An air duct system control method based on LSTM prediction model

Abstract

The present invention discloses an air duct system control method based on an LSTM prediction model, belonging to the technical field of control and regulation. The present invention constructs a neural network model that predicts the pressure and flow rate when the system is stable based on the current system state parameters and the opening of the pipeline valve after adjustment. The historical data of the system operation is preprocessed to obtain a data set, and the model with the best performance is selected through multiple training. When formulating a strategy, the valve to be adjusted is specified and the valve adjustment range is limited. The valve adjustment direction is set according to the relationship between the target pressure, flow and the current value. The adjustment range of all valves is traversed with a certain step size to obtain a valve opening combination, and the air pressure and flow rate when the system is stable under each valve opening combination are predicted. The present invention predicts the changes in the system air pressure and flow rate in advance by a method of first predicting and then adjusting, and then accurately controls the valve opening, which can improve the work efficiency of state regulation of the air duct system.

Description

Technical Field

[0001] The present invention relates to predictive control technology, and in particular discloses an air duct system control method based on an LSTM prediction model, belonging to the technical field of control and regulation. Background Art

[0002] Ground testing of aircraft engines is a crucial step in their R&D, mass production, and operational lifecycle, revealing the vast majority of technical and quality issues. The air duct system provides air at a certain pressure and flow rate for test bench engines. Traditionally, the control strategy involves manually adjusting valves and observing the pressure and flow rate of the test bench air after the system stabilizes. The observed pressure and flow rates are then compared with the expected targets. Based on the direction and magnitude of the deviation between the observed pressure and the target flow rate, the valve adjustment strategy is adjusted and the valves are readjusted. This valve adjustment process is repeated multiple times to ultimately achieve the target pressure and flow rate. Traditional manual adjustment strategies rely heavily on experience and require multiple adjustments, resulting in low efficiency and a time-consuming and labor-intensive process.

[0003] Predictive control, also known as model predictive control (MPC), is a special type of control strategy. Its current control action is obtained at each sampling instant by solving a finite-time open-loop optimal control problem. The current state of the process is used as the initial state of the optimal control problem, and the optimal control sequence is obtained. MPC essentially solves an open-loop optimal control problem. Its concept is independent of the specific model, but its implementation is model-dependent.

[0004] Predictive models are a crucial component of predictive control. They predict future process outputs based on the system's current control inputs and historical process information. Long Short-Term Memory (LSTM) networks are a key research area in machine learning, and their introduction has significantly advanced the development of artificial intelligence. Deep learning has achieved significant success in areas such as natural language processing (NLP), data mining, and machine translation. Due to its unique design structure, LSTMs are well-suited for processing and predicting time series with specific intervals and delays.

[0005] Aiming at the air duct system of aircraft engine ground test, the present invention aims to propose an air duct system control method based on LSTM prediction model to overcome the defects of manual adjustment valves. Summary of the Invention

[0006] The purpose of the present invention is to address the shortcomings of the above-mentioned background technology, based on the idea of ​​predictive control and deep learning methods, to provide an air duct system control method based on the LSTM prediction model, to solve the technical problems of low adjustment efficiency, time and labor consumption of traditional manual adjustment strategies, and to achieve the purpose of the invention of automatic control of the air duct system.

[0007] The present invention adopts the following technical solutions to achieve the above-mentioned purpose:

[0008] An air duct system control method based on an LSTM prediction model, comprising:

[0009] Training phase: Initialize the air duct system parameter file, identify each steady state from the initialized air duct system parameter file and record the average value of each parameter in each steady state time period, record the period from the start time of a steady state to the end time of the next adjacent steady state as a single adjustment process, and obtain the direction of change of the test bench air pressure and flow in each single adjustment process according to the change trend of the average value of the same parameter under two adjacent steady states, merge two or more single adjustment processes with continuous same-direction changes into a continuous same-direction adjustment process, calculate the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period in each single adjustment process, and calculate the average value and The average value of each parameter in the terminal steady-state period, the average value of each parameter in the starting steady-state period and the average value of each parameter in the terminal steady-state period in each single adjustment process, and the average value of each parameter in the starting steady-state period and the average value of each parameter in the terminal steady-state period in each continuous unidirectional adjustment process are used as the data set of the prediction model. The data set is used to train prediction models with different structures. The prediction model with the highest comprehensive prediction accuracy is selected as the optimal prediction model among the prediction models whose loss function values ​​meet the convergence requirements. The parameters recorded in the air duct system parameter file include: the opening of each valve that affects the air pressure and flow of the test bench, the air pressure and temperature of the branch where each valve is located on the test bench intake duct, and the air pressure, temperature and flow of the test bench;

[0010] Application stage: Set the absolute position range of each valve that affects the air pressure and flow of the test bench, determine the adjustment direction of each valve that affects the air pressure and flow of the test bench according to the air pressure target value and air flow target value of the test bench, set the valve opening adjustment step, traverse all valve opening combinations according to the adjustment direction of each valve that affects the air pressure and flow of the test bench to generate a valve opening combination, input each valve opening combination and the current air temperature and pressure of the branch where each valve is located on the test bench intake pipe into the optimal prediction model, and the optimal prediction model outputs the air pressure of each valve on the branch where the test bench intake pipe is located. The predicted values ​​of the air pressure and flow of the test bench after the system is stable are used. The pressure query range and flow query range are calculated according to the air pressure target value and air flow target value of the test bench and the set pressure query fault tolerance rate and flow query fault tolerance rate. Within the pressure query range and flow query range, valve adjustment strategies that meet the safety requirements of the air pipeline system are screened from each valve opening combination, and the screened valve adjustment strategies are sorted. One of the sorted valve adjustment strategies is selected for execution. After the air pipeline system is stable, the adjustment amount of each valve executing the valve adjustment strategy is the adjustment result of each valve.

[0011] Furthermore, in an air duct system control method based on an LSTM prediction model, a specific method for initializing an air duct system parameter file is as follows: according to the time series data of each parameter in the air duct system parameter file, the adjustment direction of each valve that affects the air pressure and flow of the test bench at the previous moment is extracted, the adjustment direction of each valve that affects the air pressure and flow of the test bench at the previous moment is used as a parameter and spliced ​​with the air duct system parameter file, and the median sampling is performed on each parameter in the spliced ​​air duct system parameter file.

[0012] Furthermore, in an air duct system control method based on an LSTM prediction model, the basis for identifying each steady state from the initialized air duct system parameter file includes: the range of each parameter in the current steady-state time period is less than 1% of the corresponding parameter minimum value, the pressure change rate at each moment in the current steady-state time period is less than 3kPa / s, the opening change of each valve affecting the test bench air pressure and flow in the current steady-state time period is less than 1%, and the duration of the current steady-state time period is greater than 60s.

[0013] Furthermore, in an air duct system control method based on an LSTM prediction model, a data set consisting of the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each single adjustment process, as well as the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each continuous unidirectional adjustment process is used as the data set of the prediction model. The specific method is: the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each single adjustment process, as well as the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each continuous unidirectional adjustment process The average values ​​of the various parameters in the terminal steady-state period are normalized, and 70% of the normalized results of the average values ​​of the various parameters in the starting steady-state period and the average values ​​of the various parameters in the terminal steady-state period of the continuous unidirectional adjustment process, as well as the average values ​​of the various parameters in the starting steady-state period and the average values ​​of the various parameters in the terminal steady-state period of the single adjustment process are used as the training data set, and 30% of the normalized results of the average values ​​of the various parameters in the starting steady-state period and the average values ​​of the various parameters in the terminal steady-state period of the single adjustment process are used as the test data set. The training data set and the test data set constitute the data set of the prediction model.

[0014] Furthermore, in an air duct system control method based on an LSTM prediction model, a weighted square error method is used to calculate the loss function value in the process of using a data set to train prediction models of different structures.

[0015] Furthermore, in an air duct system control method based on an LSTM prediction model, a specific method for determining the adjustment direction of each valve affecting the air pressure and flow of a test bench according to the air pressure target value and air flow target value of the test bench is as follows: a group of valves with opposite correlation with the air flow of the test bench are selected as valves for adjusting the air flow of the test bench, and a group of valves with negative correlation with the air pressure of the test bench are selected as valves for adjusting the air pressure of the test bench; the adjustment direction of a group of valves with opposite correlation with the air flow of the test bench is determined according to the relationship between the air flow target value of the test bench and the current air flow; the adjustment direction of a group of valves with negative correlation with the air pressure of the test bench is determined according to the relationship between the air pressure target value of the test bench and the current air pressure; and the remaining valves maintain their original openings.

[0016] Furthermore, in an air duct system control method based on an LSTM prediction model, a specific method for sorting the screened valve adjustment strategies is as follows: sorting the screened valve adjustment strategies according to the number of adjustment valves; when the number of adjustment valves is the same, sorting them in ascending order according to the valve adjustment amount.

[0017] The present invention adopts the above-mentioned technical solution, which has the following beneficial effects: the present invention constructs a neural network prediction model for predicting the pressure and flow of the system test bench after valve adjustment, which can predict in advance the pressure and flow of the air duct flowing through the test bench after valve adjustment. By traversing all possible valve opening combination strategies as the input of the prediction model and predicting the output results corresponding to different valve opening combination strategies, the optimal adjustment strategy is obtained, and the valve is automatically adjusted to complete the control. The present invention adopts a method of first predicting and then adjusting. Compared with the traditional strategy of first outputting and then feedback and then controlling, this method has the technical advantage of accurately predicting the pressure and flow of the test bench and then accurately controlling the valve opening, thereby improving the work efficiency of state adjustment of the air duct system, liberating manpower to a certain extent and saving the adjustment time of the air duct system, thereby improving the efficiency of ground testing of aircraft engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a large air duct system.

[0019] Figure 2(a) and Figure 2(b) are the flow charts of the control strategy training process and application process.

[0020] Figure 3 It is the original timing curve of pressure and flow of the test bench.

[0021] Figure 4 It is the timing curve diagram of the median pressure and flow rate of the test bench after sampling.

[0022] Figure 5 The following is an example diagram of the changes in the test bench pressure and flow timing curves during a single adjustment process.

[0023] Figure 6 This is an example diagram of the changes in the test bench pressure and flow timing curves during a continuous unidirectional adjustment process after a certain merger.

[0024] Figure 7 Graph showing the average relative error of the predicted test bench pressure for different prediction models.

[0025] Figure 8 Schematic diagram of the average relative error of the test bench flow predicted for different prediction models. DETAILED DESCRIPTION

[0026] The embodiment of the present invention is a large air duct system, such as Figure 1 For the convenience of explanation, the symbols used are shown in Table 1.

[0027] symbol meaning K1,K2,…,K9 Opening of valves K1, K2, ..., K9 P Test bench pressure W Air flow through the test bench T Test bench air temperature P1 The pressure measured by the instrument on the branch where K1 is located P2 The pressure measured by the instrument on the branch where K2 is located T1 The temperature measured by the instrument on the branch where K2 is located

[0028] Table 1

[0029] With respect to the embodiments of the present invention, the specific implementation manner of the present invention will be further described below in conjunction with the accompanying drawings.

[0030] The present application discloses an air duct system control method based on an LSTM prediction model. The method includes a training process and an application process of the prediction model. The flow chart of the training process control strategy is shown in Figure 2(a).

[0031] The training process of the prediction model includes the following 9 steps.

[0032] Step 1: Based on the historical operation records of the air duct system, export the time series data of relevant parameters in each operating time period.

[0033] For the embodiment of the present invention, all parameters of the historical operation record of the air duct system are stored in the database in chronological order, including: the opening of each valve, the air temperature, air pressure, air flow of each branch, and the operation information of the fuel subsystem and the cooling water subsystem. The branch refers to the air duct from one node (the intersection of three or more pipelines) to another node. The present invention does not control the fuel subsystem and the cooling water subsystem, so it will not be discussed. All parameters related to the pressure and flow control of the air duct system are derived from the database, including the valve opening of the 9 related valves K1 to K9, the pressure P1 measured by the instrument of the branch where K1 is located, the pressure P2 measured by the instrument of the branch where K2 is located, the temperature T1 measured by the instrument of the branch where K2 is located, and the temperature T, pressure P and flow W of the air at the test bench, and these data are saved as different files by date. The data visualization of different files is realized through programming, the data is analyzed, and the data files for normal operation are screened out. The data file is saved in the form of a table, the table header is the parameter symbol, and its content is the value of the corresponding parameter at each moment. The timing curve of the air pressure and air flow at the test bench is shown as follows. Figure 3 shown.

[0034] Step 2: Based on the air duct system operation timing data obtained in step 1, extract the adjustment direction of each valve at the last moment, and use the adjustment direction of each valve at the last moment as a parameter to splice with the file saved in step 1.

[0035] The data in the spliced ​​files are sampled at a median interval of one second to reduce the amount of data and improve the stability and representativeness of the data. The time series curves of the test bench air pressure and air flow after median sampling are as follows: Figure 4 shown.

[0036] Step 3: Based on the median-sampled data from Step 2, identify the steady state, record the start and end times of the steady state, and calculate the average value of each parameter during the period from the start to the end of the steady state. Based on the start and end times of each steady state, separate the single adjustment process, that is, the process from the start of one steady state to the end of the next adjacent steady state.

[0037] The definition of steady state is as follows: (1) the range of the process is less than 1% of the minimum value; (2) the rate of change of pressure at each moment is less than 3kPa / s; (3) the change of the opening of all valves is less than 1%; (4) the duration.

[0038] Figure 5 The following is an example diagram of the changes in the test bench air pressure and air flow timing curves during a single adjustment process.

[0039] Step 4: Based on the average values ​​of each parameter within each steady-state period obtained in Step 3 and the direction of change of air pressure and air flow during a single adjustment process, the direction of change of the test bench air pressure and air flow during each single adjustment process is obtained based on the trend of the average values ​​of the same parameter under two adjacent steady states. If the adjustment direction of the air pressure and air flow remains unchanged during two or more consecutive single adjustment processes, these two or more consecutive single adjustment processes can be merged to obtain a merged continuous same-direction adjustment process, thereby expanding the data volume. Figure 6 This is an example of the changes in the test bench air pressure and air flow during a combined continuous unidirectional regulation process.

[0040] Step 5: Based on the single adjustment process obtained in step 3 and the continuous same-direction adjustment process obtained in step 4, calculate the average value of each parameter in the initial steady-state period and the average value of each parameter in the terminal steady-state period of the two types of processes, and use them as the input data of the training model.

[0041] Step 6: Based on the input data obtained in step 5 for model training, convert the input data into a format suitable for supervised training. This includes normalizing the data and dividing it into training and test datasets.

[0042] The normalization algorithm used is the standard normalization algorithm, and the conversion function is as follows:

[0043]

[0044] Where x represents the unnormalized data, and x' represents the normalized data. μ is the mean of all sample data, and σ is the standard deviation of all sample data. Normalized data conforms to a standard normal distribution, with a mean of 0 and a standard deviation of 1.

[0045] The divided training data set contains 70% of the data volume of the average value of each parameter in the initial steady-state period and the average value of each parameter in the final steady-state period of the merged continuous unidirectional adjustment process, and the average value of each parameter in the initial steady-state period and the average value of each parameter in the final steady-state period of a single adjustment process; the divided test data set is 30% of the data volume of the average value of each parameter in the initial steady-state period and the average value of each parameter in the final steady-state period of a single adjustment process.

[0046] Step 7: As shown in Table 2, select different numbers of neural network layers or different numbers of neurons in neural network layers, and preset multiple prediction model structures and other hyperparameters of each prediction model.

[0047] parameter value Number of network layers 1，2，3，4，5，6 Number of nodes 100，300，500，800，1000，1500，2000，2500，3000 Network layer type LSTM Optimizer Adam Initial learning rate 0.001

[0048] Table 2

[0049] Step 8: Based on the training dataset obtained in Step 6, train the prediction model structure preset in Step 7. Based on the test dataset obtained in Step 6, calculate the loss of the prediction model on the test set. If the loss function value of the prediction model does not decrease for multiple consecutive rounds, after reaching a certain number of times, training is terminated, and the prediction model trained under the corresponding model structure is obtained. If all preset model structures have been traversed, proceed to Step 9. Otherwise, return to Step 7 to construct another prediction model structure and repeat the training process in Step 8.

[0050] The loss function of the model in this embodiment is calculated as weighted sum of squared differences. The loss function wmse of r predictor variables is expressed as:

[0051]

[0052] Among them, y i is the true result value of the i-th variable; is the prediction result of the i-th variable under the action of the model; w i is the weight of the i-th variable. The larger its value is, the higher the prediction accuracy requirement of the corresponding variable is.

[0053] The number of prediction variables of the model of this embodiment is r=4, including P, W, P1 and P2, and their corresponding weights are 10, 10, 1, 1 respectively.

[0054] Step 9: Based on the multiple prediction models trained in step 8, compare the prediction accuracy of each prediction model for each prediction variable, and select the prediction model with the best overall performance as the final prediction model.

[0055] The average relative errors of different prediction models in predicting the test bench air pressure and air flow are as follows: Figure 7 and Figure 8As shown in the figure, the average relative error results of the comprehensive pressure and flow rate show that the prediction model with 1000 hidden layer nodes and 3 layers has the smallest average relative error of the predicted values ​​of each parameter. Therefore, the optimal parameters of the model are determined to be 1000 hidden layer nodes and 3 layers.

[0056] The application process of the prediction model is shown in Figure 2(b), which includes the following 8 steps.

[0057] Step 1: For the air piping system, select the valves that affect the system pressure and flow, and set their absolute adjustment position range to avoid entering the valve's insensitive zone, that is, the valve opening range where the valve adjustment has no significant effect on the system air flow and air pressure.

[0058] In this embodiment, there are a total of 9 valves that affect the system pressure and flow, and each valve is denoted as K1 to K9. The absolute position range of the valve setting is shown in Table 3.

[0059]

[0060]

[0061] Table 3

[0062] Step 2: Set the target pressure and target flow of the test bench. Based on this, set the adjustment direction of the valve selected in step 1. The selected adjustment direction should match the adjustment direction of the target pressure and target flow.

[0063] The correlation between the 9 valves selected in step 1 and the test bench pressure and test bench flow is shown in Table 4.

[0064] valve Relationship with test bench pressure Relationship with test bench flow rate K1 Positive correlation Positive correlation K2 Positive correlation Positive correlation K3 Positive correlation Positive correlation K4 Positive correlation Positive correlation K5 Positive correlation Positive correlation K6 negative correlation Positive correlation K7 negative correlation Positive correlation K8 negative correlation Positive correlation K9 negative correlation negative correlation

[0065] Table 4

[0066] To resolve the conflict between simultaneously regulating pressure and flow, four key valves are selected for adjustment: K9, K4, K6, and K7. K9 and K4 are primarily used to regulate flow, while K6 and K7 are primarily used to regulate pressure. The other five valves remain at a normal opening.

[0067] When the target air flow of the test bench is greater than the current air flow, the K9 adjustment direction is set to reduce the valve opening and the K4 adjustment direction is set to increase the valve opening; when the target air flow of the test bench is less than the current air flow, the K9 adjustment direction is set to increase the valve opening and the K4 adjustment direction is set to reduce the valve opening.

[0068] When the target air pressure of the test bench is greater than the current air pressure, set the adjustment direction of K6 and K7 to reduce the valve opening; when the target air pressure of the test bench is less than the current air pressure, set the adjustment direction of K6 and K7 to increase the valve opening.

[0069] Step 3: Based on the valve selected in step 1 and the valve adjustment direction (increase, decrease or remain unchanged) set in step 2, within the valve opening range set in step 1, with a valve opening of 1% as a step size, traverse and generate all possible valve opening combinations.

[0070] Step 4: Input the valve opening combination obtained in Step 3 into the prediction model along with the current state parameters of the air duct system. The current state parameters of the system include: the pressure of the gas in the branch pipe where K1 is located, the temperature of the gas in the branch pipe where K2 is located, the temperature of the gas in the branch pipe where K3 is located, and the temperature of the gas in the branch pipe where K4 is located.

[0071] Step 5: Based on step 4, for each valve opening combination, the prediction model outputs the pressure P and flow W of the test bench after the system stabilizes when the valve openings of the system are adjusted to the valve opening combination, as well as the other two related pressures P1 and P2.

[0072] Step 6: Set the pressure query tolerance and flow query tolerance. Based on the target pressure and flow set in Step 2, and the pressure query tolerance and flow query tolerance, determine the pressure query range and flow query range. Within this range, based on system safety and measurement requirements, and other constraints, select available valve adjustment strategies from each valve opening combination. Each valve adjustment strategy corresponds to a set of predicted data: the pressure P and flow W of the test bench after system stabilization, as well as two other related pressures, P1 and P2.

[0073] In this embodiment, the query error tolerance rate of pressure is set to 1%, and the query error tolerance rate of flow is set to 2%.

[0074] According to the safety requirements of the air pipeline system, the pressure P1 cannot be lower than 3300kPa and is usually not higher than 4500kPa. According to the use requirements of the flow orifice plate to measure flow, the difference between the pressure P2 and the test bench pressure P2-P cannot be greater than 250kPa.

[0075] Query the valve adjustment strategy that meets the requirements based on the above conditions.

[0076] Step 7: Based on the valve adjustment strategy obtained in step 6, sort the valves from small to large according to the number of adjustment valves. When the number of adjustment valves is the same, sort the valves from small to large according to the adjustment amount.

[0077] Step 8: Based on the sorted valve adjustment strategies obtained in step 7, select one to execute, wait for the system to stabilize, and obtain the adjustment result.

[0078] When the air duct system model has sufficient data samples, the flow control strategy error is within 2%. The effectiveness of the strategy is verified by simulation. The results are shown in Table 5:

[0079]

[0080] Table 5

[0081] When the air duct system model has sufficient data samples, the pressure control strategy error is within 1%. The effectiveness of the strategy is verified by simulation. The results are shown in Table 6:

[0082]

[0083] Table 6

[0084] The above description is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiment. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the scope of protection of the present invention.

Claims

1. An air duct system control method based on LSTM prediction model, characterized in that: include: Training phase: Initialize the air duct system parameter file, identify each steady state from the initialized air duct system parameter file and record the average value of each parameter in each steady state time period, record the period from the start time of a steady state to the end time of the next adjacent steady state as a single adjustment process, and obtain the direction of change of the test bench air pressure and flow in each single adjustment process according to the change trend of the average value of the same parameter under two adjacent steady states, merge two or more single adjustment processes with continuous same-direction changes into a continuous same-direction adjustment process, calculate the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period in each single adjustment process, and calculate the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period in each continuous same-direction adjustment process. The average value of each parameter in the steady-state period, the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period in each single adjustment process, and the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period in each continuous unidirectional adjustment process are used as the data set of the prediction model, and the data set is used to train prediction models with different structures. The prediction model with the highest comprehensive prediction accuracy is selected as the optimal prediction model among the prediction models whose loss function values ​​meet the convergence requirements. The parameters recorded in the air duct system parameter file include: the opening of each valve affecting the air pressure and flow of the test bench, the air pressure and temperature of the branch where each valve on the test bench intake duct is located, and the air pressure, temperature and flow of the test bench; Application stage: Set the absolute position range of each valve that affects the air pressure and flow of the test bench, determine the adjustment direction of each valve that affects the air pressure and flow of the test bench according to the air pressure target value and air flow target value of the test bench, set the valve opening adjustment step, traverse all valve opening combinations according to the adjustment direction of each valve that affects the air pressure and flow of the test bench to generate a valve opening combination, input each valve opening combination and the current air temperature and pressure of the branch where each valve is located on the test bench intake pipe into the optimal prediction model, and the optimal prediction model outputs the air pressure of each valve on the branch where the test bench intake pipe is located. The predicted values ​​of the air pressure and flow of the test bench after the system is stable are used. The pressure query range and flow query range are calculated according to the air pressure target value and air flow target value of the test bench and the set pressure query fault tolerance rate and flow query fault tolerance rate. Within the pressure query range and flow query range, valve adjustment strategies that meet the safety requirements of the air pipeline system are screened from each valve opening combination, and the screened valve adjustment strategies are sorted. One of the sorted valve adjustment strategies is selected for execution. After the air pipeline system is stable, the adjustment amount of each valve executing the valve adjustment strategy is the adjustment result of each valve.

2. The air duct system control method based on the LSTM prediction model according to claim 1 is characterized in that: The specific method for initializing the air duct system parameter file is as follows: based on the time series data of each parameter in the air duct system parameter file, the adjustment direction of each valve that affected the air pressure and flow of the test bench at the previous moment was extracted, the adjustment direction of each valve that affected the air pressure and flow of the test bench at the previous moment was used as a parameter to be spliced ​​with the air duct system parameter file, and the median sampling of each parameter in the spliced ​​air duct system parameter file was performed.

3. The air duct system control method based on the LSTM prediction model according to claim 2 is characterized in that: The basis for identifying each steady state from the initialized air duct system parameter file includes: the range of each parameter in the current steady-state time period is less than 1% of the corresponding parameter minimum value, the pressure change rate at each moment in the current steady-state time period is less than 3kPa / s, the opening change of each valve affecting the test bench air pressure and flow in the current steady-state time period is less than 1%, and the duration of the current steady-state time period is greater than 60s.

4. The air duct system control method based on the LSTM prediction model according to claim 3 is characterized in that: The specific method of using the data set consisting of the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each single adjustment process, and the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each continuous unidirectional adjustment process as the data set of the prediction model is: normalizing the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each single adjustment process, and the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of each continuous unidirectional adjustment process, and using 70% of the normalized results of the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of the continuous unidirectional adjustment process and the normalized results of the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of the single adjustment process as the training data set, and using 30% of the normalized results of the average value of each parameter in the starting steady-state period and the average value of each parameter in the ending steady-state period of the single adjustment process as the test data set, and the training data set and the test data set constitute the data set of the prediction model.

5. The air duct system control method based on the LSTM prediction model according to claim 4 is characterized in that: In the process of using the data set to train different structure prediction models, the weighted square difference method is used to calculate the loss function value.

6. The air duct system control method based on the LSTM prediction model according to claim 5 is characterized in that: The specific method for determining the adjustment direction of each valve affecting the air pressure and flow of the test bench based on the air pressure target value and the air flow target value of the test bench is as follows: selecting a group of valves with an opposite correlation with the test bench air flow as valves for adjusting the test bench air flow, selecting a group of valves with a negative correlation with the test bench air pressure as valves for adjusting the test bench air pressure, determining the adjustment direction of a group of valves with an opposite correlation with the test bench air flow according to the relationship between the test bench air flow target value and the current air flow, determining the adjustment direction of a group of valves with a negative correlation with the test bench air pressure according to the relationship between the test bench air pressure target value and the current air pressure, and maintaining the original openings of the remaining valves.

7. The air duct system control method based on the LSTM prediction model according to claim 6 is characterized in that: The specific method for sorting the selected valve adjustment strategies is: sorting the selected valve adjustment strategies according to the number of adjustment valves, and sorting them in ascending order according to the valve adjustment amount when the number of adjustment valves is the same.

Images