A method for screening industrial process datasets based on data attribution

Abstract

In an embodiment of the present invention, a method for screening industrial process datasets based on data attribution is provided, belonging to the field of data processing technology. The method specifically includes the following steps: Step 1, obtaining industrial process data within a preset time period; Step 2, constructing a deep network regression model based on the industrial process data; Step 3, defining a calculation formula for the linear leave-one-out effect; Step 4, locally linearizing the deep network regression model to obtain a linear model; Step 5, reducing the dimensionality of the linear model and obtaining a nonlinear leave-one-out effect based on the calculation formula for the linear leave-one-out effect; Step 6, considering the interaction between samples, calculating the marginal effect of each sample by sampling different data subsets; Step 7, defining sample utility based on the marginal effect, and screening the industrial process data based on the sample utility. The scheme of the present invention improves the efficiency and accuracy of data screening.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and more particularly to a method for screening industrial process data sets based on data attribution. Background Art

[0002] Currently, data attribution methods aim to address a fundamental question: which data significantly influence the performance of data-driven models? By addressing this question, optimal training subset selection can be achieved, thereby improving model performance. These methods are derived from the leave-one-out influence calculation framework. The leave-one-out influence calculation framework evaluates the impact of individual data points on the training of machine learning models through a systematic process. Its core principle is to quantify the impact of each example by iteratively removing individual examples from the dataset, retraining the model, and comparing the prediction results before and after removal. Although this concept was proposed as early as 1983, its practical implementation in complex deep networks has been constrained by two key limitations: the need for iterative retraining and the large amount of randomness inherent in deep neural network training.

[0003] It can be seen that there is an urgent need for an industrial process data set screening method based on data attribution that can improve data screening efficiency and accuracy. Summary of the Invention

[0004] In view of this, an embodiment of the present invention provides an industrial process data set screening method based on data attribution, which at least partially solves the problem of poor data screening efficiency and accuracy in the prior art.

[0005] In a first aspect, an embodiment of the present invention provides a method for screening industrial process data sets based on data attribution, comprising:

[0006] Step 1: Obtain industrial process data within a preset time period;

[0007] Step 2: Build a deep network regression model based on industrial process data;

[0008] Step 3, define the calculation formula of the linear leave-one-out effect;

[0009] Step 4: Locally linearize the deep network regression model to obtain a linear model;

[0010] Step 5: Reduce the dimensionality of the linear model and obtain the nonlinear leave-one-out effect based on the calculation formula of the linear leave-one-out effect;

[0011] Step 6: Consider the interaction between samples and calculate the marginal impact of each sample by sampling different data subsets;

[0012] In step 7, sample utility is defined based on marginal impact, and industrial process data is screened based on the sample utility.

[0013] According to a specific implementation of the embodiment of the present invention, step 4 specifically includes:

[0014] Step 4.1: Perform a first-order Taylor expansion near the optimal parameters of the deep network regression model.

[0015] ;

[0016] in, is a deep network regression model, is the optimal network parameter, For the input industrial process data, Represents the deep network regression model in The approximate value of the Taylor expansion at Represents the deep network regression model in The gradient at Represents the parameters of the deep network regression model;

[0017] Step 4.2, define the input of the linear model as

[0018] ;

[0019] in, represents the number of parameters of the deep network regression model, represents the field of real numbers;

[0020] In step 4.3, the nonlinear mapping of the industrial process data input is transformed into the input of the linear model.

[0021] According to a specific implementation of the embodiment of the present invention, step 5 specifically includes:

[0022] Step 5.1, multiply the input of the linear model by the random projection matrix to reduce the input dimension of the linear model to obtain the input variable after dimension reduction, wherein the expression for reducing the input dimension of the linear model is

[0023] ;

[0024] in, Represents a random projection matrix for dimensionality reduction, the dimension of the random projection matrix satisfies ;

[0025] The expression of the input variable after dimension reduction is

[0026] ;

[0027] Step 5.2: Substitute the reduced-dimensional linear model into the linear leave-one-out effect calculation formula to obtain the nonlinear leave-one-out effect of the nonlinear model on a single sample.

[0028] ;

[0029] in, Represented by vector The matrix formed, For samples The corresponding label.

[0030] According to a specific implementation of the embodiment of the present invention, step 6 specifically includes:

[0031] Step 6.1: Define the marginal effects that take into account interactions between samples

[0032] ;

[0033] in, Indicates that the sampling process follows a uniform distribution, is a training sample subset obtained by sampling industrial process data according to uniform distribution, Represents the leave-one-out effect of deep networks trained on a subset of training samples;

[0034] Step 6.2: randomly select a subset of training samples according to uniform distribution to train a nonlinear deep network regression model;

[0035] Step 6.3, calculate the nonlinear leave-one-out effect of the nonlinear model on a single sample in the training sample subset;

[0036] In step 6.4, repeat steps 6.2 to 6.3 to estimate the marginal effects based on the truncated Monte Carlo sampling method.

[0037] According to a specific implementation of the embodiment of the present invention, step 7 specifically includes:

[0038] Step 7.1: Define sample utility based on the marginal impact of the sample

[0039] ;

[0040] Step 7.2, based on sample utility in the validation set Comprehensive evaluation of samples

[0041] ;

[0042] Step 7.3, based on the sample utility vector From industrial process data A preset proportion of low-utility samples is removed to obtain high-quality training samples.

[0043] The industrial process data set screening scheme based on data attribution in an embodiment of the present invention includes: step 1, obtaining industrial process data within a preset time period; step 2, constructing a deep network regression model based on the industrial process data; step 3, defining a calculation formula for the linear leave-one-out effect; step 4, locally linearizing the deep network regression model to obtain a linear model; step 5, reducing the dimension of the linear model, and obtaining the nonlinear leave-one-out effect based on the calculation formula for the linear leave-one-out effect; step 6, considering the interaction between samples, calculating the marginal effect of each sample by sampling different data subsets; step 7, defining sample utility based on the marginal effect, and screening the industrial process data based on the sample utility.

[0044] The beneficial effects of the embodiments of the present invention are as follows: through the scheme of the present invention, a data attribution method based on deep network linearization is proposed, which effectively transfers the analytical calculation advantages of the leave-one-out effect in the linear regression model to the deep network, thereby significantly improving the computational efficiency; the present invention uses the truncated Monte Carlo method to estimate the marginal influence of the sample, which solves the limitation of the leave-one-out influence estimation that ignores the synergistic effect between samples; the present invention defines the sample utility based on the obtained marginal influence, quantifies the overall influence of each training sample on the model performance, and improves the data screening efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A schematic diagram of a process flow of an industrial process data set screening method based on data attribution provided by an embodiment of the present invention;

[0047] Figure 2 A comparison chart of the effects of different methods provided by the embodiments of the present invention on the validation set;

[0048] Figure 3 A comparison chart of the effects of different methods provided by the embodiments of the present invention on the test set. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present invention, those skilled in the art will appreciate that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0052] It should also be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. The illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0053] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0054] An embodiment of the present invention provides an industrial process data set screening method based on data attribution, which can be applied to the data screening process in industrial production scenarios.

[0055] See also Figure 1 , is a flow chart of a method for screening industrial process data sets based on data attribution provided by an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps:

[0056] Step 1: Obtain industrial process data within a preset time period;

[0057] For implementation, 3,728 process data records were collected from the Salt Lake Chemical Plant between 2022 and 2023. Each record contained 72 process variables and one quality indicator to be predicted. The dataset was partitioned according to a strict 70-15-15 split: the first 70% was used for model training, the next 15% for validation, and the remaining 15% for final testing.

[0058] Step 2: Build a deep network regression model based on industrial process data;

[0059] In specific implementation, in this embodiment of the present invention, three models with different architectures were constructed: an MLP (Multi-layer Perceptron) model, an RNN (Recurrent Neural Network) model, and a Transformer model. Each model had an input feature dimension of 72. The specific structural designs are as follows: The MLP (Multi-layer Perceptron) model consists of multiple fully connected layers with an input dimension of 72. After passing through several hidden layers (each with 128 neurons), it ultimately outputs a predicted value. To prevent overfitting, the ReLU activation function and Dropout layer were used. The RNN (Recurrent Neural Network) model used a standard single-layer RNN structure with an input dimension of 72 and a hidden dimension of 64. The tanh activation function was used. To capture temporal relationships in sequence data, a time step size of 64 was set, and the output layer contained a single regressor. The Transformer model was based on a self-attention mechanism with an input dimension of 72. It used a multi-head self-attention mechanism (8 heads), with each head having a dimension of 64. The Transformer model also included a positional encoding layer to process the sequential information of the input sequence. In addition, the model is composed of several stacked encoder layers, each with 4 attention heads, and features are processed through fully connected layers and ReLU activation functions. The output layer is a single predicted value.

[0060] Given a dataset Include samples, A nonlinear model (multilayer perceptron, recurrent neural network, convolutional neural network, etc.), denoted as By minimizing the mean square error of the prediction, the optimal model parameters can be obtained:

[0061] ;

[0062] Step 3, define the calculation formula of the linear leave-one-out effect;

[0063] In specific implementation, the linear regression model assumes that the dependent variable With the independent variable vector The relationship between is linear. The standard linear regression model can be defined as:

[0064] ;

[0065] in represents the model parameters, Represents the difference between the observed value and the predicted value. Usually, linear regression can be represented by a matrix as:

[0066] ;

[0067] in:

[0068] ;

[0069] The optimal parameters of linear regression can be obtained by the least squares method, and its optimization goal is to minimize the sum of the mean square errors on the training set, that is:

[0070] ;

[0071] To find the minimum value, Take the derivative and set it to zero:

[0072] ;

[0073] in .

[0074] Data attribution in linear regression can be expressed in analytical form, the key is to evaluate the removal of specific training samples The new regression coefficient is defined as:

[0075] ;

[0076] Among them, remove the sample The data matrix after , the label vector is denoted as . Remove The influence of a sample on the regression coefficient estimation can be expressed as:

[0077] ;

[0078] in, It can be expressed as Single-rank update of :

[0079] ;

[0080] It can be simplified based on the Sherman-Morrison formula, and its simplified expression is defined as:

[0081] ;

[0082] From this we can deduce:

[0083] ;

[0084] ;

[0085] Sample pair The leave-one-out effect on the prediction results of a sample can be directly obtained by the change of the regression coefficient, which is defined as:

[0086] ;

[0087] The above formula simplifies the effect of removing samples on the linear regression results into a simple analytical form. It can be seen that the large matrix multiplication and inverse operation in the formula are similar to It doesn't matter, it means It only needs to be calculated once and can be reused when calculating the impact of different samples. Therefore, the calculation of LOO impact of linear regression model is very efficient.

[0088] Step 4: Locally linearize the deep network regression model to obtain a linear model;

[0089] In the specific implementation, considering the effect of removing a single sample on the optimal model parameters The effect can be neglected and the nonlinear model can be obtained by Perform a first-order Taylor expansion to approximate:

[0090] ;

[0091] Taylor expansion will be the original The nonlinear mapping is transformed into and The input of the new linear model can be defined as in Indicates the number of model parameters. In this case, The leave-one-out effect of the linear model can be directly substituted to calculate the removed sample For samples The predicted impact, denoted as The linear model input dimension for the MLP model is 17666, the linear model input dimension for the RNN model is 8770, and the linear model input dimension for the Transformer model is 187969.

[0092] Step 5: Reduce the dimensionality of the linear model and obtain the nonlinear leave-one-out effect based on the calculation formula of the linear leave-one-out effect;

[0093] In the specific implementation, although the linearization of the model significantly simplifies the calculation of the leave-one-out effect, the dimension of the input variables of the linear model is essentially equivalent to the number of parameters in the nonlinear model. This results in the model complexity seriously affecting the computational efficiency when the leave-one-out effect formula of the linear model is used to calculate the leave-one-out effect. For even moderately complex models (usually with millions of parameters), calculating the leave-one-out effect becomes infeasible. In order to reduce the dimension of the input variables, the present invention converts the model gradient With random projection matrix Multiplication ensures the inner product Maintain a high degree of similarity with the original inner product. This high similarity of the inner product ensures that when using the mapped gradient to update the model in a deep network, the effect of using a smaller learning rate is basically the same as using the original gradient. The dimension of the random projection matrix satisfies , effectively reducing the dimension of the input variables. This process is defined as:

[0094] ;

[0095] The input variable after dimension reduction can be expressed as

[0096] .

[0097] The calculation of the leave-one-out effect of nonlinear models is feasible because the parameter dimension reduction makes it possible to calculate the leave-one-out effect using the linear model leave-one-out formula. The input matrix after dimensionality reduction is defined as Substituting this into the leave-one-out method formula for linear models, the leave-one-out effect for nonlinear models is defined as:

[0098] ;

[0099] in in Not Replace. This is because Represents the regression error, and its computational complexity is not high. If it is replaced by the mapped gradient, it will increase the estimation error of the leave-one-out effect. Since each sample needs to calculate the leave-one-out effect on the validation set sample, the final result of this effect is a 2609 559 matrix, where each row represents the leave-one-out effect of a training sample on each validation set sample.

[0100] Step 6: Consider the interaction between samples and calculate the marginal impact of each sample by sampling different data subsets;

[0101] In specific implementation, the sample influence estimation of sample interaction is considered. The influence obtained by the nonlinear leave-one-out influence estimation formula is essentially a quantitative estimate of the impact on the model when a sample is removed from the complete dataset. However, the leave-one-out influence ignores the interaction between samples. For example, if two samples are very similar, removing one sample may have little effect on the prediction, which will cause the LOO method to believe that the two samples have no effect on the prediction. In order to incorporate the interaction between samples into the calculation of marginal influence, the LOO influence from different training subsets is used. In this case, the marginal impact is defined as:

[0102] ;

[0103] in represents the uniform distribution of all data points, Indicates that from the subset The estimated LOO effect in the deep network model trained on . It is not feasible to directly calculate this expectation. Therefore, the present invention proposes to use the truncated Monte Carlo sampling method to estimate the sample The marginal impact The sampling probability of each sample is set to 0.8. If the estimated value obtained by this sampling iteration is and the estimated value obtained from the previous sampling iteration If the difference is less than 10% of itself, truncation is triggered, the Monte Carlo estimation is stopped, and the sample impact obtained is output. , which is formally expressed as

[0104] ;

[0105] in Represents the cutoff threshold, which is set to 0.1 here.

[0106] In step 7, sample utility is defined based on marginal impact, and industrial process data is screened based on the sample utility.

[0107] In the specific implementation, the dimension of the final influence matrix is ​​2609 559. Based on this matrix, it is assumed that removing a sample that reduces the prediction error should have a positive utility. The larger the absolute value of the utility, the more significant the impact of the sample on the model performance. Therefore, the sample utility is defined as:

[0108] ;

[0109] The above equation defines the Sample pair The utility of the sample predictions. The resulting sample utility matrix dimension is still 2609 559. In most cases, the utility of the sample needs to be over the entire validation set A comprehensive assessment is conducted on the following, which is defined as:

[0110] ;

[0111] The obtained sample utility becomes a one-dimensional vector, that is, each training sample corresponds to a value. From the training set Remove the specified ratio The low-utility samples. This filtering process is defined as:

[0112] ,

[0113] in, Represents the sample utility value The quantile threshold is defined as:

[0114] ;

[0115] The industrial process data set screening method based on data attribution provided in this embodiment effectively transfers the analytical computational advantages of the leave-one-out effect in the linear regression model to the deep network by proposing a data attribution method based on deep network linearization, thereby significantly improving computational efficiency; the present invention uses the truncated Monte Carlo method to estimate the marginal influence of samples, solving the limitation of the leave-one-out influence estimation that ignores the synergistic effect between samples; the present invention defines the sample utility based on the obtained marginal influence, quantifies the overall influence of each training sample on the model performance, and improves data screening efficiency and accuracy.

[0116] The method of the present invention will be further described below with reference to a specific embodiment. In order to rigorously evaluate the comparative performance of the method of the present invention, the present invention systematically introduces six benchmark methods: Full set, Random drop, DataShapley, UIDS, DataModel and IF-DS. Their specific details are as follows: (1) Full set directly uses the entire training dataset without any data screening or removal. (2) Random drop implements a random subsampling strategy to discard a predetermined percentage of training instances before model training. (3) Data Shapley applies the principles of cooperative game theory to identify key training samples through Shapley value calculation. The selection threshold is determined by the attribution score obtained based on repeated model retraining. (4) UIDS optimizes model performance through unweighted subsampling guided by influence functions and strategically selects training subsets by evaluating the influence of individual data points. (5) DataModel combines iterative neural network retraining and least squares regression to derive sample-specific utility scores for selective retention. (6) IF-DS systematically selects high-influence training data through influence function analysis, optimizes model performance and interpretability, strategically retains samples that contribute most to model accuracy, and removes redundant or harmful examples.

[0117] In order to verify the performance of the algorithm under different quantile thresholds, the set of quantile threshold selection is set to {0.01, 0.02, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}, which includes 12 different sample removal ratios ranging from 0.01% to 90%. Figure 2 , 3 respectively show the improvement effect of the present invention on the validation set and test set in actual industrial cases. The performance of the proposed method and the baseline method is shown using the Transformer architecture as an example under different data removal ratios. The black dotted line in the figure represents the verification performance using the complete dataset, while the cyan dotted line represents the verification performance of the corresponding UIDS (due to its fixed ratio of data removal). The bar charts of different colors show the verification performance of other algorithms. The gray bar chart serves as a baseline comparison, indicating the verification performance after randomly discarding a certain proportion of training samples.

[0118] Table 1 comprehensively compares the performance of six dataset selection strategies on an industrial validation set. The optimal MSE values ​​are recorded for different sample removal ratios, and the performance advantage over random removal is quantified using a fidelity score. The full dataset and random dropout methods serve as baselines for evaluating the accuracy of data attribution. Both methods incur no computational overhead, with random dropout establishing a fidelity baseline of Fidelity = 0. Empirical results show that random data removal generally fails to improve model performance. In contrast, all data attribution-based strategies significantly improve model accuracy. UIDS and IF-DS, two influence function (IF)-based methods, perform comparably in terms of computational efficiency because they both leverage IF results. However, because IF only approximates the leave-one-out (LOO) effect of nonlinear models and fails to account for multi-sample interactions, their accuracy remains limited. Data Shapley and DataModel, retraining-based attribution methods, achieve higher accuracy. DataModel provides the second-best results across multiple architectures, trailing only the present invention. However, their need to repeatedly train models results in high computational costs—often taking over 24 hours per execution on small datasets and models.

[0119] Our proposed method excels in three key areas: validation set improvement, attribution accuracy, and runtime efficiency. It achieves strong results on both datasets for most model architectures. Furthermore, our method is computationally efficient, taking up only 1% of the computation time of DataModel. This advantage stems from its high-precision Taylor expansion approximation within a local parameter neighborhood (where the parameter perturbation caused by removing a single example is minimal) and analytical LOO computation for linear models. The main approximation error comes from estimating the global sample utility across different subsets via LOO effects. This explains why our method significantly outperforms state-of-the-art methods in accuracy while maintaining computational efficiency.

[0120] Table 1

[0121]

[0122] As shown in the figure, DataModel and the present invention achieve comparable performance to using the full dataset when retaining only 10%-20% of the key samples. The proposed method significantly outperforms baseline and state-of-the-art methods at low data removal ratios, achieving the best validation performance on both datasets at 10% and 20% removal ratios, respectively. When the proportion of removed samples is small, almost all data attribution methods effectively identify anomalous or low-utility samples. However, as the proportion of retained training data decreases, the performance of Data Shapley and IF-DS decreases significantly, while DataModel and the present invention maintain relatively good performance. Notably, at low data retention ratios, all data attribution-based filtering methods perform worse than random removal. This may be due to the over-concentration of key samples identified by the attribution method within a specific distribution, resulting in insufficient data diversity. Therefore, data attribution methods are more suitable for filtering potential anomalous samples (low-utility samples) rather than selecting representative samples to improve training efficiency.

[0123] The model performance analysis on the validation set only evaluates the data attribution accuracy of the proposed method. The model performance on the test set demonstrates the effectiveness of our method in enhancing model generalization. We evaluated the best-performing model in the validation phase and validated it on the test set. The results are reported in Table 2. It is important to note that improved validation performance does not guarantee superior results on the test set.

[0124] Specifically, on the Salt Lake dataset, UIDS demonstrated comparable test performance to training using the full dataset, while achieving significantly better results on validation metrics. As shown in Table 2, most dataset filtering methods based on data attribution were able to effectively improve the model's test performance. The present invention demonstrated the most significant improvement, consistently achieving optimal or suboptimal results across multiple model architectures. These results demonstrate the effectiveness of the present invention in identifying potential anomalies in industrial process data.

[0125] Table 2

[0126]

[0127] Intuitively, Figure 3The test performance of the Transformer architecture at different data retention ratios is demonstrated. The chart shows that DataModel, IF-DS, and the present invention can still maintain performance close to that of the full dataset when only 20-30% of the key samples are retained. However, the cross-dataset analysis reveals a fundamental limitation: when 10% of the core samples are retained, the test error of most attribution methods is higher than the baseline. This highlights a key shortcoming of sample utility evaluation through data attribution - high utility samples may be concentrated in a specific distribution. Since sample utility is defined by validation performance, which is consistent with the characteristics of the validation set, the model's generalization ability to the test set is impaired when trained on a concentrated distribution. This phenomenon is manifested in a significant drop in performance when relying only on representative samples.

[0128] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware or a combination thereof.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for screening industrial process data sets based on data attribution, characterized in that: include: Step 1: Obtain industrial process data within a preset time period; Step 2: Build a deep network regression model based on industrial process data; Step 3, define the calculation formula of the linear leave-one-out effect; Step 4: Locally linearize the deep network regression model to obtain a linear model; The step 4 specifically includes: Step 4.1: Perform a first-order Taylor expansion near the optimal parameters of the deep network regression model. ； in, is a deep network regression model, is the optimal network parameter, For the input industrial process data, Represents the deep network regression model in The approximate value of the Taylor expansion at Represents the deep network regression model in The gradient at Represents the parameters of the deep network regression model; Step 4.2, define the input of the linear model as ； in, represents the number of parameters of the deep network regression model, represents the field of real numbers; Step 4.3, convert the nonlinear mapping of industrial process data input into the input of the linear model; Step 5: Reduce the dimensionality of the linear model and obtain the nonlinear leave-one-out effect based on the calculation formula of the linear leave-one-out effect; The step 5 specifically includes: Step 5.1, multiply the input of the linear model by the random projection matrix to reduce the input dimension of the linear model to obtain the input variable after dimension reduction, wherein the expression for reducing the input dimension of the linear model is ； in, Represents a random projection matrix for dimensionality reduction, the dimension of the random projection matrix satisfies ; The expression of the input variable after dimension reduction is ； Step 5.2: Substitute the reduced-dimensional linear model into the linear leave-one-out effect calculation formula to obtain the nonlinear leave-one-out effect of the nonlinear model on a single sample. ； in, Represented by vector The matrix formed, For samples Corresponding labels; Step 6: Consider the interaction between samples and calculate the marginal impact of each sample by sampling different data subsets; In step 7, sample utility is defined based on marginal impact, and industrial process data is screened based on the sample utility.

2. The method according to claim 1, characterized in that The step 6 specifically includes: Step 6.1: Define the marginal effects that take into account interactions between samples ； in, Indicates that the sampling process follows a uniform distribution, is a training sample subset obtained by sampling industrial process data according to uniform distribution, Represents the leave-one-out effect of deep networks trained on a subset of training samples; Step 6.2: randomly select a subset of training samples according to uniform distribution to train a nonlinear deep network regression model; Step 6.3, calculate the nonlinear leave-one-out effect of the nonlinear model on a single sample in the training sample subset; In step 6.4, repeat steps 6.2 to 6.3 to estimate the marginal effects based on the truncated Monte Carlo sampling method.

3. The method according to claim 2, characterized in that The step 7 specifically includes: Step 7.1: Define sample utility based on the marginal impact of the sample ； in Representation sample For the prediction sample The utility of Step 7.2, based on sample utility in the validation set Comprehensive evaluation of samples ； in Representation sample The comprehensive effect of Representation sample Forecast validation set samples The utility of Step 7.3, based on the sample utility vector From industrial process data A preset proportion of low-utility samples is removed to obtain high-quality training samples.

Images