Aero-engine gas path system digital twinborn virtual-real synchronization method based on deep reinforcement learning

Abstract

The invention discloses an aero-engine gas path system digital twin virtual-real synchronization method based on deep reinforcement learning, and aims to realize virtual-real synchronization and monitor deep parameters in an aero-engine gas path system in real time. According to the method, parameters are divided into parameters which can be directly measured and parameters which cannot be directly measured, and the parameters are synchronized by adopting different algorithms respectively. An unscented Kalman filtering algorithm is adopted for parameters which can be directly measured; and for parameters which cannot be directly measured, a reinforcement learning method is adopted to deduce. Through the method, the inference precision of deep parameters can be effectively improved, and meanwhile, the real-time performance and robustness of the inference process are ensured. Experimental results show that the method has significant performance improvement in parameter inference and virtual-real synchronization, and can be applied to state monitoring and optimal scheduling of an aero-engine gas path system.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engine gas path monitoring, and is used to realize the monitoring and parameter inference of the operating state of the aero-engine gas path system. Specifically, it is a digital twin virtual-real synchronization method for the aero-engine gas path system based on deep reinforcement learning. Background Art

[0002] The aero-engine gas path system is vulnerable to damage under harsh conditions, such as fatigue, wear, and corrosion of gas path components. Therefore, timely monitoring of the health status of the aero-engine gas path system is crucial for the safe operation of the aero-engine. How to achieve more comprehensive monitoring of the health status of the aero-engine gas path system has become a highly concerned issue. In recent years, the monitoring scheme for the aero-engine gas path system developed based on the digital twin theory has made certain progress ("Zhou, Liang, Huawei Wang, and Shanshan Xu. Aero-engine gas path system health assessment based on depth digital twin."). However, most of the existing digital twin monitoring schemes only consider monitoring gradual parameters such as temperature, pressure, and flow rate, and lack the monitoring of deep-level parameters such as the efficiency parameter and pressure ratio parameter of gas path components. These deep-level parameters can better reflect the health status of the aero-engine gas path system itself.

[0003] In view of the problem of incomplete parameter monitoring of the aero-engine gas path system faced, it is necessary to study a digital twin virtual-real synchronization method for the aero-engine gas path system based on deep reinforcement learning to solve it. Summary of the Invention

[0004] In order to solve the problem of incomplete parameter monitoring of the aero-engine gas path system, the purpose of the present invention is to provide a digital twin virtual-real synchronization method for the aero-engine gas path system based on deep reinforcement learning, which is used to realize more comprehensive parameter monitoring of the aero-engine gas path system.

[0005] The technical solution of the present invention:

[0006] A method for virtual-real synchronization of an aero-engine gas path system digital twin based on deep reinforcement learning includes virtual-real synchronization of non-directly measurable parameters and virtual-real synchronization of directly measurable parameters; the virtual-real synchronization of non-directly measurable parameters realizes the monitoring of non-directly measurable parameters; the virtual-real synchronization of directly measurable parameters realizes the monitoring of directly measurable parameters; inferring non-directly measurable parameters depends on a parameter inference model and inferring directly measurable parameters depends on the unscented Kalman filter algorithm; the inferred non-directly measurable parameters are respectively used for parameter update of the non-directly measurable parameters in the parameter table 1 to be synchronized and parameter correction of the corresponding non-directly measurable parameters in the aero-engine gas path mechanism model inside the unscented Kalman filter algorithm; the inferred directly measurable parameters are used for parameter update of the directly measurable parameters in the parameter table 1 to be synchronized; the parameter table 1 to be synchronized synchronizes the parameter values of non-directly measurable parameters and directly measurable parameters to the parameter table 2 to be synchronized at a certain update frequency; the parameter table 2 to be synchronized is used for subsequent construction of the aero-engine gas path digital twin; the input data of the parameter inference model is the sensor data sequence, and the output is the inferred non-directly measurable parameters; the input of the unscented Kalman filter is the sensor data at the current moment and the inferred non-directly measurable parameters, and the output is the inferred directly measurable parameters; the sensor data sequence is obtained after data stacking and data normalization of the data collected by the sensors arranged on the aero-engine gas path physical entity; the sensor data at the current moment is factually updated by the data collected by the sensors arranged on the aero-engine gas path physical entity; the accurate inference of the parameter inference model depends on the training of the parameter inference model; the training of the parameter inference model realizes the accurate inference of the non-directly measurable parameters by the parameter inference model; before the training of the parameter inference model starts, a parameter inference model is designed, a value evaluation model is designed, and a data acquisition method for the data required for training the parameter inference model is proposed; the training of the parameter inference model adopts the soft actor-critic algorithm; the parameter inference model is built by a fully connected neural network and a Relu activation function, its input is the sensor data sequence, and its output is the mean of the non-directly measurable parameters and the variance of the non-directly measurable parameters; the value evaluation model is built by a fully connected neural network and a Relu activation function, its input is the sensor data sequence and the non-directly measurable parameters, and its output is the value evaluation of the non-directly measurable parameters; the mean of the non-directly measurable parameters and the variance of the non-directly measurable parameters constitute the distribution of the non-directly measurable parameters; the value of the non-directly measurable parameters is randomly sampled from the distribution of the non-directly measurable parameters; a data acquisition method for the data required for training the parameter inference model realizes the collection of the sensor data sequence data, the inferred non-directly measurable parameter data, and the reward return data required for training the parameter inference model;

[0007] The working process of a data acquisition method for the data required for training the parameter inference model is as follows:

[0008] Step 1: An aero-engine control input sequence is injected into the physical entity of the aero-engine gas path according to the time sequence;

[0009] Step 2: Sensors arranged on the physical entity of the aero-engine gas path collect the gas path data of the aero-engine, stack the data to form a data sequence, and then normalize the sensor data sequence to obtain the sensor data sequence where is the normalized sensor data of the i-th sensor at time t, and N represents the length of this sensor data sequence; A total of M such normalized sensor data sequences are collected;

[0010] Step 3: Input the M collected normalized sensor data sequences into the parameter inference model, and the parameter inference model outputs the inferred distribution of the non-directly measurable parameters after calculation;

[0011] Step 4: Sample an inferred non-directly measurable parameter from the distribution of the non-directly measurable parameters;

[0012] Step 5: Input the sampled non-directly measurable parameter and the aero-engine control input sequence in Step 1 into the aero-engine gas path mechanism model;

[0013] Step 6: Simulate the aero-engine mechanism model and collect its simulation output; Normalize the simulation sensor data sequence output of the simulation to obtain the normalized simulation sensor data sequence where is the normalized simulation data of the i-th sensor corresponding to the gas path state at time t, and N represents the length of this simulation sensor data sequence; A total of M normalized simulation sensor data sequences corresponding to the sensor measurement states are collected;

[0014] Step 7: Calculate the reward return according to the M normalized sensor data sequences obtained in Step 2 and the M normalized simulation sensor data sequences obtained in Step 6; The specific calculation formula is

[0015] Step 8: Combine the M normalized sensor data sequences generated in Step 2, the non-directly measurable parameter generated in Step 4, and the reward return generated in Step 7 into a data tuple, and push the data tuple into the data collection container;

[0016] Step 9: Repeat Steps 1 - 8 to sample multiple data tuples for subsequent training of the parameter inference model by the soft actor-critic algorithm.

[0017] The beneficial effects of the present invention are as follows: The parameter inference model realizes the accurate inference of parameters that cannot be directly measured; the parameters that cannot be directly measured inferred by the parameter inference model are not only synchronized to the digital twin of the aero-engine gas path, but also used to correct the aero-engine gas path mechanism model in the unscented Kalman filter algorithm; the unscented Kalman filter algorithm with the corrected mechanism model realizes better inference of directly measurable parameters, and synchronizes the inferred directly measurable parameters to the digital twin of the aero-engine gas path; finally, the digital twin virtual-real synchronization method of the aero-engine gas path system based on deep reinforcement learning realizes the synchronization of two types of parameters, namely parameters that cannot be directly measured and directly measurable parameters, and solves the problem of incomplete parameter monitoring of the aero-engine gas path system; Description of the Drawings

[0018] Figure 1 is the operation logic diagram of the digital twin virtual-real synchronization method of the aero-engine gas path system based on deep reinforcement learning;

[0019] Figure 2 is the flow chart of the parameter inference model training;

[0020] Figure 3 is the schematic diagram of the neural network structure of the parameter inference model and the value evaluation model;

[0021] Figure 4 is the flow chart of the parameter inference model training data collection method. Detailed Embodiments

[0022] The following further illustrates the detailed embodiments of the present invention in conjunction with the drawings and technical solutions.

[0023] The detailed embodiment of the present invention is a digital twin virtual-real synchronization method for the aero-engine gas path system based on deep reinforcement learning, aiming to effectively monitor the parameters that cannot be directly measured and directly measurable parameters in the aero-engine gas path system. First, by classifying the parameters in the aero-engine gas path system, they are divided into two categories: directly measurable and non-directly measurable according to the difficulty of measurement. Directly measurable parameters can be directly obtained by a single sensor, such as rotational speed, flow rate, temperature, pressure, etc.; while non-directly measurable parameters need to be jointly measured by multiple sensors and obtained through calculation, such as pressure ratio constant, efficiency constant, etc.

[0024] The virtual-real synchronization method of the present invention mainly includes two aspects: one is the virtual-real synchronization of parameters that cannot be directly measured, and the other is the virtual-real synchronization of directly measurable parameters. As Figure 1As shown in the figure, for the non-directly measurable parameters, the present invention performs inference through a parameter inference model, which infers the values of non-directly measurable parameters based on the sensor data sequence of the aero-engine gas path. For the directly measurable parameters, the present invention uses the unscented Kalman filter algorithm to infer the values of directly measurable parameters based on real-time sensor data and the inferred non-directly measurable parameters.

[0025] Specifically, the inferred non-directly measurable parameters will be used to update the corresponding parameters in the parameter table 1 to be synchronized, and correct the corresponding aero-engine gas path mechanism model parameters in the unscented Kalman filter algorithm; while the inferred directly measurable parameters are used to update the relevant values in the parameter table 1 to be synchronized. The parameter table 1 to be synchronized will transfer the synchronized parameter values to the parameter table 2 to be synchronized at a certain update frequency, and the latter is used to construct the digital twin of the aero-engine gas path.

[0026] The training of the parameter inference model is one of the core steps in the implementation of the present invention. Combining Figure 2 and Figure 3 , the soft actor-critic algorithm is adopted in the training process, and a network architecture composed of a fully connected neural network and a Relu activation function is designed. The input of the parameter inference model is the sensor data sequence, and the output is the mean and variance of the non-directly measurable parameters. A non-directly measurable parameter is sampled from the distribution of non-directly measurable parameters composed of the mean and variance of non-directly measurable parameters. In addition, the value evaluation model is also constructed by a fully connected neural network, whose input is the sensor data sequence and the non-directly measurable parameters, and the output is the value evaluation of the non-directly measurable parameters, which is used to optimize the inference result of the parameter inference model.

[0027] Before the training of the parameter inference model, the present invention proposes a method for collecting training data of the parameter inference model, combining Figure 2 and Figure 4 . First, the aero-engine control input sequence is injected into the aero-engine gas path physical entity in chronological order, and the data collected by the sensors is formed into a sensor data sequence after data stacking and data normalization processing. Then, the sensor data sequence is input into the parameter inference model, and the parameter inference model calculates the distribution of non-directly measurable parameters. Next, sampling is performed from this distribution to obtain the inferred non-directly measurable parameters, and they are input into the aero-engine gas path mechanism model for simulation to obtain simulation data. By comparing the differences between the real sensor data and the simulation data, the reward return is calculated. The sensor data sequence, the inferred non-directly measurable parameters, and the reward return are saved to the data collection container. The collection process is repeated multiple times to collect sufficient data for the training of the parameter inference model. Finally, the trained parameter inference model can infer the non-directly measurable parameters in the aero-engine gas path system in practical applications.

[0028] The working principle of the virtual-real synchronization method for the digital twin of an aero-engine gas path system based on deep reinforcement learning is as follows:

[0029] A virtual-real synchronization method for the digital twin of an aero-engine gas path system based on deep reinforcement learning includes the virtual-real synchronization of non-directly measurable parameters and the virtual-real synchronization of directly measurable parameters; a parameter inference model is designed for the virtual-real synchronization of non-directly measurable parameters and the parameter correction of the non-directly measurable parameters corresponding to the internal aero-engine gas path mechanism model of the unscented Kalman filter algorithm; the input of the parameter inference model is the normalized sensor data sequence, and the output is the inferred non-directly measurable parameters; the unscented Kalman filter algorithm is used to realize the inference of non-direct parameters; the input of the unscented Kalman filter is the sensor data at the current moment and the inferred non-directly measurable parameters, and the output is the inferred directly measurable parameters; the normalized sensor data sequence is obtained after data stacking and data normalization processing of the data collected by the sensors arranged on the physical entity of the aero-engine gas path; the sensor data at the current moment is updated by the data fact collected by the sensors arranged on the physical entity of the aero-engine gas path; finally, the non-directly measurable parameters inferred by the parameter inference model and the directly measurable parameters inferred by the unscented Kalman filter algorithm are synchronized to the digital twin of the aero-engine gas path;

[0030] Combined with Figures 1 - 4 , the working process of the virtual-real synchronization method for the digital twin of an aero-engine gas path system based on deep reinforcement learning is as follows:

[0031] Step 1: Combined with Figure 1 , the parameters in the aero-engine gas path system are divided into directly measurable parameters and non-directly measurable parameters according to the measurement difficulty; the directly measurable parameters include rotational speed, flow rate, temperature, pressure, etc., which can be measured by a single sensor; the non-directly measurable parameters include pressure ratio constant, efficiency constant, etc., which need to be jointly measured and calculated by multiple sensors; the non-directly measurable data and directly measurable data to be synchronized are represented by two identical parameter tables to be synchronized; they are named parameter table to be synchronized 1 and parameter table to be synchronized 2 respectively; parameter table to be synchronized 1 is located on the side of the virtual-real synchronization method for the digital twin of the aero-engine gas path system based on deep reinforcement learning; parameter table to be synchronized 2 is located in the aero-engine gas path digital twin; parameter table to be synchronized 1 is required to update the parameter data to parameter table to be synchronized 2 at a certain frequency;

[0032] Step 2: A digital twin virtual-real synchronization method for an aero-engine gas path system based on deep reinforcement learning is divided into two parts: virtual-real synchronization of non-directly measurable parameters and virtual-real synchronization of directly measurable parameters; virtual-real synchronization of non-directly measurable parameters realizes the monitoring of non-directly measurable parameters; virtual-real synchronization of directly measurable parameters realizes the monitoring of directly measurable parameters; inferring non-directly measurable parameters depends on a parameter inference model; the unscented Kalman filter algorithm is used to realize the inference of non-directly measurable parameters;

[0033] Step 3: Parameter inference model training is designed to accurately infer non-directly measurable parameters of the parameter inference model; the soft actor-critic algorithm is adopted during the training of the parameter inference model; a parameter inference model, a value evaluation model, and a data acquisition method for the data required to train the parameter inference model are proposed to carry out the training of the parameter inference model by the soft actor-critic algorithm;

[0034] Step 4: Combine Figure 2 , design a parameter inference model and a value evaluation model; the parameter inference model is built by a fully connected neural network and a Relu activation function, its input is the sensor data sequence, and the output is the mean of non-directly measurable parameters and the variance of non-directly measurable parameters; the parameter inference model starts from the sensor data sequence as the input; the sensor data sequence first passes through two layers of fully connected networks, each layer contains 512 neurons and each layer uses the ReLU activation function; subsequently, the obtained features are calculated in two branches; the network structures of the two branches are the same and both contain two layers of fully connected networks, the first layer of the branch contains 512 neurons, the second layer of the branch contains 256 neurons, and each layer of the branch uses the ReLU activation function; finally, the two branches of the parameter inference model respectively output "the mean of non-directly measurable parameters" and "the variance of non-directly measurable parameters; the value evaluation model is built by a fully connected neural network and a Relu activation function, its input is the sensor data sequence and non-directly measurable parameters, and the output is the value evaluation of non-directly measurable parameters; the input of the value evaluation model includes "sensor data sequence" and "non-directly measurable parameters"; the sensor data sequence first passes through two layers of fully connected networks, each layer contains 512 neurons, and the ReLU activation function is used; the non-directly measurable parameters pass through two layers of fully connected networks, each layer also contains 512 neurons, and the ReLU activation function is used; the features output by the two-way networks are concatenated in subsequent layers, and the concatenated features are processed by two layers of fully connected networks, each layer contains 256 neurons, and the activation function is ReLU; finally, the value evaluation model outputs the value evaluation of non-directly measurable parameters;

[0035] Step 5: Combine Figure 3 , design a data acquisition method for the data required to train the parameter inference model, and the working process of this method is as follows:

[0036] 1) An aero-engine control input sequence is injected into the physical entity of the aero-engine gas path according to the time sequence.

[0037] 2) Sensors arranged on the physical entity of the aero-engine gas path collect the gas path data of the aero-engine, stack the data to form a data sequence, and then normalize the sensor data sequence to obtain the sensor data sequence where is the normalized sensor data of the i-th sensor at time t, and N represents the length of this sensor data sequence; a total of M such normalized sensor data sequences are collected.

[0038] 3) Input the M collected normalized sensor data sequences into the parameter inference model, and the parameter inference model outputs the inferred distribution of non-directly measurable parameters after calculation.

[0039] 4) Sample an inferred non-directly measurable parameter from the distribution of non-directly measurable parameters.

[0040] 5) Input the sampled non-directly measurable parameter and the aero-engine control input sequence in step 1 into the aero-engine gas path mechanism model.

[0041] 6) Simulate the aero-engine mechanism model and collect its simulation output; normalize the simulation output of the sensor data sequence to obtain the normalized simulation sensor data sequence where is the normalized simulation data of the i-th sensor corresponding to the gas path state at time t, and N represents the length of this simulation sensor data sequence; a total of M normalized simulation sensor data sequences corresponding to the sensor measurement states are collected.

[0042] 7) Calculate the reward return according to the M normalized sensor data sequences obtained in 2) and the M normalized simulation sensor data sequences obtained in 6); the specific calculation formula is

[0043] 8) Combine the M normalized sensor data sequences generated in step 2, the non-directly measurable parameter generated in step 4, and the reward return generated in step 7 into a data tuple, and push the data tuple into the data collection container.

[0044] 9) Repeat 1)-8) to sample multiple data tuples for subsequent training of the parameter inference model by the soft actor-critic algorithm.

[0045] Step 6: Combine Figure 2, a parameter inference model and a value evaluation model in (3), and a data acquisition method for the data required to train the parameter inference model in step 5 are used to train the parameter inference model by the soft actor-critic algorithm; the soft actor-critic algorithm outputs the trained parameter inference model;

[0046] Step 7: Combine Figure 1 , the trained parameter inference model is deployed to the digital twin virtual-real synchronization method of the aero-engine gas path system based on deep reinforcement learning;

[0047] Step 8: The data collected by the sensors arranged on the aero-engine gas path physical entity are processed by data stacking and data normalization to obtain a sensor data sequence; the normalized sensor data sequence is input into the parameter inference model; the parameter inference model outputs the inferred non-directly measurable parameters through calculation; the inferred non-directly measurable parameters are respectively used for the parameter update of the non-directly measurable parameters in the to-be-synchronized parameter table 1 and the parameter correction of the corresponding non-directly measurable parameters in the internal mechanism model of the unscented Kalman filter algorithm;

[0048] Step 9: The data fact of the data collected by the sensors arranged on the aero-engine gas path physical entity is updated to obtain the sensor data at the current moment; the sensor data at the current moment and the non-directly measurable parameters inferred by the parameter inference model in step 7 are input into the unscented Kalman filter algorithm; the unscented Kalman filter algorithm outputs the inferred directly measurable parameters;

[0049] Step 10: Combine Figure 1 , the non-directly measurable parameters inferred in step 8 and the directly measurable parameters inferred in step 9 are used to update the to-be-synchronized parameter table 1; the to-be-synchronized parameter table 1 synchronizes the parameter values of the non-directly measurable parameters and the directly measurable parameters to the to-be-synchronized parameter table 2 at a certain update frequency; the to-be-synchronized parameter table 2 is used for subsequent construction of the aero-engine gas path digital twin.

Claims

1. A virtual-real synchronization method for digital twins of an aircraft engine gas path system based on deep reinforcement learning, characterized in that: The virtual-real synchronization method of the digital twin of the aero-engine gas path system includes virtual-real synchronization of parameters that cannot be directly measured and virtual-real synchronization of parameters that can be directly measured; A parameter inference model is designed for the virtual-real synchronization of the parameters that cannot be directly measured and the parameter correction of the internal aircraft engine gas path mechanism model of the unscented Kalman filter algorithm corresponding to the parameters that cannot be directly measured; the input of the parameter inference model is the normalized sensor data sequence, and the output is the inferred indirect measurement parameters; the unscented Kalman filter algorithm is used to realize the inference of the parameters that cannot be directly measured; the input of the unscented Kalman filter is the current sensor data and the inferred indirect measurement parameters, and the output is the inferred direct measurement parameters; The normalized sensor data sequence is obtained by stacking and normalizing the data collected by the sensors arranged on the physical entity of the aircraft engine gas path; the sensor data at the current moment is obtained by updating the data collected by the sensors arranged on the physical entity of the aircraft engine gas path in real time; Finally, the non-directly measurable parameters inferred by the parameter inference model and the directly measurable parameters inferred by the unscented Kalman filter algorithm are synchronized to the digital twin of the aircraft engine gas path; Directly measurable parameters include speed, flow, temperature, and pressure, which are measured by a single sensor.

2. The virtual-real synchronization method of the digital twin of the aero-engine gas path system according to claim 1 is characterized in that: The workflow of the virtual-real synchronization method of the digital twin of the aero-engine gas path system based on deep reinforcement learning is as follows: Step 1: Divide the parameters in the air path system of an aircraft engine into directly measurable parameters and indirect measurable parameters according to the difficulty of measurement; directly measurable parameters include speed, flow, temperature, and pressure, which are measured by a single sensor; indirect measurable parameters include flow constant, pressure ratio constant, efficiency constant, and speed constant, which require multiple sensors to measure and calculate together; synchronize the indirect measurable data and the directly measurable data that need to be synchronized using two identical parameter tables to be synchronized; they are named parameter table 1 to be synchronized and parameter table 2 to be synchronized respectively; parameter table 1 to be synchronized is located on the virtual-real synchronization method side of the digital twin of the air path system of an aircraft engine based on deep reinforcement learning; parameter table 2 to be synchronized is located in the air path twin of the aircraft engine; The parameter table 1 to be synchronized is required to update parameter data to the parameter table 2 to be synchronized at a certain frequency; Step 2: The virtual-real synchronization method of the digital twin of the air path system of the aircraft engine is divided into two parts: the virtual-real synchronization of the parameters that cannot be directly measured and the virtual-real synchronization of the parameters that can be directly measured; the virtual-real synchronization of the parameters that cannot be directly measured realizes the monitoring of the parameters that cannot be directly measured; the virtual-real synchronization of the parameters that can be directly measured realizes the monitoring of the parameters that can be directly measured; the inference model of the parameters that cannot be directly measured is inferred; the unscented Kalman filter algorithm is used to realize the inference of the parameters that cannot be directly measured; Step 3: Design parameter inference model training to achieve accurate inference of the parameter inference model. The parameters cannot be directly measured. The soft actor-critic algorithm is used in the parameter inference model training. A parameter inference model, a value assessment model and a data collection method for training the parameter inference model are proposed to train the parameter inference model using a soft actor-critic algorithm. Step 4: Design a parameter inference model and a value assessment model; the parameter inference model is built by a fully connected neural network and a Relu activation function, with the input of the sensor data sequence and the output of the mean of the parameter that cannot be directly measured and the variance of the parameter that cannot be directly measured; the parameter inference model starts with the sensor data sequence as input; the sensor data sequence first passes through a two-layer fully connected network, each layer of which contains 512 neurons and each layer uses a ReLU activation function; then, the obtained features are divided into two branches for calculation; the network structures of the two branches are the same and both contain a two-layer fully connected network, the first layer of the branch contains 512 neurons, the second layer of the branch contains 256 neurons, and each layer of the branch uses a ReLU activation function; finally, the two branches of the parameter inference model output the "mean of the parameter that cannot be directly measured" and "the variance of the parameter that cannot be directly measured" respectively. The variance of the directly measured parameter; the value assessment model is built by a fully connected neural network and a Relu activation function, with the input of the sensor data sequence and the parameter that cannot be directly measured, and the output is the value assessment of the parameter that cannot be directly measured; the input of the value assessment model includes "sensor data sequence" and "parameter that cannot be directly measured"; the sensor data sequence first passes through two layers of fully connected networks, each layer contains 512 neurons, and uses the ReLU activation function; the parameter that cannot be directly measured passes through two layers of fully connected networks, each layer also contains 512 neurons, and uses the ReLU activation function; the features output by the two networks are spliced ​​in the subsequent layers, and the spliced ​​features are processed by two layers of fully connected networks, each layer contains 256 neurons, and the activation function is ReLU; finally, the value assessment model outputs the value assessment of the parameter that cannot be directly measured; Step 5: Design a data collection method for training the parameter inference model. The workflow of this method is as follows: 1) An aircraft engine control input sequence is injected into the aircraft engine gas path physical entity according to the time sequence; 2) The sensors arranged on the physical entity of the aircraft engine gas path collect the gas path data of the aircraft engine, stack the data to form a sensor data sequence, and then normalize the sensor data sequence to obtain the sensor data sequence in is the normalized sensor data of the i-th sensor at time t, N represents the length of the sensor data sequence; a total of M normalized sensor data sequences are collected; 3) Input the collected M normalized sensor data sequences into the parameter inference model, and the parameter inference model calculates and outputs the distribution of the inferred non-directly measurable parameters; 4) sampling an inferred non-directly measurable parameter from a non-directly measurable parameter distribution; 5) inputting the sampled non-directly measurable parameters and the aircraft engine control input sequence of step 1 into the aircraft engine gas path mechanism model; 6) Simulate the aircraft engine gas path mechanism model and collect its simulation output; Normalize the simulated sensor data sequence output to obtain a normalized simulated sensor data sequence in is the normalized simulation data of the gas path state corresponding to the i-th sensor at time t, N represents the length of the simulation sensor data sequence; a total of M normalized simulation sensor data sequences corresponding to the sensor measurement state are collected; 7) Calculate the reward return based on the M normalized sensor data sequences obtained in step 2) and the M normalized simulated sensor data sequences obtained in step 6); the specific calculation formula is: 8) combining the M normalized sensor data sequences generated in step 2), the non-directly measurable parameters generated in step 4), and the reward returns generated in step 7) into a data tuple, and pushing the data tuple into a data collection container; 9) Repeat steps 1)-8) to sample multiple data tuples for subsequent soft actor-critic algorithm training parameter inference model; Step 6: Combining a parameter inference model and a value assessment model, a data collection method for the data required for training the parameter inference model in step 5) is used for training the parameter inference model by the soft actor-critic algorithm; the soft actor-critic algorithm outputs the trained parameter inference model; Step 7: The trained parameter inference model is deployed to the virtual-real synchronization method of the digital twin of the aircraft engine gas path system based on deep reinforcement learning; Step 8: Data collected by sensors arranged on the physical entity of the aircraft engine gas path are stacked and normalized to obtain a sensor data sequence; the normalized sensor data sequence is input into the parameter inference model; the parameter inference model outputs the inferred non-directly measurable parameters after calculation; the inferred non-directly measurable parameters are respectively used for parameter update of the non-directly measurable parameters in the parameter table 1 to be synchronized and parameter correction of the non-directly measurable parameters corresponding to the internal mechanism model of the unscented Kalman filter algorithm; Step 9: The data collected by the sensors arranged on the physical entity of the aircraft engine gas path is updated to obtain the sensor data at the current moment; The current moment sensor data and the non-directly measurable parameters inferred by the parameter inference model in step 7) are input into the unscented Kalman filter algorithm; the unscented Kalman filter algorithm outputs the inferred directly measurable parameters; Step 10: The non-directly measurable parameters inferred in step 8) and the directly measurable parameters inferred in step 9) are combined to update the parameter table 1 to be synchronized; The parameter table 1 to be synchronized is synchronized with the parameter table 2 to be synchronized with the parameter values ​​of the non-directly measurable parameters and the directly measurable parameters at a certain update frequency; The parameter table 2 to be synchronized is used for the subsequent construction of the digital twin of the aircraft engine gas path.

Images