Process industry soft measurement modeling method for cross-production-line transverse information fusion

Abstract

The invention provides a process industry soft measurement modeling method for cross-production-line transverse information fusion, and the method is based on an end-to-end neural network model which is formed by connecting a time sequence feature extraction network in a self-encoder form and a regression network in a front-and-back manner. A two-stage soft measurement model training strategy of pre-training-time sequence feature information fusion training is provided, and a performance index soft measurement model is safely, effectively and accurately established on a process industrial production line with insufficient data accumulation by introducing data of the production line with rich data accumulation. Therefore, soft measurement is carried out on some performance indexes which are difficult to directly measure or high in measurement cost in the process industry, and reference and guidance are provided for monitoring, optimizing and controlling various parameters in the production process of the process industry.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology and artificial intelligence, and in particular to a process industry soft measurement modeling method for cross-line horizontal information fusion. The method is mainly used to establish a soft measurement model for a series of important performance indicators in the production process on a process industry production line with insufficient data accumulation, and use the established soft measurement model to carry out real-time soft measurement. Background Art

[0002] Some terms frequently used in the present invention are introduced below.

[0003] (1) Process Industry: Process Industry refers to an industrial system that produces products for specific purposes by physically or chemically processing raw materials in a continuous manner, including steel, cement, chemicals, petroleum and other industries. Process industries are generally characterized by large-scale, continuous, and complex production principles, and therefore have high requirements for real-time monitoring, optimization and control of the production process.

[0004] (2) Process variables: Process variables refer to key parameters used to monitor and control production processes in process industries. These variables usually cover a variety of physical and chemical properties, such as temperature, pressure, flow, concentration, current, voltage, etc. The real-time data of these process variables can reflect the real-time status of the process industry production process.

[0005] (3) Performance Index: Performance Index refers to the key parameters used in process industries to evaluate and measure product quality, production efficiency and process stability. These indicators cover the physical and chemical properties of the product, as well as summary parameters of the overall production process, such as product conversion rate and production energy consumption. By monitoring and analyzing these performance indicators, the efficiency of the process industry production process and the quality of the output products can be ensured.

[0006] (4) Temporal Features: Temporal features refer to the attributes or characteristics used to describe the changes of time series data over time, including the average value, trend, dynamics, autocorrelation, etc. of time series data. Temporal features can be calculated from time series data using mathematical statistics methods, or can be adaptively extracted using machine learning methods.

[0007] (5) Soft Sensing: Soft sensing refers to a technical method in process industry that uses some process variables that are easy to measure or measure frequently to indirectly estimate performance indicators that are difficult to measure directly or frequently through mathematical models and algorithms. The mathematical model used for soft sensing modeling can be a mechanism model based on the mechanism of process industry, or a data-driven model based on data statistics and machine learning methods. Soft sensing integrates the data of multiple process variables and uses regression algorithms to establish a prediction model, thereby achieving real-time fitting and estimation of key performance indicators.

[0008] The following is an introduction to the background technology of the present invention.

[0009] In the field of process industry, due to the mutual coupling of physical and chemical reactions in the production process and the complex thermal state, there is often a complex nonlinear relationship between process variables and the performance indicators of the required soft measurement. Such a relationship is difficult to describe through precise expressions or mechanism models, so it is often necessary to use data-driven modeling methods, that is, to establish a performance indicator soft measurement model based on statistical learning or deep learning methods to obtain real-time estimation results of performance indicators. However, when the amount of data accumulated for soft measurement modeling on a production line is insufficient, the established soft measurement model will face problems such as underfitting of the training process, poor generalization ability, and unstable model parameter estimation, and its soft measurement performance will be seriously negatively affected. In this case, introducing new data sources to provide information support for soft measurement modeling is an important solution.

[0010] The homogeneous design of process industry production lines provides a potential solution to the problem of insufficient modeling data for a single production line. Taking my country's cement industry as an example, my country's cement plants are widely distributed, with nearly 1,800 production lines across the country, and the production line designs are all from several major research institutes, with a high degree of consistency in the process and production line structure. Therefore, when the amount of data for a process industry production line that requires soft measurement modeling is insufficient, the data of other production lines with large data accumulation can be introduced to participate in the soft measurement modeling process to improve the soft measurement performance of the established model. However, considering that the data of each production line usually contains some sensitive production process information, such as specific formulas and processes, the data of each production line cannot usually be directly shared with other production lines, which leads to the problem of "data islands", making it difficult to apply machine learning methods that rely on direct data sharing or data concentration of multiple production lines, such as ensemble learning, transfer learning, and traditional domain adaptation methods, under such conditions.

[0011] In response to the problem of data silos, technicians in related fields have developed a series of modeling methods that can cross data silos, such as federated learning algorithms and passive domain adaptation methods. The core of the federated learning algorithm is to train a model separately on each data set and then aggregate the parameters of all models. However, in the field of soft measurement, the parameters of the model can only vaguely describe the relationship between the input process variables and the performance indicators of the required soft measurement. Therefore, the federated learning algorithm can only show a certain effect when processing tasks with simple mechanisms, such as bearing life prediction. However, the process industry has changeable working conditions and complex mechanisms, so the federated learning method is not effective in modeling process industry soft measurements across data silos. Passive domain adaptation methods achieve modeling across data silos by transferring the assumptions of the source data domain and constructing corresponding pseudo labels. However, most of the existing passive domain adaptation methods are based on classification tasks, and many of the assumptions established by these methods cannot be effectively applied to soft measurement modeling problems where labels are continuously distributed in space.

[0012] At present, the technical solutions for cross-line complex process industry soft sensor modeling are still relatively limited. However, with the development of computer technology and the increasing application of data-driven modeling soft sensor technology in process industry scenarios, the demand for cross-line data-driven soft sensor modeling technology is growing. Cross-line information fusion soft sensor modeling based on time series features is a potential feasible path, because compared with the federated learning method, it can start from the feature level that is closer to the original data and contains richer information to achieve information fusion between multi-line data. In addition, since process variable data is usually standardized before entering the soft sensor model, it is difficult to combine the soft sensor model and the time series features extracted by the model to obtain the original production data through reverse engineering methods. Therefore, the sharing of soft sensor model parameters and time series features is relatively safe. Combined with the above technical background and analysis and demonstration, a cross-line horizontal information fusion process industry soft sensor modeling method based on time series features is a solution with strong data security and application prospects. Summary of the invention

[0013] In view of the above technical background, the present invention proposes a process industry soft sensor modeling method for cross-production line horizontal information fusion. The technical solution of the present invention is as follows:

[0014] A process industry soft sensor modeling method for cross-production line horizontal information fusion includes the following steps:

[0015] Step 1: Read offline data from the DCS database of the data source production line and the target production line respectively, match the process variable time series based on the tag value of the performance indicator, and build the source data set and the target data set;

[0016] Step 2: Use the source data set and the target data set to pre-train a source soft measurement model and a target soft measurement model respectively, and save the parameters of the two soft measurement models and the extracted source time series feature set and target time series feature set;

[0017] Step 3: Send the source time series feature set and the source soft measurement model parameters to the online platform to which the target production line belongs, aggregate the parameters of the source soft measurement model and the target soft measurement model, and construct a parameter aggregation soft measurement model;

[0018] Step 4: Through the time series feature information fusion loss function, the target data set, the source time series feature set and the target time series feature set are used to perform time series feature information fusion training on the parameter aggregation soft sensor model to obtain the information fusion soft sensor model;

[0019] Step 5: Deploy the information fusion soft measurement model on the online platform of the target production line to realize online and real-time soft measurement of the corresponding performance indicators, and upload the soft measurement results of the model to the DCS database for storage.

[0020] In step 1, the data source production line refers to a production line with sufficient data accumulation for providing information support for soft measurement modeling, and the target production line refers to a production line with insufficient data accumulation for which a soft measurement model needs to be established; the method for matching the process variable time series based on the performance indicator label value is: first, the process variables involved in the soft measurement modeling are selected from the DCS database, and a suitable time series length L is set, and then the timestamp of the soft measurement performance indicator label value acquisition in the DCS database is used as the starting point, and the process variable time series with a length of L is intercepted forward, so that each performance indicator label value within the data set range can correspond to the corresponding process variable time series; the source data set and the target data set are data sets constructed based on the data of the data source production line and the target production line, and the sequence length L of the time series of the two data sets and the category and number of process variables are consistent; the construction process of the two data sets in step 1 is completed on the online platforms corresponding to the two production lines respectively, and no interaction between the data sets is generated; the online platform refers to a system platform for deploying and running soft measurement models and their incremental learning programs, including cloud computing platforms, industrial Internet of Things platforms, and digital twin platforms.

[0021] In step 2, the soft measurement model is an end-to-end neural network model composed of a time series feature extraction network in the form of an autoencoder and a regression network connected front and back, the source soft measurement model and the target soft measurement model are trained based on the source data set and the target data set respectively, and the source time series feature set and the target time series feature set are extracted from the source data set and the target data set by the time series feature extraction networks of the source soft measurement model and the target soft measurement model respectively; the pre-training of the two models in step 2 and the time series feature set extraction process are completed on the online platforms corresponding to the two production lines respectively, and the structures of the two models are consistent. For each data set, the data set needs to be divided into a training set and a verification set during the training process, the soft measurement model performs parameter training on the training set, and the model soft measurement performance is verified on the verification set; the expression of the loss function used in the pre-training process is:

[0022]

[0023] Among them, λ1 and λ2 are the regularization coefficients of the pre-training loss function, is the supervision loss function of the soft sensor model, and its expression is:

[0024]

[0025] Among them, B is the number of training batches, Y b is the true value of the performance indicator label in the bth batch, is the prediction result of the model for the performance index in the bth batch. is the unsupervised loss function of the soft sensor model, and its expression is:

[0026]

[0027] Where N is the variable dimension of the process variable time series, is the value of the input process variable time series at the position (b, l, n), The value of the process variable time series at the position (b, l, n) reconstructed by the time series feature extraction network.

[0028] In step 3, the method for obtaining the parameters of the parameter aggregation soft sensor model is expressed as:

[0029] M agg =w1M source +w2m target

[0030] Among them, m source and M targetare the parameters of the source soft sensor model and the target soft sensor model respectively, w1 and w2 are the aggregation weights of the two model parameters, satisfying the condition of w1+w2=1. Generally, w1=w2. The model parameter aggregation operation in step 3, as well as the subsequent steps 4 and 5 are all completed on the online platform of the target production line.

[0031] In step 4, the expression of the temporal feature information fusion loss function is:

[0032]

[0033] Among them, λ′1, λ′2, λ3, and λ4 are the regularization coefficients of the time series feature information fusion training loss function, and Z source and Z target are the source time series feature set and the target time series feature set respectively, Z is the information fusion time series feature set output by the soft sensor model during the training process, and D(Z,Z source ) and D(Z,Z target ) are Z and Z respectively source , Z target The distribution difference measurement results between two feature sets are expressed as:

[0034]

[0035] Among them, MMD(Z b ,Z source ) and MMD(Z b ,Z target ) are Z and Z of the bth batch respectively. source and Z target The maximum multi-core mean difference between , is expressed as:

[0036]

[0037]

[0038] Among them, z A , z′ A and z B , z′ B Represents Z A and Z B Two randomly selected samples in , p represents the number of kernels, E represents the mean, and σ represents the standard deviation.

[0039] The calculation method of the relative size of λ3 and λ4 is expressed as follows:

[0040]

[0041] Among them, exp() is the exponential function, and are the goodness of fit of the test results of the pre-trained source soft-sensing model and the target soft-sensing model on their corresponding validation sets, and the calculation formula is:

[0042]

[0043] Among them, Y i and are the true value and predicted result of the i-th performance indicator, respectively. is the mean of the true value of the performance indicator. The beneficial effects of the present invention are:

[0044] This process industry soft measurement modeling method can establish a soft measurement model for performance indicators on process industry production lines with insufficient data accumulation by introducing data from production lines with relatively rich data accumulation. This can soft measure some performance indicators in the process industry that are difficult to measure directly or are costly to measure, thereby providing reference and guidance for the monitoring, optimization and control of various parameters in the process industry production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a process industry soft sensor modeling method for cross-production line horizontal information fusion proposed by the present invention.

[0046] Figure 2 It is a schematic diagram of the soft sensor model pre-training process of a process industry soft sensor modeling method for cross-production line horizontal information fusion proposed by the present invention.

[0047] Figure 3 It is a schematic diagram of the training process of the time series feature information fusion of the soft sensor model of the process industry soft sensor modeling method for cross-production line horizontal information fusion proposed by the present invention.

[0048] Figure 4 It is a schematic diagram of the expected effect of the time series feature information fusion training of the process industry soft sensor modeling method for cross-production line horizontal information fusion proposed by the present invention. DETAILED DESCRIPTION

[0049] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings. It should be noted that the various symbols in the expressions involved in the present invention are only for the purpose of clear expression, and different expressions can be used in different examples. Therefore, the key to the present invention lies in the method of horizontal information fusion across production lines and the soft measurement modeling idea based on the information fusion results.

[0050] The purpose of the present invention is to provide a safe, efficient and versatile process industry soft sensor modeling method for cross-production line horizontal information fusion. Figure 1As shown, it includes the following 5 steps:

[0051] Step 1: Read offline data from the DCS database of the data source production line and the target production line respectively, match the process variable time series based on the tag value of the performance indicator, and construct the source data set and the target data set.

[0052] Step 2: Use the source data set and the target data set to pre-train a source soft measurement model and a target soft measurement model respectively, and save the parameters of the two soft measurement models and the extracted source time series feature set and target time series feature set.

[0053] Step 3: Send the source time series feature set and the source soft measurement model parameters to the online platform to which the target production line belongs, aggregate the parameters of the source soft measurement model and the target soft measurement model, and construct a parameter aggregation soft measurement model.

[0054] Step 4: Through the time series feature information fusion loss function, use the target data set, the source time series feature set and the target time series feature set to perform time series feature information fusion training on the parameter aggregation soft sensor model to obtain the information fusion soft sensor model.

[0055] Step 5: Deploy the information fusion soft measurement model on the online platform of the target production line to realize online and real-time soft measurement of the corresponding performance indicators, and upload the soft measurement results of the model to the DCS database for storage.

[0056] The purpose of step 1 is to prepare the corresponding modeling data set for soft measurement modeling. In order to achieve the effect of introducing a new data source to provide information support for soft measurement modeling, in the embodiment of the present invention, two process industrial production lines, namely the data source production line and the target production line, and their corresponding data are usually involved. Among them, the data source production line refers to a production line with a large amount of data accumulation, which is used to provide information support for soft measurement modeling; and the target production line refers to a production line with insufficient data accumulation and a soft measurement model needs to be established. The method of matching the process variable time series based on the performance indicator label value is: first, select the process variables involved in soft measurement modeling from the DCS (Distributed Control System) database, and set a suitable time series length L, and then take the timestamp of the soft measurement performance indicator label value acquisition in the DCS database as the starting point, and intercept the process variable time series with a length of L forward, so that each performance indicator label value within the data set range can correspond to the corresponding process variable time series. Through this data set construction method, the source data set and the target data set can be constructed based on the data source production line and the target production line data respectively. It should be noted that the construction process of the two data sets in step 1 is completed on the online platforms corresponding to the two production lines respectively, and no interaction between the data sets will be generated. The online platform refers to the system platform used to deploy and run the soft sensor model and its incremental learning program, such as cloud computing platform, industrial Internet of Things platform, digital twin platform, etc. In addition, the sequence length L of the time series of the two data sets and the category and number of process variables should be consistent to support the subsequent cross-production line horizontal information fusion.

[0057] The purpose of step 2 is to provide the necessary soft-sensing model parameters and time series features for the horizontal information fusion across production lines. The soft-sensing model used is an end-to-end neural network model consisting of a time series feature extraction network in the form of an autoencoder and a regression network connected forward and backward. Figure 2 As shown in the figure, the time series feature extraction network of the soft measurement model receives the time series of the process variables in the data set, extracts the time series features from it, and the regression network receives the extracted time series features, and obtains the true value of the soft measurement through fitting. The source soft measurement model and the target soft measurement model are trained based on the source data set and the target data set respectively, while the source time series feature set and the target time series feature set are extracted from the source data set and the target data set by the source soft measurement model and the target soft measurement model respectively. In step 2, the pre-training of the two models and the time series feature set extraction process are also completed on the online platforms corresponding to the two production lines. In addition, the structures of the two models should be consistent to support the subsequent model parameter aggregation. Finally, for each data set, the data set needs to be divided into a training set and a validation set during the training process. The soft measurement model trains parameters on the training set, and the model soft measurement performance is verified on the validation set.

[0058] The expression of the pre-training loss function of the soft sensor model is:

[0059]

[0060] Among them, λ1 and λ2 are the regularization coefficients of the pre-training loss function, which are used to adjust the influence of each part of the pre-training loss function; is the supervision loss function of the soft sensor model, and its expression is:

[0061]

[0062] Among them, B is the number of training batches, Y b is the true value of the performance indicator label in the bth batch, and is the prediction result of the model for the performance index in the bth batch. In addition, is the unsupervised loss function of the soft measurement model. This loss function is related to the input and output of the temporal feature extraction network in the form of an autoencoder in the soft measurement model, and its expression is:

[0063]

[0064] Where N is the variable dimension of the process variable time series, is the value of the input process variable time series at the position (b, l, n), and The value of the process variable time series reconstructed by the time series feature extraction network at the position (b, l, n). Compared with simply using the supervised loss function to pre-train the network, using this semi-supervised pre-training loss function can further improve the prediction accuracy of the soft sensor model using unlabeled time series data.

[0065] The purpose of step 3 is to fuse the information of the two production lines at the model level and provide a set of initial model parameters for the cross-production line time series feature information fusion training. The method for obtaining the parameters of the parameter aggregation soft sensor model can be expressed as:

[0066] M agg =w1M source +w2M target

[0067] Among them, M source and M target are the parameters of the source soft sensor model and the target soft sensor model respectively, w1 and w2 are the aggregation weights of the two model parameters, satisfying the condition of w1+w2=1. In general, w1=w2 is set. It should be noted that the model parameter aggregation operation in step 3, as well as the subsequent steps 4 and 5 are all completed on the online platform of the target production line.

[0068] The purpose of step 4 is to further train the parameter aggregation soft sensor model based on the target data set by introducing the source time series feature set and the target time series feature set, so as to further enhance the effect of cross-production line horizontal information fusion from the feature level. Figure 3 As shown in the figure, the target data set, source time series feature set, and target time series feature set all participate in the model training process at the same time. Among them, the time series feature information fusion training loss function partially inherits the structure of the pre-training loss function, but adds to achieve the fusion at the time series feature level. Its expression is:

[0069]

[0070] Among them, λ′1, λ′2, λ3, and λ4 are the regularization coefficients of the time series feature information fusion training loss function, and Z source and Z target are the source time series feature set and the target time series feature set respectively, Z is the information fusion time series feature set output by the soft sensor model during the training process, D(Z A ,Z B ) is Z A , Z B The distribution difference loss function between two feature sets, in this invention, is measured by the multi-core maximum mean difference between the two feature sets:

[0071]

[0072] Among them, z A , z′ A and z B , z′ B Represents Z A and Z B where p represents the number of kernels, E represents the mean, and σ represents the standard deviation. Therefore, D(Z,Z source ) and D(Z,Z target ) are:

[0073]

[0074]

[0075] It is worth noting that, in the designed time series feature information fusion loss function, the relative size of λ3 and λ4 will greatly affect the migration direction of the distribution of information fusion time series features output by the information fusion soft sensor model, such as Figure 4As shown, the migration direction of this characteristic distribution has a direct impact on the soft measurement performance of the soft measurement model. In this regard, the present invention provides a calculation method for determining the relative size of λ3 and λ4, and its expression is:

[0076]

[0077] Among them, exp() is the exponential function, and are the goodness of fit of the test results of the pre-trained source soft-sensing model and the target soft-sensing model on their corresponding validation sets, and the calculation formula is:

[0078]

[0079] Among them, Y i and are the true value and predicted result of the i-th performance indicator, respectively. is the mean of the true values ​​of the performance indicators. 2 The range of is from negative infinity to 1. The larger its value, the better the soft sensor performance of the soft sensor model on the validation set. The exp() function has a 2 The mapping relationship between the performance of the soft measurement model and the performance of the soft measurement model has a smoothing effect. Through this method of determining the relative size of λ3 and λ4, the distribution of the information fusion time series features output by the information fusion soft measurement model is migrated toward the distribution of time series features with better soft measurement effects, but at the same time, the distribution of some target time series feature sets can be retained, thereby comprehensively considering the distribution of the two groups of time series features and avoiding the problems of "negative migration" and "over-migration" to a certain extent.

[0080] Through steps 1-4, an information fusion soft measurement model that integrates the data information of the source production line and the target production line can be trained. This model not only contains the characteristic information of the target production line data, but also introduces the supporting information in the source data production line that is beneficial to improving the soft measurement performance. The obtained information fusion soft measurement model is deployed on the online platform of the target production line to receive the real-time data generated by the target production line, and the process variable time series is constructed in real time to input into the soft measurement model to obtain the real-time soft measurement results of the corresponding performance indicators. The obtained soft measurement results are uploaded to the DCS database for storage, so as to provide information reference for various downstream tasks based on the soft measurement results of performance indicators, such as monitoring, optimization, control and other tasks.

[0081] The present invention proposes a soft measurement modeling method for process industries with cross-production line horizontal information fusion. By introducing data from production lines with relatively rich data accumulation, a soft measurement model for performance indicators can be established safely, effectively and accurately on process industry production lines with insufficient data accumulation. This can perform soft measurement on some performance indicators in the process industry that are difficult to measure directly or are costly to measure, thereby providing reference and guidance for the monitoring, optimization and control of various parameters in the production process of the process industry.

[0082] It should also be noted that the terms "such as", "including" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentences "such as...", "including..." do not exclude the presence of other identical elements in the process, method or device including the elements.

[0083] For those skilled in the art, the present invention may have various changes and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A soft sensor modeling method for process industry with cross-production line horizontal information fusion, characterized in that: The steps include: Step 1: Read offline data from the DCS database of the data source production line and the target production line respectively, match the process variable time series based on the tag value of the performance indicator, and build the source data set and the target data set; Step 2: Use the source data set and the target data set to pre-train a source soft measurement model and a target soft measurement model respectively, and save the parameters of the two soft measurement models and the extracted source time series feature set and target time series feature set; Step 3: Send the source time series feature set and the source soft measurement model parameters to the online platform to which the target production line belongs, aggregate the parameters of the source soft measurement model and the target soft measurement model, and construct a parameter aggregation soft measurement model; Step 4: Through the time series feature information fusion loss function, the target data set, the source time series feature set and the target time series feature set are used to perform time series feature information fusion training on the parameter aggregation soft sensor model to obtain the information fusion soft sensor model; Step 5: Deploy the information fusion soft measurement model on the online platform of the target production line to realize online and real-time soft measurement of the corresponding performance indicators, and upload the soft measurement results of the model to the DCS database for storage.

2. The method according to claim 1, characterized in that In step 1, the data source production line refers to a production line with sufficient data accumulation for providing information support for soft measurement modeling, and the target production line refers to a production line with insufficient data accumulation for which a soft measurement model needs to be established; the method for matching the process variable time series based on the performance indicator label value is: first, the process variables involved in the soft measurement modeling are selected from the DCS database, and a suitable time series length L is set, and then the timestamp of the soft measurement performance indicator label value acquisition in the DCS database is used as the starting point, and the process variable time series with a length of L is intercepted forward, so that each performance indicator label value within the data set range can correspond to the corresponding process variable time series; the source data set and the target data set are data sets constructed based on the data of the data source production line and the target production line, and the sequence length L of the time series of the two data sets and the category and number of process variables are consistent; the construction process of the two data sets in step 1 is completed on the online platforms corresponding to the two production lines respectively, and no interaction between the data sets is generated; the online platform refers to a system platform for deploying and running soft measurement models and their incremental learning programs, including cloud computing platforms, industrial Internet of Things platforms, and digital twin platforms.

3. The method according to claim 1 or 2, characterized in that: In step 2, the soft measurement model is an end-to-end neural network model composed of a time series feature extraction network in the form of an autoencoder and a regression network connected front and back, the source soft measurement model and the target soft measurement model are trained based on the source data set and the target data set respectively, and the source time series feature set and the target time series feature set are extracted from the source data set and the target data set by the time series feature extraction networks of the source soft measurement model and the target soft measurement model respectively; the pre-training of the two models in step 2 and the time series feature set extraction process are completed on the online platforms corresponding to the two production lines respectively, and the structures of the two models are consistent. For each data set, the data set needs to be divided into a training set and a verification set during the training process, the soft measurement model performs parameter training on the training set, and the model soft measurement performance is verified on the verification set; the expression of the loss function used in the pre-training process is: Among them, λ1 and λ2 are the regularization coefficients of the pre-training loss function, is the supervision loss function of the soft sensor model, and its expression is: Among them, B is the number of training batches, Y b is the true value of the performance indicator label in the bth batch, is the prediction result of the model for the performance index in the bth batch. is the unsupervised loss function of the soft sensor model, and its expression is: Where N is the variable dimension of the process variable time series, is the value of the input process variable time series at the position (b, l, n), The value of the process variable time series at the position (b, l, n) reconstructed by the time series feature extraction network.

4. The method according to any one of claims 1 to 3, characterized in that In step 3, the method for obtaining the parameters of the parameter aggregation soft sensor model is expressed as: M agg =w1M source +w2M target Among them, M source and M target are the parameters of the source soft sensor model and the target soft sensor model respectively, w1 and w2 are the aggregation weights of the two model parameters, satisfying the condition of w1+w2=1. Generally, w1=w2. The model parameter aggregation operation in step 3, as well as the subsequent steps 4 and 5 are all completed on the online platform of the target production line.

5. The method according to claim 4, characterized in that In step 4, the expression of the temporal feature information fusion loss function is: Among them, λ ′ 1. Lambda ′ 2. λ3 and λ4 are the regularization coefficients of the time series feature information fusion training loss function, Z source and Z target are the source time series feature set and the target time series feature set respectively, Z is the information fusion time series feature set output by the soft sensor model during the training process, and D(Z,Z source ) and D(Z,Z target ) are Z and Z respectively source , Z target The distribution difference measurement results between two feature sets are expressed as: Among them, MMD(Z b ,Z source ) and MMD(Z b ,Z target ) are Z and Z of the bth batch respectively. source and Z target The maximum multi-core mean difference between , is expressed as: Among them, z A 、z A ′ and z B 、z ′ B Represents Z A and Z B Two randomly selected samples in , p represents the number of kernels, E represents the mean, and σ represents the standard deviation.

6. The method according to claim 5, characterized in that The calculation method of the relative size of λ3 and λ4 is expressed as follows: Among them, exp() is the exponential function, and are the goodness of fit of the test results of the pre-trained source soft-sensing model and the target soft-sensing model on their corresponding validation sets, and the calculation formula is: Among them, Y i and are the true value and predicted result of the i-th performance indicator, respectively. is the mean of the true values ​​of the performance indicators.

Images