Aerospace mass spectrometer multi-state parameter prediction method based on ARIMA-LSTM combination model

Abstract

The invention discloses a spaceflight mass spectrometer multi-state parameter prediction method based on an ARIMA-LSTM combination model, and belongs to the field of aerospace, and the method comprises the steps: collecting six state parameters of a spaceflight mass spectrometer, and building a data set according to the collected state parameter data; dividing into a training set and a test set; eMD mode decomposition is carried out on the training set, and a plurality of linear eigenmode functions of a linear part of the data and a nonlinear residual error of a nonlinear part of the data are extracted; the linear eigenmode function is used for training an ARIMA model and predicting a data linear part; the nonlinear residual part is used for training an LSTM model and predicting the nonlinear part of the data; the test set is divided into two subsets, the first subset selects 500 pieces of data, the weight of the combined model is determined through a grey wolf optimization algorithm according to the prediction error, and the second subset is used for verifying the model prediction effect; and fusing the LSTM model and the ARIMA model according to the combined model weight. By adopting the method, the prediction precision is improved, and the method is suitable for different state parameters.

Description

Technical Field

[0001] The present invention relates to the field of aerospace technology, and in particular to a method for predicting multi-state parameters of an aerospace mass spectrometer based on an ARIMA-LSTM combined model. Background Art

[0002] The mass spectrometer uses mass spectrometry to obtain the volume fraction of trace harmful gases in spacecraft. The basic principle is to produce ions of different charge-to-mass ratios in the gas to be tested by ionization, separate ions of different charge-to-mass ratios and measure the signal intensity of different ions to generate a mass spectrum, and conduct qualitative or quantitative analysis of the gas to be tested. With the development of space science and technology, China, the United States, the Soviet Union, Europe and other countries in the world have frequently carried out deep space exploration and manned space research. Astronauts need to stay in orbit for a long time. The mass spectrometer is an instrument for monitoring the volume fraction of trace harmful gases in the confined space where astronauts work and live for a long time. Predicting multiple state parameters to ensure the reliable operation of the mass spectrometer is of great significance to ensuring the safety of astronauts and the smooth completion of space missions. During the operation of the mass spectrometer, the monitoring data of each state parameter is usually a complex non-stationary and nonlinear time series, and the different state parameters of different mass spectrometer components have completely different data characteristics, and different models have different prediction accuracy for each state parameter. Summary of the invention

[0003] The purpose of the present invention is to provide a method for predicting multi-state parameters of aerospace mass spectrometer based on an ARIMA-LSTM combined model, which improves the prediction accuracy and is applicable to different state parameters.

[0004] To achieve the above object, the present invention provides a method for predicting multi-state parameters of aerospace mass spectrometer based on an ARIMA-LSTM combined model, comprising the following steps:

[0005] S1. Collect six state parameters of the aerospace mass spectrometer, including multiplier voltage, filament current, ion pump voltage, +200v voltage, DC output and RF output, and establish a data set based on the state parameter data collected by each sensor;

[0006] S2, divide the data set into training set and test set;

[0007] S3, perform EMD modal decomposition on the training set, extract the linear part and nonlinear part of the data, and divide the data into multiple linear eigenmode functions and nonlinear residuals;

[0008] S4, the linear intrinsic mode function is used to train the ARIMA model, and the trained ARIMA model is used to predict the linear part of the data;

[0009] S5. The nonlinear residual part is used to train the LSTM model, and the trained LSTM model is used to predict the nonlinear part of the data;

[0010] S6. Divide the test set into two subsets. The first subset selects 500 data and uses the Grey Wolf Optimization Algorithm to determine the weight of the combined model according to the prediction error. The second subset is used to verify the prediction effect of the combined model.

[0011] S7. The LSTM model and the ARIMA model are integrated according to the combined model weights, and the prediction results are verified using the second subset.

[0012] Preferably, the steps of EMD modal decomposition in S3 are as follows:

[0013] A1. The original signal is x(t). Find the maximum and minimum points of x(t), and use the cubic spline interpolation function to fit the upper and lower envelopes of x(t).

[0014] A2. Find the mean of the upper envelope and the lower envelope, and draw the mean envelope.

[0015] A3. Subtract the mean envelope from x(t) to obtain the intermediate signal.

[0016] A4. Determine whether the intermediate signal meets the two conditions of IMF:

[0017] a. The number of extreme points and the number of zero-crossing points in the entire data segment must be equal or differ by no more than one;

[0018] b. At any time, the average value of the upper envelope formed by the local maximum point and the lower envelope formed by the local minimum point is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis. If it is satisfied, it is the first IMF component. If not, repeat steps A1-A3;

[0019] A5. After obtaining the IMF component, subtract the first IMF component from x(t) to obtain r(t), and determine whether r(t) is monotonic. If not, repeat the above steps. If the final remaining signal is monotonic, the decomposition ends, and the remaining signal is the nonlinear part.

[0020] Preferably, the steps of ARIMA time series prediction in S4 are as follows:

[0021] B1. Use unit root test (ADF) to test the stability of mass spectrometer state parameter data;

[0022] B2. If the data is a non-stationary time series, perform d-order difference on the data to obtain a stationary time series;

[0023] B3. Select the p-order and q-order of the ARIMA model, and use the Akaike Information Criterion (AIC) and the Minimization Bayesian Criterion (BIC) to determine the order p and q. In order to reduce the amount of calculation, the maximum order of p is limited to no more than 6, and the maximum order of q is limited to no more than 4;

[0024] B4. Fit the model and test whether the residual is white noise;

[0025] B5. Predict the state parameters of the mass spectrometer and perform differential restoration on non-stationary data.

[0026] Preferably, the network structure of the LSTM algorithm in S5 is: including LSTM layer 1, LSTM layer 2, fully connected layer 1 and fully connected layer 2, and LN layers are connected between LSTM layer 1 and LSTM layer 2, and between LSTM layer 2 and fully connected layer 1, wherein LSTM layer 1 and LSTM layer 2 are both separate storage units, each storage unit has three layers: a forget gate, an input gate and an output gate, and each layer performs a separate function.

[0027] Preferably, the step of using the Gray Wolf Optimization Algorithm in S6 to determine the weight of the combined model according to the prediction error is as follows:

[0028] C1. Social stratification, the gray wolf group is divided into four types, which are α wolf, β wolf, δ wolf and ω wolf in descending order. Among them, α wolf, β wolf and δ wolf represent the three solutions with the best fitness value, and the remaining gray wolf in the wolf pack is ω wolf;

[0029] C2. During the prey search process, the update formula for the wolf pack's position is:

[0030]

[0031] Where: is the distance between the gray wolf and its prey; and is the coefficient vector; is the position vector of the prey; is the position vector of the gray wolf; t is the current iteration number; convergence factor As the number of iterations increases, it decreases linearly in the interval [0,2]; and is a random vector in [0,1];

[0032] C3, calculate the fitness function of all wolves in the wolf pack, and select the three best solutions, namely the three leading gray wolves, set as α wolf, β wolf and δ wolf;

[0033] C4, surround the prey, update the position of the ω gray wolf according to the position information of the three leading gray wolves, and the update expression is:

[0034]

[0035]

[0036] Where: and are the distances between α wolf, β wolf, δ wolf and the rest of ω wolf, and is the coefficient vector of formula (13), and are the position vectors of α wolf, β wolf and δ wolf respectively, and are the position vectors updated by ω wolf to α wolf, β wolf and δ wolf respectively, and is the coefficient vector obtained by equation (3);

[0037] C5. Perturb the current optimal solution to obtain a new solution;

[0038] C6. Determine whether the maximum number of iterations is met.

[0039] Therefore, the present invention adopts the above-mentioned aerospace mass spectrometer multi-state parameter prediction method based on the ARIMA-LSTM combined model, which has the following beneficial effects:

[0040] (1) The linear data processing module designed by the present invention is an ARIMA model, which can convert a non-stationary time series into a stationary time series by difference;

[0041] (2) The ARIMA model consists of an autoregression (AR) model, a difference model, and a moving average (MA) model, and can effectively process and predict linear time series;

[0042] (3) The nonlinear data processing module uses the LSTM model. The LSTM introduces a gate structure to alleviate the gradient explosion and gradient disappearance phenomena during the RNN network training process. It can better capture the nonlinear trend in the time series and is used to capture the nonlinear errors in the mass spectrometer state parameter data.

[0043] (4) The two algorithms are fused using the Grey Wolf Optimization Algorithm. The fused algorithm can better capture the linear and nonlinear trends of the data and has a good prediction effect on various state parameters of the mass spectrometer.

[0044] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of an embodiment of the present invention;

[0046] Figure 2This is a specific flow chart of the ARIMA model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1 , a multi-state parameter prediction method for aerospace mass spectrometer based on ARIMA-LSTM combined model, comprising the following steps:

[0049] S1. Collect six state parameters of the aerospace mass spectrometer, including multiplier voltage, filament current, ion pump voltage, +200v voltage, DC output and RF output, and establish a data set based on the state parameter data collected by each sensor.

[0050] S2. Divide the dataset into training set and test set.

[0051] S3. Perform EMD modal decomposition on the training set, extract the linear and nonlinear parts of the data, and divide the data into multiple linear eigenmode functions and nonlinear residuals.

[0052] The steps of EMD modal decomposition are as follows:

[0053] A1. The original signal is x(t). Find the maximum and minimum points of x(t), and use the cubic spline interpolation function to fit the upper and lower envelopes of x(t).

[0054] A2. Find the mean of the upper envelope and the lower envelope, and draw the mean envelope.

[0055] A3. Subtract the mean envelope from x(t) to obtain the intermediate signal.

[0056] A4. Determine whether the intermediate signal meets the two conditions of IMF:

[0057] a. The number of extreme points and the number of zero-crossing points in the entire data segment must be equal or differ by no more than one;

[0058] b. At any time, the average value of the upper envelope formed by the local maximum point and the lower envelope formed by the local minimum point is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis. If it is satisfied, it is the first IMF component. If not, repeat steps A1-A3;

[0059] A5. After obtaining the IMF component, subtract the first IMF component from x(t) to obtain r(t), and determine whether r(t) is monotonic. If not, repeat the above steps. If the final remaining signal is monotonic, the decomposition ends, and the remaining signal is the nonlinear part.

[0060] S4. The linear eigenmode function is used to train the ARIMA model, and the trained ARIMA model is used to predict the linear part of the data.

[0061] like Figure 2 , the steps of ARIMA time series forecasting are as follows:

[0062] B1. Use unit root test (ADF) to test the stability of mass spectrometer state parameter data;

[0063] B2. If the data is a non-stationary time series, perform d-order difference on the data to obtain a stationary time series;

[0064] B3. Select the p-order and q-order of the ARIMA model, and use the Akaike Information Criterion (AIC) and the Minimization Bayesian Criterion (BIC) to determine the order p and q. In order to reduce the amount of calculation, the maximum order of p is limited to no more than 6, and the maximum order of q is limited to no more than 4;

[0065] B4. Fit the model and test whether the residual is white noise;

[0066] B5. Predict the state parameters of the mass spectrometer and perform differential restoration on non-stationary data.

[0067] S5. The nonlinear residual part is used to train the LSTM model, and the trained LSTM model is used to predict the nonlinear part of the data.

[0068] The network structure of the LSTM algorithm is as follows: including LSTM layer 1, LSTM layer 2, fully connected layer 1 and fully connected layer 2. LN layers are connected between LSTM layer 1 and LSTM layer 2, and between LSTM layer 2 and fully connected layer 1. LSTM layer 1 and LSTM layer 2 are separate storage units. Each storage unit has three layers: forget gate, input gate and output gate. Each layer performs a separate function.

[0069] S6. Divide the test set into two subsets. The first subset selects 500 data and uses the Grey Wolf Optimization Algorithm to determine the weight of the combined model based on the prediction error. The second subset is used to verify the prediction effect of the combined model.

[0070] Preferably, the step of using the Gray Wolf Optimization Algorithm in S6 to determine the weight of the combined model according to the prediction error is as follows:

[0071] C1. Social stratification, the gray wolf group is divided into four types, which are α wolf, β wolf, δ wolf and ω wolf in descending order. Among them, α wolf, β wolf and δ wolf represent the three solutions with the best fitness value, and the remaining gray wolf in the wolf pack is ω wolf;

[0072] C2. During the prey search process, the update formula for the wolf pack's position is:

[0073]

[0074]

[0075] Where: is the distance between the gray wolf and its prey; and is the coefficient vector; is the position vector of the prey; is the position vector of the gray wolf; t is the current iteration number; convergence factor As the number of iterations increases, it decreases linearly in the interval [0,2]; and is a random vector in [0,1];

[0076] C3, calculate the fitness function of all wolves in the wolf pack, and select the three best solutions, namely the three leading gray wolves, set as α wolf, β wolf and δ wolf;

[0077] C4, surround the prey, update the position of the ω gray wolf according to the position information of the three leading gray wolves, and the update expression is:

[0078]

[0079] Where: and are the distances between α wolf, β wolf, δ wolf and the rest of ω wolf, and is the coefficient vector of formula (13), and are the position vectors of α wolf, β wolf and δ wolf respectively, and are the position vectors updated by ω wolf to α wolf, β wolf and δ wolf respectively, and is the coefficient vector obtained by equation (3);

[0080] C5. Perturb the current optimal solution to obtain a new solution;

[0081] C6. Determine whether the maximum number of iterations is met.

[0082] S7. The LSTM model and the ARIMA model are integrated according to the combined model weights, and the prediction results are verified using the second subset.

[0083] The proposed method was used to predict six state parameters respectively. It was verified that the minimum RMSE of the prediction results was 0.0568, and the minimum MAE was 0.0382, which were 21.1% and 6.3% higher than those of the ARIMA model, and 10.1% and 36.5% higher than those of the LSTM model.

[0084] Table 1 Prediction results of each model for six parameters

[0085]

[0086] Therefore, the present invention adopts the above-mentioned aerospace mass spectrometer multi-state parameter prediction method based on the ARIMA-LSTM combined model, and integrates the ARIMA and LSTM algorithms through the Grey Wolf optimization algorithm, which can effectively track and predict 6 mass spectrometer state parameters with different characteristics.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting multi-state parameters of aerospace mass spectrometer based on ARIMA-LSTM combined model, characterized in that: The following steps are involved: S1. Collect six state parameters of the aerospace mass spectrometer, including multiplier voltage, filament current, ion pump voltage, +200v voltage, DC output and RF output, and establish a data set based on the state parameter data collected by each sensor; S2, divide the data set into training set and test set; S3, perform EMD modal decomposition on the training set, extract the linear part and nonlinear part of the data, and divide the data into multiple linear eigenmode functions and nonlinear residuals; S4, the linear intrinsic mode function is used to train the ARIMA model, and the trained ARIMA model is used to predict the linear part of the data; S5. The nonlinear residual part is used to train the LSTM model, and the trained LSTM model is used to predict the nonlinear part of the data; S6. Divide the test set into two subsets. The first subset selects 500 data and uses the Grey Wolf Optimization Algorithm to determine the weight of the combined model according to the prediction error. The second subset is used to verify the prediction effect of the combined model. S7. The LSTM model and the ARIMA model are integrated according to the combined model weight, and the prediction results are verified using the second subset.

2. The method for predicting multi-state parameters of aerospace mass spectrometer based on ARIMA-LSTM combined model according to claim 1, characterized in that: The steps of EMD modal decomposition in S3 are as follows: A1. The original signal is x(t). Find the maximum and minimum points of x(t). Use the cubic spline interpolation function to fit the upper and lower envelopes of x(t). A2. Find the mean of the upper envelope and the lower envelope, and draw the mean envelope; A3, subtract the mean envelope from x(t) to get the intermediate signal; A4. Determine whether the intermediate signal meets the two conditions of IMF: a. The number of extreme points and the number of zero-crossing points in the entire data segment are equal or differ by no more than one; b. At any time, the average value of the upper envelope formed by the local maximum point and the lower envelope formed by the local minimum point is zero, and the upper and lower envelopes are locally symmetrical with respect to the time axis; if satisfied, it is the first IMF component, if not satisfied, repeat steps A1-A3; A5. After obtaining the IMF component, use x(t) to subtract the first IMF component to obtain r(t), and determine whether r(t) is monotonic. If not, repeat steps A1-A4. If the final remaining signal is monotonic, the decomposition ends, and the remaining signal is the nonlinear part.

3. The method for predicting multi-state parameters of aerospace mass spectrometer based on ARIMA-LSTM combined model according to claim 2, characterized in that: The steps for ARIMA time series forecasting in S4 are as follows: B1. Use unit root test ADF to test the stability of mass spectrometer state parameter data; B2. If the data is a non-stationary time series, perform d-order difference on the data to obtain a stationary time series; B3. Select the p-order and q-order of the ARIMA model, and use the Akaike Information Criterion AIC and the Minimization Bayesian Criterion BIC to determine the order p and q. The maximum order of p is not greater than 6, and the maximum order of q is not greater than 4. B4. Fit the model and test whether the residual is white noise; B5. Predict the state parameters of the mass spectrometer and perform differential restoration on non-stationary data.

4. The method for predicting multi-state parameters of aerospace mass spectrometer based on ARIMA-LSTM combined model according to claim 3, characterized in that: The network of the LSTM algorithm in S5 includes LSTM layer 1, LSTM layer 2, fully connected layer 1 and fully connected layer 2. LN layers are connected between LSTM layer 1 and LSTM layer 2, and between LSTM layer 2 and fully connected layer 1. LSTM layer 1 and LSTM layer 2 are separate storage units, and each storage unit is provided with a forget gate, an input gate and an output gate.

5. The method for predicting multi-state parameters of aerospace mass spectrometer based on ARIMA-LSTM combined model according to claim 4, characterized in that: The steps in S6 to use the Gray Wolf Optimization Algorithm to determine the weight of the combined model based on the prediction error are as follows: C1. Social stratification, the gray wolf group is divided into four types, which are α wolf, β wolf, δ wolf and ω wolf in descending order. Among them, α wolf, β wolf and δ wolf represent the three solutions with the best fitness value, and the remaining gray wolf in the wolf pack is ω wolf; C2. During the prey search process, the update formula for the wolf pack's position is: In the formula, is the distance between the gray wolf and its prey; and is the coefficient vector; is the position vector of the prey; is the position vector of the gray wolf; t is the current iteration number; convergence factor As the number of iterations increases, it decreases linearly in the interval [0,2]; and is a random vector in [0,1]; C3, calculate the fitness function of all wolves in the wolf pack, select three optimal solutions, set as α wolf, β wolf and δ wolf; C4, surround the prey, update the position of the ω gray wolf according to the position information of the three leading gray wolves, and the update expression is: In the formula, and are the distances between α wolf, β wolf, δ wolf and the rest of ω wolf respectively; and is the coefficient vector of formula (13); and are the position vectors of α wolf, β wolf and δ wolf respectively; and are the position vectors updated by ω wolf to α wolf, β wolf and δ wolf respectively; and is the coefficient vector obtained by equation (3); C5. Perturb the current optimal solution to obtain a new solution; C6. Determine whether the maximum number of iterations is met.

Images