Strong coupling technological process prediction method based on large model fine tuning

Abstract

The invention discloses a strong coupling process prediction method based on large model fine tuning, and relates to the technical field of industrial process intelligent optimization, and the method comprises the steps: S1, carrying out the cleaning, coupling feature extraction, time sequence standardization and VAE small sample enhancement of strong coupling process original data, and obtaining standardized time sequence data; s2, selecting a Transform architecture industrial time sequence pre-training large model as a base model; s3, through a layered freezing-fine tuning mechanism, freezing a bottom-layer general knowledge layer, performing fine tuning on a middle-layer coupling feature adaptation layer and a top-layer multi-branch output layer, and completing model adaptation in combination with a weighted loss function containing a coupling constraint regular term; and S4, inputting data to obtain a prediction result, and realizing rapid iteration of the model through increment fine tuning. The method is suitable for strong coupling scenes of chemical engineering, metallurgy and the like, an iron-making and steel-making process is taken as an experimental case, the predicted MAE of the molten iron temperature reaches 9.2 DEG C, the predicted MAE of the carbon content of the molten steel reaches 0.012%, the generalization accuracy rate is 93%, the prediction precision and the industrial suitability are remarkably improved, and the production energy consumption and cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology for industrial processes, specifically to a process prediction method for multivariate strongly coupled characteristics, which is particularly suitable for strongly coupled industrial scenarios such as continuous chemical reactions and multi-component smelting in metallurgy. It achieves high-precision prediction of process parameters and product quality through fine-tuning of a large model. Background Technology

[0002] Traditional methods for predicting strongly coupled processes suffer from the following technical challenges:

[0003] Traditional mechanistic models have drawbacks: mechanistic models that rely on physical or chemical formulas are difficult to quantify the interactions between variables in strongly coupled scenarios, have high model complexity and poor adaptability, and exhibit significant prediction errors when faced with process fluctuations.

[0004] Traditional machine learning models have limitations: Random forests, ordinary neural networks such as CNN and LSTM require a large amount of labeled domain data for training. However, in industrial scenarios, the situation of "small sample size + high noise" is common, resulting in weak model generalization ability and inability to be transferred to different process conditions.

[0005] Problems with the direct application of general large models: Although general large models are quite mature in handling language and image problems at present, and have the "knowledge transfer ability" of learning from massive amounts of data, they have not been optimized for the domain characteristics of "strongly coupled processes", resulting in problems such as "domain knowledge disconnect" and large deviations between prediction results and industrial reality.

[0006] The shortcomings of existing fine-tuning methods: Existing large model fine-tuning is mostly aimed at general domains such as text and images, and no fine-tuning strategy has been designed to adapt to the temporal, high-dimensional and strongly correlated nature of strongly coupled process data. For example, directly using full fine-tuning will lead to overfitting and high training costs. Summary of the Invention

[0007] The purpose of this invention is to provide a strongly coupled process prediction method based on large model fine-tuning to overcome the above-mentioned defects.

[0008] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0009] A strongly coupled process prediction method based on large model fine-tuning includes the following steps;

[0010] 1) Obtain the raw data of the strongly coupled process and preprocess it to obtain standardized time series data;

[0011] 2) Select a large pre-trained industrial time series model with time series modeling capabilities and industrial knowledge transfer capabilities as the base model;

[0012] 3) Construct a hierarchical freeze-fine-tuning mechanism to perform domain-adaptive fine-tuning on the base model to obtain a strongly coupled process prediction model;

[0013] 4) Input standardized time series data into the strongly coupled process prediction model to obtain the target parameter prediction results; at the same time, establish a model iterative update mechanism, regularly collect actual production data of the strongly coupled process, and use incremental fine-tuning to update the parameters of the strongly coupled process prediction model.

[0014] Step 1) includes the following steps:

[0015] 1.1) Data cleaning: The original data is processed sequentially using the 3σ criterion and the DBSCAN clustering algorithm to remove outliers; an interpolation method based on parameter coupling correlation is used to calculate the coupling correlation between missing parameters and other parameters, and the 1-2 parameters with the highest correlation are selected to construct an interpolation model, which is then used to fill in the missing values.

[0016] 1.2) Coupling Feature Extraction: The mutual information entropy algorithm is used to calculate the correlation between the original input parameters and the target prediction parameters in the cleaned data, and parameters with a correlation higher than the threshold are retained as core coupling variables; the core coupling variables are decomposed using partial least squares to obtain multiple uncorrelated principal component factors;

[0017] 1.3) Data Time Series and Standardization: The principal component factors are divided into time series windows according to the process cycle to form a time series input format of "[number of samples, time step, feature dimension]", and the time series is standardized using the Z-score standardization method.

[0018] Step 3) includes the following steps:

[0019] 3.1) Divide the base model of the Transformer architecture into three functional layers according to network depth: bottom layer, middle layer, and top layer;

[0020] 3.2) Freeze the underlying layers of the base model to preserve the general temporal patterns learned by the large model;

[0021] 3.3) Fine-tuning the middle layer of the base model by introducing a coupled attention module and using a weight allocation mechanism to assign weights to the principal component factors;

[0022] 3.4) Fine-tune the top layer of the base model, integrate the output of the middle layer with the actual constraints of the strongly coupled process, and add a coupling constraint regularization term to the weighted loss function during the training process of the base model so that the model actively avoids prediction results that violate process constraints during training, and outputs the prediction results through the fully connected layer.

[0023] Step 3.1) specifically refers to:

[0024] The first 60% to 70% of the Transformer architecture's network layers are used as the bottom layer to learn general temporal patterns. The middle 15% to 20% of the Transformer architecture's network layers are used as the middle layer to perform coupled feature adaptation. The last 10% to 15% of the Transformer architecture's network layers are used as the top layer to predict the output. A multi-branch structure design is adopted to adapt to the needs of multi-target prediction.

[0025] The weight allocation mechanism is as follows:

[0026] A parameter correlation matrix is ​​constructed by input and output parameters. The mutual information entropy of each pair of parameters in the parameter correlation matrix is ​​calculated. The mutual information entropy is normalized to obtain the basic correlation weight between each parameter and the target prediction parameter.

[0027] A dynamic adjustment factor is introduced to adjust the basic association weights in real time based on the prediction error feedback during model training, forming a dynamic attention weight matrix. The value range of the dynamic adjustment factor is 0.8 to 1.2. When the prediction error corresponding to a certain parameter increases, the adjustment factor is greater than 1 to increase the weight of that parameter, and vice versa.

[0028] The incremental fine-tuning method is as follows:

[0029] A variational autoencoder is used to construct a sample generation model. The cleaned original data is used as training samples to train the sample generation model, learn the parameter distribution law and coupling relationship characteristics of the strongly coupled process, generate virtual samples that conform to the process characteristics through the trained sample generation model, and fuse the virtual samples with the cleaned original data to form an enhanced sample set.

[0030] During incremental fine-tuning, a coupling constraint regularization term is added to the weighted loss function as a conditional penalty term. When the predicted parameter exceeds the process threshold, the loss value is increased; otherwise, the loss value is decreased. The fine-tuning loss function allocates loss weights based on the process importance of each target prediction parameter. The sum of the weight coefficients of all parameters is 1. The mean square error is used to calculate the single-parameter prediction loss, and the total loss is obtained by weighted summation.

[0031] The present invention has the following beneficial effects and advantages:

[0032] 1. This invention uses techniques such as coupled feature extraction and layered freeze-fine-tuning to control the prediction error of key parameters of strongly coupled processes within the industrially permissible range, improving the accuracy by more than 60% compared with traditional LSTM and XGBoost models, thus meeting the stringent requirements of production for parameter control.

[0033] 2. This invention utilizes a fusion design of VAE small sample enhancement and process constraints to achieve a generalization accuracy of over 90% under complex operating conditions such as raw material fluctuations and equipment load adjustments, thus solving the pain point of poor adaptability of traditional models in scenarios with small samples and multiple disturbances.

[0034] 3. This invention supports direct connection with industrial DCS systems, incremental fine-tuning can be completed within 30 minutes, the prediction results meet process compliance requirements, improve training efficiency, and reduce the cost of industrial intelligent upgrades. Attached Figure Description

[0035] Figure 1 Overall architecture diagram of the invention;

[0036] Figure 2 The core flowchart of this invention;

[0037] Figure 3 Flowchart of the layered freezing-fine-tuning mechanism of this invention;

[0038] Figure 4 The flowchart of the model iterative optimization of this invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0040] like Figure 1 As shown, the overall framework of this invention is divided into four parts: a data preprocessing layer, a large model pre-training layer, a domain adaptation fine-tuning layer, and a prediction and inference layer. Figure 2 As shown, the specific steps are as follows:

[0041] Step 1: Perform data preprocessing on strongly coupled process data to solve the problems of poor data quality and difficulty in extracting coupling relationships.

[0042] The core characteristics of strongly coupled process data are time-series nature, high dimensionality, strong correlation, and high noise. This data includes multi-dimensional input parameters and corresponding target prediction parameters for the strongly coupled process. The strongly coupled process is an industrial continuous production process with multiple parameters interacting and significant dynamic interference. To address these data characteristics, targeted processing is performed on the data. The multi-dimensional input parameters for the strongly coupled process include, but are not limited to, material characteristic parameters, equipment operating parameters, process operation parameters, and environmental impact parameters. The target prediction parameters include, but are not limited to, product quality parameters, process status parameters, and energy efficiency parameters.

[0043] (1) Data cleaning: The “3σ criterion + DBSCAN clustering algorithm” is used to remove sensor outliers; for missing values, the “interpolation method based on coupling relationship” is used. By calculating the coupling correlation between the missing parameter and other parameters, the 1-2 parameters with the highest correlation are selected to construct an interpolation model to achieve accurate filling of missing values, rather than traditional mean interpolation. The outlier removal adopts a combination strategy of 3σ criterion and DBSCAN clustering algorithm. The 3σ criterion initially screens suspected outliers that deviate from the statistical distribution, and the DBSCAN clustering algorithm further confirms and removes isolated outliers based on data density characteristics; the missing value filling adopts an interpolation method based on parameter coupling correlation. By calculating the coupling correlation between the missing parameter and other parameters, the 1-2 parameters with the highest correlation are selected to construct an interpolation model to achieve accurate filling of missing values.

[0044] (2) Coupling Feature Extraction: Core coupling variables are selected using mutual information entropy (MI) + partial least squares (PLS):

[0045] a. The mutual information entropy algorithm is used to calculate the correlation degree between each original input parameter and the target prediction parameter. A correlation degree threshold is set, and parameters with a correlation degree higher than the threshold are retained as core coupling variables, while redundant and irrelevant parameters are eliminated. The correlation degree threshold is set according to the number and complexity of parameters in the strongly coupled process, and the value range is 0.3~0.5.

[0046] b. Use PLS to decompose high-dimensional coupled variables and extract "principal component factors" (e.g., convert "temperature, pressure, and reactant concentration" into three uncorrelated principal components, retaining more than 95% of the original information) to reduce the dimensionality of subsequent model inputs.

[0047] (3) Data time series and standardization: The processed data is divided into time series windows according to the process cycle to form a time series input format of "[number of samples, time step, feature dimension]"; "Z-score standardization" is used to eliminate the influence of dimensions. The standardization formula is: z=(x-μ) / σ, where x is the original data, μ is the data mean, and σ is the data standard deviation.

[0048] Step 2: Select a pre-trained large model that is compatible with the strongly coupled process to solve the problem of insufficient domain knowledge in general large models.

[0049] Pre-trained models are the foundation for fine-tuning. It is necessary to select or improve models with "temporal modeling capability + industrial knowledge transfer capability". Core options include: large industrial time series models, large general time series models, and self-developed hybrid architecture models.

[0050] This invention selects the large-scale industrial time series model ProcessBERT, which uses a Transformer encoder as its base model and has been pre-trained with massive amounts of industrial time series data, possessing basic time series modeling capabilities. By adding a coupling relationship attention mechanism, the learning of cross-influences between variables is strengthened.

[0051] Step 3: Large Model Fine-Tuning Strategy Guided by Strongly Coupled Processes

[0052] This invention addresses the characteristics of "limited data in strongly coupled processes and complex variable relationships" by designing a hierarchical fine-tuning scheme. This avoids the shortcomings of full-scale fine-tuning or simple freezing, and solves the problems of "poor fine-tuning adaptability and overfitting on small samples." Figure 3 As shown, the base model of the Transformer architecture is divided into three functional layers according to network depth: bottom layer, middle layer, and top layer. The bottom layer consists of the first 60% to 70% of the network layers, which is used to learn general temporal patterns and performs a complete freeze. The middle layer consists of the middle 15% to 20% of the network layers, which serves as a coupled feature adaptation layer, with a built-in coupled attention module and performs full parameter fine-tuning. The top layer consists of the last 10% to 15% of the network layers, which serves as the prediction output layer. It adopts a multi-branch structure design to adapt to multi-target prediction needs and performs full parameter fine-tuning. The coupled attention module dynamically allocates the feature contribution of different parameters by calculating the attention weight matrix between parameters.

[0053] (1) Layered Freeze-Fine-Tuning Strategy:

[0054] a. Freeze the bottom layer of the pre-trained model (the first 6 layers of the Transformer): preserve the general temporal patterns of large model learning.

[0055] b. Fine-tuning the middle layer of the model, “coupled feature adaptation layer”: Introducing a coupled attention module to assign weights to the principal component factors extracted in step 1, and setting different weight coefficients according to the degree of influence of input parameters on output parameters to strengthen the learning of domain coupling relationships;

[0056] The weight allocation mechanism of the coupled attention module is as follows: First, a parameter correlation matrix is ​​constructed, and the basic correlation weights between each parameter and the target prediction parameter are determined based on the mutual information entropy calculation results. Then, a dynamic adjustment factor is introduced, and the basic correlation weights are adjusted in real time according to the prediction error feedback during the model training process to form a dynamic attention weight matrix. The value range of the dynamic adjustment factor is 0.8~1.2. When the prediction error corresponding to a certain parameter increases, the adjustment factor is greater than 1 to increase the weight of the parameter, and vice versa.

[0057] Construct a parameter correlation matrix and determine the basic correlation weights between each parameter and the target prediction parameter based on the mutual information entropy calculation results. Specifically, construct an (m+n)×(m+n) dimensional matrix using m input parameters and n output parameters, calculate the mutual information entropy of each pair of parameters using the mutual information entropy algorithm, and the larger the value, the tighter the correlation. Normalize the mutual information entropy to obtain the correlation weight (if the minimum value of all MI values ​​is 0.1 and the maximum value is 0.7, then the normalized value of MI=0.62 is (0.62-0.1) / (0.7-0.1)=0.52 / 0.6≈0.87).

[0058] c. Fine-tuning the top-level "prediction output layer": The mid-level output is fused with the "process constraints" and the prediction results are output through a fully connected layer.

[0059] (2) Small sample augmentation and regularization:

[0060] a. A generative data augmentation approach is adopted, which generates virtual samples that conform to the characteristics of strong coupling based on VAE, increasing the sample size of the domain by 2-3 times and solving the problem of overfitting with small samples. The specific process is as follows: a sample generation model is constructed using variational autoencoder (VAE), and the cleaned original data is used as training samples to train the sample generation model to learn the parameter distribution law and coupling relationship characteristics of the strong coupling process; virtual samples that conform to the process characteristics are generated through the sample generation model, and the virtual samples are fused with the original samples to form an augmented sample set. The augmented sample size is 2-3 times that of the original sample size, which is used to improve the generalization ability of subsequent model fine-tuning.

[0061] b. Add a "coupling constraint regularization term" to the fine-tuning loss function: For example, when the predicted parameter exceeds the process threshold, increase the loss value to ensure the prediction result conforms to industrial reality. The fine-tuning loss function assigns loss weights based on the process importance of each target prediction parameter; parameters with higher importance are assigned higher weight coefficients. The sum of the weight coefficients of all parameters is 1. The mean squared error (MSE) is used to calculate the single-parameter prediction loss, and the weighted sum is used to obtain the total loss. The coupling constraint regularization term is a conditional penalty term. A reasonable process range for each target prediction parameter is preset. When the prediction result exceeds the reasonable range, an additional penalty is applied to the total loss based on the magnitude of the exceedance. The penalty coefficient ranges from 0.2 to 0.3; the larger the exceedance, the larger the penalty coefficient.

[0062] Step 4: Predictive Inference and Iterative Model Optimization

[0063] (1) Real-time prediction: Input the pre-processed real-time process data into the fine-tuned large model and output the process parameter prediction results for the next 5-30 minutes. The delay is controlled within 100ms to meet the real-time requirements of industry.

[0064] (2) Model iteration: The deviation between actual process data and prediction results is collected periodically, and the model is updated by incremental fine-tuning. Only the top output layer is fine-tuned with new data, without the need for full training, to ensure long-term prediction accuracy.

[0065] like Figure 4 As shown, the specific execution process of the incremental fine-tuning is as follows: First, an incremental dataset is constructed, newly added actual production data is filtered, and a preprocessing process consistent with step S1 is executed; then, the parameters of the bottom and middle layers of the strongly coupled process prediction model are fixed, and only the parameters of the top prediction output layer are unfrozen; using the incremental dataset as training data, the model is trained using a mini-batch gradient descent algorithm, with a batch size of 16~32 and a learning rate of 1e. -5 ~5e -5 The number of training iterations is determined based on the amount of incremental data, ensuring that the training time is controlled within 30 minutes, and enabling rapid model iteration and updates.

[0066] Example:

[0067] This invention uses the prediction of molten iron temperature and molten steel carbon content in a strongly coupled ironmaking and steelmaking process as an implementation example, providing detailed parameters, steps, and experimental data. Specific contents include:

[0068] (1) Experimental scenario and data

[0069] 1) Scenario: A large steel plant's continuous production process of "blast furnace ironmaking - converter steelmaking" requires prediction of two strongly coupled key parameters:

[0070] a. The temperature of molten iron during the blast furnace ironmaking stage;

[0071] b. Carbon content of molten steel at the end of the converter steelmaking stage.

[0072] Among them, there is a strong coupling relationship between ironmaking and steelmaking: if the temperature of molten iron is too low, the energy consumption for heating the converter will increase. The fluctuation of the initial carbon content of molten iron will directly affect the decarburization time and the accuracy of the final carbon content in the converter. The cross-influence between variables is significant. The main parameter relationship is: blast furnace blast volume → molten iron temperature → converter oxygen consumption → carbon content of molten steel.

[0073] 2) Data: Production data for three consecutive months was collected through the factory's DCS system, with a sampling frequency of 1 minute / time, resulting in a total of 14,400 sets of historical data (including 200 sets of abnormal data caused by equipment maintenance and raw material fluctuations).

[0074] a. Input variables: Ironmaking stage: Iron ore grade (56%~62%), fixed carbon content of coke (85%~88%), blast furnace blast volume (2200~2600 m³ / h). 3 / min), hot blast temperature (1100~1250℃), furnace top pressure (180~220kPa); steelmaking stage: molten iron charge (80~100t), initial carbon content of molten iron (4.2%~4.8%), oxygen flow rate (280~320m³ / min), hot blast temperature (1100~1250℃), furnace top pressure (180~220kPa); 3 / h), lime addition (4~6t / furnace), oxygen blowing time (12~18min);

[0075] b. Target variables: molten iron temperature (1450~1550℃), final carbon content of molten steel (0.08%~0.25%).

[0076] c. Data partitioning: 11,520 sets (80%) were used for fine-tuning of the large model, and 2,880 sets (20%) were used for testing. The test set included three different operating conditions (normal production, raw material grade fluctuation, and equipment load adjustment) to verify the model's generalization ability.

[0077] (2) Model parameter settings

[0078] 1) Pre-trained model: The ProcessBERT industrial time series pre-trained model was selected;

[0079] 2) Data preprocessing parameters:

[0080] a. Outlier handling: The “3σ criterion + DBSCAN clustering algorithm” is used to remove outliers such as molten iron temperature fluctuations exceeding 50°C within 1 minute and molten steel carbon content fluctuations >0.3% within 1 minute.

[0081] b. Coupling Feature Extraction: The correlation between the input variables and the target variables is calculated using mutual information entropy (MI), and core variables with MI values ​​> 0.4 are retained; then, the dimensionality of the 10 input variables is reduced to 5 principal component factors through partial least squares (PLS), retaining 95% of the original information;

[0082] c. Data augmentation: A VAE (variable autoencoder) was used to generate virtual samples that conform to the coupling characteristics to simulate the correlation between blast volume and molten iron temperature when the iron ore grade drops from 56% to 54%, increasing the number of fine-tuning samples from 11,520 to 23,040.

[0083] 3) Fine-tuning parameters:

[0084] a. Layered Freezing - Fine-tuning Strategy: Freeze the first 6 layers of ProcessBERT (retaining general industrial timing knowledge), fine-tune the middle 4 layers "coupled feature adaptation layer" (add a coupled attention module, set the weight of the iron temperature-related variable to 0.7, and the weight of the steel carbon content-related variable to 0.65), and fine-tune the top 2 layers "predictive output layer", outputting the iron temperature and steel carbon content in two branches;

[0085] b. Training parameters: Learning rate 3e -5 The batch size is 32, the training epochs are 25, the optimizer is AdamW, and the weight decay coefficient is 0.01.

[0086] c. Loss function: The "weighted MSE loss" is adopted. The loss weight for predicting molten iron temperature is set to 0.4, and the loss weight for predicting molten steel carbon content is set to 0.6. Coupled constraint regularization terms are added. When the molten iron temperature is <1450℃, the loss value increases by 20% to ensure that it meets the industrial production threshold.

[0087] (3) Experimental results and comparison

[0088] The experiment compares the proposed method with traditional methods such as LSTM, XGBoost, and the general large model GPT-4 Turbo to verify its advantages in predicting strongly coupled processes in ironmaking and steelmaking. Key indicators are shown in the table below:

[0089] Prediction methods Mean Absolute Error (MAE) for predicting molten iron temperature Prediction of carbon content in molten steel (MAE) Generalization accuracy (TFe 60%→56%) Training time XGBoost 32.5℃ 0.045% 72% 15min LSTM 25.8℃ 0.038% 79% 40min GPT-4 Turbo (Direct Prediction) 28.1℃ 0.042% 75% None (API call) Method of the present invention 9.2℃ 0.012% 93% 90 minutes (initial fine-tuning)

[0090] (4) Experimental conclusions

[0091] 1) The method of this invention reduces the prediction error of molten iron temperature to within ±10℃, which is 64.3% higher than the LSTM model; and controls the prediction error of carbon content in molten steel to within ±0.015%, which is 68.4% higher than the LSTM model, thus meeting the control accuracy requirements of key parameters in industrial production.

[0092] 2) Under the fluctuating condition of raw iron ore grade decreasing from 60% to 56%, the generalization accuracy of this patented method, through "VAE small sample enhancement + partial least squares (PLS) dimensionality reduction", still reaches 90%, which is 28.6% higher than the LSTM model. It can cope with the typical scenario of "small sample + multiple interference" in steel production and verify the strong coupling and adaptation capability under small sample conditions.

[0093] 3) After the initial fine-tuning, an incremental fine-tuning strategy is adopted, using only 1,000 newly collected data sets to fine-tune the top output layer, reducing the training time to 25 minutes, which meets the rapid iteration needs of industrial scenarios.

Claims

1. A strongly coupled process prediction method based on large model fine-tuning, characterized in that, Includes the following steps; 1) Obtain the raw data of the strongly coupled process and preprocess it to obtain standardized time series data; 2) Select a large pre-trained industrial time series model with time series modeling capabilities and industrial knowledge transfer capabilities as the base model; 3) Construct a hierarchical freeze-fine-tuning mechanism to perform domain-adaptive fine-tuning on the base model to obtain a strongly coupled process prediction model; 4) Input standardized time series data into the strongly coupled process prediction model to obtain the target parameter prediction results; at the same time, establish a model iterative update mechanism, regularly collect actual production data of the strongly coupled process, and use incremental fine-tuning to update the parameters of the strongly coupled process prediction model.

2. The strongly coupled process prediction method based on large model fine-tuning according to claim 1, characterized in that, Step 1) includes the following steps: 1.1) Data cleaning: The original data is processed sequentially using the 3σ criterion and the DBSCAN clustering algorithm to remove outliers; an interpolation method based on parameter coupling correlation is used to calculate the coupling correlation between missing parameters and other parameters, and the 1-2 parameters with the highest correlation are selected to construct an interpolation model, which is then used to fill in the missing values. 1.2) Coupling Feature Extraction: The mutual information entropy algorithm is used to calculate the correlation between the original input parameters and the target prediction parameters in the cleaned data, and parameters with a correlation higher than the threshold are retained as core coupling variables; the core coupling variables are decomposed using partial least squares to obtain multiple uncorrelated principal component factors; 1.3) Data temporalization and standardization: The principal component factors are divided into time series windows according to the process cycle to form a time series input format of "[number of samples, time step, feature dimension]", and the time series is standardized using the Z-score standardization method.

3. The strongly coupled process prediction method based on large model fine-tuning according to claim 1, characterized in that, Step 3) includes the following steps: 3.1) Divide the base model of the Transformer architecture into three functional layers according to network depth: bottom layer, middle layer, and top layer; 3.2) Freeze the underlying layers of the base model to preserve the general temporal patterns learned by the large model; 3.3) Fine-tuning the middle layer of the base model by introducing a coupled attention module and using a weight allocation mechanism to assign weights to the principal component factors; 3.4) Fine-tune the top layer of the base model, integrate the output of the middle layer with the actual constraints of the strongly coupled process, and add a coupling constraint regularization term to the weighted loss function during the training process of the base model so that the model actively avoids prediction results that violate process constraints during training, and outputs the prediction results through the fully connected layer.

4. The strongly coupled process prediction method based on large model fine-tuning according to claim 3, characterized in that, Step 3.1) specifically refers to: The first 60% to 70% of the Transformer architecture's network layers are used as the bottom layer to learn general temporal patterns. The middle 15% to 20% of the Transformer architecture's network layers are used as the middle layer to perform coupled feature adaptation. The last 10% to 15% of the Transformer architecture's network layers are used as the top layer to predict the output. A multi-branch structure design is adopted to adapt to the needs of multi-target prediction.

5. The strongly coupled process prediction method based on large model fine-tuning according to claim 3, characterized in that, The weight allocation mechanism is as follows: A parameter correlation matrix is ​​constructed by input and output parameters. The mutual information entropy of each pair of parameters in the parameter correlation matrix is ​​calculated. The mutual information entropy is normalized to obtain the basic correlation weight between each parameter and the target prediction parameter. A dynamic adjustment factor is introduced to adjust the basic association weights in real time based on the prediction error feedback during model training, forming a dynamic attention weight matrix. The value range of the dynamic adjustment factor is 0.8 to 1.

2. When the prediction error corresponding to a certain parameter increases, the adjustment factor is greater than 1 to increase the weight of that parameter, and vice versa.

6. The strongly coupled process prediction method based on large model fine-tuning according to claim 1, characterized in that, The incremental fine-tuning method is as follows: A variational autoencoder is used to construct a sample generation model. The cleaned original data is used as training samples to train the sample generation model, learn the parameter distribution law and coupling relationship characteristics of the strongly coupled process, generate virtual samples that conform to the process characteristics through the trained sample generation model, and fuse the virtual samples with the cleaned original data to form an enhanced sample set.

7. A strongly coupled process prediction method based on large model fine-tuning according to claim 1, 3, or 6, characterized in that, During incremental fine-tuning, a coupling constraint regularization term is added to the weighted loss function as a conditional penalty term. When the predicted parameter exceeds the process threshold, the loss value is increased; otherwise, the loss value is decreased. The fine-tuning loss function allocates loss weights based on the process importance of each target prediction parameter. The sum of the weight coefficients of all parameters is 1. The mean square error is used to calculate the single-parameter prediction loss, and the total loss is obtained by weighted summation.

Images