An industrial soft-sensing method based on variable-rate complementary cyclic unit network

Abstract

The application discloses an industrial soft measurement method based on a variable-rate complementary cycle unit network and belongs to the technical field of soft measurement. The method comprises the following steps: a rate self-adaptive adjusting door is designed in a GRU architecture, nonlinear exponential normalization transformation is performed on the rate self-adaptive adjusting door and an updating door, a variable-rate forgetting door and a variable-rate input door are formed, and the two doors are embedded into a GRU hidden state updating function to construct an SVCRU neural network prediction model. The model adopts a nonlinear variable-rate adjusting mechanism to replace inherent linear constant-rate adjusting constraints of the GRU, accelerates the model training convergence speed, and improves the prediction precision.

Description

Technical Field

[0001] This invention relates to an industrial soft measurement method based on a variable-rate complementary cyclic unit network, belonging to the field of soft measurement technology. Background Technology

[0002] In complex industrial processes, it is necessary to measure certain key target variables that are difficult to measure directly or in real time but are closely related to product quality and production safety in real time to meet industrial process control requirements. Soft measurement technology can use easily measurable process variables in industrial processes to make real-time and effective estimates of difficult-to-measure target variables, and it is widely used in process manufacturing and process control fields.

[0003] With the continuous development of industrial intelligence and the widespread application of distributed control systems, data-driven soft measurement technology has become a research focus in this field. Typical data-driven soft measurement methods include principal component analysis, support vector regression, and artificial neural networks. Among them, artificial neural networks have attracted much attention due to their excellent nonlinear modeling capabilities. However, actual industrial process data usually exhibits strong nonlinearity, dynamism, and temporal correlation as time series. Traditional static feedforward neural networks, due to the inherent limitations of their structure, struggle to effectively capture the dynamic temporal characteristics of industrial process data. Therefore, the soft measurement models established by them often fail to meet actual production needs.

[0004] In recent years, gated recurrent neural networks (RNNs), by introducing temporal feedback mechanisms and gated unit structures, have not only effectively described the dynamic temporal characteristics of data but also alleviated the gradient vanishing and exploding problems inherent in traditional RNNs. Among them, the gated recurrent unit (GRU) network integrates information by linearly interpolating the hidden state function through updating gates. Its single-gated adjustment mechanism effectively reduces model complexity, attracting widespread attention from researchers in the field. However, the linear constant-rate adjustment constraint imposed by the GRU when calculating its hidden state hinders the full transmission of internal information, thus affecting its ability to process large-scale time series data. To address this issue, Xie et al. proposed a novel soft sensor modeling method based on a two-stream λ-gated recurrent unit network with soft sensor application in their paper "Data-driven modeling based on two-stream λ-gated recurrent unit network with soft sensor application." This method, by designing a dual-stream structure to handle time and dynamic causal relationships separately, can more effectively extract dynamic features from the data, thereby improving the model's prediction accuracy and generalization ability. In their paper "A soft sensor model based on long & short-term memory dual pathways convolutional gated recurrent unit network for predicting cement specific surface area," Sun et al. first combined dual-channel LSTM and GRU, taking into account both long and short-term features and spatiotemporal correlation. By combining long short-term memory and local feature extraction capabilities, they significantly improved prediction accuracy. The methods mentioned above all introduce hyperparameters into the GRU hidden state to change its inherent linear constant-rate adjustment constraint, effectively enhancing the model's feature extraction capability. However, directly introducing additional hyperparameters not only complicates parameter tuning but also increases the risk of overfitting. Furthermore, inappropriate hyperparameter settings can degrade model prediction performance, leading to excessive model dependence on specific hyperparameter settings. Therefore, achieving adaptive adjustment of the rate parameter while changing the inherent linear adjustment constraint of the GRU hidden state has become an urgent scientific challenge. Summary of the Invention

[0005] To address the issue that existing soft measurement techniques based on GRU networks fail to adequately consider the inherent linear constant-rate adjustment constraint of the hidden state, which hinders effective information transmission and thus reduces prediction performance, this invention provides an industrial soft measurement method based on a variable-rate complementary cyclic unit (SVCRU) network and applies it to SO2 concentration prediction in the flue gas desulfurization process of a coal-fired power plant. This method designs a rate-adaptive adjustment gate to control the information transmission rate. The rate-adaptive adjustment gate and the update gate are then subjected to nonlinear exponential normalization transformation to form a variable-rate forget gate and a variable-rate input gate, which are embedded into the GRU hidden state update function to construct an SVCRU neural network model. The nonlinear variable-rate adjustment mechanism replaces the inherent linear constant-rate adjustment constraint of the GRU, accelerating model training convergence and improving prediction accuracy.

[0006] An industrial soft sensing method based on variable-rate complementary cyclic cell networks, the method comprising:

[0007] Step 1: Obtain historical data of process variables and target variables in the industrial production / processing process as a training sample database; and preprocess the training sample database to obtain the training dataset and test dataset.

[0008] Step 2: Construct the initial GRU network architecture consisting of gate controllers such as reset gates, update gates, and candidate hidden states;

[0009] Step 3: Based on the GRU's use of an update gate to linearly interpolate its candidate hidden states to achieve information fusion, design the general expression for the SVCRU hidden state update function;

[0010] Step 4: Design a new rate adaptive adjustment gate st, and combine it with the update gate from Step 2 to obtain the SVCRU network. By performing a nonlinear normalized exponential transform, a variable rate input gate is formed. and variable rate forget gate

[0011] Step 5: Input the variable rate gate from Step 4 and variable rate forget gate By embedding the general expression of the SVCRU candidate hidden state update function in step 3, the updated SVCRU candidate hidden state update function expression is obtained, and thus the SVCRU network is obtained.

[0012] Step 6: Combine the SVCRU network obtained in Step 5 with a multilayer perceptron to construct an SVCRU neural network prediction model;

[0013] Step 7: Using historical data from the industrial production / processing process, train the SVCRU neural network prediction model constructed in Step 6 to obtain the trained SVCRU neural network prediction model.

[0014] Step 8: Obtain real-time data of process variables in the industrial production / processing process, input them into the trained SVCRU neural network prediction model obtained in Step 7, and obtain the predicted value of the target variable in the process.

[0015] Furthermore, the industrial production / processing procedures in step 1 include desulfurization processes in coal-fired power plants, butane removal tower processes, fermentation processes, etc.

[0016] Furthermore, the preprocessing in step 1 includes handling missing values, outlier handling, and data standardization;

[0017] Furthermore, in step 1, the ratio of the training dataset to the test dataset is 8:2;

[0018] Furthermore, in the initial GRU network architecture constructed in step 2, the expressions for each gate controller of the GRU are as follows:

[0019]

[0020] Where, r t k, and Let be the reset gate, update gate, and candidate hidden state outputs in the topological dimension at time t, respectively; σ(·) and tanh(·) are the sigmoid activation function and hyperbolic tangent activation function, respectively; and k represents the topological output dimension. This represents the current input at the i-th input node. This represents the hidden weight vector of the reset gate at the i-th input node. This represents the hidden state output of the j-th hidden node at time t-1. The network input weight vector representing the reset gate of the j-th hidden node. The bias vector representing the reset gate of the j-th hidden node. This represents the hidden weight vector of the gate that updates the i-th input node. The network input weight vector representing the update gate of the j-th hidden node. The bias vector representing the update gate of the j-th hidden node. Let ri represent the hidden weight vector of the candidate hidden state of the i-th input node, and rjt represent the output of the reset gate of the j-th hidden node. The network input weight vector represents the candidate hidden state of the j-th hidden node. represents the bias vector of the candidate hidden state of the j-th hidden node, ⊙ represents the Hadamard product operation, and p and q represent the total number of input nodes and hidden nodes of the GRU neural network, respectively.

[0021] Furthermore, the general expression for the SVCRU hidden state update function designed in step 3 is:

[0022]

[0023] Here, α and β represent the weights of the current candidate hidden state output in the topological dimension and the output of the hidden state in the previous time step, respectively, playing the roles of forgetting and updating. The complementary adjustment mechanism between the two enables it to incorporate current information while preserving long-term dependencies, thereby more effectively capturing dynamic changes and long-term patterns in continuous data.

[0024] Furthermore, the calculation method for the rate adaptive adjustment gate st designed in step 4 is as follows:

[0025]

[0026] in, This represents the network input weight vector of the rate adaptive adjustment gate for the i-th input node. The network input weight vector represents the rate adaptive adjustment gate for the j-th hidden node. This is the corresponding bias vector. Similar to other gate controllers, the rate-adaptive regulating gate s... t It is determined through the network training process. The main difference between the two lies in the rate adaptive adjustment gate s. t It consists of only a single neuron, so it has a small impact on the computational complexity of SVCRU.

[0027] Furthermore, in step 4, the rate adaptive adjustment gate st is compared with the update gate. By combining these methods, a nonlinear exponential normalization transformation is performed to form a variable-rate input gate. and variable rate forget gate The calculation method is as follows:

[0028]

[0029] in, Represents a complementary update gate. The exponential change forms representing the rate-adaptive adjustment gate and the update gate, This represents the exponential change form of the rate-adaptive regulation gate and the complementary update gate. In the rate-adaptive regulation gate s... t With the update gate Under coordinated control, variable rate input gate and variable rate forget gate SVCRU will adjust the network information with a non-linear rate of change.

[0030] Furthermore, in step 5, the variable rate input gate is used... and variable rate forget gate By embedding these values ​​into the general expression of the SVCRU hidden state update function, and replacing the inherent linear constant rate adjustment constraint of the GRU with a nonlinear variable rate adjustment mechanism, the expression for the SVCRU hidden state update function is obtained as follows:

[0031]

[0032] in, pass Instead of α, it serves as a memory regulation gate, selectively extracting relevant and important information from candidate hidden states using a variable rate mechanism. pass SVCRU replaces β as the forgetting regulation gate to dynamically adjust and discard irrelevant information from the previous hidden state.

[0033] Furthermore, in step 6, the SVCRU network is combined with a multilayer perceptron to construct an SVCRU network model, the output of which is shown below:

[0034]

[0035] Where g(·) is the activation function of the output layer of the SVCRU network, W y Let b be the weight matrix of the network output layer. y This is the corresponding bias vector.

[0036] Furthermore, in step 7, historical data from the industrial production / processing process is used to train the SVCRU neural network prediction model constructed in step 6, resulting in a trained SVCRU neural network prediction model.

[0037] Furthermore, in step 8, real-time data of process variables in the industrial production / processing process are obtained and input into the trained SVCRU neural network prediction model obtained in step 7 to obtain the predicted value of the target variable in the process.

[0038] The beneficial effects of this invention are:

[0039] This invention proposes an industrial soft measurement method based on a variable-rate complementary recurrent unit (GRU) network. This method combines an adaptive rate adjustment gate with an update gate, performing a nonlinear normalized exponential transformation to improve the model's nonlinear feature extraction capability, thus forming a variable-rate forget gate and a variable-rate input gate. Compared to existing methods that use hyperparameters to adjust the information flow rate, this method can adaptively learn rate adjustment factors at different time steps based on the information flow characteristics during model training. Finally, the variable-rate forget gate and variable-rate input gate are embedded into the GRU hidden state update function, using an adaptive information flow adjustment mechanism to replace the inherent linear constant-rate adjustment constraint of the GRU. This method effectively achieves a dynamic balance between current and historical input information, maintaining their complementary adjustment relationship while modifying the inherent linear constant-rate adjustment constraint. This effectively improves the model's information transmission rate and feature extraction capability, thereby enhancing the model's predictive performance. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0041] Figure 1 This is a structural diagram of the gated loop unit in the industrial soft measurement modeling method of the present invention;

[0042] Figure 2 The diagram shows the internal physical structure and gate topology of the gated cyclic unit in an industrial soft measurement method based on a variable-rate complementary cyclic unit network provided by this invention.

[0043] Figure 3 A schematic diagram of the variable-rate complementary cyclic unit structure constructed in an industrial soft measurement method based on a variable-rate complementary cyclic unit network provided by the present invention.

[0044] Figure 4 The diagram shows the internal physical structure and gate topology of the variable-rate complementary cyclic unit (CRCU) constructed in the industrial soft measurement method based on the CRCU network provided by this invention.

[0045] Figure 5 A schematic diagram of the flue gas desulfurization process in a coal-fired power plant;

[0046] Figure 6 This is a schematic diagram showing the changes in the loss function of prediction methods based on different algorithms on the training set;

[0047] Figure 7AThe image shows the SO2 concentration prediction fitting curve obtained using the existing GRU-based prediction method.

[0048] Figure 7B The image shows the SO2 concentration prediction fitting curve obtained using the existing TS-λGRUs prediction method.

[0049] Figure 7C The image shows the SO2 concentration prediction fitting curve obtained using the existing L / S-ConvGRU prediction method.

[0050] Figure 7D The SO2 concentration prediction fitting curve obtained by the SVCRU neural network prediction method based on the industrial soft measurement method of variable rate complementary cyclic unit network provided in this application.

[0051] Figure 8A This is a frequency distribution map of SO2 concentration prediction errors obtained using the existing GRU-based prediction method;

[0052] Figure 8B This is a frequency distribution map of SO2 concentration prediction errors obtained using the existing TS-λGRUs prediction method;

[0053] Figure 8C This is a frequency distribution map of SO2 concentration prediction errors obtained using the existing L / S-ConvGRU prediction method;

[0054] Figure 8D The frequency distribution of SO2 concentration prediction error obtained by the SVCRU neural network prediction method based on the industrial soft measurement method of variable rate complementary cyclic unit network provided in this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0056] First, let's introduce the basic theories involved:

[0057] GRU Neural Network Algorithm

[0058] GRU is a gated recurrent neural network that utilizes time feedback and gate controllers to regulate information flow, aiming to solve the vanishing and exploding gradient problems of traditional recurrent neural networks. Compared to LSTM, GRU has a simpler structure and fewer parameters, thus accelerating training speed while maintaining comparable performance. Figure 1 As shown. The GRU network threshold mechanism consists of a reset threshold r t and update gate z t Two threshold controllers and candidate hidden states The system comprises the following components: a reset gate, which determines whether the current candidate hidden state depends on the previous hidden state and the degree of dependence; an update gate, which controls the update of the hidden state from time t-1 to time t; and a candidate hidden state, which controls the retention of relevant input information. The formulas for each component are defined as follows:

[0059] r t =σ(U r x t +W r h t-1 +b r (1)

[0060] z t =σ(U z x t +W z h t-1 +b z (2)

[0061]

[0062] Where σ(·) and tanh(·) are the sigmoid activation function and the hyperbolic tangent activation function, respectively; x t and h t-1 W represents the current input and the previous hidden state output, respectively. r W z and W h U is the input weight vector to the network. r U z and U h To hide the weight vector, b r b z and b h Here, is the bias vector; ⊙ represents the Hadamard product operation. The hidden state h of the GRU at time t. t It can be represented as:

[0063]

[0064] As mentioned earlier, GRU effectively regulates information flow through the cooperation of gating controllers, achieving dynamic updates and continuous information retention. When z t When z approaches zero, GRU prioritizes short-term dependencies, tending to extract features from candidate states at the current time step; while when z approaches zero... t As the value approaches 1, GRU tends to retain previous hidden states to capture long-term dependencies. This single-gated adjustment mechanism maintains a dynamic balance between long-term dependencies and short-term memory, effectively mitigating the vanishing and exploding gradient problems. However, this mechanism is specifically manifested as a linear constant-rate adjustment constraint, i.e., when z... t When changes occur, 1-z tThey will inevitably change at the same rate, which will hinder the full transmission of information and affect the learning speed of model training.

[0065] Example 1

[0066] This embodiment provides an industrial soft measurement modeling method based on variable-rate complementary cyclic unit networks, the method comprising:

[0067] Step 1: Obtain historical data of process variables and target variables in the industrial production / processing process as a training sample database; and preprocess the training sample database to obtain the training dataset and test dataset.

[0068] Historical data of process variables and target variables were collected to form a training sample database. The training dataset and test dataset were obtained by preprocessing the training sample database by handling missing values, outliers, and data standardization. The ratio of the training dataset to the test dataset was 8:2.

[0069] Industrial production / processing technologies include desulfurization processes in coal-fired power plants, butane removal tower processes, and fermentation processes.

[0070] Step 2: Construct an initial GRU network architecture based on a deep learning server, consisting of gate controllers such as reset gates, update gates, and candidate hidden states. Its internal physical architecture is as follows: Figure 2 As shown;

[0071] The expression for the gate controller in the initial GRU network architecture is:

[0072]

[0073] Where, r t k, and Let represent the outputs of the reset gate, update gate, and candidate hidden state in the topological dimension, respectively; σ(·) and tanh(·) represent the sigmoid activation function and the hyperbolic tangent activation function, respectively; and k represents the topological output dimension. This represents the current input at the i-th input node. This represents the hidden weight vector of the reset gate at the i-th input node. This represents the hidden state output of the j-th hidden node at time t-1. The network input weight vector representing the reset gate of the j-th hidden node. The bias vector representing the reset gate of the j-th hidden node. This represents the hidden weight vector of the gate that updates the i-th input node. The network input weight vector representing the update gate of the j-th hidden node. The bias vector representing the update gate of the j-th hidden node. Let ri represent the hidden weight vector of the candidate hidden state of the i-th input node, and rjt represent the output of the reset gate of the j-th hidden node. The network input weight vector represents the candidate hidden state of the j-th hidden node. represents the bias vector of the candidate hidden state of the j-th hidden node, ⊙ represents the Hadamard product operation, and p and q represent the total number of input nodes and hidden nodes of the GRU neural network, respectively.

[0074] Step 3: Based on the GRU's use of update gates to linearly interpolate its hidden states to achieve information fusion, the general expression of the SVCRU hidden state update function is designed as follows:

[0075]

[0076] Here, α and β represent the weights of the current candidate hidden state output in the topological dimension and the output of the hidden state in the previous time step, respectively, playing the roles of forgetting and updating. The complementary adjustment mechanism between the two enables it to incorporate current information while preserving long-term dependencies, thereby more effectively capturing dynamic changes and long-term patterns in continuous data.

[0077] Step 4: Design a new rate adaptive adjustment gate st, and combine it with the update gate from Step 2 to obtain the SVCRU network. By performing a nonlinear normalized exponential transform, a variable rate input gate is formed. and variable rate forget gate

[0078] The calculation method for the rate adaptive control gate st is as follows:

[0079]

[0080] in, This represents the network input weight vector of the rate adaptive adjustment gate for the i-th input node. The network input weight vector represents the rate adaptive adjustment gate for the j-th hidden node. This is the corresponding bias vector. Similar to other gate controllers, the rate-adaptive regulating gate s... t It is determined through the network training process. The main difference between the two lies in the rate adaptive adjustment gate s. t It consists of only a single neuron, so it has a small impact on the computational complexity of SVCRU.

[0081] By adjusting the rate adaptive adjustment gate st with the update gate z t By combining these methods, a nonlinear exponential normalization transformation is performed to form a variable-rate input gate. and variable rate forget gate The calculation method is as follows:

[0082]

[0083] in, Represents a complementary update gate. The exponential change forms representing the rate-adaptive adjustment gate and the update gate, This represents the exponential change form of the rate-adaptive regulation gate and the complementary update gate. In the rate-adaptive regulation gate s... t With the update gate Under coordinated control, variable rate input gate and variable rate forget gate SVCRU will adjust the network information with a non-linear rate of change.

[0084] A schematic diagram of the variable-rate complementary cyclic unit structure is shown below. Figure 3 As shown, in its core calculation module, the rate adaptive adjustment gate and the update gate are transformed into a variable rate input gate by adopting a nonlinear exponential normalization transformation. and variable rate forget gate Both factors together determine the variable rate adjustment of the hidden state in the SVCRU network.

[0085] Step 5: Input the variable rate gate from Step 4 and variable rate forget gate By embedding the general expression of the SVCRU hidden state update function in step 3, the updated SVCRU hidden state update function is obtained, and thus the SVCRU network is obtained.

[0086] By using variable rate input gates and variable rate forget gate By embedding these values ​​into the general expression of the SVCRU hidden state update function, and replacing the inherent linear constant rate adjustment constraint of the GRU with a nonlinear variable rate adjustment mechanism, the updated expression of the SVCRU hidden state update function is obtained as follows:

[0087]

[0088] in, pass Instead of α, it serves as a memory regulation gate, selectively extracting relevant and important information from candidate hidden states using a variable rate mechanism. pass SVCRU replaces β as the forgetting regulation gate to dynamically adjust and discard irrelevant information from the previous hidden state.

[0089] The internal physical and gate topology of the SVCRU network are as follows: Figure 4As shown, the topology of the gate controller in the SVCRU network is illustrated using the reset gate as an example. Its input node is p, and the hidden node is q, employing a feedforward fully connected approach for information transmission. Furthermore, the topology connection method for the rate adaptive adjustment gate and the update gate is demonstrated. This method uses a nonlinear exponential transformation followed by normalization to obtain the variable rate input gate. and variable rate forget gate

[0090] Step 6: Combine the SVCRU network obtained in Step 5 with a multilayer perceptron to construct an SVCRU neural network prediction model;

[0091] By combining the SVCRU network with a multilayer perceptron, an SVCRU network model is constructed. The model output is shown below:

[0092]

[0093] Where g(·) is the activation function of the output layer of the SVCRU network, W y Let b be the weight matrix of the network output layer. y This is the corresponding bias vector.

[0094] Step 7: Train the SVCRU neural network prediction model from Step 6 using the training dataset obtained in Step 1 to obtain the trained SVCRU neural network prediction model.

[0095] Step 8: Obtain real-time data of process variables during the process, input them into the trained SVCRU neural network prediction model obtained in Step 7, and obtain the predicted value of the target variable during the process.

[0096] Example 2

[0097] This embodiment provides a method for dynamically predicting SO2 content in the flue gas desulfurization process of coal-fired power generation based on a variable-rate complementary cyclic unit network. This method is based on the SVCRU neural network prediction model proposed in Embodiment 1 to dynamically predict the SO2 content in the flue gas desulfurization process of coal-fired power generation.

[0098] First, the flue gas desulfurization process of a certain coal-fired power plant is introduced, such as... Figure 5As shown, this power plant employs limestone-gypsum wet flue gas desulfurization technology. The reaction principle involves the reaction of CaCO3 and oxygen in the desulfurizing agent with SO2 in the flue gas to generate gypsum dihydrate. Lump limestone is ground into limestone powder using a wet ball mill, which is then mixed with water to generate a limestone slurry of a certain concentration. This slurry is then pumped to the spray layer by a variable frequency circulating water pump to form the desulfurization circulating slurry. The desulfurization slurry, atomized from the spray layer, is sprayed downwards in a staggered manner, reacting with the counter-current sulfur-containing flue gas to generate gypsum slurry. The gypsum slurry meeting the density requirements enters the gypsum dewatering system, where it is dewatered to form gypsum, a desulfurization byproduct. The SO2 concentration in the emitted flue gas is a key quality variable that the enterprise needs to monitor and control to ensure compliance with national pollutant emission standards. However, in actual production, SO2 concentration fluctuates significantly, leading to unstable operation of the online analyzer and high maintenance costs. Therefore, designing suitable soft sensors to provide verification data is crucial for improving the stability of the data monitoring system.

[0099] The SVCRU neural network prediction model proposed in Example 1 is used to predict the dynamic SO2 content in the flue gas desulfurization process of coal-fired power generation. The prediction steps include:

[0100] Step 1: Simulation experiment data was collected from the plant-level monitoring information system of the flue gas desulfurization process in the coal-fired power plant. The data sampling period was 5 minutes, and the sampling time was 31 days, including 52 process variables and 1 target variable. The target variable was the net flue gas SO2 concentration at the outlet of the secondary absorption tower, with a total of 8928 sets of sample data.

[0101] Step 2: Collect process variables in the flue gas desulfurization process of coal-fired power plants and input them into the SVCRU neural network prediction model to predict the target variables.

[0102] To analyze the prediction performance of the SVCRU neural network prediction model proposed in this invention from multiple perspectives, its performance is compared with three soft sensor prediction methods: GRU, two-stream λ-factor GRU network prediction method (TS-λGRUs), and two-path convolutional GRU prediction method (L / S-ConvGRU). The TS-λGRUs and L / S-ConvGRU prediction methods work by introducing additional hyperparameters to change the inherent linear constant-rate adjustment constraint of the GRU. For details on GRU, please refer to the literature "Learning phrase representations using RNN encoder-decoder for statistical machine translation"; for details on the two-stream λ-factor GRU network prediction method, please refer to the literature "Data-driven modeling based on two-stream λ-gated recurrent unit network with soft sensor application"; for details on the two-path convolutional GRU prediction method, please refer to the literature "A soft sensor model based on long & short-term memory dual pathways convolutional gated recurrentunit network for predicting cement specific surface area".

[0103] In addition, the present invention uses the following four parameters as model performance evaluation indicators, and their calculation formulas are as follows:

[0104] (1) Root mean square error (MSE): reflects the degree of deviation between the model's predicted value and the true value. The smaller the value, the higher the accuracy of the model's prediction.

[0105]

[0106] (2) Mean absolute percent error (MAPE): The average of the absolute values ​​of the deviations between a single observation and the arithmetic mean, used to evaluate the accuracy of the model's prediction results.

[0107]

[0108] (3) Correlation coefficient (COR): It reflects the tracking and fitting ability of the model data prediction. The closer its value is to 1, the better the model prediction performance.

[0109]

[0110] (4) Jensen-Shannon divergence (JSD): It reflects the similarity of the probability distribution between the model's predicted value and the true value. The closer its value is to 0, the better the model's prediction performance.

[0111]

[0112] Where N is the number of samples in the test set, y i and These represent the actual and predicted values ​​of the output variables in the test dataset samples, respectively. For y i The average value, P y and They are y i and The probability distribution of P i and Q i They are P y and The probability value of the i-th sample. It is P i and Q i The average probability distribution between them.

[0113] The simulation environment for each algorithm was a deep learning server: Intel Cor(TM) i7-12700H 2.30GHz CPU processor, NVIDIA GeForce RTX 3060 GPU graphics card, 64GB memory, and Python 3.9 under a 64-bit Windows 11 operating system. The network structure and hyperparameter settings for different algorithms are shown in Table 1.

[0114] Table 1. Network structure and hyperparameter settings for different algorithms

[0115]

[0116]

[0117] To ensure the rationality of the simulation settings, this invention employs a network search method to optimize the hyperparameters of different algorithms, and all algorithms use the same GRU basic network configuration. The learning loss curves of different algorithms on the training set are shown below. Figure 6As shown, the convergence performance of other models is significantly better than that of GRU during training. This indicates that breaking the linear constant rate adjustment constraint of the GRU hidden state can significantly improve its information transmission efficiency and learning performance. Among them, the SVCRU algorithm has the fastest convergence speed and reaches a stable state first in the fast convergence phase. In the final stable phase, the SVCRU prediction method has the smallest loss function value. This result shows that the SVCRU algorithm proposed in this application has the best convergence performance compared with other algorithms.

[0118] In addition, the present invention uses the same dataset and simulation environment to run the prediction methods based on each algorithm 10 times, and uses the optimal value of the statistical results as the final index of the model. The simulation results of different prediction methods are shown in Table 2.

[0119] Table 2 Simulation results of SO2 concentration prediction using different prediction methods

[0120]

[0121] Table 2 shows that, compared with the prediction method based on the traditional GRU algorithm, the prediction methods based on other algorithms show significant reductions in all performance evaluation indicators. This result indicates that breaking the linear constant-rate adjustment constraint in the GRU hidden state can effectively improve the model information transmission efficiency and enhance its prediction performance. Furthermore, compared with the prediction methods based on the TS-λGRU algorithm and the L / S-ConvGRU algorithm, the prediction method based on the SVCRU algorithm proposed in this invention achieves lower model error evaluation indicators MAPE and RMSE, and a lower coefficient of determination R0. 2 The higher accuracy and lower JSD indicate that the prediction method based on the SVCRU algorithm has higher prediction accuracy, and the probability distributions of the predicted and true values ​​show a high degree of similarity, fully verifying the superiority of the proposed prediction method. Furthermore, to more intuitively demonstrate the prediction performance of different methods, this invention presents the SO2 concentration prediction fitting curves based on different algorithms, as shown below. Figure 7A , Figure 7B , Figure 7C and Figure 7D As shown; the histograms of the frequency distribution of prediction errors with normal fitting are respectively as follows: Figure 8A , Figure 8B , Figure 8C and Figure 8D As shown.

[0122] The SO2 concentration prediction fitting curves based on different algorithms clearly show that, compared with other prediction methods, the SO2 concentration prediction method based on the SVCRU algorithm proposed in this invention yields more accurate prediction results. Its prediction fitting curve matches the actual observed values ​​better, indicating that the SVCRU algorithm-based method can more effectively capture the dynamic changes in SO2 concentration. Furthermore, the error frequency distribution histograms of different algorithms show that the prediction error of the SVCRU algorithm-based method is more concentrated near zero, with smaller standard deviation (σ = 1.7455) and average deviation (μ = 0.3969), resulting in a smaller difference between the predicted and actual values ​​and higher prediction accuracy.

[0123] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial soft sensing method based on a variable-rate complementary cyclic unit network, characterized in that, The method includes: Step 1: Obtain historical data of process variables and target variables in the industrial production / processing process as a training sample database, and preprocess the training sample database to obtain the training dataset and test dataset; Step 2: Construct the initial GRU network architecture, which includes reset gates, update gates, and candidate hidden states; Step 3: Based on the update gate in the initial GRU network architecture constructed in Step 2, linear interpolate the candidate hidden states to achieve information fusion, and design the SVCRU candidate hidden state update function expression: in, and These represent the weights of the current candidate hidden state output in the topological dimension and the output of the hidden state in the previous time step, respectively, serving the functions of forgetting and updating. Step 4: Design a rate adaptive adjustment gate This is then combined with the update gate in the initial GRU network architecture constructed in step 2 to obtain the SVCRU network. Through nonlinear normalized exponential transformation, a variable rate input gate is formed. and variable rate forget gate : in, Represents a complementary update gate. The exponential change forms representing the rate-adaptive adjustment gate and the update gate, The exponential change forms of the rate-adaptive adjustment gate and the complementary update gate; Step 5: Input the variable rate gate obtained in Step 4 and variable rate forget gate The SVCRU candidate hidden state update function expression obtained in step 3 is embedded into the updated SVCRU candidate hidden state update function expression, and the SVCRU network is obtained. Step 6: Combine the SVCRU network obtained in Step 5 with a multilayer perceptron to construct an SVCRU neural network prediction model; Step 7: Train the SVCRU neural network prediction model obtained in Step 6 using the training dataset obtained in Step 1 to obtain the trained SVCRU neural network prediction model. Step 8: Input the process variable corresponding to the target variable to be predicted into the SVCRU neural network prediction model obtained in step 7 to predict the target variable.

2. The method according to claim 1, characterized in that, In step 2, the reset gate, update gate, and output expressions for the candidate hidden states in the initial GRU network architecture are as follows: in, , and These are the topological outputs of the reset gate, update gate, and candidate hidden state, respectively. and These are the sigmoid activation function and the hyperbolic tangent activation function, respectively. k Represents the topology output dimension. Representing the i The current input of each input node. Representing the i The hidden weight vector of the gate is reset for each input node. represent t -1 moment j Output the hidden state of each hidden node. Representing the j The hidden node resets the network input weight vector of the gate. Representing the j The bias vector of the reset gate for each hidden node. Representing the i Update the hidden weight vector of the gate for each input node. Representing the j The network input weight vector of the update gate for each hidden node. Representing the j The bias vector of the update gate for each hidden node. Representing the i The hidden weight vector of each input node's candidate hidden state. Representatives of the 1st j The output of the reset gate for each hidden node. Representing the j The network input weight vector of the candidate hidden states of each hidden node. Representing the j The bias vector of the candidate hidden states of each hidden node. Represents the Hadamard product operation. p and q These represent the total number of input nodes and hidden nodes in the GRU neural network, respectively.

3. The method according to claim 2, characterized in that, The rate adaptive adjustment gate in step 4 The expression is: in, Representing the i The network input weight vector of an input node rate adaptive adjustment gate Representing the j The network input weight vector of a hidden node rate adaptive adjustment gate This is the corresponding bias vector.

4. The method according to claim 3, characterized in that, In step 5, the variable rate input gate is used. and variable rate forget gate These are respectively embedded into the SVCRU hidden state update function obtained in step 3, and the specific calculation method is as follows: in, ,pass replace As a memory regulation gate, it selectively extracts relevant and important information from candidate hidden states using a variable rate mechanism; ,pass replace It serves as a forgetting regulation gate, used to dynamically adjust and discard the rate at which irrelevant information from the previous hidden state is discarded.

5. The method according to claim 4, characterized in that, The expression output by the SVCRU neural network prediction model in step 6 is: in, The activation function for the output layer of the SVCRU network. This is the weight matrix of the network output layer. This is the corresponding bias vector.

6. The method according to claim 5, characterized in that, The preprocessing in step 1 includes handling missing values, outliers, and data standardization.

7. The method according to claim 6, characterized in that, In step 1, the ratio of the training dataset to the test dataset is 8:

2.

8. A method for predicting the dynamic SO2 content during flue gas desulfurization in coal-fired power generation, characterized in that, The method is implemented based on the method described in any one of claims 1-7 above. For the desulfurization process of flue gas from coal-fired power generation, the process variables include the pH value of the desulfurization slurry, the flue gas velocity, the oxygen content of the flue gas, the slurry supply flow rate of the secondary absorption tower, the net flue gas flow rate at the inlet of the furnace chimney, the net flue gas pressure at the inlet of the furnace chimney, the O2 concentration of the original flue gas, the outlet pressure of the secondary absorption tower, the flow rate of the limestone slurry absorption tower, and the flue gas temperature at the inlet of the absorption tower. The target variable is the SO2 concentration in the emitted clean flue gas.

Images