A method, system, and storage medium for intelligent correction of deviations in heat treatment process parameters.

Abstract

This invention discloses an intelligent correction method, system, and storage medium for deviations in heat treatment process parameters, belonging to the field of production optimization technology. The method includes: augmenting the original production data; training a first prediction model locally based on the augmented data to obtain a second prediction model; calculating the deviation of network parameters at the same location in different second prediction models for each factory; determining adjustment values ​​based on the deviations; correcting the network parameters based on the adjustment values; defining the corrected second prediction model as a third prediction model; inputting real-time production data into the third prediction model; and using a dynamic adjustment mechanism in the third prediction model to adjust the number of network layers according to the product type and outputting the process adjustment amount and corresponding confidence level. When the confidence level is less than a second threshold, the historical case library is called to correct the process adjustment amount. This invention achieves real-time dynamic adjustment of heat treatment process parameters through a collaborative correction mechanism between local and cloud models.

Description

Technical Field

[0001] This invention belongs to the field of production optimization technology, specifically relating to an intelligent correction method, system, and storage medium for deviations in heat treatment process parameters. Background Technology

[0002] Hot processing manufacturing is a typical long-process manufacturing industry, which generally includes processes such as smelting, casting, forging, welding, and heat treatment. There are many process factors that affect manufacturing quality, and it is difficult to determine the quantitative relationship between process data and product quality. In addition, traditional verification-type process design optimization has a long cycle and high cost, so there is a problem of difficulty in adjusting process parameters. To solve this technical problem, the following method has been proposed in the prior art.

[0003] For example, Chinese patent document CN117590810A discloses a control method, device, equipment and storage medium for iron and steel smelting. The method records the process parameters and corresponding output quality indicators of each process such as coking, sintering and blast furnace. The actual quality indicators are compared with the theoretically predicted data. If the deviation is within the allowable range, it is determined to be consistent. Subsequently, only the data that are consistent in all processes are retained to form a quality indicator database. From this database, 5-10 sets of parameters with the highest matching degree are selected as the optimal process parameters.

[0004] However, this method is essentially a static optimization based on historical data, and cannot dynamically adjust process parameters according to real-time production conditions. When production conditions such as raw material composition and equipment status change, the database needs to be manually updated, resulting in poor adaptability. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent correction method, system, and storage medium for deviations in heat treatment process parameters, thereby resolving the issues present in the background art.

[0006] To achieve the aforementioned objectives, this invention proposes an intelligent correction method for deviations in heat treatment process parameters, comprising:

[0007] Obtain the original production data of each process in each factory, and augment the original production data to obtain augmented data;

[0008] A first prediction model with a dynamic adjustment mechanism is established. Based on the expanded data, the corresponding first prediction model is trained locally to obtain a second prediction model. The deviation of the network parameters at the same position of the second prediction model in different factories is calculated.

[0009] If the deviation exceeds the first threshold, an adjustment value is determined based on the deviation, the network parameters of each of the second prediction models at the same position are corrected based on the adjustment value, and it is determined whether to retain the corrected network parameters based on the change in the prediction accuracy of the second prediction models before and after the correction.

[0010] The revised second prediction model is defined as the third prediction model. Real-time production data is input into the third prediction model. The third prediction model determines the product type of the target product for this production. The dynamic adjustment mechanism adjusts the number of network layers according to the product type and outputs the process adjustment amount and the corresponding confidence level.

[0011] When the confidence level is less than the second threshold, real-time production data is uploaded to the adjustment model in the cloud, and the adjustment model calls the historical case library to correct the process adjustment amount.

[0012] Furthermore, data augmentation of the original production data includes the following steps:

[0013] Obtain a first quantity of the original production data, set an interpolation quantity, determine a second quantity based on the first quantity and the interpolation quantity, and generate interpolated data of the second quantity with the same dimension as the original production data under the constraints of the production process, which is defined as the basic data;

[0014] The basic data most similar to each of the original production data is selected and defined as the first data. A first correction amount is calculated based on the similarity between the original production data and the first data and the data sensitivity of each dimension of the data. The first data is corrected based on the first correction amount to obtain the second data. Multiple interpolated data most similar to the second data are obtained and defined as the third data. A second correction amount is determined based on the similarity between the second data and the third data. The third data is corrected based on the second correction amount to obtain the fourth data. The similarity between the original production data and each of the second data and the fourth data is calculated again and updated. This step is repeated until a preset number of times is reached. The original production data and multiple data most similar to the original production data in the basic data are used as the expanded data.

[0015] Furthermore, the confidence level of the process adjustment amount output by the third prediction model includes the following steps:

[0016] The output of the third prediction model includes intermediate quality parameters for each process and final quality parameters after the product is completed. A validation model is set up, which includes multiple calculation formulas. These formulas calculate the intermediate quality parameters after each process and the final quality parameters after production based on the real-time production data. Each calculation formula includes multiple unknown coefficients. The validation model is trained based on the original production data to determine the values ​​of the unknown coefficients. Corrected data is obtained by adjusting the real-time production data based on the process adjustment amount. The corrected data is input into the validation model to calculate the intermediate quality parameters and the final quality parameters after each process. The intermediate quality parameters and the final quality parameters of the third prediction model and the validation model are compared to determine the confidence level.

[0017] Furthermore, the adjustment model's process adjustment based on historical case libraries includes the following steps:

[0018] The historical case library includes multiple best production cases. The adjustment model calculates the similarity of raw material composition, equipment status, and environmental parameters between the real-time production data and each production case. If the similarity of a production case with the real-time production data in all three parameters is greater than the reliable threshold, the process adjustment amount is generated based on the process parameters in that production case and sent to the factory; otherwise, a manual review process is triggered.

[0019] Furthermore, the third prediction model outputs the process adjustment amount by including the following steps:

[0020] The third prediction model is a graph neural network. After the real-time production data is input into the third prediction model, the third prediction model obtains the data sensitivity, divides the real-time production data into high-sensitivity data and low-sensitivity data based on the data sensitivity, and predicts the process adjustment amount based on the high-sensitivity data.

[0021] Furthermore, calculating the deviation of the network parameters at the same location in the second prediction model for different factories includes the following steps:

[0022] After the number of the second prediction models reaches a first value, the cloud-based correction model obtains the network parameters of each of the second prediction models, aggregates the network parameters at the same location into a data group, selects one network parameter from the data group as the first data, and uses the remaining network parameters as the second data, calculates the average value of the second data, and uses the difference between the first data and the average value as the deviation.

[0023] Further, determining the adjustment value based on the deviation includes the following steps:

[0024] The number of data including network parameters in the data group is obtained. If the number of data is not greater than a third threshold, an adjustment coefficient is set. The deviation is reduced based on the adjustment coefficient to obtain a transition value. The network parameters are corrected based on the transition value to obtain the adjustment value.

[0025] If the amount of data exceeds the third threshold, density estimation is performed on the data based on the Gaussian kernel function. Based on the density estimation result, the mode of the network parameters in the data group is determined, the kurtosis coefficient of each mode is calculated, a correction coefficient is determined based on the kurtosis coefficient, the dominant mode corresponding to the network parameter is determined, and the distribution distance between the network parameter and the dominant mode is obtained. The network parameter is corrected based on the distribution distance and the correction coefficient to obtain the adjusted value. The difference in information entropy before and after the correction of the data group is taken as the entropy change. When the entropy change is less than the preset entropy tolerance, the correction of the data group is stopped; otherwise, the correction of the network parameters in the data group continues.

[0026] Furthermore, the graph neural network divides the real-time production data into high-sensitivity data and low-sensitivity data based on the data sensitivity, including the following steps:

[0027] If the data sensitivity is greater than the critical threshold, the real-time production data is classified as high-sensitivity data; otherwise, it is classified as low-sensitivity data.

[0028] The present invention also provides an intelligent correction system for deviations in heat treatment process parameters. This system is used to implement the methods described above, and includes:

[0029] An expansion module is used to acquire the original production data of each process in each factory, and to expand the original production data to obtain expanded data.

[0030] The training module is used to establish a first prediction model containing a dynamic adjustment mechanism, train the corresponding first prediction model locally based on the expanded data to obtain a second prediction model, calculate the deviation of network parameters at the same position of the second prediction model in different factories, if the deviation exceeds a first threshold, determine an adjustment value based on the deviation, correct the network parameters of each second prediction model at the same position based on the adjustment value, and determine whether to retain the corrected network parameters based on the change in prediction accuracy of the second prediction model before and after correction.

[0031] The prediction module defines the corrected second prediction model as the third prediction model. The prediction module inputs real-time production data into the third prediction model. The third prediction model determines the product type of the target product for this production. The dynamic adjustment mechanism adjusts the number of network layers according to the product type and outputs the process adjustment amount and the corresponding confidence level.

[0032] The correction module uploads real-time production data to the adjustment model in the cloud when the confidence level is less than the second threshold. The adjustment model then calls the historical case library to correct the process adjustment amount.

[0033] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the method described above.

[0034] The beneficial effects of this invention are as follows:

[0035] This invention first augments the original production data, increasing the number of training samples to ensure the model receives complete training. By establishing a prediction model with a dynamic adjustment mechanism, the model can automatically adjust the number of network layers based on product type, adapting to the differences in process characteristics of different products and balancing prediction speed and accuracy. By calculating and collaboratively correcting the deviations of network parameters at the same location in different factories, the impact of data deviations from a single factory on the model is reduced, improving the model's generalization ability across multiple scenarios. By calculating confidence levels and combining them with a historical case library to correct low-confidence results, the frequency of manually recalibrating error thresholds is reduced. Furthermore, through an architecture combining local training and cloud-based collaborative correction, cross-factory knowledge sharing is achieved while protecting factory data privacy, improving the accuracy and real-time performance of overall process parameter correction. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the steps of an intelligent correction method for deviations in heat treatment process parameters according to the present invention.

[0037] Figure 2 This is a schematic diagram of the structure of an intelligent correction system for deviations in heat treatment process parameters according to the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this application, a first script may be referred to as a second script, and similarly, a second script may be referred to as a first script.

[0040] like Figure 1As shown, a method for intelligent correction of deviations in heat treatment process parameters includes:

[0041] S1: Obtain the original production data of each process in each factory, expand the original production data, and obtain expanded data.

[0042] The following explanation uses steel manufacturing as an example. The manufacturing process includes coking, sintering, smelting, steelmaking, and rolling. To protect the privacy of factory production data as much as possible, this embodiment only collects data from within the factory and trains the prediction model locally. To obtain the influence relationships between process parameters, a graph neural network is used to establish the first prediction model. In other embodiments, a backpropagation (BP) neural network can also be used. Raw material characteristics, process parameters, and quality indicators are used as nodes in the graph neural network, and process adjustment amounts are obtained through fully connected layers within it. Considering the updates and iterations of steelmaking equipment and changes in environmental data, only production data from the last three years is used for training. Since only data from within the factory is used to train the model, the amount of data is relatively small. In addition, the limitation of the years further reduces the amount of usable data. Therefore, this embodiment needs to expand the original production data before training the model. The specific expansion method will be described in detail later.

[0043] S2: Establish a first prediction model with a dynamic adjustment mechanism, train the corresponding first prediction model locally based on the expanded data, obtain a second prediction model, and calculate the deviation of network parameters at the same location in the second prediction model of different factories.

[0044] The first prediction model is the pre-trained model, which includes a pre-set dynamic adjustment mechanism. This mechanism dynamically adjusts the number of layers in the neural network based on the production target. For example, the production target is divided into low-carbon steel, medium-carbon steel, and high-carbon steel. Low-carbon steel has a wide rolling temperature range, low deformation resistance, and relatively simple quality control. High-carbon steel has strict rolling temperature control, high deformation resistance, and requires precise control of production process parameters. When the production target is low-carbon steel, fewer network layers are used, such as two graph convolutional layers, to speed up the model's prediction. When the production target is high-carbon steel, more network layers are used, such as four graph convolutional layers. The more graph convolutional layers, the better the model can capture the influence relationships between distant nodes. In other words, using a more complex network structure can achieve higher prediction accuracy.

[0045] In this embodiment, calculating the deviation of network parameters at the same location in the second prediction model for different factories includes the following steps:

[0046] After the number of second prediction models reaches the first value, the cloud-based correction model obtains the network parameters of each second prediction model, aggregates the network parameters at the same location into a data group, selects one network parameter from the data group as the first data, and uses the remaining network parameters as the second data, calculates the average value of the second data, and uses the difference between the first data and the average value as the deviation.

[0047] Different factories have varying equipment status and environmental conditions. Local data may overfit specific equipment, raw materials, or operating habits, resulting in low generalization ability of the model. When production equipment is changed or environmental conditions change, the prediction accuracy of the model may be significantly affected. To solve this problem, it is necessary to detect and correct the network parameters of the second prediction model. Specifically, by calculating the deviation of parameters at the same location in different factory models, the network parameters are corrected based on the value of the deviation. The more second prediction models involved in the correction, the more reliable the correction result. In this embodiment, the first value is set to 10. When the number of second prediction models does not reach the first value, that is, when the number of factories with second prediction models deployed is less than 10, the factories that have already deployed second prediction models use their own trained second prediction models to correct the production process parameters.

[0048] When the number of second prediction models reaches a certain threshold, the cloud-based adjustment model acquires the network parameters of each second prediction model. For graph neural networks, the network parameters are the edge weights. The network parameters at the same position are selected from all the second prediction models where both ends of the nodes are identical; the weight of this edge is the network parameter. For example, edges connecting endpoints A and B are selected from all the second prediction models, and the weights of these edges are combined into a data set. For fully connected neural layers, "same position" refers to edges where neurons at both ends are identical.

[0049] Extract the first network parameter from the data set and use it as the first value. Calculate the average value of the remaining network parameters in the data set and use it as the second value. Calculate the difference between the first value and the second value and use this difference as the deviation of the first network parameter. The larger the deviation, the greater the difference between the first data and the distribution of the overall network parameters. Then continue to calculate the deviation of the second network parameter and repeat the process until the deviation of all network parameters in all data sets is calculated.

[0050] S3: If the deviation exceeds the first threshold, the adjustment value is determined based on the deviation, the network parameters of each second prediction model at the same position are corrected based on the adjustment value, and the change in the prediction accuracy of the second prediction model before and after the correction is determined to determine whether to retain the corrected network parameters.

[0051] The adjustment value refers to how much the network parameters need to be modified. This is achieved by adding the adjusted values ​​to the original network parameters. Each time the network parameters of the second prediction model need to be modified, the modified second prediction model is defined as the experimental model. The prediction performance of the experimental model is verified. Specifically, validation data, which is different from the training data, is pre-divided from the original production data. This validation data is input into both the second prediction model and the experimental model, and the accuracy of their outputs is compared. If the accuracy of the experimental model is greater than that of the second prediction model, the network parameter adjustment is retained. If the accuracy of the experimental model is significantly lower than that of the second prediction model, the network parameters are not modified at that point. This step avoids reducing the model's prediction accuracy after modifying the network parameters.

[0052] S4: Define the corrected second prediction model as the third prediction model. Input real-time production data into the third prediction model. The third prediction model determines the product type of the target product for this production. The dynamic adjustment mechanism adjusts the number of network layers according to the product type and outputs the process adjustment amount and the corresponding confidence level.

[0053] The modified second prediction model is defined as the third prediction model. Real-time production data is fed into the third prediction model. The real-time generated data includes the currently set process parameters, the target quality parameters of the product, the actual quality parameters of the produced product, environmental parameters, and raw material characteristic parameters. Based on the difference between the actual quality parameters and the target quality parameters, and combined with the current production process parameters, the third prediction model outputs the process adjustment amount and the corresponding confidence level. The confidence level represents the reliability of the process adjustment amount; the higher the confidence level, the higher the reliability of the process adjustment amount. Generally, a Softmax function can be added to the graph neural network to output the confidence level. This application obtains the confidence level by setting an additional validation model, which will be described in detail later.

[0054] S5: When the confidence level is less than the second threshold, the real-time production data is uploaded to the adjustment model in the cloud. The adjustment model calls the historical case library to correct the process adjustment amount.

[0055] When the confidence level is low, i.e., below the second threshold, it indicates that the reliability of the process adjustment amount output by the second prediction model is low. In this case, real-time production data is uploaded to the cloud. The cloud has an adjustment model that selects the optimal process adjustment amount from a historical case library through similarity comparison. If the optimal process adjustment amount is close to the process adjustment amount output by the third prediction model, the adjustment is made based on the process adjustment amount output by the third prediction model; otherwise, the adjustment is made manually. Notably, the prediction result of the third prediction model is not always low in confidence, so production data is only uploaded to the cloud occasionally, as the privacy of factory data is still largely protected.

[0056] In this embodiment, data augmentation of the original production data includes the following steps:

[0057] Obtain the first quantity of the original production data, set the interpolation quantity, determine the second quantity based on the first quantity and the interpolation quantity, and generate the second quantity of interpolated data with the same dimension as the original production data under the constraints of the production process, which is defined as the basic data.

[0058] There are several methods to determine the second quantity. For example, the second quantity can be determined by the first formula, which is N=M*(n+1), where N is the second quantity, M is the first quantity, and n is the interpolation quantity. When M is 1000 and n is 1, the second quantity is 2000. The generated interpolated data has the same data dimension as the original production data. For example, if the original production data is 40-dimensional data, including melting temperature, furnace pressure, and cooling rate, the generated interpolated data will also be 40-dimensional.

[0059] This embodiment generates interpolated data in the following way: First, based on the upper and lower limits of the data, it is divided into K subspaces, where the value of K can be set according to actual needs. For example, if the melting temperature is between 800-1200℃, it is divided into intervals such as [800, 840) and [840, 880). Then, a sample point is randomly selected in each interval. Finally, a random permutation method is used to combine the sample points of each dimension to avoid correlation between dimensions. After generating the interpolated data, the interpolated data is also verified using a process rule library. For example, if the process rule library contains a rule that the quenching temperature must not be lower than the tempering temperature by 150℃, and the generated interpolated data does not meet this rule, it is regenerated, and the finally retained interpolated data is defined as the basic data.

[0060] Select the basic data most similar to each original production data and define it as the first data. Calculate the first correction amount based on the similarity between the original production data and the first data and the data sensitivity of each dimension. Correct the first data based on the first correction amount to obtain the second data. Obtain multiple interpolated data most similar to the second data and define them as the third data. Determine the second correction amount based on the similarity between the second data and the third data. Correct the third data based on the second correction amount to obtain the fourth data. Calculate the similarity between the original production data and each of the second and fourth data again and update it. Repeat this step until a preset number of times is reached. Use the original production data and multiple data most similar to the original production data from the basic data as supplementary data.

[0061] After obtaining the basic data, the similarity between the original production data and the basic data is calculated. For example, if there are 1000 original production data entries and 2000 basic data entries, the similarity between each of the 1000 original production data entries and each of the 2000 basic data entries is calculated. This embodiment uses cosine similarity for calculation. Before calculating the similarity, the original production data and basic data need to be standardized to eliminate the influence of dimensions. After calculating the similarity between each piece of original production data and each piece of basic data, the basic data most similar to the original production data is selected as the first data.

[0062] Then, the first correction amount is calculated by combining similarity and data sensitivity. Specifically, the difference between each dimension of the original production data and the first data is first obtained. Then, the correction weight for each dimension is determined by data sensitivity. The higher the data sensitivity, the greater the impact of the data on product quality, and therefore the smaller the correction weight, avoiding excessive correction to the data values. Before training the graph neural network, data sensitivity can be obtained through statistical analysis of historical data, such as using the Pearson correlation coefficient between data points. After training the graph neural network, when data expansion is needed, parameter analysis of the graph neural network can be performed to determine a more accurate data sensitivity. The specific analysis method is existing technology and will not be described here.

[0063] In this embodiment, the first correction amount is calculated using the following second formula: Si = Ci * (1-X) * Wi, where Si is the first correction amount for the i-th dimension of the first data, C is the difference between the i-th dimension of the original production data and the i-th dimension of the first data, X is the similarity between the original production data and the first data, and Wi is the correction weight for the i-th dimension. In the second formula, the higher the similarity or the greater the data sensitivity, the smaller the first correction amount is obtained. That is, the greater the similarity, the smaller the correction value for the first data; the greater the data sensitivity, the smaller the correction value for that dimension, thus avoiding large corrections to sensitive data. The first data after the final correction is defined as the second data.

[0064] After obtaining the second data, several data points most similar to the second data are extracted from the base data, such as 10 or 20 data points most similar to the second data. These are defined as the third data. Based on the method of correcting the second data according to the first data, the third data is corrected using the second data to obtain the fourth data. The values ​​of the interpolated data are continuously updated in this way, making them increasingly closer to the original production data. The update stops after a preset number of times. Finally, several data points most similar to each original production data point are selected from the updated base data as target data. The original production data and the target data are used as supplementary data.

[0065] In this embodiment, the confidence level of the process adjustment amount output by the third prediction model includes the following steps:

[0066] The output of the third prediction model includes intermediate quality parameters for each process and final quality parameters after the product is completed. A validation model is set up, which includes multiple calculation formulas. These formulas calculate the intermediate quality parameters after each process and the final quality parameters after the product is completed based on real-time production data. The calculation formulas include multiple unknown coefficients. The validation model is trained based on the original production data to determine the values ​​of the unknown coefficients. Corrected data is obtained by correcting the real-time production data based on process adjustment. The corrected data is input into the validation model to calculate the intermediate and final quality parameters after each process. The intermediate and final quality parameters of the third prediction model and the validation model are compared to determine the confidence level.

[0067] As previously introduced, the steelmaking process involves multiple steps. After each step, the raw materials are processed into semi-finished products. The quality parameters of these semi-finished products are called intermediate quality parameters. For hot processing, there are clear boundaries between steps, making these intermediate quality parameters readily available. Finally, after production is complete, finished steel products are obtained, and the quality parameters of these finished products are called final quality parameters. In this embodiment, the intermediate and final quality parameters are used as the prediction results of the graph neural network in constructing the first prediction model.

[0068] To verify the aforementioned process adjustment amounts, this embodiment also establishes a verification model, specifically a formula library. This library includes multiple physical and empirical formulas for calculating various parameters of the production process, such as converter charge formulas, oxygen flow rate formulas, and formulas for the addition of various substances. These formulas contain unknown coefficients that need to be calibrated, requiring analysis of historical production data. Substituting the currently used process parameters into the calculation formulas yields intermediate and final quality parameters, and the results obtained through the verification model have high interpretability. Therefore, by correcting real-time production data using process adjustment amounts to obtain corrected data, and inputting this corrected data into the verification model for calculation, intermediate and final quality parameters can be obtained. If the output results of the third prediction model regarding quality parameters are similar to the output results of the verification model, it is determined that the output results of the third prediction model have a high confidence level. Specifically, the confidence level can be obtained as follows: first, calculate the first similarity of the two intermediate quality parameters and the second similarity of the two final quality parameters; then, use the average of the first and second similarities as the confidence level. The larger the average value, the higher the confidence level.

[0069] In this embodiment, adjusting the model by calling the historical case library to correct the process adjustment amount includes the following steps:

[0070] The historical case library includes multiple best production cases. The adjustment model calculates the similarity of raw material composition, equipment status, and environmental parameters between real-time production data and each production case. If the similarity of a production case and real-time production data in all three parameters is greater than the reliable threshold, then the process adjustment amount is generated based on the process parameters in that production case and sent to the factory; otherwise, a manual review process is triggered.

[0071] By using multidimensional similarity comparison, the reliability of the comparison results is ensured. For example, the process parameters in the production case can only be used when all three similarities are greater than 0.85, thus avoiding decision-making bias caused by single-dimensional matching.

[0072] In this embodiment, the process adjustment amount output by the third prediction model includes the following steps:

[0073] The third prediction model is a graph neural network. After real-time production data is input into the third prediction model, the third prediction model obtains the data sensitivity, divides the real-time production data into high-sensitivity data and low-sensitivity data based on the data sensitivity, and predicts the process adjustment amount based on the high-sensitivity data.

[0074] Graph neural networks classify real-time production data into high-sensitivity and low-sensitivity data based on data sensitivity, including the following steps:

[0075] If the data sensitivity is greater than the critical threshold, the real-time production data will be classified as high-sensitivity data; otherwise, it will be classified as low-sensitivity data.

[0076] The third prediction model includes a fully connected layer. After real-time production data is input into the third prediction model, it is divided into high-sensitivity data and low-sensitivity data based on the network parameters in the graph neural network. For example, if the data sensitivity is between 0 and 1, the critical threshold can be set to 0.6. Data with a sensitivity greater than 0.6 is classified as high-sensitivity data, and the rest are classified as low-sensitivity data. The specific critical threshold can be set according to actual needs. High-sensitivity data is data that has a significant impact on production quality. The fully connected layer predicts and calculates the process adjustment amount based on the value and sensitivity of the high-sensitivity data. Through this step, only high-sensitivity data is input into the model for processing and prediction, which can reduce computational complexity and increase data processing efficiency.

[0077] In this embodiment, determining the adjustment value based on the deviation includes the following steps:

[0078] After the number of second prediction models reaches the first value, the cloud-based correction model obtains the network parameters of each second prediction model, aggregates the network parameters at the same location into a data group, selects one network parameter from the data group as the first data, and uses the remaining network parameters as the second data, calculates the average value of the second data, and uses the difference between the first data and the average value as the deviation.

[0079] For example, if the third threshold is set to 20, when the number of network parameters in the data set is less than 20, the network parameters are adjusted based on the following method: First, set the adjustment coefficient, which can be set based on experience, for example, set the adjustment coefficient to 0.5. Then, obtain the deviation and multiply the deviation by the adjustment coefficient, that is, reduce the deviation by 1 / 2 to obtain the transition value. Finally, add the network parameter to the transition value to obtain the adjustment value.

[0080] When the number of second prediction models deployed exceeds 20, the diversity of network parameter distribution increases. Using only simple correction methods may cause the corrected network parameters to deviate far from the original values. Therefore, this embodiment also proposes the following technical solutions.

[0081] If the amount of data exceeds the third threshold, density estimation is performed on the data based on the Gaussian kernel function. Based on the density estimation results, the modes of network parameters in the data set are determined, the kurtosis coefficient of each mode is calculated, and the correction coefficient is determined based on the kurtosis coefficient. The dominant mode corresponding to the network parameter is determined, and the distribution distance between the network parameter and the dominant mode is obtained. The network parameter is corrected based on the distribution distance and the correction coefficient to obtain the adjustment value. The difference in information entropy before and after the data set is corrected is taken as the entropy change. When the entropy change is less than the preset entropy tolerance, the correction of the data set is stopped; otherwise, the correction of the network parameters in the data set continues.

[0082] The following example illustrates the above steps. First, a Gaussian kernel function is used for density estimation. This density estimation determines the modal distribution of the data, thus identifying whether the data is unimodal or multimodal. A unimodal distribution indicates that the network parameters have only one numerical tendency, while a multimodal distribution indicates that the network parameters have multiple numerical tendencies. For example, density estimation determines that the network parameters in the data set are multimodal, with values ​​of 0.8 and 0.3, indicating that some network parameters tend towards 0.8, and others towards 0.3. Here, these two distributions are defined as the dominant peak and the secondary peak, respectively. Next, multiple network parameters are selected on both sides of the dominant and secondary peaks to calculate their kurtosis coefficients. The kurtosis coefficient measures the steepness of the data distribution; a larger kurtosis value results in a more angular probability distribution, indicating a more concentrated distribution of network parameter values.

[0083] Next, a table is established to compare the kurtosis coefficient with the correction coefficient. The larger the kurtosis coefficient, the smaller the correction coefficient. For example, when the kurtosis coefficient is between 1.5 and S ≤ 2.5, the correction coefficient is 0.5. The larger the correction coefficient, the greater the correction ratio for the network parameters. If the dataset is multimodal, the distance between the network parameters and each mode is obtained. The distance is the absolute value of the difference between the network parameter and the mode. Based on the distance, the mode closest to the network parameter is selected as the dominant mode of that network parameter. For example, if the dominant mode of the network parameter is 0.8, the network parameter is corrected towards 0.8.

[0084] The network parameters can be corrected using the following formula: x'=x+D*F, where: xi' is the corrected network parameter, x is the original network parameter, D is the distribution distance between the network parameter and the dominant mode, and F is the correction coefficient.

[0085] After correcting all network parameters in the data set, the change in information entropy before and after the correction is calculated. As the network parameters are continuously corrected, the value of information entropy becomes smaller and smaller because the network parameters become more and more concentrated. When the entropy change is less than the preset entropy tolerance, and when the information entropy no longer changes significantly, it indicates that the distribution of network parameters tends to be stable. Otherwise, the network parameters are corrected again.

[0086] In this embodiment, if there is no dominant mode in the data group, the network parameters corresponding to the data group will not be modified.

[0087] The absence of a dominant mode indicates that the distribution of data is relatively uniform, suggesting that the network parameters at this point do not exhibit a regular distribution pattern. To avoid reducing the prediction accuracy of the prediction model, the network parameters at this point will not be modified.

[0088] like Figure 2 As shown, the present invention also provides an intelligent correction system for deviations in heat treatment process parameters. This system is used to implement the above-described method and includes:

[0089] The expansion module is used to acquire the original production data of each process in each factory, expand the original production data, and obtain expanded data.

[0090] The training module is used to establish a first prediction model with a dynamic adjustment mechanism, train the corresponding first prediction model locally based on the expanded data to obtain a second prediction model, calculate the deviation of network parameters at the same position of the second prediction models in different factories, if the deviation exceeds a first threshold, determine the adjustment value based on the deviation, correct the network parameters of each second prediction model at the same position based on the adjustment value, and determine whether to retain the corrected network parameters based on the change in the prediction accuracy of the second prediction models before and after the correction.

[0091] The prediction module defines the corrected second prediction model as the third prediction model. The prediction module inputs real-time production data into the third prediction model. The third prediction model determines the product type of the target product for this production. The dynamic adjustment mechanism adjusts the number of network layers according to the product type and outputs the process adjustment amount and the corresponding confidence level.

[0092] The correction module uploads real-time production data to the adjustment model in the cloud when the confidence level is less than the second threshold. The adjustment model then calls the historical case library to correct the process adjustment amount.

[0093] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the method described above.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent correction of process parameter deviation for thermal processes, characterized in that, The method comprises the following steps: obtaining original production data of each process of each factory, and performing data augmentation on the original production data to obtain augmented data; establishing a first prediction model comprising a dynamic adjustment mechanism, training the corresponding first prediction model locally based on the augmented data to obtain a second prediction model, and calculating the deviation of network parameters at the same position of the second prediction model of different factories; if the deviation exceeds a first threshold value, determining an adjustment value based on the deviation, correcting the network parameters of the second prediction model at the same position based on the adjustment value, and determining whether to retain the corrected network parameters based on the change in prediction accuracy of the second prediction model before and after correction; defining the second prediction model after correction as a third prediction model, inputting real-time production data into the third prediction model, and determining the product type of the target product of this production by the third prediction model, wherein the dynamic adjustment mechanism adjusts the number of network layers according to the product type, and outputs a process adjustment amount and a corresponding confidence level; when the confidence level is less than a second threshold value, uploading the real-time production data to a cloud adjustment model, and calling a historical case library to correct the process adjustment amount by the adjustment model; calculating the deviation of network parameters at the same position of the second prediction model of different factories comprises the following steps: after the number of second prediction models reaches a first value, a cloud correction model obtains the network parameters of each second prediction model, aggregates the network parameters at the same position into a data group, selects one network parameter as first data and the remaining network parameters as second data in the data group, calculates the average value of the second data, and takes the difference between the first data and the average value as the deviation; determining the adjustment value based on the deviation comprises the following steps: obtaining the number of data including network parameters in the data group, if the number of data is not greater than a third threshold value, setting an adjustment coefficient, reducing the deviation based on the adjustment coefficient to obtain a transition value, correcting the network parameters based on the transition value to obtain the adjustment value; if the number of data is greater than the third threshold value, performing density estimation on the data based on a Gaussian kernel function, determining the mode of the network parameters in the data group based on the density estimation result, calculating the kurtosis coefficient of each mode, determining a correction coefficient based on the kurtosis coefficient, determining the main mode corresponding to the network parameters, and obtaining the distribution distance between the network parameters and the main mode, correcting the network parameters based on the distribution distance and the correction coefficient to obtain the adjustment value, taking the difference in information entropy before and after correction of the data group as an entropy change, and stopping the correction of the data group when the entropy change is less than a preset entropy tolerance, otherwise continuing the correction of the network parameters in the data group.

2. The method of claim 1, wherein, The data augmentation on the original production data comprises the following steps: obtaining a first number of the original production data, setting an interpolation number, determining a second number based on the first number and the interpolation number, generating interpolation data of the second number under the constraint of the production process and having the same dimension as the original production data, and defining the interpolation data as basic data; Screening the basic data most similar to each of the original production data and defining as first data, calculating a first correction amount based on the similarity of the original production data and the first data and the data sensitivity of each dimension data, modifying the first data based on the first correction amount, obtaining second data, obtaining a plurality of interpolation data most similar to the second data and defining as third data, determining a second correction amount based on the similarity of the second data and the third data, obtaining fourth data after modifying the third data based on the second correction amount, recalculating and updating the similarity between the original production data and each of the second data and the fourth data, repeating the step until a preset number of times is reached, and taking the original production data and a plurality of data in the basic data most similar to the original production data as the expanded data.

3. The method of claim 1, wherein, The third prediction model outputs the process adjustment amount includes the following steps: The output result of the third prediction model includes the intermediate quality parameter of each process and the final quality parameter after the product completes production, a verification model is set, the verification model includes a plurality of calculation formulas, the calculation formulas calculate the intermediate quality parameter and the final quality parameter after the product completes production according to the real-time production data, a plurality of unknown coefficients are included in the calculation formulas, the verification model is trained based on the original production data to determine the values of the unknown coefficients, the modified data is obtained after the real-time production data is modified based on the process adjustment amount, the modified data is input into the verification model, the intermediate quality parameter and the final quality parameter after each process are calculated, and the intermediate quality parameter and the final quality parameter of the third prediction model and the verification model are compared to determine the confidence.

4. The method of claim 3, wherein, The adjustment model calls a historical case library to modify the process adjustment amount includes the following steps: The historical case library includes a plurality of best production cases, the adjustment model calculates the similarity of raw material composition, equipment state and environmental parameters between the real-time production data and each production case, if the similarity of the three parameters of a production case and the real-time production data is greater than a reliable threshold, the process adjustment amount is generated based on the process parameters in the production case and is issued to the factory, otherwise an artificial audit process is triggered.

5. The method of claim 2, wherein, The third prediction model outputs the process adjustment amount includes the following steps: The third prediction model is a graph neural network, after the real-time production data is input into the third prediction model, the third prediction model obtains the data sensitivity, divides the real-time production data into high-sensitivity data and low-sensitivity data based on the data sensitivity, and predicts the process adjustment amount based on the high-sensitivity data.

6. The method of claim 5, wherein, The graph neural network divides the real-time production data into high-sensitivity data and low-sensitivity data based on the data sensitivity includes the following steps: If the data sensitivity is greater than a critical threshold, the real-time production data is divided into the high-sensitivity data, otherwise it is divided into the low-sensitivity data.

7. A system for intelligent correction of process parameter deviations for thermal processes for implementing the method according to any one of claims 1 to 6, characterized in that The system comprises: An expansion module is configured to obtain original production data of each process of each factory, expand the original production data, and obtain expanded data; A training module is configured to establish a first prediction model including a dynamic adjustment mechanism, train the corresponding first prediction model locally based on the expanded data, obtain a second prediction model, calculate a deviation of network parameters at the same position of the second prediction model of different factories, determine an adjustment value based on the deviation if the deviation exceeds a first threshold, correct the network parameters of each second prediction model at the same position based on the adjustment value, and determine whether to retain the corrected network parameters based on a change in prediction accuracy of the second prediction model before and after correction. A prediction module defines the second prediction model after correction as a third prediction model, inputs real-time production data into the third prediction model, determines a product type of a target product of this production by the third prediction model, adjusts the number of network layers according to the product type by the dynamic adjustment mechanism, and outputs a process adjustment amount and a corresponding confidence. A correction module uploads the real-time production data to a cloud adjustment model when the confidence is less than a second threshold, and the adjustment model corrects the process adjustment amount by calling a historical case library.

8. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement the method of any one of claims 1-6.

Images