An aero-engine digitized test running method based on multi-model deep learning

Abstract

This invention belongs to the field of aero-engine performance prediction technology, specifically relating to a digital test method for aero-engines based on multi-model deep learning. The specific technical solution is as follows: Key parameter prediction sub-models are trained and obtained using environmental parameters and throttle lever angles to predict key operating state parameters of various components; a performance prediction model is constructed, and trained and obtained using the predicted key operating state parameters and environmental parameters to predict the engine's performance parameters. Thus, multiple models are used to complete the digital test process, solving the problem of low performance prediction accuracy caused by directly using environmental parameters and throttle lever parameters to predict performance parameters.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine performance prediction technology, specifically relating to a digital test method for aero-engines based on multi-model deep learning. Background Technology

[0002] Engine testing is essential during the research, development, manufacturing, and overhaul / maintenance of aero-engines. Its primary purpose is to verify performance and reliability. Extensive testing is required before an aero-engine leaves the factory. By utilizing massive amounts of historical test data, a digital twin model of the aero-engine can be learned and used for subsequent digital testing. This model can obtain the engine's expected performance and component operating status in real time during actual physical testing, enabling full-stage condition monitoring of the aero-engine during testing. Based on this, the performance and status parameters obtained through digital testing can evaluate the performance of the engine during actual physical testing and guide the engine troubleshooting process. This will significantly improve the level of aero-engine testing and verification.

[0003] The core of conducting digital testing of aero-engines is to establish an accurate digital twin model of engine performance. Typically, the established digital twin model of aero-engine performance is achieved by directly training a neural network using performance-related input parameters to learn the mapping relationship from input data to engine performance parameters.

[0004] Current methods for predicting aero-engine performance utilize neural network models trained with sensor monitoring data related to performance parameters. In 2022, He Yongbo et al. proposed a hybrid model based on a feature- and temporal-series-based two-sided attention mechanism and a convolutional neural network-gated recurrent unit, selecting five feature parameters highly correlated with exhaust temperature: high and low pressure rotor speeds, fuel flow rates, atmospheric temperature, and atmospheric pressure, to predict exhaust temperature. In 2020, Kong Chenyi et al. extracted exhaust temperature-related features from flight data, including rotor guide vane angle, rotor speed, fuel flow rate, flight altitude, Mach number, and nozzle diameter, establishing an aero-engine exhaust temperature prediction model combining an autoencoder and a temporal convolutional neural network. Before physical engine testing, known performance-related parameters only include environmental parameters and throttle lever angle. However, the throttle lever angle is not directly related to performance parameters; rather, it adjusts the fuel supply by regulating the fuel flow valve, thereby regulating the engine's operating state, and does not contain information about the operating states of specific engine components. Therefore, predicting engine performance parameters solely based on environmental parameters and throttle lever angle has the problem of low prediction accuracy and cannot be used for digital testing. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a digital test method for aero-engines based on multi-model deep learning. First, a sub-model for predicting the state parameters of key components is constructed, using environmental parameters and throttle lever angle to predict the operating state parameters of key components. Then, an engine performance prediction model is established, selecting the operating state parameters of key components obtained from the sub-model and existing environmental parameters as model input parameters to predict engine performance parameters. Finally, the results obtained from multiple models are used to conduct digital test of the aero-engine.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a digital test method for aero-engines based on multi-model deep learning, which uses environmental parameters and throttle lever angle to train and obtain key parameter prediction sub-models to predict key operating state parameters of each component; and constructs a performance prediction model, which uses the predicted key operating state parameters and environmental parameters to train and obtain a performance prediction model to predict the engine's performance parameters.

[0007] The performance prediction model construction process is as follows: clustering the input feature parameters and embedding them to obtain the component feature matrix; using a multi-head self-attention mechanism to fuse the component features; automatically learning the correlation between the component features; and then performing feature learning to predict the performance parameters.

[0008] Preferably, the steps include:

[0009] S1. Construct key parameter prediction sub-model datasets, wherein the sub-model datasets are named D1, D2, ..., D... m D i ={x i ,y i}, i = 1, 2, ..., m, where m represents the number of key parameters of the component, Di is the dataset of the prediction sub-model for the key parameters of the i-th component, and x i The input features of the prediction sub-model for the key parameters of the i-th component are represented, including engine inlet temperature, engine inlet pressure, and throttle lever angle; y i This represents the output feature of the parameter prediction sub-model for the i-th component, which is the actual value of the key working state parameter of the i-th component;

[0010] S2. Use a multilayer perceptron network to build a key parameter prediction sub-model, and use the sub-model dataset from step S1 to train the key parameter prediction sub-model.

[0011] S3. Use the key parameter prediction sub-model to predict the key operating state parameters of each component and construct a performance prediction dataset; the predicted key operating state parameters and environmental parameters are the input features of the performance prediction model, the environmental parameters include inlet temperature and inlet pressure; the performance parameters are the output features of the performance prediction model, including thrust, fuel consumption rate and exhaust temperature.

[0012] S4. A performance prediction model is built using a multi-head self-attention mechanism, and the performance prediction model is trained using the performance prediction dataset in step S3.

[0013] S5. Use key parameter prediction sub-models and performance prediction models to conduct digital engine testing.

[0014] Preferably, step S4 includes the following steps:

[0015] S41. Cluster the components according to the input feature parameters to obtain the feature parameters of each component;

[0016] S42. Embed the feature parameters of each component obtained in step S41: Perform feature mapping on the feature parameters of each component using a fully connected network to obtain the component feature matrix Xc:

[0017]

[0018] Where C i Let represent the feature vector corresponding to the i-th component, i = 1, 2, ..., m`, where m` represents the number of components; the vector dimension is the number of nodes in the output layer of the fully connected network, d.

[0019] S43. Use a multi-head self-attention module to fuse component spatial features;

[0020] S44. Perform feature learning and performance parameter mapping on the component features after spatial feature fusion: Use a bidirectional long short-term memory neural network to extract features, and then use a fully connected network to perform feature mapping and output performance parameters.

[0021] Preferably, step S43 includes the following steps:

[0022] S431. Three fully connected layers are used to map the component feature matrix into Q, K, and V matrices, each with the shape [m`, d... k Its mathematical description is:

[0023] Q = XcW Q ;

[0024] K = XcW K ;

[0025] V = XcW V ;

[0026] Among them, W Q W K W V This is the weight matrix;

[0027] S432. Using scaled dot product attention to fuse component space features: Perform a dot product of Q with all K values, and use the softmax function to obtain the weights of V. The calculation formula is as follows:

[0028]

[0029] S433. Multiple scaled dot product attention heads are used in parallel to fuse particulate spatial features. The calculation formula is as follows:

[0030] MultiHead(Xc,Xc,Xc)=Concat(head1,head2,...,head h W o ;

[0031] head i =Attention(XcW i Q ,XcW i K ,XcW i V );

[0032] Where h represents the number of heads, W Q W K W V W O This is the weight matrix.

[0033] Preferably, in step S5, the digital engine test process involves: using environmental parameters and throttle lever angles recorded by sensors to input them into the key parameter prediction sub-model to obtain the real-time operating status parameters of engine components; then inputting the engine operating status parameters and environmental parameters into the performance prediction model to obtain the real-time performance parameters of the engine; and comparing them with the measured component operating status parameters and performance parameters of the engine to monitor the engine operating status and performance in real time, thus completing the digital test process.

[0034] Preferably, the engine components include an intake duct, a fan, a high-pressure compressor, a combustion chamber, a high-pressure turbine, a low-pressure turbine, and an exhaust nozzle; the key operating parameters of the intake duct include inlet temperature and inlet pressure; the key operating parameters of the fan include the adjustable guide vane angle at the fan inlet and low-pressure rotational speed; the key operating parameters of the high-pressure compressor include the adjustable guide vane angle at the high-pressure compressor inlet and high-pressure rotational speed; the key operating parameters of the combustion chamber include fuel flow rate; the key operating parameters of the high-pressure turbine include high-pressure rotational speed; the key operating parameters of the low-pressure turbine include low-pressure rotational speed; and the key operating parameters of the exhaust nozzle include minimum cross-sectional diameter.

[0035] Preferably, in step S2, multi-input single-output network models M1, M2, ..., M6 are established respectively; the multilayer perceptron network includes two hidden layers, each with 32 nodes.

[0036] Preferably, in step S41, the characteristic parameters of each component are arranged according to the order in which air flows through each component of the engine, namely, intake manifold, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and exhaust nozzle, to obtain the sorted component characteristic parameters.

[0037] Accordingly, an electronic device includes:

[0038] One or more processors;

[0039] Storage device for storing one or more programs;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement the digital test method for aero-engines based on multi-model deep learning.

[0041] Accordingly, a computer-readable medium stores a computer program that, when executed by a processor, implements the digital test method for aero-engines based on multi-model deep learning.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention first constructs a component state parameter prediction sub-model. Using a historical test dataset with satisfactory performance, the sub-model learns the engine's control laws, thereby obtaining the operating state parameters of key components. Then, an engine performance prediction model is constructed, using the key component state parameters obtained from the sub-model and existing environmental parameters to predict performance parameters. Thus, multiple models are used to complete the digital test process, solving the problem of low performance prediction accuracy caused by directly using environmental parameters and throttle parameters to predict performance parameters.

[0044] 2. The engine performance parameter prediction model proposed in this invention, which incorporates knowledge from the aero-engine field, obtains a component feature matrix by clustering input feature parameters and embedding them. Then, it utilizes a multi-head self-attention mechanism to fuse component features, automatically learn the correlations between component features, and finally performs feature learning to predict performance parameters. This model incorporates knowledge of aero-engine components, enabling the establishment of spatial correlations between features of different components, improving the accuracy of performance parameter prediction, and making the model interpretable. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the implementation of the digital test method for aero-engines of the present invention.

[0046] Figure 2 This is a diagram illustrating the overall architecture of the digital test method for aero-engines according to the present invention.

[0047] Figure 3 This is a diagram of the single-head attention structure of the attention module of the present invention;

[0048] Figure 4 This is a structural diagram of the engine performance parameter prediction model of the present invention;

[0049] Figure 5 This is a schematic diagram illustrating the scaling dot product attention calculation of the present invention;

[0050] Figure 6 This is a schematic diagram of the multi-head attention calculation of the present invention. Detailed Implementation

[0051] 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.

[0052] like Figure 1-6 As shown, this invention discloses a digital test method for aero-engines based on multi-model deep learning. Its core idea is: to train and obtain key parameter prediction sub-models using environmental parameters and throttle lever angles to predict key operating state parameters of each component; to construct a performance prediction model, and to train and obtain a performance prediction model using the predicted key operating state parameters and environmental parameters to predict the engine's performance parameters; the performance prediction model construction process involves: clustering input feature parameters and performing feature embedding to obtain a component feature matrix; fusing component features using a multi-head self-attention mechanism; automatically learning the correlation between component features; and then performing feature learning to predict performance parameters.

[0053] Before physical engine testing, the known input data only includes environmental parameters and the throttle lever angle controlling the engine's operating state. However, this data does not include the state parameters of key components affecting engine performance. Directly predicting performance parameters would result in low accuracy, failing to meet the accuracy requirements of digital testing. Multi-model deep learning can fully utilize data to train multiple models and finally predict target parameters. In this invention, a sub-model for predicting key component state parameters is first constructed. This sub-model is trained using a historical test dataset with acceptable performance to learn the engine's control laws and predict the operating state parameters of key components. Then, an engine performance prediction model is constructed, using the key component state parameters obtained from the sub-model and existing environmental parameters to predict engine performance parameters. Thus, multiple models are used to complete the digital engine testing process.

[0054] Furthermore, the specific methods and steps were disclosed, including the following:

[0055] S1. Construct the dataset for the prediction sub-model of key parameters.

[0056] The sub-model datasets are D1, D2, ..., D... m D i ={x i ,y i}, i = 1, 2, ..., m, where m represents the number of parts, D i For the dataset of the sub-model for predicting key parameters of the i-th component, x i y represents the input features of the prediction sub-model for the key parameters of the i-th component, including engine inlet temperature, engine inlet pressure, and throttle lever angle; i Let represent the output feature of the prediction sub-model for the key parameter of the i-th component, where is the actual value of the key parameter of the i-th component. The input feature x... i They are all the same, and the corresponding key component parameter y i different.

[0057] Specifically, the sub-model dataset is derived from test data of the same model engine that has passed performance tests. The key operating parameters include key operating parameters of the main engine components, such as: key operating parameters of the compressor component include adjustable guide vane angle and speed; key operating parameters of the turbine component include speed; key operating parameters of the combustion chamber component include fuel flow rate; and key operating parameters of the tailpipe component include minimum cross-sectional diameter.

[0058] S2. A key parameter prediction sub-model is built using a multilayer perceptron network, and the key parameter prediction sub-model is trained using the sub-model dataset from step S1.

[0059] Specifically, based on the sub-model dataset obtained in step S1, multi-input single-output network models M1, M2, ..., M are established respectively. m This model uses a multilayer perceptron network, and the model structure diagram is shown below. Figure 3 As shown, its input parameters are inlet temperature, inlet pressure, and throttle lever angle, and its output parameters are the key operating status parameters of each component. The Adam optimizer is used during training, with mean squared error loss selected as the loss function, its mathematical description as follows:

[0060]

[0061] Where N represents the number of samples, y i These represent the key operating parameters of actual engine components. This represents the predicted values ​​of key operating parameters for engine components.

[0062] S3. Use the key parameter prediction sub-model from step S2 to predict the key operating state parameters P1, P2, ..., P of each component. m A performance prediction dataset is constructed; the key operating state parameters and environmental parameters predicted are the input features of the performance prediction model, including inlet temperature T1 and inlet pressure P1; the performance parameters are the output features of the performance prediction model, including thrust, fuel consumption rate and exhaust temperature, etc.

[0063] S4. Build and train the performance prediction model. This invention provides an engine performance parameter prediction model that incorporates knowledge from the field of aero-engines. The model structure is as follows: Figure 4 As shown, the specific steps include the following:

[0064] S41. Cluster the components according to the input feature parameters to obtain the feature parameters of each component.

[0065] Specifically, the characteristic parameters clustered into a component can represent the component's operating state. For example, the characteristic parameters of the intake duct component are inlet temperature and inlet pressure, the characteristic parameters of the compressor component are inlet guide vane angle and rotational speed, and the characteristic parameter of the combustion chamber is fuel flow rate. Then, the component characteristic parameters are arranged according to the order in which air flows through each engine component, generally in the order of intake duct, compressor, combustion chamber, turbine, and exhaust nozzle, resulting in the sorted component characteristic parameters.

[0066] S42. Embed the feature parameters of each component obtained in step S41.

[0067] Specifically, the feature embedding method involves using a fully connected network to perform feature mapping on the feature parameters of each component. The number of input layer nodes in the fully connected network is equal to the number of feature parameters for the corresponding component, and the number of output nodes is d. This yields the component feature matrix Xc, mathematically described as follows:

[0068]

[0069] Where C i Let represent the feature vector corresponding to the i-th component, i = 1, 2, ..., m`, where m` represents the number of components; the vector dimension is the number of nodes d in the output layer of the fully connected network.

[0070] S43. Use a multi-head self-attention module to fuse component spatial features.

[0071] Specifically, the calculation steps for the multi-head self-attention module are as follows:

[0072] S431. Obtain the Querry (Q) matrix, Key (K) matrix, and Value (V) matrix. Three fully connected layers are used to map the component feature matrices into Q, K, and V matrices, each with the shape [m`, d... k Its mathematical description is:

[0073] Q = XcW Q ;

[0074] K = XcW K ;

[0075] V = XcW V ;

[0076] Among them, W Q W K W V This is the weight matrix;

[0077] S432. Scaled dot product attention is used to fuse part space features. Specifically, Q is multiplied by all K values, and the softmax function is used to obtain the weights of V. The calculation process is as follows: Figure 5 As shown, the calculation formula is as follows:

[0078]

[0079] S433, Parallel use of multiple scaled dot product attention heads to fuse component spatial features, such as Figure 6 As shown, the calculation formula is as follows:

[0080] MultiHead(Xc,Xc,Xc)=Concat(head1,head2,...,head h W o ;

[0081] head i =Attention(XcW i Q ,XcW i K ,XcW i V );

[0082] Where h represents the number of heads, W Q W K W V W O This is the weight matrix.

[0083] S44. Perform feature learning and performance parameter mapping on the component features after spatial feature fusion. First, a bidirectional long short-term memory neural network (LSTM) is used to extract features, and then a fully connected network is used for feature mapping to output performance parameters. Similar to the training method in step S2, the Adam optimizer is used to optimize the model weight parameters, and the mean squared error loss is selected as the loss function.

[0084] S5. Use key parameter prediction sub-models and performance prediction models to conduct digital engine testing.

[0085] Specifically, the key parameter prediction sub-model obtained in step S2 and the engine performance parameter prediction model obtained in step S4 are embedded into the engine physical test system. During actual testing, environmental parameters and throttle lever angles recorded in real time by sensors are input into the key parameter prediction sub-model to obtain the real-time operating state parameters of engine components. Then, the engine operating state parameters and environmental parameters are input into the performance prediction model to obtain the engine's real-time performance parameters. By comparing these parameters with the measured component operating state parameters and performance parameters of the engine, the engine's operating state and performance are monitored in real time, completing the digital test process.

[0086] This invention also discloses an electronic device, comprising: one or more processors; a storage device for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the aforementioned digital test method for aero-engines based on multi-model deep learning. The electronic device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers.

[0087] This invention also discloses a computer-readable medium storing a computer program that, when executed by a processor, implements the aforementioned digital test method for aero-engines based on multi-model deep learning. Embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing execution procedures. Figure 1 The program code for the method shown.

[0088] It should be noted that the computer-readable medium of this disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0089] In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0090] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0091] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0092] Example 1

[0093] The following describes in detail the implementation examples of the present invention using a twin-shaft turbofan engine as an example. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0094] Reference Figure 1 and Figure 2 The flowchart and overall architecture of the digital test method for aero-engines based on multi-model deep learning are as follows:

[0095] Step S1: Construct a key parameter prediction sub-model dataset. Specifically, the dataset is derived from test data of a qualified engine of the same model, and the key operating state parameters include key operating state parameters of the main engine components. In this embodiment, the main engine components include the intake manifold, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and exhaust nozzle. The key operating state parameters of the intake manifold components include the engine inlet temperature T1 and engine inlet pressure P1, which are known environmental parameters; the key operating state parameters of the fan components include the adjustable guide vane angle α1 and low-pressure speed n1; the key operating state parameters of the high-pressure compressor components include the adjustable guide vane angle α2 and high-pressure speed n2; the key operating state parameters of the combustion chamber components include the fuel flow rate W. f The key operating parameters for the high-pressure turbine components include the high-pressure speed n2; the key operating parameters for the low-pressure turbine components include the low-pressure speed n1; and the key operating parameters for the exhaust nozzle components include the minimum cross-sectional diameter D8. The resulting sub-model datasets are D1, D2, ..., D6, D... i ={x i ,y i}, i = 1, 2, ..., 6, D i The sub-model dataset for the key parameters of the i-th component is α1, n1, α2, n2, W. f D8, x i y represents the input features of the prediction sub-model for the key parameters of the i-th component, including engine inlet temperature, engine inlet pressure, and throttle lever angle; i This represents the output feature of the prediction sub-model for the key parameters of the i-th component, which is the actual value of the key working state parameter of the i-th component.

[0096] It should be noted that the high-pressure turbine and the high-pressure compressor are coaxial components, so their corresponding rotational speeds are both n2. Similarly, the low-pressure turbine and the fan are coaxial components, so their corresponding low-pressure rotational speeds are both n1.

[0097] Step S2: Build and train the key parameter prediction sub-model. Based on the sub-model dataset obtained in Step S1, build multi-input single-output network models M1, M2, ..., M6 respectively. This model uses a multilayer perceptron network, and the model structure diagram is shown below. Figure 3 As shown, its input parameters are inlet temperature, inlet pressure, and throttle lever angle, and its output parameters are the key operating status parameters of the component.

[0098] The multilayer perceptron used in this embodiment contains two hidden layers, each with 32 nodes. During training, the Adam optimizer is used, and the mean squared error loss is selected as the loss function, which is mathematically described as follows:

[0099]

[0100] Where N represents the number of samples, y i These represent the key operating parameters of actual engine components. This represents the predicted values ​​of key operating parameters for engine components.

[0101] Step S3: Construct a performance prediction dataset based on the predicted key operating state parameters and environmental parameters. Specifically, step S2 obtains the key operating state parameters P1, P2, ..., P6 of the component, and adds inlet temperature T1 and inlet pressure P1 as input parameters to the engine performance prediction model dataset. The model outputs performance parameters, including thrust, fuel consumption rate, exhaust temperature, etc.

[0102] Step S4: Build and train the performance prediction model. This invention provides an engine performance parameter prediction model that incorporates knowledge from the field of aero-engines. The model structure is as follows: Figure 4 As shown, the specific steps include:

[0103] S41: Cluster the components to which the input feature parameters belong to to obtain the feature parameters of each component. Specifically, the feature parameters clustered to a component can represent the working state of that component. In this embodiment, the feature parameters of the intake duct component are the inlet temperature T1 and the inlet pressure P1; the feature parameters of the fan component are the fan inlet guide vane angle α1 and the rotational speed n1; the feature parameters of the high-pressure compressor component are the high-pressure compressor inlet guide vane angle α2 and the rotational speed n2; and the feature parameter of the combustion chamber component is the fuel flow rate W. f The characteristic parameter of the high-pressure turbine component is the high-pressure speed n2; the characteristic parameter of the low-pressure turbine component is the low-pressure speed n1; and the characteristic parameter of the exhaust nozzle component is the minimum cross-sectional diameter D8. Next, the characteristic parameters of each component are arranged according to the order in which air flows through them: intake duct, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and exhaust nozzle, resulting in the ordered component characteristic parameters.

[0104] S42. Embed the feature parameters of each component obtained in step S41. Specifically, the feature embedding method involves using a fully connected network to map the feature parameters of each component. The number of input layer nodes in the fully connected network is equal to the number of feature parameters for the corresponding component, and the number of output nodes is 32. This yields the component feature matrix Xc, mathematically described as:

[0105]

[0106] Where C i Let i represent the feature vector corresponding to the i-th component, where i = 1, 2, ..., 7, and the vector dimension is 32.

[0107] S43. Use a multi-head self-attention module to fuse component spatial features.

[0108] Specifically, the calculation steps for the multi-head self-attention module are as follows:

[0109] S431. Obtain the Querry (Q) matrix, Key (K) matrix, and Value (V) matrix. Three fully connected layers are used to map the component feature matrices into Q, K, and V matrices, each with a shape of [7,32]. Their mathematical description is as follows:

[0110] Q = XcW Q ;

[0111] K = XcW K ;

[0112] V = XcW V ;

[0113] Among them, W Q W K W V This is the weight matrix;

[0114] S432. Scaled dot product attention is used to fuse part space features. Specifically, Q is multiplied by all K values, and the softmax function is used to obtain the weights of V. The calculation process is as follows: Figure 5 As shown, the calculation formula is as follows:

[0115]

[0116] S433, Parallel use of multiple scaled dot product attention heads to fuse component spatial features, such as Figure 6 As shown, the calculation formula is as follows:

[0117] MultiHead(Xc,Xc,Xc)=Concat(head1,head2,...,head h W o ;

[0118] head i =Attention(XcW i Q ,XcW i K ,XcW i V );

[0119] Where h represents the number of heads, W Q W K W V W O This is the weight matrix.

[0120] S44. The component features after spatial feature fusion are used for feature learning and performance parameter mapping. First, a bidirectional long short-term memory neural network (LSTM) is used to extract features, and then a fully connected network is used for feature mapping to output performance parameters. In this embodiment, the LSTM layer has 32 input nodes and 64 output nodes. Finally, the features output at the last time step are mapped to performance parameters through a fully connected layer.

[0121] Similar to the training method in step S2, the Adam optimizer is used to optimize the model weight parameters, and the mean squared error loss is selected as the loss function.

[0122] S5. Digital engine testing using multiple performance prediction models. Specifically, the key parameter prediction sub-model obtained in step S2 and the engine performance parameter prediction model obtained in step S4 are embedded into the engine physical testing system. During actual testing, environmental parameters and throttle parameters recorded in real time by sensors are input into the key parameter prediction sub-model to obtain the real-time operating status parameters of engine components. Then, the engine operating status parameters and environmental parameters are input into the performance prediction model to obtain the real-time performance parameters of the engine. By comparing these parameters with the measured component status parameters and performance parameters of the engine, the engine's operating status and performance are monitored in real time, completing the digital testing process.

[0123] 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 test method for aero-engines based on multi-model deep learning, characterized in that: By using environmental parameters and throttle lever angle, a key parameter prediction sub-model is trained and obtained to predict the key operating state parameters of each component; a performance prediction model is constructed, and the performance prediction model is trained and obtained using the predicted key operating state parameters and environmental parameters to predict the engine's performance parameters. The performance prediction model construction process is as follows: clustering the input feature parameters and embedding them to obtain the component feature matrix; using a multi-head self-attention mechanism to fuse the component features; automatically learning the correlation between the component features; and then performing feature learning and predicting the performance parameters. Includes the following steps: S1. Construct a key parameter prediction sub-model dataset, wherein the sub-model datasets are respectively... m represents the number of key parameters of the component. For the dataset of the sub-model for predicting key parameters of the i-th component, This represents the input features of the prediction sub-model for the key parameters of the i-th component, including engine inlet temperature, engine inlet pressure, and throttle lever angle. This represents the output feature of the parameter prediction sub-model for the i-th component, which is the actual value of the key working state parameter of the i-th component; S2. Use a multilayer perceptron network to build a key parameter prediction sub-model, and use the sub-model dataset from step S1 to train the key parameter prediction sub-model. S3. Use the key parameter prediction sub-model to predict the key operating state parameters of each component and construct a performance prediction dataset; the predicted key operating state parameters and environmental parameters are the input features of the performance prediction model, the environmental parameters include inlet temperature and inlet pressure; the performance parameters are the output features of the performance prediction model, including thrust, fuel consumption rate and exhaust temperature. S4. A performance prediction model is built using a multi-head self-attention mechanism, and the performance prediction model is trained using the performance prediction dataset in step S3. S5. Use key parameter prediction sub-models and performance prediction models to conduct digital engine testing. Step S4 includes the following steps: S41. Cluster the components according to the input feature parameters to obtain the feature parameters of each component; S42. Feature embedding of the feature parameters of each component obtained in step S41: Feature mapping is performed on the feature parameters of each component using a fully connected network to obtain the component feature matrix. : ； in This represents the feature vector corresponding to the i-th component. , This represents the number of components; the vector dimension is d, which represents the number of nodes in the output layer of the fully connected network. S43. Use a multi-head self-attention module to fuse component spatial features; S44. Perform feature learning and performance parameter mapping on the component features after spatial feature fusion: Use a bidirectional long short-term memory neural network to extract features, and then use a fully connected network to perform feature mapping and output performance parameters.

2. The digital test method for aero-engines based on multi-model deep learning according to claim 1, characterized in that: Step S43 includes the following steps: S431. Three fully connected layers are used to map the component feature matrix into Q, K, and V matrices, each with the shape of... Its mathematical description is: ； ； ； in, , , This is the weight matrix; S432. Using scaled dot product attention to fuse component space features: Perform a dot product of Q with all K values, and use the softmax function to obtain the weights of V. The calculation formula is as follows: ； S433. Multiple scaled dot product attention heads are used in parallel to fuse particulate spatial features. The calculation formula is as follows: ； ； Where h represents the number of heads, , , , This is the weight matrix.

3. The digital test method for aero-engines based on multi-model deep learning according to claim 1, characterized in that: The digital test process of the engine in step S5 is as follows: using the environmental parameters and throttle lever angle recorded by the sensors, the real-time working status parameters of the engine components are obtained by inputting them into the key parameter prediction sub-model. Then, the engine working status parameters and environmental parameters are input into the performance prediction model to obtain the real-time performance parameters of the engine. By comparing them with the measured working status parameters and performance parameters of the engine components, the engine working status and performance are monitored in real time, thus completing the digital test process.

4. The digital test method for aero-engines based on multi-model deep learning according to claim 3, characterized in that: The engine components include an intake manifold, a fan, a high-pressure compressor, a combustion chamber, a high-pressure turbine, a low-pressure turbine, and an exhaust nozzle. Key operating parameters of the intake manifold include inlet temperature and inlet pressure. Key operating parameters of the fan include the adjustable guide vane angle at the fan inlet and low-pressure rotational speed. Key operating parameters of the high-pressure compressor include the adjustable guide vane angle at the high-pressure compressor inlet and high-pressure rotational speed. Key operating parameters of the combustion chamber include fuel flow rate. Key operating parameters of the high-pressure turbine include high-pressure rotational speed. Key operating parameters of the low-pressure turbine include low-pressure rotational speed. Key operating parameters of the exhaust nozzle include minimum cross-sectional diameter.

5. The digital test method for aero-engines based on multi-model deep learning according to claim 4, characterized in that: In step S2, multi-input single-output network models are established respectively. The multilayer perceptron network consists of two hidden layers, each with 32 nodes.

6. The digital test method for aero-engines based on multi-model deep learning according to claim 5, characterized in that: In step S41, the characteristic parameters of each component are arranged according to the order in which air flows through each component of the engine, namely, intake manifold, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and exhaust nozzle, to obtain the sorted component characteristic parameters.

7. An electronic device, characterized in that: include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the digital test method for aero-engines based on multi-model deep learning as described in any one of claims 1-6.

8. A computer-readable medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the digital test method for aero-engines based on multi-model deep learning as described in any one of claims 1-6.

Images