Aero-engine performance digital twin model fused with assembly data and establishment method

Abstract

The application belongs to the technical field of aero-engine performance prediction, and particularly relates to an aero-engine performance digital twin model fusing assembly data and a building method. The specific technical scheme is as follows: collecting assembly test data of multiple engines of the same type in different test procedures; determining sensor monitoring data and assembly data related to target performance parameters; constructing a performance digital twin model fusing assembly data; and training the performance digital twin model by using the sensor monitoring data and the assembly data. The performance digital twin model considers the influence of assembly parameters on the overall performance of the engine, thereby improving the prediction accuracy of the performance prediction model. Meanwhile, the strategy of separately learning the time sequence sensor monitoring data and the assembly data can avoid the problem that the assembly feature information is covered in the training process due to the small number of engine sets in the training data and the too small number of assembly data sample combinations.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine performance prediction technology, specifically involving a digital twin model of aero-engine performance and its establishment method that integrates assembly data. Background Technology

[0002] Aero engines have complex structures, and the cumulative assembly errors between different components result in poor assembly precision. Simultaneously, the rotor components suffer from severe initial imbalance. Because multi-stage rotors in aero engines operate under high speed, high temperature, and high pressure for extended periods, substandard assembly precision can generate enormous forces and torques on the entire engine, causing vibration, accelerating mechanical fatigue of engine components, and even leading to malfunctions. Therefore, the assembly quality of the multi-stage rotor directly affects the overall performance of the aero engine, thus determining its reliability and lifespan. Before leaving the factory, engines require multiple test runs to verify whether their performance meets field service requirements. This typically involves multiple disassembly and reassembly processes to improve assembly precision and meet factory performance requirements.

[0003] Digital twins of aero-engine performance are typically based on deep learning and data-driven methods. They utilize physics knowledge from the engine field to build deep learning models, such as neural networks, which learn from vast amounts of historical data to capture complex nonlinear relationships and patterns, thereby enabling the modeling and prediction of engine performance. Then, using the trained performance digital twin model that meets accuracy requirements, the model directly outputs the corresponding performance parameters by inputting sensor monitoring parameters and assembly parameters for a given state. This reduces the number of subsequent test runs and lowers production costs.

[0004] However, current digital twin models for aero-engine performance primarily rely on monitoring data recorded by sensors deployed on-board or during testing to train deep models and predict engine performance parameters. For example, patent CN115688609A discloses an intelligent prediction and real-time early warning method for aero-engines. This method utilizes on-board engine measurement parameters to construct a real-time thrust prediction model and a baseline thrust prediction model, obtaining real-time thrust prediction values ​​and baseline thrust prediction values ​​under real-time flight conditions, respectively. This method does not consider the impact of assembly parameters on the overall engine performance. In actual production, the assembly precision of different batches of the same model varies, resulting in differences in overall performance during testing. Therefore, the model input parameters do not include assembly parameters, ignoring the impact of assembly on the overall engine performance, leading to reduced accuracy in engine performance prediction.

[0005] Furthermore, the typical approach to modeling performance digital twins involves inputting all relevant input feature parameters into a deep learning network model for learning. However, considering the limited number of engine test runs available for training performance digital twins, and the fact that the assembly parameters of a single engine remain constant during testing while sensor monitoring data dynamically changes with different operating conditions, inputting assembly data along with engine sensor monitoring data into the performance digital twin model would mask the impact of assembly information on performance parameters, hindering the network's ability to learn the correlation between assembly information and performance parameters. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention addresses the problem of low performance prediction accuracy in digital twin models of engine performance due to the lack of consideration for assembly parameters. It proposes a digital twin model for aero-engine performance that integrates assembly data and its establishment method. This model comprises two parts: one part learns engine state characteristics using sensor monitoring data, and the other part extracts assembly features using assembly data. Finally, the two parts of features are fused for further learning to obtain predicted performance parameters. Furthermore, by learning sensor monitoring data and assembly features separately during engine testing and then fusing them at the end, the invention avoids the problem of assembly feature information being masked during training due to a limited number of engines and consequently, a small number of assembly data sample combinations in the training data.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for establishing a digital twin model of aero-engine performance that integrates assembly data, comprising the following steps:

[0008] S1. Collect assembly and test data of multiple engines of the same model under different test procedures;

[0009] S2. Determine the sensor monitoring data and assembly data related to the target performance parameters;

[0010] S3. Construct a performance digital twin model that integrates assembly data;

[0011] S4. Train the performance digital twin model using the sensor monitoring data and assembly data from step S2;

[0012] S5. Calculate the prediction accuracy of the performance digital twin model after training. If the prediction accuracy meets the requirements, the process ends. If the prediction accuracy does not meet the requirements, proceed to step S4 again to retrain the performance digital twin model.

[0013] Preferably, step S3 includes:

[0014] S31. Use the sensor monitoring feature learning module to learn and extract engine status information from sensor monitoring data;

[0015] S32. Use the assembly feature learning module to learn and extract assembly feature information from the assembly data;

[0016] S33. Perform feature fusion between the engine status information extracted by the sensor monitoring feature learning module and the assembly feature information extracted by the assembly feature learning module during the current working period;

[0017] S34. A feature mapping layer is used to map the fused features to the target performance parameters.

[0018] Preferably, in step S31, the input of the sensor monitoring feature learning module is time-series sensor monitoring data, and the sampling method adopts a sliding window method. The shape of each sample input is [T, d], where T is the sliding window size and d is the number of sensor monitoring parameters related to the target performance parameters. Then, two recurrent network layers are used to learn the engine status information of the time-series sensor monitoring data. The recurrent network layers include RNN, GRU, LSTM or other recurrent neural networks.

[0019] Preferably, step S32 includes:

[0020] Step S321: Cluster the assembly parameters related to the target performance parameters according to the component and parameter attributes. The cluster categories include the axial clearance of each stage of compressor rotor and stator, the radial clearance of each stage of compressor rotor and stator, the axial clearance of each stage of turbine rotor and stator, and the radial clearance of each stage of turbine rotor and stator.

[0021] Step S322: Use a feature embedding layer to map the clustered assembly parameters into high-dimensional vectors of the same dimension;

[0022] Step S323: Input the component assembly feature vector after feature embedding into the attention module to obtain the weighted assembly feature vector. The attention module uses a multi-head attention mechanism to learn the weight coefficients of the component.

[0023] The single-head attention module is described as follows:

[0024]

[0025] v = HW - b

[0026]

[0027] oTH

[0028] Where v represents the component weight vector obtained after mapping through the fully connected layer, H represents the input assembly feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, b represents the bias; α represents the weight parameters after softmax normalization; x iLet v represent the assembly feature vector of the i-th component input unit, where N is the number of components. i represents the weights of each component learned through the fully connected layer; o represents the assembly feature vector obtained by weighted summation of the feature vectors of all components;

[0029] The above calculations are performed in parallel using m attention heads to obtain o1, o2, ..., o m Finally, the assembly feature vectors of m attention heads are merged to obtain the high-dimensional feature vector O' output by the assembly feature learning module, which is:

[0030] O' = concat(o1, o2, ..., o m ).

[0031] Preferably, in step S33, the feature fusion method is either additive fusion or splicing fusion.

[0032] Correspondingly, the aero-engine performance digital twin model that integrates assembly data includes a sensor monitoring feature learning module, an assembly feature learning module, and a feature fusion layer and a feature mapping layer connected in parallel. The input of the sensor monitoring feature learning module is time-series sensor monitoring data, and the input of the assembly feature learning module is assembly data. The outputs of the sensor monitoring feature learning module and the assembly feature learning module are jointly fed into the feature fusion layer. The output of the feature fusion layer is fed into the feature mapping layer, and the output of the feature mapping layer is the target performance parameters.

[0033] Preferably, the sensor monitoring feature learning module includes two recurrent network layers connected in sequence, wherein the recurrent network layer is one of RNN, GRU, LSTM or other recurrent neural networks.

[0034] Preferably, the assembly feature learning module is sequentially connected to a clustering module, a feature embedding layer, and an attention module. The clustering module clusters the assembly parameters related to the target performance parameters according to the component and parameter attributes. The feature embedding layer maps the clustered assembly parameters into high-dimensional vectors of the same dimension. The attention module analyzes the component assembly feature vectors after feature embedding to obtain a weighted assembly feature vector.

[0035] Preferably, the clustering categories of the clustering module include the axial clearance of the compressor stage rotor and stator, the radial clearance of the compressor stage rotor and stator, the axial clearance of the turbine stage rotor and stator, and the radial clearance of the turbine stage rotor and stator.

[0036] Preferably, the attention module employs a multi-head attention mechanism to learn the weight coefficients of the components; wherein the single-head attention module is described as follows:

[0037]

[0038] v = HW T +b

[0039]

[0040] =α T H

[0041] Where v represents the component weight vector obtained after mapping through the fully connected layer, H represents the input assembly feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, b represents the bias; α represents the weight parameters after softmax normalization; x i Let v represent the assembly feature vector of the i-th component input unit, where N is the number of components. i represents the weights of each component learned through the fully connected layer; o represents the assembly feature vector obtained by weighted summation of the feature vectors of all components;

[0042] The above calculations are performed in parallel using m attention heads to obtain o1, o2, ..., o m Finally, the assembly feature vectors of m attention heads are merged to obtain the high-dimensional feature vector O' output by the assembly feature learning module, which is:

[0043] O' = concat(o1, o2, ..., o m ).

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention proposes a digital twin model for aero-engine performance that integrates assembly data. The model comprises two modules: a sensor monitoring feature learning module that learns engine operating state characteristics using time-series sensor monitoring data, and an assembly feature learning module that extracts assembly features from assembly data. Finally, the features extracted by the two modules are fused for further learning to obtain predicted performance parameters. The proposed performance digital twin model considers the impact of assembly parameters on the overall engine performance, improving the prediction accuracy of the performance prediction model. Furthermore, the strategy of learning time-series sensor monitoring data and assembly data separately avoids the problem of assembly feature information being masked during training due to a small number of engines in the training data and consequently, a limited number of assembly data sample combinations. Additionally, an attention module is incorporated into the assembly feature learning module, which learns the importance of assembly feature information for each component, thereby obtaining weighted assembly feature information, which further enhances the model's prediction accuracy. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the implementation process of the digital twin model of aero-engine performance that integrates assembly data, as described in this invention.

[0047] Figure 2 This is a structural diagram of the digital twin model of aero-engine performance that integrates assembly data according to the present invention.

[0048] Figure 3 This is a diagram of the single-head attention structure of the attention module of the present invention;

[0049] Figure 4 This is a structural diagram of the digital twin model of aero-engine performance that integrates assembly data in Embodiment 1 of the present invention. Detailed Implementation

[0050] 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.

[0051] like Figure 1-3 As shown, this invention discloses a method for establishing a digital twin model of aero-engine performance that integrates assembly data, comprising the following steps:

[0052] S1. Collect assembly and test data of multiple engines of the same model under different test procedures, including sensor monitoring data and assembly data. Divide the data into training set and test set according to different engine batches. That is, the assembly and test data of the first m engines is used as the training set, and the assembly and test data of the remaining engines is used as the test set.

[0053] Specifically, the engine factory test run checks various conditions, including safety, flushing, lubricating oil assessment, power delivery, break-in, submission, and main engine functionality. During engine testing, multiple test runs are executed to verify performance under different operating conditions. Using a larger dataset for training allows the engine to learn the characteristics of each operating condition, resulting in higher model accuracy.

[0054] S2. Determine the sensor monitoring data and assembly data related to the target performance parameters. Target performance parameters include thrust, fuel consumption rate, exhaust temperature, etc.

[0055] Specifically, the sensor monitoring data parameters related to the target performance parameters include environmental parameters, control parameters, and status parameters. Among them, environmental parameters include engine inlet temperature and inlet pressure; control parameters include compressor inlet guide vane angle, throttle lever angle, and exhaust nozzle throat cross-sectional area; and status parameters include engine rotor speed and lubricating oil pressure.

[0056] The assembly data parameters related to the target performance parameters mainly include the radial and axial clearances of each stage of the rotor.

[0057] S3. Construct a performance digital twin model that integrates assembly data. The structure diagram of the aero-engine performance digital twin model integrating assembly data is shown below. Figure 2 As shown.

[0058] Specifically, step S3 includes:

[0059] S31. Use the sensor monitoring feature learning module to learn and extract engine status information from the sensor monitoring data.

[0060] Specifically, the input to the sensor monitoring feature learning module is time-series sensor monitoring data. The sampling method adopts a sliding pane approach, and the shape of each sample input is [T, d], where T is the sliding window size and d is the number of sensor monitoring parameters related to the target performance parameters, i.e., the input feature dimension.

[0061] Two recurrent network layers are then used to learn the engine status information contained in the time-series sensor monitoring data, wherein the recurrent network layers include, but are not limited to, other recurrent neural networks such as RNN, GRU, and LSTM.

[0062] S32. Utilize the assembly feature learning module to learn and extract assembly feature information from the assembly data. Specifically, step S32 includes:

[0063] Step S321: Cluster the assembly parameters related to the target performance parameters according to the components and parameter attributes, such as the axial clearance of the compressor stage rotor and stator, the radial clearance of the compressor stage rotor and stator, the axial clearance of the turbine stage rotor and stator, and the radial clearance of the turbine stage rotor and stator.

[0064] Step S322: Use a feature embedding layer to map the clustered assembly parameters into high-dimensional vectors of the same dimension. The feature embedding layer can use a fully connected network or a convolutional neural network. After clustering, the number of assembly parameters differs for different parts. For each part, a feature embedding layer is used to map the low-dimensional vector to a high-dimensional vector. It is necessary to ensure that the output dimension of the feature embedding layer for each part is the same to facilitate subsequent feature learning.

[0065] Step S323: Input the component assembly feature vector after feature embedding into the attention module to obtain the weighted assembly feature vector. Specifically, the attention module uses a multi-head attention mechanism to learn the weight coefficients of the components.

[0066] The structure diagram of the single-head attention module is as follows: Figure 3 As shown, the specific mathematical description of the single-head attention module is as follows:

[0067]

[0068] v = HW T +b

[0069]

[0070] =α T H

[0071] Where v represents the component weight vector obtained after mapping through the fully connected layer, H represents the input assembly feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, b represents the bias, and α represents the weight parameters after softmax normalization. i Let v represent the assembly feature vector of the i-th component input unit, where N is the number of components. i represents the weights of each component learned through the fully connected layer; o represents the assembly feature vector obtained by weighted summation of the feature vectors of all components.

[0072] The above calculations are performed in parallel using m attention heads to obtain o1, o2, ..., o m Finally, the assembly feature vectors of m attention heads are merged to obtain the high-dimensional feature vector O' output by the assembly feature learning module, which is mathematically described as follows:

[0073] O' = concat(o1, o2, ..., o m ).

[0074] S33. The engine operating status information extracted by the sensor monitoring feature learning module and the assembly feature information extracted by the assembly feature learning module are fused. Specifically, the feature fusion methods include additive fusion (element-level addition) and concatenation fusion.

[0075] S34. A feature mapping layer is used to map the fused features to the target performance parameters. Specifically, the feature mapping layer typically uses a fully connected network.

[0076] S4. Train the performance digital twin model using the training set (sensor monitoring data and assembly data) obtained in step S2. Specifically, the MSE loss function is used during training, and the network weights are updated using the Adam optimizer. The formula for calculating MSE is as follows:

[0077]

[0078] Where N represents the number of samples, y i This represents the actual engine performance parameters. This represents the predicted values ​​of engine performance parameters.

[0079] S5. Calculate the prediction accuracy of the performance digital twin model after training. If the prediction accuracy meets the requirements, the process ends. If the prediction accuracy does not meet the requirements, proceed to step S4 again to retrain the performance digital twin model.

[0080] Specifically, the test set is input into the trained digital twin model of engine performance to predict engine performance parameters, and the model's prediction accuracy is verified using an evaluation metric. The evaluation metric is the Mean Percentage Error (MAPE), calculated as follows:

[0081]

[0082] Where m represents the number of samples, y i This represents the actual engine performance parameters. This represents the predicted value of the performance parameter.

[0083] If the model accuracy calculated based on the evaluation indicators meets the requirements of the factory test run, the performance digital twin model can be used for subsequent digital test runs; if it does not meet the requirements, return to step S4, re-initialize the model and adjust the hyperparameters, and start training again.

[0084] like Figure 2 and Figure 3 As shown, this invention also discloses a digital twin model of aero-engine performance that integrates assembly data, including a sensor monitoring feature learning module, an assembly feature learning module, and a feature fusion layer and a feature mapping layer connected in sequence. The input of the sensor monitoring feature learning module is time-series sensor monitoring data, and the input of the assembly feature learning module is assembly data. The outputs of the sensor monitoring feature learning module and the assembly feature learning module are jointly fed into the feature fusion layer. The output of the feature fusion layer is fed into the feature mapping layer, and the output of the feature mapping layer is the target performance parameter.

[0085] Furthermore, the sensor monitoring feature learning module includes two recurrent network layers connected in sequence. These recurrent network layers are one of RNN, GRU, LSTM, or other recurrent neural networks. The structures of the two recurrent network layers can be the same or different.

[0086] Furthermore, the assembly feature learning module is sequentially connected to a clustering module, a feature embedding layer, and an attention module. The clustering module clusters the assembly parameters related to the target performance parameters according to the component and parameter attributes. The feature embedding layer maps the clustered assembly parameters into high-dimensional vectors of the same dimension. The attention module analyzes the component assembly feature vectors after feature embedding to obtain a weighted assembly feature vector.

[0087] Furthermore, the clustering categories of the clustering module include the axial clearance of the compressor stage rotor and stator, the radial clearance of the compressor stage rotor and stator, the axial clearance of the turbine stage rotor and stator, and the radial clearance of the turbine stage rotor and stator.

[0088] Furthermore, the attention module employs a multi-head attention mechanism to learn the weight coefficients of the components. The mathematical description of the single-head attention module is as follows:

[0089]

[0090] v = HW T +b

[0091]

[0092] =α T H

[0093] Where v represents the component weight vector obtained after mapping through the fully connected layer, H represents the input assembly feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, b represents the bias; α represents the weight parameters after softmax normalization; x i Let v represent the assembly feature vector of the i-th component input unit, where N is the number of components. i represents the weights of each component learned through the fully connected layer; o represents the assembly feature vector obtained by weighted summation of the feature vectors of all components.

[0094] The above calculations are performed in parallel using m attention heads to obtain o1, o2, ..., o m Finally, the assembly feature vectors of m attention heads are merged to obtain the high-dimensional feature vector O' output by the assembly feature learning module, which is:

[0095] O' = concat(o1, o2, ..., o m ).

[0096] Example 1

[0097] The following example uses a twin-shaft turbofan engine, such as Figure 4 As shown, embodiments of the present invention are described in detail. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0098] Step S1: Collect assembly and test data of multiple engines of the same model under different test procedures, including sensor monitoring data and assembly data. Divide the data into training set and test set according to different engine batches. That is, the assembly and test data of the first m engines is used as the training set, and the assembly and test data of the remaining engines is used as the test set.

[0099] Specifically, the engine factory test run checks various states, including safety, flushing, lubricating oil assessment, power delivery, break-in, submission, and main engine functionality. During engine testing, multiple test runs are executed to verify performance under different conditions. Using more datasets for training allows the engine to learn the characteristics of each state, resulting in higher model accuracy.

[0100] Step S2: Filter sensor monitoring data and assembly data related to the target performance parameters. Target performance parameters include thrust, fuel consumption rate, exhaust temperature, etc.

[0101] Specifically, the sensor monitoring data related to the target performance parameters include environmental parameters, control parameters, and state parameters. In this embodiment, environmental parameters include engine inlet temperature and inlet pressure; control parameters include fan inlet guide vane angle, high-pressure compressor inlet guide vane angle, throttle lever angle, and tailpipe critical cross-sectional area; and state parameters include low-pressure speed, high-pressure speed, and fuel flow rate.

[0102] In this embodiment, the assembly parameters related to the target performance parameters include the fan's various stages of rotation, stator axial clearance and radial clearance, and the high-pressure compressor's various stages of rotation, stator axial clearance and radial clearance.

[0103] Step S3: Construct a performance digital twin model that integrates assembly data. The structure diagram of the aero-engine performance digital twin model that integrates assembly data in this embodiment is as follows: Figure 4 As shown.

[0104] Specifically, step S3 includes:

[0105] S31: Use the sensor monitoring feature learning module to learn and extract engine operating status information contained in the monitoring data.

[0106] Specifically, the sensor monitoring feature learning module takes time-series sensor monitoring data as input, and its sampling method uses a sliding pane approach. Each sample input has a shape of [T, d]. Here, T is the sliding window size, and d is the number of sensor monitoring parameters related to the performance parameters, i.e., the feature dimension of the input. Here, T is set to 5-10, and d is set to 9; all sensor parameters have already been listed in the preceding section.

[0107] In this embodiment, a two-layer bidirectional long short-term memory neural network (Bi-LSTM) is used to learn the engine operating status information contained in the timing sensor monitoring data.

[0108] S32: Use the assembly feature learning module to learn and extract the assembly feature information contained in the assembly parameters.

[0109] Specifically, the selected assembly parameters related to the target performance parameters are first clustered according to the components and parameter attributes. In this embodiment, they are divided into fan rotation and stator axial clearance parameter input units, fan rotation and stator radial clearance parameter input units, high-pressure compressor stage rotation and stator axial clearance parameter input units, high-pressure compressor stage rotation and stator radial clearance parameter input units, high-pressure turbine stage rotation and stator axial clearance parameter input units, high-pressure turbine stage rotation and stator radial clearance parameter input units, low-pressure turbine stage rotation and stator axial clearance parameter input units, and low-pressure turbine stage rotation and stator radial clearance parameter input units.

[0110] Next, a feature embedding layer is used to map the clustered assembly parameters into high-dimensional vectors of the same dimension. In this embodiment, a fully connected network is used for the feature embedding layer.

[0111] Finally, the component assembly feature vector after feature embedding is input into the attention module to obtain the weighted assembly feature vector. Specifically, the attention module uses a multi-head attention mechanism to learn the weight coefficients of the components. The structure diagram of the single-head attention module is shown below. Figure 3 As shown. Where x i This represents the assembly feature vector of the i-th component input unit, where N is the number of components. In this embodiment, N = 8. i α represents the weights of each component learned through the fully connected layer. i This represents the component weight coefficients after normalization by the softmax layer, and o represents the assembly feature vector obtained by weighted summation of all component feature vectors. The specific mathematical description is as follows:

[0112]

[0113] v = HW T +b

[0114]

[0115] =α T H

[0116] Where H represents the input assembly feature matrix after feature embedding, v represents the component weight vector obtained after mapping through the fully connected layer, W represents the weight parameters of the fully connected layer, b represents the bias, α represents the weight parameters after softmax normalization, and o represents the weighted assembly feature vector.

[0117] In this embodiment, eight attention heads are used in parallel to perform the above calculations to obtain o1, o2, ..., o8 respectively. Finally, the assembly feature vectors of the eight heads are merged to obtain the high-dimensional feature vector O' output by the assembly feature learning module, which is mathematically described as follows:

[0118] O' = concat(o1,o2,…,o8).

[0119] S33: Perform feature fusion between the engine operating status information extracted by the sensor monitoring feature learning module and the assembly feature information extracted by the assembly feature learning module.

[0120] Specifically, feature fusion methods include additive fusion (element-level addition) and splicing fusion.

[0121] The mathematical description of additive fusion is:

[0122] t = t1 + t2

[0123] The mathematical description of splicing and fusion is as follows:

[0124] t = concat(t1, t2)

[0125] Where t1 represents engine operating status information, t2 represents assembly feature information, and t represents the fused features.

[0126] S34: Use a feature mapping layer to map the fused features to the target performance parameters.

[0127] Specifically, the feature mapping layer typically uses a fully connected network.

[0128] Step S4: Train the digital twin model of aero-engine performance using the training set obtained in steps S1 and S2. The MSE loss function is used during training, and the network weights are updated using the Adam optimizer. The formula for calculating MSE is as follows:

[0129]

[0130] Where N represents the number of samples, y i This represents the actual engine performance parameters. This represents the predicted values ​​of engine performance parameters.

[0131] Step S5: Input the test set into the trained digital twin model of engine performance to obtain the predicted engine performance parameters, and verify the model's prediction accuracy according to the evaluation index. The evaluation index uses the mean percentage error (MAPE), and its calculation formula is as follows:

[0132]

[0133] Where m represents the number of samples, y i This represents the actual engine performance parameters. This represents the predicted value of the performance parameter.

[0134] If the model accuracy calculated based on the evaluation metrics meets the factory test requirements, the performance digital twin model can be used for subsequent digital test runs; if it does not meet the requirements, return to step S4 to re-initialize the model and adjust the hyperparameters, and restart training. In this embodiment, the model accuracy MAPE requirement is no more than 3%.

[0135] 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 method for establishing a digital twin model of aero-engine performance that integrates assembly data, characterized in that: Includes the following steps: S1. Collect assembly and test data of multiple engines of the same model under different test procedures; S2. Determine the sensor monitoring data and assembly data related to the target performance parameters; S3. Construct a performance digital twin model that integrates assembly data; S4. Train the performance digital twin model using the sensor monitoring data and assembly data from step S2; S5. Calculate the prediction accuracy of the performance digital twin model after training. If the prediction accuracy meets the requirements, the process ends. If the prediction accuracy does not meet the requirements, proceed to step S4 again and retrain the performance digital twin model. The sensor monitoring data includes environmental parameters, control parameters, and status parameters; The assembly data includes the axial and radial clearances of each stage of the fan and the stator, as well as the axial and radial clearances of each stage of the high-pressure compressor. Step S3 includes: S31. Use the sensor monitoring feature learning module to learn and extract engine status information from sensor monitoring data; S32. Use the assembly feature learning module to learn and extract assembly feature information from the assembly data; S33. Perform feature fusion between the engine operating status information extracted by the sensor monitoring feature learning module and the assembly feature information extracted by the assembly feature learning module; S34. A feature mapping layer is used to map the fused features to the target performance parameters; In step S31, the input to the sensor monitoring feature learning module is time-series sensor monitoring data, and the sampling method uses a sliding pane approach, with each sample input having the following shape: , For the size of the sliding window, The number of sensor monitoring parameters related to the target performance parameters is input; then two recurrent network layers are used to learn the engine state information of the time-series sensor monitoring data. The recurrent network layers include RNN, GRU, and LSTM recurrent neural networks. Step S32 includes: Step S321: Cluster the assembly parameters related to the target performance parameters according to the component and parameter attributes. The cluster categories include the axial clearance of each stage of compressor rotor and stator, the radial clearance of each stage of compressor rotor and stator, the axial clearance of each stage of turbine rotor and stator, and the radial clearance of each stage of turbine rotor and stator. Step S322: Use a feature embedding layer to map the clustered assembly parameters into high-dimensional vectors of the same dimension; Step S323: Input the component assembly feature vector after feature embedding into the attention module to obtain the weighted assembly feature vector. The attention module uses a multi-head attention mechanism to learn the weight coefficients of the component.

2. The method for establishing a digital twin model of aero-engine performance according to claim 1, characterized in that: The single-head attention module is described as follows: ； in, This represents the component weight vector obtained after mapping through the fully connected layer. denoted as the input assembled feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, and b represents the bias; This represents the weight parameters after softmax normalization; This represents the assembly feature vector of the i-th component input unit, where N is the number of components. This represents the weights of each component learned through the fully connected layer; This represents the assembly feature vector obtained by weighted summation of the feature vectors of all components; The above calculations were performed in parallel using m attention heads to obtain the results. Finally, the assembly feature vectors from m attention heads are merged to obtain the high-dimensional feature vector output by the assembly feature learning module. ,for: 。 3. The method for establishing a digital twin model of aero-engine performance according to claim 1, characterized in that: In step S33, the feature fusion method is either additive fusion or splicing fusion.

4. A digital twin model of aero-engine performance integrating assembly data, established based on the method for establishing a digital twin model of aero-engine performance as described in claim 1, characterized in that: It includes a sensor monitoring feature learning module and an assembly feature learning module arranged in parallel, and a feature fusion layer and a feature mapping layer connected in sequence. The input of the sensor monitoring feature learning module is time-series sensor monitoring data, and the input of the assembly feature learning module is assembly data. The outputs of the sensor monitoring feature learning module and the assembly feature learning module are jointly fed into the feature fusion layer. The output of the feature fusion layer is fed into the feature mapping layer, and the output of the feature mapping layer is the target performance parameter.

5. The digital twin model of aero-engine performance according to claim 4, characterized in that: The sensor monitoring feature learning module includes two recurrent network layers connected in sequence. The recurrent network layer is one of RNN, GRU, or LSTM recurrent neural networks.

6. The digital twin model of aero-engine performance according to claim 4, characterized in that: The assembly feature learning module is sequentially connected to the clustering module, the feature embedding layer, and the attention module. The clustering module clusters the assembly parameters related to the target performance parameters according to the component and parameter attributes. The feature embedding layer maps the clustered assembly parameters into high-dimensional vectors of the same dimension. The attention module analyzes the component assembly feature vectors after feature embedding to obtain a weighted assembly feature vector.

7. The digital twin model of aero-engine performance according to claim 6, characterized in that: The clustering module includes clustering categories such as axial clearance of compressor stage rotors and stators, radial clearance of compressor stage rotors and stators, axial clearance of turbine stage rotors and stators, and radial clearance of turbine stage rotors and stators.

8. The digital twin model of aero-engine performance according to claim 6, characterized in that: The attention module employs a multi-head attention mechanism to learn the weight coefficients of the components; the single-head attention module is described as follows: ； in, This represents the component weight vector obtained after mapping through the fully connected layer. denoted as the input assembled feature matrix after feature embedding, W represents the weight parameters of the fully connected layer, and b represents the bias; This represents the weight parameters after softmax normalization; This represents the assembly feature vector of the i-th component input unit, where N is the number of components. This represents the weights of each component learned through the fully connected layer; This represents the assembly feature vector obtained by weighted summation of the feature vectors of all components; The above calculations were performed in parallel using m attention heads to obtain the results. Finally, the assembly feature vectors from m attention heads are merged to obtain the high-dimensional feature vector output by the assembly feature learning module. ,for: 。

Images