Tire production key parameter prediction method based on autonomous optimization learning

Abstract

The application discloses a tire production key parameter prediction method based on autonomous optimization learning, which comprises the following steps: collecting key parameter data and influencing factor data on a tire production line, training a deep learning model LSTNet and verifying the deep learning model, constructing an environment for training a reinforcement learning intelligent agent, writing a reward function, defining an action space, an observation space and a state in the environment, training a reinforcement learning PPO intelligent agent based on test data, enabling the intelligent agent to continuously learn and adjust production parameters in a virtual dynamic production environment, optimizing the prediction accuracy of key parameters, and using the trained reinforcement learning intelligent agent and the LSTNet model to predict and autonomously optimize the key parameters.

Description

Technical Field

[0001] This invention relates to a method for predicting key parameters in tire production based on autonomous optimization learning. This method combines reinforcement learning algorithms to reduce the prediction error of key parameters by autonomously optimizing the remaining parameters in the data, and belongs to the fields of deep learning and tire production technology. Background Technology

[0002] Industrial manufacturing processes typically involve complex structures and massive scales, with multiple subsystems working collaboratively to complete the overall manufacturing task. With the development of manufacturing technology, the number of operating parameters within subsystems is increasing, and redundancy exists, with some parameters becoming uncontrollable. Therefore, how to achieve autonomous parameter optimization while ensuring efficient collaboration among subsystems has become an important research topic. However, due to the complexity of the parameters and operating mechanisms within each subsystem, effectively implementing autonomous optimization learning to improve the overall system performance remains a challenging problem.

[0003] With the rapid development of artificial intelligence technology, autonomous learning methods based on supervised learning and reinforcement learning have gradually become an important research direction in intelligent systems. In traditional industrial manufacturing processes, especially in complex production environments, predictive networks typically rely on traditional supervised learning methods. However, supervised learning methods usually require a large amount of labeled data and have certain limitations when dealing with dynamically changing and high-dimensional complex environments. Therefore, how to effectively cope with the ever-changing industrial manufacturing environment and improve the system's adaptability has become a current challenge.

[0004] Taking the tire manufacturing industry as an example, the production process is complex, involving multiple subsystems and key parameters (such as the width and thickness of the tire film, and the temperature of the conveyor belt). Traditional production processes cannot effectively adapt to the real-time changes in these key parameters, leading to significant resource waste. Existing parameter optimization methods typically rely on traditional mathematical models, which are often too complex to adapt to changes in the production process. Summary of the Invention

[0005] This invention provides a method for predicting key parameters in tire production based on autonomous optimization learning. The aim is to optimize the prediction error of key parameters by interacting with the dynamically changing tire production environment. This method employs the PPO algorithm of reinforcement learning for real-time parameter adjustment, enabling the prediction network LSTNet to adapt to the changing factors in the production process, gradually reducing the prediction error of key parameters, thereby improving the accuracy of prediction and the stability of the production process.

[0006] A method for predicting key parameters in tire production based on autonomous optimization learning includes:

[0007] Step 1. Collect key parameter data and influencing factor data on the tire production line. Clean and preprocess the collected data and divide it into training data and test data.

[0008] The data should ideally consist of no fewer than 10,000 records, with a training data to test data ratio of 3:1.

[0009] The key parameters include tire rubber width, tire rubber thickness, and conveyor belt temperature during tire descent.

[0010] The factors affecting tire rubber sheet width include: calender main motor current (A), calender speed (mpm), calender outlet temperature (°C), calender pull-out roller speed (mpm), conveyor belt speed (mpm), calender trimmer blade spacing (mm), calender upper roller temperature control (°C), calender lower roller temperature control (°C), extruder main motor current (A), extruder speed (rpm), extruder die pressure (MPa), extruder die temperature (°C), extruder screw end pressure (MPa), extruder die temperature control (°C), extruder plasticizing 1 temperature control (°C), extruder plasticizing 2 temperature control (°C), extruder barrel temperature control (°C), and extruder screw temperature control (°C). Among these, width is a key parameter that needs to be predicted; the calender trimmer blade spacing is a manipulated variable and is also a parameter that needs to be optimized.

[0011] Factors affecting tire rubber sheet thickness include: extruder head temperature (°C), extruder head pressure (MPa), extruder screw end pressure (MPa), extruder current (A), extruder speed (mpm), calender roll gap (mm), calender outlet temperature (°C), calender current (A), calender speed (mpm), extruder screw section temperature (°C), extruder plasticizing temperature (°C), extruder extrusion section temperature (°C), calender upper roll temperature (°C), and calender lower roll temperature (°C). Thickness is a key parameter that needs to be predicted; calender roll gap is a manipulated variable and is also a parameter that needs to be optimized.

[0012] The factors affecting the temperature of the descent conveyor belt include: width at the contact point (mm), return air temperature (°C), supply air temperature (°C), mixed air temperature (°C), natural air temperature (°C), air-cooled 1 temperature (°C), air-cooled 2 temperature (°C), air-cooled 3 temperature (°C), air-cooled 4 temperature (°C), return air valve opening (°C), natural air valve opening (°C), and air-cooling percentage (%). Among these, the descent conveyor belt temperature is the key parameter that needs to be predicted; the return air valve opening, natural air valve opening, and air-cooling percentage are manipulated variables and are also parameters that we need to optimize.

[0013] Step 2. Using the training data collected in Step 1, a deep learning model LSTNet is developed, enabling the model to accurately predict key parameters from the given input parameters.

[0014] The LSTNet structure is as follows: LSTNet is a hybrid neural network architecture combining Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) network, and Skip-GRU, primarily used for processing time series data. The core idea of ​​this model is to extract local features of the time series through convolutional layers, capture temporal dependencies through LSTM, and enhance the modeling ability of local temporal dependencies through Skip-GRU. Finally, the model outputs the prediction results through fully connected layers. This model uses a sliding window technique (window size of 96) to process time series data. It extracts local features through three one-dimensional convolutional layers, introduces non-linearity using the ReLU activation function after each convolution, and reduces dimensionality through max pooling layers. The output dimension of the convolutional layers gradually decreases, ultimately extracting the local features of the time series. The output of the convolutional layers models temporal dependencies through a unidirectional LSTM layer, and then undergoes layer normalization and Dropout regularization. The Skip-GRU module further enhances the modeling ability of local dependencies through a skip window. The outputs of LSTM and Skip-GRU are concatenated along the feature dimension (the last dimension). This means the outputs of LSTM and Skip-GRU are connected along the feature dimension. A fully connected layer then maps the outputs of LSTM and Skip-GRU to a single output value, yielding the final prediction result.

[0015] Step 3. Use the test data collected in Step 1 to validate the deep learning model trained in Step 3, evaluate its prediction error, and thus obtain performance feedback of the prediction model.

[0016] Step 4. Build an environment for training the reinforcement learning agent and write the reward function;

[0017] The action space in the environment is the set of actions that an agent can choose at each time step;

[0018] The observation space in the environment defines the environmental information that an agent can acquire at each time step.

[0019] The state in an environment represents the configuration of parameters that need to be optimized at a certain time step.

[0020] The reward function calculates the error based on the prediction results of LSTNet, and its calculation formula is as follows:

[0021] When the error decreases, the calculation formula is:

[0022]

[0023] When the error increases, the calculation formula is:

[0024]

[0025] Where gap represents the current time step, gap old The gap represents the error of the previous time step. The formula for calculating gap is:

[0026] gap=-abs(predict-true)# (3)

[0027] Here, predict represents the key parameter values ​​predicted by the LSTNet model, and true represents the actual key parameter values.

[0028] When the key parameter is film width, the action space has a dimension of 2, meaning the agent has two actions: increasing or decreasing the trimming blade spacing. The observation space is represented as an array containing multiple trimming blade spacing values. Each value corresponds to the trimming blade spacing at a certain moment in the production process. The state in the environment represents the trimming blade spacing configuration at a certain time step.

[0029] When the key parameter is film thickness, the action space has two dimensions, meaning the agent has two actions: increasing or decreasing the calender roll gap. The observation space is represented as an array containing multiple calender roll gaps. Each value corresponds to the calender roll gap at a specific moment in the production process. The state in the environment represents the calender roll gap configuration at a given time step.

[0030] When the key parameter is the temperature of the conveyor belt during tire descent, the action space has a dimension of 8, meaning the agent has 8 actions. Each parameter has two actions: increasing or decreasing the return air valve opening, the natural air valve opening, and the air cooling percentage, respectively. The observation space is represented as an array containing multiple return air valve openings, natural air valve openings, and air cooling percentages. Each value corresponds to the return air valve opening, natural air valve opening, and air cooling percentage at a specific moment in the production process. The state in the environment represents the configuration of the return air valve opening, natural air valve opening, and air cooling percentage at a given time step.

[0031] Step 5. Train the reinforcement learning PPO agent based on the test data, enabling it to continuously learn and adjust production parameters in a virtual dynamic production environment to optimize the prediction accuracy of key parameters. This training process specifically includes:

[0032] (1) Select one data point;

[0033] (2) The agent selects an action to adjust the parameters that need to be optimized in the data, obtains a new optimized parameter, updates the current data, and inputs the updated data into the model for prediction, and obtains a new key parameter.

[0034] (3) Calculate the current reward based on the prediction error of the model and update the strategy periodically.

[0035] (4) Iterate steps (2) and (3) continuously to enable the agent to gradually improve its decision-making ability through continuous trial and error and learning, and update the strategy regularly.

[0036] (5) When the time for the agent to train this data reaches the preset number of time steps, or the prediction error of the model is less than the specified error, the training of this data is completed, and the process jumps to step (1) until all data training is completed.

[0037] Step 6. Use the reinforcement learning agent trained in Step 5 and the LSTNet model validated in Step 3 to predict and autonomously optimize key parameters.

[0038] New data is collected on the tire production line, cleaned, and preprocessed before being input into the prediction network LSTNet and the reinforcement learning agent for parameter optimization. After inputting the data, the optimization process is similar to the training process, but the policy is not updated. The training process specifically includes:

[0039] (1) Select one data point;

[0040] (2) The agent selects an action to adjust the parameters that need to be optimized in the data, obtains a new optimized parameter, updates the current data, and inputs the updated data into the model for prediction, and obtains a new key parameter.

[0041] (3) Calculate the current reward based on the prediction error of the model.

[0042] (4) Iterate through steps (2) and (3).

[0043] (5) When the time for the agent to train this data reaches the preset number of time steps, or the prediction error of the model is less than the specified error, the optimization of this data is completed, and the process jumps to step (1) until all data is optimized.

[0044] Through autonomous decision-making and parameter adjustments by the intelligent agent, the predictive model can be continuously optimized. Optimization ends once all data has been iterated through. This ensures that key parameters remain within predetermined ranges during actual production, improving production efficiency, reducing scrap rates, and optimizing resource utilization.

[0045] This invention innovatively combines two advanced technologies: deep learning and reinforcement learning. By introducing reinforcement learning, the agent can learn and optimize its behavioral strategies through interaction with the environment. Compared to traditional supervised learning, reinforcement learning does not rely on large amounts of labeled data. Instead, it optimizes the decision-making process through exploration and feedback mechanisms, thereby continuously self-adjusting and optimizing parameters in complex and dynamic environments, and constantly improving prediction accuracy. This self-optimizing learning method is particularly suitable for industrial manufacturing environments with complex system structures, dynamic changes, and difficulties in complete modeling.

[0046] Deep learning models can effectively extract the complex relationships between key parameters and various production parameters from large amounts of historical data, while reinforcement learning dynamically adjusts parameters through interaction with the environment. The combination of these two approaches enables the prediction network to not only make accurate predictions under static conditions but also continuously optimize through real-time feedback during the production process, further improving prediction accuracy. In traditional tire production, manual adjustment of production parameters based on experience is often required, which is not only inefficient but also susceptible to human error. This invention allows the intelligent agent to automatically judge and adjust production parameters, reducing human error, improving production efficiency, and ensuring consistent product quality. Through the agent's learning and optimization, the production process can remain in an optimal state, significantly reducing scrap rates, lowering production costs, and increasing efficiency. Especially in the precise control of key parameters, it avoids resource waste caused by parameter instability or over-adjustment, promoting efficient resource utilization. This invention, through the application of intelligent optimization algorithms, promotes the intelligentization and automation of industrial manufacturing processes, aligning with the intelligent manufacturing concept of Industry 4.0. Through autonomous learning and real-time optimization, production systems can respond more flexibly and efficiently to changes in market demand, enhancing the adaptability of production lines and improving the industry's level of intelligence. Developing a parameter autonomous optimization learning method based on a combination of reinforcement learning and traditional supervised learning can not only effectively solve key problems in current industrial manufacturing but also provide a new technological path for the future intelligent and automated processes in industrial production. Attached Figure Description

[0047] Figure 1 This is a flowchart of the present invention.

[0048] Figure 2 This is a diagram of the deep learning model architecture of the present invention.

[0049] Figure 3 This is the prediction effect of the model using tire rubber width as an example in this invention.

[0050] Figure 4 This is a flowchart illustrating the parameter self-optimization process using tire rubber width as an example in this invention.

[0051] Figure 5 This is the optimization effect of the reinforcement learning agent in this invention, taking the width of tire rubber as an example. Detailed Implementation

[0052] The process of this invention is as follows Figure 1 As shown in the diagram. First, key parameter data and influencing factor data are collected on the tire production line. The collected data is then cleaned and preprocessed, and divided into training data and test data.

[0053] The required data includes key parameters and their influencing factors, as shown in the table below. The correspondence between the labels and parameters is as follows: A: Tire rubber sheet width (mm), B: Calender main motor current (A), C: Calender speed (mpm), D: Calender outlet temperature (°C), E: Calender pull-out roller speed (mpm), F: Conveyor belt speed (mpm), G: Calender trimming knife spacing (mm), H: Calender upper roller temperature control (°C), I: Calender lower roller temperature control (°C), J: Extruder main motor current (A), K: Extruder speed (rpm), L: Extruder head pressure (MPa), M: Extruder head temperature (°C), N: Extruder screw end pressure (MPa), O: Extruder head temperature control (°C), P: Extruder plasticizing 1 temperature control (°C), Q: Extruder plasticizing 2 temperature control (°C), R: Extruder barrel temperature control (°C), S: Extruder screw temperature control (°C). Where A: tire sheet width (mm) is the key parameter that needs to be predicted, and G: calender trimmer blade spacing (mm) is the manipulated variable, which is also the parameter that we need to optimize. The collected data is shown in Table 1 below.

[0054] Table 1 shows the key parameter data and influencing factors collected on the tire production line.

[0055]

[0056] These data reflect the changing patterns of key parameters under different production conditions. The collected data needs to be cleaned and preprocessed to ensure its integrity and accuracy. Then, the data is divided into training and testing data in a 3:1 ratio to ensure effective validation of the model during training.

[0057] The LSTNet network is trained using training data. LSTNet is a neural network architecture based on sliding window Long Short-Term Memory (LSTM) units, with a sliding window size of 96. Its network architecture diagram is shown below. Figure 2As shown, this model combines a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a Skip-GRU network to effectively capture local features and long-term dependencies in time-series data for processing. A sliding window technique is used to generate a fixed-size window on the time-series data, ensuring the network can handle local features. This approach allows the model to make predictions using historical data in the time series without relying on the entire sequence. The network first extracts local features from the time-series data through three one-dimensional convolutional layers. Each convolutional layer is followed by a max-pooling layer to reduce the data dimensionality while retaining the most important features. The activation function uses ReLU to introduce non-linearity. The features extracted by the convolutional layers are fed into a unidirectional LSTM layer (GRU1) for time dependency modeling. LSTM excels at capturing long-term dependencies in data, ensuring the data input is processed according to the specified format. The output is layer normalized and Dropout regularization is used to prevent overfitting. The network further enhances its ability to model local dependencies in the time series using a Skip-GRU module. This module captures features at different time steps by dividing time series data into multiple skip windows and modeling them using a GRU. Finally, the outputs of the LSTM and Skip-GRU are concatenated as input to the next layer. After feature extraction and modeling, the network's output passes through a fully connected layer for final prediction. This layer maps high-dimensional temporal features to a single output value. Finally, the outputs of the LSTM and GRU are further mapped to the final prediction result.

[0058] The deep learning model is trained using the collected training data. During training, the model adjusts the weights and biases in the network by optimizing the loss function (mean squared error) to minimize prediction error. Through backpropagation and gradient descent, the model progressively optimizes its parameters so that the prediction results are as close as possible to the true values. During training, the model continuously updates its internal weights to better learn the relationship between the input parameters and key parameters. After training, the LSTNet deep learning model can predict key parameters based on different production conditions, providing a preliminary prediction result.

[0059] The trained deep learning model LSTNet was validated using collected test data. Its prediction error was evaluated by inputting the test data into the model. The results are as follows: Figure 3 As shown. The mean absolute error (MAE) and the pass rate of the prediction results are used to measure the model's predictive performance. The purpose of validation is to determine the model's generalization ability on unseen data and ensure that its prediction results have sufficient accuracy. If the prediction error is too large, it may be necessary to adjust the deep learning model or increase the amount of training data to improve the model's performance.

[0060] An environment is set up for training a reinforcement learning agent. Taking tire film width as an example, the action space in the environment is the set of actions the agent can choose at each time step. Specifically, the agent can perform the following two actions: 0: increase the trimming blade spacing; 1: decrease the trimming blade spacing. The action space is discrete, and the agent can only choose one action at a time, either increasing or decreasing the trimming blade spacing. The observation space in the environment defines the environmental information that the agent can acquire at each time step. The observation space is represented as an array containing multiple trimming blade spacing values. Each value corresponds to the trimming blade spacing at a certain moment in the production process. The dimension of this array is determined by the sliding window size of LSTNet. When performing each action, the agent's adjustment amount for the trimming blade spacing is set to 0.1 mm. That is, each time the trimming blade spacing is increased or decreased, the actual change in the trimming blade is 0.1 mm. The state in the environment represents the trimming blade spacing configuration at a certain time step.

[0061] Write a reward function that calculates the error based on the prediction results of LSTNet. The calculation formula is as follows:

[0062] When the error decreases, the calculation formula is:

[0063]

[0064] When the error increases, the calculation formula is:

[0065]

[0066] Where gap represents error, and the formula for calculating gap is:

[0067] gap=-abx(predict-true)# (3)

[0068] Where predict is the LSTNet model's predicted value, and true is the actual value.

[0069] Based on the obtained model test results, the reinforcement learning agent is further trained using the test data. The agent's task is to continuously learn and adjust production parameters in a virtual dynamic production environment to optimize the prediction accuracy of key parameters. One set of data is selected for training each time.

[0070] Reinforcement learning agents learn through interaction with their environment. The agent's decision-making process includes the following steps: Action Selection: The agent selects an action based on its current state. This action adjusts production parameters related to key parameters; for example, adjusting the trimming blade spacing to change the state of the production environment (e.g., tire sheet width). Reward Calculation: The agent evaluates the selected action based on the current prediction error and a pre-defined reward function. For example, a positive reward is given when the selected production parameter reduces the prediction error of the key parameter; conversely, a negative reward is given. State Update and Policy Adjustment: After selecting an action, the agent updates its state and adjusts its policy based on the new reward information. Through this trial-and-error process, the agent continuously optimizes its decision-making process, gradually improving the accuracy of production parameter adjustments. Through continuous interaction and learning, the agent can gradually identify the optimal parameter adjustment strategy during production, ultimately achieving precise control of key parameters. The agent performs this operation once per time step. Training for a given data point ends when the time steps for that data point are exhausted or the model's prediction error is less than 1 millimeter. This process is repeated until all test data has been collected and the training is complete.

[0071] Once the reinforcement learning agent has been sufficiently trained and the prediction error of the LSTNet model is small enough, it can enter the autonomous optimization phase. The optimization process is as follows: Figure 4 As shown. The optimization process is similar to the training process of the agent, but the policy is not updated. The optimization of a data point ends when the time steps for that data point are exhausted or the model's prediction error is less than a predetermined error. This process is repeated until all data is optimized. At this point, the agent can use its learned optimization strategy and deep learning model to make real-time adjustments to ensure that key parameters remain within the predetermined standard range. In actual production, the deep learning model continuously monitors changes in key parameters and makes predictions based on current production data. Simultaneously, the reinforcement learning agent continuously adjusts production parameters based on real-time prediction results; for example, adjusting the trimming blade spacing to minimize the prediction error of key parameters for tire sheet width. The optimization process ends when the model's prediction error is lower than the predetermined minimum error. The optimization results are shown below. Figure 5 As shown in the table below, through this self-optimization process, the production process can automatically adjust production parameters without human intervention, maintaining high accuracy of key parameters and stability of the production process. This not only improves production efficiency but also effectively reduces scrap rate and resource waste, ultimately achieving optimization and intelligentization of the production process. The actual width value, model predicted value, and optimized width predicted value are shown in Table 2 below.

[0072] Table 2 Comparison of actual width value, model prediction value, and optimized width prediction value

[0073]

[0074]

[0075] After incorporating an agent for optimization, the error between the predicted and actual width values ​​was significantly reduced. Specifically, for both smaller width values ​​(e.g., 488.80 mm) and larger width values ​​(e.g., 523.20 mm), the prediction error was effectively controlled to within 1 mm in the agent-optimized prediction model. For example, with an actual width of 488.80 mm, the original LSTNet model's prediction might fluctuate between 487.02 mm and 487.91 mm, resulting in a relatively large error. However, after incorporating agent optimization, the prediction was precisely adjusted to a range extremely close to the actual value, reducing the error to less than 1 mm. This improvement is evident at every data point; regardless of the width value, agent optimization ensures the accuracy and consistency of the predictions.

Claims

1. A method for predicting key parameters in tire production based on autonomous optimization learning, characterized by: Includes the following steps: Step 1. Collect key parameter data and influencing factor data on the tire production line, clean and preprocess the collected data, and divide it into training data and test data; Step 2. Use the data collected in Step 1 to train the deep learning model LSTNet, so that the model can accurately predict key parameters from the given input parameters; Step 3. Use the test data collected in Step 1 to validate the deep learning model trained in Step 3, evaluate its prediction error, and thus obtain performance feedback of the prediction model. Step 4. Build an environment for training the reinforcement learning agent and write the reward function; The action space in the environment is the set of actions that an agent can choose at each time step; The observation space in the environment defines the environmental information that the agent can acquire at each time step; The State in the environment represents the configuration of parameters that need to be optimized at a certain time step; Step 5. Train the reinforcement learning PPO agent based on the test data, so that it can continuously learn and adjust production parameters in a virtual dynamic production environment to optimize the prediction accuracy of key parameters; Step 6. Use the reinforcement learning agent trained in Step 5 and the LSTNet model validated in Step 3 to predict and autonomously optimize key parameters. That is, collect new data on the tire production line, clean and preprocess it, and then input it into the prediction network LSTNet and the reinforcement learning agent to optimize the parameters. The LSTNet structure described in step 2 is as follows: LSTNet is a hybrid neural network architecture combining a convolutional neural network (CNN), a long short-term memory network (LSTM), and a skip-GRU (Skip-GRU) for processing time series data. The core idea of ​​this model is to extract local features of the time series through convolutional layers, capture temporal dependencies through LSTM, and enhance the modeling ability of local temporal dependencies through Skip-GRU. Finally, the model outputs the prediction results through fully connected layers. This model uses a sliding window technique to process time series data, with a window size of 96. It extracts local features through three one-dimensional convolutional layers, and uses the ReLU activation function after each convolution. The input is nonlinear and dimensionality is reduced by max pooling layers, gradually decreasing the output dimension of the convolutional layers, and then local features of the time series are extracted. The output of the convolutional layers is used to model time dependencies through a unidirectional LSTM layer, and then processed by layer normalization and Dropout regularization. The Skip-GRU module further enhances the ability to model local dependencies through a skip window. The outputs of LSTM and Skip-GRU are concatenated in the feature dimension, i.e., the last dimension, that is, the outputs of LSTM and Skip-GRU are connected in the feature dimension. A fully connected layer is used to map the outputs of LSTM and Skip-GRU to a single output value to obtain the final prediction result.

2. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized in that: The data collected in step 1 should preferably be no less than 10,000 records, and the ratio of training data to test data should be 3:

1.

3. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized by the following steps: The key parameters mentioned in 1 include tire rubber width, tire rubber thickness, and conveyor belt temperature during tire descent; among which, Factors affecting tire rubber sheet width include: extruder temperature in each temperature control channel (°C), extruder head pressure (MPa), extruder head temperature (°C), extruder screw motor current (A), extruder screw speed (rpm), calender roller temperature (°C), calender pressure roller speed (mpm), calender pressure roller current (A), calender outlet rubber temperature (°C), calender trimming knife distance (mm), calender roller distance (mm), calender pull-out roller speed (mpm), and conveyor belt speed (mpm). Among these, width is a key parameter that needs to be predicted; the calender trimming knife distance is a manipulated variable and also a parameter that needs to be optimized. Factors affecting tire rubber thickness include: extruder head temperature (°C), extruder head pressure (MPa), extruder screw end pressure (MPa), extruder current (A), extruder speed (mpm), calender roll gap (mm), calender outlet temperature (°C), calender current (A), calender speed (mpm), extruder screw section temperature (°C), extruder plasticizing temperature (°C), extruder extrusion section temperature (°C), calender upper roll temperature (°C), and calender lower roll temperature (°C). Thickness is a key parameter that needs to be predicted; calender roll gap is a manipulated variable and also a parameter that needs to be optimized. The factors affecting the temperature of the descent conveyor belt include: width at the contact point (mm), return air temperature (°C), supply air temperature (°C), mixed air temperature (°C), natural air temperature (°C), air-cooled 1 temperature (°C), air-cooled 2 temperature (°C), air-cooled 3 temperature (°C), air-cooled 4 temperature (°C), return air valve opening (°C), natural air valve opening (°C), and air-cooled percentage (%). Among these, the descent conveyor belt temperature is the key parameter that needs to be predicted; the return air valve opening (°C), natural air valve opening (°C), and air-cooled percentage (%) are manipulated variables and also parameters that need to be optimized.

4. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized in that: In step 4 When the key parameter is film width, the action space has a dimension of 2, meaning the agent has two actions: increasing or decreasing the trimming blade spacing. The observation space is represented as an array containing multiple trimming blade spacing values. Each value corresponds to the trimming blade spacing at a certain moment in the production process, and the State in the environment represents the trimming blade spacing configuration at a certain time step. When the key parameter is film thickness, the action space has a dimension of 2, meaning the agent has two actions: increasing or decreasing the calender roll gap. The observation space is represented as an array containing multiple calender roll gaps; each value corresponds to the calender roll gap at a certain moment in the production process; the State in the environment represents the calender roll gap configuration at a certain time step. When the key parameter is the temperature of the conveyor belt as the tires descend, the action space has a dimension of 8, meaning the agent has 8 actions. Each parameter has two actions: increasing or decreasing the opening of the return air valve, the opening of the natural air valve, and the air cooling percentage, respectively. The observation space is represented as an array containing multiple return air valve openings, natural air valve openings, and air cooling percentages. Each value corresponds to the return air valve opening, natural air valve opening, and air cooling percentage at a certain moment in the production process. The State in the environment represents the configuration of the return air valve opening, natural air valve opening, and air cooling percentage at a certain time step.

5. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized in that: In step 4, the reward function calculates the error based on the prediction results of LSTNet, and the calculation formula is as follows: When the error decreases, the calculation formula is: When the error increases, the calculation formula is: in, gap This indicates the current time step, while the other indicates the error of the previous time step. gap The calculation formula is: Where is the key parameter value predicted by the LSTNet model, and is the actual key parameter value.

6. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized in that: In step 5, the training process specifically includes: (1) Select one data point; (2) The agent selects an action to adjust the parameters that need to be optimized in the data, obtains a new optimized parameter, updates the current data, and inputs the updated data into the model for prediction, and obtains a new key parameter. (3) Calculate the current reward based on the prediction error of the model and update the strategy periodically; (4) Iterate steps (2) and (3) continuously to enable the agent to gradually improve its decision-making ability through continuous trial and error and learning, and update the strategy regularly; (5) When the time for the agent to train this data reaches the preset number of time steps, or the prediction error of the model is less than the specified error, the training of this data is completed, and the process jumps to step (1) until all data training is completed.

7. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 1, characterized in that: Step 6, which involves prediction and autonomous optimization, is similar to the training process after data input, but without updating the strategy. Specifically, it includes: (1) Select one data point; (2) The agent selects an action to adjust the parameters that need to be optimized in the data, obtains a new optimized parameter, updates the current data, and inputs the updated data into the model for prediction, and obtains a new key parameter. (3) Calculate the current reward based on the model's prediction error; (4) Iterate through steps (2) and (3) continuously; (5) When the time for the agent to train this data reaches the preset number of time steps, or the prediction error of the model is less than the specified error, the optimization of this data is completed, and the process jumps to step (1) until all data is optimized.

8. The method for predicting key parameters of tire production based on autonomous optimization learning as described in claim 7, characterized in that: In step 6, the prediction model is continuously optimized through the autonomous decision-making of the agent and the adjustment of parameters. The optimization ends after all the data has been iterated once.

Images