A krypton-xenon refining method and device based on temperature and pressure multi-source data fusion

Through temperature and pressure multi-source data fusion and MXC-LSTM network prediction, the problem of abnormal temperature and pressure abnormality during krypton xenon refining is solved, and the reliability and efficiency of krypton xenon refining process is improved.

CN118197448BActive Publication Date: 2024-12-03HANGZHOU ZHONGTAI HYDROGEN TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410304105.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-18
Publication Date
2024-12-03
Estimated Expiration
2044-03-18

AI Technical Summary

Technical Problem

During the existing krypton xenon refining process, it is difficult to timely detect abnormal temperature pressure trends in the refining tower, which affects the purity and refining efficiency of krypton xenon.

Method used

Using a method based on temperature pressure multi-source data fusion, data is collected through pressure sensors and temperature sensors, normalization, outlier processing, missing value processing, multi-source data fusion and wavelet decomposition feature extraction, and input the attention mechanism The MXC-LSTM network is trained to predict temperature and pressure, and sets a threshold to issue an alarm in advance to prevent abnormalities.

Benefits of technology

An early warning of abnormal temperature pressure in the refining tower is achieved, the reliability and efficiency of the krypton xenon refining process is improved, and the high purity production of krypton xenon is ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118197448B_ABST
    Figure CN118197448B_ABST
Patent Text Reader

Abstract

The invention discloses a krypton-xenon refining method and device based on temperature and pressure multi-source data fusion, which is carried out according to the following steps: step S1: a pressure sensor and a temperature sensor collect pressure and temperature data in a refining tower; step S2: the temperature data and the pressure data are normalized, outlier processing is processed, missing value processing is processed, multi-source data fusion is processed, and wavelet decomposition feature extraction is processed to obtain a data set, and the data set is divided to obtain a training set and a verification set; step S3: the training set and the verification set are input into an MXC-LSTM network with an attention mechanism for training, and the trained MXC-LSTM network with an attention mechanism outputs a predicted temperature T pre and predicted pressure F pre , and record the training weights at the same time; the present invention can obtain the predicted temperature and predicted pressure based on the current temperature and pressure in the refining tower, so as to facilitate the early discovery of abnormal trends in temperature and pressure in the refining tower and assist in subsequent production.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of rare gas refining, and in particular to a krypton-xenon refining method and device based on temperature and pressure multi-source data fusion. Background Art

[0002] Krypton-xenon is a rare gas with very stable chemical properties. It is not easy to react with other substances. Its content in nature is extremely rare, with a content of only about 0.000087% in the air. As an inert gas, it plays a role in protecting objects and preventing oxidation in the fields of semiconductors, automobile industry, and medical treatment. After crude extraction, krypton-xenon needs to enter a refining tower for refining to obtain high-purity krypton-xenon. The existing refining process generally involves pressurizing and vaporizing the krypton-xenon concentrate and then adjusting it to a suitable temperature to allow the oxygen and methane doped therein to react chemically to achieve purification. The pressure and temperature in the process directly affect the purity of krypton-xenon, and it is necessary to promptly detect abnormal temperature and pressure trends in the refining tower. Summary of the invention

[0003] The purpose of the present invention is to provide a krypton-xenon refining method and device based on temperature and pressure multi-source data fusion. The present invention can obtain predicted temperature and predicted pressure based on the current temperature and pressure in the refining tower, so as to facilitate the early detection of abnormal temperature and pressure trends in the refining tower and assist subsequent production.

[0004] The technical solution of the present invention is a krypton-xenon refining method based on temperature and pressure multi-source data fusion, which is carried out according to the following steps:

[0005] Step S1: The pressure sensor and the temperature sensor collect the pressure and temperature data in the refining tower;

[0006] Step S2: The temperature data and pressure data are normalized, outlier processed, missing value processed, multi-source data fused, and wavelet decomposition feature extracted to obtain a data set, and the data set is divided to obtain a training set and a validation set;

[0007] Step S3: The training set and the validation set are input into the MXC-LSTM network with attention mechanism for training, and the trained MXC-LSTM network with attention mechanism outputs the predicted temperature T pre and predicted pressure F pre , while recording the training weights;

[0008] Step S4: Setting the temperature threshold T in the refining tower set and pressure threshold F set , if the predicted temperature T pre Meet T set -0.5<T pre <T set +0.5, while predicting pressure F preMeet F set -0.5<F pre <F set +0.5, the refining tower operates normally to achieve krypton-xenon refining. If the temperature threshold T set Or pressure threshold F set If any of the above conditions are not met, the corresponding temperature alarm and pressure alarm will sound an alarm in advance;

[0009] Step S5: After the alarm is issued, the data is rechecked for outliers and missing values. If they exist, outliers are removed, missing values ​​are supplemented, and the data set is enhanced, and steps S1 to S4 are repeated. If they do not exist, retrain and update the training weights until the refining tower operates normally to achieve krypton-xenon refining.

[0010] In the above-mentioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the normalization process in step S2 is to firstly make the temperature data and the pressure data dimensionless:

[0011]

[0012] Where L is the measured value, L' is the value after transformation, and S is the transformation method based on standard deviation:

[0013] Then standardize:

[0014]

[0015] The processed data is normalized to [0, 1], and the original data is mapped to the [0, 1] interval according to the linear transformation method. The formula is as follows:

[0016]

[0017] Among them, max(L) is the maximum value of L, and min(L) is the minimum value of L.

[0018] In the above-mentioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the abnormal value processing in step S2 adopts the Laida criterion in the hypothesis testing method, assuming that the sample data are x1, x2, ... x n , the average value is The deviation is The standard deviation is calculated according to the Bessel formula, which is as follows:

[0019]

[0020] If the sample data x i Deviation v i Satisfy:|v i |>3σ, then x iIt is abnormal data and should be removed.

[0021] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the missing value processing in step S2 adopts single interpolation, and the model of single interpolation is expressed as follows:

[0022] Y i =f(x i )+ε i ;

[0023] Among them, Y i is the interpolation value, x i is the auxiliary variable of the ith data, f is the auxiliary variable function, ε i is the residual;

[0024] Missing values ​​are filled by single mean interpolation, using the mean of valid data in the data unit to replace the value of the vacant part, and all missing values ​​are replaced by the mean of the overall data. The formula is as follows:

[0025]

[0026] Among them, yi is all valid data, and n is the total number of all data.

[0027] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the multi-source data fusion processing in step S2 is performed using a batch estimation arithmetic average method: first, multiple sensors of the same type are divided into two groups of odd-numbered measuring points and even-numbered measuring points according to their position distribution in the refining tower. The average value formula of the odd-numbered measuring points and the even-numbered measuring points is as follows:

[0028]

[0029] Among them, T1 is the measurement point data of odd-numbered measurement points, T2 is the measurement point data of even-numbered measurement points, m is the number of measurement points of odd-numbered measurement points, and n is the total number of measurement points;

[0030] Then calculate the standard deviation of odd and even measurement points, the formula is as follows:

[0031]

[0032] Finally, the fusion average is calculated, and the formula is as follows:

[0033]

[0034] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the wavelet decomposition feature extraction process of step S2 is to decompose the original data x(t) of pressure and temperature, filter through the filter to obtain high-frequency signals and low-frequency signals, retain the high-frequency signal, and continue to decompose the low-frequency signal into a high-frequency signal and a low-frequency signal, repeat the decomposition action until the designed decomposition layer number is reached. In the decomposition process, as the number of decompositions changes, the decomposition result obtained is:

[0035]

[0036]

[0037] Among them, ψ m (·) is the time response of the filter during filtering, N is the number of samples detected by the sensor, and n is the number of layers of wavelet decomposition;

[0038] Wavelet analysis separates the pressure and temperature signals detected by the sensor into a series of wavelet functions ψ(t) obtained by translation or scaling of the mother wavelet function, and uses the dbN wavelet function for decomposition, where N is the order of the wavelet, the support area of ​​the wavelet function ψ(t) and the scaling function is 2N-1, and the vanishing moment of ψ(t) is N. After wavelet decomposition, the detail coefficients and approximate coefficients of each signal are obtained, and the features of the decomposed approximate coefficients of pressure and temperature are extracted to form a feature vector.

[0039] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the MXC-LSTM network with attention mechanism in step S3 is constructed according to the following steps:

[0040] Step S3.1: Build the LSTM network architecture and use the multi-objective arithmetic optimization algorithm to optimize the number of neurons and learning rate of the hidden layer of the LSTM network;

[0041] Step S3.2: Use the XGBoost algorithm to establish a direct multi-step prediction model based on the LSTM network and combine it with CNN neural network optimization to have the ability to optimize and extract the input feature vector;

[0042] Step S3.3: Add attention mechanism in LSTM network to improve prediction accuracy and stability.

[0043] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the construction and optimization process of step S3.1 is carried out according to the following steps:

[0044] Step S3.1.1: Build the LSTM network architecture, set the hidden layer to a single layer, set the number of training generations, initialize the number of hidden layer neurons and the learning rate;

[0045] Step S3.1.2: Initialize the parameters of the multi-objective arithmetic optimization algorithm, including the initial position of the population, the archive capacity, the spatial dimension, the population size, and the number of iterations;

[0046] Step S3.1.3: The parameters to be optimized are formed into corresponding individuals, the fitness value of the individuals is calculated according to the optimized objective function, the dominance relationship is determined to obtain the optimal position of the individuals, and an external archive is created;

[0047] Step S3.1.4: Randomly select an individual as the global optimal individual from the external archive in a grid with low crowding density according to the roulette strategy;

[0048] Step S3.1.5: Execute the AOA arithmetic optimization process, update the individual positions, and obtain a new generation of population;

[0049] Step S3.1.6: Repeat steps S3.1.3 to S3.1.5. When the maximum number of iterations is reached, stop the iteration and obtain the Pareto solution set.

[0050] Step S3.1.7: Select the optimal hyperparameters of the LSTM network through fuzzy decision making based on the Pareto solution set, then input the training set into the LSTM network for learning and training, and use it to predict the test set data.

[0051] In the aforementioned krypton-xenon refining method based on temperature and pressure multi-source data fusion, the multi-step prediction model establishment process of step S3.2 is carried out according to the following steps:

[0052] Step S3.2.1: Initialize the XGBoost algorithm parameters, input the training sample set, set the number of iterations, and regularize the objective function;

[0053] Step S3.2.2: Start iteration, transform the objective function and learn the first tree to predict the sample value;

[0054] Step S3.2.3: Expand the loss function using the second-order derivative Taylor rule;

[0055] Step S3.2.4: Ignore the constant term to simplify the objective function;

[0056] Step S3.2.5: splitting the nodes of the tree by scoring the tree structure;

[0057] Step S3.2.6: If the iteration requirements are met, stop growing and generate a decision tree. If not, repeat steps SS3.2.2 to S3.2.5.

[0058] The CNN neural network includes three convolutional layers and two maximum pooling layers that are alternately connected. The last convolutional layer is connected to a flat layer for output. The decision tree of the multi-step prediction model is applied to the flat layer to optimize the input data.

[0059] The aforementioned krypton-xenon refining device based on temperature and pressure multi-source data fusion includes a hydraulic system for pressure regulation arranged in the refining tower, the hydraulic system includes an asymmetric axial piston pump, a hydraulic accumulator and a differential cylinder, the asymmetric axial piston pump has a pump port A, a pump port B and a pump port T, the pump port A is connected to the rodless chamber of the differential cylinder, the pump port B is connected to the rod chamber of the differential cylinder, and the pump port T is connected to the hydraulic accumulator.

[0060] Compared with the prior art, the present invention firstly extracts the temperature data of the collected pressure data set reliably through normalization, outlier processing, missing value processing, multi-source data fusion, and wavelet decomposition feature extraction processing, and inputs it into the MXC-LSTM network with attention mechanism for high-precision and high-stability training to obtain the predicted temperature T pre and predicted pressure F pre , by judging the predicted temperature T pre and predicted pressure F pre Whether the temperature of the refining tower exceeds the set threshold value T set and pressure threshold F set To sound an alarm and achieve early warning, so that the operator can intervene in the control in advance and improve the reliability of the krypton-xenon refining process. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the process of the present invention;

[0062] Figure 2 It is a schematic diagram of the decomposition process of wavelet decomposition feature extraction;

[0063] Figure 3 This is a flow chart of the LSTM network;

[0064] Figure 4 This is the structural diagram of the MXC-LSTM network;

[0065] Figure 5 It is a flowchart of the AOA algorithm;

[0066] Figure 6 This is a flowchart of the XGBoost algorithm;

[0067] Figure 7 This is a schematic diagram of XGBoost tree ensemble;

[0068] Figure 8 It is the unit structure diagram of the attention mechanism;

[0069] Fig. 9 It is a structural diagram of the hydraulic system.

[0070] The markings in the attached drawings are: 1-hydraulic system; 2-asymmetric axial piston pump; 3-hydraulic accumulator; 4-differential cylinder; 5-overflow valve; 6-check valve; 7-oil tank. DETAILED DESCRIPTION

[0071] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0072] Embodiment: A krypton-xenon refining method based on temperature and pressure multi-source data fusion, as shown in the attached Figure 1 As shown, follow the steps below:

[0073] Step S1: The pressure sensor and the temperature sensor collect the pressure and temperature data in the refining tower;

[0074] Step S2: The temperature data and pressure data are normalized, outlier processed, missing value processed, multi-source data fused, and wavelet decomposition feature extracted to obtain a data set, and the data set is divided to obtain a training set and a validation set;

[0075] The normalization process in step S2 is to first make the temperature data and the pressure data dimensionless:

[0076]

[0077] Where L is the measured value, L' is the value after transformation, and S is the transformation method based on standard deviation:

[0078] Then standardize:

[0079]

[0080] The processed data is normalized to [0, 1], and the original data is mapped to the [0, 1] interval according to the linear transformation method. The formula is as follows:

[0081]

[0082] Among them, max(L) is the maximum value of L, and min(L) is the minimum value of L.

[0083] Outlier processing means that when processing data, highly abnormal outliers should be eliminated. Whether outliers should be eliminated depends on the reasons for their occurrence and whether they should be discarded. Usually, there are several reasons for outliers: data input errors, measurement errors, experimental errors, data processing errors, etc. In the actual production of refining equipment, outliers will be generated due to serious errors in technical personnel's records or sudden fluctuations in instruments. Even if there is only a small part of abnormal data, it will affect the structure of the entire data sample and even lead to poor modeling results of the later data mining algorithm. Therefore, abnormal data must be eliminated before data mining. For outlier processing, the Laida criterion in the hypothesis testing method is used. Suppose the sample data is x1, x2, ... x n , the average value is The deviation is The standard deviation is calculated according to the Bessel formula, which is as follows:

[0084]

[0085] If the sample data x i Deviation v i (1≤i≤n) satisfies:|v i |>3σ, then x i It is abnormal data and should be removed.

[0086] Missing value processing uses single interpolation, and the model of single interpolation is expressed as follows:

[0087] Y i =f(x i )+ε i ;

[0088] Among them, Y i is the interpolation value, x i is the auxiliary variable of the ith data, f is the auxiliary variable function, ε i is the residual;

[0089] Missing values ​​are filled by single mean interpolation, using the mean of valid data in the data unit to replace the value of the vacant part, and all missing values ​​are replaced by the mean of the overall data. , the formula is as follows:

[0090]

[0091] Among them, y i is all valid data, and n is the total number of all data.

[0092] The multi-source data fusion processing is carried out by batch estimation arithmetic mean method. First, multiple sensors of the same type are divided into two groups of odd-numbered measuring points and even-numbered measuring points according to their position distribution in the refining tower. The average value formula of odd-numbered measuring points and even-numbered measuring points is as follows:

[0093]

[0094] Among them, T1 is the measurement point data of odd-numbered measurement points, T2 is the measurement point data of even-numbered measurement points, m is the number of measurement points of odd-numbered measurement points, and n is the total number of measurement points;

[0095] Then calculate the standard deviation of odd and even measurement points, the formula is as follows:

[0096]

[0097] Finally, the fusion average is calculated, and the formula is as follows:

[0098]

[0099] As attached Figure 2 As shown in the figure, the wavelet decomposition feature extraction process is to decompose the original data x(t) of pressure and temperature, filter out high-frequency signals and low-frequency signals through filters, retain the high-frequency signal, and continue to decompose the low-frequency signal into a high-frequency signal and a low-frequency signal. Repeat the decomposition action until the designed decomposition layer number is reached. In the decomposition process, as the number of decompositions changes, the decomposition results are:

[0100] The wavelet decomposition result after point C is:

[0101] Among them, ψ m (·) is the time response of the filter during filtering, N is the number of samples detected by the sensor, and n is the number of layers of wavelet decomposition;

[0102] Wavelet analysis separates the pressure and temperature signals detected by the sensor into a series of wavelet functions ψ(t) obtained by translation or scaling of the mother wavelet function, and uses the dbN wavelet function for decomposition, where N is the order of the wavelet, the support area of ​​the wavelet function ψ(t) and the scaling function is 2N-1, and the vanishing moment of ψ(t) is N. After wavelet decomposition, the detail coefficients and approximate coefficients of each signal are obtained, and the features of the decomposed approximate coefficients of pressure and temperature are extracted to form a feature vector for input.

[0103] Step S3: The training set and the validation set are input into the MXC-LSTM network with attention mechanism for training, and the trained MXC-LSTM network with attention mechanism outputs the predicted temperature T pre and predicted pressure F pre , while recording the training weights;

[0104] As attached Figure 4 As shown, the MXC-LSTM network with attention mechanism is built according to the following steps:

[0105] Step S3.1: Build the LSTM network architecture and use the multi-objective arithmetic optimization algorithm to optimize the number of neurons and learning rate of the hidden layer of the LSTM network;

[0106] As attached Figure 3 As shown, the process of LSTM protecting and controlling the cell state through three gate structures is as follows:

[0107] (1) Forget gate: determines which information is forgotten.

[0108] f t =σ(W f ·[h t-1 ,x t ]+b f ), where w represents the weight matrix and b represents the bias.

[0109] (2) Input gate: determines what information is stored in the cell state;

[0110]

[0111] (3) Output gate: determines which information of the cell state is output;

[0112]

[0113] Among them, C t-1 and C t Respectively represent the cell shape at time t-1 and time t; h t-1 and h t Respectively represent the module output at time t-1 and time t; σ represents the sigmoid layer; f t represents the output of the forget gate; x t represents external input; tanh represents the hyperbolic tangent activation function.

[0114] The construction and optimization process of step S3.1 is carried out as follows:

[0115] Step S3.1.1: Build the LSTM network architecture, set the hidden layer to a single layer, set the number of training epochs to 100, and initialize the number of hidden layer neurons h and the learning rate α;

[0116] Step S3.1.2: Initialize the parameters of the multi-objective arithmetic optimization algorithm, including the initial position of the population, the archive capacity, the spatial dimension, the population size, and the number of iterations;

[0117] Step S3.1.3: The parameters to be optimized are formed into corresponding individuals, and the optimization objective function is used to optimize the parameters. and Calculate individual fitness values, determine dominance relationships to obtain optimal individual positions, and create external archives;

[0118] Step S3.1.4: Randomly select an individual as the global optimal individual from the external archive in a grid with low crowding density according to the roulette strategy;

[0119] Step S3.1.5: Execute the AOA arithmetic optimization process, update the individual positions, and obtain a new generation of population;

[0120] Step S3.1.6: Repeat steps S3.1.3 to S3.1.5. When the maximum number of iterations is reached, stop the iteration and obtain the Pareto solution set.

[0121] Step S3.1.7: Select the optimal hyperparameters of the LSTM network through fuzzy decision making based on the Pareto solution set, then input the training set into the LSTM network for learning and training, and use it to predict the test set data.

[0122] As attached Figure 5 As shown in the figure, the AOA algorithm uses multiplication and division operations for global exploration to ensure global dispersion when updating the position; and uses addition and subtraction operations for local development to ensure local accuracy when updating the position. The specific steps are as follows:

[0123] (1) Update the mathematical acceleration optimizer, the formula is as follows:

[0124]

[0125] Among them, t represents the current number of iterations; T represents the maximum number of iterations set; Max represents the upper boundary of the mathematical acceleration optimizer MOA value, and Max is 1; Min represents the upper and lower boundaries of the mathematical acceleration optimizer MOA value, and Min is 0.2

[0126] Update the mathematical probability optimizer, the formula is as follows:

[0127]

[0128] Wherein, α represents the control parameter, and α=5;

[0129] (2) updating individual positions based on the mutual transformation between global exploration and local exploitation;

[0130] The random number r1 created in the range [0,1] is used to switch between global exploration and local development. iWhen \(MOA\), \(AOA\) enters the global exploration stage and updates the position according to the formula When \(r\) i When \(<MOA\), \(AOA\) enters the local development stage and updates the position according to the formula Update the position;

[0131] Repeat the above process until the iteration stops

[0132] Among them, \(x\) i,j (t + 1) represents the value of the \(i\)-th dimension position of the \(i\)-th solution in one iteration; \(best(x\) j ) represents the best solution of the \(i\)-th dimension position obtained so far; \(UB\) j represents the upper bound value of the \(j\)-th dimension position; \(LB\) j represents the lower bound value of the \(j\)-th dimension position; \(\mu\) represents the control parameter for adjusting the search process, with a fixed value of 0.5, and \(r2\), \(r3\) represent random numbers in \([0, 1]\).

[0133] The multi-objective arithmetic optimization algorithm is the multi-objective version of \(AOA\). To perform multi-objective optimization, \(AOA\) needs to be upgraded using the same two mechanisms as in multi-objective particle swarm optimization (MOPSO); the first upgrade is to set an archive containing all non-dominated Pareto-optimal solutions obtained so far. Among them, the archive is updated according to the following rules:

[0134] (1) If there is no solution saved in the archive, put the Pareto solutions of the current population into it;

[0135] (2) If there are already solutions saved in the archive, compare the Pareto solutions of the current population with the solutions in the archive. If the new solution is dominated by a certain solution in the archive, discard the new solution;

[0136] (3) If the new solution is not dominated by any solution in the archive, put the new solution into the archive;

[0137] (4) If the new solution can dominate a certain solution in the archive, delete that solution from the archive.

[0138] When the archive capacity reaches the upper limit, use the grid mechanism to randomly select and eliminate the existing solutions in the area with a large crowding density to create space for the new solutions to be stored.

[0139] The second upgrade is that the leader selection mechanism changes. Different from the single-objective optimization algorithm, the population optimal position of the multi-objective optimization algorithm is obtained from the archive. If there is more than one solution in the archive, the least crowded area is selected according to the roulette wheel strategy, and a new leader is randomly selected from the non-dominated solutions it contains. The probability that each area may generate a leader is calculated by the following formula:

[0140]

[0141] Where c is a constant greater than 1, N i is the number of particles contained in region i.

[0142] Step S3.2: Use the XGBoost algorithm to establish a direct multi-step prediction model based on the LSTM network and combine it with CNN neural network optimization to have the ability to optimize and extract the input feature vector;

[0143] XGBoost is a tree learning algorithm that uses decision trees as the basic model unit. The decision forest composed of multiple decision trees is the final learning model of XGBoost. Its model basic unit is the classification and regression tree (CART). XGBoost is a tree ensemble model that combines multiple CART base classifiers into a strong classifier. Figure 7 As shown, K is the number of trees (model selection times), and G is the maximum depth of the tree

[0144] As attached Figure 6 As shown in the figure, the multi-step prediction model building process is carried out in the following steps:

[0145] (1) Regularization: Assume that the given dataset D = {(X i ,y i )}(|D|=n,X i ∈R m ,y i ∈R), the data set D contains n samples and m features, where |D| is the number of samples; X i is the feature vector of the i-th sample; y i is the true value of the i-th sample;

[0146] The prediction output of the XGBoost ensemble model consisting of K trees is as follows:

[0147]

[0148] Where, F = {f(X) = ω q(X)}(q:R m →T,ω q(X) ∈R T ) is the function space of CART, q is the structure of each tree, T is the number of leaf nodes, X=(X1,X2,...,X n )

[0149] XGBoost's regularization objective function:

[0150] Where: loss function term Regularization term The regularization term is a penalty term for model complexity. Its function is to avoid overfitting. is the output value of the prediction model, y i is the true value of the sample, f k is the k-th tree model, It is the sum of the losses of each sample. The loss function of XGBoost can be set manually. T is the number of leaf nodes, γ is the regularization term of the leaf nodes of the tree, ω is the leaf node weight value of the kth tree, and λ ensures that the score of the leaf node does not exceed a certain range. Both parts of the penalty term can prevent overfitting.

[0151] (2) Iterative solution

[0152] The XGBoost model is built by iteratively adding trees. is the predicted value of the i-th sample at the t-th generation. The model minimizes the objective function by continuously iteratively adding trees:

[0153]

[0154] In order to quickly optimize the target, the second-order Taylor expansion of the above formula is used to obtain the following formula

[0155]

[0156] Where: g i is the first-order derivative of the loss function; h i is the second-order derivative of the loss function. In the process of adding the t-th tree, the previous t-1 trees have been trained, that is, is a constant term. Removing this term yields the simplified objective function of the tth step:

[0157]

[0158] Definition I j ={i|q(X i )=j} is the sample set of leaf node j. By expanding the regularization term Ω, the simplified objective function Written as:

[0159]

[0160] Where ω j is the weight of leaf node j, and the optimal weight of the leaf node is obtained by derivation As shown below:

[0161]

[0162] in,

[0163] The corresponding optimal objective function value is shown as follows:

[0164]

[0165] Split the tree nodes and determine the iteration process

[0166]

[0167] When the loss function decreases by a value greater than γ, the tree can be split. The γ parameter added to the denominator when calculating the gain can make it smooth. At the same time, γ can help prevent overfitting and is optional.

[0168] The CNN neural network is a convolutional neural network, which includes three convolutional layers and two maximum pooling layers connected alternately. The last convolutional layer is connected to the flat layer for output. The decision tree of the multi-step prediction model is applied to the flat layer to optimize the input data. The input feature of the CNN-LSTM model is constructed as (11, 59, 1). The feature extraction layer, that is, the convolutional layer (Conv2D) of the CNN layer part, extracts and maps features through the convolution kernel. The maximum pooling layer (MaxPooling2D) is used to compress the network parameters to reduce the model complexity and prevent overfitting. The number of convolution kernels of the CNN network part is set to 64, 128 respectively. , 256, the convolution kernel size of the first two convolution layers is set to 3, the convolution step is 1, the convolution layer uses the Relu activation function, the pooling layer uses the maximum pooling, the size is set to 2, the step size of the first pooling layer is 1, the second layer is set to 2, and the convolution kernel size of the third layer is set to (3, 27). After three convolution operations, a (1,1,256) three-dimensional feature vector array is finally output, and then through the flatten layer (Flatten) operation, the three-dimensional vector feature array is finally compressed into a one-dimensional time series vector feature with a length of 1x256. The extracted feature vector is input into the LSTM network for prediction and the temperature and pressure prediction results are directly output.

[0169] Step S3.3: Add an attention mechanism to the LSTM network to improve prediction accuracy and stability; the unit structure of the attention mechanism is shown in the attached figure. Figure 8 As shown, the expression formula of the attention mechanism is as follows:

[0170] S ti =V tanh(Wh t +Uh i +b),i=1,2,3,…,t-1

[0171]

[0172]

[0173] h′ t =f(F,h t ,y t );

[0174] Where: a ti Output value h for the hidden layer of the MXC-LSTM network t Attention weight value for the current input; y1,y2,y3,…,y t is the input sequence; h1,h2,h3,…,h t corresponds to the input sequence y1,y2,y3,…,y t The hidden layer state value, that is, h t corresponds to the input y t The hidden layer state value; h t is the final feature vector; V, w, U, and b are the learning parameters of the model.

[0175] Step S4: Setting the temperature threshold T in the refining tower set and pressure threshold F set , if the predicted temperature T pre Meet T set -0.5<T pre <T set +0.5, while predicting pressure F pre Meet F set -0.5<F pre <F set +0.5, the refining tower operates normally to achieve krypton-xenon refining. If the temperature threshold T set Or pressure threshold F set If any of the above conditions are not met, the corresponding temperature alarm and pressure alarm will sound an alarm in advance;

[0176] Step S5: After the alarm is issued, the data is rechecked for outliers and missing values. If they exist, outliers are removed, missing values ​​are supplemented, and the data set is enhanced, and steps S1 to S4 are repeated. If they do not exist, retrain and update the training weights until the refining tower operates normally to achieve krypton-xenon refining.

[0177] Taking temperature information as an example, the model prediction result index method adopted in this embodiment is as follows:

[0178] Mean absolute error (MAE), the formula is as follows

[0179]

[0180] Among them, y t Indicates the actual temperature value. Represents the predicted temperature value, N is the number of samples, and MAE represents the average absolute error between the actual power value and the predicted value, which can better reflect the deviation degree of the predicted value from the actual value. The smaller the MAE, the lower the deviation degree;

[0181] Root mean square error (RMSE), the formula is as follows

[0182]

[0183] RMSE represents the average value of the square root of the deviation between the actual power value and the predicted value, which can be used to measure the closeness between the actual power value and the predicted value. The smaller the RMSE, the higher the model accuracy.

[0184] Standard deviation error (SDE), the formula is as follows

[0185]

[0186] Among them, std represents the standard deviation, SDE represents the degree of dispersion between a set of error data, and reflects the fluctuation of the overall prediction error. The smaller the SDE, the more stable the model prediction.

[0187] The temperature prediction indicators of the four refining towers using the LSTM network, CNN-LSTM network, MAOA-LSTM network and the MCX-LSTM network of this embodiment are shown in the following table:

[0188]

[0189] It can be seen from the above table that the prediction effect of the embodiment provided by the method of the present invention is better than that of the other three traditional network structures.

[0190] The krypton-xenon refining device based on temperature and pressure multi-source data fusion is as shown in the attached Fig. 9As shown, it includes a hydraulic system 1 for pressure regulation arranged in the refining tower, the hydraulic system 1 includes an asymmetric axial piston pump 2, a hydraulic accumulator 3 and a differential cylinder 4, the asymmetric axial piston pump 2 has a pump port A, a pump port B and a pump port T, the pump port A is connected to the rodless cavity of the differential cylinder 4, the pump port B is connected to the rod cavity of the differential cylinder 4, and the pump port T is connected to the hydraulic accumulator 3. When the hydraulic system 1 is working, the pressure oil enters the rodless cavity of the differential cylinder 4 from the pump port A, and the differential cylinder 4 drives the load to rise; when the differential cylinder 4 descends, the pressure oil in the rodless cavity of the differential cylinder 4 enters the asymmetric axial piston pump 2 through the pump port A. The symmetrical axial piston pump 2, the pressure oil discharged through the pump port B enters the rod chamber of the differential cylinder 4 to drive the differential cylinder 4 to descend, and the pressure oil discharged through the pump port T enters the hydraulic accumulator 3 for use when the differential cylinder 4 rises again; when the differential cylinder 4 rises again, the pressure oil in the rod chamber of the differential cylinder 4 enters the asymmetrical axial piston pump 2 through the pump port B and the pressure oil in the hydraulic accumulator 3 through the T port to drive the differential cylinder 4 to rise. The direct pump control technology is used to realize the driving of the differential cylinder 4, avoiding the throttling loss caused by the use of valve control technology, reducing energy loss, and the hydraulic accumulator 3 The pressure oil when the differential cylinder 4 descends is stored for use when the hydraulic system 1 works again, which reduces energy loss, improves energy utilization efficiency, and achieves low-carbon energy saving and energy recovery and reuse in the gas refining process. The hydraulic energy storage system has a high power density and can store a large amount of energy in a small volume. It is suitable for occasions where the internal space of the rare gas refining tower is limited. In addition, the hydraulic energy storage system has the characteristics of fast response speed and can quickly release the stored energy. It is suitable for working conditions that require fast response (such as quickly switching to the control of the next refining tower); finally, the hydraulic energy storage system can reduce the energy consumption of the system and reduce energy waste by storing energy under zero load or light load conditions; by balancing the pressure pulsation of the hydraulic system, the hydraulic energy storage technology can reduce the operating frequency of the hydraulic pump and valve components in the system and extend the service life of each equipment in the refining tower; the rodless chamber of the differential cylinder, the rod chamber of the differential cylinder and the hydraulic accumulator are also connected to the oil tank 7 through the overflow valve 5 and the one-way valve 6. The overflow valve and the one-way valve are used to prevent the pressure overload generated by the hydraulic system and play a load unloading protection role.

[0191] In summary, the present invention can obtain predicted temperature and predicted pressure based on the current temperature and pressure in the refining tower, so as to facilitate early detection of abnormal trends of temperature and pressure in the refining tower and assist subsequent production.

Claims

1. A krypton-xenon refining method based on temperature and pressure multi-source data fusion, characterized in that: Follow these steps: Step S1: The pressure sensor and the temperature sensor collect the pressure and temperature data in the refining tower; Step S2: The temperature data and pressure data are normalized, outlier processed, missing value processed, multi-source data fused, and wavelet decomposition feature extracted to obtain a data set, and the data set is divided to obtain a training set and a validation set; Step S3: The training set and the validation set are input into the MXC-LSTM network with attention mechanism for training, and the trained MXC-LSTM network with attention mechanism outputs the predicted temperature T pre and predicted pressure F pre , while recording the training weights; Step S4: Setting the temperature threshold T in the refining tower set and pressure threshold F set , if the predicted temperature T pre Meet T set -0.5<T pre <T set +0.5, while predicting pressure F pre Meet F set -0.5<F pre <F set +0.5, the refining tower operates normally to achieve krypton-xenon refining. If the temperature threshold T set Or pressure threshold F set If any of the above conditions are not met, the corresponding temperature alarm and pressure alarm will sound an alarm in advance; Step S5: After the alarm is issued, the data is re-checked for abnormal values ​​and missing values. If they exist, the data is re-tested for abnormal values, missing values ​​are supplemented, and the data set is enhanced, and then steps S1 to S4 are repeated. If they do not exist, re-training is performed and the training weights are updated until the refining tower operates normally to achieve krypton-xenon refining; The MXC-LSTM network with attention mechanism in step S3 is constructed according to the following steps: Step S3.1: Build the LSTM network architecture and use the multi-objective arithmetic optimization algorithm to optimize the number of neurons and learning rate of the hidden layer of the LSTM network; Step S3.2: Use the XGBoost algorithm to establish a direct multi-step prediction model based on the LSTM network and combine it with CNN neural network optimization to have the ability to optimize and extract the input feature vector; Step S3.3: Add attention mechanism in LSTM network to improve prediction accuracy and stability; The multi-step prediction model building process of step S3.2 is carried out according to the following steps: Step S3.2.1: Initialize the XGBoost algorithm parameters, input the training sample set, set the number of iterations, and regularize the objective function; Step S3.2.2: Start iteration, transform the objective function and learn the first tree to predict the sample value; Step S3.2.3: Expand the loss function using the second-order derivative Taylor rule; Step S3.2.4: Ignore the constant term to simplify the objective function; Step S3.2.5: splitting the nodes of the tree by scoring the tree structure; Step S3.2.6: If the iteration requirements are met, stop growing and generate a decision tree. If not, repeat steps SS3.2.2 to S3.2.

5. The CNN neural network includes three convolutional layers and two maximum pooling layers that are alternately connected. The last convolutional layer is connected to the flattening layer for output. After three convolution operations, a three-dimensional feature vector array is finally output. Then, through the flattening layer operation, the three-dimensional vector feature array is finally compressed into a one-dimensional time series vector feature. The decision tree of the multi-step prediction model is applied to the flattening layer to optimize the input data. The extracted feature vector is input into the LSTM network for prediction, and then the temperature and pressure prediction results are directly output.

2. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The normalization process in step S2 is to first make the temperature data and the pressure data dimensionless: Where L is the measured value, L is the value after transformation, and S is the transformation method based on standard deviation; Then standardize: The processed data is normalized to [0, 1], and the original data is mapped to the [0, 1] interval according to the linear transformation method. The formula is as follows: Among them, max(L) is the maximum value of L, and min(L) is the minimum value of L.

3. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The outlier processing in step S2 adopts the Laida criterion in the hypothesis testing method. Suppose the sample data is x1, x2, ... x n , the average value is The deviation is The standard deviation is calculated according to the Bessel formula, which is as follows: If the sample data x i Deviation v i Satisfy:|v i |>3σ, then x i It is abnormal data and should be removed.

4. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The missing value processing in step S2 adopts single interpolation, and the model of single interpolation is expressed as follows: Y i =f(x i )+ε i ; Among them, Y i is the interpolation value, x i is the auxiliary variable of the ith data, f is the auxiliary variable function, ε i is the residual; Missing values ​​are filled by single mean interpolation, using the mean of valid data in the data unit to replace the value of the vacant part, and all missing values ​​are replaced by the mean of the overall data. The formula is as follows: Among them, yi is all valid data, and n is the total number of all data.

5. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The multi-source data fusion processing in step S2 is performed by batch estimation arithmetic averaging method: first, multiple sensors of the same type are divided into two groups of odd-numbered measuring points and even-numbered measuring points according to the position distribution in the refining tower. The average value formula of the odd-numbered measuring points and the even-numbered measuring points is as follows: Among them, T1 is the measurement point data of odd-numbered measurement points, T2 is the measurement point data of even-numbered measurement points, m is the number of measurement points of odd-numbered measurement points, and n is the total number of measurement points; Then calculate the standard deviation of odd and even measurement points, the formula is as follows: Finally, the fusion average is calculated, and the formula is as follows:

6. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The wavelet decomposition feature extraction process of step S2 is to decompose the original data x(t) of pressure and temperature, obtain high-frequency signals and low-frequency signals through filtering, retain the high-frequency signal, and continue to decompose the low-frequency signal into a high-frequency signal and a low-frequency signal, and repeat the decomposition action until the designed decomposition layer number is reached. In the decomposition process, as the number of decompositions changes, the decomposition result obtained is: Among them, ψ m (·) is the time response of the filter during filtering, N is the number of samples detected by the sensor, and n is the number of layers of wavelet decomposition; Wavelet analysis separates the pressure and temperature signals detected by the sensor into a series of wavelet functions ψ(t) obtained by translation or scaling of the mother wavelet function, and uses the dbN wavelet function for decomposition, where N is the order of the wavelet, the support area of ​​the wavelet function ψ(t) and the scaling function is 2N-1, and the vanishing moment of ψ(t) is N. After wavelet decomposition, the detail coefficients and approximate coefficients of each signal are obtained, and the features of the decomposed approximate coefficients of pressure and temperature are extracted to form a feature vector.

7. The krypton-xenon refining method based on temperature and pressure multi-source data fusion according to claim 1, characterized in that: The construction and optimization process of step S3.1 is carried out according to the following steps: Step S3.1.1: Build the LSTM network architecture, set the hidden layer to a single layer, set the number of training generations, initialize the number of hidden layer neurons and the learning rate; Step S3.1.2: Initialize the parameters of the multi-objective arithmetic optimization algorithm, including the initial position of the population, the archive capacity, the spatial dimension, the population size, and the number of iterations; Step S3.1.3: The parameters to be optimized are formed into corresponding individuals, the fitness value of the individuals is calculated according to the optimized objective function, the dominance relationship is determined to obtain the optimal position of the individuals, and an external archive is created; Step S3.1.4: Randomly select an individual as the global optimal individual from the external archive in a grid with low crowding density according to the roulette strategy; Step S3.1.5: Execute the AOA arithmetic optimization process, update the individual positions, and obtain a new generation of population; Step S3.1.6: Repeat steps S3.1.3 to S3.1.

5. When the maximum number of iterations is reached, stop the iteration and obtain the Pareto solution set. Step S3.1.7: Select the optimal hyperparameters of the LSTM network through fuzzy decision making based on the Pareto solution set, then input the training set into the LSTM network for learning and training, and use it to predict the test set data.