Aero-engine digital twin modeling method based on data mode conversion

Abstract

The invention discloses an aero-engine digital twin modeling method and device based on data mode conversion, a medium and equipment, and the method comprises the steps: constructing an input matrix and an output matrix, and converting each row vector in the input matrix into an N-dimensional polar coordinate radar map; and constructing an aero-engine performance prediction and fault diagnosis digital twinborn model, taking the N-dimensional polar coordinate radar map at each moment as an input feature, taking the output parameter as a supervision signal, and carrying out iterative training to enable the aero-engine performance prediction and fault diagnosis digital twinborn model to converge so as to obtain a final digital twinborn model. According to the method, the multi-source and multi-dimensional aero-engine time series data are converted into the image data, the image data are acquired, the acquired common monitoring parameters of the aero-engine are processed, and the performance and health state detection result of the aero-engine is obtained. And processing the imaging data based on an aero-engine performance prediction and fault diagnosis digital twin model to obtain a performance prediction and fault diagnosis result of the aero-engine.

Description

Technical Field

[0001] The present application relates to the field of digital twin modeling of aircraft engines, and in particular to a method, device, medium and equipment for digital twin modeling of aircraft engines based on data modal conversion. Background Art

[0002] As the power plant of modern aircraft, monitoring and predicting the performance and health of aircraft engines is crucial for aviation safety. With the rapid development of intelligent technologies, digital twin technology is increasingly being applied to routine aircraft engine maintenance, particularly in the areas of engine health management and remaining life prediction. Digital twin technology builds a virtual digital twin model of the engine, monitors its operating status in real time, and compares it with the actual engine itself, enabling prediction of engine health and maintenance decisions.

[0003] In the digital twin modeling process for aircraft engines, time series data is the primary data source. Typically collected by the engine's sensor system, it reflects engine operating parameters (such as speed, pressure ratio, and temperature) under different operating conditions. Time series data exhibits strong temporal and high-dimensional characteristics. Effectively processing and deeply mining this data are key to digital twin modeling. While statistical analysis or machine learning methods can, to a certain extent, uncover the potential information in time series data, there are still issues with insufficient information mining and the effective extraction of decay information. Summary of the Invention

[0004] The main purpose of this application is to provide an aero-engine digital twin modeling method, device, medium and equipment based on data modality conversion, aiming to solve the technical problems in the existing technology that the model lacks accuracy, reliability and generalization in tasks such as performance prediction, state prediction and fault diagnosis due to the single data modality and the inability to fully explore the deep features implicit in the data.

[0005] To achieve the above-mentioned purpose, the present application provides an aircraft engine digital twin modeling method based on data modal conversion, including: determining N common monitoring parameters of the aircraft engine, wherein the common monitoring parameters are selected from the intersection of parameters obtained by the aircraft engine in ground bench test, high-altitude bench test and airborne flight data scenarios; collecting N common monitoring parameters at T moments and constructing an input matrix, wherein the number of row vectors of the input matrix is T and the number of column vectors is N, synchronously collecting M-dimensional output parameters at T moments, and constructing an output matrix, and converting the normalized row vectors at each moment in the input matrix into an N-dimensional polar coordinate radar map; constructing an aircraft engine performance prediction and fault diagnosis digital twin model based on a self-attention mechanism, a convolutional neural network and a multi-layer perceptron, using the N-dimensional polar coordinate radar map at each moment as the input feature and the output parameter as the supervision signal, iteratively training to converge the aircraft engine performance prediction and fault diagnosis digital twin model to obtain the final digital twin model, and processing the obtained common monitoring parameters of the aircraft engine based on the aircraft engine performance prediction and fault diagnosis digital twin model to obtain the performance and health status detection results of the aircraft engine.

[0006] Optionally, the commonly monitored parameters include: air static temperature, atmospheric pressure, total air temperature, low-pressure guide vane angle, low-pressure speed, high-pressure guide vane angle, high-pressure speed, fuel flow, tail nozzle throat diameter, lubricating oil pressure and throttle lever angle.

[0007] Optionally, in the N-dimensional polar coordinate radar chart, each polar coordinate axis is arranged clockwise according to the order of the corresponding sensors before and after the engine.

[0008] Optionally, the digital twin model for aircraft engine performance prediction and fault diagnosis includes: a first detection module to a fourth detection module, and a first convolution module to a fourth convolution module; wherein the first detection module to the fourth detection module each include a multi-head self-attention mechanism layer and a multi-layer perceptron connected in series in sequence; the output of the multi-layer perceptron of the previous detection module is connected to the input of the multi-head self-attention mechanism layer of the subsequent detection module; the outputs of the first convolution module to the fourth convolution module are respectively connected to the inputs of the multi-head self-attention mechanism layers of the first detection module to the fourth detection module.

[0009] Optionally, the digital twin model for aircraft engine performance prediction and fault diagnosis is constructed based on the self-attention mechanism, convolutional neural network and multi-layer perceptron, with the N-dimensional polar coordinate radar map at each moment as the input feature and the output parameter as the supervision signal. Iterative training is performed to converge the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the final digital twin model, including: inputting the N-dimensional polar coordinate radar map at each moment into the first convolution module to the fourth convolution module respectively to obtain the first feature map to the fourth feature map; inputting the first feature map to the fourth feature map into the multi-head self-attention mechanism layer of the first detection module to the fourth detection module respectively, and outputting the aircraft engine performance and health status in the multi-layer perceptron of the fourth detection module; constructing a loss function based on the aircraft engine performance and health status and the input first feature map to the fourth feature map; iteratively training the digital twin model until the error of the loss function is lower than the first preset value, or until the number of iterations reaches the second preset value.

[0010] To achieve the above-mentioned purpose, the present application also provides an aero-engine digital twin modeling method based on data modal conversion, including: a parameter determination module for determining N common monitoring parameters of the aero-engine, wherein the common monitoring parameters are selected from the intersection of the parameters obtained by the aero-engine in ground bench test, high-altitude bench test and airborne flight data scenarios; a matrix construction module for collecting N common monitoring parameters at T moments and constructing an input matrix, wherein the number of row vectors of the input matrix is T and the number of column vectors is N, synchronously collecting M-dimensional output parameters at T moments, and constructing an output matrix, and converting each parameter in the input matrix into a matrix. The moment-normalized row vector is converted into an N-dimensional polar coordinate radar chart; the detection module is used to construct a digital twin model for aircraft engine performance prediction and fault diagnosis based on the self-attention mechanism, convolutional neural network and multi-layer perceptron, with the N-dimensional polar coordinate radar chart at each moment as the input feature and the output parameter as the supervision signal. Iterative training is used to converge the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the final digital twin model. The common monitoring parameters of the aircraft engine obtained are processed based on the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the performance and health status detection results of the aircraft engine.

[0011] To achieve the above-mentioned objectives, the present application also provides a computer-readable storage medium, which includes instructions. When the instructions are run on a computer, the computer executes the aerospace engine digital twin modeling method based on data modal conversion provided in the above-mentioned embodiment.

[0012] To achieve the above-mentioned objectives, the present application also provides an electronic device, which includes: at least one processor, a memory and an input and output unit; wherein the memory is used to store computer programs, and the processor is used to call the computer program stored in the memory to execute the digital twin modeling method of an aircraft engine based on data modal conversion provided in any of the aforementioned embodiments.

[0013] The embodiments of the present application propose a method, device, medium and equipment for digital twin modeling of an aircraft engine based on data modal conversion, which determines N common monitoring parameters of the aircraft engine, and the common monitoring parameters are selected from the intersection of parameters obtained by the aircraft engine in ground bench test, high-altitude bench test and airborne flight data scenarios; collects N common monitoring parameters at T moments and constructs an input matrix, wherein the number of row vectors of the input matrix is T and the number of column vectors is N, synchronously collects M-dimensional output parameters at T moments, constructs an output matrix, and converts the normalized row vectors at each moment into an N-dimensional polar coordinate radar map; constructs a digital twin model for performance prediction and fault diagnosis of an aircraft engine based on a self-attention mechanism, a convolutional neural network and a multi-layer perceptron, with the N-dimensional polar coordinate radar map at each moment as the input feature and the output parameter as the supervision signal , iterative training enables the digital twin model of aircraft engine performance prediction and fault diagnosis to converge and obtain the final digital twin model. Based on the digital twin model of aircraft engine performance prediction and fault diagnosis, the common monitoring parameters of the aircraft engine obtained are processed to obtain the performance and health status detection results of the aircraft engine. This application converts the multi-source and multi-dimensional aircraft engine time series data into graphical data, and then processes the graphical data based on the digital twin model of aircraft engine performance prediction and fault diagnosis to obtain the performance prediction, status prediction and fault diagnosis results of the aircraft engine. This application uses deep learning algorithms such as image processing to mine deep-level features in the time series that are difficult to be directly mined, including decay information, etc., which can improve the accuracy, reliability and generalization of the model, thereby better meeting engineering needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart illustrating an embodiment of an aero-engine digital twin modeling method based on data modality conversion provided by this application;

[0015] Figure 2 A radar chart of 11 engine monitoring parameters after modal conversion at a certain moment, provided in accordance with an embodiment of the present invention for an aircraft engine digital twin modeling method based on data modal conversion;

[0016] Figure 3 A schematic diagram of converting time series data into radar images through modal conversion, provided in accordance with an embodiment of the aero-engine digital twin modeling method based on data modal conversion of the present application;

[0017] Figure 4 A schematic diagram of the functional modules provided for an embodiment of an aero-engine digital twin modeling device based on data modal conversion in this application.

[0018] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0020] In recent years, deep learning technology in the field of computer vision has made breakthrough progress. Advanced network algorithms such as convolutional neural networks and self-attention mechanisms in image processing can effectively extract feature information from image data. Therefore, this application proposes an innovative technical solution to convert the modalities of multi-source, multi-dimensional aircraft engine time series data into image data. Then, based on image processing deep learning algorithms, a digital twin model for aircraft engine performance prediction and fault diagnosis is established to improve the model's capabilities in performance prediction, state prediction, and fault diagnosis.

[0021] Reference Figure 1 The first embodiment of the present application provides an aero-engine digital twin modeling method based on data modal conversion. The aero-engine thrust prediction digital twin modeling method based on data modal conversion may include:

[0022] S10. Determining N common monitoring parameters of the aircraft engine, where the common monitoring parameters are selected from the intersection of parameters obtained from the aircraft engine during ground test bench testing, high-altitude test bench testing, and airborne flight data scenarios;

[0023] Specifically, the engine monitoring parameter dimension is determined. Based on actual aircraft engine operating data, including ground test bench data, high-altitude test bench data, and airborne flight data, common monitoring parameters are found in these data types. The number N of common monitoring parameters is the target monitoring parameter dimension.

[0024] In one embodiment of the present application, the commonly monitored parameters include at least 11 of: air static temperature, atmospheric pressure, total air temperature, low-pressure guide vane angle, low-pressure rotational speed, high-pressure guide vane angle, high-pressure rotational speed, fuel flow, tail nozzle throat diameter, lubricating oil pressure and throttle lever angle.

[0025] S20, collecting N common monitoring parameters at T moments and constructing an input matrix, where the number of row vectors in the input matrix is T and the number of column vectors is N, synchronously collecting M-dimensional output parameters at T moments and constructing an output matrix, and converting the normalized row vectors at each moment in the input matrix into an N-dimensional polar coordinate radar chart;

[0026] refer to Figure 3 , the processor needs to draw the N-dimensional polar coordinate radar chart in a loop. The processor uses the normalized time series matrix M obtained in the above steps to generate the N-dimensional polar coordinate radar chart. norm , extract each row of data, which is a row vector of shape (1, N), and draw an N-dimensional polar coordinate radar chart. The coordinate dimension of the N-dimensional polar coordinate radar chart is the number of monitoring parameters N. Loop through T time points and draw N-dimensional polar coordinate radar charts in sequence to obtain T N-dimensional polar coordinate radar charts.

[0027] Collect time series data. For the N target monitoring parameters determined in step 1, collect the above N parameters at T moments to obtain a time series matrix M with a shape of (T, N) ori At the same time, according to the application purpose of the digital twin model for aircraft engine performance prediction and fault diagnosis, the output parameters of the model are determined, the time series data of the output parameters are collected, and the time series matrix M with the shape of (T, M) is obtained. output , where M represents the dimension of the target output parameter.

[0028] During the specific execution process, the processor uses various sensors to collect 2 hours of actual engine operation data. The sampling frequency of each sensor is 1HZ, so the sensor collects a total of 7200 data, and the time series matrix M with the shape of (7200,11) is obtained. ori The application purpose of the thrust prediction digital twin model is thrust prediction. Therefore, the sampled data at each moment is processed to obtain an output parameter. That is, the processor collects a total of column vectors M with a shape of (7200,1) for the thrust prediction digital twin model. output .

[0029] Before drawing the N-dimensional polar coordinate radar chart, the processor can first calculate the time series matrix M ori Normalize in the time feature dimension. The processor obtains the time series matrix M ori The data in the matrix are normalized to unify the values of all parameters to between 0 and 1, eliminating the influence of the dimensions of different parameters. In this way, the processor can obtain the normalized new time series matrix M. norm .

[0030] During normalization, the processor can obtain the minimum and maximum values of the 11 parameters at each moment, and then normalize the 11 parameters at each moment within the minimum and maximum value range, so that the value range of the 11 parameters is unified to between 0 and 1.

[0031] refer to Figure 2Specifically, in the N-dimensional polar coordinate radar map, each polar coordinate axis is arranged clockwise according to the order of the corresponding sensor before and after the engine.

[0032] S30. Construct a thrust prediction digital twin model based on the self-attention mechanism, convolutional neural network and multi-layer perceptron. Take the N-dimensional polar coordinate radar chart at each moment as the input feature and the corresponding output parameters of the output matrix as the supervision signal. Iterative training is performed to converge the thrust prediction digital twin model to obtain the final thrust prediction digital twin model. Based on the thrust prediction digital twin model, the common monitoring parameters of the aircraft engine are processed to obtain the performance and health status test results of the aircraft engine.

[0033] In one embodiment of the present application, the thrust prediction digital twin model includes:

[0034] First to fourth detection modules, and first to fourth convolution modules;

[0035] The first detection module to the fourth detection module each include a multi-head self-attention mechanism layer and a multi-layer perceptron connected in series;

[0036] The output of the multi-layer perceptron of the previous detection module is connected to the input of the multi-head self-attention mechanism layer of the subsequent detection module;

[0037] The outputs of the first to fourth convolution modules are respectively connected to the inputs of the multi-head self-attention mechanism layers of the first to fourth detection modules.

[0038] In one embodiment of the present application, the thrust prediction digital twin model is constructed based on the self-attention mechanism, convolutional neural network and multi-layer perceptron, with the N-dimensional polar coordinate radar chart at each moment as the input feature, the corresponding output parameter of the output matrix as the supervision signal, and iterative training to converge the thrust prediction digital twin model to obtain the final thrust prediction digital twin model, including:

[0039] Input the N-dimensional polar coordinate radar map at each moment into the first convolution module to the fourth convolution module respectively to obtain the first feature map to the fourth feature map;

[0040] Input the first to fourth feature maps into the multi-head self-attention mechanism layers of the first to fourth detection modules respectively, and output the thrust of the aircraft engine in the multi-layer perceptron of the fourth detection module;

[0041] Constructing a loss function based on the thrust of the aircraft engine and the first to fourth characteristic maps of the input;

[0042] The thrust prediction digital twin model is iteratively trained until the error of the loss function is lower than a first preset value, or until the number of iterations reaches a second preset value.

[0043] In the actual execution process, the processor obtains the output parameter time series matrix M in step S20. output , and T N-dimensional polar coordinate radar charts, to iteratively train the thrust prediction digital twin model. The thrust prediction digital twin model is constructed using a multi-layer perceptron, a convolutional neural network, and a self-attention mechanism.

[0044] Specifically, during iterative training of the thrust prediction digital twin model, the N-dimensional polar coordinate radar chart corresponding to each moment is cyclically input over T time periods. The thrust prediction digital twin model then maps the N-dimensional polar coordinate radar chart to the target output parameter at that moment. This cycle is repeated until the digital twin model reaches the required accuracy or the number of iterations reaches the maximum set number of training steps, at which point training is terminated to obtain the digital twin model.

[0045] For example, the processor cyclically inputs the corresponding N-dimensional polar coordinate radar chart at each moment in the thrust prediction digital twin model over 7200 time periods, mapping it to the thrust at that moment. The processor constructs a loss function based on the aircraft engine thrust and the first through fourth input feature maps. This cycle is repeated until the error of the digital twin model falls below 1e-4, or the number of iterations reaches 500, at which point training is terminated to obtain the digital twin model.

[0046] During execution, the processor uses various sensors to collect aircraft engine time series data, assuming it contains T moments, each containing N characteristic parameters, to generate a time series matrix with a shape of (T, N). Simultaneously, the thrust prediction digital twin model processes the time series matrix to generate an output matrix. The processor then uses a continuous sliding window sampling method (assuming the window length is w) to segment the data in the time and output matrices into multiple three-dimensional tensors encompassing the time dimension and all characteristic dimensions (N). The thrust prediction digital twin model is then iteratively trained using the three-dimensional tensors of the time series matrix and output matrix to generate a trained thrust prediction digital twin model.

[0047] Based on the above method embodiments, this application also proposes an aero-engine digital twin modeling method based on data modality conversion, including:

[0048] a parameter determination module, configured to determine N common monitoring parameters of the aircraft engine, wherein the common monitoring parameters are selected from the intersection of parameters obtained from the aircraft engine during ground test benches, high-altitude test benches, and airborne flight data scenarios;

[0049] The matrix construction module is used to collect N common monitoring parameters at T moments and construct an input matrix, where the number of row vectors in the input matrix is T and the number of column vectors is N. It also collects M-dimensional output parameters at T moments, constructs an output matrix, and converts the normalized row vectors at each moment into an N-dimensional polar coordinate radar chart.

[0050] The detection module is used to build a thrust prediction digital twin model based on the self-attention mechanism, convolutional neural network and multi-layer perceptron. It uses the N-dimensional polar coordinate radar chart at each moment as the input feature and the corresponding output parameters of the output matrix as the supervision signal. Iterative training is used to converge the thrust prediction digital twin model to obtain the final thrust prediction digital twin model. The common monitoring parameters of the aircraft engine obtained are processed based on the thrust prediction digital twin model to obtain the performance and health status detection results of the aircraft engine.

[0051] Based on the above method embodiments, the present application also proposes a computer-readable storage medium, which includes instructions. When the instructions are run on a computer, the computer executes the aerospace engine digital twin modeling method based on data modal conversion provided in the method embodiments.

[0052] Based on the above method embodiment, the present application further proposes an electronic device, comprising:

[0053] at least one processor, memory, and input-output unit;

[0054] In which, the memory is used to store computer programs, and the processor is used to call the computer programs stored in the memory to execute the aerospace engine digital twin modeling method based on data modal conversion provided in the method embodiment.

[0055] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A digital twin modeling method for aircraft engines based on data modal conversion, characterized in that: include: Determining N common monitoring parameters of the aircraft engine, where the common monitoring parameters are selected from the intersection of parameters obtained from the aircraft engine during ground test bench testing, high-altitude test bench testing, and airborne flight data scenarios; Collect N common monitoring parameters at T moments and construct an input matrix, where the number of row vectors in the input matrix is T and the number of column vectors is N. Simultaneously collect M-dimensional output parameters at T moments and construct an output matrix. The normalized row vectors at each moment in the input matrix are converted into an N-dimensional polar coordinate radar chart. A digital twin model for aircraft engine performance prediction and fault diagnosis is constructed based on the self-attention mechanism, convolutional neural network, and multi-layer perceptron. The N-dimensional polar coordinate radar chart at each moment is used as the input feature, and the corresponding output parameters of the output matrix are used as the supervision signal. Iterative training is used to converge the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the final digital twin model. The common monitoring parameters of the aircraft engine obtained are processed based on the digital twin model to obtain the performance and health status test results of the aircraft engine.

2. The aerospace engine digital twin modeling method based on data modal conversion according to claim 1, characterized in that: The common monitoring parameters include: Air static temperature, atmospheric pressure, total air temperature, low-pressure guide vane angle, low-pressure speed, high-pressure guide vane angle, high-pressure speed, fuel flow, tail nozzle throat diameter, oil pressure and throttle lever angle.

3. The aerospace engine digital twin modeling method based on data modal conversion according to claim 1, characterized in that: In the N-dimensional polar coordinate radar chart, each polar coordinate axis is arranged clockwise according to the order of the corresponding sensors before and after the engine.

4. The aerospace engine digital twin modeling method based on data modal conversion according to claim 1, characterized in that: The aero-engine performance prediction and fault diagnosis digital twin model includes: First to fourth detection modules, and first to fourth convolution modules; The first detection module to the fourth detection module each include a multi-head self-attention mechanism layer and a multi-layer perceptron connected in series; The output of the multi-layer perceptron of the previous detection module is connected to the input of the multi-head self-attention mechanism layer of the subsequent detection module; The outputs of the first to fourth convolution modules are respectively connected to the inputs of the multi-head self-attention mechanism layers of the first to fourth detection modules.

5. The aerospace engine digital twin modeling method based on data modal conversion according to claim 4 is characterized in that: The digital twin model for aircraft engine performance prediction and fault diagnosis is constructed based on the self-attention mechanism, convolutional neural network, and multi-layer perceptron. The N-dimensional polar coordinate radar chart at each moment is used as input features, and the corresponding output parameters of the output matrix are used as supervision signals. Iterative training is performed to converge the digital twin model for aircraft engine performance prediction and fault diagnosis, and the final digital twin model is obtained, including: Input the N-dimensional polar coordinate radar map at each moment into the first convolution module to the fourth convolution module respectively to obtain the first feature map to the fourth feature map; Input the first to fourth feature maps into the multi-head self-attention mechanism layers of the first to fourth detection modules, respectively, and output the aircraft engine performance and health status in the multi-layer perceptron of the fourth detection module; constructing a loss function based on the aircraft engine performance and health status and the first to fourth feature maps of the input; The digital twin model is iteratively trained until the error of the loss function is lower than a first preset value, or until the number of iterations reaches a second preset value.

6. A digital twin modeling method for an aero-engine based on data modal conversion, characterized in that: include: a parameter determination module, configured to determine N common monitoring parameters of the aircraft engine, wherein the common monitoring parameters are selected from the intersection of parameters obtained from the aircraft engine during ground test benches, high-altitude test benches, and airborne flight data scenarios; The matrix construction module is used to collect N common monitoring parameters at T moments and construct an input matrix, where the number of row vectors in the input matrix is T and the number of column vectors is N. It also collects M-dimensional output parameters at T moments and constructs an output matrix. It also converts the normalized row vectors of the input matrix at each moment into an N-dimensional polar coordinate radar chart. The detection module is used to build a digital twin model for aircraft engine performance prediction and fault diagnosis based on the self-attention mechanism, convolutional neural network and multi-layer perceptron. It uses the N-dimensional polar coordinate radar chart at each moment as the input feature and the corresponding output parameters of the output matrix as the supervision signal. Iterative training is used to converge the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the final digital twin model. The common monitoring parameters of the aircraft engine obtained are processed based on the digital twin model for aircraft engine performance prediction and fault diagnosis to obtain the performance and health status detection results of the aircraft engine.

7. A computer-readable storage medium, characterized in that It includes instructions, which, when run on a computer, enable the computer to execute the aero-engine digital twin modeling method based on data modal conversion as described in any one of claims 1 to 6.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor, memory, and input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the aerospace engine digital twin modeling method based on data modal conversion according to any one of claims 1 to 6.

Images