An aeroengine digital twin model

Abstract

This invention specifically relates to a digital twin model of an aero-engine, comprising a data input layer, a coupling network layer, and a mapping network layer. The coupling network layer extracts and couples data from the data input layer, and then outputs the target value through the mapping network layer. The data input layer includes a current sortie performance network layer and a sortie-influenced physical characteristic extraction network layer. The current sortie performance network layer extracts physical information from the current sortie, and the sortie-influenced physical characteristic extraction network layer extracts physical information from the k previous sorties. The aim is to achieve higher accuracy than traditional models by considering the performance impact between previous sorties throughout the aircraft's flight lifecycle.

Description

Technical Field

[0001] This invention belongs to the field of aircraft performance tracking / health monitoring technology, specifically relating to a digital twin model of an aircraft engine. Background Technology

[0002] For aircraft, the core component is the aircraft engine. The performance of the aircraft engine directly affects flight safety and economic efficiency. Therefore, aircraft health monitoring systems primarily rely on real-time monitoring and prediction of aircraft engine performance parameters. Statistics show that in the past 30 years, at least 60 accidents have occurred due to engine fires, shutdowns, blade failures, and other causes. If the engine's operating status could be monitored or predicted before flight or before an accident, allowing for an assessment of the engine and aircraft's health level, and enabling appropriate measures to prevent the spread or occurrence of faults and eliminate potential flight safety hazards, this would be of paramount importance for preventing and eliminating major safety risks.

[0003] Aircraft maintenance theory is mainly divided into three types: restorative maintenance, preventive maintenance, and predictive maintenance. For predictive maintenance, traditional modeling does not consider flight process information, treating all data as equivalent. However, in actual flight, after several flights, the performance of a single flight (also called a flight sequence, referring to a complete flight mission from takeoff and steady-state flight to landing) has already degraded. This degradation occurs gradually with the accumulation of flights, and the generation relationship is not linear. Because the influence of each flight is ignored, the prediction results of this traditional modeling method deviate from the current state that needs to be predicted (for example, the prediction result for the nth flight is actually the result for the n / 2th flight), resulting in low overall prediction accuracy. Summary of the Invention

[0004] To address the issues of inaccuracy and low precision in traditional prediction models for aero-engines, this invention provides a digital twin model for aero-engines. This model considers the performance impact between flight sorties throughout the aircraft's flight lifecycle, thus achieving higher precision than traditional models.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: an aircraft health monitoring model, comprising a data input layer, a coupling network layer, and a mapping network layer. The coupling network layer extracts and couples the data from the data input layer, and then outputs the target value through the mapping network layer. The data input layer includes a current sortie performance network layer and a sortie impact physical characteristic extraction network layer. The current sortie performance network layer is used to extract physical information from the current sortie, and the sortie impact physical characteristic extraction network layer is used to extract physical information from the k sorties preceding the current sortie, where k is an integer not less than 1.

[0006] Preferably, the current sortie performance network layer is the nth sortie performance network layer. The physical information extracted by the nth sortie performance network layer is the performance status parameter corresponding to any moment in the nth sortie, and the extracted performance status parameter corresponding to any moment is transmitted to the coupling network layer. The sortie impact physical characteristic extraction network layer is composed of the (n-1-k)th sortie performance network layer to the (n-1)th sortie performance network layer. The sortie impact physical characteristic extraction network layer extracts the performance status parameters corresponding to partial or all moments in each sortie from the (n-1-k)th sortie to the (n-1)th sortie, and transmits the extracted performance status parameters corresponding to partial or all moments to the coupling network layer.

[0007] Preferably, when the physical characteristics of the flight are extracted from the (n-1-k)th to (n-1)th flights, the performance state parameters corresponding to local moments within each flight are randomly extracted, and the randomly extracted data covers all flight states in the entire flight.

[0008] Preferably, when the physical characteristics extraction network layer of the flight number extracts the performance state parameters corresponding to local moments, it is selected according to the proportion of flight state data during the flight process. The extracted time state data is consistent with the data proportion of each state in the flight state data of the flight number. The flight state includes takeoff, cruise flight, and descent.

[0009] Preferably, the proportion of takeoff and landing data extracted by the flight impact physical characteristic extraction network layer is based on a multiple of the takeoff and landing data throughout the entire flight process; the total amount of takeoff, cruise, and landing data extracted by the flight impact physical characteristic extraction network layer for a flight is 0.1%-1% of the total data volume for a flight; the time interval for collecting performance status parameters in the performance network layer of the nth flight is 0.01s~1s.

[0010] Preferably, from the (n-1-k)th sortie to the (n-1)th sortie, the total amount of data extracted for takeoff, cruise flight, and descent gradually increases in each sortie.

[0011] Preferably, the physical information includes the physical operational characteristics and patterns of various components and subsystems during the operation of the aircraft.

[0012] Preferably, the current flight performance network layer and the flight impact physical characteristic extraction network layer adopt any one of LSTM time series network, GRU time series network, ARIMA time series network, WaveNet network and attention mechanism network.

[0013] Preferably, when the current flight performance network layer and the flight impact physical characteristic extraction network layer are time-series networks, the inputs are the characteristic parameters of each subsystem at the corresponding time. The characteristic parameters include the characteristic parameters of the inlet network, fan component network, compressor component network, combustion chamber component network, lubricating oil system network, turbine component network, mixing chamber component network, and tail nozzle component network.

[0014] Preferably, the coupling network layer is coupled using a neural network method; the mapping relationship in the mapping network layer is mainly the mapping relationship of the fully connected network layer in the neural network; the target value is the performance parameter that needs to be monitored and predicted during actual flight, including engine exhaust temperature, thrust and fuel consumption rate.

[0015] The present invention has the following beneficial effects:

[0016] 1. Based on predictive maintenance technology, this invention provides an aircraft health monitoring model. This model considers the cumulative effects between flights during actual flight and combines data from historical flights with information from the current flight to analyze and predict the health status of a specific moment within the current flight. It is built using neural network methods in machine learning, and then automatically analyzes data to obtain patterns, using these patterns to predict unknown data. Therefore, it has higher accuracy, efficiency, and physical characteristic representation. The aircraft health monitoring technology proposed in this invention has higher safety and reliability compared to traditional maintenance methods. At the same time, this technology can significantly reduce aircraft maintenance costs and extend and improve the service life of aircraft.

[0017] 2. This invention employs a local memory learning model, which selects corresponding performance state parameters for each flight stage according to the proportion of the flight stages in a single flight. These performance state parameters are used to characterize the overall performance state of the flight, thereby avoiding the huge computational resources required when taking overall performance characteristics and improving the overall learning and operation speed.

[0018] 3. In a local memory learning model of the present invention, for different moments in the current flight, the data sets in the network layer corresponding to the flight's physical characteristics at each moment are different. Therefore, during training, multiple iterative calculation operations are performed, so the overall data of the previous flight can be traversed many times, ensuring that the overall situation is learned and memorized through multiple random local memories, making the constructed model more accurate. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the present invention;

[0020] Figure 2 Data flow for a local memory learning model that considers the impact between sorties;

[0021] Figure 3 For the nth flight performance network layer;

[0022] Figure 4 This is the network layer for the first flight's performance. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0024] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0025] like Figure 1As shown, a digital twin model of an aero-engine includes a data input layer, a coupling network layer, and a mapping network layer. The coupling network layer S21 and the mapping network layer S31 extract and couple data from the data input layer, then output the target value through the mapping network layer S31. The data input layer includes a current sortie performance network layer (the nth sortie performance network layer S03) and a sortie-affected physical characteristic extraction network layer S11. The current sortie performance network layer extracts physical information from the current sortie, and the sortie-affected physical characteristic extraction network layer S11 extracts physical information from the k previous sorties, where k is an integer not less than 1. Specifically, it extracts the performance change patterns of different sorties from the (n-1-k)th sortie performance network layer S01 to the (n-1)th sortie performance network layer S02, where k < n. Due to the massive amount of time data in a single sortie, these k historical sortie performance network layers employ local memory learning to learn thematic knowledge and key knowledge points. The coupling network layer S21 couples the physical information extracted from the flight impact physical characteristics extraction network layer S11 and the nth flight performance network layer S03, and extracts performance-related features containing flight impact factors. Specifically, the feature extraction here mainly uses deep learning methods (mainly neural networks) in artificial intelligence technology. This method is currently the mainstream technology in the field of artificial intelligence, and its specific details will not be elaborated here.

[0026] The current sortie performance network layer is the nth sortie performance network layer. The physical information extracted by the nth sortie performance network layer is the performance status parameter corresponding to any moment in the nth sortie, and the extracted performance status parameter corresponding to any moment is transmitted to the coupling network layer. The sortie impact physical characteristic extraction network layer is composed of the (n-1-k)th sortie performance network layer to the (n-1)th sortie performance network layer. The sortie impact physical characteristic extraction network layer extracts the performance status parameter corresponding to local or all moments in each sortie from the (n-1-k)th sortie to the (n-1)th sortie, and transmits the extracted performance status parameter corresponding to local or all moments to the coupling network layer.

[0027] For the local memory learning model, the specific implementation details mainly focus on the physical feature extraction network layer S11, which is influenced by the flight number, and its subordinate sub-networks, as well as the performance network layer S03 for the nth flight (i.e., the current flight). Here, the physical feature extraction network layer S11 has k subordinate sub-networks, including the performance network layers S01 for the (n-1-k)th flight to S02 for the (n-1)th flight. Since the current flight is most affected by its most recent flight, in order to reduce computational resources and improve computational speed, the optimal value of k is 1-3.

[0028] The performance correlation characteristics containing the impact of sortie rate mainly refer to the impact of the operating status of the previous k sorties on the performance of the current sortie. That is, the internal changes of the engine caused by environmental changes and the interaction between internal components during the previous k sorties, such as wear and thermal deformation. If this change is not considered, the prediction of the current sortie's status will have a large deviation. This deviation is the performance degradation impact deviation.

[0029] The function of the mapping network layer S31 is to map the information of the coupled network layer S21 to the final target output value. The mapping relationship in the mapping network layer is mainly the mapping relationship of the fully connected network layer in the neural network. The target value is the performance parameter that needs to be monitored and predicted during actual flight, including engine exhaust temperature, thrust and fuel consumption rate.

[0030] Furthermore, the performance state parameters corresponding to local moments within each flight from the (n-1-k)th to the (n-1)th flight are randomly selected in the flight-influence-physical-characteristic extraction network layer. The randomly selected data covers all flight states throughout the entire flight. Specifically, in the (n-1-k)th flight, performance state parameters corresponding to several moments are randomly selected. These performance state parameters are used to characterize the overall information of the (n-1-k)th flight, and so on up to the (n-1)th flight. For each flight, several performance state parameters corresponding to different moments are randomly selected.

[0031] In other words, since the amount of data in the entire flight is large, random sampling can not only fully reflect the information of the entire flight, but also greatly reduce the amount of data to be calculated.

[0032] Furthermore, when the physical characteristic extraction network layer extracts performance state parameters corresponding to local moments in the flight sequence, it selects them according to the proportion of flight state data during the flight process. The extracted time state data is consistent with the data proportion of each state in the flight state data of the flight sequence, whereby the flight states include takeoff, cruise flight, and descent. In each flight, x time state data points are randomly selected, and the selected x time state data points are selected according to the proportion of flight state data during the flight process, that is, the extracted x time state data points are consistent with the data proportion of each state in the flight state data during the flight process, whereby the flight states include takeoff, cruise flight, and descent.

[0033] Taking flight n-1 as an example, x time state data points are randomly selected from all time data of flight n-1 to characterize the overall performance status of flight n-1. These x time state data points are not taken from all time state data points, but are taken proportionally according to the state during the flight process, so as to ensure that the x time state data points include all flight states, which mainly include takeoff, cruise flight, and descent.

[0034] Preferably, the proportion of each state in x time data points is consistent with the proportion of each state in all time data points of n-1 flights, and each state should have at least one time data point.

[0035] Furthermore, the proportion of takeoff and landing data extracted by the flight impact physical characteristic extraction network layer is based on a multiple of the takeoff and landing data throughout the entire flight process; the total amount of takeoff, cruise, and landing data extracted by the flight impact physical characteristic extraction network layer for a flight is 0.1%-1% of the total data volume for a flight; the time interval for collecting performance status parameters in the performance network layer of the nth flight is 0.01s~1s.

[0036] Specifically, since there is significantly less data for takeoff and descent compared to cruise, the proportion of takeoff and descent data can be appropriately increased. That is, let i:j:k represent the proportion of takeoff, cruise, and descent data in a randomly selected flight, and m:n:l represent the proportion of takeoff, cruise, and descent data in the entire flight. The data is extracted in multiples, i:j:k = (1.1m):n:(1.1l). The sum of i+j+k is 0.1%-1% of the total data volume for a single flight.

[0037] For the current flight, flight n, its task is to monitor and predict the performance status at each time point. The time interval for data collection is often on the order of 0.01 seconds to 1 second. For example, a one-hour flight often requires thousands of performance predictions. For ease of explanation later, we will use n_t to represent the total number of monitoring / prediction attempts required for flight n, and n_k to represent the k-th monitoring attempt in flight n. To accurately predict the performance status of all n_t time points in flight n, the performance network layer S03 of flight n is constructed based on the performance status at each n_k time point, and therefore its input is the performance status parameters at n_k time points.

[0038] Unlike the performance network layer S03 of the nth sortie, the performance network layers of the first k sorties do not extract the performance features of a certain state. In order to better reflect the actual operating state, they extract the performance features of the entire process in the corresponding sortie. For the current sortie (the nth sortie), each time point needs to be predicted. Therefore, for the current sortie itself, its focus is on the performance state at each time point. As for the influence of its previous sorties, it should be the overall influence. That is, any time point in n_t should be affected by all states of the n-1 sorties. In order to avoid a large amount of computation, the k most recent sorties with the greatest influence are selected.

[0039] For each time state of the current flight, the x time states of the preceding flights should be different. This way, after the current flight has traversed n_t states, all states of the preceding flights have also been traversed, and the number of traversals is x times that of the current flight. In this way, the performance characteristics of the previous k flights, as well as the current flight itself, can be learned through a local learning strategy.

[0040] Furthermore, from the (n-1-k)th sortie to the (n-1)th sortie, the total amount of data extracted for takeoff, cruise flight, and descent gradually increases in each sortie.

[0041] Specifically, considering that the impact of previous flights on the current flight will weaken as the time interval increases, the data values ​​for earlier flights are reduced to save computing resources. That is, the amount of data taken for the (n-1-k)th flight is less than the amount of data taken for the (n-1)th flight.

[0042] Furthermore, the physical information includes the physical operating characteristics and laws of various components and subsystems during the operation of an aircraft, such as the physical laws of thermal flow, control laws, altitude characteristics, speed characteristics, and throttling characteristics inside the engine.

[0043] Furthermore, the current flight performance network layer and the flight impact physical characteristic extraction network layer adopt any one of the following: LSTM time series network, GRU time series network, ARIMA time series network, WaveNet network, and attention mechanism network.

[0044] Furthermore, when the current flight performance network layer and the flight impact physical characteristic extraction network layer are time-series networks, the inputs are the characteristic parameters of each subsystem at the corresponding time. The characteristic parameters include the characteristic parameters of the intake network, fan component network, compressor component network, combustion chamber component network, lubricating oil system network, turbine component network, mixing chamber component network, and tail nozzle component network.

[0045] Furthermore, the coupling network layer adopts a neural network method for coupling; the mapping relationship in the mapping network layer is mainly the mapping relationship of the fully connected network layer in the neural network; the target value is the performance parameter that needs to be monitored and predicted during actual flight, including engine exhaust temperature, thrust and fuel consumption rate.

[0046] Example 1

[0047] Taking k=2 as an example, that is, in this embodiment, the network layer S11 that extracts the physical characteristics of the flight number takes the n-2 and n-1th flights from the first two historical flight numbers, such as... Figure 2As shown, the flight data collected by an aero-engine in one flight is stored in a two-dimensional data space, where each row of data represents the characteristic parameters F = (F1, ..., F2) at the same time. n The characteristic parameters in this invention are those that play a key role in overall performance, such as atmospheric parameters: altitude, Mach number, etc.; state parameters: compressor inlet guide angle, high and low pressure speeds, throttle lever position, exhaust nozzle diameter and position, etc.; performance parameters: fuel flow rate, low-pressure turbine afterburner total gas temperature, etc.; in particular, (F 1,0 ,...,F n,0 (F) represents the characteristic parameter at time 0 (t=0), and so on, (F) 1,m ,...,F n,m The denoted ) represents the feature parameters at time m. Here, the feature parameters at t=1 do not refer to the feature parameters collected in the first second of the flight, but rather to the feature parameters collected after the first time interval of the sampling frequency. The sampling frequency in this invention is 0.01s to 1s. Specifically, since the flight conditions of different sorties are different, their total duration is not necessarily the same; that is, the number of data entries for each single sortie is not necessarily m. To better understand and explain the scheme of this invention, Figure 2 The total number of data entries for the (n-2), (n-1), and nth flights is m.

[0048] To obtain the physical correlation features between the feature parameters F at each time point in the nth flight, we will take time point 0 as an example, and its performance network layers are as follows: Figure 3 As shown (the same applies to times t=1, t=2...t=n). The time-series mapping network contains time information, and the input to this network layer can be parameters related to time information, such as time. The intake duct network, fan component network, compressor component network, combustion chamber component network, lubrication system network, turbine component network, mixing chamber component network, and exhaust nozzle component network are respectively engine component networks, and their inputs are the feature parameters of each component at the corresponding time in the time-series network. Specifically, the time-series network contains... Figure 2 The data space of the nth sortie contains the longitudinal time information, while the input of the other component networks is the lateral feature parameter F of the data space, which respectively represent the operating rules of each component of the engine at that time.

[0049] Specifically, the characteristic parameters corresponding to the intake duct network are intake duct parameters, such as intake pressure and temperature; the characteristic parameters corresponding to the fan component network are fan parameters, such as low-pressure fan speed; the characteristic parameters corresponding to the compressor component network are compressor parameters, such as compressor guide angle and high-pressure speed; the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, such as fuel flow and pressure; the characteristic parameters corresponding to the lubricating oil system network are lubricating oil system parameters, such as lubricating oil differential pressure, lubricating oil flow, and lubricating oil temperature; the characteristic parameters corresponding to the turbine component network are turbine component parameters, such as turbine exhaust temperature and speed; and the characteristic parameters corresponding to the exhaust nozzle component network are exhaust nozzle component parameters, such as exhaust nozzle diameter or exhaust nozzle distance.

[0050] When performing data coupling, feature parameters of all times in the (n-2)th and (n-1)th flights are extracted, as well as feature parameters of a certain time (the time to be predicted) in the performance network layer of the nth flight. All the extracted feature parameter information is coupled through the coupling network layer S21, and finally the target value is output through the mapping network layer S31. By combining the cumulative influence of historical flights, the flight characteristics at a certain time are predicted.

[0051] Example 2

[0052] The difference from the above embodiments is that, in this embodiment, the feature parameters extracted from the (n-2)th and (n-1)th flights are feature parameters from local moments. Including all features from all moments in the data space of the (n-1)th and (n-2)th flights would make the overall model structure too large, leading to slow computation or fitting problems. To improve the model's accuracy, this embodiment uses a local memory method, that is, extracting data points from a subset of moments for each flight state. This simplifies the calculation and prevents overfitting.

[0053] like Figure 4 The diagram shows the network layer for the first flight, which consists of i climb state random networks, j cruise state networks, and k descent state random networks. The input to the climb state random network is the characteristic parameter data at random moments during the climb; the input to the cruise state random network is the characteristic parameter data at random moments during the cruise; and the input to the descent state random network is the characteristic parameter data at random moments during the descent. Specifically, the characteristic parameters here refer to all characteristic parameters, not just the parameters of a single component; climb state refers to the state before the aircraft takes off and climbs, i.e., before reaching cruise; cruise state refers to the state of stable flight at a specific altitude; and descent state refers to the state of diving and landing, i.e., leaving cruise.

[0054] Specifically, for the ascending state random network T1 to T iIt should satisfy a temporal sequence (for example, assuming the period from 0 to 200 is the takeoff state, i is 3, and feature parameters are randomly selected from 0 to 200 at times t=2, 54, and 100, then the feature parameter at t=2 is used as the feature parameter input for T1, the feature parameter at t=54 is used as the feature parameter input for T2, and the feature parameter at t=100 is used as the feature parameter input for T3). For the cruise state random network C1 to C... j and descent state random network D1 to D k Similarly, in this way, Figure 4 The network layers are arranged strictly in chronological order from left to right.

[0055] Specifically, the i:j:k ratio is related to the data ratio of takeoff, cruise, and descent states in the corresponding flight sortie. Assuming the data ratio of takeoff, cruise, and descent states in a sortie is m:n:l, then i:j:k can be consistent with m:n:l. However, since takeoff and descent states are much less frequent than cruise states, the ratio of i and k can be appropriately increased. Here, we suggest adjusting the ratio as follows:

[0056] i:j:k=(1.1m):n:(1.1l)

[0057] In addition, the sum of i+j+k is (0.1%-1%) * the total number of flight data. The formula for the value of i+j+k for the first k flights can be calculated consistently. However, considering that the influence of the previous flights on the current flight will weaken as the time interval increases, it is recommended to reduce the value of the data for earlier flights to save computing resources (for example, assuming the number of data entries for the first n-1 and n-2 flights is m1 and m2, respectively, if i+j+k for the n-1 flight is 1% * m1, then i+j+k for the n-2 flight is 0.09% * m1, and the same applies to earlier flights).

[0058] Specifically, as mentioned above, for the nth flight at time 0, only 0.1%-1% of the data from the previous n-1 and n-2 flights are considered. Since global memory is not performed, this approach is defined as a local memory method in this invention.

[0059] The specific implementation method is as follows:

[0060] S1: Perform the first data coupling and randomly extract the feature parameters corresponding to x times in the (n-2)th flight. x is extracted according to the proportion of takeoff, cruise flight and descent states in the whole flight. The number of extractions in each state is not less than 2. In this way, randomly extract the feature parameters corresponding to x times in the (n-1)th flight.

[0061] S2: Extract the feature parameters of a certain time (the time to be predicted) in the performance network layer of the nth flight;

[0062] S3: Couple all the information extracted in steps S1 and S2 through the coupling network layer S21, and output the target value through the mapping network layer S31;

[0063] S4: Perform the second data coupling. Based on the method in step S1, randomly extract the feature parameters corresponding to x times in the (n-2)th and (n-1)th flights. The feature parameters corresponding to x times extracted in this step are not exactly the same as those corresponding to x times extracted in step S1, and are all randomly selected. Extract the feature parameters corresponding to the next time to be predicted in the performance network layer of the nth flight, and repeat step S3.

[0064] This explanation only uses two data coupling operations as an example. Subsequent data coupling operations will follow the same steps. By sequentially taking data up to time n, the data states of all times in both the (n-2)th and (n-1)th sorties can be traversed. During training, this operation involves multiple iterative calculations, allowing the overall data from previous sorties to be traversed many times. This means that through multiple random local memory operations, the overall situation is learned and memorized, thereby improving the accuracy of the model's predictions.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, substitutions, or variations made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A digital twin device for an aircraft engine, comprising a digital twin model of the aircraft engine, characterized in that: It includes a data input layer, a coupling network layer, and a mapping network layer. The coupling network layer extracts and couples the data from the data input layer, and then outputs the target value through the mapping network layer. The data input layer includes a current sortie performance network layer and a sortie impact physical characteristic extraction network layer. The current sortie performance network layer is used to extract physical information from the current sortie, and the sortie impact physical characteristic extraction network layer is used to extract physical information from the k sorties preceding the current sortie, where k is an integer not less than 1. The current sortie performance network layer is the nth sortie performance network layer. The physical information extracted by the nth sortie performance network layer is the performance state parameter corresponding to any moment in the nth sortie, and the extracted performance state parameter corresponding to any moment is transmitted to the coupling network layer. The sortie impact physical characteristic extraction network layer is composed of the (n-1-k)th sortie performance network layer to the (n-1)th sortie performance network layer. The sortie impact physical characteristic extraction network layer extracts the performance state parameters corresponding to partial or all moments in each sortie from the (n-1-k)th sortie to the (n-1)th sortie, and transmits the extracted performance state parameters corresponding to partial or all moments to the coupling network layer. The extracted performance state parameters cover all flight states in the entire sortie.

2. The digital twin device for an aero-engine according to claim 1, characterized in that: When extracting the physical characteristics of a flight, the network layer extracts the performance state parameters corresponding to local moments within each flight from flight number (n-1-k) to flight number (n-1), and these parameters are randomly selected.

3. The digital twin device for an aero-engine according to claim 1, characterized in that: When extracting performance state parameters corresponding to local moments in the physical characteristics extraction network layer, the parameters are selected based on the proportion of flight state data during the flight process. The extracted time state data is consistent with the proportion of data for each state in the flight state data of the flight sortie. The flight states include takeoff, cruise flight, and descent.

4. The digital twin device for an aero-engine according to claim 3, characterized in that: The proportion of takeoff and landing data extracted by the physical characteristic extraction network layer based on the number of flights affects the physical characteristic extraction network layer is extracted based on a multiple of the takeoff and landing data throughout the entire flight process; the total amount of takeoff, cruise, and landing data extracted by the physical characteristic extraction network layer for a flight is 0.1%-1% of the total data volume for a flight; the time interval for collecting performance status parameters in the performance network layer of the nth flight is 0.01s to 1s.

5. The digital twin device for an aero-engine according to claim 4, characterized in that: From the (n-1-k)th sortie to the (n-1)th sortie, the total amount of data extracted from takeoff, cruise flight, and descent in each sortie gradually increases.

6. The digital twin device for an aero-engine according to claim 1, characterized in that: The physical information includes the physical operational characteristics and patterns of various components and subsystems during the operation of an aircraft.

7. The digital twin device for an aero-engine according to claim 6, characterized in that: The current flight performance network layer and the flight impact physical characteristic extraction network layer adopt any one of the following: LSTM time series network, GRU time series network, ARIMA time series network, WaveNet network, and attention mechanism network.

8. The digital twin device for an aero-engine according to claim 7, characterized in that: When the current sortie performance network layer and the sortie impact physical characteristic extraction network layer are time-series networks, the inputs are the characteristic parameters of each subsystem at the corresponding time. The characteristic parameters include the characteristic parameters of the inlet network, fan component network, compressor component network, combustion chamber component network, lubricating oil system network, turbine component network, mixing chamber component network, and tail nozzle component network.

9. The digital twin device for an aero-engine according to claim 1, characterized in that: The coupling network layer is coupled using a neural network method; the mapping relationship in the mapping network layer is mainly the mapping relationship of the fully connected network layer in the neural network; the target value is the performance parameter that needs to be monitored and predicted during actual flight, including engine exhaust temperature, thrust and fuel consumption rate.

Images