Aluminum wheel die-casting mold temperature closed-loop regulation system and method

Abstract

The application discloses an aluminum wheel die-casting mold temperature closed-loop regulation and control system and method, belongs to the field of aluminum alloy die-casting forming and intelligent control technology, and comprises a data acquisition module, specifically temperature sensors, flow sensors and pressure sensors arranged at key positions of a mold, a prediction module in communication connection with the data acquisition module, a CNN-LSTM composite neural network model adopted by the prediction module, an optimization decision module in communication connection with the prediction module and providing a prediction value, an integrated tree structure Pareto estimator TPE adopted by the optimization decision module, an execution control module provided with optimal decisions by the optimization decision module, the execution control module being a PLC controller, and a feedback learning module arranged for online updating of the prediction model and the optimization strategy according to control effect data. The application can effectively reduce energy consumption and manual intervention cost, and has remarkable practical value and economic value.

Description

Technical Field

[0001] This invention relates to the field of aluminum alloy die casting and intelligent control technology, and in particular to a closed-loop temperature control system and method for aluminum wheel die casting molds based on machine learning prediction and optimization. Background Technology

[0002] In the field of automotive lightweighting and high-end manufacturing, aluminum wheels have become core components due to their combination of high strength and lightweight characteristics. Low-pressure casting is the mainstream production process for them. The uniformity and stability of the mold temperature field directly determine the solidification rate, microstructure and internal quality of the casting. If the mold temperature is not properly controlled, it is easy to cause defects such as shrinkage cavities, shrinkage porosity, deformation and surface cracks in the casting, which seriously affect the mechanical properties and service life of the aluminum wheel. Therefore, precise control of mold temperature is the core technical link in the low-pressure casting process of aluminum wheels.

[0003] Currently, temperature control technologies for aluminum wheel die-casting molds in the industry are mainly divided into four categories: empirical control, PID control, preset program control, and offline optimization. Each method has significant technical limitations. Empirical control, as a traditional method, relies on skilled workers manually adjusting the opening and on / off duration of cooling water valves by observing intuitive characteristics such as the surface color and forming state of the casting. While it can maintain basic production under fixed conditions, this method is highly subjective, lacks standardized operating procedures, and suffers from significant response lag in manual judgment and adjustment, making it unsuitable for the high-speed, high-precision production demands of modern assembly lines. PID control, using a combination of proportional, integral, and derivative parameters to correct temperature deviations, is widely used in simple industrial control scenarios. However, the aluminum wheel die-casting process has complex characteristics of strong nonlinearity, large time delay, and multi-variable coupling. Mold temperature is affected by various factors such as fluctuations in raw material composition, changes in ambient temperature and humidity, and equipment wear and tear. The fixed parameters of traditional PID controllers are insufficient to cope with dynamic changes in operating conditions, requiring frequent manual parameter tuning. Neither adaptability nor control accuracy can meet the requirements of high-end casting production. Pre-programmed control is an open-loop control method that pre-sets cooling timing and parameter combinations according to a predetermined process procedure. It is simple to operate and highly repeatable, but lacks dynamic adjustment capabilities. When faced with unexpected situations such as batch differences in raw materials or disturbances in the workshop environment, it cannot correct the control strategy in real time, easily leading to mold temperature deviations from the optimal range and consequently causing fluctuations in casting quality. Offline optimization methods, on the other hand, use numerical simulation software such as ProCAST and MAGMA, or experimental design methods such as orthogonal experiments and response surface methodology, to statically optimize process parameters before production. While this can obtain the theoretically optimal parameter combination, the optimization cycle is long, the experimental and simulation costs are high, and the optimization results are a fixed set of parameters, unable to respond in real time to dynamic changes in operating conditions during production, making it difficult to guarantee optimal mold temperature control throughout the entire production cycle.

[0004] Further analysis reveals five core technical deficiencies in existing mold temperature control solutions. First, insufficient predictive capability: all methods rely on passive adjustments based on current temperature deviations, lacking the ability to anticipate future trends in the mold temperature field, resulting in control lag and significant temperature fluctuations. Second, slow optimization response: traditional methods cannot quickly solve multi-parameter collaborative optimization problems within the stringent time constraints of production cycles, hindering true real-time optimization control. Third, lack of self-learning capability: control strategies remain fixed, unable to extract process patterns from massive amounts of historical production data and automatically improve control logic, making continuous iteration of control effects difficult. Fourth, low system integration: data acquisition, data analysis and decision-making, and actuator control are independent processes, creating barriers to information flow and lacking efficient collaborative mechanisms, limiting overall system response speed. Fifth, poor adaptability to operating conditions: existing methods struggle to quickly adjust control parameters and strategies in the face of product specification changes, equipment aging, and sudden changes in environmental temperature and humidity, easily leading to control failures. With the automotive industry demanding increasingly stringent quality requirements for aluminum wheels and continuously improving the level of intelligent manufacturing, existing mold temperature control technologies can no longer meet the production needs of high precision, high stability, and high adaptability. There is an urgent need for a closed-loop control system and method with predictive capabilities, real-time optimization, autonomous learning, and multi-stage collaborative characteristics to overcome many pain points of traditional technologies and achieve precise, dynamic, and intelligent control of the temperature of aluminum wheel die-casting molds. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide a closed-loop temperature control system and method for aluminum wheel die-casting molds based on machine learning prediction and optimization. It seeks to address the shortcomings of existing aluminum wheel die-casting mold temperature control methods, such as the inability to perform high-precision dynamic prediction of the mold temperature field; the difficulty of completing real-time optimization of multiple parameters within limited production cycles using traditional control methods; the rigidity of control strategies, which cannot adapt to changes in operating conditions; and the independent data acquisition, decision-making, and execution stages, resulting in slow system response speeds, large mold temperature fluctuations, poor product quality consistency, and high scrap rates. This invention can effectively achieve precise temperature control and improve yield.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a closed-loop temperature control system for aluminum wheel die-casting molds based on machine learning prediction and optimization, including a data acquisition module, specifically temperature sensors, flow sensors, and pressure sensors arranged at key positions of the mold. A prediction module is communicatively connected to the data acquisition module. The prediction module employs... The composite neural network model has an optimization decision module connected to the prediction module and provides the predicted value. The optimization decision module adopts an integrated tree-structured Pareto estimator (TPE). The optimization decision module provides the optimal decision to the execution control module, which is a PLC controller. It also includes a feedback learning module for online updates of the prediction model and optimization strategy based on control effect data.

[0007] A further improvement to the technical solution of this invention lies in: The composite neural network model consists of the following layers from input to output: input layer, Conv1D layer, MaxPooling1D layer, LSTM layer, Dropout layer, LSTM layer, Flatten layer, Dense layer, Dense layer, and output layer.

[0008] A further improvement to the technical solution of this invention lies in: The input layer of the composite neural network model adopts an ultra-long time window design, that is, a 1250 time step retrieval window design, which is specifically designed to capture the thermal inertia characteristics and complete process cycle of the die casting process.

[0009] A further improvement of the technical solution of this invention lies in the following: the input of the tree-structured Pareto estimator TPE is the manually set standard process template data, namely the ideal temperature curve and the cooling parameters of the previous round; the output is the optimal control parameters, wherein the cooling parameters of the previous round and the optimal control parameters are both parameters for controlling the cooling equipment; at the same time, a hot start mechanism is introduced into TPE, that is, the process time is divided into two stages: controllable and uncontrollable. A differentiated processing strategy is adopted in TPE optimization: the baseline process prediction data is reused for the uncontrollable stage, and parameter optimization search is performed for the controllable stage, thereby achieving a dual improvement in optimization efficiency and process stability.

[0010] A closed-loop control method for die-casting mold temperature of aluminum wheels based on machine learning prediction and optimization includes the following steps:

[0011] Step 1: Collect historical data and use historical data to train. Composite neural network model;

[0012] Step 2: Use the acquisition module to collect relevant data about the mold in real time. After data preprocessing, input the data into the trained database. A composite neural network model outputs predicted temperature values ​​for each measuring point in the future.

[0013] Step 3: Based on the predicted temperature value and prediction results, if an anomaly is found in the prediction results, the TPE module is intervened according to the ideal temperature curve. The TPE module provides an initial cooling parameter prediction combination. This initial cooling parameter prediction combination is converted into time series data and input into the prediction module to predict whether the prediction result after adjusting this initial cooling parameter combination can meet the requirements of the ideal temperature curve. If not, this process is iterated until the optimal control parameters that meet the ideal temperature curve are given.

[0014] Step 4: Input the optimal control parameters into the execution control module, and use the PLC controller to drive the electric regulating valve and variable frequency water pump according to the optimal control parameters to accurately adjust the cooling water flow and on / off timing.

[0015] Step 5: Collect data. The data is fed back into the data collection system of the prediction model to complete the closed loop and synchronously and adaptively adjust the weight coefficients of the objective function. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization.

[0016] A further improvement to the technical solution of the present invention is as follows: Step 1 is specifically as follows:

[0017] Step 1.1: Collect production data over a period of time, including: temperature time-series data of each measuring point within each casting cycle, with a sampling interval of 1 second; corresponding cooling process parameters, specifically including flow rate, temperature, and on / off timing; product quality inspection results, specifically including pass rate and defect type; and environmental parameters, specifically including workshop temperature and humidity.

[0018] Step 1.2: Data Cleaning and Feature Engineering: Remove abnormal data during equipment failure; perform receipt window processing on temperature data with a window width of 1250 steps; perform cumulative calculation on flow data to enhance flow characteristics; normalize all features to the [0,1] interval;

[0019] Step 1.3: Training A composite neural network model: The input layer receives 1250 seconds of historical data; the CNN layer contains two convolutional blocks to extract spatial features; the LSTM layer contains two layers with 128 neurons to capture temporal dependencies; and the output layer predicts the temperature value for the next cycle. It employs the Adam optimizer with a learning rate of 0.001, a batch size of 64, and is trained for 30 epochs. It achieves [performance / achievements] on the validation set. Save the model afterwards.

[0020] A further improvement to the technical solution of the present invention is as follows: Step 2 is specifically as follows:

[0021] Step 2.1: Establish a real-time data stream processing pipeline, configure the data acquisition module to acquire all sensor data at a frequency of 1Hz; implement edge computing preprocessing, including data verification, outlier removal, and missing value interpolation; construct a sliding time window and maintain a cache of historical data for the most recent 1250 seconds; push the preprocessed data to the prediction module in real time.

[0022] Step 2.2: Perform online temperature prediction by inputting the current time window data into the deployed... Model; Model outputs predicted temperature values ​​for each measuring point within the next 245 seconds; Calculates the deviation vector between the predicted temperature and the set target temperature. When any measuring point deviates When this happens, the optimization decision module is triggered;

[0023] Step 2.3: Implement dynamic correction of the prediction results, use Kalman filtering to fuse the predicted and measured values ​​to improve the reliability of the prediction; dynamically adjust the prediction confidence based on the error statistics of the most recent 10 predictions; when the prediction error exceeds the threshold for 5 consecutive predictions, trigger the online fine-tuning mechanism of the model.

[0024] A further improvement to the technical solution of the present invention is as follows: Step 3 is specifically as follows:

[0025] Step 3.1: Construct the Bayesian optimization search space; define the decision variable space. This includes i key cooling control parameters: delay_time and opening_time; and setting the objective function. This refers to the root mean square error between the predicted temperature and the set temperature curve; defining the early stop threshold. The optimization will automatically terminate when the required accuracy is reached.

[0026] Step 3.2: Execute the TPE algorithm for intelligent search; implement the TPE sampler using the Optuna framework, randomly sample 5 parameter combinations during the initialization phase, and evaluate their RMSE values ​​as prior knowledge; the TPE algorithm automatically sorts historical observation points according to the target value, selects the best-performing observation points to construct the excellent sample set L(x), and constructs the poor sample set G(x) from the rest; fit kernel density estimation models to L(x) and G(x) respectively, and construct probability density functions. and Improve by maximizing expectations Select the next sampling point; set the maximum number of evaluations to 25, and terminate when RMSE < 1.0 or the maximum number of evaluations is reached;

[0027] Step 3.3: Implement a downsampling acceleration strategy, downsampling the generated 1Hz high-frequency process data by a factor of 10 to convert it into 0.1Hz data, significantly reducing... The model's input data volume; after downsampling, it retains complete casting cycle data for 2450 time steps to ensure that key process features are not lost; by downsampling, the single temperature prediction time is shortened from 30 seconds to less than 3 seconds, enabling the TPE algorithm to complete multiple iterative evaluations within the production cycle; an incremental evaluation strategy is adopted, and the early stop conditions are checked immediately after each test to avoid unnecessary calculations;

[0028] Step 3.4: Generate optimized control parameters; After TPE optimization is completed, extract the optimal parameter combination best_x and update it in the process_data dictionary; Convert the optimized parameters into a complete time-series control sequence using the Creating_time_series_data function, including the flow rate settings for each cooling water line and precise switching timing, with the flow rate settings ranging from 5-80 L / min; Add feature engineering processing to calculate the cumulative flow actual_flow_V for accurate model prediction; Save the optimal parameters to a JSON file for easy subsequent calling and traceability;

[0029] Step 3.5: Verify the optimization effect; re-execute the complete prediction process using the optimal parameters and calculate the final RMSE value for verification; evaluate the maximum absolute error (MAE) and root mean square error (RMSE) of the predicted temperature versus the set temperature curve using the calculate_errors function; record the convergence curve of the optimization process and analyze the performance improvement of the TPE algorithm in the first 5 random explorations and the subsequent 20 Bayesian optimizations; calculate the total optimization time to ensure that the real-time control requirements are met.

[0030] A further improvement to the technical solution of the present invention is as follows: Step 4 is specifically as follows:

[0031] Step 4.1: Issue and execute control commands; encode the optimal control parameters optimized by TPE into Modbus RTU protocol frames; send precise opening commands to each electric regulating valve via PLC, with a resolution of 0.1%; coordinate the control of the variable frequency water pump to maintain stable system pressure. It records the execution status and response time of instructions in real time.

[0032] A further improvement to the technical solution of the present invention is as follows: Step 5 is as follows:

[0033] Step 5.1: Monitor and evaluate the control effect and performance; collect temperature response data after control execution, calculate the root mean square error between the actual temperature and the target temperature; evaluate the temperature convergence time and overshoot; statistically analyze the convergence speed and calculation time of the TPE algorithm; record cooling water consumption and energy consumption data for cost analysis.

[0034] Step 5.2: Implement online updates for Bayesian optimization; store the parameter-effect pairs of each control operation in the empirical database; update the prior distribution of TPE using new data every 30 casting cycles; employ transfer learning to transfer optimization experience from similar working conditions to new tasks; adaptively adjust the weight coefficients of the objective function based on the cumulative control effect. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization.

[0035] Step 5.3: Construct a knowledge graph to assist decision-making; extract high-frequency optimal parameter combinations in the TPE optimization process and construct a process knowledge base; establish a mapping relationship between "temperature deviation and optimal parameters" to accelerate parameter selection for similar operating conditions; analyze the correlation and coupling relationships between parameters to simplify the search space dimension; generate an optimization report to visualize the parameter distribution and convergence process.

[0036] Due to the adoption of the above technical solution, the technical progress achieved by this invention is: by carrying out... The model enables real-time, high-precision prediction of mold temperature change trends with a prediction error of less than 5℃. It achieves the advantage of predicting mold temperature fluctuations in advance, breaking the limitation of traditional control that can only passively adjust based on the current temperature deviation. It effectively eliminates control lag and provides a reliable predictive basis for precise mold temperature control.

[0037] By combining intelligent optimization algorithms, the collaborative optimization of multiple cooling process parameters can be completed within seconds, achieving the advantage of rapid response to production cycle time. This solves the pain point that traditional methods cannot solve multi-parameter optimization problems in real time, significantly improving the system response speed and mold temperature control accuracy, and ensuring that the mold temperature field is always maintained in the optimal range.

[0038] By endowing the system with online learning capabilities, it can extract process patterns from historical production data, achieving the advantage of adapting to changes in operating conditions and product specifications. This overcomes the shortcomings of traditional fixed control strategies, and allows for continuous iterative optimization of control logic. Even in the face of raw material fluctuations, equipment status changes, environmental disturbances, and other situations, it can quickly adjust control strategies to ensure stable control performance.

[0039] By constructing an integrated closed-loop control architecture of "perception-prediction-optimization-execution-feedback," efficient collaboration among all stages is achieved, resulting in smooth information flow and timely response throughout the entire process. This solves the problems of low integration and poor collaboration in traditional systems, reducing mold temperature fluctuations by more than 60%. This not only significantly reduces casting defects such as shrinkage cavities, porosity, and deformation caused by mold temperature imbalance, but also improves product qualification rates. It effectively reduces energy consumption and manual intervention costs, and has significant practical and economic value as well as broad industry application prospects. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the structure of a closed-loop temperature control system for aluminum wheel die-casting molds based on machine learning prediction and optimization according to the present invention.

[0042] Figure 2 This is the present invention. Schematic diagram of the composite neural network model structure;

[0043] Figure 3 This is a schematic diagram of the optimization and control algorithm of this invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to embodiments:

[0045] like Figure 1 The diagram shows a structural schematic of a closed-loop temperature control system for aluminum wheel die-casting molds based on machine learning prediction and optimization. It includes a data acquisition module, specifically temperature sensors, flow sensors, and pressure sensors arranged at key locations on the mold. The system is characterized by a prediction module connected to the data acquisition module, which employs… The composite neural network model has an optimization decision module connected to the prediction module and provides the predicted value. The optimization decision module adopts an integrated tree-structured Pareto estimator (TPE). The optimization decision module provides the optimal decision to the execution control module, which is a PLC controller. It also includes a feedback learning module for online updates of the prediction model and optimization strategy based on control effect data.

[0046] Composite neural network models, such as Figure 2 As shown, the sequence from input to output is as follows:

[0047] Input layer: Shape (1250, 19) (1250 time steps, 19 features)

[0048] Conv1D layer: 32 filters, convolution kernel size of 2, activation function of ReLU.

[0049] MaxPooling1D: Pooling size is 2.

[0050] LSTM layer: 50 units, returns a sequence (True).

[0051] Dropout layer: Drop rate is 0.2%.

[0052] LSTM layer: 50 units, returns a sequence (False).

[0053] Flatten layer: Flattens the output.

[0054] Dense layer: 50 neurons, activation function ReLU.

[0055] Dense layer: 25 neurons, activation function ReLU.

[0056] Output layer: 1 neuron, no activation function (regression problem).

[0057] The receipt window is designed as a key input parameter for the temperature prediction model.

[0058] Based on the time-delay characteristics of low-pressure cast aluminum wheels, an innovative 1250-time-step feedback window design was adopted in the temperature prediction model. This window length covers approximately five cycles from filling to cooling, effectively solving the modeling challenge of the cumulative effect of delayed heat conduction during casting and significantly improving the accuracy of terminal temperature prediction. It accurately captures the delayed cumulative effect of temperature changes, thereby achieving precise prediction of the terminal temperature during the casting process.

[0059] like Figure 3 As shown, a closed-loop control method for aluminum wheel die-casting mold temperature based on machine learning prediction and optimization is implemented according to the above system. The specific steps are as follows:

[0060] Step 1: Collect historical data and use historical data to train. Composite neural network model;

[0061] Step 1.1: Collect production data over a period of time, including: temperature time-series data of each measuring point within each casting cycle, with a sampling interval of 1 second; corresponding cooling process parameters, specifically including flow rate, temperature, and on / off timing; product quality inspection results, specifically including pass rate and defect type; and environmental parameters, specifically including workshop temperature and humidity.

[0062] Step 1.2: Data Cleaning and Feature Engineering: Remove abnormal data during equipment failure; perform receipt window processing on temperature data with a window width of 1250 steps; perform cumulative calculation on flow data to enhance flow characteristics; normalize all features to the [0,1] interval;

[0063] Step 1.3: Training The composite neural network model consists of an input layer that receives 1250 seconds of historical data, a CNN layer containing two convolutional blocks to extract spatial features, an LSTM layer containing two layers of 128 neurons to capture temporal dependencies, and an output layer that predicts the temperature value for the next cycle. The model is trained for 30 epochs using the Adam optimizer with a learning rate of 0.001 and a batch size of 64. The model is saved after achieving a MAE of <5℃ on the validation set.

[0064] Step 2: Use the acquisition module to collect relevant data about the mold in real time. After data preprocessing, input the data into the trained database. A composite neural network model outputs predicted temperature values ​​for each measurement point in the future. During training, an early stopping strategy and dropout ratio suitable for the characteristics of die-casting data are configured. The dropout ratio is set to dropout_rate=0.2, and the early stopping mechanism automatically checks the availability of the validation set during the first training epoch and automatically switches the monitoring mode based on the detection result.

[0065] Validation set available → Validate loss monitoring mode;

[0066] Validation set unavailable → Training loss monitoring mode;

[0067] Automatic switching of monitoring metrics:

[0068] In validation mode: Decisions are made based on test set loss;

[0069] In training mode: Decisions are made based on the training set loss.

[0070] Step 2.1: Establish a real-time data stream processing pipeline, configure the data acquisition module to acquire all sensor data at a frequency of 1Hz; implement edge computing preprocessing, including data verification, outlier removal, and missing value interpolation; construct a sliding time window and maintain a cache of historical data for the most recent 1250 seconds; push the preprocessed data to the prediction module in real time.

[0071] Step 2.2: Perform online temperature prediction by inputting the current time window data into the deployed... Model; Model outputs predicted temperature values ​​for each measuring point within the next 245 seconds; Calculates the deviation vector between the predicted temperature and the set target temperature. When any measuring point deviates When this happens, the optimization decision module is triggered;

[0072] Step 2.3: Implement dynamic correction of the prediction results. Use Kalman filtering to fuse predicted and measured values ​​to improve prediction reliability. Dynamically adjust the prediction confidence level based on the error statistics of the most recent 10 predictions. When the prediction error exceeds a threshold for 5 consecutive predictions, trigger the online fine-tuning mechanism. The fine-tuning mechanism involves... The predicted temperature is weighted and fused with the actual temperature measured by the sensor. The Kalman filter automatically adjusts the weight ratio by calculating the confidence level (uncertainty) of the two data points.

[0073] When the prediction model has high accuracy, it relies more on predicted values; when the measurement data has high reliability, it relies more on measured values. Error statistics from the last 10 predictions are used to adaptively adjust the confidence parameter based on the prediction accuracy. Triggering condition: Fine-tuning is performed when prediction performance remains poor. The method is to use the latest production data to... The model is retrained quickly. A lightweight design is used, adjusting only the parameters of the last few layers to avoid complete retraining; this ensures the system has both predictive accuracy and measurement precision, creating complementary advantages.

[0074] Step 3: Based on the predicted temperature value and the prediction results, if an anomaly is detected, the TPE module is intervened according to the ideal temperature curve. The TPE module provides an initial cooling parameter prediction combination, which is converted into time-series data and input into the prediction module. The module predicts whether the predicted result after adjusting this initial cooling parameter combination can meet the requirements of the ideal temperature curve. If not, this process iterates until the optimal control parameters that meet the ideal temperature curve are given. Introducing a hot-start mechanism into the TPE module can save optimization control time in this area. For the process characteristics of the cooling process of low-pressure cast aluminum wheels, the process time is divided into two stages: controllable and uncontrollable. A differentiated processing strategy is adopted in TPE optimization: the baseline process prediction data is reused for the uncontrollable stage, while parameter optimization search is performed for the controllable stage, achieving a dual improvement in optimization efficiency and process stability.

[0075] Step 3.1: Construct the Bayesian optimization search space; define the decision variable space. This includes i key cooling control parameters: delay_time and opening_time; and setting the objective function. This refers to the root mean square error between the predicted temperature and the set temperature curve; defining the early stop threshold. The optimization will automatically terminate when the required accuracy is reached.

[0076] Step 3.2: Execute the TPE algorithm for intelligent search; implement the TPE sampler using the Optuna framework, randomly sample 5 parameter combinations during the initialization phase, and evaluate their RMSE values ​​as prior knowledge; the TPE algorithm automatically sorts historical observation points according to the target value, selects the best-performing observation points to construct the excellent sample set L(x), and constructs the poor sample set G(x) from the rest; fit kernel density estimation models to L(x) and G(x) respectively, and construct probability density functions. and Improve by maximizing expectations Select the next sampling point; set the maximum number of evaluations to 25, and terminate when RMSE < 1.0 or the maximum number of evaluations is reached;

[0077] Step 3.3: Implement a downsampling acceleration strategy, downsampling the generated 1Hz high-frequency process data by a factor of 10 to convert it into 0.1Hz data, significantly reducing... The model's input data volume; after downsampling, it retains complete casting cycle data for 2450 time steps to ensure that key process features are not lost; by downsampling, the single temperature prediction time is shortened from 30 seconds to less than 3 seconds, enabling the TPE algorithm to complete multiple iterative evaluations within the production cycle; an incremental evaluation strategy is adopted, and the early stop conditions are checked immediately after each test to avoid unnecessary calculations;

[0078] Step 3.4: Generate optimized control parameters; After TPE optimization is completed, extract the optimal parameter combination best_x and update it in the process_data dictionary; Convert the optimized parameters into a complete time-series control sequence using the Creating_time_series_data function, including the flow rate settings for each cooling water line and precise switching timing, with the flow rate settings ranging from 5-80 L / min; Add feature engineering processing to calculate the cumulative flow actual_flow_V for accurate model prediction; Save the optimal parameters to a JSON file for easy subsequent calling and traceability;

[0079] Step 3.5: Verify the optimization effect; re-execute the complete prediction process using the optimal parameters and calculate the final RMSE value for verification; evaluate the maximum absolute error (MAE) and root mean square error (RMSE) of the predicted temperature versus the set temperature curve using the calculate_errors function; record the convergence curve of the optimization process and analyze the performance improvement of the TPE algorithm in the first 5 random explorations and the subsequent 20 Bayesian optimizations; calculate the total optimization time to ensure that the real-time control requirements are met.

[0080] Step 4: Input the optimal control parameters into the execution control module, and use the PLC controller to drive the electric regulating valve and variable frequency water pump according to the optimal control parameters to accurately adjust the cooling water flow and on / off timing.

[0081] Step 4.1: Issue and execute control commands; encode the optimal control parameters optimized by TPE into Modbus RTU protocol frames; send precise opening commands to each electric regulating valve via PLC, with a resolution of 0.1%; coordinate the control of the variable frequency water pump to maintain stable system pressure. It records the execution status and response time of instructions in real time.

[0082] Step 5: Collect data. The data is fed back into the data collection system of the prediction model to complete the closed loop and synchronously and adaptively adjust the weight coefficients of the objective function. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization.

[0083] Step 5.1: Monitor and evaluate the control effect and performance; collect temperature response data after control execution, calculate the root mean square error between the actual temperature and the target temperature; evaluate the temperature convergence time and overshoot; statistically analyze the convergence speed and calculation time of the TPE algorithm; record cooling water consumption and energy consumption data for cost analysis.

[0084] Step 5.2: Implement online updates for Bayesian optimization; store the parameter-effect pairs of each control operation in the empirical database; update the prior distribution of TPE using new data every 30 casting cycles; employ transfer learning to transfer optimization experience from similar working conditions to new tasks; adaptively adjust the weight coefficients of the objective function based on the cumulative control effect. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization.

[0085] Step 5.3: Construct a knowledge graph to assist decision-making; extract high-frequency optimal parameter combinations in the TPE optimization process and construct a process knowledge base; establish a mapping relationship between "temperature deviation and optimal parameters" to accelerate parameter selection for similar operating conditions; analyze the correlation and coupling relationships between parameters to simplify the search space dimension; generate an optimization report to visualize the parameter distribution and convergence process.

[0086] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A closed-loop temperature control system for aluminum wheel die-casting molds, comprising a data acquisition module, specifically a temperature sensor, a flow sensor, and a pressure sensor arranged at key locations in the mold, characterized in that: The data acquisition module has a communication connection to the prediction module, which uses... The composite neural network model has an optimization decision module connected to the prediction module and provides the prediction value. The optimization decision module adopts an integrated tree-structured Pareto estimator (TPE). The optimization decision module provides the optimal decision to the execution control module, which is a PLC controller. It also has a feedback learning module for online updating of the prediction model and optimization strategy based on the control effect data. The control method based on this closed-loop control system includes the following steps: Step 1: Collect historical data and use historical data to train. Composite neural network model; Step 2: Use the acquisition module to collect relevant data about the mold in real time. After data preprocessing, input the data into the trained system. A composite neural network model outputs predicted temperature values ​​for each measuring point in the future. Step 3: Based on the temperature prediction value and prediction results, if an anomaly is found in the prediction results, the TPE module is intervened according to the ideal temperature curve. The TPE module provides an initial cooling parameter prediction combination. This initial cooling parameter prediction combination is converted into time series data and input into the prediction module to predict whether the prediction result after adjusting this initial cooling parameter combination can meet the requirements of the ideal temperature curve. If not, this process is iterated until the optimal control parameters that meet the ideal temperature curve are given. Step 4: Input the optimal control parameters into the execution control module, and use the PLC controller to drive the electric regulating valve and variable frequency water pump according to the optimal control parameters to accurately adjust the cooling water flow and on / off timing. Step 5: Collect data. The data is fed back into the data collection system of the prediction model to complete the closed loop and synchronously and adaptively adjust the weight coefficients of the objective function. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization.

2. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: The CNN-LSTM composite neural network model consists of the following layers from input to output: input layer, Conv1D layer, MaxPooling1D layer, LSTM layer, Dropout layer, LSTM layer, Flatten layer, Dense layer, Dense layer, and output layer.

3. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 2, characterized in that: The input layer of the composite neural network model adopts an ultra-long time window design, namely a 1250 time step retrieval window design, to capture the thermal inertia characteristics and complete process cycle of the die-casting process.

4. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: The tree-structured Pareto estimator (TPE) takes as input manually set standard process template data, namely the ideal temperature curve and the previous cooling parameters, and outputs the optimal control parameters, which are the parameters for controlling the cooling equipment. A hot-start mechanism is introduced into the TPE, dividing the process time into two stages: controllable and uncontrollable. A differentiated processing strategy is adopted in the TPE optimization: the baseline process prediction data is reused for the uncontrollable stage, while parameter optimization search is performed for the controllable stage.

5. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: Step 1 is as follows: Step 1.1: Collect production data over a period of time, including: temperature time-series data of each measuring point within each casting cycle, with a sampling interval of 1 second; corresponding cooling process parameters, specifically including flow rate, temperature, and on / off timing; product quality inspection results, specifically including pass rate and defect type; and environmental parameters, specifically including workshop temperature and humidity. Step 1.2: Data Cleaning and Feature Engineering: Remove abnormal data during equipment failures; perform receipt window processing on temperature data with a window width of 1250 steps; perform cumulative calculation on flow data to enhance flow characteristics; normalize all features to... interval; Step 1.3: Train the CNN-LSTM composite neural network model: The input layer receives 1250 seconds of historical data; the CNN layer contains two convolutional blocks to extract spatial features; the LSTM layer contains two layers with 128 neurons to capture temporal dependencies; the output layer predicts the temperature value for the next cycle; the Adam optimizer is used with a learning rate of 0.001, a batch size of 64, and training for 30 epochs; the model achieves [specific results] on the validation set. Save the model afterwards.

6. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: Step 2 is as follows: Step 2.1: Establish a real-time data stream processing pipeline, configure the data acquisition module to acquire all sensor data at a frequency of 1Hz; implement edge computing preprocessing, including data verification, outlier removal, and missing value interpolation; construct a sliding time window and maintain a cache of historical data for the most recent 1250 seconds; The preprocessed data is pushed to the prediction module in real time; Step 2.2: Perform online temperature prediction by inputting the current time window data into the deployed... Model; Model outputs predicted temperature values ​​for each measuring point within the next 245 seconds; Calculates the deviation vector between the predicted temperature and the set target temperature. When any measuring point deviates When this happens, the optimization decision module is triggered; Step 2.3: Implement dynamic correction of the prediction results by using Kalman filtering to fuse the predicted and measured values; dynamically adjust the prediction confidence based on the error statistics of the last 10 predictions; and trigger the online fine-tuning mechanism of the model when the prediction error exceeds the threshold for 5 consecutive predictions.

7. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: Step 3 is detailed below: Step 3.1: Construct the Bayesian optimization search space; define the decision variable space. This includes i key cooling control parameters: delay time Opening time Set the objective function This refers to the root mean square error between the predicted temperature and the set temperature curve; defining the early stop threshold. The optimization will automatically terminate when the early stop threshold is reached. Step 3.2: Execute the TPE algorithm for intelligent search; implement the TPE sampler using the Optuna framework, randomly sample 5 parameter combinations during the initialization phase, and evaluate their RMSE values ​​as prior knowledge; the TPE algorithm automatically sorts historical observation points according to the target value, selects the best-performing observation points to construct the excellent sample set L(x), and constructs the poor sample set G(x) from the rest; fit kernel density estimation models to L(x) and G(x) respectively, and construct probability density functions. and Improve by maximizing expectations To select the next sampling point; Set the maximum number of evaluations to 25. Or it may terminate when the maximum number of evaluations is reached; Step 3.3: Implement a downsampling acceleration strategy, downsampling the generated 1Hz high-frequency process data by a factor of 10 to convert it into 0.1Hz data, significantly reducing the amount of input data for the CNN-LSTM model; after downsampling, retain the complete casting cycle data for 2450 time steps; through downsampling, shorten the single temperature prediction time from 30 seconds to less than 3 seconds, enabling the TPE algorithm to complete multiple iterative evaluations within the production cycle; adopt an incremental evaluation strategy, checking the early stop conditions immediately after each test. Step 3.4: Generate optimized control parameters; after TPE optimization is completed, extract the optimal parameter combination. Updated to In the dictionary; through The function converts the optimized parameters into a complete timing control sequence, including the flow rate settings for each cooling water line and precise switching timing. The range of the flow rate settings is... ; Add feature engineering processing to calculate cumulative flow. Used for accurate predictions by the model; saves the optimal parameters to a JSON file; Step 3.5: Verify the optimization effect; re-execute the complete prediction process using the optimal parameters and calculate the final RMSE value for verification; through The function evaluates the maximum absolute error (MAE) and root mean square error (RMSE) of the predicted temperature versus the set temperature curve; the convergence curve of the optimization process is recorded, and the performance improvement of the TPE algorithm in the first 5 random explorations and the subsequent 20 Bayesian optimizations is analyzed. The total optimization time is calculated to ensure that real-time control requirements are met.

8. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: Step 4 is as follows: Step 4.1: Issue and execute control commands; encode the optimal control parameters optimized by TPE into Modbus RTU protocol frames; send precise opening commands to each electric regulating valve via PLC, with a resolution of 0.1%; coordinate the control of the variable frequency water pump to maintain stable system pressure. It records the execution status and response time of instructions in real time.

9. The closed-loop temperature control system for aluminum wheel die-casting molds according to claim 1, characterized in that: Step 5 is as follows: Step 5.1: Monitor and evaluate the control effect and performance; collect temperature response data after control execution, calculate the root mean square error between the actual temperature and the target temperature; evaluate the temperature convergence time and overshoot; statistically analyze the convergence speed and calculation time of the TPE algorithm; record cooling water consumption and energy consumption data for cost analysis. Step 5.2: Implement online updates for Bayesian optimization; store the parameter-effect pairs for each control into the experience database; After every 30 casting cycles, the prior distribution of TPE is updated using new data; a transfer learning strategy is employed to transfer optimization experience from similar working conditions to new tasks; and the weight coefficients of the objective function are adaptively adjusted based on the cumulative control effect. and Regularly analyze the sampling distribution of TPE to identify parameter-sensitive areas for focused optimization. Step 5.3: Construct a knowledge graph to assist decision-making; extract high-frequency optimal parameter combinations in the TPE optimization process and construct a process knowledge base; establish a mapping relationship between "temperature deviation and optimal parameters"; analyze the correlation and coupling relationship between parameters; and generate an optimization report.

Images