Rapid generation design method and system for redundant signals of aero-engine rotating speed sensor

Abstract

The invention discloses a rapid generation design method and system for redundant signals of an aero-engine rotating speed sensor, and belongs to the field of aero-engine modeling and simulation. The method comprises the following steps: simplifying a component-level model by introducing actual measurement sensor parameters, and establishing a model-based rotating speed generation module; the deep learning generation module is used for constructing a data set based on flight data and sensor parameters, and training a Transform architecture; dynamic switching logic is analyzed and designed through the error change rate, and signal smooth transition is achieved through weighted fusion. The system comprises a corresponding data acquisition module, a signal generation module and a fusion control module. According to the method, the advantages of model calculation and deep learning are fused, high-precision rotating speed sensor redundant signals can be quickly generated, and the fault-tolerant capability of an engine control system is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of aero-engine modeling and simulation, and particularly relates to a method and system for quickly generating redundant signals of aero-engine speed sensors. Background Art

[0002] As the core sensing unit of the aero-engine control system, the speed sensor undertakes the key functions of monitored control variables and feedback, and its measurement accuracy directly determines the stability of core control loops such as fuel regulation and surge suppression. In a typical control architecture, the speed parameter not only serves as the reference input for closed-loop regulation, but also is the basic data source for multi-dimensional decisions such as gas path parameter estimation and health state assessment. However, due to the extreme operating conditions of the engine, the speed sensor has a high probability of failure, seriously threatening the closed-loop reliability of the control system. Therefore, it is very necessary to realize the generation of speed sensor signals. Compared with the hardware redundancy technology, the analytical redundancy technology does not need to increase the number of hardware, so it does not need to increase the overall weight of the engine, and no new installation positions will be generated. Therefore, using the analytical redundancy technology to realize the generation of speed sensor signals is a more suitable method.

[0003] Traditional analytical redundancy technologies mainly use component-level models or neural networks to provide analytical redundancy to realize sensor signal generation. However, the gas path calculation of a complete component-level model is relatively complex and time-consuming, making it difficult to be implemented on board. The shallow neural network has the phenomenon of insufficient fitting ability when facing a system with a high degree of non-linearity. Based on this idea, the present invention separately designs a model-based speed fast generation sub-module and a Transformer architecture fast generation sub-module, and designs an effective fusion strategy, comprehensively considering the accuracy and time consumption of the signals provided by the two modules, and realizes the fast generation of aero-engine speed sensor signals based on signal fusion. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to aim at the defects of the background art, according to the model-based speed fast generation sub-module and the Transformer architecture fast generation sub-module, and design a fusion strategy to realize the generation of aero-engine speed sensor signals based on signal fusion. The present invention adopts the following technical solutions to solve the above technical problems:

[0005] A method for quickly generating redundant signals of aero-engine speed sensors includes the following steps:

[0006] (1) Simplify the complete component-level model by introducing measured sensor parameters, establish a model-based speed fast generation sub-module for the speed sensor, and realize the generation of speed sensor signals based on the model;

[0007] (2) Construct a dataset based on flight conditions and measured sensor data, train and establish a rotational speed rapid generation sub-module under the Transformer architecture to realize the generation of rotational speed sensor signals based on a deep learning network;

[0008] (3) Compare the output accuracy and time consumption of the two sub-modules, design a switching logic based on the error change rate within a sliding window, and achieve signal smooth fusion through a weighting coefficient.

[0009] Preferably, step (1) includes:

[0010] Introduce the measured values of the fan outlet temperature and pressure sensors to generate the low-pressure rotor speed;

[0011] Introduce the measured values of the total temperature and total pressure sensors at the inlet and outlet of the compressor and the measured value of the low-pressure turbine outlet pressure sensor, and combine the low-pressure rotor speed to generate the high-pressure rotor speed.

[0012] Preferably, in step (2):

[0013] The dataset construction uses the sliding window method to extract time series features;

[0014] The Transformer network includes an encoder, a fully connected layer, and a pooling layer, and processes the sensor time series data through the multi-head self-attention mechanism.

[0015] Preferably, the specific steps in step 3 are as follows:

[0016] Step 3.1: Quantitatively compare and evaluate the output signal accuracy and time consumption of the rotational speed rapid generation sub-module under the Transformer architecture and the output signal of the model-based rotational speed rapid generation sub-module. Select the rotational speed rapid generation sub-module under the Transformer architecture with superior comprehensive performance as the main operating unit, and switch to the model-based rotational speed rapid generation sub-module when the error shows an increasing trend;

[0017] Step 3.2: Statistically analyze the output signal error of the rotational speed rapid generation sub-module under the Transformer architecture within a time window, and calculate the percentage change rate of the error at adjacent moments. Determine whether to switch to the model-based rotational speed rapid generation sub-module based on the maximum change rate within the window;

[0018] Step 3.3: Achieve a smooth transition of the output signal from the rotational speed rapid generation sub-module under the Transformer architecture to the output signal of the model-based rotational speed rapid generation sub-module through dynamic weight calculation, specifically including:

[0019] Set the step size of the weighting coefficient change during the transition stage;

[0020] Adjust the output weights of the two models step by step during the switching process;

[0021] Output the final signal after weighted fusion.

[0022] The present invention also provides an aero-engine speed sensor signal generation system, comprising:

[0023] A data acquisition module for acquiring engine sensor data;

[0024] A model-based generation module for performing the generation calculation of step (1) described in claim 1;

[0025] A deep learning generation module for performing the generation calculation of step (2) described in claim 1;

[0026] A fusion control module for performing the fusion strategy of step (3) described in claim 1.

[0027] Preferably:

[0028] The model-based generation module includes an air path calculation unit and an iterative solution unit;

[0029] The deep learning generation module includes a data preprocessing unit and a Transformer encoding unit.

[0030] Preferably:

[0031] The fusion control module includes an error analysis unit and a switching control unit;

[0032] The system further includes an output module for outputting the final fused signal.

[0033] The present invention also provides an aero-engine control system, comprising:

[0034] The signal generation system described above;

[0035] A control unit for adjusting the engine operating parameters based on the generated speed signal.

[0036] The present invention also provides a computer-readable storage medium storing a computer program, which implements the steps of the method when executed by a processor.

[0037] The present invention also provides an aero-engine installed with the signal generation system provided by the present invention for providing a speed signal.

[0038] Beneficial effects: Model-based generation simplifies the gas path calculation model, reducing the number of initial guess values for the six balance equations required by the traditional component-level model to one; gas path calculation only needs to run four core components such as the intake duct and fan; the time consumption of model-based generation is reduced by more than 75% compared with the complete component-level model; through parallel computing, the single inference time of the Transformer architecture is controlled within 10 ms. Description of the Drawings

[0039] Figure 1 It is a schematic diagram of the fast generation design method and system for redundant signals of an aero-engine speed sensor;

[0040] Figure 2 It is a schematic diagram of gas path calculation for generating the low-pressure rotor speed;

[0041] Figure 3 It is a schematic diagram of gas path calculation for generating the high-pressure rotor speed;

[0042] Figure 4 It is a network structure diagram of the speed fast generation sub-module of the Transformer architecture;

[0043] Figure 5 It is a comparison chart of the accuracy and time consumption of two generation modules;

[0044] Figure 6 It is a dynamic simulation diagram of generating sensor signals;

[0045] Figure 7 It is a schematic diagram of the time consumption for generating sensor signals. Detailed Implementation Modes

[0046] The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

[0047] The detailed implementation modes of the present invention take a twin-spool turbofan engine as the research object, Figure 1 It is a flowchart of the design method of a fast generation model for an aero-engine speed sensor based on signal fusion. The design of this signal generation scheme includes the following steps:

[0048] Step 1), introducing measured sensor parameters to simplify the complete component-level model, establishing a corresponding model-based speed fast generation sub-module for the speed sensor signal, and realizing the generation of the speed sensor signal based on the model.

[0049] Step 2), constructing a data set according to flight conditions and measured sensor data, training and establishing a speed fast generation sub-module under the Transformer architecture, and realizing the generation of the speed sensor signal based on the deep learning network.

[0050] Step 3), analyze and compare the accuracy and time consumption of the two signal generation methods, design a switching logic based on the error change rate within the sliding window, and to ensure the smoothness of the switching process, design a weighted coefficient calculation strategy during the switching process to achieve the fusion of the generated signals.

[0051] The detailed steps of step 1) are as follows:

[0052] Step 1.1), introduce the measured values of the fan outlet temperature sensor and the fan outlet pressure sensor to generate the low-pressure rotor speed sensor signal NL.

[0053] Taking the dual-rotor mixed-exhaust turbofan engine as the research object, the definitions of each section are as shown in the following table

[0054] Table 1 Definitions of each section of the turbofan engine

[0055]

[0056] When generating the low-pressure rotor speed signal, since the fan and low-pressure turbine components are associated with the low-pressure shaft, the inlet and outlet temperature and pressure sensor signals of the two are considered for introduction. During the gas path calculation process, the air flow passes through components such as the inlet duct and the fan from front to back in sequence. If the measured parameters of the low-pressure turbine inlet and outlet temperature and pressure are introduced, the gas flow rate passing through the low-pressure turbine is unknown, and all components in front of its air flow path need to be considered, thus increasing the number of initial guess values of the balance equation and the complexity of the gas path calculation, and further leading to a decrease in the stability of the generation process and an increase in time consumption. However, there are no rotating components in front of the fan component, and the aerodynamic thermodynamics calculation is simple. Therefore, the healthy sensor signals selected for introduction are the fan outlet temperature sensor signal T 22,real and the fan outlet pressure sensor signal P 22,real to generate the low-pressure rotor speed sensor signal NL.

[0057] According to the flight altitude and Mach number, calculate the fan inlet temperature T2 and the fan inlet pressure P2 in the inlet duct model. After introducing the measured value P 22,real of the fan outlet pressure sensor, the fan component pressure ratio can be calculated.

[0058] (P2, T2) = f(H, Ma)

[0059] π fan = P 22,real / P2

[0060] Given the fan pressure ratio and the inlet and outlet temperature and pressure, run the fan component model, select the initial guess value as the low-pressure rotor speed NL to be generated, and the balance equation is the fan outlet temperature T 22,model calculated by the mainstream parameter model and the measured value T 22,real of the real engine fan outlet pressure sensor to be in balance, as shown in the following formula:

[0061] e = T 22,model / T 22,real -1

[0062] In the formula, e is the residual between the two. Iterative solution is carried out by the Newton - Raphson method. When the residual is less than the convergence accuracy, the low - pressure rotor speed NL at this time is output as the generated value of the low - pressure rotor speed sensor.

[0063] When generating the low - pressure rotor speed NL signal using the mainstream parameter model, the dimension of the initial guess value is reduced by 1. And compared with the traditional component - level model, only the inlet duct and the fan need to be run during the gas path calculation, and the complexity of the gas path calculation is greatly reduced. The gas path calculation process is as Figure 2 shown.

[0064] Step 1.2), introduce the measured values of the total temperature and total pressure sensors at the inlet and outlet sections of the compressor, the measured value P 6,real of the total pressure sensor at the outlet of the low - pressure turbine, and the measured value NL real of the low - pressure rotor speed sensor to generate the high - pressure rotor speed NH.

[0065] After introducing the pressure at the outlet of the low - pressure turbine, by calculating the bypass duct components, according to the static pressure conservation between the outlet of the bypass duct and the outlet of the low - pressure turbine, the gas flow rate through the bypass duct is calculated. Select to introduce the fan outlet temperature sensor, the fan outlet pressure sensor, the compressor outlet temperature sensor, the compressor outlet pressure sensor, the pressure at the outlet of the low - pressure turbine, and the low - pressure rotor speed to jointly generate the high - pressure rotor speed sensor signal NH.

[0066] In the inlet duct model, the fan inlet temperature T2 and the fan inlet pressure P2 are calculated according to the flight altitude and Mach number. After introducing the fan outlet pressure P 22,real and the compressor outlet pressure P 3,real , the pressure ratios of the fan and compressor components can be calculated. The calculation process is shown in the following formula:

[0067] (P2, T2) = f(H, Ma)

[0068]

[0069] Given the low - pressure rotor speed, the fan pressure ratio, and the inlet and outlet temperature and pressure, run the fan component model to obtain the gas flow rate through the fan. According to the pressure P 6,real at the outlet of the low - pressure turbine and the bypass duct components, calculate the flow rate through the bypass duct. From the flow continuity, the gas flow rate through the compressor can be obtained. For the compressor component, the gas flow rate through it, the pressure ratio, and the inlet and outlet section temperature and pressure are known. Select the initial guess value as the high - pressure rotor speed NH to be generated, and the balance equation is the compressor outlet temperature T 3,modelEqual to the measured value T of the real engine compressor outlet pressure sensor, as shown in the following formula: 3,real Equal, as shown in the following formula:

[0070] e = T 3,model / T 3,real -1

[0071] Iterative solution is carried out by the Newton - Raphson method. When the residual is less than the convergence accuracy, the high - pressure rotor speed NH at this time is output as the generated value of the high - pressure rotor speed sensor.

[0072] When generating the high - pressure rotor speed NH signal using the mainstream parameter model, the dimension of the initial guess value is reduced from 6 to 1. And compared with the traditional component - level model, during the gas path calculation, only four components, namely the inlet duct, fan, compressor, and bypass duct, need to be run. The complexity of the gas path calculation is greatly reduced. The process of generating the gas path calculation is as Figure 3 shown.

[0073] The detailed steps of step 2) are as follows:

[0074] Step 2.1) Traverse the fuel flow rate within the control plan at different envelope operating points, collect sensor signal data, perform data processing and data set division, and construct an analytical redundancy deep learning network model data set.

[0075] Traverse the fuel flow rate from low to high within the control plan at different envelope operating points, and determine other model inputs such as the adjustable area according to the control plan, thereby obtaining measurable sensor data corresponding to different fuel flow rates at different envelope operating points.

[0076] During the process of traversing the fuel flow rate, the fuel flow rate steps up. Each step amplitude is randomly selected within {2%, 3%, 4%}. After the step - up, the fuel remains unchanged. Wait for the model to stabilize and run for a period of time in the steady state, then the fuel flow rate steps up again. This cycle continues until the fuel flow rate reaches the maximum value in the control plan. The above fuel traversal scheme enables the collected data to contain both sensor data at different change rates of the fuel flow rate during the dynamic process and sensor data in the steady state at different fuel flow rates.

[0077] After the data collection is completed, since the data magnitudes and units of various types of sensors in aero - engines are different, if they are not normalized, the pressure sensors with larger values will have too much influence on the parameters during the network training process. Therefore, it is necessary to normalize the collected data. The min - max normalization method is used, and the calculation formula is as shown below:

[0078]

[0079] where, x ij′ represents the element in the i - th row and j - th column after the data matrix is normalized, represents the element at the \(i\)-th row and \(j\)-th column of the data matrix before normalization, represents the minimum value of all elements in the \(j\)-th column of the data matrix before normalization, represents the maximum value of all elements in the \(j\)-th column of the data matrix before normalization.

[0080] After collecting and normalizing the data, the dataset feature data and label data are divided according to different types of generated sensors. Since the measurement data of the health sensor contains a large amount of valid information, which can provide support for generating the measurement signals of other sensors, the constructed dataset should include the health sensor data. At the same time, due to the time correlation of the aero-engine sensor data, that is, the data at the current moment is affected by the data at the previous moment, the sliding window method is used to select the time series data of measurable sensors when constructing the dataset. After analysis, the dataset feature data is selected as all measurable sensor data except the generated sensor, flight altitude, Mach number, and fuel flow at the current moment and the previous 9 sampling moments; the dataset label data is the generated sensor data at the current moment.

[0081] Step 2.2) Design and analyze the redundant deep learning network model in combination with the Transformer network architecture to realize the generation of the rotational speed sensor signal based on the deep learning network.

[0082] The self-attention mechanism introduced by Transformer can effectively capture the long-range dependence relationships in the sequence information and allows the model to perform parallel computing, making the computing efficiency of the model higher and having unique advantages in processing long sequences. For the sensor signal generation task, the training feature data is fuel flow, afterburner fuel flow, all measurable sensors except the rotational speed sensor, and flight altitude and Mach number, while the training label data is the value of the rotational speed sensor at the current moment.

[0083] During the network training and inference process, only the feature extraction and learning of the input sensor sequence data are required, and there is no need to generate a new sequence. Therefore, only the encoder is needed to process the input and output data, and there is no need for the decoder to generate step by step. Therefore, taking the encoder part in the Transformer network architecture as the core, a deep learning network model for rotational speed sensor parsing redundancy estimation is designed, and its structure is as Figure 4 shown.

[0084] The multi-head self-attention layer is the core of Transformer, which is composed of multiple self-attention layers and plays the role of feature extraction. The multi-head self-attention layer is composed of the splicing linear transformation of multiple self-attention layers.

[0085] In the self-attention layer, the Q (query), K (key value), and V (value) matrices need to be calculated first, and their calculation formulas are as follows:

[0086] Q = W Q X

[0087] K = W K X

[0088] V = W V X

[0089] Among them, W Q , W K , W V is the weight matrix, belonging to learnable parameters, and is updated with the backpropagation algorithm during the training of the network. X is the feature vector matrix obtained after position encoding of the original input.

[0090] After the calculation of the Q, K, and V matrices, the attention matrix is calculated using the following formula

[0091]

[0092] Among them, d k is the number of columns of the Q and K matrices.

[0093] The multi-head self-attention layer consists of h self-attention layers. During the actual calculation process, first, the feature vector matrix X is input into h self-attention layers respectively to obtain h output matrices Z1 to Z h . After concatenating and linearly transforming all output matrices, the final output matrix Z is obtained, and its calculation formula is as follows:

[0094] Z = MultiHead(Q, K, V) = Concat(Z1,..., Z h )W O

[0095] Among them, Concat represents the matrix concatenation operation, and W O represents the weight matrix, which is updated with the backpropagation during the network training.

[0096] It should be noted that in the multi-head attention layer, d k is not equal to d model , but is calculated by the following formula:

[0097]

[0098] Among them, h represents the number of self-attention layers.

[0099] After the eigenvector matrix X is input into the encoder module, the output matrix Z of the encoder layer with the same dimension is obtained, which serves as the input to the parameter regression layer. To match the dimensions of the input and output prediction data of the parameter regression layer, a fully connected layer and a pooling layer are added after the encoder layer. Among them, the pooling layer adopts the average pooling operation. By averaging the features in the local area, the spatial dimension of the feature map is effectively reduced, thus ensuring the matching of the input dimension of the subsequent layer and the model output dimension. Average pooling not only helps reduce the computational complexity but also preserves the global information of the input features to a certain extent, avoiding overfitting.

[0100] The calculation formula of the fully connected layer is as follows:

[0101] f3(x) = W3x + b3

[0102] The calculation formula of the pooling layer is as follows:

[0103]

[0104] Among them, i is the index after pooling, x i·s+j is an element in the input vector, k is the size of the pooling window, and s is the stride of the pooling.

[0105] After establishing the analytical redundancy deep learning network model, the constructed dataset is used for network training. Based on the flight conditions and the sensor measurement data within the time sliding window, the generation of the rotational speed sensor signal based on the deep learning network can be achieved.

[0106] The detailed steps of step 3) are as follows:

[0107] Step 3.1), quantitatively compare and evaluate the output signal accuracy and time consumption of the rotational speed rapid generation sub-module of the Transformer architecture and the output signal of the model-based rotational speed rapid generation sub-module. According to the evaluation results, select the rotational speed rapid generation sub-module under the Transformer architecture with superior comprehensive performance as the main operating unit, and switch to the model-based rotational speed rapid generation sub-module when the error shows an increasing trend.

[0108] Taking the high-pressure rotational speed as an example, quantitatively compare and evaluate the output signal accuracy and time consumption of the rotational speed rapid generation sub-module of the Transformer architecture and the output signal of the model-based rotational speed rapid generation sub-module. As Figure 5 shown, it can be seen from the simulation results that the output signal accuracy of the rotational speed rapid generation sub-module of the Transformer architecture is slightly worse, but it meets the accuracy requirements and has better real-time performance. Therefore, when its accuracy meets the requirements, only this module is run to save time consumption. When the error trend increases, it switches to the model-based rotational speed rapid generation sub-module with better accuracy, and in order to ensure the smoothness of the signal during the switching process, a weight calculation strategy for the two is designed.

[0109] In step 3.2), count the output signal error of the rotational speed fast generation sub-module of the Transformer architecture within the time window, calculate the percentage change rate of the error at adjacent moments, and determine whether to switch to the model-based rotational speed fast generation sub-module according to the maximum change rate within the window.

[0110] First, count the output signal error of the rotational speed fast generation sub-module of the Transformer architecture within the time window, as shown in the following formula:

[0111]

[0112] where y real represents the true signal, y trans represents the output signal of the rotational speed fast generation sub-module of the Transformer architecture, e represents the error, and the subscript t represents the moment.

[0113] Secondly, calculate the percentage change rate of the error at adjacent moments within the time window, as shown in the following formula:

[0114]

[0115] where ò is a very small positive number set to prevent division by zero, and r is the percentage change rate of the error.

[0116] Next, count the maximum percentage change rate of the error within the window, as shown in the following formula:

[0117] r max = max(|r t-1 |,|r t-2 |,...,|r t-5 |)

[0118] where r max is the maximum percentage change rate of the error within the window.

[0119] Finally, determine whether to switch to the model-based rotational speed fast generation sub-module or switch back to the rotational speed fast generation sub-module of the Transformer architecture according to whether r max is greater than the set threshold, as shown in the following formula:

[0120]

[0121] where θ is the set threshold.

[0122] Step 3.3), in order to maintain the smoothness of the switching process, a switching process weight calculation strategy is designed to achieve a smooth transition of the output signal from the rotational speed rapid generation sub-module of the Transformer architecture to the model base rotational speed rapid generation sub-module. The switching logic and the switching process weight calculation are integrated to construct a fusion strategy.

[0123] During the switching process, in order to keep the signal stable, a weighted method is adopted to output the fused signal. The output signal during the switching process is shown in the following formula:

[0124]

[0125] where ω is the weight of the rapid generation sub-module of the Transformer architecture, and y is the final output signal.

[0126] The switching process weight calculation is shown in the following formula:

[0127]

[0128] where Δ is the set step size of the weight change process.

[0129] The final fused signal strategy is shown in the following formula:

[0130]

[0131] Verify the performance of the three signal generation methods. When the main fuel flow at the ground point increases linearly with time, verify the accuracy and time consumption of the three sensor signal generation methods under the fuel dynamic change process. The simulation results are as Figure 6 、 Figure 7 shown.

[0132] In summary, the designed method for rapidly generating a model of an aero-engine rotational speed sensor based on signal fusion can better provide generated signals and reflect the true working state of the engine, which has a positive significance for providing rotational speed sensor signals.

[0133] It should be noted that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes and substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A fast generation design method for redundant signals of an aero-engine speed sensor, characterized in that It includes the following steps: (1) Simplify the complete component-level model by introducing measured sensor parameters, establish a speed rapid generation sub-module based on the model for the speed sensor, and realize the generation of speed sensor signals based on the model; (2) Construct a data set according to flight conditions and measured sensor data, train and establish a speed rapid generation sub-module under the Transformer architecture, and realize the generation of speed sensor signals based on the deep learning network; (3) Compare the output accuracy and time consumption of the two sub-modules, design a switching logic based on the error change rate within the sliding window, and realize signal smooth fusion through the weighting coefficient.

2. The method according to claim 1, wherein Step (1) includes: Introduce the measured values of the fan outlet temperature and pressure sensors to generate the low-pressure rotor speed; Introduce the measured values of the total temperature and total pressure sensors at the inlet and outlet of the compressor and the measured value of the low-pressure turbine outlet pressure sensor, and combine the low-pressure rotor speed to generate the high-pressure rotor speed.

3. The method according to claim 1, wherein In step (2): The data set construction uses the sliding window method to extract time series features; The Transformer network includes an encoder, a fully connected layer and a pooling layer, and processes the sensor time series data through the multi-head self-attention mechanism.

4. The method according to claim 1, wherein The specific steps in step 3 are as follows: Step 3.1: Quantitatively compare and evaluate the output signal accuracy and time consumption of the speed rapid generation sub-module under the Transformer architecture and the speed rapid generation sub-module based on the model. According to the evaluation results, select the speed rapid generation sub-module under the Transformer architecture with superior comprehensive performance as the main operation unit, and switch to the speed rapid generation sub-module based on the model when the error shows an increasing trend; Step 3.2: Statistically analyze the output signal error of the speed rapid generation sub-module under the Transformer architecture within the time window, and calculate the percentage change rate of the error at adjacent moments. Judge whether to switch to the speed rapid generation sub-module based on the model according to the maximum change rate within the window; Step 3.3: Realize the smooth transition of the output signal from the speed rapid generation sub-module under the Transformer architecture to the speed rapid generation sub-module based on the model through dynamic weight calculation, specifically including: Set the step size of the change of the weighting coefficient in the transition stage; Adjust the output weights of the two models step by step during the switching process; Output the final signal after weighted fusion.

5. An aero-engine speed sensor signal generation system, characterized in that, It includes: A data acquisition module for obtaining engine sensor data; A model-based generation module for performing the generation calculation of step (1) described in claim 1; A deep learning generation module for performing the generation calculation of step (2) described in claim 1; A fusion control module for performing the fusion strategy of step (3) described in claim 1.

6. The system according to claim 5, wherein: The model-based generation module includes an air path calculation unit and an iterative solution unit; The deep learning generation module includes a data preprocessing unit and a Transformer encoding unit.

7. The system according to claim 5, wherein: The fusion control module includes an error analysis unit and a switching control unit; The system further includes an output module for outputting the final fusion signal.

8. An aeroengine control system, characterized in that, It includes: The signal generation system according to any one of claims 5-7; A control unit for adjusting engine operating parameters based on a generated rotational speed signal.

9. A computer-readable storage medium, characterized in that, A computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of claims 1-4 are implemented.

10. An aeroengine, characterized in that, Installed with the signal generation system according to any one of claims 5-7 for providing a rotational speed signal.

Images