Powder particle distribution prediction method and equipment based on machine learning and medium

Abstract

The invention discloses a powder particle distribution prediction method and device based on machine learning and a medium, and relates to the technical field of industrial control, and the method comprises the steps: collecting multi-source time sequence data related to particle distribution, learning a causal dependency relationship among parameters in the multi-source time sequence data through a causal discovery algorithm, and generating a causal graph; taking a causal graph as a guidance basis of data coordination, and adopting a game theory coordination mechanism to execute dynamic weight distribution for the multi-source time sequence data to obtain a coordination data vector; constructing a powder particle distribution prediction model, inputting the coordination data vector into the powder particle distribution prediction model, and outputting a powder particle distribution prediction value at a future moment; and calculating the difference quantity between the intervened particle distribution simulation result and the powder particle distribution predicted value, and generating a process parameter adjustment suggestion. According to the method, through forward propagation calculation of the causal diagram, accurate quantitative simulation of a process intervention effect is realized, and interpretability and reliability of a prediction result are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial control, and in particular to a powder particle distribution prediction method, device and medium based on machine learning. BACKGROUND

[0002] The powder particle distribution prediction technology has important application value in pharmaceutical, metallurgical and chemical industrial processes. In recent years, with the development of machine learning technology, the prediction method based on time series data has gradually become a research hotspot. Among them, the causal discovery algorithm can identify the causal dependence relationship between variables from multi-source data, and the game theory coordination mechanism provides a theoretical framework for multi-modal data fusion. These technologies lay the foundation for building more accurate prediction models and promote the transformation from traditional statistical methods to data-driven methods.

[0003] However, the existing technology has deficiencies in multi-source data integration and intervention effect simulation, especially the most methods lack in-depth use of causal structure, which cannot effectively support counterfactual reasoning, leading to difficulty in predicting the change of particle distribution after intervention in process parameter optimization. This is mainly due to the fact that existing models usually rely on correlation rather than causality, ignoring the causal mechanism between variables, thereby limiting the practicality and interpretability of the prediction results in dynamic industrial scenarios. SUMMARY

[0004] In view of the above existing problems, the present application provides a powder particle distribution prediction method based on machine learning to solve the problem that the existing powder particle distribution prediction method cannot accurately evaluate the process intervention effect due to the lack of causal reasoning and counterfactual simulation capability. In order to solve the above technical problems, the present application provides the following technical scheme:

[0005] In a first aspect, the present application provides a powder particle distribution prediction method based on machine learning, which comprises: collecting multi-source time series data related to particle distribution; learning the causal dependence relationship between parameters in the multi-source time series data using a causal discovery algorithm to generate a causal graph; using the causal graph as a guide for data coordination, performing dynamic weight allocation for multi-source time series data using a game theory coordination mechanism to obtain a coordinated data vector; constructing a powder particle distribution prediction model, inputting the coordinated data vector into the powder particle distribution prediction model, and outputting a powder particle distribution prediction value at a future time; performing counterfactual reasoning on the predefined process parameter intervention instruction and the powder particle distribution prediction value at the future time to generate a basic counterfactual distribution state; performing forward propagation calculation of intervention effect on the basic counterfactual distribution state by coupling the structural causal mechanism contained in the causal graph to generate a particle distribution simulation result after intervention; calculating the difference between the particle distribution simulation result after intervention and the powder particle distribution prediction value to generate a process parameter adjustment suggestion.

[0006] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the generation of the causal graph is as follows,

[0007] The multi-source time series data includes process parameter time series, online measurement parameter time series and target particle distribution parameter time series.

[0008] The process parameter time series, the online measurement parameter time series and the target particle distribution parameter time series are constructed into three-dimensional tensors, and each parameter time series in the three-dimensional tensors is denoised to obtain denoised time series data.

[0009] Based on the denoised time series data, the conditional mutual information between each parameter time series is determined to construct the causal relationship.

[0010] The PC algorithm is used to verify and optimize the causal relationship to generate a candidate causal graph structure.

[0011] The confidence of the causal edges in the candidate causal graph structure is evaluated, and the causal edges below the preset confidence threshold are removed to generate the causal graph.

[0012] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the generation of the coordinated data vector is as follows,

[0013] Each parameter time series corresponding to the nodes in the causal graph is constructed into a cooperative game model, and the Shapley value of each parameter time series is calculated as an initial weight based on the cooperative game model.

[0014] The real-time data quality indicators of the process parameter time series, the online measurement parameter time series and the target particle distribution parameter time series are calculated, and combined with the initial weight to generate a dynamic weight.

[0015] The dynamic weight is iteratively optimized by multiple rounds of negotiation based on the cooperative game model until convergence, and an optimized weight is generated. The optimized weight is weighted and fused with the multi-source time series data to generate a coordinated data vector.

[0016] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the construction of the powder particle distribution prediction model is as follows,

[0017] A time series feature extraction network is constructed through an LSTM network and an attention mechanism.

[0018] A continuous time prediction architecture is established based on the time series feature extraction network, and the continuous time prediction architecture is pre-trained through historical particle distribution data to complete the construction of the powder particle distribution prediction model.

[0019] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the output of the future time powder particle distribution prediction value is as follows,

[0020] The coordination data vector is input into the time sequence feature extraction network in the powder particle distribution prediction model to extract the time sequence feature.

[0021] The time sequence feature is input into the continuous time prediction architecture in the powder particle distribution prediction model for numerical solution, and the future time powder particle distribution prediction value is output.

[0022] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the output of the future time powder particle distribution prediction value is as follows,

[0023] Based on the causal diagram and the pre-defined process parameter intervention instruction, an intervention-aware causal state encoder is constructed to encode the state of the future time powder particle distribution prediction value under causal constraints, and a state code is generated.

[0024] The state code under causal constraints is input into the causal evolution based on neural ordinary differential equation for multi-step forward state propagation to generate an anti-factual state evolution trajectory.

[0025] The anti-factual state evolution trajectory is subjected to time sequence smoothing processing to output the basic anti-factual distribution state.

[0026] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the output of the future time powder particle distribution prediction value is as follows,

[0027] Based on the causal diagram, an intervention propagation network is constructed, and the basic anti-factual distribution state is input into the intervention propagation network for structured propagation of intervention effects to obtain a dynamic propagation result.

[0028] The dynamic propagation result is input into the neural ordinary differential equation framework for multi-path intervention effect integral analysis to output the cumulative integral result of the intervention effect.

[0029] The cumulative integral result and the basic anti-factual distribution state are weighted and fused to generate the intervention-based particle distribution simulation result.

[0030] As a preferred scheme of the powder particle distribution prediction method based on machine learning, the output of the future time powder particle distribution prediction value is as follows,

[0031] Based on the multi-scale distribution difference measurement method, the intervention-based particle distribution simulation result and the powder particle distribution prediction value are subjected to difference quantification to obtain a multi-dimensional difference measurement result.

[0032] input the multi-dimensional difference measurement result into a multi-objective optimization framework with causal constraints, and output a process parameter adjustment strategy;

[0033] verify the process parameter adjustment strategy through a reinforcement learning mechanism, and generate an executable process parameter adjustment suggestion.

[0034] In a second aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the machine learning-based powder particle distribution prediction method according to the first aspect of the present application.

[0035] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the machine learning-based powder particle distribution prediction method according to the first aspect of the present application.

[0036] The present application has the following beneficial effects: through forward propagation calculation of the causal graph, the precise quantitative simulation of the process intervention effect is realized, and the interpretability and reliability of the prediction result are improved; the structural information of the causal graph is converted into a calculable propagation path, the cumulative influence of the intervention effect is captured through a multi-path integration method, and a particle distribution simulation result conforming to the physical law is generated, which can accurately reflect the dynamic influence process of the process parameter adjustment on the particle distribution, and improve the accuracy of the intervention effect prediction under complex working conditions, thereby providing a reliable theoretical basis and decision support for process optimization. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0038] Fig. 1 Flowchart of the machine learning-based powder particle distribution prediction method.

[0039] Fig. 2 Flowchart for generating a causal graph.

[0040] Fig. 3 Flowchart for obtaining a coordination data vector.

[0041] Fig. 4 Flowchart for constructing a powder particle distribution prediction model and outputting a predicted value of the powder particle distribution at a future time. DETAILED DESCRIPTION

[0042] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0043] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other ways that are not exactly as described in this description. It is understood that variations can be made in view of what is described and understood that it can be done in different but equivalent ways. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0044] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in this specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0045] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a powder particle distribution prediction method based on machine learning, comprising the following steps:

[0046] S1, collect multi-source time series data related to particle distribution, learn the causal dependence relationship between parameters in multi-source time series data by using causal discovery algorithm, and generate causal graph.

[0047] S1.1: The multi-source time series data includes process parameter time series, online measurement parameter time series and target particle distribution parameter time series;

[0048] It should be noted that the process parameter time series refers to the data sequence of dynamic parameters directly related to process operation recorded by equipment over time in the powder production or processing process, such as stirring speed, feeding rate, temperature, pressure, grinding time and air flow rate, etc., which reflects the running state of the production process and is the key controllable variable affecting the particle distribution;

[0049] The online measurement parameter time series refers to the sequence of intermediate state data continuously collected by real-time sensors or online monitoring instruments (such as laser particle size analyzer, optical spectrum sensor, etc.) installed on the production line over time, such as suspension concentration, particle flow rate, optical scattering signal, humidity or conductivity, etc., which can indirectly reflect the real-time evolution of particle morphology or distribution, and has high timeliness and process feedback value;

[0050] The target particle distribution parameter time series refers to the time series of particle size distribution characteristics obtained by particle size analysis of powder products at different time points, which is the core index directly representing product quality, the output target of the prediction model, and the benchmark basis for evaluating process effect and conducting counterfactual reasoning.

[0051] S1.2: Construct the process parameter time series, online measurement parameter time series and target particle distribution parameter time series into a three-dimensional tensor, and denoise each parameter time series in the three-dimensional tensor to obtain denoised time series data;

[0052] Specifically, the process parameter time series, online measurement parameter time series and target particle distribution parameter time series are aligned according to time steps, and a three-dimensional tensor is formed with time steps as the first dimension, parameter types as the second dimension, and observation values of each parameter at corresponding time points as the third dimension. The process parameter time series, online measurement parameter time series and target particle distribution parameter time series contained in the three-dimensional tensor are denoised respectively. Wavelet transform method is used to implement filtering operation on each parameter time series, and noise components are removed through decomposition and reconstruction to obtain denoised time series data.

[0053] S1.3: Based on the denoised time series data, determine the conditional mutual information between each parameter time series, and construct the causal relationship;

[0054] Specifically, based on the denoised time series data, for any two parameter time series in the process parameter time series, online measurement parameter time series and target particle distribution parameter time series, under the premise that the remaining parameter time series are taken as conditional variables, according to the definition of conditional mutual information in information theory, the joint occurrence frequency and the marginal occurrence frequency of each parameter time series under different value combinations are counted to obtain the conditional mutual information for representing the information sharing degree of one parameter time series to another parameter time series under given conditions. The conditional mutual information is used as the basis for judging the dependence strength between variables. When the conditional mutual information is higher than the preset conditional mutual information threshold, it is considered that there is a causal relationship between the corresponding two parameter time series.

[0055] It should be noted that the definition of conditional mutual information in information theory refers to the degree of mutual dependence between one random variable and another random variable under the condition that the value of the third variable (or a group of variables) is known, which reflects the amount of shared information that two variables still retain after excluding the influence of known variables.

[0056] It should be noted that the conditional mutual information threshold is set through the following steps: Based on the denoised time series data, the time series of process parameters, online measurement parameters, and target particle distribution parameters are subjected to multiple random shuffling operations to generate multiple sets of pseudo-data without real time series dependencies; the conditional mutual information between the time series of each parameter is repeatedly obtained on each set of pseudo-data to form an empirical distribution of conditional mutual information dominated by random noise; the percentile of the empirical distribution (e.g., the 95th percentile) is taken as the conditional mutual information threshold to ensure that only values ​​higher than the random level are used. Only conditional mutual information is considered a potential causal relationship; the exemplary value range is between 0.01 and 0.1, and the value is determined by controlling the false positive rate so that in variable pairs without a real causal relationship, only a small probability (e.g., 5%) exceeds the conditional mutual information threshold due to random fluctuations; if it is lower than 0.01, a large number of weak associations caused by noise will be misjudged as causal relationships, resulting in too many false edges in the causal graph and reducing the reliability of the structure; if it is higher than 0.1, real weak causal relationships may be ignored, causing the causal graph to miss key dependency paths and affecting the integrity of prediction and intervention simulation.

[0057] S1.4: The PC algorithm is used to verify and optimize the causal relationship and generate a candidate causal graph structure;

[0058] Specifically, using all parameter time series as nodes, undirected edges are established for variable pairs whose conditional mutual information is higher than a preset conditional mutual information threshold, forming a fully connected initial graph structure. The PC (causal discovery) algorithm is used to perform conditional independence tests under conditional sets of different orders, gradually removing edges between variable pairs that are independent under the given conditional time series of other parameters. The remaining ternary structures (i.e., X—Y—Z and X and Z are not connected) are subjected to V-structure determination. If Y is not in the conditional set of X and Z, the edge is directed to X→Y←Z. The derivable directions are expanded through the transitive closure rules of the graph structure to eliminate loops and unify directional consistency, generating candidate causal graph structures.

[0059] It should be noted that condition sets of different orders refer to sets that contain zero, one, two or more other parameter time series as condition variables in the conditional independence test, respectively. The order corresponds to the number of variables in the condition set.

[0060] S1.5: Evaluate the confidence of causal edges in the candidate causal graph structure, remove causal edges that are below the preset confidence threshold, and generate the causal graph.

[0061] Specifically, for each causal edge in the candidate causal graph structure, the Bootstrap resampling method is used to generate multiple resampled subsets from the denoised time series data. The PC algorithm is repeatedly executed on each resampled subset to obtain the frequency of the causal edge in multiple resamplings, and the frequency is used as the confidence of the causal edge. The confidence of each causal edge is compared with a preset confidence threshold, and causal edges with confidence below the preset confidence threshold are removed, while the remaining causal edges are retained to form the causal graph.

[0062] It should be noted that the confidence threshold is set based on the requirement for the stability of causal edges. The specific setting steps are as follows: determine the initial confidence level range through rules of thumb or domain knowledge; use a set of data with known causal relationships to test and adjust the confidence threshold until a balance point is found that can exclude unstable causal edges while retaining stable causal relationships. The exemplary value range is usually between 0.7 and 0.9, which can ensure that causal edges with sufficient statistical significance are selected without being too strict and mistakenly deleting real but weak causal relationships. If it is lower than 0.7, too many unstable causal edges may be included in the causal graph, increasing the false positive rate; if it is higher than 0.9, effective but slightly weak real causal relationships may be mistakenly removed, resulting in the loss of causal graph information.

[0063] S2. Using causal graphs as the guiding principle for data coordination, a game theory coordination mechanism is adopted to perform dynamic weight allocation for multi-source time series data, resulting in a coordinated data vector.

[0064] S2.1: Construct a cooperative game model from the time series of each parameter corresponding to the node in the causal graph, and calculate the Shapley value of each parameter time series as the initial weight based on the cooperative game model;

[0065] Specifically, each node in the causal graph is mapped to the time series of process parameters, the time series of online measurement parameters, and the time series of target particle distribution parameters, respectively. A cooperative game model is constructed with each parameter time series as a participant. The characteristic function of the cooperative game model is defined as the joint contribution of any parameter time series in predicting particle distribution. Based on the cooperative game model, the Shapley value of each parameter time series is calculated, and the normalized Shapley value is used as the initial weight.

[0066] Based on the cooperative game theory model, the Shapley values ​​of each parameter's time series are calculated using the following expression:

[0067] ;

[0068] In the formula, Indicates the first Shapley values ​​for a time series with one parameter. Indicates the time series index of the parameter. This indicates a selection from all parameter time series that does not contain parameter time series. any subset, Represents a set of time series parameters. Representing a subset Number of elements factorial, This represents the total number of time series parameters involved in the cooperative game. Representing a subset The number of time series parameters included. This represents the total number of time series parameters involved in the cooperative game. factorial, The characteristic function representing a cooperative game. This indicates that only the index is included. A set of parameter time series, Indicates when subset Adding parameters to time series Then, in the characteristic function The joint contribution value under, Indicates only by subset When participating in prediction, in the feature function The combined contribution value.

[0069] S2.2: Calculate the real-time data quality indicators of the time series of process parameters, online measurement parameters, and target particle distribution parameters, and combine them with the initial weights to generate dynamic weights;

[0070] Specifically, the real-time data quality indicators of the time series of process parameters, online measurement parameters, and target particle distribution parameters are calculated; the real-time data quality indicators corresponding to each parameter time series are combined with the initial weights to obtain adjustment factors that take into account the reliability of the current data; the adjustment factors of all parameter time series are normalized to form dynamic weights.

[0071] The real-time data quality index for calculating the time series of process parameters, online measurement parameters, and target particle distribution parameters is expressed as follows:

[0072] The sampling integrity index is expressed as:

[0073] ;

[0074] In the formula, Indicates at time Above, the first Sampling integrity index for time series with one parameter. Indicates at time The length of the end is Within the time window, the first The number of valid sampling points actually collected in the time series of each parameter. This indicates the number of sampling points corresponding to the length of the time window.

[0075] The mean drift stability index is expressed as:

[0076] ;

[0077] In the formula, Indicates at time Above, the first Mean drift stability index for time series with multiple parameters Indicates at time Within the time window at the end, the first The mean of a time series of parameters. Indicates the first Reference mean of each parameter time series, This represents an extremely small positive number introduced to avoid the denominator being zero or too small (e.g., (magnitude).

[0078] The volatility level indicator is expressed as follows:

[0079] ;

[0080] In the formula, Indicates at time Above, the first A volatility level indicator for a time series with one parameter. Indicates at time Within the time window at the end, the first Standard deviation of each parameter time series Indicates the first Reference standard deviation of each parameter time series.

[0081] The comprehensive real-time data quality index is expressed as follows:

[0082] ;

[0083] In the formula, Indicates at time Above, the first A comprehensive real-time data quality index for time series with multiple parameters. This indicates the weighting coefficient of the sampling integrity index in the overall data quality index. This represents the weighting coefficient of the mean drift stability index in the overall data quality index. This indicates the weighting coefficient of the volatility level indicator in the overall data quality index.

[0084] It should be noted that, This setting is based on an assessment of the importance of completeness of sampling. The specific setting steps are as follows: First, use a batch of historical windows with manual data quality levels to calculate the corresponding... and In order to satisfy Under the premise of adjusting through simple grid search Select to make The value that best matches human evaluation; in the scenario of powder particle distribution prediction, as long as the proportion of effective sampling points is within the process allowable range (e.g., ≥80%), the impact of slightly missing samples on the prediction is less than mean drift and abnormal fluctuations. An example value that can be taken is... This reflects a setting that emphasizes sampling integrity but has a slightly lower weight;

[0085] The setting is based on the impact of mean drift on process deviation and prediction accuracy. The specific setting steps are as follows: Select time periods from historical data where the mean deviates significantly from the reference value, statistically analyze the particle distribution prediction error and batch quality changes during these periods, and then adjust... The size of makes When it decreases, It can sensitively reflect a decline in data quality; since deviation from the mean often means that the process center value has deviated, it directly leads to an overall coarser or finer particle distribution, and the impact is usually significant. An example can be taken... In the comprehensive index, a higher weight is given to mean drift stability;

[0086] Based on the setting that focuses on characterizing whether the level of fluctuation has increased abnormally, the specific setting steps are as follows: extract high-noise periods with significantly amplified standard deviations from historical data, and analyze these periods. Exceed The impact on predicted residuals and non-conforming rate, and in Adjustment under constraints This causes the fluctuations to amplify abnormally. It can be significantly reduced; considering that short-term drastic fluctuations in powder processing often indicate unstable feeding or control loop oscillations, the impact is comparable to mean drift, and an example can be taken. ,and The result was the same.

[0087] S2.3: Based on the cooperative game model, the dynamic weights are iteratively optimized through multiple rounds of negotiation until convergence, generating optimized weights. The optimized weights are then weighted and fused with multi-source time series data to generate a coordinated data vector.

[0088] Specifically, based on the cooperative game model, dynamic weights are used as the weight allocation scheme for each parameter time series in the current round. In each round of negotiation, the weights of the process parameter time series, online measurement parameter time series, and target particle distribution parameter time series are adjusted according to the causal dependency direction between nodes in the causal graph. This ensures that the parameter time series upstream in the causal chain affects the weights of its downstream parameter time series, and the magnitude of the weight change is used to determine whether it is less than a preset change threshold. Multiple rounds of weight adjustment are repeated until the weight changes of all parameter time series in two adjacent rounds are less than the preset change threshold. The weights at this point are the optimized weights. The optimized weights are then combined with the observations of the corresponding process parameter time series, online measurement parameter time series, and target particle distribution parameter time series at the same time point to form a coordinated data vector.

[0089] It should be noted that the change threshold is set through the following steps: Multiple rounds of weight negotiation experiments are conducted on historical operating data to record the changing trends of the time series weights of each parameter during the iteration process; the fluctuation amplitude when the weight changes tend to stabilize is statistically analyzed, and the upper limit of the fluctuation amplitude is selected as the initial reference value; then, combined with the requirements of actual industrial scenarios for the stability of the coordinated data vector, a small-range adjustment is made near the initial reference value to determine the change threshold that ensures the timely termination of the negotiation process without sacrificing weight accuracy; the exemplary value range is between 0.005 and 0.02; the value is chosen to balance algorithm efficiency and weight stability; if it is below 0.005, the negotiation process may continue for too many rounds due to excessive sensitivity to small fluctuations, prolonging the response time and affecting real-time performance; if it is above 0.02, the negotiation may terminate prematurely before the weights are sufficiently stable, causing the optimized weights to deviate from the reasonable allocation and reducing the reliability of the coordinated data vector.

[0090] S3. Construct a powder particle distribution prediction model, input the coordinated data vector into the powder particle distribution prediction model, and output the predicted powder particle distribution value at future time.

[0091] S3.1: Construct a temporal feature extraction network using an LSTM network and an attention mechanism;

[0092] Specifically, an LSTM network is used as the basic temporal model to capture long-term dependencies in the time series of process parameters, online measurement parameters, and target particle distribution parameters step by step. An attention mechanism is superimposed on the hidden state sequence of the LSTM network, enabling the LSTM network to assign different levels of attention to the hidden states at different time steps, highlighting historical moments that have a key impact on particle distribution prediction. The LSTM network and the attention mechanism are connected sequentially to form an end-to-end structure, completing the construction of the temporal feature extraction network.

[0093] S3.2: Establish a continuous-time prediction architecture based on a time-series feature extraction network, and pre-train the continuous-time prediction architecture using historical particle distribution data to complete the construction of a powder particle distribution prediction model.

[0094] Specifically, based on the temporal feature extraction network, an ordinary differential equation network layer is connected. The temporal features output by the temporal feature extraction network are used as the initial state. The ordinary differential equation network layer describes the dynamic process of particle distribution evolving continuously over time, forming a continuous time prediction architecture. By utilizing the historical coordinated data vector sequence and its corresponding historical target particle distribution parameter time series contained in the historical particle distribution data, the internal parameters of the LSTM network, attention mechanism, and ordinary differential equation network layer in the continuous time prediction architecture are adjusted. This enables the continuous time prediction architecture to reproduce the corresponding historical target particle distribution parameter time series given the historical coordinated data vector sequence, thus completing the construction of the powder particle distribution prediction model.

[0095] S3.3: Input the coordinated data vector into the temporal feature extraction network in the powder particle distribution prediction model to extract temporal features;

[0096] Specifically, the coordinated data vector is fed into the temporal feature extraction network in the powder particle distribution prediction model in chronological order. The temporal feature extraction network updates its internal state step by step through the LSTM network to capture the long-term dynamic dependence between the time series of process parameters, the time series of online measurement parameters, and the time series of target particle distribution parameters. An attention mechanism is applied to the LSTM hidden state at all time steps to focus on the time segments that have a key impact on the evolution of particle distribution, thus forming temporal features.

[0097] It should be noted that the key impact time segment refers to the historical time point or period in the coordinated data vector that has a significant effect on the change of powder particle distribution. It is usually manifested as a sudden change in process parameters, abnormal measurement signals, or a period in which the particle distribution changes significantly.

[0098] S3.4: Input the temporal features into the continuous-time prediction framework of the powder particle distribution prediction model for numerical solution, and output the predicted powder particle distribution values ​​for future times. The expression is:

[0099] ;

[0100] In the formula, Indicates the start time of the prediction. Indicates the prediction time step. Indicates in Predicted values ​​of powder particle distribution, Indicates the starting time The actual powder particle distribution state vector. Represents a continuous-time dynamic function. This represents the set of trainable parameters, including the internal parameters of the LSTM network, attention mechanism, and ordinary differential equation network layers. Indicates the time variable The distribution state of powder particles on the surface, Represents the temporal feature vector. Represents the time variable The tiny time increment during integration.

[0101] S4. Perform counterfactual reasoning on the predefined process parameter intervention instructions and the predicted values ​​of powder particle distribution at future times to generate the basic counterfactual distribution state.

[0102] S4.1: Based on the causal graph and predefined process parameter intervention instructions, construct an intervention-aware causal state encoder, perform counterfactual state encoding on the predicted value of powder particle distribution at future time, and generate a state code under causal constraints;

[0103] Specifically, based on the causal dependencies between nodes in the causal graph and predefined process parameter intervention instructions, the predicted values ​​of powder particle distribution at future times are structurally aligned with the time series of the intervened process parameters. Under the variable dependency structure constraints specified by the causal graph, the original values ​​of the uninterrupted parameter time series are maintained, while the intervened process parameter time series are replaced with the values ​​specified by the predefined process parameter intervention instructions, forming a counterfactual initial state that conforms to the causal mechanism. The counterfactual initial state is encoded by an intervention-aware causal state encoder, and the intervention information is propagated layer by layer according to the topological structure of the causal graph to ensure that the encoding results satisfy the conditional independence constraints in the causal graph, generating a state code under causal constraints.

[0104] It should be noted that the predefined process parameter intervention instruction refers to the specific adjustment instruction that is pre-set by the control equipment during the powder production process to change the value of one or more process parameters in the future time period.

[0105] S4.2: Input the state code under causal constraints into the causal evolutioner based on the divine ordinary differential equation for multi-step forward state propagation to generate the counterfactual state evolution trajectory;

[0106] Specifically, the state encoding under causal constraints is fed as the initial state into a causal evolver based on neural ordinary differential equations. The causal evolver is composed of ordinary differential equations parameterized by a neural network, and its dynamic evolution rules are consistent with the structural causal mechanism between variables in the causal graph. Starting from the current moment on the time axis, the state evolution is gradually advanced according to the preset time step. Each step of the state update follows the continuous dynamic law described by the neural ordinary differential equations and is constrained by the influence direction of the parent node on the child node in the causal graph. After forward propagation of multiple time steps, counterfactual states are formed in chronological order, constituting the counterfactual state evolution trajectory.

[0107] It should be noted that the time step is set based on the original time interval of data acquisition during the powder production process and the physical response speed of the dynamic changes in particle distribution. The exemplary value range is between 0.1 seconds and 5 seconds. The value is determined by balancing the accuracy of numerical integration with computational efficiency, while matching the dynamic characteristics of the actual industrial process. If it is less than 0.1 seconds, it will result in too many evolution steps, introducing unnecessary small fluctuations, increasing redundancy, and possibly amplifying the impact of noise. If it is more than 5 seconds, it may skip the critical transition stage of particle distribution change, causing the counterfactual state evolution trajectory to be distorted and unable to accurately reflect the dynamic response process after intervention.

[0108] S4.3: Perform time-series smoothing on the counterfactual state evolution trajectory and output the basic counterfactual distribution state.

[0109] Specifically, for the particle distribution state at each time point in the counterfactual state evolution trajectory, a sliding window method is used to select the state values ​​of several adjacent time points along the time axis. The local weighted average method is used to fuse the state values ​​within the window, making the state change transition between adjacent time points more continuous, eliminating abrupt changes or jitters caused by numerical integration or discrete sampling, and forming a smoothed state sequence as the basic counterfactual distribution state.

[0110] S5. By using the structural causal mechanism implied in the coupled causal graph, the forward propagation calculation of the intervention effect on the basic counterfactual distribution state is performed to generate the particle distribution simulation results after intervention.

[0111] S5.1: Construct an intervention propagation network based on a causal graph, input the basic counterfactual distribution state into the intervention propagation network to propagate the intervention effect in a structured manner and obtain dynamic propagation results;

[0112] Specifically, based on the directed causal edges between nodes in the causal graph, each node is mapped to a state variable of the time series of process parameters, online measurement parameters, or target particle distribution parameters. An information transmission path from the parent node to the child node is established according to the direction of the causal edge, forming an intervention propagation network. The values ​​of each variable in the basic counterfactual distribution state are used as the initial state. The intervention effect is transmitted layer by layer along the directed path in the intervention propagation network. The state change of the parent node drives the state adjustment of the child node, and the child node then continues to transmit its updated state downstream. After multi-level cascade propagation, a dynamic propagation result reflecting the step-by-step diffusion of the intervention effect in the causal structure is obtained.

[0113] S5.2: Input the dynamic propagation results into the neural ordinary differential equation framework to perform multi-path intervention effect integral analysis, and output the cumulative integral results of the intervention effect;

[0114] Specifically, the dynamic propagation results are embedded as the initial state into the neural ordinary differential equation framework. The neural ordinary differential equation framework is composed of ordinary differential equations parameterized by a neural network, and its vector field structure is aligned with multiple causal paths in the causal graph. In the time dimension, starting from the intervention start time, continuous time evolution is performed along each causal path that starts from the time series of the intervened process parameters, passes through the online measurement parameter time series, and reaches the target particle distribution parameter time series. The trajectory of state change over time on each path is recorded. The state change of each path within the prediction time interval is integrated, and the integral contributions of all paths are summarized to form the cumulative integral result of the intervention effect, which reflects the cumulative intensity of the intervention effect under the combined action of multiple paths.

[0115] S5.3: The cumulative integral result is weighted and fused with the basic counterfactual distribution state to generate the particle distribution simulation result after intervention.

[0116] Specifically, the cumulative integral result of the intervention effect is linearly combined with the granular distribution characteristics of the basic counterfactual distribution at the same time point. The cumulative integral result of the intervention effect reflects the incremental changes brought about by multi-path intervention, and the basic counterfactual distribution represents the initial counterfactual state without considering the cumulative effect of intervention propagation. The cumulative integral result of the intervention effect and the basic counterfactual distribution are weighted and superimposed using a preset fusion weight to form the post-intervention granular distribution simulation result that includes both the initial counterfactual state and the cumulative impact of intervention.

[0117] It should be noted that the fusion weight is jointly set based on the intensity level of the intervention instruction and the number of paths from the intervened variable to the target particle distribution parameter time series in the causal graph. The specific setting steps are as follows: Based on the degree to which the parameters in the predefined process parameter intervention instruction deviate from the normal operating range, the intervention intensity level is divided (e.g., weak, medium, strong); the number of directed paths from the intervened process parameter time series to the target particle distribution parameter time series in the causal graph is counted; combined with the actual impact ratio of the intervention effect under the same intensity and path structure in historical intervention experimental data, the proportion of the cumulative integral result of the intervention effect in the fusion is determined, and the remaining part is distributed... Assigning weights to the basic counterfactual distribution states to form a fusion weight; in the exemplary value range, the weight corresponding to the cumulative integral result of the intervention effect is between 0.2 and 0.6, and the weight corresponding to the basic counterfactual distribution state is between 0.8 and 0.4; the basis for the value is to ensure that the simulation results reflect both the physical impact of the intervention and the rationality of the initial counterfactual state; if the weights of some pre-effects are lower than 0.2, the particle distribution simulation results after the intervention will not be sensitive to process adjustments and will be difficult to reflect the intervention effect; if they are higher than 0.6, the intervention effect may be over-amplified, the constraints of the basic state will be ignored, and the simulation results will deviate from the actual physical process, resulting in unreasonable fluctuations or distortions.

[0118] S6. Calculate the difference between the simulated particle distribution results after intervention and the predicted values ​​of powder particle distribution, and generate suggestions for adjusting process parameters.

[0119] S6.1: Based on the multi-scale distribution difference measurement method, the difference between the simulated particle distribution results after intervention and the predicted value of powder particle distribution is quantified to obtain multi-dimensional difference measurement results;

[0120] Specifically, based on the multi-scale distribution difference measurement method, the simulated particle distribution results after intervention are compared with the predicted powder particle distribution values ​​at multiple particle size scales, including the overall distribution morphology, key particle size indicators (such as D10, D50, D90), and the proportion of particles in different particle size ranges. At each scale, the degree of deviation between the simulated particle distribution results after intervention and the predicted powder particle distribution values ​​is determined based on the distance measurement between probability distributions. The deviations at each scale are combined to form a multi-dimensional difference measurement result covering morphology, location, and local details.

[0121] S6.2: Input the multi-dimensional difference measurement results into the causal constraint multi-objective optimization framework and output the process parameter adjustment strategy;

[0122] Specifically, the results of multi-dimensional difference measurement are used as the basis for optimization objectives and fed into a causal-constrained multi-objective optimization framework. The multi-objective optimization framework uses the adjustable range of the process parameter time series as the decision variable space and sets constraints based on the causal influence direction of each process parameter time series on the target particle distribution parameter time series in the causal diagram. Under the premise of satisfying causal logic, multiple objectives are pursued simultaneously to reduce the differences between the particle distribution simulation results after intervention and the predicted values ​​of powder particle distribution in terms of overall morphology, key particle size indicators, and local intervals. By searching within the feasible solution space for parameter combinations that satisfy all causal constraints and minimize the comprehensive multi-dimensional differences, a process parameter adjustment strategy is formed.

[0123] It should be noted that causal logic refers to the directed dependencies between variables expressed in a causal graph and the underlying structural causal mechanism. Specifically, this includes: a process parameter time series can only directly affect its child nodes in the causal graph; the influence on non-descendant nodes must be transmitted through intermediate variables; any two parameter time series that are determined to be independent of each other given certain variables by the conditional independence rule in the causal graph must not be directly associated when adjusting or inferring; all interventions and propagation paths must strictly follow the direction of the causal edges, and there must be no false influences in the opposite causal direction.

[0124] S6.3: Verify the process parameter adjustment strategy through a reinforcement learning mechanism and generate executable process parameter adjustment suggestions.

[0125] Specifically, the process parameter adjustment strategy is applied as an action sequence to the production environment simulated by the powder particle distribution prediction model. The impact of the process parameter adjustment strategy on the time series of target particle distribution parameters in multiple time steps is observed through a reinforcement learning mechanism. The results of multi-dimensional difference measurement are used as feedback signals to evaluate the effectiveness and stability of the process parameter adjustment strategy. If the process parameter adjustment strategy continuously reduces the difference in multiple simulations without violating the process safety boundary, it is retained and marked as a feasible solution. The process parameter adjustment strategy with the best performance and compliance with the operating specifications is selected from the feasible solutions to form an executable process parameter adjustment suggestion.

[0126] This embodiment also provides a computer device applicable to the machine learning-based powder particle distribution prediction method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based powder particle distribution prediction method proposed in the above embodiment.

[0127] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0128] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the powder particle distribution prediction method based on machine learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0129] In summary, this invention achieves precise quantitative simulation of the effects of process intervention through forward propagation calculation of causal graphs, improving the interpretability and reliability of prediction results; it transforms the structural information of the causal graph into a computable propagation path, captures the cumulative impact of intervention effects through multi-path integration methods, and generates particle distribution simulation results that conform to physical laws. This accurately reflects the dynamic impact of process parameter adjustments on particle distribution, improves the accuracy of intervention effect prediction under complex operating conditions, and provides a reliable theoretical basis and decision support for process optimization.

[0130] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting powder particle distribution based on machine learning, characterized in that: include, Collect multi-source time series data related to particle distribution, use causal discovery algorithms to learn the causal dependencies between parameters in the multi-source time series data, and generate a causal graph. Using causal graphs as the guiding principle for data coordination, a game-theoretic coordination mechanism is employed to perform dynamic weight allocation for multi-source time series data, resulting in a coordinated data vector. Construct a powder particle distribution prediction model, input the coordinated data vector into the powder particle distribution prediction model, and output the predicted powder particle distribution value at future time. Counterfactual reasoning is performed on predefined process parameter intervention commands and future time-to-time powder particle distribution predictions to generate a basic counterfactual distribution state; By coupling the structural causal mechanism contained in the causal graph, the forward propagation calculation of the intervention effect on the basic counterfactual distribution state is performed to generate the particle distribution simulation results after intervention. The difference between the simulated particle distribution after intervention and the predicted particle distribution of powder is calculated, and suggestions for adjusting process parameters are generated.

2. The powder particle distribution prediction method based on machine learning as described in claim 1, characterized in that: The steps for generating the cause-effect graph are as follows: The multi-source time series data includes time series of process parameters, time series of online measurement parameters, and time series of target particle distribution parameters; The time series of process parameters, online measurement parameters, and target particle distribution parameters are constructed into a three-dimensional tensor, and the time series of each parameter in the three-dimensional tensor are denoised to obtain denoised time series data. Based on the denoised time series data, the conditional mutual information between the time series of each parameter is determined, and causal relationships are constructed. The PC algorithm is used to verify and optimize causal relationships, and to generate candidate causal graph structures. The confidence level of causal edges in the candidate causal graph structure is evaluated, and causal edges with a confidence level below a preset threshold are removed to generate the causal graph.

3. The powder particle distribution prediction method based on machine learning as described in claim 2, characterized in that: The steps to obtain the coordinated data vector are as follows: The time series of each parameter corresponding to the node in the causal graph are constructed into a cooperative game model, and the Shapley value of each parameter time series is calculated as the initial weight based on the cooperative game model. Calculate the real-time data quality indicators of the time series of process parameters, online measurement parameters, and target particle distribution parameters, and combine them with the initial weights to generate dynamic weights; Based on the cooperative game model, the dynamic weights are iteratively optimized through multiple rounds of negotiation until convergence, generating optimized weights. The optimized weights are then weighted and fused with multi-source time series data to generate a coordinated data vector.

4. The powder particle distribution prediction method based on machine learning as described in claim 1, characterized in that: The steps for constructing the powder particle distribution prediction model are as follows: A temporal feature extraction network is constructed using an LSTM network and an attention mechanism; A continuous-time prediction architecture is established based on a temporal feature extraction network. The continuous-time prediction architecture is pre-trained using historical particle distribution data to complete the construction of a powder particle distribution prediction model.

5. The powder particle distribution prediction method based on machine learning as described in claim 4, characterized in that: The steps for outputting the predicted powder particle distribution at future times are as follows: The coordinated data vector is input into the temporal feature extraction network in the powder particle distribution prediction model to extract temporal features; The temporal features are input into the continuous-time prediction architecture of the powder particle distribution prediction model for numerical solution, and the predicted values ​​of powder particle distribution at future times are output.

6. The powder particle distribution prediction method based on machine learning as described in claim 5, characterized in that: The steps for generating the basic counterfactual distribution state are as follows: Based on the causal graph and predefined process parameter intervention instructions, an intervention-aware causal state encoder is constructed to perform counterfactual state encoding on the predicted value of powder particle distribution at future time, generating a state code under causal constraints. The state code under causal constraints is input into a causal evolutioner based on the divine ordinary differential equation for multi-step forward state propagation to generate a counterfactual state evolution trajectory. The evolution trajectory of the counterfactual state is smoothed over time to output the basic counterfactual distribution state.

7. The powder particle distribution prediction method based on machine learning as described in claim 6, characterized in that: The steps for generating the simulated particle distribution results after intervention are as follows: An intervention propagation network is constructed based on a causal graph. The basic counterfactual distribution state is input into the intervention propagation network to propagate the intervention effect in a structured manner and obtain dynamic propagation results. The dynamic propagation results are input into the framework of the neural ordinary differential equation to perform multi-path intervention effect integral analysis, and the cumulative integral results of the intervention effect are output. The cumulative integral results are weighted and fused with the basic counterfactual distribution state to generate the particle distribution simulation results after intervention.

8. The powder particle distribution prediction method based on machine learning as described in claim 7, characterized in that: The proposed adjustments to the generation process parameters follow these steps: Based on the multi-scale distribution difference measurement method, the difference between the simulated particle distribution results after intervention and the predicted value of powder particle distribution is quantified to obtain multi-dimensional difference measurement results. The multi-dimensional difference measurement results are input into a causal-constrained multi-objective optimization framework to output a process parameter adjustment strategy. The process parameter adjustment strategy is validated by using a reinforcement learning mechanism, and executable process parameter adjustment suggestions are generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based powder particle distribution prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based powder particle distribution prediction method according to any one of claims 1 to 8.

Images