Process optimization method based on MES data analysis

Abstract

The invention provides a process optimization method based on MES data analysis, and the method comprises the steps: obtaining historical optimization data and current dynamic factor information from a manufacturing execution system, and generating a comprehensive data set; performing grouping processing on historical optimization tracks according to the comprehensive data set, and determining a trend change mode; generating a pre-judged optimization direction vector according to the trend change mode and the current production plan; generating an optimized parameter adjustment track according to the initial solution set distribution and the dynamic factor information; according to the parameter adjustment track and the real-time production data, whether the trend change mode is updated or not is judged; training a model to generate an enhanced pre-judgment accuracy index according to the updated trend change mode; and determining a final process parameter combination according to the pre-judgment accuracy index and the updated solution set distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a process optimization method based on MES data analysis. BACKGROUND

[0002] Manufacturing is the core field driving industrial development, and its process optimization is directly related to production efficiency, product quality and enterprise competitiveness. Under the background of intelligent manufacturing, the manufacturing execution system (MES) provides important support for process optimization by collecting and storing production data in real time.

[0003] However, the existing method often faces the problem of data complexity and dynamics when using MES data for process optimization. The current solution usually relies on fixed rules or single model analysis of historical data, which is difficult to adapt to the complex scenarios of multi-variable and high-dynamics in production environment, resulting in inaccurate optimization results and difficulty in meeting actual production needs.

[0004] In the process optimization process, the core challenge is how to effectively mine the rules in the historical optimization data accumulated in the MES system and apply them to the dynamic production scene. Historical optimization data contains optimization trajectory and solution set distribution information under different process parameter combinations, and these data hide the trend change patterns of process optimization. However, due to the involvement of production planning, equipment status, raw material batch and other dynamic factors in the production environment, the complexity of data makes it difficult for traditional methods to accurately identify the trend rules. For example, in a certain automobile parts production line, process parameter adjustment needs to consider equipment wear, raw material batch difference and production rhythm change, but existing methods often cannot integrate these factors to predict the optimization direction, resulting in parameter adjustment lag, affecting production efficiency and product quality. In addition, the lack of dynamic analysis of optimization trajectory makes it difficult to quickly converge to high-quality solutions in multi-objective optimization, and the optimization efficiency is low.

[0005] Therefore, how to mine the trend change patterns based on MES historical optimization data, and combine dynamic factors such as production planning, equipment status and raw material batch to accurately predict the optimization direction, and provide heuristic guidance for multi-objective evolutionary algorithm, has become a key problem in process optimization. SUMMARY

[0006] The present application provides a process optimization method based on MES data analysis, mainly including:

[0007] The historical optimization data and current dynamic factor information are acquired from a manufacturing execution system to generate a comprehensive data set; the historical optimization trajectories are grouped and processed according to the comprehensive data set to determine a trend change mode; an optimized direction vector is generated according to the trend change mode and a current production plan; a population is initialized according to the optimized direction vector and the trend change mode is integrated to generate an initial solution set distribution; an optimized parameter adjustment trajectory is generated according to the initial solution set distribution and the dynamic factor information; whether the trend change mode is updated is judged according to the parameter adjustment trajectory and real-time production data; an enhanced prediction accuracy index is generated by training a model according to the updated trend change mode; a final process parameter combination is determined according to the prediction accuracy index and the updated solution set distribution. Further, the historical optimization data containing process parameter combination optimization trajectories are acquired from a manufacturing execution system; current dynamic factor information containing production plans, equipment states and raw material batch details is acquired; a comprehensive data set containing process parameter combination optimization trajectories, production plans, equipment states and raw material batch details is integrated and generated according to the historical optimization data and the dynamic factor information; a standardized data set format is generated by preprocessing the comprehensive data set. Further, the historical optimization trajectories in the comprehensive data set are grouped and processed by a clustering algorithm; feature data of each group of historical optimization trajectories is extracted according to the grouping result; the trend change mode of each group of historical optimization trajectories is analyzed according to the feature data; mode parameters representing the change law of similar trajectories are generated according to the trend change mode. Further, whether the trend change mode matches the current production plan is judged; if matching, the equipment state and raw material batch variables in the current production plan are acquired; the equipment state and raw material batch variables are input into a pre-trained neural network model; an optimized direction vector is generated according to the output of the neural network model. Further, a population is initialized according to the optimized direction vector by a genetic algorithm; the trend change mode is integrated into the genetic algorithm as a heuristic constraint; an initial solution set distribution is generated by multi-objective optimization of the genetic algorithm; distribution characteristic parameters of the initial solution set are generated by evaluating the initial solution set distribution.Further, the generating an optimized parameter adjustment trajectory according to the initial solution set distribution and the dynamic factor information comprises: judging whether the initial solution set distribution deviates from the trend change mode; if deviating, adjusting the crossover and mutation operators of the genetic algorithm according to the dynamic factor information; generating the optimized parameter adjustment trajectory through iterative optimization of the adjusted genetic algorithm; and verifying the parameter adjustment trajectory to generate a trajectory optimization result. Further, the judging whether to update the trend change mode according to the parameter adjustment trajectory and real-time production data comprises: obtaining real-time production data feedback; comparing the real-time production data with the parameter adjustment trajectory and the optimization direction vector; judging whether to update the trend change mode according to a comparison result; and if updating is needed, generating new trend change mode parameters according to the real-time production data. Further, the generating an enhanced prediction accuracy index by training a model according to the updated trend change mode comprises: generating a training data set according to the updated trend change mode; updating model weights by retraining a neural network model through the training data set; generating a prediction result according to the updated neural network model; and generating an enhanced prediction accuracy index according to a comparison between the prediction result and actual production data.

[0008] Further, the determining a final process parameter combination according to the prediction accuracy index and the updated solution set distribution comprises: comparing the prediction accuracy index with a preset threshold value, if the prediction accuracy index exceeds the preset threshold value, entering a next selection process, otherwise returning to iteratively updating the trend change mode; selecting a high-quality solution from the updated solution set distribution, and selecting a solution with the highest fitness from the high-quality solution as the final process parameter combination.

[0009] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0010] The application discloses a process optimization method based on MES data analysis, aiming at the problems of lack of self-adaptability and prediction accuracy of process parameter optimization in a manufacturing execution system, integrating historical optimization trajectories and current production plans, equipment states and raw material batch information to generate a comprehensive data set, and extracting a trend change mode by using a clustering algorithm, combining a neural network and a genetic algorithm to generate a prediction optimization direction vector and an initial solution set distribution, and dynamically adjusting a parameter trajectory. The application updates the trend change mode through real-time production data feedback, re-trains a neural network to improve prediction accuracy, and finally outputs an optimal process parameter combination. The application solves the problem of insufficient adaptability of a traditional optimization method to a dynamic production environment, significantly improves the intelligent level and production efficiency of process parameter optimization, and realizes self-adaptive and accurate manufacturing process optimization. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1A flowchart of a process optimization method based on MES data analysis of the present application. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only some of the embodiments of the present application.

[0013] With reference to Figure 1 , the process optimization method based on MES data analysis of the present embodiment can specifically include:

[0014] Step S1, obtaining historical optimization data and current dynamic factor information from a manufacturing execution system, generating a comprehensive data set, including: obtaining historical optimization data containing process parameter combination optimization trajectories from the manufacturing execution system; obtaining current dynamic factor information containing production plans, equipment states and raw material batch details; integrating to generate a comprehensive data set containing process parameter combination optimization trajectories, production plans, equipment states and raw material batch details according to the historical optimization data and the dynamic factor information; preprocessing the comprehensive data set to generate a standardized data set format.

[0015] In an embodiment, obtaining historical optimization data and current dynamic factor information from a manufacturing execution system, generating a comprehensive data set, includes the following steps:

[0016] Step S101, obtaining historical optimization data containing process parameter combination optimization trajectories from a manufacturing execution system.

[0017] Specifically, the historical database is queried through a system interface to extract the process parameter change sequence recorded in the past production cycle, such as the optimization path data of temperature, pressure and time combination, which reflects the trajectory evolution of the past optimization process.

[0018] According to the extracted historical optimization data, preliminary filtering is performed to remove outliers to ensure data reliability.

[0019] Step S102, obtaining current dynamic factor information containing production plans, equipment states and raw material batch details.

[0020] Specifically, the scheduling details of the current production plan, the running state of the equipment such as load and maintenance record, and the specific attributes of the raw material batch such as batch number and quality index are read from the real-time monitoring module.

[0021] During the acquisition process, the timestamps of the dynamic factors are updated synchronously to match the time sequence of the historical data.

[0022] Step S103, according to the history optimization data and the dynamic factor information, integrated to generate a comprehensive data set containing process parameter combination optimization trajectory, production plan, equipment state and raw material batch details.

[0023] Specifically, the historical optimization trajectory is associated with the current dynamic factors, for example, by key-value matching, the production plan is linked to similar trajectory paragraphs, forming a multi-dimensional data set, where each data point contains trajectory coordinates, production plan labels, equipment state vectors and raw material batch attributes.

[0024] Identify the key nodes in the historical trajectory, and inject the current equipment state as a weight factor to calculate the integrated similarity score to optimize the coherence of the data set.

[0025] Using data fusion algorithms such as weighted average method, the dynamic factors are embedded in the trajectory data, for example, the influence coefficient is applied to the equipment state variable, to ensure that the comprehensive data set captures the potential impact of real-time changes on the optimization trajectory.

[0026] Verify the integrity of the integrated data set by checking the proportion of missing values, and if it exceeds the threshold, backtrack to supplement the data.

[0027] Step S104, pre-processing the comprehensive data set to generate a standardized data set format.

[0028] Specifically, normalization techniques are applied to scale numerical features to a uniform range, for example, using min-max normalization to process process parameter values, while categorical features such as raw material batches are one-hot encoded, outputting a matrix-formatted data set for subsequent clustering or neural network input.

[0029] The selection of pre-processing parameters is based on data set statistics, such as calculating the mean and standard deviation to adjust the normalization boundaries.

[0030] For example, in the electronic component manufacturing scenario, the historical optimization data obtained in step S101 may include a trajectory change of soldering temperature from 150 degrees to 180 degrees, reflecting the path of yield optimization; the integration of step S103 combines this trajectory with the current production plan such as a daily target of 1000 pieces, the equipment state such as 80% machine load and the raw material batch purity of 95%, generating a comprehensive data set that can improve optimization accuracy by up to 15%.

[0031] In one embodiment, for the integration process of step S103, it can be understood that through multi-source data alignment, the seamless integration of historical trajectory and dynamic factors is ensured, for example, in chemical production, the integrated trajectory data can reduce the waste rate of 5% caused by parameter deviation.

[0032] For example, the preprocessing of step S104 can further include anomaly detection, such as using the boxplot method to remove points that are more than 1.5 times the interquartile range, after generating the standardized dataset, the robustness of the subsequent algorithm is improved by 10%.

[0033] In an embodiment, when calculating the similarity score in step S103, the Euclidean distance formula is used to measure the proximity of the trajectory points to the dynamic factors, for example, a distance less than 0.1 is considered a high match, and the beneficial effect is to enhance the prediction relevance of the dataset.

[0034] For example, for the extension of electronic manufacturing, the trajectory data extracted in step S101 covers multiple batches, and the raw material details obtained in step S102 include supplier variations. The standardized dataset generated after integration can be directly input into the clustering algorithm, and the trajectory grouping efficiency is improved by 20%.

[0035] Step S2, grouping processing of historical optimization trajectories according to the comprehensive dataset, determining the trend change mode, comprising: using a clustering algorithm to group process the historical optimization trajectories in the comprehensive dataset; extracting feature data of each group of historical optimization trajectories according to the grouping result; analyzing the trend change mode of each group of historical optimization trajectories according to the feature data; generating mode parameters representing the change rule of similar trajectories according to the trend change mode.

[0036] In an embodiment, the clustering algorithm is used to group process the historical optimization trajectories in the comprehensive dataset, specifically including, step S201, obtaining the historical optimization trajectories in the comprehensive dataset, these trajectories contain the optimization path of process parameter combination and the corresponding production plan equipment state and raw material batch details.

[0037] From the historical optimization data extracted from the MES system, filter out trajectory sequences similar to the current dynamic factors, such as the combined change path of temperature, pressure and time parameters.

[0038] Group these trajectory sequences using the K-means clustering algorithm, which assigns similar trajectories to the same group by iteratively calculating the Euclidean distance of data points to cluster centers until the cluster centers are stable, where K is pre-set to 5 to 10 according to the number of trajectories to cover different optimization scenarios.

[0039] In an embodiment, according to the grouping result, the feature data of each group of historical optimization trajectories is extracted, specifically including, step S202, calculating the statistical features of the trajectories from each clustering group, such as the average change rate and standard deviation, these feature data reflect the fluctuation amplitude of the trajectories in the parameter adjustment process.

[0040] For each trajectory within a group, calculate the vector difference of the parameter combination from the initial value to the optimized value as part of the feature data.

[0041] Extracting common features of trajectories within a group, such as average values of device state variables like runtime and raw material batch variables like purity level, ensures that the feature data matches the production plan.

[0042] In an embodiment, analyzing the trend change pattern of each group of historical optimized trajectories according to the feature data, specifically including, step S203, fitting a trend line to the feature data of each group using a linear regression method, which is a best-fit straight line of the feature data points solved by least squares method, obtaining a slope and an intercept representing the trend change pattern.

[0043] Performing time series analysis on the rate of change of feature data within a group to identify an upward, downward, or stable trend, for example, in a temperature parameter trajectory, if the average rate of change is greater than 0, the pattern is optimized in the direction of high value.

[0044] Comparing the trend differences between groups to determine the pattern that matches the current production plan, for example, if the current plan involves high-purity raw materials, select the corresponding group's pattern, where the pattern includes an acceleration or deceleration phase of parameter adjustment.

[0045] Verifying the stability of the trend pattern by calculating the variance of the trajectories within the group, if the variance is less than 0.05, the pattern is confirmed to be reliable, this analysis process ensures that the trend change pattern can reflect the similar laws of historical optimized trajectories, which is beneficial to improve the accuracy of predicting the optimization direction, for example, in actual processes, this pattern analysis can reduce invalid parameter adjustments and shorten the optimization period.

[0046] In an embodiment, generating pattern parameters representing the change rules of similar trajectories according to the trend change pattern, specifically including, step S204, combining the slope, intercept, and variance values of the trend change pattern into a vector form of pattern parameters, these parameters represent the rules of the trajectory from the initial state to the optimized state, for example, the pattern parameter vector is [slope value, intercept value, variance value].

[0047] Normalizing the pattern parameters to have a range of 0 to 1 to facilitate subsequent neural network input, for example, processing the slope value through the min-max normalization formula.

[0048] Storing the generated pattern parameters in the database, associated with the grouping results, forming a queryable rule representation, for example, in different raw material batch scenarios, the pattern parameters can be used to quickly match similar trajectories.

[0049] In an embodiment, for the trend change pattern analysis of step S203, in process parameter optimization, for example, a temperature control trajectory, if the feature data shows an average rate of change of 0.2 per minute, the fitted linear regression slope represents an upward trend pattern, which is beneficial for predicting the parameter adjustment direction under the device state and reducing energy consumption.

[0050] In one embodiment, regarding the generation of model parameters in step S204, if the variance in the high-pressure parameter trajectory group is 0.03, the generated model parameter vector [0.15, 1.2, 0.03] represents the stable change law. This parameter can support the initial solution set of the genetic algorithm and effectively improve the optimization efficiency.

[0051] In one embodiment, steps S203 and S204 are used in combination. For example, in a scenario where the purity of the raw material batch is 95%, the trend pattern obtained by analysis is a parameter-increasing type. The generated pattern parameters are used for iterative updates to ensure the continuity of closed-loop optimization cycle data, which is beneficial for real-time applications in the MES system.

[0052] Step S3: Generate a predicted optimization direction vector based on the trend change pattern and the current production plan, including: determining whether the trend change pattern matches the current production plan; if it matches, obtaining the equipment status and raw material batch variables in the current production plan; inputting the equipment status and raw material batch variables through a pre-trained neural network model; and generating a predicted optimization direction vector based on the output of the neural network model.

[0053] In one embodiment, generating a predicted optimization direction vector based on the trend change pattern and the current production plan includes the following steps:

[0054] Step S301: Determine whether the trend change pattern matches the current production plan.

[0055] Key features in trend change patterns are extracted, such as the trajectory change rate of process parameter combinations and the similarity of dynamic factors in production plans. The degree of matching is quantified by calculating Euclidean distance, which is defined as the square root of the sum of the squares of the differences in each dimension between two vectors. If the distance is less than a preset threshold, it is considered a match.

[0056] If the calculated Euclidean distance is less than the threshold, a match is confirmed; otherwise, the system switches to standby mode to handle mismatches.

[0057] For example, in chemical production scenarios, trend change patterns are derived from grouping historical optimization trajectories. The current production plan involves the status of specific equipment. If the rate of change of parameters in the pattern is similar to the rate of change of raw material batch details in the plan, the Euclidean distance can be used to quickly determine whether the historical pattern is applicable, thereby improving prediction efficiency and avoiding resource waste caused by invalid calculations.

[0058] In one embodiment, the judgment process in step S301 can improve the response speed of the overall optimization loop because, through quantitative matching, irrelevant historical data can be directly filtered out, ensuring that subsequent steps are based on a reliable foundation.

[0059] Step S302, if matched, obtain the equipment state and raw material batch variable in the current production plan.

[0060] For example, the equipment state includes the running time and maintenance record, and the raw material batch variable covers the batch purity and supply time, which are directly extracted from the MES system to form a standardized input vector.

[0061] Step S303, input the equipment state and raw material batch variable through the pre-trained neural network model.

[0062] The neural network model is constructed, which adopts a multi-layer perceptron structure including an input layer, a hidden layer and an output layer, wherein the input layer receives the equipment state such as temperature and pressure value, and the raw material batch variable such as ingredient ratio, the hidden layer uses the ReLU activation function to process the nonlinear relationship, and the output layer generates the preliminary direction vector component.

[0063] The training process uses historical optimization data as samples to adjust the weights through the back propagation algorithm, which involves calculating the gradient of the loss function and updating the parameters to minimize the prediction error. For example, the loss function can use mean square error, which is defined as the square average of the difference between the actual optimization direction and the predicted direction.

[0064] The obtained equipment state and raw material batch variable are normalized and input into the model, which is realized by subtracting the mean value and dividing by the standard deviation to ensure that the input data is within the range of 0 to 1.

[0065] For example, in metal processing production, if the equipment state shows high temperature operation and the raw material batch is high-purity alloy, the model input can capture the interaction between variables such as the influence of temperature on batch purity through the hidden layer, and output more accurate vector components through the ReLU function to avoid the problem of gradient disappearance, which is beneficial to handle the uncertainty brought by dynamic factors and improve the robustness of prediction.

[0066] In one embodiment, the neural network model training in step S303 is based on historical optimization trajectories and dynamic factors, which can be extended to infer the use of batch gradient descent to optimize weights. Batch gradient descent is a process of calculating the gradient and updating the parameters of all training samples at once, which helps to efficiently converge on large data sets such as MES historical data, bringing higher prediction accuracy.

[0067] For example, from multiple perspectives, if the production scene involves different raw material batches, the model can learn the influence of batch variables on the optimization direction, such as when the batch purity is high, the direction vector tends to reduce energy consumption; if the equipment state is after maintenance, the direction vector tends to increase production, which ensures that the model has strong adaptability and is beneficial to the stability of the closed-loop optimization cycle.

[0068] Step S304, according to the output of the neural network model, a pre-judgment optimization direction vector is generated.

[0069] The model output is post-processed, such as vector normalization and threshold filtering, to ensure that each component represents the relative intensity of the optimization direction.

[0070] The processed components are combined to form a complete vector, which is used for subsequent genetic algorithm initialization.

[0071] For example, in pharmaceutical production, the model output may be a three-dimensional vector representing the adjustment direction of temperature, pressure and time parameters. After post-processing to filter out noise components, the generated direction vector can directly guide the parameter optimization trajectory, which is beneficial to reduce the number of iterations and improve production efficiency.

[0072] In one embodiment, the generation process of step S304 is closely related to the goal of generating a pre-judgment optimization direction vector. By comparing with the optimization direction vector, it is ensured that the generated vector integrates the trend pattern as a constraint, thereby forming a closed-loop process optimization cycle data.

[0073] Step S4, initializing a population according to the optimization direction vector and integrating the trend change pattern to generate an initial solution set distribution, including: initializing a population according to the optimization direction vector using a genetic algorithm; integrating the trend change pattern as a heuristic constraint into the genetic algorithm; generating an initial solution set distribution by performing multi-objective optimization using the genetic algorithm; and evaluating the initial solution set distribution to generate distribution characteristic parameters of the initial solution set.

[0074] In one embodiment, step S401, a population is initialized according to the optimization direction vector using a genetic algorithm, specifically including calculating the initial position coordinates of each individual in the population according to the vector component values in the optimization direction vector, wherein the optimization direction vector is obtained by pre-judgment using a neural network model and contains the adjustment direction and amplitude of the process parameters.

[0075] The vector components are mapped into the solution space of the genetic algorithm to generate chromosome encoding of the initial population.

[0076] Specifically, this embodiment can be applied in the process optimization scenario in the MES system, and the optimization direction vector obtained from historical data is used to guide the population initialization.

[0077] For example, in a chemical production batch, the optimization direction vector represents the pre-judged adjustment direction of temperature and pressure parameters, such as a vector value of (0.8, -0.5), indicating that the temperature needs to be increased by 0.8 units and the pressure needs to be decreased by 0.5 units, so that when initializing the population, individual codes close to this direction are generated first. This initialization method can make the population deviate from the historical trend from the beginning, avoid the search inefficiency caused by random initialization, and is beneficial to quickly converge to a high-quality solution set.

[0078] In one embodiment, step S402, the trend change pattern is integrated into the genetic algorithm as a heuristic constraint, specifically including extracting key trajectory points in the trend change pattern, such as inflection points and slope values of parameter changes, and converting them into constraint conditions of the fitness function of the genetic algorithm, wherein the trend change pattern is a similar pattern obtained from the clustering algorithm grouping processing of historical optimization trajectories.

[0079] According to the constraint condition, the selection probability in the population evolution process is adjusted, so that individuals deviating from the trend pattern are reduced in fitness.

[0002] The pattern constraint is integrated into the crossover operation to ensure that the offspring individuals inherit the characteristics of the trend trajectory.

[0081] Specifically, this embodiment is applicable to the scenario of dynamic changes of equipment states and raw material batches in production planning. By taking the trend pattern as a heuristic constraint, the genetic algorithm can be guided to avoid invalid search paths.

[0082] For example, in a metal processing process, the trend change pattern shows that the temperature parameter presents a linear rising trajectory under a certain equipment state, and the constraint will punish individuals with sharp fluctuations in temperature, so that the evolution process is more consistent with historical optimization experience. This integration improves the adaptability of the algorithm to dynamic factors, reduces the number of iterations, and improves the stability of the optimization trajectory.

[0083] In one possible implementation, for different raw material batches, the trend pattern can be further subdivided into sub-patterns, such as a pattern for high batch purity emphasizing slow adjustment and a pattern for low batch purity allowing greater variation, so as to dynamically switch constraints in the genetic algorithm, enhance prediction accuracy, and reduce production deviation.

[0084] Step S403, performing multi-objective optimization by the genetic algorithm to generate an initial solution set distribution, specifically including using the initialized population and the integrated constraint to perform multi-generation evolution, and the optimization objectives include minimizing production cost and maximizing yield, wherein the multi-objective optimization generates a non-dominated solution set as the distribution using the Pareto frontier method.

[0085] Specifically, this step directly uses the aforementioned initial population and constraint for iteration to generate a distribution for subsequent evaluation.

[0086] Step S404, evaluating the initial solution set distribution to generate distribution characteristic parameters of the initial solution set, specifically including calculating center position, dispersion, and Pareto frontier coverage rate of the solution set distribution, wherein the distribution characteristic parameters are used to judge whether the distribution deviates from the historical trend pattern.

[0087] If the distribution center deviates from the optimization direction vector beyond the threshold, it is marked as a parameter that needs to be adjusted.

[0088] Specifically, this embodiment is applied in process parameter combination optimization, and by evaluating the generated characteristic parameters, such as a dispersion of 0.2, it is beneficial to subsequent mutation adjustment of the genetic algorithm in the solution set.

[0089] For example, in an electronic component production scenario, if the evaluation shows that the distribution coverage rate reaches 85%, the characteristic parameters can be directly input to the adjustment step to ensure that the optimization trajectory matches the current production plan. This evaluation process forms a closed loop from distribution generation to parameter feedback, improving the robustness of overall process optimization.

[0090] Step S5, according to the initial solution set distribution and the dynamic factor information, generating an optimized parameter adjustment trajectory, including: judging whether the initial solution set distribution deviates from the trend change mode; if it deviates, adjusting the crossover and mutation operators of the genetic algorithm according to the dynamic factor information; generating an optimized parameter adjustment trajectory by iterative optimization of the adjusted genetic algorithm; verifying the parameter adjustment trajectory to generate a trajectory optimization result.

[0091] In one embodiment, judging whether the initial solution set distribution deviates from the trend change mode includes: step S501, determining the deviation degree by calculating the deviation value between the statistical indicators of the initial solution set distribution and the corresponding indicators of the trend change mode, wherein the deviation value is calculated based on the Euclidean distance, and if the deviation value exceeds the preset threshold, it is determined that it deviates.

[0092] Step S502, if it deviates, adjusting the crossover and mutation operators of the genetic algorithm according to the dynamic factor information, including: calculating the adjustment coefficient of the crossover probability according to the dynamic factor information such as the equipment state and the raw material batch details, wherein the initial value of the crossover probability is 0.8, and if the dynamic factor shows that the equipment load is high, the coefficient is multiplied by 1.2 to increase the diversity;

[0093] Adjusting the mutation operator by associating the mutation rate with the fluctuation amplitude of the dynamic factor, for example, the basic value of the mutation rate is 0.1, and if the raw material batch change rate exceeds 10%, the mutation rate is increased to 0.15, ensuring that the algorithm adapts to the current production environment.

[0094] For example, in process parameter optimization, assuming that the trend change mode shows that the temperature parameter in the historical trajectory tends to stabilize in the 200-220 degree Celsius interval, if the temperature solution of the initial solution set distribution deviates to 180-200 degrees Celsius, the average difference is calculated as a deviation value of 15 degrees Celsius, which exceeds the threshold of 10 degrees Celsius and is determined to deviate. At this time, according to the dynamic factor such as the increase of 5% in the moisture content of the raw material batch, adjusting the crossover and mutation operators can make the algorithm converge faster to the temperature combination that adapts to the new batch, thereby improving the optimization efficiency.

[0095] In an embodiment, step S503, the iterative optimization is performed by the adjusted genetic algorithm to generate the optimized parameter adjustment trajectory, including: applying the adjusted crossover mutation operator to the population iteration, wherein each generation iteration evaluates the fitness function based on the multi-objective of yield and quality, and after 50 iterations, the solution sequence on the Pareto optimal frontier is selected as the trajectory.

[0096] Specifically, this iterative optimization can incorporate the trend change pattern as a constraint, for example, preferentially selecting individuals close to the historical trajectory when initializing the population, and ensuring that the trajectory gradually converges to the pre-judgment optimization direction vector through the adjusted operator, which is beneficial to reducing invalid search and improving the stability of parameter adjustment.

[0097] For example, for the scenario where the device state is high load, the trajectory generated after iteration may show that the speed parameter is gradually adjusted from the initial 15 meters per minute to 18 meters per minute, and the trajectory length covers 10 adjustment points, each point verifies a 2% yield increase, which can support the accuracy of subsequent verification.

[0098] In an embodiment, step S504, the parameter adjustment trajectory is verified to generate a trajectory optimization result, including: comparing the matching degree of each parameter point in the trajectory with the real-time production data feedback, if the matching degree is higher than 90%, the trajectory is confirmed as the optimization result, otherwise, return to iterative optimization.

[0099] For example, in the verification, if the pressure parameter predicted by the trajectory is 500 Pa and the real-time data is 510 Pa, the matching degree is calculated as 98%, and the result is generated for the MES system application, which improves the prediction accuracy of the process closed loop.

[0100] Step S6, according to the parameter adjustment trajectory and real-time production data, judging whether to update the trend change pattern, including: obtaining real-time production data feedback; comparing the real-time production data with the parameter adjustment trajectory and the optimization direction vector; judging whether to update the trend change pattern according to the comparison result; if it needs to be updated, generating new trend change pattern parameters according to the real-time production data.

[0101] In an embodiment, the step of judging whether to update the trend change pattern according to the parameter adjustment trajectory and real-time production data includes the following steps:

[0102] Step S601, obtaining real-time production data feedback.

[0103] Specifically, through the MES system interface, real-time collection of key indicators in the production process is performed, such as device running state, raw material batch change and process parameter value, to form a feedback data set. This data set is directly used for subsequent comparison to ensure data timeliness.

[0104] Step S602, compare the real-time production data with the parameter adjustment trajectory and the optimization direction vector.

[0105] For example, after obtaining the feedback dataset, calculate the Euclidean distance between the real-time data and the parameter adjustment trajectory, and compare it with the cosine similarity of the optimization direction vector. The Euclidean distance is used to quantify the numerical deviation, and the cosine similarity measures the consistency of the direction. If the distance exceeds the preset threshold or the similarity is lower than the threshold, it is marked as significant deviation. This comparison process integrates the historical optimization trajectory, ensuring that dynamic factors such as equipment status are included in the evaluation.

[0106] Calculate the deviation index based on the comparison result, for example, by weighted average deviation value, where the weight is based on the importance of the raw material batch.

[0107] Generate a comparison report using the deviation index to determine the need for updating.

[0108] In one possible implementation, this step is applied to a chemical production scene, for example, in a polymerization reaction process, the real-time data includes temperature and pressure values, and the adjustment trajectory is compared. If the deviation causes yield fluctuations, this comparison identifies the trend pattern that needs to be optimized, thereby improving the prediction accuracy.

[0109] Step S603, determine whether to update the trend change pattern based on the comparison result.

[0110] Specifically, if the deviation index exceeds a preset threshold such as 0.1, it is determined that it needs to be updated; otherwise, the original pattern is maintained. The preset threshold is determined by historical data statistics to ensure the robustness of the judgment.

[0111] If the deviation index is within the threshold range, output a maintenance signal.

[0112] If it exceeds, trigger the update process.

[0113] For example, in electronic manufacturing, when the device state changes cause parameter drift, this judgment can respond in time to avoid production interruption and improve the efficiency of the overall optimization cycle.

[0114] Step S604, if it needs to be updated, generate a new trend change pattern parameter based on the real-time production data.

[0115] For example, use clustering algorithms to group real-time data and extract trend features such as slope and inflection point to form new parameters. The clustering algorithm is specifically the K-means algorithm, which first initializes the cluster center, then iteratively assigns data points and updates the center until convergence, thereby generating pattern parameters such as average change rate.

[0116] Preprocess the real-time data to remove noise by moving average filtering.

[0117] Apply K-means algorithm to group, calculate the trend parameters of each cluster, such as fitting the slope by linear regression.

[0118] Fusion history mode, generate final new parameters, ensure alignment with optimization direction vector.

[0119] In an embodiment, this step is applied in food processing scenarios, such as in the baking process, real-time data feedback temperature trajectory, if the deviation from the predicted vector is large, generate new mode parameters such as heating rate adjustment value 0.5 degrees / min, which can reduce energy waste and improve product quality consistency.

[0120] Specifically, in pharmaceutical production, the process of generating new parameters can be extended to multi-objective optimization, such as considering yield and energy consumption at the same time, the number of clusters of K-means algorithm is set to 3, corresponding to low, medium and high change modes respectively, the parameters include peak value and period length, in this way, the updated mode can better match the current dynamic factors, bring higher prediction accuracy and production stability.

[0121] For example, on the other side, when the raw material batch varies, the generated parameters emphasize the weight of batch details, the calculation result shows that the new mode slope increases from 0.2 to 0.3, which supports the subsequent initialization of genetic algorithm and improves the adaptability of the optimized trajectory.

[0122] Step S7, according to the updated trend change mode, train the model to generate enhanced prediction accuracy index, including: generating training data set according to the updated trend change mode; retrain neural network model through the training data set, update model weight; generate prediction result according to the updated neural network model; according to the comparison between the prediction result and the actual production data, generate enhanced prediction accuracy index.

[0123] In an embodiment, step S701, according to the updated trend change mode, generate training data set, specifically including, from the updated trend change mode, extract feature data containing process parameter combination optimization trajectory, production plan equipment state and raw material batch details, wherein these feature data are derived from historical optimization data in MES system and current dynamic factor information, and form similar trajectory trend after grouping processing by clustering algorithm.

[0124] Organize the extracted feature data into input-output pairs to form a training data set for subsequent model training.

[0125] Specifically, this embodiment can quickly generate a dataset in a production environment, for example, in a chemical production scenario, after updating the trend change pattern, the temperature, pressure, and other parameter combinations of the past optimization trajectory are directly extracted from the pattern as input, and the corresponding device state and raw material batch are taken as labels, thereby ensuring that the dataset reflects real dynamic changes and is beneficial to improving the adaptability of the model to similar production plans.

[0126] In one embodiment, step S702, the neural network model is retrained by the training dataset, and the model weights are updated, specifically including inputting the training dataset into the neural network model, wherein the neural network model adopts a multi-layer perceptron structure including an input layer, a hidden layer, and an output layer, the input layer receives the process parameter combination and the dynamic factor, the hidden layer processes the nonlinear relationship through an activation function such as ReLU, and the output layer generates a pre-judged optimization direction vector.

[0127] The loss function such as mean square error is calculated using the back propagation algorithm, and the model weights are iteratively adjusted based on the input-output pairs of the training dataset, wherein the weight value is optimized by gradient descent at each iteration until the loss function converges below a preset threshold.

[0128] The updated weights are saved to form a retrained model for subsequent pre-judgment.

[0129] Specifically, this embodiment is applicable to process optimization in the manufacturing field, for example, on an electronic product assembly line, when the trend change pattern is updated to a trajectory containing new device states, through retraining, the model weight adjustment can better capture the impact of raw material batch changes on parameters, which can improve the robustness of pre-judgment, reduce production deviations, and ensure that the weights reflect the latest trends through multiple iterations, which is beneficial to realizing a closed-loop optimization cycle.

[0130] For example, in one possible implementation, considering a specific production batch, the training dataset includes 100 historical trajectory samples, each sample has 10 process parameter dimensions, the retraining process sets the learning rate to 0.01, and after 50 iterations, the weight update improves the pre-judgment accuracy of the model for similar batches from 80% to 95%, which reflects the benefits of improving the generalization ability of the model through dynamic updates.

[0131] In one embodiment, step S703, a pre-judgment result is generated according to the updated neural network model, specifically including inputting the device state and raw material batch variables of the current production plan into the updated neural network model, and calculating the pre-judged optimization direction vector through forward propagation to output as the result.

[0132] Specifically, this step ensures that the pre-judgment result is directly derived from the retrained model, which is beneficial to quickly responding to production changes.

[0133] In an embodiment, step S704, according to the comparison of the pre-judgment result and the actual production data, an enhanced pre-judgment accuracy index is generated, specifically including obtaining real-time production data such as actual process parameter adjustment trajectory and optimization effect from the feedback from the MES system.

[0134] The similarity between the pre-judgment result and the actual data is calculated, for example, the cosine similarity formula is used to compare the optimization direction vector, wherein the cosine similarity is defined as the product of the dot product of two vectors divided by the product of their module lengths, and the value closer to 1 indicates higher similarity.

[0135] According to the similarity value, an accuracy index is generated, such as converting the similarity into a percentage form and comparing it with the historical index, if the improvement is more than 10%, it is marked as an enhanced pre-judgment accuracy index.

[0136] The generated index is recorded in the system for judging whether further iteration is needed to update the trend change mode.

[0137] Specifically, this embodiment plays a key role in production optimization, for example, in the food processing scene, after comparing the temperature adjustment vector predicted by the pre-judgment result with the actual data, the accuracy index is calculated to be 92%, which is enhanced compared with the previous mode of 85%, which helps to confirm the effectiveness of the model, and guides the genetic algorithm to select high-quality solutions through the index, and finally determines the process parameter combination, forming a closed-loop process optimization cycle data.

[0138] For example, in one possible implementation, for a metal processing batch, the pre-judgment result vector is [0.8, 0.6, 0.9], the actual data vector is [0.75, 0.65, 0.85], the cosine similarity is calculated to be about 0.98, and the generated index is 98%, which indicates that the accuracy is enhanced by more than 90% of the threshold, which is beneficial to reduce production waste and improve overall efficiency. Through this comparison, it can also reveal the influence of dynamic factors such as raw material batch on the trajectory, further supporting the iteration update of the trend mode.

[0139] Step S8, according to the pre-judgment accuracy index and the updated solution set distribution, determine the final process parameter combination, including: comparing the pre-judgment accuracy index with the preset threshold, if the pre-judgment accuracy index exceeds the preset threshold, then enter the next selection process, otherwise return to iterate and update the trend change mode; select high-quality solutions from the updated solution set distribution, and select the solution with the highest fitness from the high-quality solutions as the final process parameter combination.

[0140] In an embodiment, according to the pre-judgment accuracy index and the updated solution set distribution, the final process parameter combination is determined, including the following sub-steps:

[0141] Step S801, compare the pre-judgment accuracy index with the preset threshold.

[0142] The preset threshold is an average accuracy obtained by statistics after multiple iterations in the historical optimization trajectory, for example, set to 0.85, indicating that the matching degree of the model prediction and the actual optimization direction needs to reach 85% or more.

[0143] If the prediction accuracy index exceeds the preset threshold, the next selection process is entered, otherwise the trend change mode is returned to iteration and update.

[0144] Step S802, selecting a high-quality solution from the updated solution set distribution.

[0145] The updated solution set distribution is a parameter combination set obtained after genetic algorithm optimization, and each solution represents a process parameter combination, including temperature, pressure and time variables.

[0146] The selection of high-quality solutions is based on the fitness function, which calculates the multi-objective value of each solution, such as maximizing yield and minimizing energy consumption.

[0147] The specific calculation process of the fitness function is to calculate the weighted sum of the yield target value and the energy consumption target value for each solution, and the weighting coefficients are dynamically adjusted according to the production plan, for example, the yield weight is 0.6 and the energy consumption weight is 0.4.

[0148] Step S803, determining the final process parameter combination.

[0149] Select the solution with the highest fitness from the high-quality solutions as the final process parameter combination.

[0150] For example, in a chemical production scenario, if the solution set distribution contains multiple temperature-pressure combinations, select the one with the highest fitness, such as temperature 150 degrees and pressure 2.5 MPa.

[0151] In one embodiment, in step S802, a high-quality solution is selected from the updated solution set distribution, specifically, the genetic algorithm is a simulation of natural evolution optimization algorithm, which iterates the population through selection, crossover and mutation operations.

[0152] Here, the population of the genetic algorithm evolves from the initial solution set distribution, and each individual is encoded as a vector of process parameters, such as [temperature, pressure, time].

[0153] The selection operation uses the roulette method, and the probability is proportional to the fitness, ensuring that high-quality individuals are retained.

[0154] The crossover operation exchanges the parameter vectors of two parent individuals at a random point to generate offspring.

[0155] The mutation operation randomly changes the values in the vector with a small probability, such as increasing or decreasing the temperature value by 5 degrees.

[0156] By these operations, solutions with fitness exceeding the average value are screened from the updated solution set distribution as high-quality solutions.

[0157] The beneficial effect of this is that the global optimization capability of the parameter combination is improved, local optimum is avoided, and more stable process parameters are ensured in complex production environments.

[0158] For example, in the same field of semiconductor manufacturing, the solution set distribution may contain wafer processing time and chemical concentration combinations. After genetic algorithm iteration, the solution with the highest fitness is selected, which can reduce the defect rate by 10%.

[0159] In one embodiment, when comparing the prediction accuracy index with the preset threshold in step S801, if the index is exactly equal to the threshold, it can be considered as exceeding, so as to include the boundary case.

[0160] The calculation of the prediction accuracy index is based on the deviation between the output of the neural network model and the real-time feedback. The smaller the deviation, the higher the index.

[0161] The neural network model is a multi-layer perceptron, including an input layer, a hidden layer and an output layer. The input is the device state and the raw material batch variable, and the output is the optimization direction vector.

[0162] The training process uses the backpropagation algorithm to adjust the weights, and the loss function is the mean square error.

[0163] The updated solution set distribution incorporates the trend change pattern as a constraint to ensure that the solution set conforms to the historical trajectory.

[0164] The beneficial effect is that the robustness of the decision is improved, and in the case of large changes in raw material batches, the parameter combination can still be accurately determined.

[0165] For example, assuming the preset threshold is 0.85 and the calculated index is 0.87, then the solution is selected from the solution set distribution; if it is 0.83, then the iteration update mode is used.

[0166] In one embodiment, after determining the final process parameter combination in step S803, it can be further verified for its matching degree with the production plan.

[0167] The matching degree is obtained by calculating the similarity of the parameter combination and the plan target. If the similarity is high, it is directly applied.

[0168] This forms a closed-loop optimization that can continuously improve process efficiency.

[0169] For example, in pharmaceutical production, the final combination such as stirring speed 200 rpm and temperature 37 degrees can increase production by 15% and reduce energy consumption.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A process optimization method based on MES data analysis, characterized in that, The method comprises the following steps: obtaining historical optimization data and current dynamic factor information from a manufacturing execution system to generate a comprehensive data set; grouping the historical optimization trajectories according to the comprehensive data set to determine a trend change mode; generating a predicted optimization direction vector according to the trend change mode and the current production plan; initializing a population according to the optimization direction vector and incorporating the trend change mode to generate an initial solution set distribution; generating an optimized parameter adjustment trajectory according to the initial solution set distribution and the dynamic factor information; determining whether to update the trend change mode according to the parameter adjustment trajectory and real-time production data; training a model to generate an enhanced prediction accuracy index according to the updated trend change mode; determining a final process parameter combination according to the prediction accuracy index and the updated solution set distribution.

2. The method of claim 1, wherein, The method of obtaining historical optimization data and current dynamic factor information from a manufacturing execution system to generate a comprehensive data set comprises the following steps: obtaining historical optimization data containing process parameter combination optimization trajectories from a manufacturing execution system; obtaining current dynamic factor information containing production plans, equipment states, and raw material batch details; integrating the historical optimization data and the dynamic factor information to generate a comprehensive data set containing process parameter combination optimization trajectories, production plans, equipment states, and raw material batch details; and preprocessing the comprehensive data set to generate a standardized data set format.

3. The method of claim 1, wherein, The method of grouping the historical optimization trajectories according to the comprehensive data set to determine a trend change mode comprises the following steps: grouping the historical optimization trajectories in the comprehensive data set using a clustering algorithm; extracting feature data of each group of historical optimization trajectories according to the grouping results; analyzing the trend change mode of each group of historical optimization trajectories according to the feature data; and generating mode parameters representing the change rules of similar trajectories according to the trend change mode.

4. The method of claim 1, wherein, The method of generating a predicted optimization direction vector according to the trend change mode and the current production plan comprises the following steps: determining whether the trend change mode matches the current production plan; if it matches, obtaining equipment state and raw material batch variables in the current production plan; inputting the equipment state and raw material batch variables into a pre-trained neural network model; and generating a predicted optimization direction vector according to the output of the neural network model.

5. The method of claim 1, wherein, The method of initializing a population according to the optimization direction vector and incorporating the trend change mode to generate an initial solution set distribution comprises the following steps: initializing a population according to the optimization direction vector using a genetic algorithm; incorporating the trend change mode as a heuristic constraint into the genetic algorithm; performing multi-objective optimization through the genetic algorithm to generate an initial solution set distribution; and evaluating the initial solution set distribution to generate distribution characteristic parameters of the initial solution set.

6. The method of claim 1, wherein, The generating an optimized parameter adjustment trajectory according to the initial solution set distribution and the dynamic factor information comprises: judging whether the initial solution set distribution deviates from the trend change mode; if deviating, adjusting the crossover and mutation operators of the genetic algorithm according to the dynamic factor information; generating the optimized parameter adjustment trajectory through iterative optimization of the adjusted genetic algorithm; and verifying the parameter adjustment trajectory to generate a trajectory optimization result.

7. The method of claim 1, wherein, The judging whether to update the trend change mode according to the parameter adjustment trajectory and real-time production data comprises: obtaining real-time production data feedback; comparing the real-time production data with the parameter adjustment trajectory and the optimization direction vector; judging whether to update the trend change mode according to the comparison result; and if updating is needed, generating new trend change mode parameters according to the real-time production data.

8. The method of claim 1, wherein, The generating an enhanced prediction accuracy index by training a model according to the updated trend change mode comprises: generating a training data set according to the updated trend change mode; updating the model weight by retraining a neural network model through the training data set; generating a prediction result according to the updated neural network model; and generating an enhanced prediction accuracy index according to the comparison between the prediction result and actual production data.

9. The method of claim 1, wherein, The determining a final process parameter combination according to the prediction accuracy index and the updated solution set distribution comprises: comparing the prediction accuracy index with a preset threshold value, if the prediction accuracy index exceeds the preset threshold value, entering a next step selection process, otherwise returning to iteratively update the trend change mode; selecting a high-quality solution from the updated solution set distribution, and selecting a solution with the highest fitness from the high-quality solution as the final process parameter combination.

Images