A soft measurement modeling method based on deep spatio-temporal autoencoder

Abstract

The application discloses a kind of soft measurement modeling methods based on deep spatio-temporal auto-encoder, belong to industrial process soft measurement modeling field.First, select and need the auxiliary variable related to the quality index to be predicted, data set is constructed using sliding window method.Then, design a memory module and embed into auto-encoder, construct time series auto-encoder for extracting the time series feature of data;Then, a kind of feature calibration module is developed and embedded into time series auto-encoder to construct spatio-temporal auto-encoder, for capturing the spatial feature of data.Finally, stack multiple spatio-temporal auto-encoders to constitute deep spatio-temporal auto-encoder, the spatio-temporal feature extracted is input to a fully connected layer, realize the real-time prediction of key quality index in industrial process.Testing through factory real-time data, the result shows that the proposed deep spatio-temporal auto-encoder can extract more robust high-level semantic features, more accurate than existing benchmark model in soft measurement modeling.

Description

Technical Field

[0001] This invention pertains to soft measurement methods for key quality indicators in industrial processes, specifically involving a soft measurement modeling method based on a deep spatiotemporal autoencoder. Background Technology

[0002] In recent years, with the introduction of the concepts of "carbon peaking" and "carbon neutrality," green and intelligent manufacturing has received increasing attention from academia and industry. To achieve this goal, there is an urgent need for real-time and accurate prediction, control, and optimization of key performance indicators in modern industrial processes. However, due to the limitations of testing instruments and the harshness of the industrial environment, many key variables related to product quality are difficult to measure. To address this, researchers have proposed soft measurement techniques, which use easily measurable auxiliary variables to predict difficult-to-measure quality variables. Currently, there are two typical soft measurement methods: mechanism-based models and data-driven models. Mechanism-based modeling methods are based on physicochemical theories and require a large amount of prior knowledge to establish accurate mathematical models. However, the complexity and difficulty in obtaining mechanistic knowledge of industrial processes make mechanism modeling extremely challenging. Therefore, data-driven soft measurement modeling has been widely applied in industrial processes.

[0003] Over the past few decades, with the rise of artificial intelligence and big data, data-driven soft measurement models have been widely applied in process industries such as metallurgy, chemical engineering, and oil refining. Traditional data-driven modeling methods mainly include various multivariate statistical analysis methods and machine learning models, such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Partial Least Squares (PLS). However, these models are generally suitable for extracting linear correlations from process data. To extract complex nonlinear features, some well-known machine learning methods, such as Support Vector Machines (SVM) and Artificial Neural Networks (ANN), have also been applied to industrial processes. Today, with the increasing demands for product quality and environmental protection, modern industrial processes have become extremely complex and large-scale. Shallow learning models struggle to handle these strongly nonlinear, high-dimensional, strongly coupled, and dynamic process data. To overcome the inherent limitations of these methods, many researchers have begun to seek deep architectures to model these complex industrial data. Hinton et al. first proposed an unsupervised hierarchical pre-training and supervised fine-tuning method in 2006, successfully solving the gradient vanishing and exploding problems. Since then, deep learning, with its superior ability to automatically capture nonlinear features, has gained favor among many researchers. At the same time, deep network models have become a research hotspot for soft measurement in the process industry. Summary of the Invention

[0004] This invention addresses the problem that strong spatiotemporal coupling between variables in industrial processes makes it difficult to predict key quality changes. It innovatively proposes a soft measurement modeling method based on a deep spatiotemporal autoencoder.

[0005] This invention is achieved using the following technical solution:

[0006] This invention first provides a soft measurement modeling method based on a depth spatiotemporal autoencoder, which includes the following steps:

[0007] 1) Read relevant data from the real-time industrial process database and preprocess the data to construct the original dataset; assuming there are n samples in the dataset, where M auxiliary variables are easily measurable and one is the quality variable to be predicted, then the original dataset can be represented as D(x,y)={(x j ,y j Let x = |j = 1, 2, 3, ..., n}, where j represents the j-th sample, x j Let y represent the auxiliary variable. j This represents the quality variable to be measured;

[0008] 2) The data is divided into segments using the sliding window technique to construct training, validation, and test sets. If the length of the sliding window segment is T, then the serialized data segment is represented as follows:

[0009] 3) The autoencoder is unfolded into a chain structure in chronological order, and a memory module is embedded between two adjacent autoencoders to construct a temporal autoencoder. The memory module is used to store useful historical information and pass it to the next autoencoder.

[0010] 4) After obtaining the hidden layer features in the temporal autoencoder, the feature calibration module is connected to the encoding module of the temporal autoencoder, and the hidden layer features are input into the feature calibration module to quantify the importance between the extracted hidden layer features and the quality variables, thus building a spatiotemporal autoencoder; then the extracted spatiotemporal features are input into the fully connected layer to make real-time predictions of the quality variables.

[0011] 5) Stack individual spatiotemporal autoencoders to build a deep spatiotemporal autoencoder model, train the built deep spatiotemporal autoencoder model, and validate it with a test dataset to finally obtain a validated deep spatiotemporal autoencoder model.

[0012] Preferably, according to a preferred embodiment of the present invention, in step 1), relevant data are collected from actual industrial processes, and the main auxiliary variables are screened using a combination of mathematical analysis and mechanistic analysis. Then, the data is preprocessed, including data labeling, data cleaning, data conversion, feature engineering, and other steps.

[0013] As a preferred embodiment of the present invention, in step 2), in order to better preserve historical information and effectively acquire temporal features, autoencoders are chained together to form a chain structure, and a memory module is designed and embedded therein to store the valid information from the previous moment and pass it to the next moment. Following this structure, a TAE network can be constructed by connecting them sequentially.

[0014] Preferably, according to a preferred embodiment of the present invention, in step 3), for a given input sequence X = {x1, x2, x3, ..., x...} t}, where x t The hidden layer of the temporal autoencoder represents the sequence of auxiliary variables at time t as follows:

[0015]

[0016] in, These represent the weights and biases of the autoencoder at time t, respectively, and D is the hidden layer h. t The dimension, m t f represents the memory state at the current moment. t h represents the encoding function of a temporal autoencoder. t represents the hidden layer features obtained by the temporal autoencoder, and [] represents concatenating the vectors.

[0017] Preferably, according to a preferred embodiment of the present invention, in step 3), to accurately describe the correlation between the features extracted by TAE and the quality variable, one-dimensional global average pooling is first used to process the features extracted by TAE. Then, the obtained matrix is ​​input into a two-layer fully connected network to learn the contribution of each feature. Here, in order to better learn the features related to the quality variable, label information is introduced into the network and used as a loss function for training. This can obtain the spatiotemporal features related to the target and improve the accuracy of soft measurement modeling.

[0018] Specifically, the processing flow of the memory module in step 3) is as follows:

[0019] The memory module contains a gating unit g. t-1 An autoencoder is used to dynamically control how much historical information needs to be retained or forgotten, so that useful historical information can be passed to the next state.

[0020]

[0021] Feature constructor This is used to construct a large feature vector, which is then used as input to the scoring function G. The expression is:

[0022]

[0023] Where, x t-1 h represents the input from the previous time step. t-1 x represents the hidden layer features of the previous time step. t The current input variables are represented by [], which indicates that these vectors are concatenated. It can be represented as: W α W β , and W λ These are the weight parameters that need to be learned;

[0024] The gating unit uses a two-layer simple fully connected network for computation. and These represent the weights that need to be learned. and These represent the biases of the two network layers, and σ and tanh represent the sigmoid activation function and the hyperbolic sine activation function, respectively.

[0025]

[0026] The state m of the memory module t Expressed as:

[0027] m t =g t-1 GRU(x t-1 )+(1-g t-1 )m t-1 (6)

[0028] Here, GRU stands for memory function, which is implemented by a gated loop unit and is used to capture the historical information that needs to be remembered from the previous time step.

[0029] Preferably, according to a preferred embodiment of the present invention, in step 4), a feature calibration module is designed to calculate the importance of hidden layer features and quality variables, so as to extract hidden layer features related to quality variables more efficiently;

[0030] Suppose a batch of input variables are B represents the batch size, so the output of the time-series autoencoder is expressed as: The feature calibration module includes two operations. The first operation is to compress H in the batch dimension using one-dimensional global average pooling and then calculate the average value. The second operation involves inputting Q into a two-layer neural network to adaptively learn the weights of each hidden feature, which represent the importance of the feature. Then, the label information is introduced into the spatiotemporal autoencoder network for supervised learning to capture the spatial relationship between hidden features and quality variables.

[0031] The deep spatiotemporal autoencoder described in this invention is constructed by hierarchically stacking multiple spatiotemporal autoencoders, enabling layer-by-layer learning of hierarchical features related to quality variables. First, the original data is input into the first spatiotemporal autoencoder. After a nonlinear transformation, the first hidden layer feature is obtained by minimizing the error between the input and output. This first hidden layer feature is then used as the input to the second spatiotemporal autoencoder to obtain the second hidden layer feature, and so on, thus constructing the deep spatiotemporal autoencoder. Finally, the obtained spatiotemporal features are pre-trained and fed into a fully connected layer for online prediction of quality indicators.

[0032] Preferably, according to a preferred embodiment of the present invention, in step 5), after constructing the spatiotemporal autoencoder, multiple spatiotemporal autoencoders are stacked to form a depth spatiotemporal autoencoder, specifically as follows:

[0033] First, the first spatiotemporal autoencoder is pre-trained to obtain hidden layer features H1. Then, the parameters of the first spatiotemporal autoencoder are frozen, and the first hidden layer H1 is used as the input of the next autoencoder. Similarly, by minimizing the error between the input and output, the hidden layer features H2 of the second spatiotemporal autoencoder are obtained and used as the input of the third spatiotemporal autoencoder. This process continues, with a series of spatiotemporal autoencoders stacked layer by layer to form a deep spatiotemporal autoencoder. Finally, the extracted spatiotemporal features are input into a fully connected layer to achieve online prediction of quality indicators.

[0034] Preferably, in step 5), the parameters of the online monitoring model for key quality indicators are adjusted in real time based on real-time data from the industrial process, and continuously optimized and iterated to make the model more robust.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. For process industries, the entire production process is a continuous physical and chemical reaction process, and the data often has strong temporal sequence characteristics. Autoencoders are static network structures, and for this type of temporal data, it is difficult to effectively extract high-level semantic features, posing challenges to soft measurement modeling. Therefore, this invention proposes a Temporal Autoencoder (TAE) to capture the inherent temporal correlation of data and improve the model's feature representation capability.

[0037] 2. In actual industrial production, there are numerous process variables, each with varying degrees of importance to quality indicators. However, traditional autoencoders do not consider label information during feature learning, resulting in features extracted during pre-training failing to learn their correlation with the labels. To address this issue, this invention embeds a novel Feature Recalibration Module into the Temporal Autoencoder (STAE) to explicitly quantify the differences in feature contributions across spatial dimensions. A feature reconstruction module is also embedded within the spatiotemporal autoencoder to extract latent features related to quality indicators, improving the model's robustness. Attached Figure Description

[0038] Figure 1 A schematic diagram of soft measurement modeling based on a depth spatiotemporal autoencoder;

[0039] Figure 2 Schematic diagram of a timing autoencoder;

[0040] Figure 3 Schematic diagram of a spacetime autoencoder;

[0041] Figure 4 Schematic diagram of a depth spatiotemporal autoencoder;

[0042] Figure 5 Comparison of prediction results of soft measurement model based on depth spatiotemporal autoencoder in test samples. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0044] Figure 1 This invention provides specific construction and application steps for a soft measurement model based on a deep spatiotemporal autoencoder. The process can be summarized as follows: First, through mechanism analysis and expert knowledge, variables related to quality indicators are selected, and relevant data are read and preprocessed from an actual factory, using a sliding window approach to construct a dataset. Then, a temporal autoencoder and a spatiotemporal autoencoder are built to obtain the temporal and spatial correlations between data. Subsequently, multiple spatiotemporal autoencoders are stacked to form a deep spatiotemporal autoencoder, which is used to extract high-level semantic features of the data. The obtained spatiotemporal features are then input into a fully connected layer to achieve real-time prediction of key quality indicators. Finally, the model results are debugged and verified in an actual industrial field.

[0045] This embodiment provides the specific steps of the soft measurement modeling method based on a depth spatiotemporal autoencoder, including the following steps:

[0046] 1) Read relevant data from the real-time industrial process database and preprocess the data to construct the original dataset; assuming there are n samples in the dataset, where M auxiliary variables are easily measurable and one is the quality variable to be predicted, then the original dataset can be represented as D(x,y)={(x j ,y j Let x = |j = 1, 2, 3, ..., n}, where j represents the j-th sample, x j Let y represent the auxiliary variable. j This represents the quality variable to be measured;

[0047] 2) The data is divided into segments using the sliding window technique to construct training, validation, and test sets. If the length of the sliding window segment is T, then the serialized data segment is represented as follows:

[0048] 3) The autoencoder is unfolded into a chain structure in chronological order, and a memory module is embedded between two adjacent autoencoders to construct a temporal autoencoder. The memory module is used to store useful historical information and pass it to the next autoencoder, such as... Figure 2 As shown. For a given input sequence X = {x1, x2, x3, ..., x...} t}, where x t The hidden layer of the temporal autoencoder represents the sequence of auxiliary variables at time t as follows:

[0049]

[0050] in, These represent the weights and biases of the autoencoder at time t, respectively, and D is the hidden layer h. t The dimension, m t f represents the memory state at the current moment. t h represents the encoding function of a temporal autoencoder. t represents the hidden layer features obtained by the temporal autoencoder, and [] represents concatenating the vectors.

[0051] The calculation process for the memory module is as follows:

[0052] The memory module contains a gating unit g. t-1 An autoencoder is used to dynamically control how much historical information needs to be retained or forgotten, so that useful historical information can be passed to the next state.

[0053]

[0054] Feature constructor This is used to construct a large feature vector, which is then used as input to the scoring function G. The expression is:

[0055]

[0056] Where, x t-1 h represents the input from the previous time step. t-1 x represents the hidden layer features of the previous time step. t The current input variables are represented by [], which indicates that these vectors are concatenated. It can be represented as: W α W β , and W λ These are the weight parameters that need to be learned;

[0057] The gating unit uses a two-layer simple fully connected network for computation. and These represent the weights that need to be learned. and These represent the biases of the two network layers, and σ and tanh represent the sigmoid activation function and the hyperbolic sine activation function, respectively.

[0058]

[0059] The state m of the memory module t Expressed as:

[0060] m t =g t-1 GRU(x t-1 )+(1-g t-1 )m t-1 (6)

[0061] Here, the GRU function stands for memory function, used to capture historical information that needs to be remembered from the previous time step. In this embodiment, the GRU is implemented by a gated recurrent unit, which is a type of recurrent neural network.

[0062] Therefore, the decoder part of TAE can be represented as:

[0063]

[0064] in, and These represent the weights and biases that TAE needs to learn at time t, respectively.

[0065] The loss function for TAE is:

[0066]

[0067] Where N represents the number of samples.

[0068] 4) Based on the temporal autoencoder, develop a feature calibration module to quantify the importance between extracted features and the target variable, and build a spatiotemporal autoencoder, such as... Figure 3 As shown. The extracted spatiotemporal features are then input into a fully connected layer for real-time prediction of the quality variable. A feature calibration module is designed to calculate the importance of features to the quality variable, in order to extract latent features related to the quality variable more efficiently. Assume a batch of input variables is... B represents the batch size, so the output of the encoded part can be represented as: The feature calibration module includes two operations. The first operation is to compress H in the batch dimension using one-dimensional global average pooling (GAP) and calculate the average value. The second operation involves inputting Q into a two-layer neural network to adaptively learn the weights of each feature, representing their importance. Next, label information is introduced into the network for supervised learning, capturing the spatial relationship between features and labels. The expression for Q can be formalized as:

[0069]

[0070] Where Q(t,m) represents the value of the m-th variable at time t after the GAP operation, and B is the number of samples in each batch.

[0071] Then, a gating mechanism with two fully connected layers is used to extract the correlation of features, where S represents the weight coefficient tensor of each feature.

[0072]

[0073] in, and These are the weight parameters of the two fully connected layers, and σ represents the Sigmoid activation function.

[0074] Next, the weight coefficient tensor S is scaled using an exponential function exp.

[0075]

[0076] At this point, the scaled tensor The importance coefficients of the extracted features can be used to obtain new hidden layer feature representations through Hadamard product.

[0077]

[0078] Then, the newly generated hidden layer features are input into a fully connected layer to predict the quality variable. Conduct supervised learning:

[0079]

[0080] W h and b h These represent the weights and biases of the fully connected layer, respectively.

[0081] The loss function in this supervised learning process can be expressed as:

[0082]

[0083] Where y(i) represents the true value of the i-th quality variable.

[0084] Finally, the cost function for the entire STAE is expressed as:

[0085] loss STAE =loss1 + loss2. (15)

[0086] Then, the entire network continuously optimizes the loss function through backpropagation, learning the optimal parameters of the model. This allows the network to learn features more effectively and explicitly express the importance of each feature.

[0087] 5) According to Figure 4 This method involves stacking individual spatiotemporal autoencoders to build a deep spatiotemporal autoencoder model, which is then trained and validated using a test dataset. First, the first spatiotemporal autoencoder is pre-trained to obtain hidden layer features H1. Then, the parameters of the first spatiotemporal autoencoder are frozen, and the first hidden layer H1 is used as the input to the next autoencoder. Similarly, by minimizing the input-output error, the hidden layer features H2 of the second spatiotemporal autoencoder are obtained and used as the input to the third spatiotemporal autoencoder. This process continues, with a series of spatiotemporal autoencoders stacked layer by layer to form a deep spatiotemporal autoencoder. Finally, the extracted spatiotemporal features are input into a fully connected layer to achieve online prediction of quality metrics.

[0088] The present invention will be further explained below with reference to specific examples.

[0089] (1) Description of the desulfurization process and determination of variables

[0090] This technology is mainly a dry HCl removal system developed for the waste incineration industry. The byproduct is sodium chloride, which is recycled and used as a raw material to produce soda ash. The SDS dry desulfurization spray technology involves uniformly spraying a high-efficiency desulfurizing agent (20-25μm) into the pipeline. The desulfurizing agent is thermally activated in the pipeline, and its specific surface area increases rapidly. It comes into full contact with the acidic flue gas and undergoes physical and chemical reactions. The acidic substances such as SO2 in the flue gas are absorbed and purified. The specific chemical reaction process is shown in equations (16) and (17). The desulfurization process is a dynamic time-varying process with complex mechanisms, numerous influencing factors, uncertainty, strong nonlinearity, large lag, and high coupling. It is difficult to establish an accurate mathematical model. Data-driven soft measurement methods will be an effective solution. Through mechanism analysis and expert knowledge, 14 auxiliary variables related to the sintering endpoint BTP were selected, as shown in Table 1.

[0091]

[0092] 2NaHCO3+SO3→Na2SO4+2CO2+H2O (17) Table 1 Input and Output Variables

[0093]

[0094]

[0095] (2) Dataset creation

[0096] For this process, 5000 samples were collected at equal intervals from the desulfurization plant. The first 3000 samples were used as the training set, 1000 as the validation set, and the remaining 1000 as the test set. Sampling was then performed using a sliding window with a window size of 10 to obtain different data segments for constructing the dataset. The model results were analyzed using the coefficient of determination R0. 2 The evaluation is measured by three metrics: root mean square error (RMSE) and mean absolute error (MAE). y represents the predicted value, and y represents the actual value.

[0097]

[0098]

[0099]

[0100] (3) Model building

[0101] Step 1: Offline modeling;

[0102] Step 1.1: Through analysis of the desulfurization mechanism, 15 key variables can be identified as input features for the model, such as inlet temperature, inlet pressure, and inlet benchmark flow rate. Then, data is read from the database in real time and preprocessed, including data filtering, smoothing, and normalization.

[0103] Step 1.2: Input the processed data into the first spatiotemporal autoencoder model. The resulting hidden layer features are used as the input to the second spatiotemporal autoencoder, and so on. By learning layer by layer, high-level semantic features of the original data are obtained. Finally, an online prediction is performed through a fully connected layer, which can achieve the prediction of sulfide concentration after desulfurization. The model parameters are then fine-tuned and its structure adjusted using the above three evaluation metrics until an accurate prediction model is obtained.

[0104] Step 2: Online detection;

[0105] Step 2.1: Real-time online data is read using sensors and a database to collect auxiliary variables, which are then preprocessed to construct a dataset. The established prediction model is then deployed within the sintering expert system to perform online predictions based on this real-time data.

[0106] Step 3: Model Update

[0107] Based on real-time data during the sintering process, the parameters of the encoding and decoding model are also adjusted in real time, continuously optimized and iterated, and step 1 is repeated to make the model have strong robustness.

[0108] (4) Model performance comparison

[0109] To compare the model established in this invention with the deep spatial-temporal autoencoder (S) model... 2 To compare the advantages and disadvantages of TAEs, the comparison models used here are two traditional machine learning models: Multilayer Perceptron (MLP) and Stacked Autoencoder (SAE), and Variable Weighted Autoencoder (VW-SAE).

[0110] Table 2 shows the prediction performance based on the deep spatiotemporal autoencoder (SAE) model. It can be seen that the traditional machine learning model MLP performs better in multi-step prediction. The SAE model, due to its ability to stack features to obtain high-level semantic features and its pre-trained weights, outperforms MLP. However, the original SAE cannot guarantee the relevance between the acquired features and the prediction target. Therefore, VW-SAE performs targeted learning by calculating the relevance between features and the prediction target, resulting in more effective features and thus better performance than SAE. The proposed SAE model... 2 The TAE model, which considers both temporal and spatial correlations, performs best, with a direct correlation between the actual and predicted values. Figure 5 As shown in the figure. The results indicate that this model provides desulfurization operators with sufficient time to adjust process parameters, which can improve product yield and quality, bringing significant economic benefits to enterprises.

[0111] Table 2 Comparison of Model Prediction Results

[0112]

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A soft measurement modeling method based on a depth spatiotemporal autoencoder, characterized in that, Includes the following steps: 1) Read relevant data from the real-time industrial process database and preprocess the data to construct the original dataset; assuming there are n samples in the dataset, where M auxiliary variables are easily measurable and one is the quality variable to be predicted, then the original dataset can be represented as follows: Where j represents the j-th sample, Represents auxiliary variables, The auxiliary variables are: inlet temperature, inlet pressure, inlet benchmark flow rate, inlet oxygen, inlet SO2 concentration, inlet dust concentration, inlet humidity, flue gas outlet temperature, flue gas outlet pressure, flue gas outlet flow rate, flue gas outlet O2 content, flue gas outlet dust concentration, sodium bicarbonate feed A, and sodium bicarbonate feed rate B; the mass variable to be predicted is the flue gas outlet converted SO2 concentration. X concentration; 2) The data is divided into segments using the sliding window technique to construct training, validation, and test sets. If the length of the sliding window segment is T, then the serialized data segment is represented as follows: ; 3) The autoencoder is unfolded into a chain structure in chronological order, and a memory module is embedded between two adjacent autoencoders to construct a temporal autoencoder. The memory module is used to store useful historical information and pass it to the next autoencoder. In step 3), for a given input sequence , in The hidden layer of the temporal autoencoder represents the sequence of auxiliary variables at time t as follows: （1） in, , These represent the weights and biases of the autoencoder at time t, respectively. It is a hidden layer dimensionality Represents the current state of memory. This represents the encoding function of a temporal autoencoder, used to compute hidden layer features. represents the hidden features obtained by the temporal autoencoder, and [] represents concatenation of vectors; 4) After obtaining the hidden layer features in the temporal autoencoder, the feature calibration module is connected to the encoding module of the temporal autoencoder, and the hidden layer features are input into the feature calibration module to quantify the importance between the extracted hidden layer features and the quality variables, thus building a spatiotemporal autoencoder; then the extracted spatiotemporal features are input into the fully connected layer to make real-time predictions of the quality variables. In step 4), a feature calibration module is designed to calculate the importance of hidden layer features and quality variables, so as to extract hidden layer features related to quality variables more efficiently. Suppose a batch of input variables are Let B represent the batch size. Then the output of the timing autoencoder is expressed as: The feature calibration module includes two operations. The first operation is to use one-dimensional global average pooling to... Compression is performed along the batch dimension, and the average value is obtained. The second operation is to... The input is fed into a two-layer neural network to adaptively learn the weights of each hidden layer feature, which represent the feature importance. Then, the label information is introduced into a spatiotemporal autoencoder network for supervised learning to capture the spatial relationship between hidden layer features and quality variables. 5) Stack individual spatiotemporal autoencoders to build a deep spatiotemporal autoencoder model, train the built deep spatiotemporal autoencoder model, and verify it with a test dataset to finally obtain a verified deep spatiotemporal autoencoder model.

2. The soft measurement modeling method based on a depth spatiotemporal autoencoder according to claim 1, characterized in that, The processing flow of the memory module in step 3) is as follows: The memory module contains a gating unit. An autoencoder is used to dynamically control how much historical information needs to be retained or forgotten, so that useful historical information can be passed to the next state. （2） Feature constructor This is used to construct a large feature vector, which is then used as input to the scoring function G. The expression is: （3） in, Represents the input from the previous moment. This represents the hidden layer features from the previous time step. The current input variables are represented by [], which indicates that these vectors are concatenated. It can be represented as: , , and These are the weight parameters that need to be learned; The gating unit uses a two-layer simple fully connected network for computation. and These represent the weights that need to be learned. and These represent the biases of the two network layers, respectively. and These represent the sigmoid activation function and the hyperbolic sine activation function, respectively. （5） The state of the memory module Expressed as: （6） Here, GRU stands for memory function, used to compute the previous time step. Historical information that needs to be memorized.

3. The soft measurement modeling method based on a depth spatiotemporal autoencoder according to claim 1, characterized in that, In step 5), after constructing the spatiotemporal autoencoder, multiple spatiotemporal autoencoders are stacked to form a depth spatiotemporal autoencoder, specifically as follows: First, the first spatiotemporal autoencoder is pre-trained to obtain hidden layer features. Then freeze the parameters of the first spatiotemporal autoencoder and set the first hidden layer... This serves as the input to the next autoencoder; similarly, by minimizing the error between the input and output, the hidden layer features of the second spatiotemporal autoencoder are obtained. The extracted spatiotemporal features are then used as input to the third spatiotemporal autoencoder. This process is repeated, with a series of spatiotemporal autoencoders stacked layer by layer to form a deep spatiotemporal autoencoder. Finally, the extracted spatiotemporal features are input to a fully connected layer to achieve online prediction of quality indicators.

Images