An edge control layer and cloud cooperation-based process parameter optimization method

Abstract

The embodiment of the application discloses a process parameter optimization method based on edge control layer and cloud cooperation, relates to the technical industrial data analysis and process control field, and comprises the following steps: a cloud end receives a sample set; the cloud end constructs an agent model based on the samples in the set; the cloud end calculates candidate process parameters from a process parameter region and performs safety checking, and if the checking fails, the cloud end returns to recalculation, if the checking is passed, the cloud end issues the parameters to an edge control layer; the edge control layer controls a production line to produce by using the candidate process parameters and monitors a production state in real time; if an abnormality is monitored, a rollback is triggered, the cloud end generates a product quality observation value which is worse than a highest product quality value in a current set for the candidate process parameters, based on an updated set, the foregoing steps are repeatedly executed until a preset convergence condition is reached. The method can realize automatic searching for optimal parameters and safety checking.

Description

Technical Field

[0001] This invention relates to the field of industrial data analysis and process control technology, and in particular to a method for optimizing process parameters based on edge control layer and cloud collaboration. Background Technology

[0002] In polymer material processing such as plastic extrusion and injection molding, the setting of process parameters, such as barrel temperature, screw speed, injection pressure, and traction speed, directly affects product quality, including dimensional accuracy, appearance defects, mechanical properties, production efficiency, and production costs. Currently, most production lines adopt a fixed process parameter mode, where process parameters remain stable near set values. While this mode can ensure short-term production stability, it is often difficult to continuously adapt to actual production conditions due to dynamic interference factors such as strong coupling between process parameters, large batch fluctuations in raw materials, and changes in equipment operating status over time. Therefore, it is essential to dynamically adjust and optimize process parameters based on real-time conditions and external interference during production to ensure stable product quality, improve production efficiency, and effectively control production costs.

[0003] Traditional methods of adjusting process parameters based on manual experience typically require experienced engineers to spend several days manually debugging and repeatedly trying to find the optimal combination of process parameters. This approach is not only slow to respond and has low precision, but it is also greatly affected by differences in operator experience, easily leading to product quality fluctuations, raw material waste, and certain safety risks. While existing machine learning (ML) proxy model-based optimization technologies can automatically recommend process parameters, they mostly output optimization suggestions in an open-loop manner, lacking real-time safety monitoring and closed-loop verification. If the recommended process parameters are directly applied to the equipment, the system cannot intervene and correct them in time if the process parameters exceed safety boundaries or cause production fluctuations. Moreover, most related optimization algorithms require a large number of blind trial productions, resulting in high trial-and-error costs, an inability to quantify decision uncertainty, and high risks when exploring unknown process regions.

[0004] Meanwhile, existing solutions lack a prior knowledge integration mechanism, relying solely on historical data for optimization and ignoring the constraints of process principles on the range of process parameters. This purely data-driven approach lacks process cognition capabilities, has poor interpretability, and insufficient reliability. It cannot leverage expert experience to mitigate known process risks, and in complex operating conditions, it struggles to balance production safety, product quality, and process rationality, easily leading to product defects, raw material waste, and excessively high trial production costs, ultimately failing to meet actual processing needs.

[0005] Furthermore, when introducing new products or switching production lines, traditional parameter optimization methods often face a "cold start" dilemma due to the lack of historical production data. On the one hand, the lack of initial samples makes it impossible to build an effective surrogate model, hindering the optimization process from starting. On the other hand, if initial data is obtained through random trial and error or manual debugging, a large number of actual experiments are required, resulting in high trial-and-error costs and long cycles. In addition, existing methods typically rely on a single data source from the current production line for modeling, failing to effectively utilize production data accumulated from the same model of equipment, historical batches, or similar process scenarios. This leads to a waste of historical knowledge and further exacerbates the modeling difficulty and low optimization efficiency in data-scarce scenarios. Summary of the Invention

[0006] In view of this, the present invention provides a process parameter optimization method based on edge control layer and cloud collaboration, aiming to solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for optimizing process parameters based on edge control layer and cloud collaboration, the method comprising: Step S1: Receive a sample set from the cloud, where each sample contains process parameters and corresponding product quality values; Step S2: The cloud platform constructs a proxy model based on the samples in the set using a preset algorithm. The proxy model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters. Step S3: The cloud-based system calculates candidate process parameters from the process parameter region by maximizing the value of a preset acquisition function based on the predicted mean and the uncertainty measure output by the proxy model. Step S4: The cloud performs a security check on the candidate process parameters. If the check fails, it returns to step S3 to recalculate. If the check passes, the candidate process parameters are sent to the edge control layer. Step S5: The edge control layer controls the production line to use the candidate process parameters for production and monitors the production status in real time. If an anomaly is detected, a rollback is triggered, and the candidate process parameters and associated process data are uploaded to the cloud. The cloud generates a product quality observation value for the candidate process parameters that is inferior to the highest product quality value in the current set, and adds a new sample consisting of the candidate process parameters and the product quality observation value to the set. If production is completed normally, the edge control layer uploads the candidate process parameters and the measured product quality value to the cloud. The cloud adds a new sample consisting of the candidate process parameters and the measured product quality value to the set. Step S6: Based on the updated set, repeat steps S2 to S5 until the preset convergence condition is reached.

[0008] Optionally, the cloud-based proxy model, constructed using a preset algorithm based on samples in the set, includes: When constructing the proxy model using a preset algorithm, the cloud also incorporates prior knowledge. The method for incorporating prior knowledge includes at least one of the following: using the safety boundary of process parameters as a constraint condition for training the proxy model; generating samples containing virtual process parameters and virtual product quality values ​​based on physical formulas or expert experience rules, and adding the samples to the set.

[0009] Optionally, the security verification in step S4 includes at least one of the following verifications: Process parameter safety boundary verification, process parameter change step size verification, and simulation verification based on the process model.

[0010] Optionally, in step S2, the preset algorithm is a Gaussian process regression algorithm.

[0011] Optionally, in step S5, the acquisition function is a desired improvement function, and its expression is: ,in: EI(x) represents the expected improvement in process parameter x; μ(x) is the mean product quality predicted by the surrogate model at process parameter x; f(x+) is the highest known product quality value in the current set; ξ is a preset non-negative exploration parameter; σ(x) is the standard deviation of the prediction made by the surrogate model at point x, reflecting the uncertainty of the surrogate model at the process parameter x; Z is a standardized variable, and its calculation formula is: Z= ; When σ(x) = 0, the formula is defined as: EI(x) = 0; Φ(Z) is the cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

[0012] Optionally, the edge control layer includes a PID controller, which controls the adjustment of process parameters.

[0013] Optionally, the product quality observations are generated in at least one of the following ways: Set to a fixed value that is inferior to the highest product quality value in the current set; The value is set as the deteriorated value based on the highest product quality value in the current set, according to a preset ratio. The data is dynamically calculated and generated based on the anomaly type and severity represented by the associated process data.

[0014] Optionally, in step S6, after reaching a preset convergence condition, the cloud outputs the globally optimal process parameters that have been historically verified, and outputs the best surrogate model constructed based on the updated sample set; wherein the convergence condition includes at least one of the following: The maximum value of the preset acquisition function is lower than a set threshold within the process parameter range; The number of iterations has reached the preset limit.

[0015] Optionally, the sample set includes target task samples from the current target to be optimized, and source task samples from other targets; The preset algorithm is a multi-task Gaussian process regression algorithm, and the proxy model is a multi-task proxy model; Step S2 includes the following sub-steps: S21. When step S2 is executed for the first time, the cloud uses the source task samples to pre-train the multi-task Gaussian process and construct a multi-task proxy model. S22. The cloud updates the multi-task agent model based on the target task sample. In subsequent iterations, step S22 is only repeated, and step S21 is not executed.

[0016] Optionally, the kernel function of the multi-task Gaussian process regression algorithm is: ,in: t and t' are task indices, where t=0 represents the source task and t=1 represents the target task; Ktask(t,t') is the task-related kernel, used to learn the correlation between the source task and the target task; Kinput(x,x') is the input kernel used to learn the similarity of the process parameter space.

[0017] Optionally, the process may further include the following step between step S21 and step S22: The cloud-based system employs a space-filling design, selecting multiple process parameters as test points within the process parameter area. The cloud-controlled edge control layer executes the production test corresponding to the test point on the production line to obtain the initial target task sample. The cloud adds the initial target task sample to the sample set and uses it for model update in step S22.

[0018] Furthermore, to achieve the above objectives, the present invention also provides a parameter optimization system based on edge control layer and cloud collaboration, characterized in that it includes a cloud optimization platform and an edge control layer, wherein the cloud optimization platform and the edge control layer communicate through an encrypted network protocol, wherein the cloud optimization platform includes: The receiving module is used to receive a sample set, wherein the sample contains process parameters and corresponding product quality values; The proxy model construction module is used to construct a proxy model based on the samples in the sample set using a preset algorithm. The proxy model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters. The optimization decision module is used to calculate candidate process parameters from the process parameter region by maximizing the value of a preset acquisition function based on the predicted mean and the uncertainty measure output by the proxy model. The safety verification module is used to perform safety verification on the candidate process parameters and output a trigger signal when the verification fails, so that the optimization decision module can recalculate the candidate process parameters. The instruction issuing module is used to issue the verified candidate process parameters to the edge control layer; The sample update and iteration control module is used to update the sample set and control the proxy model construction module, optimization decision module, security verification module and instruction issuance module to repeatedly execute related operations until the preset convergence conditions are met.

[0019] The edge control layer includes: The instruction execution module is used to control the production line to execute production based on the candidate process parameters issued by the application. The safety monitoring module is used to monitor the production status in real time. The rollback module is used to trigger a rollback operation when an anomaly is detected. The data upload module is used to upload the candidate process parameters and related data to the cloud optimization platform when production is abnormal, and to upload the candidate process parameters and measured product quality values ​​to the cloud optimization platform when production is completed normally.

[0020] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the parameter optimization method based on edge control layer and cloud collaboration as described in any of the preceding claims. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] To gain a more complete understanding of this application and its beneficial effects, the following description will be provided in conjunction with the accompanying drawings. In the following description, the same reference numerals denote the same parts.

[0023] Figure 1 This is a flowchart illustrating a parameter optimization method based on edge control layer and cloud collaboration, provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0027] Example 1: Please refer to Figure 1 , Figure 1 A flowchart of a parameter optimization method based on edge control layer and cloud collaboration is provided in this embodiment of the invention. The method can be applied to polymer material processing production lines such as plastic extrusion and injection molding. The method includes: Step S1: Receive a sample set in the cloud, where each sample contains process parameters and corresponding product quality values; Among them, the cloud refers to a remote service platform deployed outside the production site. As the decision-making center for implementing parameter optimization methods based on the collaboration between the edge control layer and the cloud, it conducts two-way data interaction and command issuance with the edge control layer of the production site through an encrypted network. It is responsible for receiving and storing sample sets, as well as subsequent operations such as building proxy models, determining candidate process parameters and performing security verification, thereby realizing the perception, decision-making and optimization guidance of the on-site production line operation status.

[0028] A sample set refers to a data set composed of multiple samples, which is the basic data source for building proxy models in the cloud. Each sample is a piece of production data, and each sample contains process parameters and corresponding product quality values. The process parameters refer to various production operation parameters that need to be set and adjusted during the operation of the production line, such as temperature, speed, and pressure. These are key variables that determine product quality, production efficiency, and production costs, and are also the objects to be optimized in this embodiment.

[0029] In this embodiment, the sample set received by the cloud is uploaded by the edge control layer deployed on the production line. During production line operation, the edge control layer collects data in real time, collects the current process parameters and the measured product quality values ​​corresponding to the process parameters, obtains samples, and uploads them to the cloud. For example, in the early stage of debugging, when the production line reaches a stable operating state under the set process parameter x_current, the edge control layer uploads the process parameter x_current and the product quality value y_current corresponding to the process parameter x_current as a sample to the cloud.

[0030] It should be noted that samples can also be imported manually, for example, by copying them from storage devices such as external hard drives. This data may include debugging data from the initial stage of production line commissioning, process parameters recorded during past manual adjustment of process parameters, and corresponding product quality data. Samples can also be received by receiving sample data stored locally in the cloud, that is, historical production data that has been collected, organized, and stored in the cloud during past parameter optimization processes, which can be directly called as the sample set for this application. This application does not impose any restrictions on this.

[0031] In step S1, the cloud receives a sample set, which includes process parameters and corresponding product quality values, thus providing a basic data source for the subsequent construction of the proxy model.

[0032] Step S2: The cloud platform constructs a proxy model based on the samples in the set using a preset algorithm. The proxy model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters.

[0033] Preferably, the preset algorithm in this embodiment is Gaussian Process Regression (GPR). This algorithm can directly output the predicted mean and predicted variance, and it still has stable modeling accuracy under small sample conditions. The model output is stable and the uncertainty estimation is reliable, making it more suitable for modeling scenarios with limited sample size and complex operating conditions in polymer material processing production lines. In other embodiments, the preset algorithm can also be Random Forest Quantile Regression (RFQR), Bayesian Neural Network (BNN), or other algorithms with uncertainty quantification capabilities. As long as the algorithm can take process parameters as input, product quality as output, and provide the predicted mean and corresponding uncertainty measure, it falls within the scope of the preset algorithms covered in this embodiment.

[0034] A surrogate model is a model trained on samples to describe the mapping relationship between process parameters and product quality. Its function is to directly output the predicted mean of product quality for any set of process parameters, along with a measure of the uncertainty of that predicted mean, without the need for actual trial production. The predicted mean of product quality refers to the expected value of the surrogate model's final product quality output under given process parameters, representing the surrogate model's comprehensive judgment of the quality level achievable with those parameters. The uncertainty measure is the surrogate model's confidence assessment of the current prediction result, reflecting the degree of its cognitive limitations at that process parameter point, and providing a basis for subsequent exploration and utilization of the balance.

[0035] In this embodiment, the cloud loads the sample set received in step S1, uses the process parameters in the samples as input features and the product quality value as the target output, optimizes the kernel function parameters based on the GPR algorithm and minimizes the prediction error, and completes the construction of a proxy model with the ability to output the product quality prediction mean and uncertainty measurement.

[0036] As a preferred implementation, when constructing the proxy model using a preset algorithm, the cloud also incorporates prior knowledge. The method for incorporating prior knowledge includes at least one of the following: using the safety boundary of process parameters as a constraint condition for training the proxy model; generating samples containing virtual process parameters and virtual product quality values ​​based on physical formulas or expert experience rules, and adding the samples to the set.

[0037] The process parameter safety boundary refers to the range of process parameter values ​​determined based on process principles, equipment capabilities, quality standards, and safety production requirements in industrial scenarios such as polymer material processing. For example, the maximum barrel temperature cannot exceed 210℃, and the minimum screw speed cannot be lower than 50 r / min. When using the GPR algorithm to construct a surrogate model, the process parameter safety boundary is used as the input constraint of the surrogate model. That is, only process parameters and corresponding product quality values ​​within the process parameter safety boundary are input to the surrogate model as training data. This restricts the input range of the surrogate model to always fall within the safety boundary, so that the learning process of the surrogate model focuses only on the feasible process parameter region that meets the safety boundary. This avoids the surrogate model learning process parameter mapping relationships that exceed equipment capacity, do not conform to process logic, or pose safety hazards, thus solving the problem that existing solutions ignore the essential constraints of process principles on the process parameter search space.

[0038] When constructing a proxy model using the GPR algorithm, physical formulas or expert rules such as thermodynamic equations, Newton's law of cooling, and melt flow models can be transformed into computable functions. For example, rules such as extending the holding time appropriately for higher temperatures and coordinating screw speed and back pressure adjustment are then generated based on these functions. These samples, containing virtual process parameters and virtual product quality values, are then added to the sample set described in step S1. The cloud platform then constructs the proxy model based on the expanded sample set.

[0039] By integrating prior knowledge, the proxy model can acquire process cognition capabilities, enabling optimization decisions to no longer rely solely on data, but to take into account both data patterns and the essence of the process. This enhances the intelligence, reliability, and interpretability of the optimization process. At the same time, it can leverage expert experience to avoid known process risks, prevent process parameters from deviating from experience standards, reduce product defects and raw material waste, and balance production safety, product quality, and process rationality under complex operating conditions and dynamic disturbances. This reduces the costs and risks associated with blind trial production, making parameter optimization results more aligned with the actual processing scenario requirements.

[0040] Step S3: The cloud-based system calculates the candidate process parameter x_candidate from the process parameter region by maximizing the value of a preset acquisition function based on the predicted mean and the uncertainty measure output by the proxy model. Preferably, the preset acquisition function in this embodiment is the Expected Improvement (EI) function, whose expression is: ,in: EI(x) is the expected improvement value of process parameter x, which is used to quantify the expected quality improvement that can be obtained by performing a production test under the process parameter; μ(x) is the mean product quality prediction of the surrogate model at process parameter x; f(x + ) represents the highest known product quality value in the current set; ξ is a preset non-negative exploration parameter used to balance the utilization and exploration intensity during the optimization process. The larger the value of ξ, the more the EI tends to explore regions with high uncertainty; the smaller the value of ξ, the more the EI tends to utilize currently known high-quality process parameter regions. In the process parameter optimization scenario of polymer material processing, the value of ξ can range from 0.01 to 0.1, for example, ξ=0.05. The specific value can be dynamically adjusted according to the parameter optimization stage: in the early stage of optimization, i.e., the cold start stage, ξ can be set to 0.08~0.1 to prioritize strengthening the exploration capability to cover more process parameter regions; in the middle and late stages of optimization, ξ can be adjusted to 0.01~0.05 to focus on utilizing the validated high-quality process parameter regions and improve the optimization convergence efficiency.

[0041] σ(x) is the standard deviation of the surrogate model's prediction at process parameter x, reflecting the uncertainty measure of the surrogate model at that point; Z is a standardized variable, and its calculation formula is: Z= ; When σ(x)=0, the formula is defined as: EI(x)=0, which means that there is no longer any value in further exploration at the observed process parameter points; Φ(Z) is the cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

[0042] EI explicitly introduces the uncertainty measure σ(x) from the surrogate model output and the preset exploration parameter ξ, enabling the optimization process to proactively select parameter points with the maximum information gain, rather than simply the highest mean, for experimentation in regions where real data is sparse and model uncertainty is high. This mechanism fundamentally solves the problems of blind trial and error in the cold start phase and premature convergence at local extrema in traditional methods, improving sample utilization efficiency and global optimization accuracy. In other implementations, the acquisition function can also be other policy functions with the ability to balance exploration and utilization, such as Entropy Search (ES) or Upper Confidence Bound (UCB), and this application does not limit this to any particular one.

[0043] By maximizing the value of EI within the process parameter range, the cloud automatically calculates and determines the candidate process parameter x_candidate that maximizes the acquisition function. This candidate process parameter x_candidate is the optimal parameter to be verified, balancing the potential for quality improvement with the value of exploring uncertainties, and provides direction for subsequent iterative optimization.

[0044] In step S3, the cloud inputs the predicted mean and uncertainty metric output by the surrogate model into a preset acquisition function. The maximum value of the acquisition function is then calculated through numerical optimization, determining the candidate process parameter x_candidate from the process parameter region. This process realizes the mapping and transformation from model cognition to experimental decision-making, providing definite parameter inputs for the subsequent production test verification in step S5 and the surrogate model update in step S2.

[0045] Step S4: The cloud performs a security check on the candidate process parameter x_candidate. If the check fails, it returns to step S3 to recalculate. If the check passes, the candidate process parameter x_candidate is sent to the edge control layer. Preferably, the security verification includes at least one of the following verifications: Process parameter safety boundary verification, process parameter change step size verification, and simulation verification based on the process model.

[0046] Among them, the process parameter safety boundary verification is used to determine whether the number of candidate process parameters x_candidate is within a reasonable range of values ​​determined by the process principle, equipment capability, quality standards and safety production requirements. By comparing the candidate process parameters x_candidate with the preset upper and lower limits of process parameters, the candidate process parameters x_candidate that exceed the physical limits of the equipment, the tolerance range of the material, or violate the basic process constraints are directly eliminated, thus preventing unsafe and unreasonable candidate process parameters x_candidate from entering the subsequent test stage from the source.

[0047] The process parameter change step size verification is used to limit the single adjustment range between the candidate process parameter x_candidate and the actual process parameters of the current production line. By setting the maximum allowable change step size, for example, the process parameter change does not exceed 1%-5% of the original process parameter, even if the candidate process parameter x_candidate passes the aforementioned step size verification, it is still subjected to limit processing to generate the final set value that can be issued, ensuring that the parameter adjustment is smooth and preventing production fluctuations, equipment shocks or quality abnormalities caused by sudden changes in process parameters.

[0048] Simulation verification based on the process model uses a built-in process mechanism model to virtually verify the candidate process parameter x_candidate, predicting in advance the feasibility, stability and potential production risks of the candidate process parameter x_candidate in actual production, and further screening out process parameters that fit the actual production conditions.

[0049] If the verification passes, the candidate process parameter x_candidate will be sent to the edge control layer.

[0050] Through the above multi-layer security verification, the cloud can identify and eliminate unreasonable, unsafe or potentially production-volatile candidate process parameters x_candidate before sending them to the production site, thereby reducing the risk of equipment wear, raw material waste and quality defects caused by blind trial production during the parameter optimization process.

[0051] Step S5: The edge control layer controls the production line to use the candidate process parameter x_candidate for production and monitors the production status in real time. If an anomaly is detected, a rollback is triggered, and the candidate process parameter x_candidate and associated process data are uploaded to the cloud. The cloud generates a product quality observation value for the candidate process parameter x_candidate that is inferior to the highest product quality value in the current set, and adds a new sample consisting of the candidate process parameter x_candidate and the product quality observation value to the set. If production is completed normally, the edge control layer uploads the candidate process parameter x_candidate and the measured product quality value to the cloud. The cloud adds a new sample consisting of the candidate process parameter x_candidate and the measured product quality value to the set. Preferably, the edge control layer includes a proportional-integral-derivative (PID) controller, which controls the adjustment of process parameters. In other embodiments, the controller of the edge control layer can also be replaced by an advanced controller with setpoint tracking and real-time adjustment capabilities, such as a fuzzy proportional-integral-derivative (Fuzzy-PID) controller, a model predictive control (MPC) controller, or an adaptive proportional-integral-derivative (Adaptive-PID) controller. This application does not limit this to any particular controller.

[0052] The edge control layer refers to the control layer deployed on the production line and directly connected to the equipment. Its functions include real-time production control, data acquisition, safety monitoring, and anomaly handling. Specifically, it receives candidate process parameters (x_candidate) from the cloud and converts them into executable instructions for the equipment. Based on these instructions, it controls equipment operation in real time, synchronously collects production data such as equipment operating status and process execution, and monitors production safety throughout the process, identifying abnormal operating conditions. The cloud and edge control layer interact and issue instructions via encrypted network protocols, forming a collaborative working mode of "global optimization in the cloud and real-time execution at the edge." Compared to remote computing in the cloud, the edge control layer offers advantages such as low latency response, high real-time performance, and independence from network status, making it suitable for the real-time management and control needs of industrial production.

[0053] The candidate process parameter x_candidate is the process parameter to be evaluated that is sent from the cloud to the edge control layer for actual production verification after passing the security verification in step S4. It is the control parameter for production testing in this round of optimization iteration, and also the process parameter whose merits and demerits need to be verified in the optimization process.

[0054] Production status refers to the comprehensive real-time operating conditions of the production line during operation, reflecting the equipment, processes, and product quality. This includes, but is not limited to, equipment operating status, process parameter execution status, and product forming status, such as melt flow and product appearance. It is the core basis for judging whether production is normal. Among them, equipment operating status includes, for example, whether the motor speed and pressure are stable; process parameter execution status includes, for example, whether the actual parameters match the candidate parameters; and product forming status includes, for example, melt flow and product appearance.

[0055] An anomaly refers to an abnormal operating condition in which the production status deviates from the preset safety or quality threshold. This includes, but is not limited to, actual values ​​of process parameters deviating from the set range of candidate parameters, equipment malfunctions, product quality anomalies, and safety hazards. Equipment malfunctions include material jams and overloads, product quality anomalies include material shortages and deformations, and safety hazards include exceeding temperature limits. Once an anomaly is detected, the test of the current candidate process parameter x_candidate will be terminated immediately to avoid losses.

[0056] Rollback refers to the immediate suspension of the execution of the current candidate process parameter x_candidate when the edge control layer detects a production anomaly, and the restoration of the production line's process parameters to the normally operating process parameters that were verified before the anomaly occurred.

[0057] Related process data refers to real-time data collected synchronously by the edge control layer when an anomaly occurs, which is related to the anomaly time and candidate process parameter x_candidate. This includes, but is not limited to, the anomaly occurrence timestamp, anomaly type identifier, actual values ​​of process parameters at the time of the anomaly, equipment operating parameters, and raw material batch information. This data is used for subsequent cloud-based analysis of the anomaly causes and optimization of candidate parameter calculations.

[0058] The product quality observation is a virtual product quality value generated by the cloud in the event of production anomalies and the inability to obtain actual measured product quality. This observation is set to be inferior to the highest known product quality in the current sample set. Its purpose is to transform the abnormal process parameter into a sample that can be used to train the surrogate model, so as to avoid the parameter being ignored or misjudged by the model due to the lack of product quality data.

[0059] The measured product quality value refers to the product quality data corresponding to the candidate process parameter x_candidate that is actually measured after normal production is completed. It is the measured data for judging the merits of the candidate process parameter x_candidate.

[0060] In this step, the edge control layer controls the production line to use the candidate process parameter x_candidate for production. The specific process is as follows: After receiving the candidate process parameter x_candidate from the cloud, the edge control layer writes it as a set value into the PID controller. The PID controller dynamically adjusts the control command output according to the deviation between the set value and the sensor feedback value, drives the operation of each execution device on the production line, and makes the actual process parameters track the set value of the candidate process parameter x_candidate, ensuring that the production line produces according to the candidate process parameter x_candidate.

[0061] During production line operation, the edge control layer continuously collects production status data at a preset frequency through sensors deployed on the production line, monitoring the actual values ​​of process parameters, equipment operating status, and product quality characteristics in real time. These sensors include temperature sensors, pressure sensors, and speed sensors. The data acquisition frequency is preset according to production process requirements, for example, it can be 1Hz; this application does not limit this.

[0062] When the edge control layer detects any production status indicator exceeding a preset threshold or violating preset logic rules, it determines the current test is abnormal and triggers a rollback mechanism. This restores the PID controller's setpoint to the process parameters corresponding to the previous safe state before the abnormality occurred, allowing the production line to return to a stable operating state. Simultaneously, the edge control layer uploads the candidate process parameters x_candidate and associated process data to the cloud. Upon receiving the data, the cloud generates a product quality observation value for the candidate process parameter x_candidate that is inferior to the highest product quality value in the current set. The new sample, composed of the candidate process parameter x_candidate and the product quality observation value, is then added to the set used to train the proxy model, completing the incremental update of samples under abnormal operating conditions.

[0063] If the edge control layer does not trigger any anomalies during the complete production cycle, i.e., production is completed normally, the edge control layer uploads the candidate process parameter x_candidate and the corresponding measured product quality value of this test to the cloud; the cloud adds the new sample composed of the candidate process parameter x_candidate and the measured product quality value to the set, completing the sample incremental update under normal operating conditions.

[0064] In step S5, the edge control layer controls the production line to apply the candidate process parameter x_candidate to produce and monitor the status in real time. If an anomaly is detected, a rollback is triggered and data is uploaded. The cloud generates a product quality observation value that is inferior to the highest value and adds it to the set, realizing rapid response and safe handling in case of anomalies, and supplementing failure scenario data to update the sample set. If production is normal, the measured product quality value is uploaded and added to the set, providing real feedback to improve the model and form a closed-loop verification.

[0065] In a preferred embodiment, the product quality observation values ​​are generated in at least one of the following ways: (1) Set to a fixed value that is inferior to the highest product quality value in the current set; The specific implementation process of this generation method is as follows: The cloud first reads all product quality values ​​in the current sample set, and then filters to obtain the highest product quality value f(x). + Subsequently, the cloud directly assigns the product quality observation value to a preset fixed value, and this fixed value is lower than f(x). + This ensures that the candidate process parameter x_candidate that triggers the exception is marked as a poor sample.

[0066] For example, in the context of polymer material processing, the product size qualification rate is used as a quality evaluation index, with a maximum score of 100. If the highest product quality value f(x) in the current sample set... +If the value is 95, a fixed value of 60 can be preset. That is, when an abnormality occurs in production, the cloud will directly set the product quality observation value corresponding to the candidate process parameter x_candidate to 60. If the quality evaluation index is melt flow rate (unit g / 10min) and the current highest value is 25, a fixed value of 5 can be preset to ensure that the value is lower than the highest product quality value.

[0067] In the aforementioned manner, when the cloud performs step S3 to calculate EI(x_candidate), since the observed quality value of the product is lower than f(x+), after the surrogate model is updated, the predicted mean μ(x) of the point and its neighboring areas will decrease accordingly, and the value of EI(x_candidate) will be less than or equal to zero. The candidate process parameter x_candidate will be judged by EI as having no exploration and utilization value, and production trials of the candidate process parameter x_candidate and similar parameters will no longer be recommended. This avoids equipment losses and raw material waste caused by ineffective trial production, guides the optimization process to converge toward a better and safer process parameter region, and improves the efficiency and reliability of overall parameter optimization.

[0068] (2) Set as the value after deterioration based on the highest product quality value in the current set according to a preset ratio; The specific implementation process of this generation method is as follows: The cloud first reads the highest product quality value f(x) in the current sample set. + Subsequently, the cloud system degrades according to a preset degradation ratio at f(x). + Linear degradation calculations are performed using f(x) as a baseline to obtain product quality observations. This degradation ratio is preset based on process experience, is a non-negative value, and can be dynamically adjusted during the optimization phase. The calculation results are ensured to be lower than f(x). + This causes the candidate process parameter x_candidate that triggers the exception to be marked as a poor sample.

[0069] In the aforementioned manner, when the cloud performs the EI(x) calculation in step S3, if the EI(x_candidate) value corresponding to the candidate process parameter x_candidate is always less than or equal to zero, it will be automatically determined to have no exploration or utilization value. The EI will no longer recommend it as a candidate process parameter, thereby avoiding repeated trial production of the candidate process parameter x_candidate and similar parameters, reducing equipment wear and raw material waste. At the same time, compared with the fixed value method, its observed value is dynamically adapted to the current highest product quality value, which can avoid sample distribution imbalance and further improve the training accuracy and optimization convergence efficiency of the proxy model.

[0070] (3) Dynamically calculate and generate based on the anomaly type and severity represented by the associated process data.

[0071] The specific implementation process of this generation method is as follows: First, the cloud extracts abnormal features from the associated process data, identifies the abnormality type (e.g., deviation of process parameters, equipment failure, safety exceeding limits, raw material abnormalities), and the severity of the abnormality (quantified and graded based on a preset threshold, such as slight, moderate, and severe). Then, based on the abnormality type and severity, a corresponding degradation coefficient is generated. Finally, the highest product quality value f(x) in the current sample set is used as the degradation coefficient. + Based on the baseline, the product quality observation value is calculated in conjunction with the degradation coefficient. The more severe the anomaly, the greater the degree of degradation of the observation value, causing the candidate process parameter x_candidate to be marked as a substandard sample of the corresponding level.

[0072] Using the aforementioned method, the EI(x_candidate) value is also less than or equal to zero, and EI automatically avoids this parameter and its neighboring regions. Compared to the previous two methods, the dynamic calculation method enables observations to distinguish the differences in the impact of different anomalies on quality, making the surrogate model's learning more closely aligned with real production patterns, and further improving the accuracy, safety, and reliability of parameter optimization.

[0073] Step S6: Based on the updated set, repeat steps S2 to S5 until the preset convergence condition is met.

[0074] Specifically, after each round of sample increment update in step S5, the cloud re-executes the proxy model construction, candidate process parameter determination, safety verification, and on-site production verification process based on the updated sample set, forming an iterative parameter optimization closed loop until the preset convergence condition is met and the iteration stops.

[0075] After reaching the preset convergence condition, the cloud outputs the globally optimal process parameters that have been verified historically, along with a surrogate model built based on the updated sample set. The optimal process parameters are selected from those that have been optimized through multiple rounds of iterations and verified in actual trial production. They can be directly applied to the production line to achieve stable and high-quality production. The finally updated surrogate model fully integrates all valid sample data from the entire optimization process. It can be directly reused in subsequent situations such as changes in operating conditions and parameter fine-tuning, without having to iterate and optimize from scratch. This improves the efficiency of subsequent process optimization and reduces testing and production costs.

[0076] In a preferred embodiment, the convergence condition includes at least one of the following: (1) The maximum value of the preset acquisition function within the process parameter range is lower than a set threshold; This condition determines whether the parameter optimization space has been fully explored by measuring the value of the acquisition function: when the maximum value of the acquisition function is lower than the set threshold, it indicates that it is difficult to find parameters with higher quality improvement potential within the process parameter range, and the iterative optimization has reached a convergence state.

[0077] Specifically, in each iteration, the cloud calculates the maximum value of the acquisition function within the safety boundary of the process parameters and compares it with a preset threshold. If the maximum value of the acquisition function is less than or equal to the preset threshold, the convergence condition is satisfied.

[0078] By using the threshold of the acquisition function as the convergence criterion, it can be ensured that the optimization process automatically stops when sufficient accuracy is reached, avoiding redundant iterations and improving the efficiency of parameter optimization.

[0079] (2) The number of iterations reaches the preset limit.

[0080] This condition is a hard constraint on the iterative process, used to avoid infinite iterations due to complex working conditions and large optimization space, and to ensure that the entire parameter optimization process is completed within the preset time and experimental cost.

[0081] Specifically, the cloud-based system pre-sets an upper limit for the number of iterations. Each time a loop from step S2 to S5 is completed, the number of iterations is incremented by one. When the number of iterations reaches the preset upper limit, the iteration process is terminated directly regardless of the value of the acquisition function.

[0082] By setting an upper limit on the number of iterations, the parameter optimization method described in this invention can better fit the actual production constraints in industrial settings, thereby improving the engineering practicality of the solution.

[0083] The above convergence conditions can be used individually or in combination. The iteration will terminate when any condition is met, which ensures the accuracy of optimization while taking into account the efficiency and cost control.

[0084] After reaching the preset convergence condition, the cloud can also output the globally optimal process parameters that have been verified in history, and output the best surrogate model built based on the updated sample set, so that it can be directly applied to the process parameter configuration, offline prediction and rapid iterative optimization of the same type of production line.

[0085] The beneficial effects of this embodiment are as follows: The technical solution of constructing a proxy model using the GPR algorithm and automatically finding candidate process parameters via EI eliminates the need for repeated manual trial and error. Parameter optimization and safety verification are automatically completed in the cloud, replacing traditional manual debugging, significantly shortening the parameter tuning cycle, avoiding product quality fluctuations caused by differences in operator experience, and reducing raw material waste and safety hazards. Through a closed-loop control technology solution involving multi-layered safety verification, real-time monitoring at the edge control layer, and anomaly rollback, candidate process parameters undergo safety verification before being issued. During production, the edge control layer monitors the operating conditions in real time, triggering rollback and marking substandard samples in case of anomalies. This avoids equipment failures and production fluctuations caused by process parameter exceeding limits, providing safety monitoring throughout the entire process from parameter issuance to production execution. This solves the problem that existing machine learning proxy model optimization technologies only provide open-loop output, lack real-time safety monitoring and closed-loop verification, and cannot intervene in time for parameter exceeding limits or production fluctuations. By employing a proxy model to quantify uncertainty, exploring and utilizing the EI function balance, and automatically labeling process parameters that cause anomalies using product quality observations, this approach solves the problems of existing optimization algorithms, such as the need for extensive blind trial production, high trial-and-error costs, inability to quantify decision uncertainty, and high risks in exploring unknown process regions. It fundamentally reduces blind trial production, lowers trial-and-error costs, and reduces safety and quality risks when exploring unknown process regions, making the overall optimization process more stable, efficient, and reliable.

[0086] Example 2: Based on Embodiment 1, in this embodiment, the sample set in step S1 further includes samples from two different sources: target task samples and source task samples.

[0087] The target task sample refers to a sample derived from the current target to be optimized, which can specifically be the production line to be optimized, such as an extruder or injection molding machine. Specifically, the target task sample is a sample generated during the production process of the target to be optimized, containing process parameters and corresponding product quality values. The process parameters are various adjustable parameters in the actual operation of the production line, such as temperature, speed, and pressure. The product quality values ​​are the measured quality data of the products produced under these process parameters, such as coating thickness tolerance. The process parameters and product quality values ​​correspond one-to-one, together constituting the target task sample. Taking an extruder as an example, several process parameters are selected as test points within the process parameter range for production, and the corresponding product quality values ​​are obtained. This set of process parameters and corresponding product quality values ​​constitutes a target task sample. The target task sample directly reflects the real data of the current production conditions and dynamically increases with optimization iterations. It is the core basis for constructing the multi-task proxy model in this embodiment.

[0088] Source task samples refer to samples from targets other than the current target to be optimized. These other targets include, but are not limited to: other production equipment of the same model as the current target to be optimized; different production lines using the same technological principles as the current target; other production lines using raw materials similar to the current processing materials; and similar production lines from different production batches under the same manufacturer. Source task samples typically have a large data volume and can provide general patterns regarding the mapping relationship between "process parameters and product quality," but may deviate somewhat from the current operating conditions.

[0089] Unlike Example 1, the preset algorithm in this example is Multi-task Gaussian Process Regression (MTGPR), the surrogate model is a multi-task surrogate model, and the kernel function of the MTGPR algorithm is: ,in: t and t' are task indices, where t=0 represents the source task and t=1 represents the target task; K task (t,t') is the task-related kernel, used to learn the correlation between the source task and the target task; K input (x,x') is the input kernel, used to learn the similarity of the process parameter region.

[0090] The source task refers to other production processes that have similar technological principles or production conditions to the target task, but are not the current optimization target. Examples include other equipment of the same model as the target production line. The source task itself does not require further optimization, but its samples can serve as prior knowledge to help model the target task. The samples corresponding to the source task are called source task samples. The target task refers to the production process currently being optimized, such as an extruder production line that needs to find the optimal process parameters. The ultimate goal of this embodiment is to find the process parameters that optimize product quality for this task; the samples corresponding to the target task are called target task samples.

[0091] This kernel function allows the model to improve prediction accuracy by leveraging source task samples when target task samples are sparse, while retaining its ability to model the target task. Multi-task Gaussian process regression retains the uncertainty quantification capability of standard Gaussian processes, outputting the predicted mean μ(x) and predicted standard deviation σ(x) for the target task. It also supports an incremental update mechanism, making it suitable for dynamically fusing new samples during iterative optimization, and is particularly applicable to data-scarce scenarios in industrial production, such as new product introductions and production line switching.

[0092] In this embodiment, step S2 includes the following sub-steps: S21. When step S2 is executed for the first time, the cloud uses the source task samples to pre-train the multi-task Gaussian process and build a multi-task agent model.

[0093] Specifically, when the system runs for the first time or when there is no actual measured sample data for the target task, the cloud loads a set of source task samples. This source task sample data can be obtained in advance through methods such as reading from local cloud storage or manual import. The cloud uses the process parameters in the source task samples as input features and the corresponding product quality values ​​as output targets to pre-train a multi-task Gaussian process. This process learns and determines the task-related kernel K by maximizing marginal likelihood estimation. task With input kernel K input The hyperparameters enable the model to learn the general rules of the "process parameters-product quality" mapping relationship in the source task, and to initially establish the prior estimate of the correlation between the source task and the target task, thereby constructing an initial multi-task agent model. After pre-training, the multi-task agent model has the general process knowledge extracted from the source task samples, providing an initial state for subsequent target task modeling.

[0094] S22. The cloud updates the multi-task agent model by combining the target task samples. In subsequent iterations, step S22 is only repeated and step S21 is no longer executed.

[0095] Specifically, the cloud integrates the acquired target task samples into the multi-task proxy model. Based on the general process knowledge obtained from the pre-training of the source task, the model further learns the process characteristics of the target to be optimized through the target task samples, gradually adapting its prediction ability from general process knowledge to the current actual production scenario. In the subsequent iteration execution of step S2, since the pre-training process corresponding to the source task samples has been completed and the relevant general process knowledge has been solidified in the multi-task proxy model, there is no need to perform the source task pre-training operation again. The cloud only performs incremental updates to the multi-task proxy model based on the currently updated target task samples, which reduces redundant calculations while ensuring the model's prediction accuracy, thereby improving the overall parameter optimization efficiency and iteration speed.

[0096] In a preferred embodiment, the method further includes the following step between step S21 and step S22: The cloud uses a space-filling design to select multiple process parameters as test points within the process parameter area, and sends the test points down to the edge control layer; The edge control layer performs production trials, acquires initial target task samples, and uploads them to the cloud. The cloud adds the initial target task sample to the sample set and uses it for model update in step S22.

[0097] Specifically, after completing the pre-training of the source task, the cloud platform has acquired general process knowledge, but it has not yet mastered the measured information of the target task. To enable the model to make accurate predictions for the current production scenario, a small number of initial experiments can be conducted within the process parameter region to obtain representative target task samples. Therefore, the cloud platform uses a space-filling design method to select multiple process parameters as experimental points within the process parameter region. The space-filling design can be a Latin hypercube design, a Sobol sequence, or a uniform design, etc. These designs can uniformly cover the entire process parameter region space with fewer sample points, avoiding sample aggregation in local areas, thereby maximizing the information gain of the initial experiments and providing a global observation basis for subsequent model updates. The Latin hypercube design method is preferred, as it can effectively reduce the correlation between parameters in different dimensions while ensuring that sample points are uniformly distributed across the various dimensions of the process parameters, avoiding overlap of experimental information due to parameter redundancy.

[0098] Subsequently, the cloud sends the selected test points to the edge control layer via an encrypted network protocol. Upon receiving this information, the edge control layer uses it as the setpoint for the PID controller or other advanced controllers, controlling the production line to execute the corresponding production tests. During production, the edge control layer monitors and records the measured product quality values ​​corresponding to each test point in real time, forming an initial target task sample. Each sample includes a set of process parameters and their corresponding product quality values, such as combinations of parameters like temperature, rotation speed, and pressure, as well as quality indicators like coating thickness tolerance.

[0099] After the experiment is completed, the edge control layer uploads the initial target task samples to the cloud. The cloud adds them to the sample set and uses them together with the source task samples for the multi-task agent model update in step S22. This process enables the model to quickly calibrate the correlation between tasks based on the prior knowledge of the source tasks and with the help of a small amount of real target data, thus achieving initial adaptation to the current production scenario.

[0100] The aforementioned preferred implementation method obtains representative target task samples with a small number of initial tests, balancing test efficiency and sample coverage, and reducing test costs. It is especially suitable for situations where there are no target task samples or very few target task samples during the cold start phase. In addition, target task samples can also be obtained by importing historical production data of the target to be optimized, manually entering process parameters and product quality measurement data confirmed by on-site testing, etc.

[0101] Consistent with Embodiment 1, after step S2, steps S3 to S6 are executed sequentially to complete the optimization of the process parameters of the optimization target. It should be noted that in this embodiment, the new samples obtained in step S5 are all target task samples.

[0102] The beneficial effects of this embodiment are: By introducing a transfer learning mechanism between the target task and the source task, and fully utilizing historical production data from similar scenarios, the modeling challenge in data-scarce scenarios is effectively addressed. Specifically, by distinguishing between target task samples and source task samples, the source task samples are used as prior knowledge for model pre-training. This allows the model to possess basic predictive capabilities even without target task samples, avoiding blind exploration in the initial optimization phase and improving optimization efficiency. Furthermore, a small number of initial experiments based on space-filling design are combined to quickly obtain globally representative initial target task samples, enabling model calibration with minimal experimentation and achieving rapid iterative optimization during the cold start phase. Simultaneously, a multi-task Gaussian process regression algorithm is employed, fusing the correlation between the source and target tasks through a kernel function. This ensures high prediction accuracy even when target task samples are sparse, effectively reducing the number of actual production experiments for process parameter verification and lowering production experiment costs. This multi-task transfer mechanism can fully reuse historical data from the same equipment model, historical batches, and similar processes, breaking the limitations of a single production line data source. It is applicable to various polymer material processing scenarios with similar historical data, demonstrating strong versatility and high promotional value.

[0103] Example 3: The parameter optimization method based on edge control layer and cloud collaboration in the embodiments of this application has been described above. The parameter optimization system based on edge control layer and cloud collaboration in the embodiments of this application is described below.

[0104] The parameter optimization system based on edge control layer and cloud collaboration in this embodiment includes a cloud optimization platform and an edge control layer. The cloud optimization platform and the edge control layer communicate through an encrypted network protocol. The cloud optimization platform includes: A receiving module is used to receive a sample set, wherein each sample contains process parameters and corresponding product quality values; The proxy model construction module is used to construct a proxy model based on the samples in the sample set using a preset algorithm. The proxy model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters. The optimization decision module is used to calculate candidate process parameters from the process parameter region by maximizing the value of a preset acquisition function based on the predicted mean and the uncertainty measure output by the proxy model. The safety verification module is used to perform safety verification on the candidate process parameters and output a trigger signal when the verification fails, so that the optimization decision module can recalculate the candidate process parameters. The instruction issuing module is used to issue the verified candidate process parameters to the edge control layer; The sample update and iteration control module is used to update the sample set and control the surrogate model construction module, optimization decision module, security verification module and instruction issuance module to repeatedly execute related operations until the preset convergence condition is reached. Finally, it outputs the globally optimal process parameters and the best surrogate model to achieve closed-loop iterative optimization.

[0105] The edge control layer includes: The data acquisition module, deployed on the production line, continuously collects production status data at a preset frequency using various sensors such as temperature sensors, pressure sensors, and speed sensors. This production status data includes equipment operating status, process parameter execution status, and product forming status, providing a real-time and accurate monitoring data source for the safety monitoring module. Simultaneously, in conjunction with the data upload module, it synchronously collects candidate process parameters, measured product quality values, and related process data under abnormal operating conditions, ensuring the comprehensiveness and timeliness of data acquisition. The instruction execution module is used to control the production line to execute production based on the candidate process parameters issued by the application. The safety monitoring module is used to monitor the production status in real time and send a trigger signal to the rollback module when an anomaly is detected. The rollback module is used to trigger a rollback operation when a trigger signal is received. The data upload module is used to upload the candidate process parameters and related data to the cloud optimization platform when production is abnormal, and to upload the candidate process parameters and measured product quality values ​​to the cloud optimization platform when production is completed normally.

[0106] Through the collaborative efforts of the various modules mentioned above, this system achieves closed-loop intelligent optimization and full-process safety control of process parameters in polymer material processing production lines. It breaks through the barriers of low efficiency in traditional manual parameter tuning, the open-loop output of existing machine learning optimization, and the lack of real-time safety monitoring. It solves the technical challenges of blind trial production, high trial-and-error costs, unquantifiable decision-making uncertainty, and high risks in exploring unknown process areas during process parameter optimization. The specific collaborative working process of each module is as follows: The receiving module, as the data entry point of the cloud optimization platform, is responsible for receiving the sample set, which includes process parameters and corresponding product quality values. The sample set comes from production data uploaded in real time by the edge control layer, or production line debugging and past operation data imported manually, or historical production data stored locally in the cloud. This application does not limit this. The receiving module synchronizes the received sample set to the proxy model construction module to provide a basic data source for the construction of the proxy model.

[0107] The surrogate model construction module receives the sample set output by the receiving module. Based on the samples in the set, it constructs a surrogate model using a preset algorithm. The surrogate model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters. The preset algorithm is preferably the GPR algorithm, but algorithms with uncertainty quantification capabilities such as RFQR and BNN can also be used. When constructing the model, prior knowledge can also be integrated, including using the safety boundary of process parameters as the input constraint of the surrogate model, or generating virtual samples based on physical formulas and expert experience rules to expand the sample set, thereby optimizing the model accuracy. This allows the model learning process to focus on the feasible process parameter region, avoid known process risks, and make up for the shortcomings of existing surrogate models that lack process knowledge, rely solely on data training, and whose optimization results are detached from actual production conditions.

[0108] The optimization decision-making module constructs the predicted mean and uncertainty measure output by the surrogate model module, and calculates candidate process parameters from the process parameter region by maximizing the value of a preset acquisition function. The preset acquisition function is preferably EI, which balances the intensity of exploration and utilization in the optimization process by introducing the uncertainty measure of the surrogate model and the preset exploration parameter ξ. The value of ξ can be dynamically adjusted to adapt to different stages of optimization, avoiding blind trial production and premature convergence of local extrema. In addition, acquisition functions such as ES and UCB, which have the ability to balance exploration and utilization, can also be selected. The maximum value of the acquisition function is solved by numerical optimization to determine the candidate process parameters that take into account both the potential for quality improvement and the value of uncertainty exploration. This solves the shortcomings of existing optimization algorithms, such as the inability to quantify decision uncertainty, high risk of exploring unknown process regions, and easy getting trapped in local optima.

[0109] The safety verification module receives candidate process parameters output by the optimization decision module, performs multi-layer safety verification on these parameters, and outputs a trigger signal when a verification fails, causing the optimization decision module to recalculate the candidate process parameters. The safety verification includes process parameter safety boundary verification, process parameter change step size verification, and simulation verification based on the process model. Safety boundary verification eliminates parameters that exceed equipment limits or material tolerance ranges. Step size verification limits the magnitude of single parameter adjustments; even if a parameter passes step size verification, it is still subject to amplitude limiting to ensure smooth adjustment. Simulation verification uses virtual operation to predict parameter feasibility and potential risks. These multi-layer verifications collectively reduce the risks of equipment wear, raw material waste, and quality defects caused by blind trial production, ensuring the safe issuance of candidate process parameters.

[0110] The instruction issuance module will send the candidate process parameters verified by the security verification module to the instruction execution module of the edge control layer through an encrypted network protocol. This ensures the secure and accurate transmission of instructions, enables the linkage between cloud-based optimization decisions and on-site production execution, provides reliable control parameters for actual production trials on the production line, avoids loss or tampering during instruction transmission, and ensures the timeliness and accuracy of parameter issuance.

[0111] The data acquisition module is deployed on the production line and continuously collects production status data at a preset frequency (e.g., 1Hz) using various sensors such as temperature sensors, pressure sensors, and speed sensors. This includes real-time operating data such as equipment operating status (e.g., motor speed, pressure stability), process parameter execution status (e.g., the fit between actual and candidate parameters), and product forming status (e.g., melt flow, product appearance). It also provides a monitoring data source for the safety monitoring module. Furthermore, when production is normal or abnormal, it works with the data upload module to collect candidate process parameters, measured product quality values, and related process data, achieving real-time acquisition of data throughout the entire production process.

[0112] The instruction execution module receives candidate process parameters from the instruction distribution module of the cloud optimization platform, converts them into control instructions that can be executed by the production line equipment, and dynamically adjusts the equipment operating parameters through PID controllers or advanced controllers such as Fuzzy-PID and MPC, so that the actual process parameters track the set values ​​of the candidate process parameters, and controls the production line to conduct production tests according to the candidate process parameters, ensuring that the optimization parameters distributed by the cloud can be executed.

[0113] The safety monitoring module receives real-time production status data uploaded by the data acquisition module, monitors the production line's operating conditions in real time, and determines whether the production status deviates from preset safety or quality thresholds. Abnormal operating conditions include actual values ​​of process parameters deviating from the candidate parameter setting range, equipment malfunctions (such as material jamming or overload), product quality abnormalities (such as material shortage or deformation), and safety hazards (such as exceeding temperature limits). If an abnormality is detected, an abnormality trigger signal is immediately sent to the rollback module, and the data upload module is simultaneously notified to prepare to upload abnormality-related data. This enables real-time safety monitoring of the production process and solves the problem that existing machine learning optimization technologies lack real-time monitoring and cannot detect production abnormalities in a timely manner.

[0114] When the rollback module receives an abnormal trigger signal from the safety monitoring module, it immediately triggers a rollback operation, stops the execution of the current candidate process parameters, restores the process parameters of the production line to the normal operating parameters that were verified before the abnormality occurred, so that the production line can quickly return to stability, avoid equipment wear, raw material waste and quality defects, and achieve rapid response and handling of abnormal operating conditions.

[0115] When production is abnormal, the data upload module uploads the candidate process parameters and related process data to the receiving module of the cloud optimization platform; when production is completed normally, the data upload module uploads the candidate process and the measured product quality values ​​collected synchronously by the data acquisition module to the cloud receiving module, so that the cloud optimization platform can update the sample set and provide measured data support for subsequent agent model updates and iterative optimization, realizing the closed-loop data flow between the cloud platform and the edge control layer.

[0116] As the closed-loop control core of the cloud-based optimization platform, the sample update and iteration control module receives new samples uploaded by the edge control layer data upload module, updates the sample set, and controls the proxy model construction module, optimization decision module, security verification module, and instruction issuance module to repeatedly execute related operations until the preset convergence conditions are met. Finally, it outputs the globally optimal process parameters and the best proxy model, forming a closed-loop dynamic optimization mechanism of "data acquisition - model construction - parameter optimization - security verification - instruction execution - real-time monitoring - sample update - iterative optimization". This replaces the traditional manual adjustment and passive optimization mode, continuously improving the accuracy and efficiency of process parameter optimization, and achieving the goals of stable quality, improved efficiency, and controllable cost in polymer material processing and production.

[0117] Example 4: According to an embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor performs the steps of the parameter optimization method based on edge control layer and cloud collaboration described in Embodiment 1.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0119] The above-mentioned products can execute any of the parameter optimization methods based on edge control layer and cloud collaboration described in Embodiment 1, and have the corresponding beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to the parameter optimization method based on edge control layer and cloud collaboration provided in Embodiment 1 of this invention.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A parameter optimization method based on edge control layer and cloud collaboration, characterized in that, include: Step S1: Receive a sample set from the cloud, where each sample contains process parameters and corresponding product quality values; Step S2: The cloud platform constructs a proxy model based on the samples in the set using a preset algorithm. The proxy model is used to output the corresponding product quality prediction mean and uncertainty measure according to the input process parameters. Step S3: The cloud-based system calculates candidate process parameters from the process parameter region by maximizing the value of a preset acquisition function based on the predicted mean and the uncertainty measure output by the proxy model. Step S4: The cloud performs a security check on the candidate process parameters. If the check fails, it returns to step S3 to recalculate. If the check passes, the candidate process parameters are sent to the edge control layer. Step S5: The edge control layer controls the production line to use the candidate process parameters for production and monitors the production status in real time. If an anomaly is detected, a rollback is triggered and the candidate process parameters and associated process data are uploaded to the cloud. The cloud generates a product quality observation value for the candidate process parameters that is inferior to the highest product quality value in the current set, and adds a new sample composed of the candidate process parameters and the product quality observation value to the set. If production is completed normally, the edge control layer uploads the candidate process parameters and the measured product quality values ​​to the cloud; the cloud adds a new sample consisting of the candidate process parameters and the measured product quality values ​​to the set. Step S6: Based on the updated set, repeat steps S2 to S5 until the preset convergence condition is met.

2. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, The cloud-based proxy model, constructed using a preset algorithm based on samples from the set, includes: When constructing the proxy model using a preset algorithm, the cloud also incorporates prior knowledge. The method for incorporating prior knowledge includes at least one of the following: using the safety boundary of process parameters as a constraint condition for training the proxy model; generating samples containing virtual process parameters and virtual product quality values ​​based on physical formulas or expert experience rules, and adding the samples to the set.

3. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, The security verification in step S4 includes at least one of the following verifications: Process parameter safety boundary verification, process parameter change step size verification, and simulation verification based on the process model.

4. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, In step S2, the preset algorithm is the Gaussian process regression algorithm.

5. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, In step S3, the acquisition function is the desired improvement function, and its expression is: ,in: EI(x) represents the expected improvement in process parameter x; μ(x) is the mean product quality predicted by the surrogate model at process parameter x; f(x + ) represents the highest known product quality value in the current set; ξ is a preset non-negative exploration parameter; σ(x) is the standard deviation of the prediction made by the surrogate model at point x, reflecting the uncertainty measure of the surrogate model at process parameter x; Z is a standardized variable, and its calculation formula is: Z= ； When σ(x) = 0, the formula is defined as: EI(x) = 0; Φ(Z) is the cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

6. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, The product quality observations are generated in at least one of the following ways: Set to a fixed value that is inferior to the highest product quality value in the current set; The value is set as the deteriorated value based on the highest product quality value in the current set, according to a preset ratio. The data is dynamically calculated and generated based on the anomaly type and severity represented by the associated process data.

7. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, In step S6, after reaching the preset convergence condition, the cloud outputs the globally optimal process parameters that have been verified historically, and outputs a proxy model constructed based on the updated sample set; wherein the convergence condition includes at least one of the following: The maximum value of the preset acquisition function is lower than a set threshold within the process parameter range; The number of iterations has reached the preset limit.

8. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 1, characterized in that, The sample set includes target task samples from the current target to be optimized, and source task samples from other targets; The preset algorithm is a multi-task Gaussian process regression algorithm, and the proxy model is a multi-task proxy model; Step S2 includes the following sub-steps: S21. When step S2 is executed for the first time, the cloud uses the source task samples to pre-train the multi-task Gaussian process and construct a multi-task proxy model. S22. The cloud updates the multi-task agent model based on the target task sample. In subsequent iterations, step S22 is only repeated, and step S21 is not executed.

9. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 8, characterized in that, The kernel function of the multi-task Gaussian process regression algorithm is: ,in: t and t' are task indices, where t=0 represents the source task and t=1 represents the target task; K task (t,t') is the task-related kernel, used to learn the correlation between the source task and the target task; K input (x,x') is the input kernel, used to learn the similarity of the process parameter region.

10. The parameter optimization method based on edge control layer and cloud collaboration as described in claim 8, characterized in that, Between step S21 and step S22, the following is also included: The cloud uses a space-filling design to select multiple process parameters as test points within the process parameter area, and sends the test points down to the edge control layer; The edge control layer performs production trials, acquires initial target task samples, and uploads them to the cloud. The cloud adds the initial target task sample to the sample set and uses it for model update in step S22.

Images