Aero-engine performance degradation prediction method based on limited airborne sensor and GA-LSTM combined architecture driving, medium and computer program

Abstract

The invention discloses an aero-engine performance degradation prediction method driven based on a limited airborne sensor and a GA-LSTM combined architecture, a medium and a computer program. The method comprises the following steps: acquiring engine operation state data by using a standard airborne sensor with limited configuration; constructing a time series data set with state parameters as input and performance parameters as output, then constructing an engine performance prediction model based on a long short-term memory (LSTM) network, and optimizing structure parameters and training hyper-parameters of the LSTM network by using a genetic algorithm (GA); and finally, performing performance prediction and degradation trend evaluation based on the optimized LSTM model. According to the invention, through a GA-LSTM combined architecture, a complex nonlinear relationship in an engine performance degradation process can be effectively modeled, and prediction precision and generalization ability are improved. The method has good engineering adaptability and deployment feasibility, and has a wide application prospect in the field of aero-engine health monitoring and fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of aero-engine health monitoring and fault diagnosis, and relates to engine performance degradation modeling and prediction technology, in particular to a genetic algorithm (GA) and long short-term memory (LSTM) combined architecture driven aero-engine performance degradation prediction model modeling method, medium and computer program, which is used for realizing high-precision prediction of key performance parameters such as engine thrust under the condition of limited on-board sensors. BACKGROUND

[0002] As the core power device of an aircraft, the performance of an advanced aero-engine will degrade due to various reasons such as carbon deposition in the combustion chamber and blade deformation during use, making it difficult to meet the thrust and specific fuel consumption requirements of the target mission, and thus requiring regular maintenance and repair. Existing engines mainly adopt a fixed-time and fixed-period maintenance and repair scheme. In order to ensure safety during use, the maintenance interval is often short, which is prone to over-maintenance, resulting in increased cost and short on-wing time. Therefore, a health management system is the main technical path for realizing the transformation from traditional fixed-time maintenance to efficient condition-based maintenance, and how to accurately predict aero-engine performance parameters represented by thrust and specific fuel consumption is the key to the landing of the health management system.

[0003] Existing aero-engine performance degradation prediction algorithms mainly rely on two categories of methods based on physical modeling and data-driven methods. The former relies on thermodynamic mechanism modeling and component-level performance simulation, and its model construction is complex, parameter acquisition is difficult, and it is difficult to adapt to individual differences and environmental changes. The latter uses statistical learning, neural networks, deep learning and other methods to process a large amount of operating data, which can achieve high modeling accuracy and flexibility, and has been widely used in engine performance parameter trend modeling. However, existing data-driven methods use model simulation data as input, which has the following two problems:

[0004] (1) The prediction algorithm requires too many inputs. Since the model calculation process does not need to consider the sensor arrangement problem, aerodynamic parameters at any position can be obtained, and research based on this often inputs a large number of parameters in order to pursue prediction accuracy. When used on real engines, the prediction algorithm fails because there is no sensor to obtain the corresponding parameters. For example, in the current prediction model structure, the input often includes the total temperature Tt4 at the outlet of the combustion chamber, which is extremely high, and no sensor can work at this position for a long time to obtain Tt4; for another example, in the design process of existing prediction algorithms, engine life is needed, and these works are mainly based on the C-MAPSS public data set on the Internet, which contains a large amount of simulation data of engines, and even contains engine life data. However, in actual engineering, the real service life of an engine is difficult to effectively obtain.

[0005] (2) Simulation data and actual results deviation. The causes of engine performance degradation are various, the mechanism is unclear, the engine performance degradation model itself has low precision, and cannot output degradation data other than known mechanism. Inaccurate model simulation makes it difficult to fully reflect the real engine degradation process, resulting in low accuracy of the prediction algorithm in actual use, making it difficult to achieve accurate prediction of thrust, and the existing public technology generally lacks the support and verification of test data.

[0006] In addition, the existing neural network model usually relies on manual experience setting in network structure design and hyperparameter optimization, and lacks systematic optimization mechanism. Especially in the context of limited data volume or complex feature distribution, traditional network is easy to fall into local optimum, which limits the further improvement of prediction performance.

[0007] In the prior art, Chinese patent CN1 05547705B discloses an engine performance degradation trend prediction method, which improves the wear prediction accuracy through oil monitoring data reconstruction and oil change compensation correction, focuses on lubricating oil wear particle concentration analysis, and belongs to different technical scenarios with aerodynamic performance parameter prediction. The engine performance degradation prediction method (METHOD FOR ENGINE PERFORMANCE DEGRADATION PREDICTION BASED ON THE EC-RBELM ALGORITHM) based on the EC-RBELM algorithm disclosed in US2022 / 0300808A1 establishes a model cluster in multiple atmospheric environments and adopts minimum variance weighted fusion prediction result, its input still needs multiple aerodynamic parameters, and relies on multiple offline trained sub-models, without optimizing the number of input parameters and the complexity of model structure.

[0008] In summary, the existing aviation engine performance prediction technology has significant deficiencies in input parameter demand, simulation and actual deviation, neural network parameter optimization, and real data verification. Therefore, how to construct a performance degradation prediction method with high prediction accuracy, strong generalization ability and optimization efficiency on the basis of limited on-board sensors is a technical problem to be solved. SUMMARY

[0009] (I) Invention purpose

[0010] The application aims to provide an aero-engine performance degradation prediction method based on a limited airborne sensor and a GA-LSTM combined architecture, a medium and a computer program, which are mainly used to solve the problem of large demand for data types and data volume in the existing aero-engine performance prediction algorithm. In the traditional algorithm, at least 15 different input signals are required for prediction, and they cannot be arranged in real aero-engines. The application uses a GA-LSTM combined network to optimize the structure and parameters of the long short-term memory network (LSTM) by using the genetic algorithm (GA), realizes modeling and prediction of the engine performance change process, reduces the input signal requirement to 7 by optimizing the input parameters, and all of them are sensors commonly configured in existing aero-engines, which has good engineering applicability and deployment feasibility. In addition, the prediction model structure in the application is verified to be the optimal structure under the current input by using real long-time test data, which reduces the influence of any input on the model prediction accuracy and ensures the scientificity and reliability of the model structure.

[0011] (II) Technical solutions

[0012] To achieve the object of the application and solve the technical problems, the application adopts the following technical solutions:

[0013] The first object of the application is to provide an aero-engine performance degradation prediction method based on a limited airborne sensor and a GA-LSTM combined architecture, which is used to realize high-precision prediction and degradation trend evaluation of key performance parameters such as engine thrust under the condition of limited airborne sensors. The method at least includes the following steps when implemented:

[0014] S100. Data acquisition and model input-output architecture:

[0015] Based on long-time testing on the table or in-flight operation, a plurality of sets of time series operation data of the aero-engine under a plurality of operation cycles are acquired by a limited configuration of airborne sensors, including state parameters representing the operation state of the engine and performance parameters representing the performance of the engine, which are respectively used as input variables and output variables of the prediction model. The acquired data are preprocessed and divided into a training set and a test set according to the time sequence;

[0016] S200. LSTM prediction model structure design and initialization:

[0017] An engine performance prediction model based on a long short-term memory (LSTM) network is constructed, including an input layer, a hidden layer and an output layer. The input layer is used to receive an input variable sequence composed of state parameters and map it to a neural network feature space. The hidden layer contains a plurality of LSTM units and is used to model the time correlation in the input variables, the state evolution characteristics and the nonlinear mapping relationship between the state evolution characteristics and the performance response. The output layer is used to output the engine target performance index value predicted based on the input state parameters. Structure parameters and hyperparameters of the LSTM network are initialized and set as optimization variables to be input into a subsequent optimization process.

[0018] S300. GA algorithm parameter encoding and initial population generation

[0019] A chromosome coding scheme is designed according to the constructed LSTM network structure, and the structure parameters and training hyperparameters of the LSTM network are encoded into a chromosome representation in a genetic algorithm (GA). GA algorithm running parameters are set, and an initial population is randomly generated. Each individual represents a complete set of parameter configurations of the LSTM network, which is used as an initial solution set for model optimization.

[0020] S400. GA-LSTM fusion optimization training

[0021] Each individual in the population is decoded into a set of LSTM network parameter configurations based on chromosome coding, and a temporary LSTM prediction model corresponding to the set of parameters is dynamically constructed. Then, each temporary model is trained using the training set data. A fitness function is constructed based on the model performance evaluation index, and the fitness value of the LSTM temporary model corresponding to each individual is calculated. The gene structure of the population individuals is updated by performing selection, crossover and mutation operations of the GA algorithm, and the parameter combination of the LSTM network is gradually optimized.

[0022] S500. Iterative optimization and convergence judgment

[0023] The genetic operations in step S400 are repeatedly performed, and the optimal fitness value and the average fitness value of the current population are evaluated after each iteration. The iteration process is terminated when the preset maximum number of iterations or the fitness function convergence condition is reached. The chromosome individual with the best fitness value is selected from the final population as the optimal LSTM network parameter configuration.

[0024] S600. Optimal prediction model construction and performance verification

[0025] Decode the optimal chromosome individual obtained in step S500 into the structure parameters and training hyperparameters of the final LSTM prediction model, and model train the LSTM network with the optimal configuration based on the complete training set to form a target prediction model with actual deployment capability, and evaluate the prediction accuracy and generalization ability of the model based on the test set data.

[0026] The second inventive purpose of the present application is to provide a computer program product comprising computer instructions for executing the above-mentioned aero-engine performance degradation prediction method based on a combination of limited on-board sensors and GA-LSTM architecture.

[0027] The third inventive purpose of the present application is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned aero-engine performance degradation prediction method based on a combination of limited on-board sensors and GA-LSTM architecture.

[0028] (III) Technical effects

[0029] Compared with the prior art, the aero-engine performance degradation prediction method based on a combination of limited on-board sensors and GA-LSTM architecture, medium and computer program of the present application have the following beneficial and significant technical effects: The present application significantly reduces the model input required by the traditional performance prediction algorithm. For example, to implement engine performance prediction using the C-MPASS data set, a total of 21 parameters including total temperature, total pressure, pressure ratio, and fuel ratio are required. Some scholars have reduced the model input to 15 by conducting correlation analysis on the data, but for real aero-engines, especially light aero-engines, the required input is still too much and cannot be met. The present application uses a real long-time test data driven method to propose a GA-LSTM combined network prediction structure, which realizes engine performance prediction under the premise of 7 inputs including engine use time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow Wf, exhaust temperature Tt5, high-pressure rotor speed NH, and low-pressure rotor speed NL, greatly optimizes the input conditions, and these 7 sensors are essential key sensors for traditional aero-engines, so the present application has a clear engineering application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The flowchart of the aero-engine performance degradation prediction method based on a combination of limited on-board sensors and GA-LSTM architecture of the present application is shown in the figure.

[0031] Figure 2 The input and output structure diagram of the prediction model is shown in the figure.

[0032] Figure 3 The LSTM network structure diagram is shown in the figure.

[0033] Figure 4 GA-LSTM performance prediction model training progress schematic diagram;

[0034] Figure 5 GA-LSTM performance prediction model result schematic diagram;

[0035] Figure 6 Performance prediction result schematic diagram of input dimension reduction, wherein (a) is a prediction result without low-pressure rotating speed input, (b) is a prediction result without high-pressure rotating speed input, (c) is a prediction result without fuel flow input, and (d) is a prediction result without exhaust gas temperature input. DETAILED DESCRIPTION

[0036] The present application aims to provide an aero-engine performance degradation prediction method, medium and computer program based on a limited on-board sensor and GA-LSTM combined architecture, mainly for solving the problem of large demand for data types and data volume in existing aero-engine performance prediction algorithms, so as to realize high-precision prediction and degradation trend evaluation of key performance parameters such as engine thrust under the condition of limited on-board sensors. In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiment of the present application will be described in more detail below in combination with the drawings in the embodiment of the present application.

[0037] Example 1: GA-LSTM combined architecture prediction method

[0038] As a specific example, as shown in Figure 1 The aero-engine performance degradation prediction method based on a limited on-board sensor and GA-LSTM combined architecture provided by the embodiment of the present application mainly includes the following steps when implemented:

[0039] S100. Data acquisition and model input-output architecture:

[0040] Based on the on-site long-time test or in-flight operation process, a plurality of sets of time series operation data of the aero-engine under a plurality of operation cycles are collected through the limited configuration of the on-board sensor, including state parameters representing the operation state of the engine and performance parameters representing the performance of the engine, which are respectively used as input variables and output variables of the prediction model; the collected data are preprocessed and divided into a training set and a test set according to the time sequence.

[0041] As preferred, the state parameters characterizing the engine operating state at least include engine service time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow Wf, exhaust temperature Tt5, high-pressure rotor speed NH and low-pressure rotor speed NL, and the performance parameters characterizing the engine performance and serving as the predicted target output parameters at least include engine thrust F. Each state parameter is a basic operating parameter directly measurable by an on-board sensor of a standard configuration of an aero-engine, without the need for additional sensor arrangement, thereby ensuring the engineering applicability and practical deployment feasibility of the method.

[0042] In addition, in the embodiment of the present application, the collected operating data is preprocessed, at least including:

[0043] S101. Structurally sorting the collected multiple sets of operating data in time sequence, to ensure that each input and output variable has a strict time sequence correspondence;

[0044] S102. For missing values, abnormal values and high-frequency noise possibly existing in the original collected data, respectively using interpolation method, sliding window filtering and abnormal point elimination for cleaning and correction;

[0045] S103. Numerical normalization of input variables and output variables respectively by using normalization, standard deviation standardization or minimum-maximum scaling method, to enhance the numerical stability and convergence efficiency of the model;

[0046] S104. Dividing the complete data set into training set and test set (the division ratio is preferably 80:20) in time sequence, and reserving part of the data as an independent validation set for model parameter tuning and overfitting control, to support the training and generalization ability evaluation of the subsequent LSTM model.

[0047] S200. LSTM prediction model structure design and initialization:

[0048] The engine performance prediction model based on LSTM network is constructed, including input layer, hidden layer and output layer. The input layer is used to receive the input variable sequence composed of state parameters and map it to the neural network feature space. The hidden layer contains several LSTM units and is used to model the time correlation, state evolution characteristics in the input variables and the nonlinear mapping relationship between them and the performance response. The output layer is used to output the engine target performance index value predicted based on the input state parameters. The structure parameters and hyperparameters of the LSTM network are initialized and set as optimization variables input into the subsequent optimization process.

[0049] In the embodiment of the present application, the LSTM prediction model further comprises a residual connection structure for inhibiting gradient disappearance, the input layer of which embeds original input variables into a high-dimensional feature space through an embedding mapping module, the hidden layer comprises a plurality of LSTM units, and each LSTM unit is followed by a fully connected layer to enhance the nonlinear fitting capability of the network, and the output layer adopts a linear activation function to adapt to the regression prediction of continuous variables; and wherein the LSTM model parameter settings at least include input and output parameter dimensions, hidden layers and fully connected layers, iteration number and training batch size, maximum training number and minimum batch, and learning rate and validation period settings.

[0050] More specifically, the LSTM network structure used in the present application is as shown in Figure 2 Each LSTM unit comprises three key components of a forgetting gate f t , an input gate i t and an output gate o t , wherein the forgetting gate determines whether to retain the cell state C t-1 at time t-1 through the formula f t =σ(W f x t +U f h t-1 +b f ), the input gate is used to determine the degree of writing new information at the current time t through the formula i t =σ(W i x t +U i h t-1 +b i ), and the output gate determines the output at time t through the formula o t =σ(W o x t +U o h t-1 +b o ), wherein x t is the input variable at the current time t, h t-1 is the hidden state at the previous time, W, U and b are respectively the weight matrix, the recurrent connection weight matrix and the bias matrix corresponding to each gate, and σ represents a sigmoid function for mapping the input into the interval of 0-1; on the basis of the gating results, the cell state C t and the hidden state h t at the current time t are updated through the formula C t =f t *C t-1 +i t *tanh(W c ·x t +U c ·h t-1 +b c) and h t = o t * tanh(C t ) is updated, * denotes unit multiplication, and tanh is a hyperbolic tangent activation function for nonlinear transformation of the current state.

[0051] S300. GA algorithm parameter encoding and initial population generation:

[0052] According to the constructed LSTM network structure, a chromosome coding scheme is designed to encode the structure parameters and training hyperparameters of the LSTM network into a chromosome representation in the GA algorithm; GA algorithm operation parameters are set, and an initial population is randomly generated, each individual representing a complete set of parameter configurations of the LSTM network, serving as an initial solution set for model optimization.

[0053] In the embodiment of the application, the chromosome coding adopts a real value coding method, the LSTM network structure parameters and training hyperparameters are represented by numerical values with a preset precision, and the individual genes are represented by floating-point chromosome structures in the GA algorithm, and the chromosome length is equal to the total number of all parameters to be optimized; the set GA algorithm operation parameters at least include population size, crossover probability, mutation probability and maximum iteration number.

[0054] S400. GA-LSTM fusion optimization training:

[0055] Each individual in the population is decoded into a set of LSTM network parameter configurations based on the chromosome coding, and a temporary LSTM prediction model corresponding thereto is dynamically constructed using the set of parameters, and then each temporary model is trained using the training set data; a fitness function is constructed based on the model performance evaluation index, and the fitness value of each individual LSTM temporary model is calculated; the gene structure of the population individuals is updated according to the fitness value by performing selection, crossover and mutation operations of the GA algorithm, and the parameter combination of the LSTM network is gradually optimized.

[0056] In the embodiment of the application, the GA-LSTM fusion optimization training process at least includes the following sub-steps:

[0057] S401. Individual decoding and temporary model construction: each individual in the population is decoded into a set of LSTM network parameter configurations based on the chromosome coding scheme, and a corresponding temporary LSTM prediction model is dynamically constructed to ensure that each individual can participate in training and evaluation as an independent model architecture;

[0058] S402. Temporary model training and performance evaluation: each temporary LSTM model is trained using the training set data, the training process updates the network weights using the backpropagation algorithm, and the model parameter optimization is realized by minimizing the loss function between the predicted output and the true value; the training termination conditions are set, including the maximum training number and the loss function convergence criterion;

[0059] S403. Fitness function construction and calculation: the fitness value is calculated based on the prediction performance of the completed LSTM model on the validation set, and the fitness function is constructed based on the root mean square error (RMSE) for comprehensive measurement of the prediction performance of the temporary LSTM model, and the fitness value is used as the basis for selection, crossover and mutation operations in GA evolution;

[0060] S404. Genetic operation execution: based on the fitness value of each individual, selection, crossover and mutation operations in the GA algorithm are performed, and a roulette selection strategy is used to retain high fitness individuals as parents; crossover operation is performed on the selected parent individuals to generate offspring individuals using arithmetic crossover; mutation operation is performed on the crossed individuals, and the chromosome genes are randomly disturbed according to the preset mutation probability;

[0061] S405. New generation population formation and iterative control: the new individuals generated by crossing and mutation and the excellent individuals reserved are combined to form the next generation population, and the current population set is updated; the optimal fitness value and the average fitness value of the current generation are recorded to provide a basis for subsequent convergence judgment and early termination.

[0062] S500. Iterative optimization and convergence judgment:

[0063] The genetic operation in step S400 is repeatedly performed, and the optimal fitness value and the average fitness value of the current population are evaluated after each iteration; the iteration process is terminated when the preset maximum iteration number is reached or the fitness function convergence condition is met; the chromosome individual with the best fitness value is selected from the final population as the optimal LSTM network parameter configuration.

[0064] In the embodiment of the application, the convergence condition includes at least one of the following cases: the iteration number reaches the preset maximum iteration number upper limit; the current optimal fitness value reaches the preset target threshold; the change amplitude of the optimal fitness value of the continuous multiple generations is less than the set micro-amplitude convergence difference; the population fitness variance converges to be lower than the preset threshold; the iteration is terminated when any convergence condition is met.

[0065] S600. Optimal prediction model construction and performance verification:

[0066] The optimal chromosome individual obtained in step S500 is decoded into the structure parameters and training hyperparameters of the final LSTM prediction model, and the optimal configured LSTM network is modeled and trained based on the complete training set to form a target prediction model with actual deployment capability, and the prediction accuracy and generalization ability of the model are evaluated based on the test set data.

[0067] S700. Model input variable sensitivity analysis and optimal input configuration:

[0068] Based on the established optimal LSTM prediction model, several sets of input variable reduction configuration schemes are constructed. After removing each single input variable in turn, the model training and performance evaluation process is repeated. The performance index changes of the model on the test set under different input configurations are compared to determine the degree of influence of each input variable on the model prediction accuracy. The smallest parameter input group with the least impact and prediction accuracy meeting the preset threshold is selected as the optimal input parameter configuration scheme. This achieves the goal of minimizing the dependence on the number of airborne sensors while ensuring model accuracy, thereby improving the model's structural versatility and airborne deployment feasibility.

[0069] Example 2: Application Example Based on Long-Term Test Data

[0070] Based on the aero-engine performance degradation prediction method driven by a finite airborne sensor and GA-LSTM combined architecture shown in Embodiment 1 above, to further verify the effectiveness and feasibility of this method in practical engineering applications, the following is an example of the implementation process based on long-term test data of a certain type of light aero-engine to further illustrate the present invention. The specific steps are as follows:

[0071] First, experimental data is prepared, and the parameters required for training the model in this invention are extracted from it. These parameters are: engine operating time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow rate Wf, exhaust temperature Tt5, high-pressure rotor speed NH, low-pressure rotor speed NL, and engine thrust F. The parameters are functionally divided into input parameters and output parameters. The input parameters include engine operating time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow rate Wf, exhaust temperature Tt5, high-pressure rotor speed NH, and low-pressure rotor speed NL. The output parameter includes engine thrust F, which is used as the prediction target. The above data is divided into training and testing sets according to the experimental time series. In this example, a total of 35 cycles of long-term test data were collected. Data from cycles 1 to 34 were used for training to obtain the performance prediction model. Data from cycle 35 was used to verify the effectiveness of the established model in predicting future performance degradation trends. The input-output structure of the constructed prediction model is as follows: Figure 2 As shown.

[0072] Secondly, based on the process of Example 1, an engine performance prediction model based on an LSTM network is built. The LSTM network structure is as follows: Figure 3 As shown in Table 1, the network includes an input layer, a hidden layer, and an output layer. The hidden layer contains several LSTM units. The settings for the input and output parameters of the LSTM network, the settings for the hidden and fully connected layers, the number of iterations and the training batch size, the maximum number of training iterations and the minimum batch size, and the learning rate and validation period are as follows.

[0073] Table 1 Key Parameter Settings for LSTM Networks

[0074]

[0075] Third, in order to achieve adaptive optimization of the predictive model structure parameters, the genetic algorithm optimization parameter settings are further given according to the method in Example 1, as shown in Table 2 below.

[0076] Table 2 Key parameter settings for the genetic algorithm

[0077]

[0078] Fourth, after initializing the LSTM network structure and configuring the genetic algorithm parameters, based on the set GA parameters and chromosome encoding method, run the GA-LSTM fusion optimization training process. Use the training set to conduct network training, output results, and evaluate prediction performance. Specifically, follow the process defined in steps S401 to S405, performing operations such as individual decoding guided by the genetic algorithm, temporary model training, fitness calculation, population evolution, and convergence determination. In this example, the input and output data are first normalized, and then the program is run in the Simulink environment to generate the entire process root mean square error (RMSE) and loss (LoS) decrease curves, as shown below. Figure 4 As shown.

[0079] Furthermore, to verify the predictive performance of the optimized GA-LSTM combined model in real-world data scenarios, the engine usage time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow rate Wf, exhaust temperature Tt5, high-pressure rotor speed NH, and low-pressure rotor speed NL from the 35th cycle were input into the trained optimal LSTM model. The thrust prediction result Fpre output by the model was recorded and compared with the actual experimental data. Figure 5 As shown in the figure, the predicted curve and the measured curve have a high degree of fit, which verifies the prediction accuracy and usability of the model in the context of real engineering data.

[0080] Finally, we examine whether the input signals of this invention's architecture can be further reduced. Since engine operating time t, inlet total temperature Tt1, and inlet total pressure Pt1 characterize the engine's operating conditions and are key input parameters, this invention does not consider reducing these three parameters. In this invention, we sequentially reduce the inputs of low-pressure rotor speed NL, high-pressure rotor speed NH, fuel flow rate Wf, and exhaust temperature Tt5. Under the same architecture and training parameters, the simulation results are as follows... Figure 6As shown, the prediction results deviate significantly at high speeds when the low-pressure rotor speed NL is missing; the prediction results deviate significantly at both low and high speeds when the high-pressure rotor speed NH is missing; the prediction results deviate significantly near throttling and intermediate states when the fuel flow rate Wf is missing; and the prediction results deviate significantly near idle and intermediate states when the exhaust temperature Tt5 is missing. Simulation results show that reducing any input parameters in the currently designed model architecture will affect the prediction accuracy. The designed input parameters—engine usage time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow rate Wf, exhaust temperature Tt5, high-pressure rotor speed NH, and low-pressure rotor speed NL—are the optimal inputs for predicting engine thrust F.

[0081] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A method for predicting the performance degradation of aero-engines based on a finite number of airborne sensors and a GA-LSTM combined architecture, characterized in that, It should include at least the following steps: S100. Collect multiple sets of time-series operational data of the aero-engine under multiple operating cycles through a limited number of airborne sensors, including state parameters and performance parameters, which are used as model input variables and output variables, respectively. Preprocess the collected data and divide it into training set and test set according to time series. S200. Construct an engine performance prediction model based on an LSTM network. Its input layer is used to receive the sequence of input variables composed of state parameters, the hidden layer is used to model the time correlation, state evolution characteristics and nonlinear mapping relationship between the input variables and the performance response, and the output layer is used to output the target performance index value of the engine; initialize and set the structural parameters and hyperparameters of the LSTM network. S300. Based on the constructed LSTM network structure, design a chromosome encoding scheme, and encode its structural parameters and hyperparameters into a chromosome representation in the GA algorithm; set the GA running parameters, randomly generate an initial population, and each individual represents a set of parameter configurations of the LSTM network; S400. Decode each individual into a set of LSTM network parameter configurations, dynamically construct the corresponding temporary LSTM prediction model, train each temporary model using the training set data; construct a fitness function, calculate the fitness value of each temporary model; update the gene structure of the population individuals based on the fitness value; S500. Repeat the genetic operation, evaluate the optimal fitness value and average fitness value of the current population after each iteration; terminate the iteration process when the preset convergence condition is met; Select the chromosome with the best fitness value from the final population as the optimal LSTM network parameter configuration; S600. Decode the optimal chromosome individual into the structural parameters and hyperparameters of the final LSTM prediction model, and model and train the optimal LSTM network based on the training set. Evaluate the prediction accuracy and generalization ability of the model based on the test set data.

2. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, It also includes step S700 for sensitivity analysis of model input variables and confirmation of optimal input configuration, including: Based on the established optimal LSTM prediction model, several sets of input variable reduction configuration schemes are constructed. After removing each single input variable in turn, the model training and performance evaluation process is repeated. The performance index changes of the model on the test set under different input configurations are compared to determine the degree of influence of each input variable on the model prediction accuracy. The smallest parameter input group with the least impact and prediction accuracy meeting the preset threshold is selected as the optimal input parameter configuration scheme. This achieves the goal of minimizing the dependence on the number of airborne sensors while ensuring model accuracy, thereby improving the model's structural versatility and airborne deployment feasibility.

3. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, In step S100, the state parameters characterizing the engine operating state include at least the engine usage time t, inlet total temperature Tt1, inlet total pressure Pt1, fuel flow rate Wf, exhaust temperature Tt5, high-pressure rotor speed NH and low-pressure rotor speed NL, and the performance parameters characterizing the engine performance and serving as the predicted target output parameters include at least the engine thrust F.

4. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1 or 3, characterized in that, In step S100, the preprocessing of the collected running data includes at least the following: S101. The collected multiple sets of operational data are structured and organized in chronological order to ensure that each input and output variable has a strict time series correspondence; S102. For missing values, outliers and high-frequency noise that may exist in the original collected data, interpolation, sliding window filtering and outlier removal are used to clean and correct them respectively. S103. Normalization, standard deviation standardization, or minimum-maximum scaling are used to numerically normalize the input and output variables respectively to enhance the numerical stability and convergence efficiency of the model. S104. Divide the complete dataset into training and testing sets in chronological order, and reserve a portion of the data as an independent validation set for model parameter tuning and overfit control, in order to support the subsequent training and generalization ability evaluation of the LSTM model.

5. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, In step S200, the LSTM prediction model further includes a residual connection structure for suppressing gradient vanishing. Its input layer embeds the original input variables into a high-dimensional feature space through an embedding mapping module. The hidden layer contains multiple LSTM units, and each LSTM unit is followed by a fully connected layer to enhance the nonlinear fitting ability of the network. The output layer uses a linear activation function to adapt to the regression prediction of continuous variables. Furthermore, the LSTM model parameter settings include at least the dimensions of input and output parameters, hidden layers and fully connected layers, number of iterations and training batch size, maximum number of training iterations and minimum batch size, as well as the learning rate and validation period settings.

6. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1 or 5, characterized in that, In step S200, each LSTM unit includes a forget gate f t Input gate i t and output gate o t Three key components, of which the forget gate is determined by formula f t =σ(W f x t +U f h t-1 +b f Decide whether to retain the cell state C at time t-1. t-1 The input gate is accessed via formula i t =σ(W i x t +U i h t-1 +b i This is used to determine the extent of new information written at the current time t. The output gate is determined by formula o. t =σ(W o x t +U o h t-1 +b o The output at time t is determined by x. t h is the input variable at the current time t. t-1 The hidden state at the previous time step is represented by W, U, and b, which are the weight matrix, circular connection weight matrix, and bias matrix corresponding to each gate, respectively. σ represents the sigmoid function, used to map the input to the interval between 0 and 1. Based on the gating result, the cell state C at the current time step t is... t and hidden state h t Through formula C t =f t *C t-1 +i t *tanh(W c ·x t +U c ·h t-1 +b c ) and h t =o t *tanh(C t The update is performed using ), where * indicates element-wise multiplication, and tanh is the hyperbolic tangent activation function, used to perform nonlinear transformations on the current state.

7. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, In step S300, chromosome encoding adopts real-valued encoding, which represents the LSTM network structure parameters and training hyperparameters with preset precision. In the GA algorithm, individual genes are represented by floating-point chromosome structures, and the chromosome length is equal to the total number of all parameters to be optimized. The GA algorithm running parameters set include at least the population size, crossover probability, mutation probability, and maximum number of iterations.

8. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, In step S400, the GA-LSTM fusion optimization training process includes at least the following sub-steps: S401. Individual Decoding and Temporary Model Construction: Based on the chromosome encoding scheme, each individual in the population is decoded into a set of LSTM network parameter configurations, and the corresponding temporary LSTM prediction model is dynamically constructed to ensure that each individual can participate in training and evaluation as an independent model architecture; S402. Temporary Model Training and Performance Evaluation: Each temporary LSTM model is trained using the training set data. The backpropagation algorithm is used to update the network weights during the training process. The model parameters are optimized by minimizing the loss function between the predicted output and the true value. Training termination conditions are set, including the maximum number of training rounds and the loss function convergence criterion. S403. Fitness Function Construction and Calculation: The fitness value is calculated based on the prediction performance of the trained LSTM model on the validation set. The fitness function is constructed based on the root mean square error (RMSE) to comprehensively measure the prediction performance of the temporary LSTM model. The fitness value is used as the basis for selection, crossover, and mutation operations in GA evolution. S404. Genetic Operation Execution: Based on the fitness value of each individual, the selection, crossover, and mutation operations in the GA algorithm are executed. A roulette wheel selection strategy is used to retain individuals with high fitness as parents. A crossover operation is performed on the selected parent individuals to generate offspring individuals using arithmetic crossover. A mutation operation is performed on the individuals after crossover, and the chromosome genes are randomly perturbed according to a preset mutation probability. S405. New Generation Population Formation and Iteration Control: The new individuals generated through crossover and mutation, along with the superior individuals retained through replication, form the next generation population, updating the current population set; the optimal fitness value and average fitness value of the current generation are recorded, providing a basis for subsequent convergence judgment and early termination.

9. The method for predicting the performance degradation of aero-engines based on a finite airborne sensor and GA-LSTM combined architecture as described in claim 1, characterized in that, In step S500, the convergence condition includes at least one of the following: the number of iterations reaches the preset maximum number of iterations; the current optimal fitness value reaches the preset target threshold. The variation range of the optimal fitness value over multiple consecutive generations is less than the set micro-convergence difference; The iteration terminates when the population fitness variance converges to below a preset threshold and any convergence condition is met.

10. A computer program product comprising computer instructions, characterized in that, The computer instructions are used to execute the aero-engine performance degradation prediction method based on a finite airborne sensor and GA-LSTM combined architecture as described in any one of claims 1 to 9.

Images