System and method for tracing non-optimal reasons in SHAP paper drying process

Abstract

The invention provides an SHAP paper drying process non-optimal reason tracing system and method. The method comprises the steps of data acquisition and preprocessing, energy efficiency level label and energy efficiency identification and non-optimal energy efficiency reason tracing. Data acquisition and preprocessing: identifying energy efficiency level labels and energy efficiency, and tracing reasons of non-optimal energy efficiency; energy efficiency level label and energy efficiency identification: ensuring the quality and consistency of model input data, and ensuring the reliability and interpretability of an analysis result through abnormal value removal, data smoothing and normalization processing; according to the energy efficiency level label and energy efficiency identification, the energy efficiency of equipment or a process is clearly identified and distinguished, so that targeted optimization measures are formulated; according to the non-optimal energy efficiency reason tracing method, key factors causing low energy efficiency are identified and analyzed, and specific problems affecting the energy efficiency are found out. The accuracy and interpretability of energy efficiency monitoring are improved, a scientific process optimization basis is provided for papermaking enterprises, and the production efficiency and the energy-saving effect are promoted to be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of paper operation status evaluation in an industrial production process, and in particular relates to a SHAP paper drying process non-optimal cause tracing method. Background Art

[0002] In the modern papermaking industry, the drying process is a key link, and its energy efficiency directly affects production costs and environmental benefits. However, the energy efficiency of the drying process is affected by many complex factors, and traditional methods are difficult to effectively identify and trace the reasons for low energy efficiency. Existing energy efficiency monitoring technology mainly relies on control systems based on "human experience + distributed computer control systems (DCS)" for regulation, but the dynamics and complexity of the drying process have surpassed the human brain's ability to analyze and optimize in a timely manner. It is often difficult to accurately locate the source of the problem, resulting in insufficient effectiveness and pertinence of improvement measures. Therefore, this control system is difficult to ensure the healthy and efficient operation of the paper drying process.

[0003] With the development of artificial intelligence and big data technologies, it has become possible to apply machine learning methods to optimize industrial processes. However, its "black box" characteristics limit the interpretability and scope of application of the model, and its ability to explain and trace the causes of low energy efficiency is limited. Summary of the invention

[0004] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method for tracing the causes of non-optimal drying of SHAP paper, thereby solving the above-mentioned technical problems in the prior art.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] A SHAP paper drying process non-optimal cause tracing system, including data collection and preprocessing, energy efficiency grade labeling and energy efficiency identification, and non-optimal energy efficiency cause tracing;

[0007] The data collection and preprocessing: energy efficiency grade labeling and energy efficiency identification, and tracing of reasons for non-optimal energy efficiency;

[0008] The energy efficiency rating label and energy efficiency identification: ensure the quality and consistency of the model input data by removing outliers, smoothing data and normalizing, and ensure the reliability and interpretability of the analysis results;

[0009] The energy efficiency rating label and energy efficiency identification mentioned above: clearly identify and distinguish the energy efficiency of equipment or processes, so as to formulate targeted optimization measures;

[0010] The method for tracing the causes of suboptimal energy efficiency is as follows: identifying and analyzing the key factors that lead to low energy efficiency, and finding out the specific problems that affect energy efficiency.

[0011] Furthermore, the data collection and preprocessing method:

[0012] S201, collecting the production line data of the drying section, and analyzing and selecting the process variables before preprocessing the original production data, and mining the process variables that have a strong correlation with the energy efficiency of the drying section;

[0013] S202, calculate the variance of the data, analyze the results of the variance characteristic calculation, obtain the degree of data dispersion, and eliminate variables with variance values ​​close to 0. The variance characteristic is expressed as:

[0014]

[0015] Among them, S 2 is the sample variance, X is the variable, is the sample mean, n is the number of samples;

[0016] S203, using the 3σ principle and box plot method to process outliers on the characteristic data and eliminate outliers;

[0017] S204, use a sliding data window to smooth the multivariate time series data, and then use a normalization operation to process the data to eliminate the impact of dimensional differences between different variables on the modeling process. The sliding data window is expressed as:

[0018]

[0019] Among them, H is the value after sliding window processing, t is the time step, w is the window width, x t is the sample characteristic value;

[0020] S205. Use the K-Means clustering algorithm to perform working condition clustering analysis. In the K-Means algorithm, the K value is a hyperparameter that directly determines the number of clusters in the data set. The SSE and average silhouette coefficient are used jointly to determine the K value.

[0021] Furthermore, in S205, the objective function of the K-Means algorithm and the K value determination method are expressed as:

[0022]

[0023] Among them: number of samples N, number of regional clusters K, sample value x n , cluster center u k and the indicator matrix r nk ;

[0024]

[0025] Where: SSE is the sum variance value, X is the jth class K j The data in Yj is the average value of all data in the jth class;

[0026]

[0027] Where: S(j) is the average silhouette coefficient value, p(j) is the average value of the dissimilarity between the j vector and other points in the same cluster, and q(j) is the minimum value of the average dissimilarity between the j vector and other clusters.

[0028] Furthermore, the energy efficiency grade label and energy efficiency identification method comprises the following steps:

[0029] S301, introduce the MLP neural network to extract data features and learn the nonlinear relationship of data, and set the structure of the input layer, hidden layer and output layer, initialize the weight and bias parameters of each layer, obtain the intermediate feature representation from the hidden layer output, and perform nonlinear transformation on the hidden layer features through the activation function; use the features processed by the activation function as the final feature output and participate in the subsequent modeling tasks;

[0030] S302, using the features extracted by MLP as new features of the training set and inputting them into the model training task; using the T-PLS method to integrate and extract multiple data modes, discover the correlation between the modes, and build a prediction model;

[0031] S303, obtaining the relationship between the papermaking output and the steam flow of the paper machine, using the Taylor expansion method to draw an energy efficiency coordinate diagram, analyzing the energy efficiency status of the drying section, and labeling the production data with energy efficiency grades of "excellent", "good", "medium", and "poor";

[0032] S304, based on the "excellent", "good", "medium" and "poor" energy efficiency grade labels of the production data, an adaptive boosting algorithm is used to iteratively train weak classifiers and combine them into strong classifiers to build an energy efficiency grade recognition model;

[0033] At this point, the process of training the weak classifier is expressed as:

[0034]

[0035] Where: e t is the training data error rate, w t (i) is the weight value of sample i in the tth iteration, err(·) is the indicator function, h t (x i ) is the prediction result of the weak classifier in the tth iteration, y i is the true label of sample i.

[0036] Furthermore, in S301, the MLP optimization process is expressed as:

[0037]

[0038] in: is the prediction result of the model, f(·) is the activation function, W (L) is the weight matrix of the Lth layer, b (L) is the bias vector of the Lth layer, and L is the number of layers of the neural network.

[0039] Further, in S302, the T-PLS calculation process is expressed as:

[0040]

[0041] Where: t y Y c The directly related score vector, p y For X c The load vector, T o and P o is the score matrix and loading matrix of the null space component, T r and P r For X c The score matrix and loading matrix in the residual space, E r and F is X c and Y c The residual term, q y Y c The load vector.

[0042] Furthermore, in S303, the relationship between the paper machine production output and the steam flow rate is expressed as:

[0043]

[0044] in: and is the current papermaking output and benchmark papermaking output of the papermaking machine, and It is the current steam flow and benchmark steam flow of the drying section.

[0045] Furthermore, the method for tracing the causes of non-optimal energy efficiency is as follows:

[0046] Based on the Shapley value theory in game theory, the marginal benefits of each process variable in the sample data are calculated to determine the importance of the process variables.

[0047] The Shapley value can be expressed as:

[0048]

[0049] in: is the Shapley value, S is the feature subset used in the model, N is the number of features, i is the sample feature value vector to be explained, is the weight value, and f(S) is the model output value under the feature combination S.

[0050] The method of the SHAP paper drying process non-optimal cause tracing system is characterized by comprising the following steps:

[0051] S1. Collect key data in the paper drying process, pre-process the collected data, use variance calculation and box plot method to deal with outliers, and use sliding data window to smooth the data to eliminate the impact of fluctuations; normalize the data to unify the dimension to ensure the consistency and reliability of the data;

[0052] S2. Input the preprocessed data into the energy efficiency evaluation model, use the Taylor expansion method to draw the energy efficiency coordinate diagram, and divide the energy efficiency levels into four categories: excellent, good, medium, and poor; use the adaptive boosting method to build an energy efficiency identification model, and accurately identify the energy efficiency level of each data point through training and testing.

[0053] S3. Conduct in-depth analysis on the data points marked as "poor" in the energy efficiency identification results, use the SHAP method to calculate the influence of each feature on energy efficiency, identify the main factors affecting energy efficiency through the SHAP value, and provide specific improvement suggestions.

[0054] Furthermore, in S1, the key data include temperature, humidity, drying time and paper thickness process parameters.

[0055] Beneficial effects of the present invention:

[0056] 1. The present invention applies SHAP to energy efficiency monitoring of the paper drying process, which can effectively identify the key factors affecting energy efficiency, explain and trace the reasons for low energy efficiency, and accurately evaluate the contribution of each factor to the model prediction results and its positive and negative impact.

[0057] 2. The present invention analyzes the energy efficiency data of the drying process, identifies and explains the main factors affecting energy efficiency, and accurately traces the reasons for low energy efficiency. It not only improves the accuracy and explainability of energy efficiency monitoring, but also provides a scientific basis for process optimization for papermaking enterprises, thereby promoting the improvement of production efficiency and energy-saving effects.

[0058] 3. The SHAP method used in the present invention not only performs well in ranking the importance of features, but also can provide local interpretation for a single sample, thereby fully understanding the overall statistical analysis and the impact of each sample point on the prediction model, making the model prediction results more transparent and understandable. The interaction analysis between features makes the contribution assessment more accurate, thus providing a powerful tool guarantee in the evaluation of the operating status of complex industrial processes and the tracing of non-optimal causes. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 is a flow chart of a tracing method according to an embodiment of the present invention;

[0061] Figure 2 is a diagram of the training process of the adaptive boosting algorithm according to an embodiment of the present invention;

[0062] Figure 3 This is a result analysis diagram of tracing the causes of poor energy efficiency according to an embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of long-time domain statistical analysis of the second type of working condition in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] like Figure 1-Figure 2 As shown, the embodiment of the present invention provides a non-optimal cause tracing system for paper drying process based on SHAP, including:

[0066] Data collection and preprocessing, energy efficiency grade labeling and energy efficiency identification, and tracing of causes of suboptimal energy efficiency.

[0067] Data collection and preprocessing ensure the quality and consistency of model input data. By removing outliers, smoothing data and normalizing, the impact of noise and errors can be reduced, thereby improving the accuracy and stability of the model. Accurate preprocessing steps can lay a solid foundation for subsequent data analysis and model building, ensuring the reliability and interpretability of analysis results.

[0068] At this time, the data collection and preprocessing methods are as follows:

[0069] S201. Collect the data of the drying section production line. Before preprocessing the original production data, it is necessary to analyze and select the process variables to find out the process variables that have a strong correlation with the energy efficiency of the drying section.

[0070] S202, calculate the variance of the data, analyze the results of the variance characteristic calculation, and obtain the degree of data dispersion. Eliminate variables whose variance values ​​are approximately 0. The variance characteristic can be expressed as:

[0071]

[0072] Where S 2 is the sample variance, X is the variable, is the sample mean, and n is the number of samples.

[0073] S203. In actual industrial operation, there are problems such as poor connection between speed sensor and machinery, large load change of mechanical equipment or paper breakage inside the drying cylinder, so there are abnormal values ​​in the characteristic data. The 3σ principle and box plot are used to deal with abnormal values ​​of characteristic data to eliminate the impact of abnormal values ​​on subsequent modeling.

[0074] S204. Considering that the environmental measurement noise is mixed in the process of sensor equipment reading and storing production data during the production process, the collected production data is subject to certain disturbances. Use the sliding data window to smooth the multivariate time series data, so that the impact of the unstable characteristics of the original production data on the subsequent modeling can be minimized, and then the normalization operation is used to process the data to eliminate the impact of the dimensional differences between different variables on the modeling process. The sliding data window can be expressed as:

[0075]

[0076] Where H is the value after sliding window processing, t is the time step, w is the window width, x t is the sample feature value.

[0077] S205. Papermaking is a typical process industry that uses human operating experience and DCS for production management and control. The operating conditions and boundary conditions in the production process change cumbersomely. Efficient and accurate capture of this feature will greatly improve the prediction accuracy of the model. In addition, the requirements for setting feature variables are different for the production of different paper products. The prediction performance of the model may be affected if the data under different production conditions are put together to train the model, so the K-Means clustering algorithm is used for operating condition clustering analysis. In the K-Means algorithm, the K value is a hyperparameter that directly determines the number of clusters in the data set. The SSE and the average silhouette coefficient are used jointly to determine the K value. The objective function of the K-Means algorithm and the K value determination method can be expressed as:

[0078]

[0079] The number of samples is N, the number of regional clusters is K, and the sample value is x n , cluster center u k and the indicator matrix rnk .

[0080]

[0081] Where SSE is the sum variance value, X is the jth class K j The data in Y j is the average value of all data in the jth class.

[0082]

[0083] Where S(j) is the average silhouette coefficient value, p(j) is the average value of the dissimilarity between the j vector and other points in the same cluster, and q(j) is the minimum value of the average dissimilarity between the j vector and other clusters.

[0084] Energy efficiency rating labels and energy efficiency identification are aimed at classifying and evaluating different energy efficiency levels, so that enterprises can clearly identify and distinguish the energy efficiency of equipment or processes, and thus formulate targeted optimization measures. The role of energy efficiency identification is to accurately identify and predict energy efficiency levels through model analysis, help enterprises to timely discover and improve low-efficiency links, and improve overall energy efficiency management efficiency.

[0085] At this time, the energy efficiency rating label and energy efficiency identification methods include:

[0086] S301, introduce MLP neural network to extract data features and learn data nonlinear relationships. Set the structure of input layer, hidden layer and output layer, initialize the weight and bias parameters of each layer. Obtain intermediate feature representation from the hidden layer output, and perform nonlinear transformation on the hidden layer features through activation function. Output the features processed by activation function as the final features and participate in subsequent modeling tasks. The MLP optimization process can be expressed as:

[0087]

[0088] in is the prediction result of the model, f(·) is the activation function, W (L) is the weight matrix of the Lth layer, b (L) is the bias vector of the Lth layer, and L is the number of layers of the neural network.

[0089] S302, use the features extracted by MLP as new features of the training set and input them into the model training task. Use the T-PLS method to integrate and extract multiple data modes, find the correlation between the modes, and build a prediction model. The T-PLS calculation process can be expressed as:

[0090]

[0091] where t y Y cThe directly related score vector, p y For X c The load vector, T o and P o is the score matrix and loading matrix of the null space component, T r and P r For X c The score matrix and loading matrix in the residual space, E r and F is X c and Y c The residual term, q y Y c The load vector.

[0092] S303, energy efficiency label data is to mark the production data of the drying section into different levels of energy efficiency status, each level represents an energy efficiency category. The relationship between the paper machine's papermaking output and steam flow is obtained, and the energy efficiency coordinate diagram is drawn using the Taylor expansion method to analyze the operating energy efficiency status of the drying section, and the production data is labeled with "excellent", "good", "medium", and "poor" energy efficiency levels. The relationship between the paper machine's papermaking output and steam flow can be expressed as:

[0093]

[0094] in and is the current papermaking output and benchmark papermaking output of the papermaking machine, and It is the current steam flow and benchmark steam flow of the drying section.

[0095] S304, based on the "excellent", "good", "medium", and "poor" energy efficiency rating labels of the production data, the weak classifiers are iteratively trained using an adaptive boosting algorithm and combined into a strong classifier to construct an energy efficiency rating identification model. The process of training the weak classifier can be expressed as:

[0096]

[0097] where e t is the training data error rate, w t (i) is the weight value of sample i in the tth iteration, err(·) is the indicator function, h t (x i ) is the prediction result of the weak classifier in the tth iteration, y i is the true label of sample i.

[0098] Tracing the causes of suboptimal energy efficiency, identifying and analyzing the key factors that lead to low energy efficiency, and helping enterprises find specific problems that affect energy efficiency. By tracing the causes, enterprises can optimize and improve the production process in a targeted manner, and improve the overall energy efficiency level and production efficiency.

[0099] Based on the Shapley value theory in game theory, the importance of process variables is determined by calculating the marginal benefits brought by each process variable in the sample data. The Shapley value can be expressed as:

[0100]

[0101] in is the Shapley value, S is the feature subset used in the model, N is the number of features, i is the sample feature value vector to be explained, is the weight value, and f(S) is the model output value under the feature combination S.

[0102] In addition, in this application, a method based on the SHAP paper drying process non-optimal cause tracing system includes:

[0103] S1, collect key data in the paper drying process, including process parameters such as temperature, humidity, drying time and paper thickness. Preprocess the collected data, use variance calculation and box plot method to deal with outliers, and use sliding data window to smooth the data to eliminate the impact of fluctuations. Normalize the data to unify the dimension, ensure the consistency and reliability of the data, and lay the foundation for subsequent analysis.

[0104] S2, input the pre-processed data into the energy efficiency evaluation model, use the Taylor expansion method to draw the energy efficiency coordinate diagram, and classify the energy efficiency levels into four categories: excellent, good, medium, and poor. Use the adaptive improvement method to build an energy efficiency identification model, and accurately identify the energy efficiency level of each data point through training and testing. Energy efficiency level labels help companies clarify the energy efficiency of equipment or processes and formulate targeted optimization measures.

[0105] S3 conducts in-depth analysis of the data points marked as "poor" in the energy efficiency identification results, and uses the SHAP method to calculate the impact of each feature on energy efficiency. The main factors affecting energy efficiency are identified through SHAP values, and specific improvement suggestions are provided. Tracing the causes of suboptimal energy efficiency helps companies find the root causes of low energy efficiency, optimize the production process in a targeted manner, and improve the overall energy efficiency level and production efficiency.

[0106] Example 1: Energy efficiency monitoring and optimization of paper drying process in a paper mill

[0107] During the paper drying process in the paper mill, process parameters such as temperature, humidity, drying time and paper thickness are collected. The collected data are preprocessed, including outlier processing, data smoothing and normalization. Data clustering analysis is performed using the K-Means clustering algorithm, and an energy efficiency prediction model is constructed using the MLP optimized T-PLS method. The energy efficiency coordinate diagram is drawn using the Taylor expansion method and divided into four categories: excellent, good, medium and poor. The ADABOOST method is used to construct an energy efficiency identification model, and the SHAP method is used to trace the causes of poor energy efficiency, identify the main influencing factors, and put forward improvement suggestions. This method can help paper mills fully understand the energy efficiency of the drying department, accurately identify and optimize inefficient links, and improve production efficiency and energy saving effects.

[0108] like Figure 3 As shown in the figure, after model identification, the drying system is in the "poor" level of energy efficiency, and the steam consumption per ton of paper in the drying section is 1.556 tons of steam / ton of paper. Process variables such as X3, X17, X10, X9, and X15 all contribute to the result judgment. X3 is the pressure of steam group 4, X17 is the exhaust temperature of T2, X10 is the exhaust flow rate of T2 2, X9 is the exhaust flow rate of T2 1, and X15 is the exhaust humidity of T1. X3 has the greatest impact, which is a negative impact, followed by X17, which is also a negative impact. The remaining process variables have similar impacts and are all negative. Analyzing from the two main influencing factors of steam group 4 pressure and T2 exhaust temperature, combined with specific data, compared with the "excellent" level of energy efficiency, the pressure of steam group 4 is relatively small and the exhaust temperature of T2 is relatively large. From the perspective of the drying process mechanism, the pressure of steam group 4 is the steam pressure of the main steam section. The steam pressure of the main steam section is too low, which can easily lead to insufficient heat transfer driving force at the drying end, hindering the evaporation of water at the drying end and reducing the drying efficiency. At the same time, the pressure difference between the sections of the drying cylinder group is small, which can hinder the removal of condensed water in the drying cylinder and affect the heat transfer efficiency, thereby reducing the energy efficiency of the drying section. The exhaust temperature T2 is the exhaust temperature of the air supply and exhaust device in the middle section of the drying section. A high temperature means that the heat carried by the exhaust hood is also high, and the heat loss increases, resulting in a decrease in the energy efficiency of the drying section. In summary, the analysis results of the data-driven model are consistent with the mechanism analysis.

[0109] Example 2: Energy efficiency analysis in the process of special paper manufacturing

[0110] In the manufacturing process of specialty paper (such as waterproof paper, art paper), the method of the present invention is used to perform energy efficiency analysis and optimization. Through detailed energy efficiency level evaluation and cause tracing, the key energy efficiency influencing factors in the drying process of specialty paper are identified. Combined with actual production needs, process parameters are optimized to improve energy efficiency levels. This method is particularly suitable for the production of specialty paper with high energy efficiency requirements, and can effectively improve product quality and production efficiency.

[0111] like Figure 4As shown in the figure, at this time, X17, X11, X5, X6 and X15 are the first five important process variables. X17 is the exhaust temperature of T2, X11 is the intake flow rate of T3, X5 is the intake flow rate of T1, X6 is the exhaust flow rate of T1, and X15 is the exhaust humidity of T1. According to the size of the sample characteristic values ​​in the figure and the degree of contribution to the energy efficiency state, it can be seen that under the second type of operating conditions, when the characteristic values ​​of the two process variables X17 and X14 are small, they make a negative contribution to the production process. Before production, the characteristic values ​​of X17 and X14 can be positively adjusted. When the characteristic values ​​of the two process variables X11 and X6 are large, they make a negative contribution to the production process. Before production, the characteristic values ​​of X11 and X6 can be negatively adjusted. The characteristic value of the process variable X5 has a negative contribution when it is large or small, and no positive or negative adjustment conclusion can be obtained. It is necessary to conduct a specific analysis in the production process to give reasonable adjustment suggestions.

[0112] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A SHAP paper drying process non-optimal cause tracing system, characterized by: Including data collection and preprocessing, energy efficiency grade labeling and energy efficiency identification, and tracing of causes of non-optimal energy efficiency; The data collection and preprocessing: energy efficiency grade labeling and energy efficiency identification, and tracing of reasons for non-optimal energy efficiency; The energy efficiency rating label and energy efficiency identification: ensure the quality and consistency of the model input data by removing outliers, smoothing data and normalizing, and ensure the reliability and interpretability of the analysis results; The energy efficiency rating label and energy efficiency identification mentioned above: clearly identify and distinguish the energy efficiency of equipment or processes, so as to formulate targeted optimization measures; The method for tracing the causes of suboptimal energy efficiency is as follows: identifying and analyzing the key factors that lead to low energy efficiency, and finding out the specific problems that affect energy efficiency.

2. The SHAP paper drying process non-optimal cause tracing system according to claim 1 is characterized in that: The data collection and preprocessing method: S201, collecting the production line data of the drying section, and analyzing and selecting the process variables before preprocessing the original production data, and mining the process variables that have a strong correlation with the energy efficiency of the drying section; S202, calculate the variance of the data, analyze the results of the variance characteristic calculation, obtain the degree of data dispersion, and eliminate variables with variance values ​​close to 0. The variance characteristic is expressed as: Among them, S 2 is the sample variance, X is the variable, is the sample mean, n is the number of samples; S203, using the 3σ principle and box plot method to process outliers on the characteristic data and eliminate outliers; S204, use a sliding data window to smooth the multivariate time series data, and then use a normalization operation to process the data to eliminate the impact of dimensional differences between different variables on the modeling process. The sliding data window is expressed as: Among them, H is the value after sliding window processing, t is the time step, w is the window width, x t is the sample characteristic value; S205. Use the K-Means clustering algorithm to perform working condition clustering analysis. In the K-Means algorithm, the K value is a hyperparameter that directly determines the number of clusters in the data set. The SSE and average silhouette coefficient are used jointly to determine the K value.

3. The SHAP paper drying process non-optimal cause tracing system according to claim 2 is characterized in that: In S205, the objective function of the K-Means algorithm and the K value determination method are expressed as follows: Among them: number of samples N, number of regional clusters K, sample value x n , cluster center u k and the indicator matrix r nk ; Where: SSE is the sum variance value, X is the jth class K j The data in Y j is the average value of all data in the jth class; Where: S(j) is the average silhouette coefficient value, p(j) is the average value of the dissimilarity between the j vector and other points in the same cluster, and q(j) is the minimum value of the average dissimilarity between the j vector and other clusters.

4. The SHAP paper drying process non-optimal cause tracing system according to claim 1 is characterized in that: The energy efficiency grade label and energy efficiency identification method comprises the following steps: S301, introduce the MLP neural network to extract data features and learn the nonlinear relationship of data, and set the structure of the input layer, hidden layer and output layer, initialize the weight and bias parameters of each layer, obtain the intermediate feature representation from the hidden layer output, and perform nonlinear transformation on the hidden layer features through the activation function; use the features processed by the activation function as the final feature output and participate in the subsequent modeling tasks; S302, using the features extracted by MLP as new features of the training set and inputting them into the model training task; using the T-PLS method to integrate and extract multiple data modes, discover the correlation between the modes, and build a prediction model; S303, obtaining the relationship between the papermaking output and the steam flow rate of the paper machine, using the Taylor expansion method to draw an energy efficiency coordinate diagram, analyzing the energy efficiency status of the drying section, and labeling the production data with energy efficiency grades of "excellent", "good", "medium", and "poor"; S304, based on the "excellent", "good", "medium", and "poor" energy efficiency grade labels of the production data, an adaptive boosting algorithm is used to iteratively train weak classifiers and combine them into strong classifiers to build an energy efficiency grade recognition model; At this point, the process of training the weak classifier is expressed as: Where: e t is the training data error rate, w t (i) is the weight value of sample i in the tth iteration, err(·) is the indicator function, h t (x i ) is the prediction result of the weak classifier in the tth iteration, y i is the true label of sample i.

5. The SHAP paper drying process non-optimal cause tracing system according to claim 4 is characterized in that: In S301, the MLP optimization process is expressed as: in: is the prediction result of the model, f(·) is the activation function, W (L) is the weight matrix of the Lth layer, b (L) is the bias vector of the Lth layer, and L is the number of layers of the neural network.

6. The SHAP paper drying process non-optimal cause tracing system according to claim 4 is characterized in that: In S302, the T-PLS calculation process is expressed as: Where: t y Y c Directly related score vector, p y For X c The load vector, T o and P o is the score matrix and loading matrix of the null space component, T r and P r For X c The score matrix and loading matrix in the residual space, E r and F is X c and Y c The residual term, q y Y c The load vector.

7. The SHAP paper drying process non-optimal cause tracing system according to claim 4 is characterized in that: In S303, the relationship between the paper machine production output and the steam flow rate is expressed as: in: and is the current papermaking output and benchmark papermaking output of the papermaking machine, and It is the current steam flow and benchmark steam flow of the drying section.

8. The SHAP paper drying process non-optimal cause tracing system according to claim 1 is characterized in that: The method for tracing the causes of non-optimal energy efficiency is as follows: Based on the Shapley value theory in game theory, the marginal benefits of each process variable in the sample data are calculated to determine the importance of the process variables. The Shapley value can be expressed as: in: is the Shapley value, S is the feature subset used in the model, N is the number of features, i is the sample feature value vector to be explained, is the weight value, and f(S) is the model output value under the feature combination S.

9. The method of the SHAP paper drying process non-optimal cause tracing system according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. Collect key data in the paper drying process, pre-process the collected data, use variance calculation and box plot method to deal with outliers, and use sliding data window to smooth the data to eliminate the impact of fluctuations; normalize the data to unify the dimension to ensure the consistency and reliability of the data; S2. Input the preprocessed data into the energy efficiency evaluation model, use the Taylor expansion method to draw the energy efficiency coordinate diagram, and divide the energy efficiency levels into four categories: excellent, good, medium, and poor; use the adaptive boosting method to build an energy efficiency identification model, and accurately identify the energy efficiency level of each data point through training and testing. S3. Conduct in-depth analysis on the data points marked as "poor" in the energy efficiency identification results, use the SHAP method to calculate the impact of each feature on energy efficiency, identify the main factors affecting energy efficiency through the SHAP value, and provide specific improvement suggestions.

10. The method of the SHAP paper drying process non-optimal cause tracing system according to claim 9, characterized in that: In S1, key data include temperature, humidity, drying time and paper thickness process parameters.

Images