A Smart Injection Method Based on Lightweight Gradient Lifter and MPC

Abstract

The application relates to an intelligent injection method based on a light gradient booster and MPC, relates to the field of injection molding machine control, and comprises the following steps: collecting historical data and current operation input data in an injection process; training a gradient boosting model by using the historical data; inputting the current operation input data into the gradient boosting model to obtain a product quality prediction value; generating a prediction error based on the current operation input data and the product quality prediction value; generating a prediction state parameter based on the current operation input data; inputting the prediction error, the prediction state parameter and the product quality prediction value into a preset MPC to generate a future optimal control input sequence, and outputting the future optimal control input sequence to an injection molding machine to perform injection molding. The application has the effect of conveniently adapting different products or equipment for injection molding.

Description

Technical Field

[0001] This invention relates to the field of injection molding machine control, and in particular to an intelligent injection method based on a lightweight gradient lifter and MPC. Background Technology

[0002] Injection molding machines are key pieces of equipment that produce various plastic products or parts by heating plastic materials to a molten state and injecting them into a mold under high pressure to cool and solidify. With the iteration of industrial products and the upgrading of market demands, the application of lightweight and economical plastic products is becoming increasingly widespread, and the high efficiency and precision of injection molding technology have become core indicators of the core competitiveness of such equipment.

[0003] Because the injection molding process using injection molding machines is susceptible to disturbances from various complex factors, it can easily lead to decreased product quality consistency, increased scrap rates, and significantly increased production costs. Currently, when controlling injection molding machines, the traditional control model based on MPC (Model Predictive Control System) is generally used. This involves establishing a state-space model to predict output fluctuations, and then adjusting input parameters based on the error feedback of the output fluctuations to stabilize the production process.

[0004] When using MPC, it is necessary to explicitly define the state space model of the system. The determination of parameters relies more on traditional empirical methods, making model building difficult and lacking in versatility, thus making it inconvenient to adapt to different products or equipment. Summary of the Invention

[0005] To facilitate injection molding of different products or equipment, this invention provides an intelligent injection method based on a lightweight gradient lifter and MPC.

[0006] This invention provides an intelligent injection method based on a lightweight gradient lifter and MPC, employing the following technical solution:

[0007] A smart injection method based on a lightweight gradient booster and MPC includes:

[0008] S1: Collect historical data and current input data during the injection process;

[0009] S2: Use the historical data to train the gradient boosting model;

[0010] S3: Input the current running input data into the gradient boosting model to obtain the product quality prediction value;

[0011] S4: Generate a prediction error based on the current operating input data and the predicted product quality value;

[0012] S5: Generate predicted state parameters based on the current operating input data;

[0013] S6: Input the prediction error, the prediction state parameter and the product quality prediction value into the preset MPC to generate the future optimal control input sequence, and output the future optimal control input sequence to the injection molding machine for injection molding.

[0014] Optionally, the training method for the gradient boosting model includes:

[0015] S21: Retrieve process input parameters, material input parameters, and product quality data during the injection process based on the historical data;

[0016] S22: Combine the process input parameters with the material input parameters to form a comprehensive input parameter, and determine the output requirement parameters based on the product quality data;

[0017] S23: Train the model based on the input comprehensive parameters and the output requirement parameters to obtain the gradient boosting model.

[0018] Optionally, the method for generating the prediction error includes:

[0019] S41: Based on the current operating input data and the product quality prediction value, input the data into the preset MPC to obtain the internal prediction value;

[0020] S42: Calculate the difference between the internal predicted value and the product quality predicted value and use it as the prediction error.

[0021] Optionally, the method for generating the predicted state parameters includes:

[0022] S51: Retrieve the historical control input information and historical system response from the previous moment based on the current operating input data;

[0023] S52: Combine the current operating input data, the historical control input information, and the historical system response as the predicted state parameter.

[0024] Optionally, the method for generating the future optimal control input sequence includes:

[0025] S61: Retrieve the current time and prediction step size based on the predicted state parameters;

[0026] S62: Select the output weight matrix, input weight matrix, and equilibrium hyperparameters;

[0027] S63: Input the prediction error, the current time, the prediction step size, the predicted product quality value, the output weight matrix, the input weight matrix, and the balance hyperparameter into a preset MPC, and use a preset joint objective function to calculate and solve to obtain the future optimal control input sequence;

[0028] The joint objective function is:

[0029] ;

[0030] N is the length of the prediction time domain;

[0031] The internal predicted value;

[0032] The preset output expectation value;

[0033] It is the future optimal control input sequence;

[0034] Input the expected value for the future;

[0035] Q is the output weight matrix;

[0036] R is the input weight matrix;

[0037] This refers to the predicted quality value of the product.

[0038] α is the equilibrium hyperparameter;

[0039] t represents the current time;

[0040] k is the prediction step size.

[0041] Optionally, the selection methods for the output weight matrix, the input weight matrix, and the balancing hyperparameters include:

[0042] The output weight matrix, the input weight matrix, and the balancing hyperparameters are determined using a grid search method. The search range of the output weight matrix is ​​[10, 100], and the search step size of the output weight matrix is ​​10. The search range of the input weight matrix is ​​[0, 1], and the search step size of the input weight matrix is ​​0.1. The search range of the balancing hyperparameters is [0.1, 10], and the search step size of the balancing hyperparameters is 0.1.

[0043] In summary, the present invention has at least one of the following beneficial technical effects:

[0044] 1. By utilizing historical data from the injection molding process, a gradient boosting model is used to replace the explicit traditional mathematical model to predict future output fluctuations, thereby determining the optimal control input sequence for the future. This approach has the advantages of strong versatility and low model construction difficulty, making it easy to adapt to different products or equipment for injection molding.

[0045] 2. Introducing the prediction error of the gradient boosting model into MPC provides a dynamic correction mechanism for the predictions of both, thereby improving the adaptability and anti-interference ability of the control strategy to complex dynamic systems. Attached Figure Description

[0046] Figure 1 This is a flowchart of a smart injection method based on a lightweight gradient booster and MPC. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0048] A smart injection method based on a lightweight gradient booster and MPC is proposed. By collecting historical data and current operating input data of the injection molding process, a gradient booster model is trained to predict product quality. The model prediction value and current operating data are combined to generate prediction error and prediction state parameters. This information is input into a preset MPC and the optimal control input sequence for the future is solved through a joint objective function. Finally, the output is sent to the injection molding machine to achieve smart injection. It has the advantages of strong versatility and low model construction difficulty, and can be easily adapted to different products or equipment for injection molding.

[0049] Reference Figure 1 This invention discloses an intelligent injection method based on a lightweight gradient booster and MPC, comprising:

[0050] S1: Collect historical data and current input data during the injection process.

[0051] Historical data refers to various operational data accumulated by the injection molding machine during past production processes. This historical data includes process input parameters, material input parameters, and product quality data during the injection process. Process input parameters refer to the key adjustable process conditions during injection molding; material input parameters refer to the characteristic parameters related to the materials used in injection molding; and product quality data refers to the indicator data that measures the qualification level of the injection-molded product.

[0052] Process input parameters include barrel temperature, back pressure of the feedstock, feedstock rotation speed, injection speed, injection pressure, transition to holding pressure position, holding pressure, and holding time. Material input parameters include material type and melt temperature.

[0053] Historical data is generated through long-term recording by sensors (such as temperature sensors, pressure sensors, speed sensors, etc.) and data acquisition systems mounted on injection molding machines. It covers operational data of different product types, machine models and production environments, and the sources may include multiple devices from multiple customer factories.

[0054] Current operating input data refers to the process and material parameters collected in real time by the injection molding machine during the current production cycle. It is real-time data reflecting the current production status. Current operating input data is obtained by synchronously acquiring the current process and material parameters during the production process through sensors pre-installed on the injection molding machine and a real-time data acquisition module.

[0055] S2: Use historical data to train a gradient boosting model.

[0056] Among them, the gradient boosting model refers to the model used to predict product quality, and the gradient boosting model is also known as the LightGBM model.

[0057] By using historical data to train the model, a gradient boosting model can be obtained, which is convenient for subsequent use.

[0058] To further ensure the rationality of the gradient boosting model, it is necessary to perform further separate analysis and calculation on the gradient boosting model, which will be explained in detail through the steps shown below.

[0059] The training method for gradient boosting models includes the following steps:

[0060] S21: Retrieve process input parameters, material input parameters, and product quality data during the injection process based on historical data.

[0061] This includes retrieving process input parameters, material input parameters, and product quality data from historical data for convenient subsequent use.

[0062] S22: Combine process input parameters with material input parameters as comprehensive input parameters, and determine output requirement parameters based on product quality data.

[0063] The input parameters refer to the set of model inputs that integrate key process conditions and material property parameters that affect product quality during injection molding. The output parameters refer to the target parameters with product quality as the core, which are the target results to be predicted or optimized during model training.

[0064] By standardizing process and material input parameters—for example, numerical process parameters such as barrel temperature and injection pressure are normalized or standardized to a unified dimensional range; categorical parameters such as material type are encoded into a format recognizable by the model—a structured feature matrix is ​​formed. Each sample corresponds to a combination of process and material parameters, thus obtaining the comprehensive input parameters. Finally, the most critical indicators in the product quality data (such as core dimensional deviation) are directly selected as the output requirement parameters.

[0065] By determining the input comprehensive parameters and output demand parameters, it is easier to provide clear learning objectives for the subsequent training of the gradient boosting model, enabling the model to learn the nonlinear relationship between process, material factors and product quality, thus laying the foundation for subsequent quality prediction.

[0066] S23: Train the model based on the input comprehensive parameters and the output requirement parameters to obtain the gradient boosting model.

[0067] The algorithm uses the integrated input parameters as input and the output requirement parameters as the target, and iteratively trains weak learners through gradient boosting. In each training round, the model first generates a prediction result based on the current combination of weak learners, calculates the error between the prediction value and the output requirement parameters, then trains a new weak learner to fit the error, and controls the weight of the new learner through the learning rate to gradually reduce the overall error until the model performance meets the target, and finally obtains the gradient boosting model, which is convenient for subsequent use.

[0068] S3: Input the current running input data into the gradient boosting model to obtain the product quality prediction value.

[0069] Among them, the product quality prediction value refers to the prediction result of product quality under the current production conditions. The product quality prediction value can be any one or a combination of quality prediction results such as product size deviation, weight fluctuation or defect probability. The product quality prediction value is used to judge the production results in advance and support subsequent control decisions.

[0070] By inputting the current running input data into the gradient boosting model, the model can quickly calculate and output the corresponding product quality prediction value for convenient subsequent use.

[0071] S4: Generate prediction error based on the current operating input data and the product quality prediction value.

[0072] The prediction error refers to the difference between the prediction result generated by the preset MPC and the prediction result generated by the gradient boosting model during the intelligent injection control process.

[0073] By analyzing the current operating input data and the predicted product quality values, a prediction error is generated for convenient subsequent use.

[0074] To further ensure the reasonableness of the prediction error, it is necessary to perform a more detailed separate analysis and calculation of the prediction error, which will be explained in detail through the steps shown below.

[0075] The method for generating prediction error includes the following steps:

[0076] S41: Input the current running input data and product quality prediction values ​​into the preset MPC to obtain internal prediction values.

[0077] Here, MPC refers to Model Predictive Control System, and the internal predicted value refers to the product quality prediction result generated by MPC based on its own dynamic prediction logic. The specific meaning of the internal predicted value is consistent with the specific meaning of the product quality predicted value. That is, when the product quality predicted value is set by the operator as the product size deviation, the internal predicted value is also the product size deviation.

[0078] By synchronizing the current operating input data and the product quality prediction value into the preset MPC, the product quality change trend over a future period under the current input conditions can be obtained, and the prediction result corresponding to the current control cycle can be used as the internal prediction value for convenient subsequent use.

[0079] S42: Calculate the difference between the internal forecast value and the product quality forecast value and use it as the forecast error.

[0080] Specifically, by calculating the difference between the internal predicted value and the product quality predicted value, and using the calculation result as the prediction error, the accuracy of the obtained prediction error is improved.

[0081] S5: Generate predicted state parameters based on the current running input data.

[0082] Among them, the predicted state parameters refer to the comprehensive set of parameters used to describe the current and associated historical operating states of the injection molding system, and are the core basis for MPC to predict and optimize future output.

[0083] By analyzing the current input data, predicted state parameters are generated for convenient subsequent use.

[0084] To further ensure the rationality of the predicted state parameters, it is necessary to perform further separate analysis and calculation on the predicted state parameters, which will be explained in detail through the steps shown below.

[0085] The method for generating predicted state parameters includes the following steps:

[0086] S51: Retrieve historical control input information and historical system response from the previous moment based on the current operating input data.

[0087] Among them, historical control input information refers to the control parameter adjustment values ​​executed by the injection molding machine in the previous control cycle. Historical control input information includes information such as the injection speed setting value, holding pressure adjustment amount, and material storage back pressure modification value at the previous moment.

[0088] Historical system response refers to the actual output feedback generated by the system under the action of historical control inputs during the previous control cycle. Historical system response includes the measured values ​​of product quality at the corresponding time (such as dimensional deviations and weight fluctuations) or changes in system operating status (such as actual change curves of pressure and temperature), reflecting the dynamic response results of the system to past control inputs.

[0089] The current time is determined by the current operating input data, which is then used to locate the previous control cycle. Finally, the preset historical database is queried to obtain historical control input information and historical system responses.

[0090] The historical database stores the control parameter adjustment values ​​executed by the injection molding machine and the actual output feedback generated by the system in real time.

[0091] S52: Combine the current operating input data, historical control input information, and historical system response as predicted state parameters.

[0092] Specifically, by standardizing the current operating input data, historical control input information, and historical system responses to ensure a unified format, and then integrating them in a structured manner according to the time dimension and logical association, a parameter matrix containing multi-dimensional information is formed. Each row of data corresponds to a control cycle, covering the complete chain of "current state - historical decision - historical feedback", thereby obtaining the predicted state parameters and improving the accuracy of the obtained predicted state parameters.

[0093] S6: Input the prediction error, prediction state parameters and product quality prediction values ​​into the preset MPC to generate the future optimal control input sequence, and output the future optimal control input sequence to the injection molding machine for injection molding.

[0094] Among them, the future optimal control input sequence refers to the optimal process parameter adjustment scheme generated by MPC through optimization calculation for multiple future control cycles, which includes the set values ​​of key control parameters such as injection speed, holding pressure, and barrel temperature at each future moment.

[0095] By inputting the prediction error, prediction state parameters, and product quality prediction values ​​into a preset MPC, the optimal control input sequence for the future is obtained through analysis and calculation. This optimal control input sequence is then output to the injection molding machine for injection molding, ensuring that the future product quality is consistently close to the expected target while avoiding drastic fluctuations in control inputs. Furthermore, by using a gradient boosting model instead of an explicit traditional mathematical model, future output fluctuations are predicted. This method has the advantages of strong versatility and low model construction difficulty, making it easy to adapt to different products or equipment for injection molding.

[0096] To further ensure the rationality of the future optimal control input sequence, it is necessary to perform further separate analysis and calculation on the future optimal control input sequence, which will be explained in detail through the steps shown below.

[0097] The method for generating the optimal control input sequence in the future includes the following steps:

[0098] S61: Retrieve the current time and prediction step size based on the predicted state parameters.

[0099] Here, "current moment" refers to the specific point in time when the system makes control decisions or processes data. In injection molding production scenarios, it typically corresponds to the start time of a certain control cycle and serves as a benchmark reference point for time series data. The prediction step size refers to the number of future prediction cycles set in the MPC (Multi-Process Control), i.e., the number of control steps the system needs to predict and plan in advance (e.g., 5 future control cycles). Its length determines the foresight of the optimization decision; a step size that is too short may lead to control lag, while a step size that is too long will increase computational complexity.

[0100] By extracting time-identifying information such as timestamps and production cycle numbers from the predicted state parameters, this identifier is bound to the acquisition time of the current operating input data, directly corresponding to the current control cycle of the system, thus determining the current time. Then, the preset prediction step size parameter in the MPC system is read and used as the prediction step size, thereby providing time-dimensional boundary conditions for generating the optimal control input sequence in the future.

[0101] S62: Select the output weight matrix, input weight matrix, and balancing hyperparameters.

[0102] The output weight matrix is ​​a matrix parameter in the MPC joint objective function used to measure the importance of the output error. The larger the value of the output weight matrix, the heavier the penalty for the deviation between the predicted product quality value and the expected target.

[0103] The input weight matrix is ​​a matrix parameter that measures the importance of changes in the control input. The larger the value of the input weight matrix, the heavier the penalty for drastic adjustments to the control input, which is used to avoid system instability caused by frequent or large adjustments.

[0104] The balancing hyperparameter is a parameter that adjusts the influence of the prediction error penalty term. It is used to balance the consistency between the predictions within the MPC model and the predictions of the gradient boosting model. The larger the value of the balancing hyperparameter, the smaller the deviation between the two predictions.

[0105] The output weight matrix, input weight matrix, and equilibrium hyperparameters are determined using a grid search method.

[0106] The search range of the output weight matrix is ​​[10, 100], and the search step size of the output weight matrix is ​​10. The search range of the input weight matrix is ​​[0, 1], and the search step size of the input weight matrix is ​​0.1. The search range of the balancing hyperparameters is [0.1, 10], and the search step size of the balancing hyperparameters is 0.1.

[0107] By employing a grid search method to select the output weight matrix, input weight matrix, and balancing hyperparameters, and by setting the search range and search step size for each of the output weight matrix, input weight matrix, and balancing hyperparameters, the accuracy of the obtained output weight matrix, input weight matrix, and balancing hyperparameters is improved, and their subsequent use is facilitated.

[0108] S63: Input the prediction error, current time, prediction step size, predicted product quality value, output weight matrix, input weight matrix, and balance hyperparameters into the preset MPC, and use the preset joint objective function to calculate and solve to obtain the optimal control input sequence in the future.

[0109] The joint objective function is:

[0110] ;

[0111] N is the length of the prediction time domain;

[0112] These are internally predicted values;

[0113] The preset output expectation value;

[0114] It is the optimal control input sequence for the future;

[0115] Input the expected value for the future;

[0116] Q is the output weight matrix;

[0117] R is the input weight matrix;

[0118] For product quality prediction;

[0119] α is the equilibrium hyperparameter;

[0120] t represents the current time.

[0121] k is the prediction step size.

[0122] Since the internal prediction value is the prediction value made by MPC, and the product quality prediction value is the prediction value made by the gradient boosting model, the prediction error introduced into MPC by the gradient boosting model is the same as the prediction error in the joint objective function. .

[0123] The output expectation value refers to the product quality target value preset by the system during the injection molding production process, that is, the ideal product quality indicator that is expected to be achieved through control strategies. The output expectation value is preset by the operator based on actual needs.

[0124] Future input expectation values ​​refer to the ideal control input target values ​​preset in MPC for multiple future control cycles, i.e., the process parameter benchmarks that the system is expected to achieve at various future times. Future input expectation values ​​are preset by the operator based on actual needs.

[0125] By inputting the prediction error, current time, prediction step size, predicted product quality, output weight matrix, input weight matrix, and balancing hyperparameters into a preset MPC, and using a preset joint objective function to calculate and solve for the prediction error, current time, prediction step size, predicted product quality, output weight matrix, input weight matrix, and balancing hyperparameters, the future optimal control input sequence is obtained, thereby improving the accuracy of the obtained future optimal control input sequence.

[0126] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A smart injection method based on light gradient boosting machine and MPC, characterized in that, Comprise: S1: collect historical data and current running input data during injection process; S2: use the historical data to train gradient boosting model; S3: input the current running input data into the gradient boosting model to obtain product quality prediction value; S4: generate prediction error based on the current running input data and the product quality prediction value; S5: generate prediction state parameter based on the current running input data; S6: input the prediction error, the prediction state parameter and the product quality prediction value into the preset MPC to generate future optimal control input sequence, and output the future optimal control input sequence to injection molding machine for injection molding; The training method of the gradient boosting model comprises: S21: based on the historical data, the process input parameters, material input parameters and product quality data during injection process are called; S22: the process input parameters and the material input parameters are combined as input comprehensive parameters, and the output required parameters are determined according to the product quality data; S23: based on the input comprehensive parameters and the output required parameters, model training is carried out to obtain the gradient boosting model.

2. The intelligent injection method based on light gradient boosting machine and MPC according to claim 1, wherein, The generation method of the prediction error comprises: S41: based on the current running input data and the product quality prediction value, the internal prediction value is obtained by inputting into the preset MPC; S42: the difference between the internal prediction value and the product quality prediction value is calculated and taken as the prediction error.

3. The intelligent injection method based on light gradient boosting machine and MPC according to claim 2, characterized in that, The generation method of the prediction state parameter comprises: S51: based on the current running input data, the historical control input information and the historical system response at the last time are called; S52: the current running input data, the historical control input information and the historical system response are combined and taken as the prediction state parameter.

4. The intelligent injection method based on light gradient boosting machine and MPC according to claim 3, wherein, The generation method of the future optimal control input sequence comprises: S61: based on the prediction state parameter, the current time and the prediction step are called; S62: the output weight matrix, the input weight matrix and the balance hyperparameter are selected; S63: the prediction error, the current time, the prediction step, the product quality prediction value, the output weight matrix, the input weight matrix and the balance hyperparameter are input into the preset MPC, and the preset joint objective function is used for calculation and solution to obtain the future optimal control input sequence; Wherein, the joint objective function is: ； N is the length of prediction time domain; is the internal prediction value; is a preset output expectation value; is the future optimal control input sequence; a desired value for a future input is preset; Q is the output weight matrix; R is the input weight matrix; is the product quality prediction value; Alpha is the balance hyperparameter; T is the current time; K is the prediction step.

5. The intelligent injection method based on light gradient boosting machine and MPC according to claim 4, wherein, The selection method of the output weight matrix, the input weight matrix and the balance hyperparameter comprises: The output weight matrix, the input weight matrix and the balance hyperparameter are determined based on a grid search method, wherein a search range of the output weight matrix is [10, 100], a search step of the output weight matrix is 10, a search range of the input weight matrix is [0, 1], a search step of the input weight matrix is 0.1, a search range of the balance hyperparameter is [0.1, 10], and a search step of the balance hyperparameter is 0.1.

Images