Intelligent feeding system for ceramic substrate casting machine

Abstract

The application belongs to the technical field of intelligent monitoring of material processing, and discloses a kind of ceramic substrate casting machine feed intelligent system, comprising: the comprehensive data in the feeding process in acquisition cycle time is collected, and the comprehensive data collected is preprocessed, and the comprehensive data set is obtained;The comprehensive data set is reduced dimension processing, and the low-dimensional feature matrix is obtained;Using low-dimensional feature matrix as input, by fusing LSTM and reinforcement learning, a feed stability prediction model is constructed to predict the stability in the feeding process, and the probability distribution and optimal action of the feed stability are obtained;According to the probability distribution of the feed stability, a feed stability state signal is generated, and based on the feed stability state signal, the operating parameters are adjusted by the optimal action, which not only improves the prediction accuracy and response speed of the feeding process, but also greatly reduces the need for manual intervention, thereby greatly optimizing the production efficiency and product quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for material processing, and more specifically, to an intelligent feeding system for a ceramic substrate casting machine. Background Technology

[0002] Ceramic substrate casting machines are key pieces of equipment used in the manufacture of precision ceramic substrates and are widely used in the electronics industry. With the increasing demand for miniaturization and high performance in electronic devices, higher requirements are being placed on the quality of ceramic substrates. Traditional casting machine feeding control systems often rely on manual adjustments based on experience, which is insufficient to meet the demands of high-precision control and is easily affected by environmental factors.

[0003] Existing feeding control systems for ceramic substrate casting machines mainly employ PID control or simple feedback control systems. While these systems are relatively simple to implement and easy to maintain, they also have significant limitations, such as: poor adaptability (traditional control systems struggle to adapt quickly to changes in process conditions or external environmental influences, leading to product quality fluctuations); limited control accuracy (due to a lack of advanced algorithms, they cannot precisely control minute changes during the feeding process); and low data utilization (a large amount of data generated during production is not effectively utilized and cannot provide a basis for optimized control).

[0004] In view of this, the present invention proposes an intelligent feeding system for a ceramic substrate casting machine to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an intelligent feeding system for a ceramic substrate casting machine, comprising: Data acquisition and processing module: used to collect comprehensive data during the feeding process within the cycle time, and to preprocess the collected comprehensive data to obtain a comprehensive dataset; Feature fusion module: used to reduce the dimensionality of the comprehensive dataset to obtain a low-dimensional feature matrix; Model building module: Used to build a feeding stability prediction model by using a low-dimensional feature matrix as input and fusing LSTM with reinforcement learning, predicting the stability during the feeding process and obtaining the probability distribution of feeding stability and the optimal action; Judgment and Adjustment Module: Used to generate a stable feeding state signal based on the probability distribution of feeding stability, and adjust the operating parameters based on the stable state signal through optimal action.

[0006] Furthermore, the comprehensive data includes indoor environmental data, outdoor environmental data, and equipment operation data; Indoor environmental data includes indoor temperature and humidity; outdoor environmental data includes outdoor temperature, humidity and solar radiation intensity; equipment operation data includes feeding equipment data, casting equipment data and coating thickness data; feeding equipment data includes slurry viscosity, slurry flow rate and slurry pump pressure; casting equipment data includes doctor blade movement speed and conveyor belt speed.

[0007] Furthermore, the preprocessing methods include data cleaning, data standardization, and data synchronization; The data synchronization methods include: The periodic time is uniformly discretized into U_i time points, and a time axis is established based on these time points. Extract the data values ​​of each data type from the comprehensive data within the period and integrate them into the corresponding category datasets. Select one category dataset as the baseline dataset AJ, calculate the optimal time offset between each of the remaining category datasets and the baseline dataset, and adjust each category dataset according to the optimal time offset so that the data values ​​of all category datasets are on the same time axis. Methods for calculating the optimal time offset include: Define a time offset range [-nj, nj], and calculate the time offset range corresponding to each τ. value, The value represents the cross-correlation function value between any other category dataset and the benchmark dataset AJ. ; in, The data value of the baseline dataset AJ at time tg. ) represents the data value of the class dataset BJ at time tg, Ap represents the average value of all data in the baseline dataset AJ, Bp represents the average value of all data in the class dataset BJ, and τ represents the time offset; Select offset range The τ corresponding to the maximum value is taken as the optimal time offset. ; The methods for adjusting each category dataset based on the optimal time offset include: Based on the optimal time offset ,like If the value is greater than 0, then the data value at tg in the category dataset BJ will shift to the right. At a specific point in time; if If <0, then the data value at tg in the category dataset BJ shifts to the left. At a specific point in time; if If = 0, then the data value at tg in the category dataset BJ will not be adjusted; Calculate the optimal time offset for the data values ​​at each of the U_i time points in the category dataset, and adjust them one by one. Collect all category datasets aligned on the time axis, and denote them as the comprehensive dataset.

[0008] Furthermore, the method for dimensionality reduction of the comprehensive dataset to obtain a low-dimensional feature matrix includes: Each category dataset within the comprehensive dataset is treated as a feature, and a BU×BU scatter plot matrix is ​​constructed, where BU represents the number of features; Construct feature SDs on the diagonal elements within the scatter plot matrix. ta Histograms are constructed by creating scatter plots between any two distinct features on off-diagonal elements, representing feature pairs (SDs). ta SD tb The scatter plot of the scatter plot matrix is ​​used to obtain the feature vector set by identification and analysis. The covariance matrix between each feature in the feature vector set is calculated. The covariance matrix is ​​decomposed into eigenvalues, and the first k eigenvectors are selected as principal components to obtain the low-dimensional feature matrix. The methods for constructing a scatter plot between any two different features include: For any two distinct features SD ta and SD tb Set one of the features SD ta For the x-axis in the scatter plot, another feature is SD. tb Let be the ordinate of the scatter plot, and let tg(SD) be the ordinate of any point on the scatter plot. ta— tg, SD tb— tg), of which SD ta— tg represents the feature SD at time point tg. ta Data values, SD tb— tg represents the feature SD at time point tg. tb Data values; The construction feature SD ta Histogram methods include: For feature SD ta Set a range [min-SD] ta max-SD ta ], of which, min-SD ta For feature SD ta Minimum value of data in the middle, max-SD ta For feature SD ta The maximum value of the data in the middle; Construct a blank coordinate system, uniformly discretize the interval range into su intervals as the horizontal axis of the blank coordinate system, count the number of data values ​​falling within each interval, and use this count as the vertical axis of the blank coordinate system to plot the feature SD.ta The histogram.

[0009] Furthermore, the method for identifying and analyzing the scatter plot matrix to obtain the feature vector set includes: Based on the scatter plot matrix, identify the distribution of points in each scatter plot; If the points in a scatter plot are arranged in a straight line, it indicates that there is a linear correlation between the feature pairs in the scatter plot. If the arrangement of points resembles a parabola or an S-shaped curve, it indicates that there is a non-linear relationship between the feature pairs in the scatter plot. For feature pairs that are linearly correlated, count the number of other linearly correlated features in the scatter plot matrix for each feature pair, and retain the feature with the most features in the feature pair. For feature pairs with nonlinear relationships, the polynomial feature construction method is used to transform the feature pair into a new combined feature, and the combined feature is preserved. The method of using the polynomial feature construction method to transform the feature pair into a new combined feature includes: For feature pairs (SD) with nonlinear relationships ta SD tb ), setting the polynomial order d=2, we construct the combinatorial characteristics of the second order [ ], SD of the original features ta and SD tb The constructed combined features are combined to form a combined feature matrix; Principal component analysis is used to process the combined feature matrix, and the first principal component is selected as a new feature to be retained. All retained features are integrated and denoted as the feature vector set.

[0010] Furthermore, the method of selecting the first k eigenvectors as principal components to obtain the low-dimensional feature matrix includes: Feature variables are extracted from the feature vector set to construct a feature matrix, and the feature matrix is ​​centered. Based on the centered feature matrix, the covariance matrix is ​​obtained. Eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; Calculate the proportion of explained variance for each feature value, set a proportion threshold, and start from the first feature value, and accumulate the proportion of explained variance for each feature value until it reaches or exceeds the preset proportion threshold. Extract the number of feature values ​​corresponding to the point where accumulation stops, and use the number of feature values ​​as the k value; Based on the value of k, the first k eigenvectors are selected as principal components to form the principal component matrix; The low-dimensional feature matrix is ​​obtained by multiplying the centered feature matrix with the principal component matrix. The centering method for the feature matrix is ​​to subtract the mean of each feature from the mean of each feature; the calculation method for the explained variance ratio is to calculate the ratio of each feature value to the sum of all feature values.

[0011] Furthermore, the method for constructing the feed stability prediction model includes: The feed stability prediction model includes an LSTM model and a policy network; The methods of using LSTM models include: The LSTM layer is set to consist of l_s LSTM units, and the output is the current state. A fully connected layer is set up, and the current state is input into the fully connected layer. The output is the main body of the feed stability, including two categories: "stable" or "unstable". The classification result of the fully connected layer is used as the input of the softmax layer, and the probability distribution p of feed stability is output through the softmax layer. The methods of using the policy network include: The output of the LSTM layer is set as the input of the policy network, and the policy network outputs the optimal action 'a'. t ; Among them, a t These are the adjustment values ​​for the equipment's operating parameters.

[0012] Furthermore, the training process of the feed stability prediction model includes: Collect a historical dataset containing historical comprehensive data and its corresponding real labels. Construct a training dataset (Y, y) using the historical dataset. Divide the training dataset into q_s groups of training data. Integrate the q_s groups of training data into three data groups according to a preset ratio, which will be used as the training set, validation set, and test set, respectively. Where Y is a low-dimensional feature matrix, y is the corresponding true label, and the true label y is defined as a binary variable, y=[y1,y2], where y1 and y2 correspond to the true labels of the "stable" and "unstable" states of the feeding process, respectively. The training set is input into the feed stability prediction model, and the output is the current state, probability distribution, and optimal action. The cross-entropy loss function is used to calculate the difference between the probability distribution and the actual label, the policy gradient method is used to calculate the gradient of the policy network with respect to the parameters, and the model parameters are updated accordingly. Repeat this process until the predetermined number of training rounds is reached; The cross-entropy loss function of the trained LSTM model is defined as: Where p is the predicted probability distribution of the model output, p=[p1,p2], p1 is the predicted probability of stable feed, and p2 is the predicted probability of unstable feed. The gradient of the policy network with respect to the parameters is calculated using the policy gradient method, and its update rule is defined as follows: ;in, For the parameters of the policy network, For the policy network with respect to parameters gradient, For the policy network, input the current state The probability distribution obtained after output This represents the cumulative return starting from the current state. It is the learning rate.

[0013] Furthermore, the method for generating the feed steady-state signal based on the probability distribution of feed stability includes: Real-time comprehensive data is collected during the feeding process to obtain a real-time low-dimensional feature matrix; Input the real-time low-dimensional feature matrix into the feed stability prediction model to obtain the probability distribution p of feed stability; If a stable probability threshold is preset, then... If the probability is greater than the stability threshold, the current feeding system is determined to be stable, and a feeding stability signal is generated. like If the probability is less than or equal to the stability threshold, the current feeding system is determined to be unstable, and an unstable feeding signal is generated.

[0014] Furthermore, the method of adjusting the operating parameters through optimal action includes: Based on the unstable feed signal, the current state is input into the strategy network to obtain the optimal action, and the operating parameters in the production process are adjusted according to the optimal action. After adjusting the operating parameters, comprehensive data was re-collected, and a new probability distribution was obtained through the feed stability prediction model. If the probability is greater than the stability threshold, the adjustment is considered beneficial; otherwise, the optimal action is re-obtained and the operating parameters are adjusted until the feeding process stabilizes.

[0015] The technical effects and advantages of the intelligent feeding system for a ceramic substrate casting machine of the present invention are as follows: This invention achieves intelligent control of the feeding process in a ceramic substrate casting machine by introducing data processing techniques and machine learning algorithms, such as LSTM and reinforcement learning. The system can automatically collect and process various data during the feeding process, including but not limited to indoor and outdoor environmental data and equipment operation data. It then predicts feeding stability through feature fusion and model building, automatically adjusting feeding parameters to maintain system stability. By combining LSTM and reinforcement learning, it effectively learns the time-series characteristics of historical data, improving the accuracy of feeding stability prediction and thus better guiding actual operation. Furthermore, it automatically adjusts operating parameters based on real-time data analysis results. Dynamic optimization control ensures the continuity and stability of the feeding process, reducing the need for manual intervention. Simultaneously, the system's data acquisition and processing module performs comprehensive data cleaning, standardization, and synchronization preprocessing, ensuring data quality and providing a reliable foundation for subsequent analysis. Through intelligent control, the system can promptly identify and resolve potential problems during the feeding process, avoiding product quality issues caused by unstable feeding, improving production efficiency and product consistency. Furthermore, precise control of the feeding process reduces raw material waste and the generation of defective products, effectively lowering production costs and demonstrating significant practical value and broad application potential. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an intelligent feeding system for a ceramic substrate casting machine according to the present invention; Figure 2 This is a schematic diagram of an intelligent feeding method for a ceramic substrate casting machine according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Please see Figure 1 As shown, the data acquisition and processing module is used to collect comprehensive data during the feeding process within a cycle time, and to preprocess the collected comprehensive data to obtain a comprehensive dataset. Feature fusion module: used to reduce the dimensionality of the comprehensive dataset to obtain a low-dimensional feature matrix; Model building module: Used to build a feeding stability prediction model by using a low-dimensional feature matrix as input and fusing LSTM with reinforcement learning, predicting the stability during the feeding process and obtaining the probability distribution of feeding stability and the optimal action; Judgment and Adjustment Module: Used to generate a stable feeding state signal based on the probability distribution of feeding stability, and adjust the operating parameters based on the stable state signal through optimal action; The comprehensive data includes indoor environmental data, outdoor environmental data, and equipment operation data; the collected comprehensive data consists of several sets. Indoor environmental data includes indoor temperature and humidity; outdoor environmental data includes outdoor temperature, humidity, and solar radiation intensity, acquired using corresponding sensors; equipment operation data includes feeding equipment data, casting equipment data, and coating thickness data; feeding equipment data includes slurry viscosity (monitored by an online viscometer), slurry flow rate (monitored by a flow meter), and slurry pump pressure (monitored by a pressure sensor); casting equipment data includes doctor blade movement speed and conveyor belt speed, acquired by the equipment encoder; coating thickness data is acquired by a non-contact thickness gauge. The preprocessing methods include data cleaning, data standardization, and data synchronization. The data cleaning process includes outlier handling, noise data processing, and missing value handling. Outlier handling involves detecting and removing outliers using statistical methods (such as standard deviation, box plots, etc.). Noise data processing involves removing noise interference from sensor data using filters (such as low-pass filters). Missing value handling involves filling in missing data using interpolation methods (such as linear interpolation, polynomial interpolation). The data standardization process involves using Z-score standardization to convert data of different dimensions into the same dimension, thereby eliminating the impact of dimension differences. The data synchronization methods include: The periodic time is uniformly discretized into U_i time points, and a time axis is established based on these time points. Extract the data values ​​of each data type from the comprehensive data within the period and integrate them into the corresponding category datasets. Select one category dataset as the baseline dataset AJ, calculate the optimal time offset between each of the remaining category datasets and the baseline dataset, and adjust each category dataset according to the optimal time offset so that the data values ​​of all category datasets are on the same time axis. Methods for calculating the optimal time offset include: Define a time offset range [-nj, nj], and calculate the time offset range corresponding to each τ. value, The value represents the cross-correlation function value between any other category dataset and the benchmark dataset AJ. ; in, The data value of the baseline dataset AJ at time tg. ) represents the data value of the class dataset BJ at time tg, Ap represents the average value of all data in the baseline dataset AJ, Bp represents the average value of all data in the class dataset BJ, and τ represents the time offset; Select offset range The τ corresponding to the maximum value is taken as the optimal time offset. ; The methods for adjusting each category dataset based on the optimal time offset include: Based on the optimal time offset ,like If the value is greater than 0, then the data value at tg in the category dataset BJ will shift to the right. At a specific point in time; if If <0, then the data value at tg in the category dataset BJ shifts to the left. At a specific point in time; if If = 0, then the data value at tg in the category dataset BJ will not be adjusted; Calculate the optimal time offset for the data values ​​at each of the U_i time points in the category dataset, and adjust them one by one. Collect all category datasets aligned on the time axis and denote them as the comprehensive dataset. The method for dimensionality reduction of the comprehensive dataset to obtain a low-dimensional feature matrix includes: Each category dataset within the comprehensive dataset is treated as a feature, and a BU×BU scatter plot matrix is ​​constructed, where BU represents the number of features; Construct feature SDs on the diagonal elements within the scatter plot matrix. ta Histograms are constructed by creating scatter plots between any two distinct features on off-diagonal elements, representing feature pairs (SDs). ta SD tb The scatter plot of the scatter plot matrix is ​​used to obtain the feature vector set by identification and analysis. The covariance matrix between each feature in the feature vector set is calculated. The covariance matrix is ​​decomposed into eigenvalues, and the first k eigenvectors are selected as principal components to obtain the low-dimensional feature matrix. The methods for constructing a scatter plot between any two different features include: For any two distinct features SD ta and SD tb Set one of the features SD ta For the x-axis in the scatter plot, another feature is SD. tbLet be the ordinate of the scatter plot, and let tg(SD) be the ordinate of any point on the scatter plot. ta— tg, SD tb— tg), of which SD ta— tg represents the feature SD at time point tg. ta Data values, SD tb— tg represents the feature SD at time point tg. tb Data values; The construction feature SD ta Histogram methods include: For feature SD ta Set a range [min-SD] ta max-SD ta ], of which, min-SD ta For feature SD ta Minimum value of data in the middle, max-SD ta For feature SD ta The maximum value of the data in the middle; The interval is uniformly discretized into su intervals. The number of data values ​​falling within each interval is counted and used as the interval frequency. The feature SD is constructed by plotting the number of intervals on the horizontal axis and the interval frequency on the vertical axis. ta Histogram; The methods for identifying and analyzing the scatter plot matrix to obtain the feature vector set include: Based on the scatter plot matrix, identify the distribution of points in each scatter plot; If the points in a scatter plot are arranged in a straight line, it indicates that there is a linear correlation between the feature pairs in the scatter plot. If the arrangement of points resembles a parabola or an S-shaped curve, it indicates that there is a non-linear relationship between the feature pairs in the scatter plot. For feature pairs that are linearly correlated, count the number of other linearly correlated features in the scatter plot matrix for each feature pair, and retain the feature with the most features in the feature pair. For feature pairs with nonlinear relationships, the polynomial feature construction method is used to transform the feature pair into a new combined feature, and the combined feature is preserved. The method of using the polynomial feature construction method to transform the feature pair into a new combined feature includes: For feature pairs (SD) with nonlinear relationships ta SD tb ), setting the polynomial order d=2, we construct the combinatorial characteristics of the second order [ ], SD of the original features ta and SD tb The constructed combined features are combined to form a combined feature matrix; Principal component analysis is used to process the combined feature matrix, and the first principal component is selected as a new feature to be retained. All retained features are integrated and denoted as the feature vector set; The method of selecting the top k eigenvectors as principal components to obtain a low-dimensional feature matrix includes: Feature variables are extracted from the feature vector set to construct the feature matrix X, and the feature matrix is ​​then centered. The centering method for the feature matrix is ​​to subtract the mean of each feature from the mean of that feature. The covariance matrix C is calculated based on the centered feature matrix Xq. covariance matrix ;in, This represents the transpose of the centered data matrix; Perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues ​​λ of the covariance matrix. i and the corresponding feature vector v i ; Calculate the proportion of variance explained for each eigenvalue. The proportion of explained variance is calculated as follows: the ratio of each eigenvalue to the sum of all eigenvalues ​​is calculated to obtain a ratio value, which is denoted as the proportion of explained variance. Set a ratio threshold, starting from the first feature value, and sequentially accumulate the proportion of the explained variance of each feature value until it reaches or exceeds the preset ratio threshold. Extract the number of feature values ​​corresponding to the point where accumulation stops, and use the number of feature values ​​as the k value; Based on the value of k, the first k eigenvectors are selected as principal components to form the principal component matrix; The low-dimensional feature matrix is ​​obtained by multiplying the centered feature matrix with the principal component matrix. The methods for constructing the feed stability prediction model include: The feed stability prediction model includes an LSTM model and a policy network; The methods of using LSTM models include: The LSTM layer is set to consist of l_s LSTM units, and the output is the current state. A fully connected layer is set up, and the current state is input into the fully connected layer. The output is the main body of the feed stability, including two categories: "stable" or "unstable". The classification result of the fully connected layer is used as the input of the softmax layer, and the probability distribution p of feed stability is output through the softmax layer. The methods of using the policy network include: The output of the LSTM layer is set as the input of the policy network, and the policy network outputs the optimal action 'a'.t ; Among them, a t These are the adjustment values ​​for the equipment's operating parameters; It should be noted that the action selection process includes: based on the current state s t The policy network π will output a policy distribution π(a|s) t Let denot , representing the probability of choosing each possible action in the current state. Based on this probability distribution, we can choose a specific action 'a'. t That is, a t =π(s t (This action can be set to a specific value, or it can be set to an adjustment direction or magnitude, and the selected action a will be executed.) t And observe the new status and immediate rewards, where the rewards reflect the immediate effect of the feeding system after the action is performed, such as whether it becomes more stable or less stable; The training process of the feed stability prediction model includes: Collect a historical dataset containing historical comprehensive data and its corresponding real labels. Construct a training dataset (Y, y) using the historical dataset. Divide the training dataset into q_s groups of training data. Integrate the q_s groups of training data into three data groups according to a preset ratio, which will be used as the training set, validation set, and test set, respectively. Where Y is a low-dimensional feature matrix, y is the corresponding true label, and the true label y is defined as a binary variable, y=[y1,y2], where y1 and y2 correspond to the true labels of the "stable" and "unstable" states of the feeding process, respectively. The training set is input into the feed stability prediction model, and the output is the current state, probability distribution, and optimal action. The cross-entropy loss function is used to calculate the difference between the probability distribution and the actual label, the policy gradient method is used to calculate the gradient of the policy network with respect to the parameters, and the model parameters are updated accordingly. Repeat this process until the predetermined number of training rounds is reached; The cross-entropy loss function of the trained LSTM model is defined as: Where p is the predicted probability distribution of the model output, p=[p1,p2], p1 is the predicted probability of stable feed, and p2 is the predicted probability of unstable feed. The gradient of the policy network with respect to the parameters is calculated using the policy gradient method, and its update rule is defined as follows: ;in, For the parameters of the policy network, For the policy network with respect to parameters gradient, For the policy network, input the current state The probability distribution obtained after output This represents the cumulative return starting from the current state. It is the learning rate; The method of generating the feed steady-state signal based on the probability distribution of feed stability includes: Real-time comprehensive data is collected during the feeding process to obtain a real-time low-dimensional feature matrix; Input the real-time low-dimensional feature matrix into the feed stability prediction model to obtain the probability distribution p of feed stability; If a stable probability threshold is preset, then... If the probability is greater than the stability threshold, the current feeding system is determined to be stable, and a feeding stability signal is generated. like If the value is less than or equal to the stability probability threshold, the current feeding system is determined to be unstable, and an unstable feeding signal is generated. The method of adjusting operating parameters through optimal action includes: Based on the unstable feed signal, the current state is input into the strategy network to obtain the optimal action, and the operating parameters in the production process are adjusted according to the optimal action. It should be noted that the feeding process of a ceramic substrate casting machine involves actions including, but not limited to, adjusting the slurry viscosity (by changing the viscosity of the slurry, the casting quality and stability are affected; reducing the viscosity increases fluidity and helps with uniform spreading; increasing the viscosity can prevent unevenness caused by excessively rapid flow), adjusting the casting speed (by changing the speed of the casting machine, the spreading rate of the slurry is adjusted; slowing down the casting speed provides more time for the slurry to spread evenly; increasing the speed can improve production efficiency, but may lead to uneven spreading), adjusting the temperature (by adjusting the setting of the heating element, the flow characteristics and drying process of the slurry are affected; increasing the temperature can accelerate the drying process, but may also accelerate the curing of the slurry, affecting uniformity; decreasing the temperature has the opposite effect), adjusting the pressure (by changing the pressure applied to the slurry, its uniformity and stability are affected; increasing the pressure can help the slurry spread better, but excessive pressure may cause material deformation or damage), and adjusting other process parameters (depending on the specific situation, this may also include other parameters that affect the stability of the feeding process, such as the proportion of slurry components). After adjusting the operating parameters, comprehensive data was re-collected, and a new probability distribution was obtained through the feed stability prediction model. If the probability is greater than the stability threshold, the adjustment is considered beneficial; otherwise, the optimal action is re-obtained and the operating parameters are adjusted until the feeding process stabilizes.

[0019] Example 2 Please see Figure 2 As shown, the parts not described in detail in this embodiment are described in Embodiment 1. An intelligent feeding method for a ceramic substrate casting machine is provided, including: S1. Collect comprehensive data during the feeding process within the cycle time, and preprocess the collected comprehensive data to obtain a comprehensive dataset; S2. Perform dimensionality reduction on the comprehensive dataset to obtain a low-dimensional feature matrix; S3. Using a low-dimensional feature matrix as input, a feeding stability prediction model is constructed by fusing LSTM with reinforcement learning to predict the stability during the feeding process and obtain the probability distribution of feeding stability and the optimal action. S4. Generate a stable feeding state signal based on the probability distribution of feeding stability, and adjust the operating parameters based on the stable state signal through optimal action.

[0020] Example 3 This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the intelligent feeding method for a ceramic substrate casting machine described above.

[0021] Since the electronic device described in this embodiment is the electronic device used to implement the intelligent feeding method for a ceramic substrate casting machine according to the embodiments of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the intelligent feeding method for a ceramic substrate casting machine described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the intelligent feeding method for a ceramic substrate casting machine according to the embodiments of this application falls within the scope of protection of this application.

[0022] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0023] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent feeding system for a ceramic substrate casting machine, characterized in that, The intelligent feeding system for the ceramic substrate casting machine includes: Data acquisition and processing module: used to collect comprehensive data during the feeding process within the cycle time, and to preprocess the collected comprehensive data to obtain a comprehensive dataset; Feature fusion module: used to reduce the dimensionality of the comprehensive dataset to obtain a low-dimensional feature matrix; Model building module: By fusing LSTM and reinforcement learning, a feeding stability prediction model is built. The low-dimensional feature matrix is ​​input into the trained feeding stability prediction model to predict the stability during the feeding process and obtain the probability distribution of feeding stability and the optimal action. Judgment and Adjustment Module: Used to generate a stable feeding state signal based on the probability distribution of feeding stability, and adjust the operating parameters based on the stable state signal through optimal action.

2. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The comprehensive data includes indoor environmental data, outdoor environmental data, and equipment operation data; Indoor environmental data includes indoor temperature and humidity; outdoor environmental data includes outdoor temperature, humidity and solar radiation intensity; equipment operation data includes feeding equipment data, casting equipment data and coating thickness data; feeding equipment data includes slurry viscosity, slurry flow rate and slurry pump pressure; casting equipment data includes doctor blade movement speed and conveyor belt speed.

3. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The preprocessing methods include data cleaning, data standardization, and data synchronization. The data synchronization methods include: The periodic time is uniformly discretized into U_i time points, and a time axis is established based on these time points. Extract the data values ​​of each data type from the comprehensive data within the period and integrate them into the corresponding category datasets. Select one category dataset as the baseline dataset AJ, calculate the optimal time offset between each of the remaining category datasets and the baseline dataset, and adjust each category dataset according to the optimal time offset so that the data values ​​of all category datasets are on the same time axis. Methods for calculating the optimal time offset include: Define a time offset range [-nj, nj], and calculate the time offset range corresponding to each τ. value, The value represents the cross-correlation function value between any other category dataset and the benchmark dataset AJ. ; in, The data value of the baseline dataset AJ at time tg. ) represents the data value of the class dataset BJ at time tg, Ap represents the average value of all data in the baseline dataset AJ, Bp represents the average value of all data in the class dataset BJ, and τ represents the time offset; Select offset range The τ corresponding to the maximum value is taken as the optimal time offset. ; The methods for adjusting each category dataset based on the optimal time offset include: Based on the optimal time offset ,like If the value is greater than 0, then the data value at tg in the category dataset BJ will shift to the right. At a specific point in time; if If <0, then the data value at tg in the category dataset BJ shifts to the left. At a specific point in time; if If = 0, then the data value at tg in the category dataset BJ will not be adjusted; Calculate the optimal time offset for the data values ​​at each of the U_i time points in the category dataset, and adjust them one by one. Collect all category datasets aligned on the time axis, and denote them as the comprehensive dataset.

4. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The method for dimensionality reduction of the comprehensive dataset to obtain a low-dimensional feature matrix includes: Each category dataset within the comprehensive dataset is treated as a feature, and a BU×BU scatter plot matrix is ​​constructed, where BU represents the number of features; Construct feature SDs on the diagonal elements within the scatter plot matrix. ta Histograms are constructed by creating scatter plots between any two distinct features on off-diagonal elements, representing feature pairs (SDs). ta SD tb The scatter plot of the scatter plot matrix is ​​used to obtain the feature vector set by identification and analysis. The covariance matrix between each feature in the feature vector set is calculated. The covariance matrix is ​​decomposed into eigenvalues, and the first k eigenvectors are selected as principal components to obtain the low-dimensional feature matrix. The methods for constructing a scatter plot between any two different features include: For any two distinct features SD ta and SD tb Set one of the features SD ta For the x-axis in the scatter plot, another feature is SD. tb Let be the ordinate of the scatter plot, and let tg(SD) be the ordinate of any point on the scatter plot. ta— tg, SD tb— tg), of which SD ta— tg represents the feature SD at time point tg. ta Data values, SD tb— tg represents the feature SD at time point tg. tb Data values; The construction feature SD ta Histogram methods include: For feature SD ta Set a range [min-SD] ta max-SD ta ], of which, min-SD ta For feature SD ta Minimum value of data in the middle, max-SD ta For feature SD ta The maximum value of the data in the middle; Construct a blank coordinate system, uniformly discretize the interval range into su intervals as the horizontal axis of the blank coordinate system, count the number of data values ​​falling within each interval, and use this count as the vertical axis of the blank coordinate system to plot the feature SD. ta The histogram.

5. The intelligent feeding system for a ceramic substrate casting machine according to claim 4, characterized in that, The methods for identifying and analyzing the scatter plot matrix to obtain the feature vector set include: Based on the scatter plot matrix, identify the distribution of points in each scatter plot; If the points in a scatter plot are arranged in a straight line, it indicates that there is a linear correlation between the feature pairs in the scatter plot. If the arrangement of points resembles a parabola or an S-shaped curve, it indicates that there is a non-linear relationship between the feature pairs in the scatter plot. For feature pairs that are linearly correlated, count the number of other linearly correlated features in the scatter plot matrix for each feature pair, and retain the feature with the most features in the feature pair. For feature pairs with nonlinear relationships, the polynomial feature construction method is used to transform the feature pair into a new combined feature, and the combined feature is preserved. The method of using the polynomial feature construction method to transform the feature pair into a new combined feature includes: For feature pairs (SD) with nonlinear relationships ta SD tb ), setting the polynomial order d=2, we construct the combinatorial characteristics of the second order [ ], SD of the original features ta and SD tb The constructed combined features are combined to form a combined feature matrix; Principal component analysis is used to process the combined feature matrix, and the first principal component is selected as a new feature to be retained. All retained features are integrated and denoted as the feature vector set.

6. The intelligent feeding system for a ceramic substrate casting machine according to claim 4, characterized in that, The method of selecting the top k eigenvectors as principal components to obtain a low-dimensional feature matrix includes: Feature variables are extracted from the feature vector set to construct a feature matrix, and the feature matrix is ​​centered. Based on the centered feature matrix, the covariance matrix is ​​obtained. Eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; Calculate the proportion of explained variance for each feature value, set a proportion threshold, and start from the first feature value, and accumulate the proportion of explained variance for each feature value until it reaches or exceeds the preset proportion threshold. Extract the number of feature values ​​corresponding to the point where accumulation stops, and use the number of feature values ​​as the k value; Based on the value of k, the first k eigenvectors are selected as principal components to form the principal component matrix; The low-dimensional feature matrix is ​​obtained by multiplying the centered feature matrix with the principal component matrix. The centering method for the feature matrix is ​​to subtract the mean of each feature from the mean of each feature; the calculation method for the explained variance ratio is to calculate the ratio of each feature value to the sum of all feature values.

7. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The methods for constructing the feed stability prediction model include: The feed stability prediction model includes an LSTM model and a policy network; The methods of using LSTM models include: The LSTM layer is set to consist of l_s LSTM units, and the output is the current state. A fully connected layer is set up, and the current state is input into the fully connected layer. The output is the main body of the feed stability, including two categories: stable or unstable. The classification result of the fully connected layer is used as the input of the softmax layer, and the probability distribution p of feed stability is output through the softmax layer. The methods of using the policy network include: The output of the LSTM layer is set as the input of the policy network, and the policy network outputs the optimal action. The optimal action is the adjustment value of the equipment operating parameters.

8. The intelligent feeding system for a ceramic substrate casting machine according to claim 7, characterized in that, The training process of the feed stability prediction model includes: Collect a historical dataset containing historical comprehensive data and its corresponding real labels. Construct a training dataset (Y, y) using the historical dataset. Divide the training dataset into q_s groups of training data. Integrate the q_s groups of training data into three data groups according to a preset ratio, which will be used as the training set, validation set, and test set, respectively. Where Y is a low-dimensional feature matrix, y is the corresponding true label, the true label y is defined as a binary variable, y=[y1,y2], y1 and y2 correspond to the true labels of the stable and unstable states of the feeding process, respectively; The training set is input into the feed stability prediction model, and the output is the current state, probability distribution, and optimal action. The cross-entropy loss function is used to calculate the difference between the probability distribution and the actual label, the policy gradient method is used to calculate the gradient of the policy network with respect to the parameters, and the model parameters are updated accordingly. Repeat this process until the predetermined number of training rounds is reached; The cross-entropy loss function of the trained LSTM model is defined as: Where p is the predicted probability distribution of the model output, p=[p1,p2], p1 is the predicted probability of stable feed, and p2 is the predicted probability of unstable feed. The gradient of the policy network with respect to the parameters is calculated using the policy gradient method, and its update rule is defined as follows: ;in, For the parameters of the policy network, For the policy network with respect to parameters gradient, For the policy network, input the current state The probability distribution obtained after output This represents the cumulative return starting from the current state. It is the learning rate.

9. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The method of generating the feed steady-state signal based on the probability distribution of feed stability includes: Real-time comprehensive data is collected during the feeding process to obtain a real-time low-dimensional feature matrix; Input the real-time low-dimensional feature matrix into the feed stability prediction model to obtain the probability distribution p of feed stability; If a stable probability threshold is preset, then... If the probability is greater than the stability threshold, the current feeding system is determined to be stable, and a feeding stability signal is generated. like If the probability is less than or equal to the stability threshold, the current feeding system is determined to be unstable, and an unstable feeding signal is generated.

10. The intelligent feeding system for a ceramic substrate casting machine according to claim 1, characterized in that, The method of adjusting operating parameters through optimal action includes: Based on the unstable feed signal, the current state is input into the strategy network to obtain the optimal action, and the operating parameters in the production process are adjusted according to the optimal action. After adjusting the operating parameters, comprehensive data was re-collected, and a new probability distribution was obtained through the feed stability prediction model. If the probability is greater than the stability threshold, the adjustment is considered beneficial; otherwise, the optimal action is re-obtained and the operating parameters are adjusted until the feeding process stabilizes.

Images