Converter steelmaking endpoint carbon temperature prediction method based on DRSupAE and application

Abstract

The application discloses a converter steelmaking endpoint carbon temperature prediction method based on a DRSupAE and application, a feature extraction model of a deep residual supervision self-encoder is formed by improving a supervision self-encoder pre-training mode and introducing a residual connection form to enable important information of each layer to participate in a deep feature learning process, the influence of important information in layer-by-layer learning is enhanced, and high-level features related to an endpoint carbon temperature are extracted, and meanwhile, based on a WSKL similarity criterion, a strategy of an adaptive real-time updating regression network is constructed to quickly adapt to production conditions of a to-be-tested sample, and the adaptive prediction performance of the model is improved.

Description

Technical Field

[0001] This invention belongs to the field of soft measurement technology, specifically relating to a method and application for predicting the final carbon temperature of converter steelmaking based on DRSupAE. Background Technology

[0002] Converter steelmaking, as a primary steelmaking technology, is characterized by high production efficiency and relatively low cost. The final carbon content and temperature of molten steel are crucial indicators in the steelmaking process, determining whether the steel meets quality standards. Therefore, accurately predicting the final carbon temperature is of great significance for improving the quality and efficiency of steel produced by converter steelmaking, minimizing production costs, and achieving energy conservation and emission reduction.

[0003] Currently, the main methods for obtaining the endpoint carbon content and temperature of molten steel are contact and non-contact measurement. The auxiliary lance is a typical contact measurement device. This method is costly due to the high price of the auxiliary lance and its long-term operation in a high-temperature corrosive environment; furthermore, this measurement method can only be used intermittently. Manual judgment of the endpoint is one of the most common non-contact methods, where workers judge whether a heat has reached the endpoint based on experience. However, this method is affected by factors such as the observer's subjective emotions and skill level, making accurate endpoint judgment difficult. Inspired by this method, some non-contact measurement models based on flame radiation images from the vessel opening have been proposed to predict the endpoint carbon content and temperature. However, because flame radiation spectrum images collected at the production site are easily interfered with by high-temperature objects such as ladles, accurate endpoint prediction becomes extremely difficult.

[0004] Due to the significant variations in the distribution of data in the converter steelmaking process and the severe nonlinearity between data points, traditional machine learning models, with their shallow architecture, are insufficient to describe the complex data structures of industrial processes, easily leading to unstable model performance or even model failure. Furthermore, in actual steelmaking production, the quality of raw materials from different batches varies greatly, and the collected process data is time-varying due to factors such as sensor aging, changes in operating conditions, and variations in raw material quality. Because the characteristics of the production process data are constantly changing, offline models cannot update in a timely manner according to actual operating conditions, resulting in significant deviations between predicted values ​​and actual endpoint carbon content and temperature, thus affecting the production quality of converter steelmaking.

[0005] Therefore, there is a current need for a method to address the highly nonlinear and complex distribution of process data in converter steelmaking endpoint prediction, in order to provide more accurate forecast results. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides a method for predicting the final carbon temperature in converter steelmaking based on DRSupAE. By improving the pre-training method of the supervised autoencoder and introducing residual connections, important information from each layer is incorporated into the deep feature learning process, forming a feature extraction model of a deep residual supervised autoencoder. This enhances the influence of important information in layer-by-layer learning and extracts high-level features related to the final carbon temperature. Simultaneously, based on the WSKL similarity criterion, an adaptive real-time update strategy for the regression network is constructed to quickly adapt to the production conditions of the test samples, thereby improving the adaptive prediction performance of the model.

[0007] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a method for predicting the final carbon temperature of converter steelmaking based on DRSupAE, characterized by comprising the following steps:

[0008] S1: Collect process data and corresponding endpoint carbon temperature data from steel plants, normalize them, and form a historical training database;

[0009] S2: Pre-train the DRSupAE feature extraction model on the historical training set of converter steelmaking to obtain high-level features of the training samples;

[0010] S3: Training a global regression network prediction model based on advanced features obtained from the DRSupAE feature extraction model;

[0011] S4: Input a new sample to be tested, normalize the sample to be tested, and obtain the high-level features of the sample to be tested through forward propagation of the trained DRSupAE feature extraction model.

[0012] S5: Calculate the similarity between the high-level features of the training samples and the high-level features of the test samples based on WSKL distance;

[0013] S6: Sort the training samples in descending order of similarity, select S samples as the update dataset, and use the update data to update the offline regression network online to obtain new model parameters;

[0014] S7: Use the updated model to predict the endpoint carbon content or temperature;

[0015] Furthermore, step S2 specifically includes the following steps:

[0016] S2.1: Input the historical training set. In the first RSupAE, use the original input data {x1,x2,...,x} N} and the corresponding labels {y1, y2, ... y N Improved supervised training can learn first-level feature data. and its network parameters;

[0017] S2.2: Before passing the extracted first-level features to the next autoencoder for deeper feature extraction, the input data is processed by the residual module defined in this invention to obtain residual block features. Then, it is fused with the first-level feature data in the form of residual connections to obtain... It is passed to the input layer of RSupAE 2 to learn secondary feature data.

[0018] S2.3: Perform supervised feature extraction of residuals based on S2.2, gradually extracting high-level features of training samples from the original input variables from low-level features;

[0019] Furthermore, step S5 specifically includes the following steps:

[0020] S5.1: Determine the weighting coefficients for the similarity criterion;

[0021] For the d-th high-level feature of the training samples The mutual information between the target variable vector y and the target variable vector y is calculated as follows:

[0022]

[0023] The MIC values ​​of each top-level feature and the target variable are obtained, and the weighting coefficient of the metric is calculated by the following formula.

[0024]

[0025] S5.2: High-level features of the sample to be tested and high-level features of training samples Perform data preprocessing as follows:

[0026]

[0027] S5.3: Based on the similarity criterion weighting coefficients obtained in S5.1, the high-level features of the sample to be tested are... and high-level features of training samples The WSKL distance between them is calculated as follows:

[0028]

[0029] The smaller the distance, the more similar the historical sample is to the sample to be tested;

[0030] To distinguish the similarity between different samples, a similarity metric function based on WSKL distance is designed as follows:

[0031]

[0032] It can be seen that the larger λ is, the more similar the two samples are. In the formula, σ is an adjustable parameter; the larger the value, the smaller the impact of WSKL distance on similarity.

[0033] Another object of the present invention is to provide a deep residual supervised autoencoder, characterized in that it comprises a three-layer network of an encoder and a decoder.

[0034] Furthermore, the deep residual supervised autoencoder training method includes simultaneously reconstructing the input and target variables during the decoding process. Since the endpoint carbon temperature is a single-dimensional target variable, the calculation of the loss function is improved during training, making the influence of labels on network learning controllable. Specifically, it includes the following steps:

[0035] S1. The encoder transforms the input variable x into a non-linear activation function, as shown in the following equation:

[0036] h = f(W) 1 x+b 1 ).

[0037] S2: The decoder will activate the hidden variable h through the function. The reconstructed input data is calculated. After activation function The reconstructed target variable was calculated. The following formula is used; the set of parameters to be optimized is expressed as follows:

[0038]

[0039] in, and These are the weight matrix and bias vector for reconstructing the input variables and predicting the target variables during the decoding process, respectively.

[0040] S3: Optimize network parameters using the backpropagation algorithm. The objective function is:

[0041]

[0042] In the formula, y i and d represents the true and predicted values ​​of the i-th data sample. x This represents the dimension of the input variable, and β represents the supervision influence parameter, which takes a value between 0 and 1.

[0043] The beneficial effects of this invention are as follows:

[0044] This invention, when extracting features from raw converter steelmaking process data, utilizes the constructed DRSupAE model to not only encode the target variable information into the hidden layer but also enhance the influence of relevant important information during layer-by-layer training. The feature extraction model can extract features related to the target variable layer by layer from the raw input data. Secondly, the adaptive real-time update regression network strategy constructed in this invention allows the prediction model to quickly adapt to current production conditions, solving the problem of decreased prediction performance caused by offline models' inability to update in a timely manner according to changes in actual operating conditions. Furthermore, since the selected samples are obtained based on the WSKL distance similarity criterion and have high similarity to the query samples, only a small number of iterations are needed for online adaptive model updates. The residual supervised feature extraction model constructed in this invention performs better in extracting features related to the endpoint carbon temperature. In addition, compared to offline static models, the real-time update mechanism of model parameters can adaptively update the model based on the characteristics of the query samples to improve prediction performance. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the present invention;

[0047] Figure 2 This is an algorithm diagram of the deep residual supervised autoencoder constructed in this invention;

[0048] Figure 3 This is a comparison between the predicted results of the final carbon temperature in converter steelmaking and the prediction results of the original SAE deep learning model. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example

[0051] See Figures 1 to 3 As shown, a method for predicting the final carbon temperature in converter steelmaking based on DRSupAE is characterized by the following steps:

[0052] Step 1: Collect process data and corresponding endpoint carbon temperature data collected during the actual production process of the steel plant, and normalize the data to form a historical training database;

[0053] Step 2: Pre-train the DRSupAE feature extraction model on the historical training set of converter steelmaking to obtain high-level features that can well represent the end point of steelmaking.

[0054] The novel deep learning model consists of multiple residual supervised autoencoders (RSupAEs). First, in each RSupAE, the features extracted from the hidden layers are obtained through supervised pre-training. Simultaneously, this invention defines a residual module (Res Model) to evaluate features in the shallow representation that are highly important to the target variable. Following the skip connection pattern in residual theory, important information from the low-level representation learned in the previous layer is fused with the features extracted in the current layer in the form of variable residual blocks, participating in the supervised feature learning process of the next layer. Figure 2 As shown, the specific flow of the feature learning algorithm is as follows:

[0055] Step 2.1: Input the historical training set. In the first RSupAE, use the original input data {x1,x2,...,x} N} and the corresponding labels {y1, y2, ... y N Improved supervised training can learn first-level feature data. and its network parameters.

[0056] The improved supervised autoencoder in this invention is a three-layer network comprising an encoder and a decoder. Unlike the original autoencoder, this invention reconstructs both the input and target variables simultaneously during the decoding process. Furthermore, since the endpoint carbon temperature is a single-dimensional target variable, the calculation of the loss function during training is improved, making the influence of labels on network learning controllable. Taking the first RSupAE as an example, the improved supervised autoencoder training method is as follows:

[0057] Step 2.1.1: The encoder transforms the input variable x into a nonlinear activation function into a hidden variable vector h, as shown in Equation (1).

[0058] h = f(W) 1 x+b 1 (1)

[0059] Step 2.1.2: The decoder passes the hidden variable h through the activation function The reconstructed input data is calculated. After activation function The reconstructed target variable was calculated. The predicted output value is shown in equation (4). The set of parameters to be optimized can be expressed as follows:

[0060]

[0061] in, and These are the weight matrix and bias vector for reconstructing the input variables and predicting the target variables during the decoding process, respectively.

[0062] Step 2.1.3: Optimize network parameters using the backpropagation algorithm. During training, to ensure the autoencoder learns features in a supervised manner, the parameter set θ1 needs to be optimized overall. The improved objective function is...

[0063]

[0064] In the formula, y i and d represents the true and predicted values ​​of the i-th data sample. x This represents the dimension of the input variable, and β represents the supervision effect parameter, which takes a value between 0 and 1.

[0065] Step 2.2: Before passing the extracted first-level features to the next autoencoder for deeper feature extraction, in order to continue utilizing the important information of the shallow features in the deeper feature learning process, the input data is processed by the residual module defined in this invention to obtain residual block features. Then, it is fused with the first-level feature data in the form of residual connections to obtain... It is passed to the input layer of RSupAE 2 to learn secondary feature data.

[0066] The process for determining the residual block characteristics using the residual module described in step 2.2 is defined as follows:

[0067] (1) Assuming there are N training samples, the feature data extracted by the (k-1)th RSupAE is represented as h k-1 The size is d h(k-1) ×N. First, the feature importance algorithm of the random forest is used to calculate the feature importance of each dimension of the extracted features in the previous layer to the target variable y, by FI(u (m) ) indicates that u (m) for h k-1 The m-th eigenvector. Specifically, for h k-1 For each decision tree in the constructed random forest, the OOB data error is calculated using the corresponding OOB data according to equation (12), denoted as OOBerror; then, the features u of the OOB data are randomly selected sequentially. (m) Adding noise interference, i.e., randomly changing the sample in feature u (m)The value above is used to calculate its OOB data error, denoted as OOBerror′; finally, u can be obtained according to equation (13). (m) The importance of features to the target variable y.

[0068]

[0069] FI(u (m) )=∑(OOB error' -OOBerror) / n tree (13)

[0070] In the formula, n tree This represents the number of decision trees; the parameters involved in the algorithm include n. tree and the number of leaf nodes n leaf It can be determined experimentally. From equation (13), it can be seen that if the characteristic u is given... (m) The addition of noise significantly increased the error in the OOB data, indicating that this feature is of high importance.

[0071] (2) Obtain shallow features h based on (1) k-1 Each dimension of the feature is scored for importance, and then sorted by importance from highest to lowest, as shown below. The important feature set that needs to be transmitted is initially determined by adaptive control threshold. As shown below:

[0072]

[0073] In the formula, δ f This is the importance control threshold, and its value is determined experimentally; n f The important feature dimension is initially determined; the numerator represents the top n after sorting. f The sum of the importance of each feature.

[0074] (3) Calculate the redundancy among the initially identified key features. The Maximum Information Coefficient (MIC) has been proven to effectively measure the degree of linear and nonlinear correlation between two variables; therefore, the redundancy between pairwise features is directly proportional to the MIC. Assume ui (a) and UI (b) Let MIC represent any two feature vectors in the feature set UI, and let MIC (ui) represent the degree of redundancy between them. (a) ,ui (b) The mutual information between any two features is represented as follows:

[0075]

[0076] In the formula, p(ui)(a) ,ui (b) p(ui) represents the joint probability distribution of two variables. (a) ) and p(ui b ) are respectively UI (a) and UI (b) The marginal probability distribution is determined. First, the scatter plot of these two variables is divided into a grid of ordered integer pairs (a, b), and the maximum MI value under the grid division is calculated, denoted as MI. * (ui (a) ,ui (b) Next, the maximum MI value is normalized to the [0,1] interval using equation (16). Finally, the first two steps are repeated for all grid pairs (a,b) to achieve the maximum grid resolution. MIC is the maximum value under different partitioning methods. The calculation method is shown in equation (17), where a·b < N. 0.6 ;

[0077]

[0078]

[0079] (4) After removing redundant information, determine the key features that ultimately need to be transmitted. If MIC(ui) (a) ,ui (b) )≥δ r And FI(ui) (a) )>FI(ui (b) ), where δ r The redundancy threshold is set as follows: if the above two conditions are met, then ui is considered... (b) This feature is redundant and will be deleted. The same redundancy removal operation is performed on the features within the feature set UI to obtain the final endpoint-related feature set, which is defined as the residual feature block, denoted by r. k Representation. Assume these features are in h k-1 The index in is but Its size is d rk ×N, which can be represented as r in a mathematical model. k =Res(h k-1 ).

[0080] (5) The residual block features obtained from the previous layer are fused with the features extracted by the autoencoder in this layer, and used as the input to the next layer encoder to further extract relevant information. That is, the features h of the k-th level... k An additional residual connection is included before passing the previous level features h to the next layer of the autoencoder for learning. k-1 The residual feature block r is obtained after processing by the Res Module as defined in Definition 1. kThen, it connects with h in the form of a skip connection. k The fusion is performed, and the resulting features are denoted as... Mathematically, it can be described as

[0081]

[0082] in, The size is (d) hk +d rk )×N. Figure 2 The detailed diagram of the Res Module in the image illustrates the specific process.

[0083] Step 2.3: Perform supervised feature extraction of residuals according to Step 2.2. Similarly, deep features can be obtained by pre-training layer by layer in a similar manner.

[0084] Suppose the feature data extracted from RSupAE k (k=2,…,K-1) is Residual block feature r k It is obtained by processing the (k-1)th level feature data through the Res Module, and the feature after the fusion of the two parts is: It will be used to extract feature data at level k+1. Therefore, RSupAE(k+1) will reconstruct the input previous-stage fused feature data. and target variable To obtain the network parameters θ k+1 The network is trained by minimizing the objective function shown in equation (20):

[0085]

[0086] Following the above training process, the original input variables gradually extract high-level features from low-level features, with the top-level feature h... K This refers to the high-level feature representation learned by the RSupAE model.

[0087] Step 3: To build a soft measurement model, a global regression network prediction model is trained based on the high-level features obtained from the feature extraction model. This model inherits the advantages of representing complex, hierarchical, and nonlinear features in the data. The specific process is as follows:

[0088] Top-level features of the DRSupAE model As input to a two-layer neural network, the predicted output is obtained through forward propagation. like:

[0089]

[0090] In the formula, (W o ,b o) represents the weights and bias parameters of the regression network. To obtain the model parameters, the model is trained by minimizing the loss function shown in equation (28) using the backpropagation algorithm.

[0091]

[0092] After training the regression network offline, its model parameters (W) o ,b o These parameters are retained as initial parameters for the model during the update phase.

[0093] Step 4: Through the offline training phase, high-level features that can well represent the steelmaking endpoint are obtained. When new test samples arrive, the features of the test samples undergo the same normalization process as the training data. The processed test samples are represented by x. t This indicates that it serves as the input variable for DRSupAE, utilizing the trained parameters. Top-level feature data is obtained through forward propagation.

[0094] Step 5: Calculate high-level features of training samples High-level features of the sample to be tested The similarity between them is calculated based on WSKL distance, and the similarity criterion calculation process is as follows:

[0095] Step 5.1: First, determine the weighting coefficients of the similarity criterion. Considering that MIC can effectively uncover linear and nonlinear relationships between variables, the weights are determined by the MIC of the top-level features and the target variable.

[0096] For the d-th high-level feature of the training samples The mutual information between the target variable vector y and the target variable vector y is calculated as follows:

[0097]

[0098] Based on the calculation process shown in equations (16)-(17), the MIC values ​​of each top-level feature and the target variable can be obtained. The weighting coefficient of the metric can be obtained by calculating the MIC ratio using equation (23).

[0099] Then w = (w1, w2, ..., w D ) represents the weight of the feature measure. The larger the weight, the greater the correlation with the target variable and the greater its contribution to the calculation of similarity.

[0100] Step 5.2: Sample to be tested and historical samples Perform the data preprocessing shown below for subsequent similarity calculation.

[0101]

[0102] Step 5.3: Based on the similarity criterion weighting coefficients obtained in Step 5.1, the sample to be tested... Compared with historical training samples The WSKL distance between them is calculated as follows:

[0103]

[0104] The smaller the distance, the more similar the historical sample is to the sample to be tested. To distinguish the similarity between different samples, a similarity metric function based on WSKL distance is designed as follows:

[0105]

[0106] It can be seen that the larger λ is, the more similar the two samples are. In the formula, σ is an adjustable parameter; the larger the value, the smaller the impact of WSKL distance on similarity. It is generally set to 1.

[0107] Step 6: Sort the training samples in descending order of similarity, select S samples as the update dataset, and use the update data to update the offline regression network online to obtain new model parameters. The specific steps are as follows:

[0108] Obtain the similarity {λ1,λ2,...,λ} N After that, the samples are sorted in descending order of similarity, and the S most similar samples are selected as the updated dataset. Assuming the indices of these related samples in the historical dataset are {e1,...,es,...,eS}, the online updated dataset UD is represented as follows:

[0109]

[0110] JITRN aims to adaptively update a globally built offline model online using an updated dataset. The loss function during the update phase is expressed as follows:

[0111]

[0112] The updated network model parameters are obtained by performing forward and backward propagation sequentially. JITRN enables rapid online model updates to adapt to the operating conditions of the test samples.

[0113] Step 7: Finally, use the updated model to predict the endpoint carbon content or temperature. When the next sample arrives, process it using the above method and obtain a new updated sample set for real-time updates and predictions.

[0114] Figure 3The results are simulations of real converter steelmaking processes, presented as scatter plots of the actual and predicted outputs for the endpoint carbon content and temperature. The samples are arranged in order of label size, with each data point representing a test sample. A comparison with SAE (Search Engine Engineering) shows that this patent delivers better prediction performance.

[0115] In summary, when extracting features from raw converter steelmaking process data, this invention, through the constructed DRSupAE model, not only encodes the information of the target variable into the hidden layer but also enhances the influence of relevant important information during layer-by-layer training. The feature extraction model can extract features related to the target variable layer by layer from the raw input data. Secondly, the adaptive real-time update regression network strategy constructed in this invention allows the prediction model to quickly adapt to current production conditions, solving the problem of decreased prediction performance caused by the offline model's inability to update in a timely manner according to changes in actual working conditions. Furthermore, since the selected samples are obtained based on the WSKL distance similarity criterion and have high similarity to the query samples, only a small number of iterations are needed for online adaptive model updates. The residual supervised feature extraction model constructed in this invention performs better in extracting features related to the endpoint carbon temperature. In addition, compared with the offline static model, the real-time update mechanism of model parameters can adaptively update the model according to the characteristics of the query samples to improve prediction performance.

Claims

1. A method for predicting the final carbon temperature in converter steelmaking based on DRSupAE, characterized in that, Specifically, the following steps are included: S1: Collect process data and corresponding endpoint carbon temperature data from steel plants, normalize them, and form a historical training database; S2: Pre-train the DRSupAE feature extraction model on the historical training set of converter steelmaking to obtain high-level features of the training samples; S3: Training a global regression network prediction model based on advanced features obtained from the DRSupAE feature extraction model; S4: Input a new sample to be tested, normalize the sample to be tested, and obtain the high-level features of the sample to be tested through forward propagation of the trained DRSupAE feature extraction model. S5: Calculate the similarity between the high-level features of the training samples and the high-level features of the query samples based on WSKL distance; S6: Sort the training samples in descending order of similarity, select S samples as the update dataset, and use the update data to update the offline regression network online to obtain new model parameters; S7: Use the updated model to predict the endpoint carbon content or temperature; Step S2 specifically includes the following steps: S2.1: Input the historical training set, and in the first RSupAE, use the original input data. and corresponding tags Improved supervised training can learn first-level feature data. and its network parameters; S2.2: Before passing the extracted first-level features to the next autoencoder for deeper feature extraction, the input data is processed by the improved residual module to obtain residual block features. Then, it is fused with the first-level feature data in the form of residual connections to obtain... This data is then passed to the input layer of RSupAE2 to learn secondary feature data. ; S2.3: Perform supervised feature extraction of residuals based on S2.2, gradually extracting high-level features of training samples from the original input variables from low-level features; Step S5 specifically includes the following steps: S5.1: Determine the weighting coefficients for the similarity criterion; For the d-th high-level feature of the training samples With the target variable vector The mutual information is calculated as follows: The MIC values ​​of each top-level feature and the target variable are obtained, and the weighting coefficient of the metric is calculated by the following formula. , ； S5.2: High-level features of the sample to be tested and high-level features of training samples Perform data preprocessing as follows: ； S5.3: Based on the similarity criterion weighting coefficients obtained in S5.1, the high-level features of the sample to be tested are... and high-level features of training samples The WSKL distance between them is calculated as follows: ； The similarity measurement function based on WSKL distance is as follows: 。 2. The method for predicting the final carbon temperature of converter steelmaking based on DRSupAE according to claim 1, characterized in that, The deep residual supervised autoencoder consists of a three-layer network of encoder and decoder.

3. The depth residual supervised autoencoder according to claim 2, characterized in that, This depth residual supervised autoencoder was applied to predict the carbon temperature at the end of converter steelmaking.

4. The deep residual supervised autoencoder according to claim 2 is applied to the feature extraction of carbon temperature prediction at the end point of converter steelmaking.

5. A deep residual supervised autoencoder training method for feature extraction of carbon temperature prediction at the end point of converter steelmaking, as described in any one of claims 1 to 4, characterized in that, The decoding process reconstructs both the input and target variables simultaneously. Furthermore, since the endpoint carbon temperature is a single-dimensional target variable, the calculation of the loss function is improved during training, making the influence of labels on network learning controllable. Specifically, the steps include: S1. The encoder transforms the input variable x into a non-linear activation function, as shown in the following equation: S2: The decoder will activate the hidden variable h through the function. The reconstructed input data is calculated. After activation function The reconstructed target variable was calculated. The following formula is used; the set of parameters to be optimized is expressed as follows: ; in, and These are the weight matrix and bias vector for reconstructing the input variables and predicting the target variables during the decoding process, respectively. S3: Optimize network parameters using the backpropagation algorithm. The objective function is: In the formula, and These are the true and predicted values ​​of the i-th data sample. Indicates the dimension of the input variable. This represents the monitoring effect parameter, with a value between 0 and 1.

Images