused by relying on only a single index.

[0040] After obtaining the cascading effect strength of each path, two types of indicators are statistically analyzed for each constraint node: path convergence. Defined as a set of conflict propagation paths Passing through nodes The number of paths reflects the structural importance of the node as a "transfer hub" in the conflict propagation network; the node's influence on the output strength. Defined as all nodes The sum of the cascading effects along the conflict propagation path originating from a given node reflects that node's ability to propagate conflict outwards. A weighted sum of these two values ​​yields the node's conflict score. ,in and These are the weighting coefficients for path convergence and node influence output intensity, respectively, with a sum of 1. They can be configured according to the relative importance of structural and output conflicts in the process scenario. When the powder feeding process is sensitive to the diffusion of local disturbances, the weighting coefficient can be appropriately increased. The proportion should be increased; when more attention is paid to the direct destructive force of a single constraint node on the overall process objective, then the proportion should be increased. The proportion of.

[0041] Node conflict scoring Exceeding the preset scoring threshold Constraint nodes are identified as conflict nodes, forming a set of conflict nodes. Threshold Determining the appropriate criteria also requires statistical analysis based on historical operating data. An adaptive setting using the mean plus a certain number of standard deviations can be employed to ensure robust identification performance under different batches and material characteristics. After identifying conflict nodes, each conflict node is calculated. Shortest propagation distance to all other constraint nodes ( For any constrained node, the shortest propagation distance is based on the set of conflict propagation paths. The resulting directed graph can be solved using breadth-first search or Dijkstra's algorithm.

[0042] Based on the shortest propagation distance, all constrained nodes are divided into three levels. Nodes with a shortest propagation distance of 0, i.e., the conflicting nodes themselves, are classified into the source conflict layer. These nodes are the origin of the conflict, and their abnormal constraint states directly drive the generation and spread of the conflict. Nodes with a shortest propagation distance of 1 are classified into the directly affected layer. These nodes have a direct first-order propagation relationship with the source conflict layer nodes and are the first to feel the transmission pressure of constraint deviation after a conflict occurs. Nodes with a shortest propagation distance greater than 1 are classified into the indirectly affected layer. The impact of the conflict on them requires multiple hops to reach them, and the intensity of the impact is relatively weak due to attenuation, but it is still not negligible in persistent conflict scenarios. The three levels together constitute a hierarchical conflict structure, providing a clear structured input for subsequent graph attention mechanisms to calculate the importance scores of constrained nodes and manage priorities. The introduction of the hierarchical conflict structure enables the control strategy to treat constrained nodes at different levels differently, prioritizing the problems of the source conflict layer while implementing preventive intervention for the directly affected layer, thereby achieving more precise and efficient multi-constraint coordinated control in the powder feeding process.

[0043] Based on the aforementioned hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node. An adaptive weight configuration is then dynamically generated by combining real-time operating condition deviations, including: Each constraint node in the hierarchical conflict structure is treated as a graph node, and the conflict propagation path between constraint nodes is treated as a graph edge to construct a constraint conflict propagation graph. The initial attention weight of the source conflict layer node is higher than that of the directly affected layer node, and the initial attention weight of the directly affected layer node is higher than that of the indirectly affected layer node. Based on the cascading effect strength and path convergence of each graph node in the constraint conflict propagation graph, the attention coefficient of each graph node relative to its neighboring graph nodes is calculated through the graph attention mechanism. The attention coefficient is then weighted and aggregated with the features of the neighboring graph nodes to obtain the structural importance score. Obtain the real-time operating parameters of each constraint node, calculate the target deviation of the real-time operating parameters relative to the target set value of the constraint node and the relative deviation of the real-time operating parameters relative to the adjacent constraint nodes in the conflict propagation path, weight and fuse the two deviations to obtain the comprehensive operating deviation value, and compare the comprehensive operating deviation value with the preset deviation sensitivity threshold to obtain the operating condition sensitivity score. The structural importance score and the operating condition sensitivity score are weighted and fused to obtain a comprehensive importance score. Based on the comprehensive importance score, resource weights are allocated and optimized to obtain an adaptive weight configuration.

[0044] After obtaining the hierarchical conflict structure, it is necessary to further quantify the importance of each constraint node in the entire conflict propagation network and combine it with the dynamic deviation information of real-time conditions to provide differentiated weight configurations for subsequent reinforcement learning optimization. To this end, each constraint node in the hierarchical conflict structure is mapped to a graph node, and the identified conflict propagation paths between constraint nodes are mapped to directed graph edges, constructing a constraint conflict propagation graph. When initializing the attention weights of each graph node, different initial values ​​are assigned according to the node's level: nodes located in the source conflict layer (i.e., nodes whose conflict scores exceed a preset score threshold and are not dominated by other conflict nodes) are given the highest initial attention weight; nodes directly affected (i.e., nodes connected to the source conflict node through a one-hop path) are given a secondary initial attention weight; and nodes indirectly affected (i.e., nodes affected through two or more hop paths) are given the lowest initial attention weight. This hierarchical initialization strategy enables the graph attention mechanism to prioritize the root node of the conflict during message passing, thereby more accurately reflecting the direction and intensity of conflict propagation in subsequent weight calculations.

[0045] After constructing the constraint conflict propagation graph, the strength of the cascading effect of each graph node is used as the basis for the analysis. With path convergence A node feature vector is constructed by concatenating two dimensions—cascade effect strength (reflecting a node's ability to propagate conflict along a path) and path convergence degree (reflecting the degree to which a node serves as a convergence point for multiple conflict paths)—into the input features of each graph node. A graph attention mechanism is then used to compute the pairing of any two adjacent graph nodes. Attention coefficient between Specifically, the nodes With nodes The eigenvectors are concatenated and then transformed by a learnable linear transformation matrix. With activation vector The mapping is converted to a scalar, and then the original attention score is obtained after LeakyReLU activation. This score is then applied to the nodes. Softmax normalization is performed on all neighboring nodes to obtain normalized attention coefficients. The formula is: ;

[0046] in For nodes The input feature vector, For nodes The input feature vector, For nodes In the set of neighbor nodes in the constrained conflict propagation graph, This represents a vector concatenation operation. The normalized attention coefficient is weighted and summed with the features of neighboring nodes to obtain the node... Aggregate representation vector After nonlinear activation, the nodes are obtained. Structural importance score This score comprehensively reflects the topological position and characteristic influence of a node in the conflict propagation network.

[0047] After calculating the structural importance score, real-time operating parameters for each constraint node are obtained, including the current feeding rate, mixing uniformity detection value, and energy consumption monitoring value. For constraint nodes... Let its real-time operating parameters be... The corresponding target setting value is Then the target deviation Defined as: Meanwhile, computing nodes Relative to adjacent constraint nodes in the conflict propagation path relative deviation Defined as: The target deviation and relative deviation are weighted by coefficients. and Perform weighted fusion to obtain nodes Comprehensive operating condition deviation value : ; in and satisfy The values ​​of both are pre-calibrated based on the process type. Typically, for uniformity constraint nodes that emphasize absolute accuracy, the values ​​are set to... The value is relatively large, and for rate-constrained nodes that emphasize relative coordination, the value is relatively large. The overall operating condition deviation value is relatively large. Sensitivity threshold for deviation When a comparison is performed, At that time, the node's operating condition sensitivity score Linear mapping amplification is used to reflect that this node is more sensitive to control adjustments under the current operating conditions; when hour, Maintaining the basic level indicates that the current operating condition of this node is stable and requires no additional attention. Specifically, the operating condition sensitivity score is calculated as follows: ,in Based on sensitivity score, This is the deviation amplification factor, used to control the rate of score increase after the deviation exceeds the threshold.

[0048] Structural importance score With operating condition sensitivity score Through configurable fusion weights and Perform weighted fusion to obtain nodes Overall importance score : ,in The dynamic adjustment of both is automatically executed based on the current operating condition fluctuation level: when the system detects that the operating condition fluctuation exceeds the set proportion within a short period of time, Automatically increase the weight configuration to respond more quickly to real-time deviations; when the operating conditions tend to stabilize, The proportion is increased to maintain the stability of structural priorities.

[0049] After obtaining the overall importance score of all constraint nodes. Then, the overall importance scores of all nodes are normalized, and the normalized scores are directly mapped to the resource weights of each constrained node in the optimization objective function. ,Right now: ; This adaptive weight configuration This will serve as a weighting coefficient for the degree of constraint satisfaction in the subsequent reinforcement learning reward function. Constraint nodes with higher overall importance scores will have a greater weight in the reward calculation, thus guiding the control strategy to prioritize satisfying the most critical constraints under the current operating conditions. In actual powder feeding scenarios, if a uniformity constraint node experiences a significant increase in overall operating condition deviation due to raw material particle size fluctuations at a certain moment, its corresponding... It will automatically improve, prompting the control strategy to prioritize the goal of mixing uniformity during this period, rather than simply pursuing the feeding rate or optimal energy consumption, thereby achieving an adaptive balance of multiple objectives under dynamic operating conditions.

[0050] When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to process safety priority, and automatically restored when the recovery conditions are met, including: The adaptive weight configuration is applied to each constraint node, the constraint satisfaction status is detected, and constraint nodes with a constraint satisfaction status of non-satisfaction are identified as residual conflict nodes. The number and distribution of residual conflict nodes are counted to obtain residual conflict features. When the residual conflict characteristics meet the preset conflict triggering conditions, the level of the residual conflict node and the conflict propagation path are extracted from the hierarchical conflict structure. The process safety priority score is calculated by combining the process safety impact range and fault propagation risk of the residual conflict node. The downgrade candidate sequence is obtained by sorting the process safety priority scores from low to high. From the candidate degradation sequence, select residual conflict nodes in sequence to reduce the constraint strength to the preset degradation level, recalculate the constraint satisfaction state after degradation, stop degradation when all constraint nodes satisfy the constraints, and record the set of degradation nodes and the degradation time. The system monitors the real-time operating parameters and constraint satisfaction status of the degraded nodes. When the real-time operating parameters of the degraded node continuously meet the original constraint strength and the duration exceeds the preset stable duration, the constraint strength of the degraded node is restored from the degraded level to the original constraint strength and the set of degraded nodes is updated.

[0051] After completing the adaptive weight configuration, the optimization resource weights for each constraint node have been dynamically allocated based on structural importance and load sensitivity. This weight configuration is then applied to the constraint solving process, checking the current satisfaction status of each constraint node one by one. The constraint satisfaction status is detected by comparing the real-time load parameters corresponding to the constraint node with its constraint boundary. If the real-time parameters fall within the allowable range of the constraint, it is determined to be satisfied; otherwise, it is determined to be unsatisfied. All constraint nodes determined to be unsatisfied are identified as residual conflict nodes, and their numbers, hierarchical levels, and positional distribution in the directed constraint dependency graph are recorded to form a residual conflict feature description. The residual conflict feature includes two dimensions: the number of residual conflict nodes. And the distribution ratio vector of these nodes in each level of the hierarchical conflict structure. ,in The The component represents the first... The proportion of residual conflicting nodes in a layer to the total number of nodes in that layer.

[0052] When residual conflict characteristics meet preset conflict triggering conditions, the priority management process is initiated. The preset conflict triggering conditions can be set as follows: Exceeding the preset threshold ,or If any component exceeds the preset distribution density threshold Once the triggering conditions are met, the hierarchical number of each residual conflict node and the set of conflict propagation paths it participates in are extracted from the hierarchical conflict structure. Based on this, a process safety priority score is calculated by combining the process safety impact range and fault propagation risk of each residual conflict node. The process safety impact range is reflected in the number of downstream devices directly associated with the node in the process flow and their safety sensitivity; the fault propagation risk is quantified by the node's convergence and output impact intensity in the conflict propagation path. For residual conflict nodes... Its process safety priority score Score based on the scope of process safety impact Risk score related to fault propagation Weighted synthesis, i.e. ,in and These are the weighting coefficients for the two types of risk indicators, satisfying... .according to The nodes are sorted from low to high to obtain a demotion candidate sequence. Nodes with lower priority scores are selected first for demotion to ensure that the constraints of high security priority are maintained as much as possible.

[0053] Residual conflict nodes are selected sequentially from the degradation candidate sequence, and their constraint strength is reduced from the original level to a preset degradation level. The preset degradation level can be pre-configured as a fixed proportion of the original constraint strength according to the process type, for example, widening the constraint boundary to a certain percentage of the original boundary. times ( This indicates an expansion of the constraint scope, or a reduction of the constraint violation penalty coefficient to its original value. times ( (This indicates an increased tolerance for violations). After each node is downgraded, the constraint satisfaction status of all current constraint nodes is immediately recalculated to determine if all residual conflicts have been eliminated. If all constraint nodes are satisfied, the downgrade operation stops, and the set of nodes involved in this downgrade is recorded. and the timing of the downgrade If there are still unmet constraint nodes, the next node is selected from the degradation candidate sequence for degradation until all constraints are satisfied or the candidate sequence is exhausted. If there are still unresolved conflicts after the candidate sequence is exhausted, an abnormal state is reported and the upper-level decision-making mechanism intervenes. At this time, the reinforcement learning strategy continues to solve for the optimal control action sequence within the currently available constraint domain without interrupting the overall control process.

[0054] After the degradation operation is completed, the set of degraded nodes is... Each node in the system is continuously monitored in real time. The monitoring includes two categories: first, whether the real-time operating parameters corresponding to the node meet the constraint boundaries under the original constraint strength; and second, whether the constraint satisfaction state of the node is stable under the current degradation level. For degraded nodes... Introducing a continuous satisfaction counter The initial value is 0. Within each sampling period, if the node... If the real-time operating parameters satisfy the constraint boundary corresponding to the original constraint strength, then Increment by 1; if not satisfied, then Reset to 0. When Exceeding the preset stable counting threshold At that time, it is considered that the real-time operating condition of the node has continuously met the original constraint strength, and the duration exceeds the preset stabilization time (the stabilization time is equal to...). Multiply by the sampling period), triggering the recovery operation. The recovery operation will then affect the node. The constraint strength is restored from the downgraded level to the original constraint strength, and From the set of degraded nodes The removed data will be used for subsequent monitoring and management.

[0055] After the recovery operation is executed, the global constraint satisfaction status needs to be re-verified to ensure that the recovery operation does not introduce new conflicts. If the global constraint satisfaction status remains normal after recovery, the recovery is complete; if new constraint dissatisfaction occurs after recovery, the node is re-added to the degraded state and included in the degraded candidate sequence, awaiting the next round of degrade and recovery judgment. The above degrade and recovery mechanism operates continuously in a closed loop throughout the entire powder feeding process, ensuring the safety and adaptability of constraint management under dynamic operating conditions. The execution result of the entire priority management process is transmitted as constraint domain information to the reinforcement learning strategy, enabling the latter to stably solve the optimal control action sequence within the effective constraint domain after priority management, thereby achieving safe, efficient, and adaptive optimization control of the powder feeding process.

[0056] Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree under the adaptive weight configuration and the process deviation, including: Based on the set of degraded nodes and the non-degraded constraint nodes, a constraint domain with priority management is constructed. The constraint boundary corresponding to the degraded node is set according to the degrade level, and the constraint boundary corresponding to the non-degraded constraint node is set according to the original constraint strength. The constraint domain is used as the action space constraint of the reinforcement learning strategy. The process state parameters at the current moment are obtained as the state input of the reinforcement learning strategy. Based on the adaptive weight configuration, the degree of constraint satisfaction of each constraint node in the current state is calculated. The process target parameters at the current moment are obtained and the deviation from the actual process state parameters is calculated to obtain the process deviation. The degree of constraint satisfaction and the process deviation are weighted and fused to construct the reward function. Based on the reward function and the constraint domain, a reinforcement learning strategy is used to iteratively search for control actions within the action space constraints, calculate the cumulative reward value corresponding to each candidate control action, and select the control action sequence with the largest cumulative reward value as the optimal control action sequence. The optimal control action sequence is converted into execution instructions for the powder feeding equipment. The process feedback parameters after execution are obtained to update the process state parameters. The updated process state parameters are used as the state input of the reinforcement learning strategy at the next moment to form a closed-loop optimization.

[0057] During the constraint domain construction phase after priority management, the set of degraded nodes will be... The constraint boundaries corresponding to each degradation node are set according to the degradation level, that is, the upper and lower bounds of the constraints of the degradation node are relaxed, and the relaxation amount is based on the relaxation ratio of the constraint boundaries determined during degradation. For constraint nodes that have not been downgraded, their constraint boundaries retain their original constraint strength and are strictly limited according to the original process constraint parameters. The constraint boundaries of these two types of nodes together constitute a complete constraint domain. and will As an action space constraint for reinforcement learning strategies, it ensures that reinforcement learning always operates within the limits of process safety when searching for control actions. The introduction of action space constraints prevents reinforcement learning strategies from violating process safety boundaries during the exploration process, thereby achieving a balance between optimization efficiency and process safety.

[0058] Obtain the process state parameters at the current moment and organize them into a state vector. As a reinforcement learning strategy, at any time The state input. Based on adaptive weight configuration, the degree of constraint satisfaction of each constraint node in the current state is calculated. Specifically, for constraint nodes... Let the measured value of the process parameter corresponding to its current state be... The lower bound of the constraint is The upper bound of the constraint is Then the constraint satisfaction degree of this node. Defined as: when Falling At that time, ;when When exceeding the constraint boundary, The penalty is calculated based on the ratio of the excess amount to the boundary range, i.e. and will Cut off to Interval. The constraint satisfaction of each constraint node is adaptively optimized according to its resource weight. We perform a weighted summation to obtain the overall constraint satisfaction level. : ; in To constrain the total number of nodes. Simultaneously, obtain the current process target parameter vector. Calculate its relationship with the actual process state parameter vector. The deviation between them yields the process deviation. : Degree of constraint satisfaction process deviation Weighted fusion to construct a reward function : ,in The reward fusion weighting coefficient is used to adjust the relative contributions of constraint satisfaction and process deviation in the reward function. The value of can be configured according to actual process requirements. When the process has higher requirements for the degree of constraint satisfaction, it should be increased appropriately. When the process requires higher target tracking accuracy, appropriately reduce... .

[0059] After the reward function and constraint domain are determined, a reinforcement learning policy is applied to constrain the action space. Iterative search is performed within the system. The reinforcement learning strategy employs a policy network. Regarding the current state Action generation is performed by the policy network. As input, output candidate control actions And through projection operation Mapping to the constraint domain Within this framework, ensure the legality of the actions. For each candidate control action sequence, calculate its corresponding cumulative reward value. : ; in The total number of time steps in a single control sequence. This is a discount factor used to balance the weighting of immediate and long-term rewards. The closer the value is to 1, the more the strategy emphasizes long-term cumulative rewards; the closer it is to 0, the more the strategy favors immediate rewards. Through multiple rounds of iterative search, the cumulative reward values ​​of each candidate control action sequence are compared. Select The longest control action sequence is the optimal control action sequence. Policy network parameters In each iteration, the policy is updated according to the policy gradient method to improve the search efficiency and optimality of the policy within the constraint domain.

[0060] Optimal control action sequence The commands are converted into execution instructions for the powder feeding equipment. During the conversion process, according to the execution interface specification of the powder feeding equipment, each component in the action vector is mapped to the corresponding control instructions of the actuators, including the feed valve opening command, the screw feeder speed command, and the agitator speed command. After the execution instructions are sent to the powder feeding equipment, the process feedback parameters are collected, including the actual feed amount, powder concentration, mixing uniformity index, and equipment energy consumption. These feedback parameters are then updated to the process state parameters, forming an updated state vector. and will As the state input for the reinforcement learning strategy in the next moment, the state is passed to complete the closed-loop optimization.

[0061] During closed-loop optimization, process state parameters and process target parameters are reacquired in each control cycle, constraint satisfaction and process deviation are recalculated, the reward function is updated, and a new round of action search is performed based on the updated reward function and the current constraint domain to drive the reinforcement learning strategy. If a degraded node is detected to meet the recovery condition within a certain control cycle, the recovered constraint boundary is synchronously updated to the constraint domain. This enables the reinforcement learning strategy to automatically adapt to stricter constraint boundaries in subsequent control cycles, achieving dynamic adaptive management of the constraint domain. The entire closed-loop optimization mechanism ensures that the powder feeding process continuously approaches the optimal process state under multi-objective constraints, while taking into account both process safety and control response speed.

[0062] A second aspect of the present invention provides an AI optimization control system for a powder feeding process, comprising: The predictive analysis unit is used to extract and predict the time-series features of the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process. The constraint graph modeling unit is used to construct a set of multi-objective constraints, including uniformity constraints, rate constraints, and energy consumption constraints, based on the evolution trend and model it as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network to perform message passing and aggregation to obtain a constraint influence propagation matrix. The conflict propagation unit is used to calculate the conflict propagation path and cascade effect intensity of each constraint node based on the constraint influence propagation matrix, identify conflict nodes, and obtain a hierarchical conflict structure. The weight adaptive unit is used to calculate the importance score of each constraint node according to the hierarchical conflict structure using a graph attention mechanism, and dynamically generate an adaptive weight configuration in combination with real-time operating condition deviations. The degradation recovery unit is used to temporarily downgrade the conflicting constraints according to the process safety priority when conflicts still exist under the adaptive weight configuration, and to automatically recover them when the recovery conditions are met. The reinforcement learning unit is used to solve the optimal control action sequence within the constraint domain after priority management using a reinforcement learning strategy, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

[0063] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0064] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0065] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

Claims

1. An AI-optimized control method for powder feeding process, characterized in that, include: Temporal features were extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process; Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network for message passing and aggregation to obtain the constraint influence propagation matrix. Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained; Based on the hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node, and an adaptive weight configuration is dynamically generated by combining real-time operating condition deviations. When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to the process safety priority, and automatically restored when the recovery conditions are met; Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

2. The method according to claim 1, characterized in that, Temporal features are extracted and predicted from the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process, including: Collect multi-dimensional status data of the powder feeding system within the current time window. The multi-dimensional status data includes feeding flow rate, silo pressure, stirring speed and material temperature. Arrange the multi-dimensional status data in chronological order to form a time-series status sequence. Multi-scale temporal feature extraction is performed on the temporal state sequence. State change rate features are extracted on a short time scale, and state trend features are extracted on a long time scale. The state change rate features and the state trend features are fused to obtain a temporal feature vector. The time-series feature vector is input into the time-series prediction neural network. The time-series prediction neural network is trained by the mapping relationship between the historical time-series feature vector and the corresponding actual evolution results, and outputs a multi-dimensional state prediction value sequence within the future time window. The multi-dimensional state prediction value sequence constitutes the evolution trend of the powder feeding process.

3. The method according to claim 1, characterized in that, Based on the evolutionary trend, a multi-objective constraint set including uniformity constraints, rate constraints, and energy consumption constraints is constructed and modeled as a directed constraint dependency graph. A graph neural network is used to perform message passing and aggregation on the directed constraint dependency graph to obtain a constraint influence propagation matrix, including: Based on the multidimensional state prediction value sequence in the evolution trend of the powder feeding process, uniformity constraints, rate constraints and energy consumption constraints are calculated and used as constraint nodes to construct a multi-objective constraint set; Extract the current working condition feature vector from the multidimensional state prediction value sequence. The current working condition feature vector includes the satisfaction margin and state change rate of each constraint node. Input the current working condition feature vector into the directed edge weight prediction network and output the time-varying coupling strength value between each pair of constraint nodes. When the time-varying coupling strength value exceeds the preset strength threshold, establish a directed edge between the pair of constraint nodes and use the time-varying coupling strength value as the weight to form a time-varying directed constraint dependency graph. A time-varying directed constraint dependency graph is input into a multi-layer graph neural network. In each graph convolutional layer, the constraint nodes perform weighted aggregation of the features of adjacent nodes through directed edge weights and update their own representations to obtain the node representation matrix of each layer. The representation difference between constraint node pairs in the first layer node representation matrix is ​​calculated to obtain the first-order constraint influence matrix. The representation difference between constraint node pairs in the second layer node representation matrix is ​​calculated to obtain the second-order constraint influence matrix. The first-order constraint influence matrix and the second-order constraint influence matrix are combined hierarchically to form a hierarchical constraint influence propagation matrix.

4. The method according to claim 1, characterized in that, Based on the constraint influence propagation matrix, the conflict propagation path and cascade effect intensity of each constraint node are calculated, conflict nodes are identified, and a hierarchical conflict structure is obtained, including: The constraint influence propagation matrix is ​​subjected to a second power operation to obtain a second-order constraint influence matrix. Elements with influence intensity exceeding a preset influence threshold are extracted from the constraint influence propagation matrix and the second-order constraint influence matrix to form a first-order conflict propagation path and a second-order conflict propagation path, respectively. The two paths are then merged to obtain a set of conflict propagation paths. The propagation attenuation coefficient is calculated based on the path length of each conflict propagation path, and the propagation attenuation coefficient is multiplied by the influence intensity of each node on the path to obtain the cascade effect intensity of each conflict propagation path. The path convergence degree is obtained by counting the number of times each constraint node is traversed by the conflict propagation path. The node influence output strength is obtained by calculating the sum of the cascade effect strength of each constraint node. The node conflict score is obtained by weighted summing of the path convergence degree and the node influence output strength. Constraint nodes whose node conflict scores exceed a preset score threshold are identified as conflict nodes. Calculate the shortest propagation distance from each conflict node to other constraint nodes, and divide the constraint nodes into source conflict layer, directly affected layer and indirectly affected layer according to the shortest propagation distance to form a hierarchical conflict structure.

5. The method according to claim 4, characterized in that, Based on the aforementioned hierarchical conflict structure, a graph attention mechanism is used to calculate the importance score of each constraint node. An adaptive weight configuration is then dynamically generated by combining real-time operating condition deviations, including: Construct a constraint conflict propagation graph by treating each constraint node in the hierarchical conflict structure as a graph node and the conflict propagation path between constraint nodes as graph edges. Based on the cascading effect strength and path convergence of each graph node in the constraint conflict propagation graph, the attention coefficient of each graph node relative to its neighboring graph nodes is calculated through the graph attention mechanism. The attention coefficient is then weighted and aggregated with the features of the neighboring graph nodes to obtain the structural importance score. Obtain the real-time operating parameters of each constraint node, calculate the target deviation of the real-time operating parameters relative to the target set value of the constraint node and the relative deviation of the real-time operating parameters relative to the adjacent constraint nodes in the conflict propagation path, weight and fuse the two deviations to obtain the comprehensive operating deviation value, and compare the comprehensive operating deviation value with the preset deviation sensitivity threshold to obtain the operating condition sensitivity score. The structural importance score and the operating condition sensitivity score are weighted and fused to obtain a comprehensive importance score. Based on the comprehensive importance score, resource weights are allocated and optimized to obtain an adaptive weight configuration.

6. The method according to claim 1, characterized in that, When conflicts still exist under the adaptive weight configuration, the conflict constraints are temporarily downgraded according to process safety priority, and automatically restored when the recovery conditions are met, including: The adaptive weight configuration is applied to each constraint node, the constraint satisfaction status is detected, and constraint nodes with a constraint satisfaction status of non-satisfaction are identified as residual conflict nodes. The number and distribution of residual conflict nodes are counted to obtain residual conflict features. When the residual conflict characteristics meet the preset conflict triggering conditions, the level of the residual conflict node and the conflict propagation path are extracted from the hierarchical conflict structure. The process safety priority score is calculated by combining the process safety impact range and fault propagation risk of the residual conflict node. The downgrade candidate sequence is obtained by sorting the process safety priority scores from low to high. From the candidate degradation sequence, select residual conflict nodes in sequence to reduce the constraint strength to the preset degradation level, recalculate the constraint satisfaction state after degradation, stop degradation when all constraint nodes satisfy the constraints, and record the set of degradation nodes and the degradation time. The system monitors the real-time operating parameters and constraint satisfaction status of the degraded nodes. When the real-time operating parameters of the degraded node continuously meet the original constraint strength and the duration exceeds the preset stable duration, the constraint strength of the degraded node is restored from the degraded level to the original constraint strength and the set of degraded nodes is updated.

7. The method according to claim 6, characterized in that, Within the constraint domain after priority management, a reinforcement learning strategy is used to solve for the optimal control action sequence, wherein the reward function is composed of the constraint satisfaction degree under the adaptive weight configuration and the process deviation, including: Based on the set of degraded nodes and the non-degraded constraint nodes, a constraint domain with priority management is constructed. The constraint boundary corresponding to the degraded node is set according to the degrade level, and the constraint boundary corresponding to the non-degraded constraint node is set according to the original constraint strength. The constraint domain is used as the action space constraint of the reinforcement learning strategy. The process state parameters at the current moment are obtained as the state input of the reinforcement learning strategy. Based on the adaptive weight configuration, the degree of constraint satisfaction of each constraint node in the current state is calculated. The process target parameters at the current moment are obtained and the deviation from the actual process state parameters is calculated to obtain the process deviation. The degree of constraint satisfaction and the process deviation are weighted and fused to construct the reward function. Based on the reward function and the constraint domain, a reinforcement learning strategy is used to iteratively search for control actions within the action space constraints, calculate the cumulative reward value corresponding to each candidate control action, and select the control action sequence with the largest cumulative reward value as the optimal control action sequence. The optimal control action sequence is converted into execution instructions for the powder feeding equipment. The process feedback parameters after execution are obtained to update the process state parameters. The updated process state parameters are used as the state input of the reinforcement learning strategy at the next moment to form a closed-loop optimization.

8. An AI-optimized control system for powder feeding process, used to implement the method as described in any one of claims 1-7, characterized in that, include: The predictive analysis unit is used to extract and predict the time-series features of the real-time status data of the powder feeding system to obtain the evolution trend of the powder feeding process. The constraint graph modeling unit is used to construct a set of multi-objective constraints, including uniformity constraints, rate constraints, and energy consumption constraints, based on the evolution trend and model it as a directed constraint dependency graph. The directed constraint dependency graph is then processed by a graph neural network to perform message passing and aggregation to obtain a constraint influence propagation matrix. The conflict propagation unit is used to calculate the conflict propagation path and cascade effect intensity of each constraint node based on the constraint influence propagation matrix, identify conflict nodes, and obtain a hierarchical conflict structure. The weight adaptive unit is used to calculate the importance score of each constraint node according to the hierarchical conflict structure using a graph attention mechanism, and dynamically generate an adaptive weight configuration in combination with real-time operating condition deviations. The degradation recovery unit is used to temporarily downgrade the conflicting constraints according to the process safety priority when conflicts still exist under the adaptive weight configuration, and to automatically recover them when the recovery conditions are met. The reinforcement learning unit is used to solve the optimal control action sequence within the constraint domain after priority management using a reinforcement learning strategy, wherein the reward function is composed of the constraint satisfaction degree and process deviation under the adaptive weight configuration.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images