A soft measurement modeling method based on pattern perception dynamic variational auto-encoding regression

Abstract

The application discloses a soft measurement modeling method based on mode perception dynamic variational autoencoding regression, and comprises the following steps: (1) acquiring three-phase flow process data with multiple mode dynamic characteristics; (2) data division and pretreatment operation; (3) establishing a mode perception dynamic variational autoencoding regression model, and realizing the prediction of an online quality variable; and (4) predicting a three-phase flow process pressure variable and performing model performance evaluation. The encoder-decoder network with a convolution-deconvolution structure learns the hidden layer feature representation of dynamic data, the multiple mode characteristics of the data are mined in the hidden space by using a Gaussian mixture distribution, the mapping relationship between the latent variable and the key quality variable is effectively learned, in addition, the reconstruction of dynamic multiple mode data is constrained by using a Wasserstein distance, and finally the purpose of improving the accuracy of key quality inference is achieved.

Description

Technical Field

[0001] This invention relates to the field of soft measurement modeling technology for industrial processes with multi-mode and dynamic characteristics, and specifically to a soft measurement modeling method based on pattern-aware dynamic variational autoencoder regression. Background Technology

[0002] The stable operation and product quality control of process industries heavily rely on the accurate acquisition of key quality variables. However, these variables often suffer from long sampling periods, high testing costs, and difficulties in online measurement, making soft measurement technology a crucial means of ensuring safe production. With the expansion of equipment scale and the increasing complexity of operating conditions, industrial processes exhibit significant dynamic, coupled, and multi-mode operating characteristics. Under these circumstances, traditional soft measurement modeling methods face severe challenges. Therefore, constructing highly generalized soft measurement models capable of adapting to dynamic multi-mode characteristics is of great significance for achieving intelligent control of industrial processes.

[0003] Due to the constant changes in raw material properties, operating strategies, and production loads, systems often exhibit multi-mode operation with significant statistical differences. The distribution of process data displays segmented and multimodal characteristics. Currently, various data-driven methods have been developed to solve multi-mode modeling problems, such as ensemble learning methods and adaptive models. Among them, the Gaussian Mixture Model (GMM) is widely used for multi-mode modeling due to its ability to capture and generate multiple Gaussian components. In actual industrial process operation, multi-mode characteristics and dynamic temporal features are often coupled and coexist; therefore, modeling research that considers both has significant practical implications.

[0004] Deep learning's powerful data feature extraction capabilities offer a new approach to soft measurement modeling of industrial processes. Variational Autoencoders (VAEs) are widely used in industrial process modeling due to their strong probabilistic modeling capabilities, structural flexibility, and good mathematical interpretability. However, standard VAEs assume that the underlying prior follows a single Gaussian distribution, which is challenging for capturing distributional differences in processes with multimodal characteristics. Although researchers have improved performance by introducing GMM models into VAEs, this approach focuses on static data modeling and lacks accurate capture of complex time dependencies.

[0005] Industrial processes typically exhibit significant time correlations, with the relationships between variables constantly changing with variations in operating conditions, load disturbances, and equipment state evolution, displaying typical dynamic characteristics. A recent invention proposes a Time-series Variational Autoencoder (TimeVAE) method. This method, based on VAE, constructs a multivariate time series generation architecture that can capture the dynamic temporal correlations and nonlinear dependencies between variables in industrial processes. However, this method still has limitations in modeling the prevalent multi-mode characteristics of industrial processes, and it struggles to fully characterize the differences in temporal features under different operating modes and the switching patterns between modes. Summary of the Invention

[0006] The purpose of this invention is to provide a soft measurement modeling method based on Mode-aware Dynamic Variational Autoencoder Regression (MADVAR) to address the difficulty of modeling key quality variables in industrial processes with multi-mode dynamic characteristics. This invention first mines the latent feature representation of dynamic data through an encoder-decoder structure; then, a Gaussian mixture distribution is introduced into the latent space to extract the distribution characteristics of the multi-mode data; furthermore, a mapping relationship between latent features and key quality variables is constructed, enabling online quality prediction. During model training, Wasserstein distance is introduced to constrain the reconstruction process of dynamic data and improve the model's ability to model multi-mode dynamic coupling features.

[0007] The technical solution adopted in this invention is as follows: S1. Acquire three-phase flow process data with multi-mode dynamic characteristics: The three-phase flow process data constructs a three-phase flow system for accurately delivering a real-time adjustable three-phase mixture of oil, water, and gas to the pressurization device; the three-phase flow system generates various flow field distributions by dynamically tuning operating parameters to adapt to different working conditions and actual application scenarios; at the same time, affected by the inherent periodicity of the production process and the feedback control mechanism, the data distribution shifts over time, exhibiting significant dynamic time-series drift characteristics. S2. Data partitioning and preprocessing operations: S21. Divide the acquired samples into training and testing datasets according to the proportions; S22. Standardize all features to eliminate dimensional differences and improve model convergence speed; S23. Data serialization is achieved through a sliding window; S3. Establish a MADVAR model, namely a pattern-aware dynamic variational autoencoder regression model, and achieve online prediction of quality variables: S31. Mining the hidden feature representation of dynamic data through encoder-decoder structure; S32. In the latent space, a Gaussian mixture distribution is introduced to extract the distribution characteristics of multi-mode data; S33. Construct a mapping relationship between latent features and key quality variables to achieve online quality prediction; during model training, Wasserstein distance is introduced to constrain the reconstruction process of dynamic data and improve the model's ability to model multi-modal dynamic coupling features. S4. Predict pressure variables in a three-phase flow process and evaluate model performance: Predict pressure variables in a three-phase flow process using an online prediction model; select root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Three indicators are used to quantitatively evaluate the effect after sample expansion.

[0008] Furthermore, in step S3, the MADVAR model includes an encoder, a decoder, and a regressor; For dynamic multimodal data An encoder consisting of one-dimensional convolutional layers is used for feature extraction to obtain multimodal latent variables. ,in For latent variable dimensions; For latent variables Let the discrete latent component label be... ;variable It is a one-hot class vector. This represents the number of Gaussian components contained in the multimodal data; the prior distribution based on GMM is defined as follows: (1) (2) in Indicates the first k The mixing coefficient of the Gaussian components; This indicates that the first [item] in the potential space has been selected. k One Gaussian component; Simplified to ; and They represent the first k The mean vector and covariance matrix of each Gaussian component.

[0009] Furthermore, the decoder reconstructs the time series using a deconvolution structure, and its conditional generation distribution is approximated by a Gaussian distribution: (3) in Indicates process variable decoder, The reconstructed sequence generated for the MADVAR decoder; This represents the variance of the Gaussian distribution reconstructed from the process variable.

[0010] Furthermore, when constructing a soft measurement model, the regressor incorporates an output quality variable. Its generation distribution is expressed in Gaussian form: (4) in, Indicates the regressor. This is a soft measurement prediction result based on latent variables; This represents the Gaussian distribution variance of the predictions for the quality variable.

[0011] Furthermore, the generation process of the MADVAR method based on encoder-regressor is as follows: (5).

[0012] Furthermore, based on the Wasserstein distance loss function, within the distribution matching modeling framework, an approximate posterior is introduced as the basis for latent distribution modeling and Wasserstein distance constraints, using a learnable variational posterior for approximation: (6) in, Representing a discrete posterior distribution, parameterized as a neural network with a softmax layer as the last layer, it is represented as: (7) in This is the classification branch of the convolutional encoder, used to calculate the probability that a sample belongs to each pattern; In formula (6) This represents the posterior distribution of a continuous latent variable, specifically: (8) Among them, the mean and variance The feature extraction branch output of the MADVAR encoder is used to sample latent variables from the above distribution through reparameterization. This is used to achieve end-to-end training. Constraining the distribution structure of continuous latent variables under different modes, approximating the posterior distribution based on the Wasserstein distance metric. With the corresponding prior distribution The difference between them is expressed as: (9) Here, the parameters represent the prior Gaussian distribution under mode c; they directly reflect the geometric differences between different latent distributions at the mean and covariance levels, and characterize the multimodal latent space structure.

[0013] Furthermore, under the generative model assumption, the MADVAR model simultaneously processes dynamic input sequences. and quality variables The model is constructed, and its reconstruction is defined as follows: (10) Its regression loss is defined as: (11) The overall optimization objective of the MADVAR model is defined as: (12) in These are weighting coefficients used to balance the relationship between sequence reconstruction, soft measurement accuracy, and potential distribution constraints.

[0014] Furthermore, after completing offline training, the MADVAR model is used to predict quality variables of real-time acquired sequence samples; given the current sample sequence... The online prediction steps are as follows: S41. Calculate the posterior probability of each mixture component using the MADVAR encoder: (13) S42. In each ingredient k Below, according to To obtain latent features , ; S43. Calculate the mean quality prediction for each component using a regressor, and obtain... ; S44. The predicted quality value is obtained by weighted summation: (14).

[0015] Compared with existing technologies, the soft measurement modeling method based on pattern-aware dynamic variational autoencoder regression provided by this invention achieves at least the following beneficial effects: This invention learns the latent feature representation of dynamic data through a convolutional-deconvolutional encoder-decoder network and utilizes Gaussian mixture distribution to mine the multi-modal characteristics of the data in the latent space, effectively learning the mapping relationship between latent variables and key quality variables. Furthermore, it employs Wasserstein distance to constrain the reconstruction of dynamic multi-modal data, ultimately improving the accuracy of key quality inference. For dynamic processes with multi-modal characteristics, a Gaussian mixture prior is introduced into the latent space to explicitly describe the distribution boundaries of different operating conditions while preserving the time-dependent features of the dynamic data. Moreover, by establishing a regression model in the latent space, the mapping of latent features to key quality variables is realized, enabling feature extraction and quality prediction to be completed within the same probabilistic framework. Attached Figure Description

[0016] Figure 1 This is a structural diagram of the MADVAR model of the present invention.

[0017] Figure 2 This is a comparison chart of the prediction performance of the present invention on the test set.

[0018] Figure 3 This is a box plot of the absolute prediction error of the present invention on the test set. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0021] Reference Figures 1-3 A pattern-aware dynamic variational autoencoder regression soft measurement modeling method, taking the construction of a pressure prediction model in a three-phase flow process as an example, includes the following steps: S1. Acquire multi-mode dynamic three-phase flow process data; The core design principle of the three-phase flow system is to precisely deliver a real-time controllable oil-water-gas mixture to pressurized equipment. The system can generate various flow field distributions by flexibly adjusting operating parameters, thereby meeting the requirements of various working conditions and application scenarios. This invention sets three modes based on the different flow rates of the water and gas phases. The water flow velocities for the three modes are set to 2 kg / s, 1 kg / s, and 2 kg / s, respectively, and the gas flow velocities are set to 0.02 m³ / s, 0.03 m³ / s, and 0.04 m³ / s, respectively. During the model building process, this invention selects pressure as the core output parameter and selects 16 process parameters with strong correlation as auxiliary input variables, specifically including: VC302 valve opening, air inlet flow rate, air supply pressure, water flow rate, water density, bottom pressure of riser, top density of riser, top pressure of riser, pressure difference between PT401 and PT408, liquid level height of top separator, top separation pressure, pressure difference of VC404, valve opening of VC501, component ratio of gas-liquid three-phase separator, water pump operating current, and valve opening of VC101.

[0022] S2. Dataset partitioning and preprocessing for three-phase flow processes: First, the collected samples are proportionally split into training and testing sets; then, all features are standardized to unify the dimensions of each feature and accelerate the model convergence speed; finally, a sliding window is used to complete the data serialization and reconstruction.

[0023] Step S21: Split the dataset The collected raw samples were divided into two parts: a training set with 1503 samples and a test set with 603 samples.

[0024] Step S22: Data normalization processing To eliminate the inconsistency in the dimensions of the features and improve the model's convergence efficiency and prediction performance, the original data is normalized, and its calculation expression is shown in equation (18): (18) In the formula: , , These represent the original sample data, the maximum value of the original data, and the minimum value of the original data, respectively. These are the feature data after normalization.

[0025] Step S23: Data serialization processing The sliding window size was set to 4 to reconstruct the time series data. After processing, 1500 time series samples were obtained for the training set and 600 time series samples for the test set.

[0026] S3. Establish a pattern-aware dynamic variational autoencoder regression model and achieve online prediction of quality variables: Step S31: The overall structure of the MADVAR model mainly consists of three parts: encoder, decoder, and regressor. First, for dynamic multi-mode data... An encoder consisting of one-dimensional convolutional layers is used for feature extraction to obtain multimodal latent variables. ,in For latent variable dimensions.

[0027] For latent variables Let the discrete latent component label be... .variable It is a one-hot class vector. This represents the number of Gaussian components contained in the multimodal data. The prior distribution based on GMM is defined as follows: (1) (2) in Indicates the first k The mixing coefficient of the Gaussian components. This indicates that the first [item] in the potential space has been selected. k Gaussian components. For simplicity, the following will refer to... Simplified to . and They represent the first k The mean vector and covariance matrix of each Gaussian component.

[0028] The decoder uses a deconvolution structure to reconstruct the time series, and its conditional generation distribution can be approximated by a Gaussian distribution: (3) in Indicates process variable decoder, The reconstructed sequence generated for the MADVAR decoder. This represents the variance of the Gaussian distribution reconstructed from the process variable.

[0029] For the regressor, in order to construct a soft measurement model, an output quality variable is introduced. Its generation distribution can also be expressed in Gaussian form: (4) in, Indicates the regressor. These are soft measurement prediction results based on latent variables. This represents the Gaussian distribution variance of the predictions for the quality variable.

[0030] Therefore, the generation process of the MADVAR method based on encoder-regressor is as follows: (5) Step S32: Within the distribution matching modeling framework, the loss function based on Wasserstein distance introduces an approximate posterior as the basis for latent distribution modeling and Wasserstein distance constraints. Because the true posterior... Since it is difficult to compute directly, this method uses a learnable variational posterior for approximation: (6) in, The discrete posterior distribution can be parameterized as a neural network with a softmax layer as its last layer, represented as: (7) in This is the classification branch of the convolutional encoder, used to calculate the probability that a sample belongs to each pattern.

[0031] In formula (10) This represents the posterior distribution of a continuous latent variable, specifically: (8) Among them, the mean and variance The output is from the feature extraction branch of the MADVAR encoder. Through reparameterization techniques, latent variables can be sampled from the aforementioned distribution. This enables end-to-end training.

[0032] To constrain the distribution structure of continuous latent variables under different modes, an approximate posterior distribution is based on the Wasserstein distance metric. With the corresponding prior distribution The difference between them is expressed as: (9) in This represents the prior Gaussian distribution parameters under mode c. This distance can directly reflect the geometric differences between different latent distributions in terms of mean and covariance, and is more conducive to characterizing the multimodal latent space structure.

[0033] Furthermore, under the generative model assumption, MADVAR simultaneously processes dynamic input sequences. and quality variables The modeling is performed, and the reconstruction and regression losses are defined as follows: (10) (11) Combining formulas (13), (14), and (15), the overall optimization objective of MADVAR is defined as: (12) in These are weighting coefficients used to balance the relationship between sequence reconstruction, soft measurement accuracy, and potential distribution constraints.

[0034] Step S33: Online prediction of key quality variables After offline training, the MADVAR model is used to predict quality variables of real-time acquired sequence samples. Given the current sample sequence... The online prediction steps are as follows: First, the posterior probabilities of each mixture component are calculated using the MADVAR encoder: (13) Next, in each component k Below, according to To obtain latent features , .

[0035] Subsequently, the mean quality prediction for each component was calculated using a regressor, resulting in... .

[0036] Ultimately, the quality prediction is obtained by weighted summation: (14) (4) Predict the pressure variables of the three-phase flow process and evaluate the model performance: Pressure variables in a three-phase flow process are predicted using an online prediction model. To objectively verify the invention, root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are selected. 2 Three indicators are used to quantitatively evaluate the effect after sample expansion.

[0037] (15) (16) (17) in: Represents the true value The corresponding model prediction value, The average value of the real samples in the test set. This represents the total number of samples in the test set. The smaller the values ​​of the root mean square error (RMSE) and mean absolute error (MAE), the higher the coefficient of determination (R²). 2The closer the value is to 1, the better the model's predictive performance and generalization ability.

[0038] To verify the effectiveness of the MADVAR method, this invention compares its predictive performance with the following methods: Variational Autoencoder Regression (VAER), Long Short-Term Memory (LSTM), Gaussian Mixture Variational Autoencoder Regression (GMVAER), and Time Variational Autoencoder Regression (TimeVAER). Table 1 presents the detailed predictive performance results of the six different methods on a three-phase flow process test set. The VAER method exhibits the worst predictive performance, achieving the highest RMSE and MAE values, and the lowest R... 2 The main reason for this is that VAER is a static model, ignoring the temporal dynamics of the data. The GMVAER method captures multi-modal characteristics in the hidden layers of the variational autoencoder, resulting in better prediction performance compared to VAER. Although LSTM and TimeVAER can learn the dynamic characteristics of time series, they struggle to capture the characteristics of multi-modal data. Therefore, the prediction accuracy of LSTM and TimeVAER is still not good enough. MADVAR achieves the best prediction performance because it can simultaneously capture the dynamic features of the data and learn the distribution characteristics among multi-modal data.

[0039] Table 1 Detailed Prediction Results of Three-Phase Flow Process

[0040] Figure 2 The results of five models fitting the time series of pressure data for a three-phase separator are presented. The MADVAR model shows the highest degree of overlap between the predicted and actual curves, accurately capturing dynamic changes and peak and trough values ​​of the pressure, demonstrating the best fitting performance. Figure 3 The absolute prediction errors of each method are presented. As shown in the figure, the performance differences among the five models are significant in terms of absolute prediction error. The MADVAR model has the lowest median error and the smallest quartile, indicating that it has the best prediction accuracy and stability. TimeVAER and GMVAER perform next, with relatively similar error distributions, both outperforming LSTM and VAER. In contrast, LSTM and VAER have larger median errors and more high-value outliers, indicating relatively weaker prediction performance and robustness. Therefore, the experimental results demonstrate the importance of fully exploring multi-mode and dynamic characteristics for establishing accurate quality prediction models.

[0041] The method of this invention employs a soft measurement modeling method based on pattern-aware dynamic variational autoencoder regression to capture the temporal dynamics and multi-mode boundary features of data, thereby realizing the mapping of latent features to key quality variables.

[0042] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A soft sensor modeling method based on pattern-aware dynamic variational autoencoder regression, characterized in that, Includes the following steps: S1. Obtain three-phase flow process data with multi-mode dynamic characteristics: The three-phase flow process data constructs a three-phase flow system, and the three-phase flow system generates various flow field distributions by dynamically tuning operating parameters; S2. Data partitioning and preprocessing operations: S21. Divide the acquired samples into training and testing datasets according to the proportions; S22. Standardize all features to eliminate dimensional differences and improve model convergence speed; S23. Data serialization is achieved through a sliding window; S3. Establish a MADVAR model, namely a pattern-aware dynamic variational autoencoder regression model, and achieve online prediction of quality variables: S31. Mining the hidden feature representation of dynamic data through encoder-decoder structure; S32. In the latent space, a Gaussian mixture distribution is introduced to extract the distribution characteristics of multi-mode data; S33. Construct a mapping relationship between latent features and key quality variables to achieve online quality prediction; during model training, Wasserstein distance is introduced to constrain the reconstruction process of dynamic data and improve the model's ability to model multi-modal dynamic coupling features. The MADVAR model includes an encoder, a decoder, and a regressor. For dynamic multimodal data An encoder consisting of one-dimensional convolutional layers is used for feature extraction to obtain multimodal latent variables. ,in For latent variable dimensions; For latent variables Let the discrete latent component label be... ;variable It is a one-hot class vector. This represents the number of Gaussian components contained in the multimodal data; the prior distribution based on GMM is defined as follows: (1) (2) in Indicates the first k The mixing coefficient of the Gaussian components; This indicates that the first [item] in the potential space has been selected. k One Gaussian component; Simplified to ; and They represent the first k The mean vector and covariance matrix of each Gaussian component; The decoder reconstructs the time series using a deconvolution structure, and its conditional generation distribution is approximated by a Gaussian distribution: (3) in Indicates process variable decoder, The reconstructed sequence generated for the MADVAR decoder; This represents the variance of the Gaussian distribution reconstructed from the process variables; The regressor incorporates an output quality variable when constructing a soft measurement model. Its generation distribution is expressed in Gaussian form: (4) in, Indicates the regressor. This is a soft measurement prediction result based on latent variables; This represents the Gaussian distribution variance of the predictions for the quality variable; S4. Predict pressure variables in a three-phase flow process and evaluate model performance: Predict pressure variables in a three-phase flow process using an online prediction model; select root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Three indicators are used to quantitatively evaluate the effect after sample expansion.

2. The method according to claim 1, characterized in that: The MADVAR method based on encoder-regressor generation process is as follows: (5)。 3. The method according to claim 2, characterized in that: A loss function based on Wasserstein distance is used. Within the distribution matching modeling framework, an approximate posterior is introduced as the basis for latent distribution modeling and Wasserstein distance constraints, employing a learnable variational posterior for approximation. (6) in, Representing a discrete posterior distribution, parameterized as a neural network with a softmax layer as the last layer, it is expressed as: (7) in This is the classification branch of the convolutional encoder, used to calculate the probability that a sample belongs to each pattern; In formula (6) This represents the posterior distribution of a continuous latent variable, specifically: (8) Among them, the mean and variance The feature extraction branch output of the MADVAR encoder is used to sample latent variables from the above distribution through reparameterization. This is used to achieve end-to-end training. Constraining the distribution structure of continuous latent variables under different modes, approximating the posterior distribution based on the Wasserstein distance metric. With the corresponding prior distribution The difference between them is expressed as: (9) in It represents the prior Gaussian distribution parameters under mode c; it directly reflects the geometric differences between different latent distributions in terms of mean and covariance, and characterizes the multimodal latent space structure.

4. The method according to claim 1, characterized in that: Under the generative model assumption, the MADVAR model simultaneously processes dynamic input sequences. and quality variables The model is constructed, and its reconstruction is defined as follows: (10) Its regression loss is defined as: (11) The overall optimization objective of the MADVAR model is defined as: (12) in These are weighting coefficients used to balance the relationship between sequence reconstruction, soft measurement accuracy, and potential distribution constraints.

5. The method according to claim 3, characterized in that: After offline training, the MADVAR model is used to predict quality variables of real-time acquired sequence samples; given the current sample sequence The online prediction steps are as follows: S41. Calculate the posterior probability of each mixture component using the MADVAR encoder: (13) S42. In each ingredient k Below, according to To obtain latent features , ; S43. Calculate the mean quality prediction for each component using a regressor, and obtain... ; S44. The predicted quality value is obtained by weighted summation: (14)。

Images