A soft measurement method for O2 concentration prediction in air separation device

Abstract

The present application relates to the field of data-driven soft measurement modeling, and particularly to a soft measurement method for O2 concentration prediction in an air separation device, comprising the following steps: S1, obtaining a historical data set: sampling air separation device data at a time interval T to form a historical sample data set D={X, Y}; S2, data preprocessing: preprocessing the data set D={X, Y} using a standardization method to eliminate the dimensional relationship between data; S3, constructing a gated recurrent unit (GRU) basic network model: using 80% of the data set as a training set and the remaining 20% as a test set to train a GRU basic network model; and S4, using a GRU model adaptive variable selection algorithm based on variance decomposition sensitivity analysis and a non-negative strangling algorithm (NNG) to realize soft measurement modeling of O2 concentration; the present application uses a GRU neural network to capture the nonlinear time sequence characteristics between data, uses a sensitivity analysis based on variance decomposition and NNG to realize adaptive input variable selection, and applies the same to soft measurement of O2 concentration in an air separation device.

Description

Technical Field

[0001] This invention relates to the field of data-driven soft measurement modeling, specifically to a soft measurement method for predicting O2 concentration in air separation devices. This method is based on an adaptive variable selection algorithm using a gated recurrent unit (GRU) and is applied to the soft measurement of O2 concentration in air separation devices. Background Technology

[0002] In actual industrial processes, to ensure safe and efficient production, it is necessary to monitor and control critical process variables that affect equipment stability and product quality in real time. However, some key variables in the production process cannot be directly measured, and offline analyzers and indirect measurement methods suffer from significant lag and difficulty in guaranteeing measurement accuracy. To effectively solve these problems, soft measurement technology has emerged. Soft measurement technology complements traditional measurement methods. It mainly relies on system identification and system modeling theories to establish predictive models for key variables that are difficult to obtain, using directly measurable process variables and other parameters.

[0003] In recent years, soft sensing technology has developed rapidly, with research covering fields such as chemical engineering, petroleum, and environmental protection. Soft sensing models have also expanded from linear and mechanistic models to "black box" based neural network models. Although the development of machine learning technology has improved the accuracy of soft sensing modeling, some significant problems still exist. On the one hand, real-world industrial process data often exhibits complex time-delay characteristics; the state of certain key parameters is not only related to the current input state but also to the input state at a past moment or time period. This makes it difficult for traditional feedforward neural network-based soft sensing models to meet actual production needs. On the other hand, with the development of industrial technology, the number of process variables collected in industrial processes has increased, resulting in a large amount of redundant information that affects the accuracy of soft sensing modeling. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a soft-sensing method for predicting O2 concentration in air separation devices. Recognizing the nonlinearity, multivariability, and time delay issues present in the raw data provided by the air separation device, this invention employs a GRU neural network as the basic model. It utilizes variance decomposition-based sensitivity analysis and NNG to achieve adaptive input variable selection for the GRU model, and applies this method to the soft measurement of O2 concentration in air separation devices. On one hand, this algorithm uses GRU as the basic network, fully considering the complex temporal relationships between data; on the other hand, it adds the total sensitivity index as an adaptive weight vector to the optimization expression of the NNG, achieving an approximately unbiased estimation of the input variable weights of the GRU model, thereby improving the reliability and stability of O2 concentration prediction in air separation devices.

[0005] This invention provides a soft measurement method for predicting O2 concentration in air separation devices, comprising the following steps:

[0006] S1. Obtain historical dataset: Sample the air separation device data at time intervals T to form a historical sample dataset D = {X, Y}; In step S1, at least 6000 data points are collected during sampling.

[0007] S2. Data preprocessing: The dataset D = {X, Y} is preprocessed using a standardization method to eliminate the dimensional relationships between the data.

[0008] S3. Construct a basic network model of a gated recurrent unit (GRU): Use 80% of the dataset D = {X, Y} as the training set and the remaining 20% ​​as the test set to train a basic GRU network model.

[0009] S4. Using sensitivity analysis based on variance decomposition and the GRU model adaptive variable selection algorithm based on nonnegative garrote (NNG), soft measurement modeling of O2 concentration in air separation device is realized.

[0010] In step S4, during the variable selection process, the NNG algorithm uses sensitivity analysis based on variance decomposition to calculate the total sensitivity index and uses the total sensitivity index as the adaptive weight vector of the NNG algorithm.

[0011] The steps for obtaining the total sensitivity index are as follows.

[0012] X A and X B Let X be two independent input variable matrices. A and X B The sizes are equal and the number of rows is X = [X] A ;X B ], j and i represent X respectively A X B The row and column indices are defined, f represents the trained GRU model, and the matrix is ​​defined. Indicates X A Replace column i with X B The new matrix after the i-th column, for:

[0013]

[0014] The overall sensitivity index is expressed as:

[0015]

[0016] Construct an adaptive weight vector: Calculate the relative importance factor of each input variable relative to the output variable based on the overall sensitivity index.

[0017]

[0018] μ i As the adaptive weight vector of the NNG algorithm, a GRU model is introduced to achieve adaptive variable selection. First, the shrinkage operator θ = [θ1, θ2, ..., θ...] of NNG is... k Introducing GRU model input weights w x* In this process, the optimized basic unit expression of the GRU model is obtained:

[0019]

[0020]

[0021]

[0022]

[0023] Secondly, the relative importance factor μ i The NNG algorithm is introduced, and constrained quadratic optimization is performed.

[0024]

[0025] Obtain the optimal shrinkage coefficient θ * =[θ1 * θ2 * , …, θ k * ].

[0026] Finally, θ * =[θ1 * θ2 * , …, θ k * Adding this to the basic unit expression of the GRU model yields the optimized GRU prediction model:

[0027]

[0028] As can be seen from the above description: 1. To address the time-lag characteristics of the raw data collected from the air separation device, this invention utilizes a GRU recurrent neural network as the basic network for modeling; 2. The presence of numerous input variables unrelated to the target variable in the original dataset D = {X, Y} leads to high complexity and poor accuracy in the GRU model. Therefore, the NNG algorithm is introduced into the GRU model to achieve input variable selection; 3. During variable selection, the NNG algorithm compresses the GRU input weights with equal intensity, resulting in biased estimation of the final model's input variable weight coefficients. This paper utilizes sensitivity analysis based on variance decomposition to calculate the total sensitivity index and uses it as the adaptive weight vector of the NNG algorithm to overcome the biased coefficient estimation of the NNG algorithm and achieve adaptive variable selection in the GRU model; 4. The selection of the strangulation parameter s in the NNG algorithm has a significant impact on the input weight coefficients of the GRU model. Block cross-validation is used to determine the optimal strangulation parameter s for the NNG algorithm; 5. Using sensitivity analysis based on variance decomposition and the NNG-based adaptive variable selection algorithm for the GRU model, soft measurement modeling of O2 concentration in the air separation device is achieved. The effectiveness of the proposed method is verified by comparing the prediction results with those of other advanced soft measurement modeling methods. Attached Figure Description

[0029] Figure 1 This is the basic unit structure of the GRU network.

[0030] Figure 2 This is a diagram of the physical structure of a GRU network.

[0031] Figure 3 This describes the process flow of an air separation unit.

[0032] Figure 4 O2 concentration prediction curve for the SANNG-GRU soft measurement modeling method. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of specific embodiments. Obviously, the described specific embodiments are merely one specific embodiment of the present invention, and not all specific embodiments. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] The attached diagram illustrates a soft measurement method for predicting O2 concentration in air separation devices, comprising the following steps: S1. Obtaining historical datasets: Data from the air separation device is sampled at time intervals T to form a historical sample dataset D = {X, Y}; S2. Data preprocessing: The dataset D = {X, Y} is preprocessed using a standardization method to eliminate dimensional relationships between data; S3. Constructing a GRU basic network model: 80% of the dataset D = {X, Y} is used as the training set, and the remaining 20% ​​is used as the test set to train a GRU basic network model; S4. Using sensitivity analysis based on variance decomposition and the NNG-based adaptive variable selection algorithm for the GRU model, soft measurement modeling of O2 concentration in the air separation device is achieved.

[0035] 1. Gated Recurrent Units (GRUs), as an advanced recurrent neural network, not only extract temporal information between data points but also retain the ability of LSTM (Long Short-Term Memory) to resist gradient vanishing and gradient exploding. Furthermore, compared to LSTM, the simpler network structure of GRUs simplifies the model training process. The basic unit structure of a GRU network is as follows: Figure 1 As shown, it only includes the reset gate r. t and update gate z t Two doors. Candidate hidden state. It mainly includes the input x at the current time. t r t Decide Including the current input x t The ratio. Hidden state h t Using z t Simultaneous forgetting and selective memorization.

[0036] According to x t and the hidden state h from the previous moment t-1 A typical basic unit of a GRU network can be implemented using the following formula:

[0037] r t =σ(x t w xr +h t-1 w hr +b r )#(1)

[0038] z t =σ(x t w xz +h t-1 w hz +b z )#(2)

[0039]

[0040]

[0041] Where w * and b * Let represent the weights and biases respectively, ⊙ represent element-wise multiplication at corresponding positions, and σ and tanh represent the sigmoid activation function and the tanh activation function respectively. The sigmoid activation function will multiply r t and z t The range is controlled within (0, 1). t The closer it is to 1, the better it represents h. t The more data you remember; t The closer to 0, the better h is. t The more data is forgotten, the better. The GRU network's model output at time t is:

[0042]

[0043] 2. Non-negative strangulation algorithm: Initially, the NNG algorithm was proposed to solve the variable selection problem in multiple linear regression. For ordinary least squares (OLS) regression models:

[0044]

[0045] in For the predicted values ​​of the output variables, x = [x1, x2, ... x2]. k Input variables, k is the number of input variables, β = [β1, β2, ..., β] k ] T It is a coefficient vector, where σ is the random error. The NNG algorithm uses the constrained shrinkage operator θ = [θ1, θ2, ..., θ... k Add this to the OLS model to enable input variable selection. The specific expression is as follows:

[0046]

[0047] Where Y∈R n×1 Output sample vector, X∈R n×k Let be the input sample matrix, y be the output variable, and n be the number of samples. When s→k, NNG can be equivalent to OLS; when s→0, all input variable coefficients are compressed to 0. The NNG algorithm utilizes the change in the strangulation parameter s to achieve input variable selection.

[0048] 3. Sensitivity analysis based on variance decomposition, assuming the model output... The input variable is X = [X1, X2, ..., X...]. k The function of ], that is Input variable X iGiven a unique and uniformly distributed system on [0, 1], define the following function...

[0049]

[0050] The same definition method can be used for higher-order problems. Model output. It can be decomposed into:

[0051]

[0052] The variance can be expressed as:

[0053]

[0054] in This represents the variance of the model output. This means only considering X. i right Contribution Indicates considering X i and X j Interaction Pairs The contribution of the higher-order terms in formula (10) has the same meaning. First-order sensitivity index:

[0055]

[0056] Only measure X i acting alone The influence of features. This paper considers the interactions between features and introduces a total sensitivity index to more comprehensively evaluate the importance of input variables. The total sensitivity index is a subset of X. i The sum of all variance terms Impact:

[0057]

[0058] Where X ～i It is divided by X i All input variables other than For containing only X i variance and Dividing X i The expected value and variance of all input variables other than the given variables.

[0059] Total Sensitivity Index Calculation: This paper uses the total sensitivity index calculation method proposed by Jansen. Assume X... A and X B Let X be two independent input variable matrices. A and X B The sizes are equal and the number of rows is X = [X]A ;X B ], j and i represent X respectively A X B The row and column indices are given, and f represents the trained GRU model. Define the matrix. Indicates X A Replace column i with X B The new matrix after the i-th column. It can be approximated as:

[0060]

[0061] The overall sensitivity index can be expressed as:

[0062]

[0063] Constructing an adaptive weight vector: Calculate the relative importance factor of each input variable relative to the output variable based on the overall sensitivity index.

[0064]

[0065] μ i As the adaptive weight vector for the NNG algorithm, the GRU model is introduced to achieve adaptive variable selection: Figure 2 This diagram illustrates the principle of NNG applied to a GRU model to achieve variable selection. First, the NNG contraction operator θ = [θ1, θ2, ..., θ...] is... k Introducing GRU model input weights w x* In this process, the optimized basic unit expression of the GRU model is obtained:

[0066]

[0067]

[0068]

[0069]

[0070] Secondly, the relative importance factor μ i The NNG algorithm is introduced, and a constrained quadratic optimization problem is solved:

[0071]

[0072] Obtain the optimal shrinkage coefficient θ * =[θ1 * θ2 * , …, θ k * ], when θ i *=0, input variable x i Deleted. Algorithm 1 demonstrates the detailed steps for determining the strangulation parameter s using block cross-validation and the Bayesian information criterion (BIC). Finally, θ is... * =[θ1 * θ2 * , …, θ k * Add the weights to the input weights of formulas (16)-(19) to obtain the optimized GRU prediction model:

[0073]

[0074]

[0075] In industrial processes, the main function of air separation devices is to separate O2 and N2 from the air while extracting rare gases such as Ar. They are widely used in chemical, metallurgical, and military fields. Figure 3 The process flow of the air separation unit is demonstrated, with five highly integrated distillation columns used to extract N2, O2, and Ar from compressed air. First, compressed air is fed into a pressurized column via a main heat exchanger. The gas at the bottom of the pressurized column passes through a secondary condenser, serving as a condenser for the crude Ar column (2) and the pure Ar column. The gas is then fed into a low-pressure column, causing N2 to precipitate at its bottom. Of the N2 at the top of the pressurized column, part passes through the main and secondary condensers to obtain liquid N2 product, while the other part is sent to the main heat exchanger to obtain N2 product. The liquid O2 at the bottom of the low-pressure column is pressurized by the main heat exchanger to obtain N2 products at a concentration of 12 kg / cm³. 3 Medium-pressure O2 products and 32kg / cm 3 High-pressure O2 product. Ar is purified after passing through crude Ar tower 1, crude Ar tower 2 and pure Ar tower.

[0076] In actual production, the O2 concentration in the air separation unit has a significant impact on production safety and product quality. When the O2 concentration in the unit is too low, a large amount of N2 accumulates at the inlet of the No. 1 crude Ar column, and the pressure at the top of the No. 2 crude Ar column gradually increases, which can easily lead to column collapse. When the O2 concentration is too high, insufficient Ar concentration will result in a decrease in Ar production. Therefore, it is crucial to monitor the O2 concentration in the air separation unit in real time and maintain it within a certain range. Figure 3 A is an online analyzer installed in an air separation device for real-time detection of O2 concentration. However, traditional online analyzers have drawbacks such as short lifespan and susceptibility to environmental influences. Soft measurement is required to correct the results and ensure accuracy.

[0077] The present invention proposes a sensitivity analysis based on variance decomposition and a GRU neural network soft sensor modeling algorithm using NNG. Experimental data is obtained from the process data acquisition system of an air separation unit. The advantages of this invention will be illustrated below with an example of O2 concentration prediction in an air separation unit. The main idea of ​​this invention is to preliminarily determine 25 measurable input variables affecting the O2 concentration in the air separation unit through process mechanism analysis (see Table 1), and then use the proposed soft sensor algorithm to select input variables highly correlated with the target variable, and establish a time series model to predict the O2 concentration in the air separation unit.

[0078] Table 1. Candidate input variables for soft-sensor modeling of O2 concentration in air separation unit.

[0079]

[0080]

[0081] The O2 concentration prediction results of different algorithms are shown in Table 2. Compared with the traditional GRU algorithm, the O2 concentration prediction models of the air separation device established by the NNG-GRU and SANNG-GRU algorithms based on variable selection have improved in all performance indicators. Furthermore, the SANNG-GRU model selects the fewest input variables, has the highest correlation coefficient, and its mean square error is significantly lower than that of the NNG-GRU algorithm, which verifies the superiority of the proposed algorithm. Figure 4 The O2 concentration prediction curve of the SANNG-GRU soft measurement modeling method shows that the proposed algorithm has a good fitting effect on the test set, and the actual value and the predicted value almost overlap, which further illustrates the effectiveness of the proposed method.

[0082] Table 2 Predictive Performance of Different Soft Sensing Modeling Methods

[0083]

[0084] In summary, the sensitivity analysis based on variance decomposition and the GRU neural network soft sensor modeling algorithm based on NNG fully utilizes the time series feature extraction capability of GRU and the variable selection capability of NNG. Furthermore, by embedding the total sensitivity index into the NNG algorithm, adaptive variable selection of the GRU model is achieved. This algorithm overcomes the biased coefficient estimation drawback of the NNG algorithm, reduces the complexity of the GRU model, and effectively improves the prediction accuracy of O2 concentration in the air separation device, providing technical support for subsequent system optimization and control.

[0085] Although specific embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these specific embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A soft measurement method for predicting O2 concentration in an air separation device, characterized in that... Includes the following steps: S1. Obtain historical dataset: Sample the air separation device data at time intervals T to form a historical sample dataset D={X,Y}; S2. Data preprocessing: The dataset D={X,Y} is preprocessed using a standardization method to eliminate the dimensional relationships between data. S3. Construct a basic GRU network model: Use 80% of the dataset D={X,Y} as the training set and the remaining 20% ​​as the test set to train a basic GRU network model. S4. Using sensitivity analysis based on variance decomposition and the adaptive variable selection algorithm of the GRU model of NNG, soft measurement modeling of O2 concentration in air separation device is realized. In step S4, during the variable selection process, the NNG algorithm uses sensitivity analysis based on variance decomposition to calculate the total sensitivity index and uses the total sensitivity index as the adaptive weight vector of the NNG algorithm. The steps for obtaining the total sensitivity index are as follows. X A and X B Given two independent input variable matrices, X A and X B The sizes are equal and the number of rows is , X =[ X A ; X B ], j and i Represent X A , X B Row index and column index, f To represent a trained GRU model, define the matrix. Indicates will X A No. i Column replacement X B The i The new matrix after the column, for: The overall sensitivity index is expressed as: ; Construct an adaptive weight vector: Calculate the relative importance factor of each input variable relative to the output variable based on the overall sensitivity index. ； μ i As the adaptive weight vector of the NNG algorithm, the GRU model is introduced to achieve adaptive variable selection. First, the shrinkage operator θ=[θ1, θ2, … , θ K Introducing input weights into the GRU model In this process, the basic unit expression of the optimized GRU model is obtained: Secondly, the relative importance factor μ i The NNG algorithm is introduced, and constrained quadratic optimization is performed. Obtain the optimal shrinkage coefficient θ =[θ1 ,θ2 ,…,θ K ], where s is the strangulation parameter, Finally, θ =[θ1 ,θ2 ,…,θ K Adding this to the basic unit expression of the GRU model yields the optimized GRU prediction model: 。 2. The soft measurement method for predicting O2 concentration in an air separation device according to claim 1, characterized in that, In step S1, no less than 6,000 data points are collected during sampling.

Images