Industrial quality variable prediction method based on parallel gating enhanced auto-encoder
Through parallel gated enhanced autoencoder model, the problem of long training time of autoencoder and insufficient feature correlation is solved, and efficient industrial quality variable prediction is achieved.
Patent Information
- Application Number
- CN202510523966.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-24
- Publication Date
- 2025-08-01
AI Technical Summary
Existing autoencoder models train for a long time in industrial mass variable prediction and do not fully consider the correlation between auxiliary variables and mass variables.
The parallel gated enhanced autoencoder (PGEAE) model is adopted, and the correlation between auxiliary variables and mass variables is embedded in the model by adding hidden layer parallel computing and gradient synchronization technology, combining data parallel computing technology.
It significantly shortens the feature extraction time, improves prediction accuracy, reduces the time cost of model training, and improves prediction effect.
Smart Images
Figure CN120408556A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial processes, and in particular, to an industrial quality variable prediction method based on a parallel gated enhanced autoencoder (PGEAE). Background Art
[0002] In modern process industrial processes, the monitoring of quality variables is crucial, which can not only ensure the safety of industrial processes but also provide effective control and optimization means. However, traditional measurement technologies are subject to many limitations such as complex production environments, high measurement costs, and low measurement frequencies. With the wide application of distributed control systems, data in industrial processes are systematically collected and saved, providing a rich data basis for data-driven models. These historical data enable the models to extract effective information, thereby improving the accuracy of prediction and the robustness of the models. Therefore, data-driven soft sensing technologies have been widely applied. It predicts quality variables that are difficult to measure or cannot be measured in real time in actual industrial production processes by using easily collectable auxiliary variables and combining data-driven methods. This technology breaks through the limitations of traditional measurements and effectively enhances the monitoring ability and efficiency of industrial processes. Especially by applying advanced machine learning algorithms such as artificial neural networks, support vector regression, and random forest, data-driven models can better cope with complex industrial process changes and achieve real-time estimation of quality variables and process optimization.
[0003] An autoencoder is a deep learning model that performs non-linear dimensionality reduction and feature extraction on data through a neural network. Its core idea is to compress high-dimensional input data into a low-dimensional latent representation and then attempt to reconstruct the original data through a decoder. In this way, the autoencoder can extract hidden features from complex data. Therefore, the industrial quality variable prediction method based on the autoencoder model has been widely applied. However, existing autoencoder models and their improved models for industrial quality variable prediction have defects. For example, the autoencoder model (Autoencoder, hereinafter referred to as AE), since AE is a shallow model, it is limited by the expression ability of the hidden layer and is difficult to capture complex non-linear features in high-dimensional data. The deep autoencoder model (Deep Autoencoder, hereinafter referred to as DAE), DAE has more hidden layers and belongs to a deep model, but DAE only focuses on extracting features of auxiliary variables, and the features it extracts have a weak correlation with quality variables. Moreover, existing models all require a long time to extract features. The parallel gated enhanced autoencoder proposed by the present invention fully considers the correlation between auxiliary variables and quality variables in the extracted features and significantly shortens the feature extraction time. Summary of the Invention
[0004] Aiming at the problems in the prior art that the autoencoder training time is long and the extracted features do not consider the correlation between auxiliary variables and quality variables, the present invention proposes an industrial quality variable prediction method based on a parallel gated enhanced autoencoder.
[0005] The present invention specifically comprises the following steps:
[0006] Step 1: Collect historical data from the industrial process and construct a training data set. The training data set includes auxiliary variables x and corresponding key quality variables y. Auxiliary variables x are multidimensional variables, x = [x(1), x(2), ..., x(u)] T ∈R u×1 , where u is the dimension of the auxiliary variable input, and x(1), x(2), and x(u) represent auxiliary variables of different dimensions.
[0007] Step 2: Normalize the collected training data set and use the Min-Max normalization method to transform the values of the training data set to between 0 and 1.
[0008] Step 3: Build a parallel gated enhanced autoencoder PGEAE:
[0009] On the basis of the autoencoder, the number of hidden layers is increased by linear function mapping, and the auxiliary variables are concatenated with the first hidden layer as the input of the second hidden layer. At the same time, the quality variables are embedded into the decoder. The parameter set to be optimized for the parallel gated enhanced autoencoder is Its loss function Where N is the total number of samples, x n is the input sample, Represented as x n Reconstruction of y n is a sample of key quality variables, Represented as y n Reconstruction.
[0010] Data parallel computing technology replicates the model across B compute nodes, then divides the dataset into B subsets and assigns them to different compute nodes to achieve data parallelism. Each node independently and simultaneously performs forward and backward propagation calculations on its assigned data subset. In each training round, each node independently calculates gradients. Before updating parameters, gradient synchronization is performed on each node to ensure that all model replicas have consistent parameters after each iteration.
[0011] Global gradient after gradient synchronization In the formula is the local gradient of the bth computing node, and B is the number of computing nodes.
[0012] After training is completed, each hidden layer in the encoder is respectively input into the gating strategy, and the gating value and alternative output of each hidden layer are calculated respectively, and finally the predicted value of the key quality variable is obtained. The expression of the gating strategy is as follows: In the formula is the weight coefficient matrix of h j , is the weight coefficient matrix of h j , is the bias of h j ; σ and tanh respectively represent the sigmoid function and the tanh function; g represents the gating value of the j-th hidden layer, and y j represents the alternative output of the j-th hidden layer; J represents the number of hidden layers; j represents the final output. Its loss function j In the formula, N is the total number of samples, and y is the true value of the key quality variable, is the predicted value of the key quality variable; n ; is the predicted value of the key quality variable;
[0013] The training is accelerated through the data parallel computing technology, and the global gradient after gradient synchronization In the formula is is the local gradient of the b-th computing node, and B is the number of computing nodes.
[0014] Step 4: Train the parallel gating enhanced autoencoder with the normalized training data set obtained in Step 2.
[0015] Step 5: After processing the auxiliary variable x collected in real time through the Min-Max normalization method, input it into the trained parallel gating enhanced autoencoder to obtain the predicted value of the key quality variable y.
[0016] The present invention also proposes an electronic device for executing the above method, including a memory and a processor. The memory of the electronic device stores executable code, and the processor executes the above industrial quality variable prediction method based on a parallel gating enhanced autoencoder.
[0017] In the pre-training stage of the model of the method of the present invention, the auxiliary variable is embedded into the hidden layer, and at the same time the quality variable is embedded into the decoder layer. Then in the fine-tuning stage, through the gating strategy, regression prediction is performed using the abstract representation of each hidden layer; and the data parallel computing technology is used, effectively shortening the time required for model training. Description of the Drawings
[0018] Figure 1It is the flowchart of the method described in the present invention;
[0019] Figure 2 It is the structural schematic diagram of a single-layer autoencoder;
[0020] Figure 3 It is the structural schematic diagram of the parallel gated enhanced autoencoder of the present invention;
[0021] Figure 4 It is the working flowchart of the primary reformer in the application example. Specific embodiments
[0022] As Figure 1 shown, an industrial quality variable prediction method based on a parallel gated enhanced autoencoder specifically includes the following steps:
[0023] Step 1: Collect historical data in the industrial process to construct a training data set. The training data set includes auxiliary variables x and corresponding key quality variables y. The auxiliary variable x is a multi-dimensional variable, x = [x(1), x(2), …, x(u)] T ∈R u×1 , where u is the dimension of the auxiliary variable input, and x(1), x(2), x(u) respectively represent auxiliary variables in different dimensions.
[0024] Step 2: Normalize the collected training data set using the Min-Max normalization method to transform the values of the training data set to between 0 and 1.
[0025] Step 3: Construct a parallel gated enhanced autoencoder PGEAE.
[0026] As Figure 2 shown, a single-layer autoencoder consists of an encoder and a decoder. The encoder maps the data x in the input layer to the hidden layer through a non-linear function, and the decoder reconstructs the input data through another non-linear function and finally transfers the result to the reconstruction layer. The specific formula is as follows: h = f e (W·x + b); where h = [h(1), h(2), …, h(m)] T ∈R m×1 is the feature of the hidden layer, m is the dimension of h; is the feature of the reconstruction layer, W is the weight coefficient matrix of x, is the weight coefficient matrix of h, b is the bias of x, is the bias of h; f e and f d represent non-linear functions.
[0027] As Figure 3As shown, the parallel gated enhanced autoencoder of the present invention increases the number of hidden layers by means of linear function mapping on the basis of the autoencoder, and splices the auxiliary variable and the first hidden layer as the input of the second hidden layer, while embedding the quality variable into the decoder. The set of parameters to be optimized for the parallel gated enhanced autoencoder is Its loss function where N is the total number of samples, x n is the input sample, is denoted as the reconstruction of x n , y n is the key quality variable sample, is denoted as the reconstruction of y n .
[0028] The data parallel computing technology copies the replicas of the model to B computing nodes, and then divides the data set into B subsets and assigns them to different computing nodes to achieve data parallelism. Each node independently performs forward and backward propagation calculations on the data subset assigned to it. In each round of training, each node independently calculates the gradient. Before updating the parameters, the gradients of each node are synchronized to ensure that the parameters of all model replicas are consistent after each iteration.
[0029] The global gradient after gradient synchronization where is the local gradient of the b-th computing node, and B is the number of computing nodes.
[0030] After training is completed, each hidden layer in the encoder is respectively input into the gating strategy, the gating value and the alternative output of each hidden layer are calculated respectively, and finally the predicted value of the key quality variable is obtained. The expression of the gating strategy is as follows: where is the weight coefficient matrix of h j , is the weight coefficient matrix of h j , is the bias of h j , is the bias of h j ; σ and tanh respectively represent the sigmoid function and the tanh function; g j represents the gating value of the j-th hidden layer, y j represents the alternative output of the j-th hidden layer; J represents the number of hidden layers; represents the final output. Its loss function where N is the total number of samples, y n is the true value of the key quality variable, is the predicted value of the key quality variable;
[0031] Accelerate training through data parallel computing technology, and the global gradient after gradient synchronization In the formula is is the local gradient of the b-th computing node, and B is the number of computing nodes.
[0032] Step 4: Train the parallel gated enhanced autoencoder using the normalized training dataset obtained in Step 2.
[0033] Step 5: After processing the real-time collected auxiliary variable x through the Min-Max normalization method, input it into the trained parallel gated enhanced autoencoder to obtain the predicted value of the key quality variable y.
[0034] The following takes the prediction of the oxygen content in the primary reformer during the hydrogen production process as an example to demonstrate the performance of the industrial quality variable prediction of the present invention.
[0035] The primary reformer is an important device in the hydrogen production process. Hydrogen is converted from the raw material methane, methane is a key component for synthesizing ammonia, and ammonia is the main component for synthesizing urea. The methane conversion device mainly includes a pre-reformer, a primary reformer device, and a secondary reformer device. The primary reformer device is the main reaction device, and optimizing the hydrogen production process control of this device is of great significance. The working flow chart of this device is as Figure 4 shown. According to the reaction mechanism, the temperature inside the tower should be 580°C, and the stability of the temperature inside the tower is the key factor to ensure hydrogen production by the primary reformer device. The temperature inside the tower is related to gas combustion. In order to make the gas combustion stable, the oxygen content in the furnace is controlled to vary within a set range. In the actual process, the oxygen content is measured by a mass spectrometer, and this measurement method is relatively expensive. Therefore, soft sensor technology can be used to reduce costs.
[0036] In this example, the oxygen content inside the furnace is selected as the key quality variable y to be predicted; Table 1 shows the 13 selected auxiliary variables x, including temperature, pressure, and flow rate. In this example, 2000 samples are randomly selected as the training dataset, and another 500 samples are randomly selected as the test set.
[0037] Table 1 Introduction to auxiliary variables [[ID=z8]]
[0038] Label Description <![CDATA[x1]]> Flow rate of combustion natural gas <![CDATA[x2]]> Flow rate of tail gas fuel <![CDATA[x3]]> Pressure of tail gas fuel <![CDATA[x4]]> Pressure of gas in the furnace <![CDATA[x5]]> Temperature of tail gas fuel <![CDATA[x6]]> Temperature of combustion natural gas <![CDATA[x7]]> Temperature of process gas at the inlet of the primary reformer <![CDATA[x8]]> Temperature of gas in the upper left corner of the furnace <![CDATA[x9]]> Temperature of gas in the upper right corner of the furnace <![CDATA[x 10 > Temperature of mixed fuel gas at the top of the furnace <![CDATA[x 11 > Temperature of converted gas at the left outlet of the primary reformer <![CDATA[x 12 > Temperature of converted gas at the right outlet of the primary reformer <![CDATA[x 13 > Temperature of converted gas at the outlet of the primary reformer y Oxygen content at the top of the furnace
[0039] Compare the quality variable prediction results of AE, DAE, gated enhanced autoencoder (hereinafter referred to as GEAE), and the PGEAE model proposed by the present invention. GEAE and PGEAE have the same model structure, and GEAE does not use data parallel technology. The network structure of AE is [13 6 13], and the network structures of DAE, GEAE, and PGEAE are all [13 10 8 6 8 10 13], where 13 represents the dimensions of the input layer and the output layer, and 10, 8, and 6 represent the dimensions of the hidden layers. Table 2 lists the prediction results of the test set under the above four methods.
[0040] Table 2 Prediction Results of Four Autoencoder Models (Primary Reformer Unit)
[0041] Model AE DAE GEAE PGEAE RMSE 0.110 0.108 0.093 0.093 <![CDATA[R 2 > 0.689 0.702 0.776 0.778 Training time (s) 15.213 19.991 24.727 13.935
[0042] It can be seen from Table 2 that the PGEAE model has a smaller root mean square error (RMSE) and a larger R 2 compared to the AE and DAE models. The PGEAE model has similar prediction performance compared to the GEAE model, but the PGEAE model requires less training time. The PGEAE industrial quality variable prediction method proposed by the present invention has the best prediction effect.
Claims
1. An industrial quality variable prediction method based on a parallel gated enhanced autoencoder, characterized in that: Specifically, it includes the following steps: Step 1: Collect historical data in the industrial process to construct a training data set; the training data set includes auxiliary variables x and corresponding key quality variables y, where the auxiliary variable x is a multi-dimensional variable, x = [x(1), x(2), …, x(u)] T ∈R u×1 , where u is the dimension of the auxiliary variable input, and x(1), x(2), x(u) represent auxiliary variables in different dimensions respectively; Step 2: Normalize the collected training data set by using the Min - Max normalization method to transform the values of the training data set to between 0 and 1; Step 3: Construct a parallel gated enhanced autoencoder PGEAE: Based on the autoencoder, the number of hidden layers is increased by means of linear function mapping, and the auxiliary variable is concatenated with the first hidden layer as the input of the second hidden layer. At the same time, the quality variable is embedded into the decoder. The set of parameters to be optimized for the parallel gated enhanced autoencoder is Its loss function In the formula, N is the total number of samples, and x n is the input sample, is denoted as x n 's reconstruction, and y n is the key quality variable sample, is denoted as y n 's reconstruction; The data parallel computing technology copies the replicas of the model to B computing nodes, then divides the data set into B subsets and assigns them to different computing nodes to achieve data parallelism; each node independently performs forward and backward propagation calculations on its assigned data subset; in each round of training, each node independently calculates the gradient; before updating the parameters, the gradients of each node are synchronized to ensure that all model replica parameters are consistent after each iteration; Global gradient after gradient synchronization where is the local gradient of the b-th computing node, and B is the number of computing nodes; After the training is completed, each hidden layer in the encoder is input into the gating strategy respectively, and the gating value and alternative output of each hidden layer are calculated respectively, and finally the predicted value of the key quality variable is obtained; the expression of the gating strategy is as follows: In the formula is the weight coefficient matrix of h j , is the weight coefficient matrix of h j , is the bias of h j ; σ and tanh represent the sigmoid function and the tanh function respectively; g represents the gating value of the j-th hidden layer, and y j represents the alternative output of the j-th hidden layer; J represents the number of hidden layers; j j represents the final output; its loss function In the formula, N is the total number of samples, and y is the true value of the key quality variable, n and is the predicted value of the key quality variable; Accelerate training through data parallel computing technology, and the global gradient after gradient synchronization In the formula is is the local gradient of the b-th computing node, and B is the number of computing nodes; Step 4: Train the parallel gated enhanced autoencoder with the normalized training data set obtained in Step 2; Step 5: After processing the real - time collected auxiliary variable x by the Min - Max normalization method, input it into the trained parallel gated enhanced autoencoder to obtain the predicted value of the key quality variable y.
2. An electronic device for performing the method according to claim 1, comprising a memory and a processor, characterized in that: The memory of the electronic device stores executable code, and the processor executes the industrial quality variable prediction method based on the parallel gated enhanced autoencoder as claimed in claim 1.