A digital modeling method for aircraft engine performance tracking

Abstract

The present invention belongs to the technical field of aircraft performance monitoring, and specifically relates to a digital modeling method for aircraft engine performance tracking. The specific technical solution is: it includes an n-1th sortie performance network layer, the input of the n-1th sortie performance network layer is the time performance status data of the n-1th sortie, and its output end is connected to a second coupling network layer; the input of the second coupling network layer also includes the migration performance learning characteristics of the n-2th sortie, and the output of the second coupling network layer is the second physical association feature of the n-1th sortie with the influence of the sortie. By introducing the influence of previous flights, the performance degradation effect caused by multiple flights during the actual use of the aircraft is taken into account, the prediction accuracy is improved, and the model is more in line with the actual physical laws.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft performance monitoring, and in particular relates to a digital modeling method for aircraft engine performance tracking. Background Art

[0002] The safe flight of aircraft (including civil aircraft, fighter jets, and helicopters) not only affects the lives of pilots and passengers on board, but also ensures the stability of combat missions. To ensure and improve the flight quality of aircraft, monitoring performance parameters and health status during flight is crucial. In addition, the development of aircraft health monitoring technology is equally important to the country's scientific and technological strategic development layout. Based on scientific research issues that have a significant impact on the national economy, health status monitoring of major projects will continue to be carried out as a key research direction. Therefore, the development of aircraft health status monitoring technology with higher safety and stability is essential for building an aviation power.

[0003] The core component of any aircraft's operation is its aircraft engine. The performance of an aircraft engine directly impacts the aircraft's 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 over the past 30 years, there have been at least 60 accidents caused by engine fires, shutdowns, blade failures, and other factors. Monitoring or predicting the engine's operating status before flight or before an accident occurs, assessing the health of the engine and the aircraft, and then taking appropriate measures to prevent the spread or occurrence of faults and eliminate potential flight safety hazards, is crucial for preventing and eliminating the potential for major safety accidents.

[0004] Over the past few decades, with the development of the aviation industry, aircraft engine performance condition monitoring technology has made significant progress, demonstrating strong innovation, reliability, and industrial application value. Currently, aircraft engine maintenance has shifted from preventive scheduled maintenance to a combination of condition-based maintenance based on engine status monitoring. The key to condition-based maintenance lies in effectively monitoring the health of engine components. Using relevant performance digital tracking / prediction technologies to monitor various parameters of each engine component in real time, comprehensive analysis of the monitoring data allows for reasonable conclusions about the operating status of each component and its development trends, providing a reliable basis for the aircraft's next steps. Aircraft health monitoring technology, particularly aircraft engine health monitoring technology, has significant practical and scientific significance for ensuring the safety and reliability of engines and even aircraft, ensuring flight safety, reducing aircraft downtime, improving aircraft utilization, and reducing maintenance investment.

[0005] Aircraft maintenance theory is broadly categorized into three approaches: corrective maintenance, preventive maintenance, and predictive maintenance. Corrective maintenance, also known as post-event maintenance, focuses solely on the integrity and usability of the equipment, disregarding its condition. Repairs are performed only after a partial or complete failure, restoring the equipment to its original usable condition. This approach, often referred to as repair after wear and tear, constitutes unplanned maintenance. Preventive maintenance, also known as scheduled maintenance, is time-based. Based on production plans and experience, the equipment is shut down for inspection, disassembly, and component replacement at prescribed intervals to prevent damage, subsequent failure, and production losses. This approach is currently widely used in planned or scheduled maintenance. Predictive maintenance first emerged in developed Western industrial countries. It uses systems engineering methods to comprehensively analyze the equipment's historical and current condition, reference the operating environment, and the performance of similar equipment. This approach identifies the equipment's internal operating conditions and physical properties, predicts its development trends, and proposes preventive measures and remedial measures.

[0006] Aircraft are equipment that operate at high altitudes. If maintenance is delayed until a malfunction occurs, it will pose a serious threat not only to the people on board but also to the lives and property of residents on the ground. In addition, by the time the aircraft exhibits obvious fault characteristics, its internal mechanical / structural damage has often reached a relatively serious state, so the success rate of repair is often low, resulting in unnecessary maintenance time and resource costs.

[0007] To fully ensure flight safety, preventive maintenance is currently the most common maintenance method used by most airports and airlines. Because it primarily relies on periodic equipment inspections by experienced ground maintenance personnel, it is often limited by equipment inspection methods and inspectors' experience. Furthermore, since fixed-cycle maintenance doesn't consider the actual condition of the equipment, maintenance plans can be inaccurate, potentially leading to excessive or insufficient maintenance. Finally, preventive maintenance carries high maintenance costs and results in extended downtime, resulting in excessive aircraft maintenance costs and potentially impacting flight plans and missions.

[0008] Predictive maintenance, a maintenance technology that has seen significant development in the aviation sector in recent years, involves periodic or continuous monitoring of equipment status, using machine learning algorithms and models to analyze and assess equipment health. This allows for predictions of the next failure and the timing of maintenance. Predictive maintenance is based on the condition of equipment / equipment, with condition monitoring and fault diagnosis as the foundation and condition prediction as the core. Maintenance decisions inform the final requirements for maintenance activities.

[0009] Manual observation-based monitoring and prediction methods are subjective and therefore limited and inaccurate. Statistical analysis methods, mathematical models, or artificial intelligence algorithms typically construct an aircraft performance model based on historical operational data, enabling monitoring and prediction of aircraft health. For example, the use of data in the traditional modeling process does not take into account the process information of the flight. Therefore, the performance analysis model of this method is a unified and averaged monitoring and prediction model, that is, the process status of all data is considered to be equal. However, in the actual flight process, the current flight sortie (flight sortie or also called flight flight, which refers to a flight mission, including a complete flight process of the aircraft from takeoff, steady-state flight to landing) has been in a state of decline after several flights. This decline is gradually generated with the accumulation of flight sorties, and the generation relationship is not a linear generation relationship. Since the influence of the flights passed down one by one is ignored, the results predicted by this traditional modeling method have a certain deviation from the current state to be predicted (for example, the prediction result for the nth flight is actually the result of n / 2 flights), resulting in low overall prediction accuracy. The starting point of the present invention is to solve this problem by considering the influence between flight sorties to construct a health status monitoring / prediction model with higher accuracy.

[0010] Therefore, if a digital modeling method for aircraft engine performance tracking that takes into account the impact between flights can be provided, it will have excellent application prospects. Summary of the Invention

[0011] To solve the above technical problems, the present invention proposes a digital modeling method for aircraft engine performance tracking, which introduces the influence of previous flights into the digital model of aircraft monitoring. It has higher model accuracy, is more in line with actual physical laws, and has stronger model interpretability.

[0012] To achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a feature transfer learning model includes an n-1th flight performance network layer, wherein the input of the n-1th flight performance network layer is the time performance status data of the n-1th flight, and the output end of the n-1th flight performance network layer is connected to a second coupling network layer; the input of the second coupling network layer also includes the transfer performance learning feature of the n-2th flight, and the output of the second coupling network layer is a second physical association feature of the n-1th flight with flight influence.

[0013] Preferably: the n-1th flight performance network layer includes multiple component sub-network layers and a timing mapping network layer, the output ends of the multiple component sub-network layers and the timing mapping network layer are connected to the first coupling network layer, and the output end of the first coupling network layer is connected to the second coupling network layer; the input of the multiple component sub-network layers is the characteristic parameters of each component at the corresponding time of the timing network, and the input of the timing mapping network layer is the timing characteristic parameters.

[0014] Preferably, the time series mapping network layer is a parameter related to the moment information, which can be one of LSTM, GRU, ARIMA, or one of wave network and attention mechanism network.

[0015] Preferably, the multiple component sub-network layers include an air inlet component network, a fan component network, a compressor component network, a combustion chamber component network, an oil system network, a turbine component network, and a tail nozzle component network.

[0016] Accordingly: A modeling method for a feature transfer learning model includes the following steps:

[0017] A1. Based on the flight data of flight 1, train the feature transfer learning model M for flight 1 and obtain the weight model M1 for flight 1.

[0018] A2. Substitute the flight data of the second flight into the weight model M1 trained in step A1 to obtain the physical correlation feature quantity of the first flight under the flight conditions of the second flight, which is regarded as the performance transfer feature quantity of the first flight;

[0019] A3. Substitute the performance transfer feature of the first flight and the flight data of the second flight into the feature transfer learning model M for training to obtain the weight model M2 for the second flight.

[0020] A4. Substitute the flight data of the third flight into the weight model M2 trained in step A3 to obtain the physical correlation feature quantity of the second flight under the flight conditions of the third flight, which is regarded as the performance transfer feature quantity of the second flight;

[0021] A5. Substitute the performance transfer feature of the j-1th flight and the flight data of the jth flight into the feature transfer learning model M for training to obtain the weight model M of the jth flight. j ; j=3, 4,..., n-1;

[0022] A6. Substitute the flight data of the j+1th flight into the weight model M trained in step A5. j , the physical correlation feature quantity of the j-th flight under the flight conditions of the j+1-th flight is obtained, which is regarded as the performance transfer feature quantity of the j-th flight.

[0023] Correspondingly: A performance tracking digital model for overall performance migration includes an n-th flight performance network layer, a second coupling network layer, and a mapping network layer connected in sequence, wherein the input of the n-th flight performance network layer is the time performance status data of the n-th flight, and its output end is connected to the second coupling network layer, and the output end of the second coupling network layer is connected to the mapping network layer; the input of the second coupling network layer also includes the migration performance learning characteristics of the (n-1)-th flight; the output of the mapping network layer is the final target output value of the model.

[0024] Preferably, the target output values ​​are performance parameters that need to be monitored and predicted during actual flight, including engine exhaust temperature, thrust, and fuel consumption rate.

[0025] Correspondingly: A performance tracking digital model that considers the migration of component performance characteristics includes a total performance network layer, a third coupling network layer and a mapping network layer connected in sequence, the total performance network layer includes the previous flight performance network layer and the current flight performance network layer, the output of the total performance network layer serves as the input of the third coupling network layer, the output of the third coupling network layer serves as the input of the mapping network layer, and the output of the mapping network layer is the target output value.

[0026] Preferably, the previous sortie performance network layer and the current sortie performance network layer each include four component sub-training network layers, namely, a compressor training network layer, a combustion chamber training network layer, a turbine training network layer and a nozzle training network layer, which are arranged in sequence.

[0027] Preferably, the input of the current compressor training network layer includes the transfer performance learning features of the previous compressor training network layer and the flight data of the current compressor training network layer; the output serves as the input of the current combustion chamber training network layer and the input of the next compressor training network layer;

[0028] The input of the current combustion chamber training network layer includes the transfer performance learning features of the previous combustion chamber training network layer, the flight data of the current combustion chamber training network layer, and the output of the current compressor training network layer; its output serves as the input of the current turbine training network layer and the input of the next combustion chamber training network layer;

[0029] The input of the current turbine training network layer includes the transfer performance learning features of the previous turbine training network layer, the flight data of the current turbine training network layer, and the output of the current combustion chamber training network layer; its output serves as the input of the current nozzle training network layer and the input of the next turbine training network layer;

[0030] The input of the current nozzle training network layer includes the transfer performance learning features of the previous nozzle training network layer, the flight data of the current nozzle training network layer, and the output of the current turbine training network layer; its output serves as the input of the current coupling network layer.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] By training and constructing a feature transfer learning model, as aircraft sorties gradually accumulate, the knowledge from the first n-1 sorties is gradually transferred and accumulated, resulting in the transferred performance learning features for the n-1 sortie. The transferred performance learning features from the n-1 sortie and the flight data from the nth sortie are used as inputs to a digital performance tracking model for predicting the performance tracking of the nth aircraft. By incorporating the influence of previous flights, the performance degradation caused by multiple sorties during actual aircraft use is taken into account. This improves prediction accuracy and achieves higher model precision based on existing health monitoring technology. Furthermore, the model is more consistent with actual physical laws and has stronger physical interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a structure for overall transfer learning of flight performance characteristics that takes into account the impact between flights;

[0034] Figure 2 It is a structural diagram of the performance network layer in the feature transfer learning model of the present invention;

[0035] Figure 3 This is a flow chart of the feature transfer learning model modeling process of the present invention;

[0036] Figure 4 This is a structural diagram of the feature transfer learning model of the present invention;

[0037] Figure 5 This is a structural diagram of the component performance feature transfer learning model of the present invention;

[0038] Figure 6 The present invention is a local transfer learning structure of flight performance characteristics that takes into account the influence between flights;

[0039] Figure 7 It is a performance tracking digital model used by the present invention to predict the performance of the current sortie. DETAILED DESCRIPTION

[0040] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0041] like Figure 1-6 As shown, the present invention discloses a digital performance tracking model that considers the influence between flights. Its core concept is to train and construct a feature transfer learning model. During the training process, as aircraft flights gradually accumulate, the knowledge of the first n-1 flights is gradually transferred and accumulated, resulting in the transferred performance learning features of the n-1 flight. The transferred performance learning features of the n-1 flight and the flight data of the nth flight are used as inputs to the digital performance tracking model for performance tracking prediction of the nth aircraft.

[0042] like Figure 4 As shown, the feature transfer learning model M includes an n-1th flight performance network layer, the input of the n-1th flight performance network layer is the time performance status data of the n-1th flight, and its output end is connected to a second coupling network layer, the input of the second coupling network layer also includes the transfer performance learning feature of the n-2th flight, and the output of the second coupling network layer is the second physical association feature of the n-1th flight with flight influence.

[0043] like Figure 2 As shown, the performance network layer for the n-1th flight includes multiple component sub-network layers and a timing mapping network layer. The outputs of the multiple component sub-network layers and the timing mapping network layer are connected to a first coupling network layer. The inputs of the multiple component sub-network layers are the characteristic parameters of the corresponding components, and the input of the timing mapping network layer is the timing characteristic parameters. The output of the first coupling network layer is a first physical correlation feature, which is connected to the second coupling network layer. Through a feature transfer learning model, the physical correlation features between the characteristic parameters F at each moment in the n-1th flight can be obtained.

[0044] Furthermore, the time series mapping network layer includes time information, and its input can be parameters related to the time information, such as time. The time series mapping network layer can be a time series network such as LSTM, GRU, and ARIMA. Alternatively, it can be a WaveNet and an attention mechanism network.

[0045] Furthermore, the multiple component sub-network layers include an air inlet component network, a fan component network, a compressor component network, a combustion chamber component network, a lubricating oil system network, a turbine component network, and a tail nozzle component network. The input of the multiple component sub-network layers is the characteristic parameters of each component at the corresponding time of the timing network. Specifically, the timing mapping network layer contains the longitudinal moment information of the data space of the n-1th flight, while the input of the multiple component sub-network layers is the transverse characteristic parameter F of the data space, which respectively represent the operating rules of each engine component at that moment. Then, through the first coupling network layer, the physical correlation features of the joint operation of each component at that moment can be extracted. The physical correlation features are the physical correlation features of the joint operation of the components. The mathematical form of the physical correlation features in the feature transfer learning model is a vector value, which is subsequently transferred and learned as a transfer feature to the network of other flights.

[0046] Specifically, the characteristic parameters corresponding to the inlet duct component network are inlet duct parameters, such as intake pressure and temperature. The characteristic parameters corresponding to the fan component network are fan parameters, such as fan low-pressure speed. The characteristic parameters corresponding to the compressor component network are compressor parameters, such as compressor guide vane angle, high-pressure speed, etc. The characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, such as fuel flow pressure, etc. The characteristic parameters corresponding to the lubricating oil system network are lubricating oil system parameters, such as lubricating oil pressure difference, lubricating oil flow, lubricating oil temperature, etc. The characteristic parameters corresponding to the turbine component network are turbine component parameters, such as turbine exhaust temperature and speed. The characteristic parameters corresponding to the tail nozzle component network are tail nozzle component parameters, such as tail nozzle diameter or tail nozzle distance, etc.

[0047] It should be noted that when n=1, the input of the feature transfer learning model only includes the time performance status data of the first flight, and its structure is as follows: Figure 2 shown.

[0048] Furthermore, in order to obtain the transfer performance learning features of n-1 flights, the transfer learning process of flight impact features is as follows:

[0049] A1. Based on the flight data of the first flight, train the feature transfer learning model M of the first flight and obtain the weight model M1 of the first flight.

[0050] A2. Substitute the flight data of the second flight into the weight model M1 trained in step A1 to obtain the physical correlation feature quantity of the first flight under the flight conditions of the second flight, which is regarded as the performance transfer feature quantity of the first flight.

[0051] A3. Substitute the performance transfer feature quantity of the first flight and the flight data of the second flight into the feature transfer learning model M for training to obtain the weight model M2 for the second flight.

[0052] A4. Substitute the flight data and transfer variables from the third flight into the weight model M2 trained in step A3 to obtain the physical correlation characteristics of the second flight under the flight conditions of the third flight. These are considered the performance transfer characteristics of the second flight. The transfer variables here are the prediction results of the first flight under the input conditions of the third flight, that is, the prediction results obtained by substituting the flight data of the third flight into the weight model M1. The performance transfer characteristics of the second flight include the performance characteristics of the second flight and also include the impact of the first flight on the second flight.

[0053] A5. Substitute the performance transfer feature of the j-1th flight and the flight data of the jth flight into the feature transfer learning model M for training to obtain the weight model M of the jth flight. j . j=3, 4,...,n-1.

[0054] A6: Substitute the flight data and migration variables of the j+1th flight into the weight model M trained in step A5. j , the physical correlation feature quantity of the jth flight under the flight conditions of the j+1th flight is obtained, which is regarded as the performance transfer feature quantity of the jth flight. j = 3, 4, ..., n-1. The transfer variable here is the prediction result of the j-1th flight under the input conditions of the j+1th flight.

[0055] like Figure 7 As shown, the present invention also discloses a performance tracking digital model for predicting the performance of the current flight, including an n-th flight performance network layer, a second coupling network layer, and a mapping network layer connected in sequence. The input of the n-th flight performance network layer is the time performance status data of the n-th flight, and its output end is connected to the second coupling network layer, and the output end of the second coupling network layer is connected to the mapping network layer. The input of the second coupling network layer also includes the migration performance learning features of the (n-1)-th flight. The output of the mapping network layer is the final target output value of the model. This model includes the above-mentioned performance tracking digital model that takes into account the influence between flights, except that a mapping network layer is added at the output end of the second coupling network layer. The other structures are the same.

[0056] It should be noted that the output of the current flight is the final target output value of the performance tracking digital model, that is, the performance prediction result of the current flight.

[0057] Furthermore, the mapping network layer maps the information from the coupled network layer to the final target output value. The mapping relationships in this layer are essentially the same as those in the fully connected network layer of the neural network. The physical meaning of this layer is to establish a relationship between the engine's physical operating rules and the target predicted values. These target output values ​​are the performance parameters that need to be monitored and predicted during actual flight, including engine exhaust temperature, thrust, and fuel consumption.

[0058] In particular, for the first flight, since it has just started flying, its input does not include transfer learning features.

[0059] Specifically, the method for building the digital model for tracking the performance of the nth flight is as follows:

[0060] B1: Initialize the model, build the performance network layer, train the performance tracking model for the first flight based on the flight data of the first flight, and obtain the transfer performance learning features of the first flight.

[0061] B2: Use the transferred performance learning features of flight 1 and the flight data of flight 2 as input to the performance network layer, train the performance tracking model for flight 2, and obtain the transferred performance learning features of flight 2. Here, the transferred performance learning features of flight 2 include the transferred performance learning features of flight 1.

[0062] B3: Use the transferred performance learning features of flight i-1 and the flight data of flight i as input to the overall performance network layer, train the performance tracking model for flight i, and obtain the transferred performance learning features of flight i. Here, the transferred performance learning features of flight i include the transferred performance learning features of flight i-1. Where i = 3, 4, 5, ..., n-2.

[0063] B4: Use the transferred performance learning features of flight n-2 and the flight data of flight n-1 as input to the overall performance network layer, train the performance tracking model for flight n-1, and obtain the transferred performance learning features of flight n-1. Here, the transferred performance learning features of flight n-1 include the transferred performance learning features of flight n-2.

[0064] B5: Use the transferred performance learning features of the n-1th flight and the flight data of the nth flight as the input of the overall performance network layer to train the performance tracking model of the nth flight for the performance tracking of the nth flight, and obtain the transferred performance learning features of the nth flight for the training of subsequent flight performance tracking models.

[0065] The aforementioned transfer learning model primarily considers the changing patterns of engine performance between flights, employing a holistic approach. However, due to varying operating environments and conditions, the performance degradation processes of different engine systems and components often vary.

[0066] like Figure 5 、 6As shown, to analyze the performance variations of engine systems and components over flight cycles, this paper proposes a component performance feature transfer learning model that considers inter-flight influences. This model employs a local transfer approach. By transferring the abstract features of the component network trained on the previous flight to the network trained on the current flight, along with the measured parameters of the components from the current flight, the influence of the previous flight on the current flight is determined. The core concept is to decompose the engine transfer learning content into the physical abstract features of each component / system and then transfer these physical abstract features to each component / system network layer of the current flight.

[0067] Specifically, the performance tracking digital model that considers the migration of component performance characteristics includes: a total performance network layer S1, a third coupling network layer S2, and a mapping network layer S3 connected in sequence. The total performance network layer S1 includes the previous flight performance network layer S11 and the current flight performance network layer S12. The output of the total performance network layer S1, i.e., the output of the previous flight performance network layer S11 and the output of the current flight performance network layer S12, serves as the input of the third coupling network layer S2. The output of the third coupling network layer S2 serves as the input of the mapping network layer S3. The output of the mapping network layer S3 is the final target output value of the model.

[0068] Furthermore, both the previous flight performance network layer S11 and the current flight performance network layer S12 include four component sub-training network layers: the compressor training network layer, the combustor training network layer, the turbine training network layer, and the nozzle training network layer. These component sub-training network layers are arranged according to the order in which the aircraft engine's main flow passes through the components, forming the compressor training network layer, the combustor training network layer, the turbine training network layer, and the nozzle training network layer, forming the backbone structure of the previous flight performance network layer and the current flight performance network layer.

[0069] Furthermore, the inputs of the previous flight performance network layer S11 and the current flight performance network layer S12 are explained using the current flight performance network layer S12 as an example: for the second-level and above component sub-training network layers, their inputs include the component flight data of the current flight, the output of the component sub-training network layer of the previous flight (i.e., the transfer performance learning features), and the output of the component sub-training network layer of the previous level of the current flight. For the first-level component sub-training network layer, its inputs include the component flight data of the current flight and the output of the component sub-training network layer of the previous flight (i.e., the transfer performance learning features). The output of the current flight component sub-training network layer serves as the input to the next-level component sub-training network layer and the corresponding component sub-training network layer of the next flight. The output of the last-level component sub-training network layer serves as the input to the coupling network layer S2.

[0070] Specifically, for the current flight, the input to the compressor training network layer for the current flight includes the transfer performance learning features from the previous flight's compressor training network layer and the flight data from the current flight's compressor training network layer. Its output serves as the input to the current flight's combustor training network layer and the next flight's compressor training network layer.

[0071] Furthermore, the input to the current combustor training network layer includes the transfer performance learning features of the previous combustor training network layer, the flight data of the current combustor training network layer, and the output of the current compressor training network layer. This output serves as the input to the current turbine training network layer and the next combustor training network layer.

[0072] Furthermore, the input to the current turbine training network layer includes the transfer performance learning features of the previous turbine training network layer, the flight data of the current turbine training network layer, and the output of the current combustor training network layer. This output serves as the input to the current nozzle training network layer and the next turbine training network layer.

[0073] Furthermore, the input of the current nozzle training network layer includes the transfer performance learning features of the previous nozzle training network layer, the flight data of the current nozzle training network layer, and the output of the current turbine training network layer. The output serves as the input of the current coupling network layer.

[0074] Furthermore, mapping network layer S3 maps the information from coupled network layer S2 to the final target output value. The mapping relationship in mapping network layer S3 is mainly the mapping relationship of the fully connected network layer in the neural network. The target output value is the performance parameter that needs to be monitored and predicted during actual flight, including engine exhaust temperature, thrust, and fuel consumption rate.

[0075] It can be understood that the migration idea of ​​the aforementioned model of the present invention is overall migration, and the migration idea of ​​the latter model is component-level physical feature migration. Overall migration and local migration each have their own advantages and application scenarios.

[0076] For overall migration, the structure is simple, it has better accuracy for simple problems (such as fewer component parameters), and its initial network layer training process is faster.

[0077] For local migration, the structure is more complex and is suitable for scenarios with more complex data (such as more component parameters). And because of its complex structure, it takes up more resources during the initial network training. The local migration model can also migrate the overall performance characteristics of the system, that is, migrate the performance characteristics of all components. When the overall system characteristics need to be migrated, more computing resources are required. Local feature migration has a better speed advantage in scenarios where only the characteristics of specific components are migrated. If it is known that only the compressor has decayed in the system, then only the network layer of the compressor needs to be feature transferred and learned, without relearning the entire system, thereby saving computing resources. The trained model is more interpretable and the determination of local faults is easier.

[0078] Example 1

[0079] Taking the transfer performance learning feature of flight 15 as an example, obtain the flight data of engine 0-16.

[0080] A1. Based on the flight data of the first flight, train the feature transfer learning model M of the first flight and obtain the weight model M1 of the first flight.

[0081] A2. Substitute the flight data of the second flight into the weight model M1 trained in step A1 to obtain the physical correlation feature quantity of the first flight under the flight conditions of the second flight, which is regarded as the performance transfer feature quantity of the first flight.

[0082] A3. Substitute the performance transfer feature quantity of the first flight and the flight data of the second flight into the feature transfer learning model M for training to obtain the weight model M2 for the second flight.

[0083] A4. Substitute the flight data of the third sortie into the trained weight model M2 to obtain the physical correlation feature quantity of the second sortie under the flight conditions of the third sortie, which is regarded as the performance transfer feature quantity of the second sortie.

[0084] A5. Substitute the performance transfer feature of the j-1th flight and the flight data of the jth flight into the feature transfer learning model M for training to obtain the weight model M of the jth flight. j . j=3,4,……,15。

[0085] A6. Substitute the flight data of the 16th flight into the trained weight model M 15 , the physical correlation characteristic quantity of the 15th sortie under the flight conditions of the 16th sortie is obtained, which is regarded as the performance migration characteristic quantity of the 15th sortie.

[0086] Example 2

[0087] Perform performance tracking and prediction for flight 16. Obtain flight data for engines from flight 0 to flight 16.

[0088] B1: Initialize the transfer learning model, build the performance network layer, train the performance tracking model for the first flight based on the flight data of the first flight, and obtain the transfer performance learning features of the first flight.

[0089] B2: Use the transfer performance learning features of the first flight and the flight data of the second flight as the input of the performance network layer of the second flight, train the performance tracking model of the second flight, and obtain the transfer performance learning features of the second flight. Here, the transfer performance learning features of the second flight include the transfer performance learning features of the first flight.

[0090] B3: Use the transferred performance learning features of the i-1th flight and the flight data of the i-th flight as the input to the i-th performance network layer, train the performance tracking model for the i-th flight, and obtain the transferred performance learning features of the i-th flight. Here, the transferred performance learning features of the i-1th flight include the transferred performance learning features of the i-1th flight. Where i = 3, 4, 5, ..., 14.

[0091] B4: Use the transfer performance learning features of the 14th flight and the flight data of the 15th flight as the input of the overall performance network layer, train the performance tracking model of the 15th flight, and obtain the transfer performance learning features of the 15th flight.

[0092] B5: The transferred performance learning features from flight 15 and the flight data from flight 16 are used as inputs to the performance network layer for flight 16 to train a performance tracking model for flight 16. At this point, the output of the third coupling network layer is fed into the mapping network layer, where the engine thrust is calculated using the mapping relationship. The transferred performance learning features from flight 16 are then derived and used to train performance tracking models for subsequent flights.

[0093] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various deformations, modifications, and substitutions made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A feature transfer learning device, comprising a feature transfer learning model, characterized in that: The feature transfer learning model includes an n-1th flight performance network layer, wherein the n-1th flight performance network layer receives as input the time performance state data of the n-1th flight, and an output end of the n-1th flight performance network layer is connected to a second coupling network layer; the input of the second coupling network layer also includes the transfer performance learning feature of the n-2th flight, and the output of the second coupling network layer is a second physical association feature of the n-1th flight with flight influence; The second physical association feature is the physical association feature of the joint operation between the components of the n-1th flight with flight impact; The n-1th flight performance network layer includes multiple component sub-network layers and a timing mapping network layer, wherein the output ends of the multiple component sub-network layers and the timing mapping network layer are connected to the first coupling network layer, and the output end of the first coupling network layer is connected to the second coupling network layer; the input of the multiple component sub-network layers is the characteristic parameters of each component at the corresponding time of the timing network, and the input of the timing mapping network layer is the timing characteristic parameters; The multiple component sub-network layers include an air inlet component network, a fan component network, a compressor component network, a combustion chamber component network, a lubricating oil system network, a turbine component network, and a tail nozzle component network. The characteristic parameters corresponding to the air inlet component network are air inlet parameters, the characteristic parameters corresponding to the fan component network are fan parameters, the characteristic parameters corresponding to the compressor component network are compressor parameters, the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, the characteristic parameters corresponding to the lubricating oil system network are lubricating oil system parameters, the characteristic parameters corresponding to the turbine component network are turbine component parameters, and the characteristic parameters corresponding to the tail nozzle component network are tail nozzle component parameters.

2. The feature transfer learning device according to claim 1, wherein: The time series mapping network layer is a parameter related to the moment information, which is one of LSTM, GRU, ARIMA, or one of wave network and attention mechanism network.

3. A modeling method for a feature transfer learning model, characterized by: The following steps are involved: A1. Based on the flight data of flight 1, a feature transfer learning model M for flight 1 is trained to obtain a weight model M1 for flight 1; the feature transfer learning model M is obtained by the feature transfer learning device according to any one of claims 1-2; A2. Substitute the flight data of the second flight into the weight model M1 trained in step A1 to obtain the physical correlation feature quantity of the first flight under the flight conditions of the second flight, which is regarded as the performance transfer feature quantity of the first flight; A3. Substitute the performance transfer feature of the first flight and the flight data of the second flight into the feature transfer learning model M for training to obtain the weight model M2 for the second flight. A4. Substitute the flight data of the third flight into the weight model M2 trained in step A3 to obtain the physical correlation feature quantity of the second flight under the flight conditions of the third flight, which is regarded as the performance transfer feature quantity of the second flight; A5. Substitute the performance transfer feature of the j-1th flight and the flight data of the jth flight into the feature transfer learning model M and train it to obtain the weight model M of the jth flight. j ; j = 3, 4, ..., n-1; A6. Substitute the flight data of the j+1th flight into the weight model M trained in step A5. j , the physical correlation feature quantity of the j-th flight under the flight conditions of the j+1-th flight is obtained, which is regarded as the performance transfer feature quantity of the j-th flight.

4. A digital performance tracking device for overall performance migration, including a digital performance tracking model for overall performance migration, characterized by: The performance tracking digital model for overall performance migration includes a sequentially connected n-th flight performance network layer, a second coupling network layer, and a mapping network layer. The input of the n-th flight performance network layer is the time performance status data of the n-th flight, and its output end is connected to the second coupling network layer. The output end of the second coupling network layer is connected to the mapping network layer. The input of the second coupling network layer also includes the migration performance learning feature of the (n-1)-th flight. The output of the mapping network layer is the final target output value of the model; The nth flight performance network layer includes multiple component sub-network layers and a timing mapping network layer, wherein the output ends of the multiple component sub-network layers and the timing mapping network layer are connected to a first coupling network layer, and the output end of the first coupling network layer is connected to the second coupling network layer; the input of the multiple component sub-network layers is the characteristic parameters of each component at the corresponding time of the timing network, and the input of the timing mapping network layer is the timing characteristic parameters; The multiple component sub-network layers include an air inlet component network, a fan component network, a compressor component network, a combustion chamber component network, a lubricating oil system network, a turbine component network, and a tail nozzle component network. The characteristic parameters corresponding to the air inlet component network are air inlet parameters, the characteristic parameters corresponding to the fan component network are fan parameters, the characteristic parameters corresponding to the compressor component network are compressor parameters, the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, the characteristic parameters corresponding to the lubricating oil system network are lubricating oil system parameters, the characteristic parameters corresponding to the turbine component network are turbine component parameters, and the characteristic parameters corresponding to the tail nozzle component network are tail nozzle component parameters. The target output values ​​are performance parameters that need to be monitored and predicted during actual flight, including engine exhaust temperature, thrust, and fuel consumption rate.

5. A digital performance tracking device that considers the migration of component performance characteristics, including a digital performance tracking model that considers the migration of component performance characteristics, characterized in that: A performance tracking digital model that considers component performance feature migration includes a total performance network layer, a third coupling network layer, and a mapping network layer connected in sequence. The total performance network layer includes a previous flight performance network layer and a current flight performance network layer. The output of the total performance network layer serves as the input of the third coupling network layer. The output of the third coupling network layer serves as the input of the mapping network layer. The output of the mapping network layer is the target output value. The sortie performance network layer includes four component sub-training network layers, namely the compressor training network layer, the combustion chamber training network layer, the turbine training network layer and the nozzle training network layer. The component sub-training network layers are arranged according to the order in which the main flow of the aircraft engine passes through the components, and are sequentially arranged as the compressor training network layer, the combustion chamber training network layer, the turbine training network layer and the nozzle training network layer. The previous flight performance network layer and the current flight performance network layer both include a compressor training network layer, a combustion chamber training network layer, a turbine training network layer, and a nozzle training network layer, which are arranged in sequence, totaling four component sub-training network layers; The target output values ​​are performance parameters that need to be monitored and predicted during actual flight, including engine exhaust temperature, thrust, and fuel consumption rate; The input of the current compressor training network layer includes the transfer performance learning features of the previous compressor training network layer and the flight data of the current compressor training network layer; Its output serves as the input of the combustion chamber training network layer of the current flight and the input of the compressor training network layer of the next flight; The input of the current combustion chamber training network layer includes the transfer performance learning features of the previous combustion chamber training network layer, the flight data of the current combustion chamber training network layer, and the output of the current compressor training network layer; its output serves as the input of the current turbine training network layer and the input of the next combustion chamber training network layer; The input of the current turbine training network layer includes the transfer performance learning features of the previous turbine training network layer, the flight data of the current turbine training network layer, and the output of the current combustion chamber training network layer; its output serves as the input of the current nozzle training network layer and the input of the next turbine training network layer; The input of the current nozzle training network layer includes the transfer performance learning features of the previous nozzle training network layer, the flight data of the current nozzle training network layer, and the output of the current turbine training network layer; its output serves as the input of the current coupling network layer.

Images