Method and device for predicting state and degradation rate of aero-engine
Through the physical information neural network combined with partial differential equations, the aero engine degradation process is simulated, which solves the problem of insufficient accuracy of aero engine state prediction in the prior art, especially in the middle and late stages of engine degradation, which has achieved significant prediction results, which improves the accuracy of degradation rate prediction.
Patent Information
- Application Number
- CN202510738731.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-08-01
AI Technical Summary
The existing aero engine state prediction methods are insufficient in processing complex time series data and high temperature and high speed environments. The neural network model training is complex and the prediction effect is poor, making it difficult to accurately capture the degradation process of aero engines.
The Physics-Informed Neural Networks (PINN) is used to combine the first multi-layer perceptron and the second multi-layer perceptron, and use partial differential equations to simulate the engine degradation process, and optimize the first multi-layer perceptron to improve prediction accuracy.
It improves the accuracy of aircraft engine degradation rate prediction, especially in the middle and late stages of engine degradation, the prediction effect is significantly better than other models, and has higher physical information feature fitting ability and generalization ability.
Smart Images
Figure CN120408319A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of engine monitoring, and particularly to a method and device for predicting the state and degradation rate of an aeroengine. Background Technique
[0002] As a representative of modern industrial products, researching the full-life fault prediction and health management (Prognostics and Health Management, PHM) of aero-engines has important practical significance. For the full-life PHM of aero-engines, in the early stage, model-based fault prediction methods received extensive attention. The Gaussian degradation model can use the maximum likelihood method for parameter estimation and is applicable to the life estimation of two-phase degradation modes. The Wiener process model, as an independent increment process, has been widely used in the reliability field together with the gamma process and the inverse Gaussian process. At the same time, the computational complexity problem of the Gaussian process model when dealing with large-scale data has troubled researchers. And the Gaussian model needs to make certain assumptions when performing parameter estimation, which also has an adverse impact on the generality of the model. The Wiener process model can only describe data with monotonically increasing dispersion, and the prediction effect is often unsatisfactory for the state recovery caused by some minor repairs and self-healing. With the improvement of computer computing power and storage capacity in recent years, a series of machine learning methods have shone in the field of aero-engine state prediction. Some basic models, such as the multi-layer perceptron (MLP), convolutional neural network (CNN), and support vector regression (SVR) models, have long started to practice the problem of predicting the remaining life of industrial equipment. While achieving some results, some problems of these shallow neural networks have also been discovered. The MLP has a simple structure and is easy to learn non-linear mapping, but its ability to process time series is poor and it cannot utilize the historical relationship of time-related data. The CNN is suitable for extracting spatial features and is still difficult to process time-related data. And for industrial equipment such as aero-engines with high rotational speeds and high temperatures, the complexity of its data is much higher than that of products with simple degradation modes such as lithium batteries and gyroscopes. In the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) dataset, the number of fault modes of its engines can exceed 2, and the actual situation is often more complex than that in the dataset. Therefore, it is generally difficult for such shallow neural networks to accurately capture the state of aero-engines. Long short-term memory networks (LSTM) and recurrent neural networks (RNN) can process time series data and capture the feature relationships in the sequence, but they also bring problems such as high training costs and gradient explosions. Other neural networks, such as kernel ridge regression (RVR), variational autoencoders (VAE), etc., have been tested on large datasets and complex data problems. Simply relying on neural networks to predict data, its training process and internal structure are unknown, which obviously limits the development of neural networks and the prediction accuracy needs to be improved. Summary of the Invention
[0003] The purpose of this application is to provide a method and device for predicting the state and degradation rate of an aeroengine, which can improve the accuracy of degradation rate prediction.
[0004] To achieve the above object, this application provides the following solutions:
[0005] In the first aspect, this application provides a method for predicting the state and degradation rate of an aeroengine, including:
[0006] Obtain an aeroengine training set; each sample in the aeroengine training set includes input data and label data, the input data is a feature set of the aeroengine, the label data is a state variable corresponding to the input data, the state variable represents the performance of the aeroengine within a preset usage period, and the smaller the value of the state variable, the worse the performance;
[0007] Train a physics-informed neural network using the aeroengine training set to obtain an aeroengine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence, the first multi-layer perceptron is used to output a predicted value of the state variable of the aeroengine according to the feature set of the aeroengine; the second multi-layer perceptron is used to simulate the degradation process of the aeroengine according to the feature set of the aeroengine and the partial differential equation of the state variable, and output the degradation rate of the aeroengine; during the training process of the physics-informed neural network, use the degradation rate output by the second multi-layer perceptron to optimize the first multi-layer perceptron, and use the trained first multi-layer perceptron as the aeroengine state prediction model;
[0008] Input the feature set to be predicted into the aeroengine state prediction model, and output a predicted value of the state variable;
[0009] Calculate the remaining life of the aeroengine according to the predicted value of the state variable.
[0010] In the second aspect, this application provides an apparatus for predicting the state and degradation rate of an aeroengine, including:
[0011] A training set acquisition module, configured to obtain an aeroengine training set; each sample in the aeroengine training set includes input data and label data, the input data is a feature set of the aeroengine, the label data is a state variable corresponding to the input data, the state variable represents the performance of the aeroengine within a preset usage period, and the smaller the value of the state variable, the worse the performance;
[0012] A prediction model training module for training a physics-informed neural network with the aero-engine training set to obtain an aero-engine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence. The first multi-layer perceptron is used to output a predicted value of the state variable of the aero-engine according to the feature set of the aero-engine. The second multi-layer perceptron is used to simulate the degradation process of the aero-engine according to the feature set of the aero-engine and the partial differential equation of the state variable, and output the degradation rate of the aero-engine. During the training process of the physics-informed neural network, the degradation rate output by the second multi-layer perceptron is used to optimize the first multi-layer perceptron, and the trained first multi-layer perceptron is used as the aero-engine state prediction model;
[0013] A state variable prediction module for inputting the feature set to be predicted into the aero-engine state prediction model and outputting a predicted value of the state variable;
[0014] A remaining life calculation module for calculating the remaining life of the aero-engine according to the predicted value of the state variable.
[0015] According to the specific embodiments provided by the present application, the present application discloses the following technical effects: The present application provides a method and device for predicting the state and degradation rate of an aero-engine. By training a physics-informed neural network (Physics-Informed Neural Networks, PINN), an aero-engine state prediction model is obtained. The degradation rate is predicted by the physics-informed neural network based on the partial differential equation of the state variable. Among them, the partial differential equation based on the state variable can more realistically simulate the degradation process of the engine, improve the fitting ability of the physics-informed features, and make up for the generalization ability of the prediction simulation, thereby improving the accuracy of the degradation rate prediction. Description of the Drawings
[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0017] Figure 1 It is a schematic flow chart of a method for predicting the state and degradation rate of an aero-engine provided by an embodiment of the present application.
[0018] [[ID=2I]] Figure 2 It is a schematic structural diagram of a multi-layer perceptron with a hidden layer provided by an embodiment of the present application.
[0019] Figure 3 It is a schematic structural diagram of a physics-informed neural network provided by an embodiment of the present application. Detailed implementation manners
[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.
[0021] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.
[0022] The present application provides a method for predicting the state and degradation rate of an aeroengine, as Figure 1 shown. The method for predicting the state and degradation rate of the aeroengine includes steps 101 to 103.
[0023] Step 101: Obtain an aeroengine training set; each sample in the aeroengine training set includes input data and label data. The input data is a feature set of the aeroengine, and the label data is a state variable corresponding to the input data. The state variable represents the performance of the aeroengine within a preset usage period. The smaller the value of the state variable, the worse the performance.
[0024] Within the preset usage period of the aeroengine, after the usage time reaches the set time, as the usage time extends, the state variable linearly decreases. That is, the label data is determined by the position of the current state in the time series. When the last data in the current state is greater than 125, the label data is 1 at this time. If it is less than 125, it linearly decreases. Therefore, the line type of the label data changing with time is a broken line. First, it is all 1, and then it decreases from 1 to 0.
[0025] Step 102: Train a physics-informed neural network using the aeroengine training set to obtain an aeroengine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence. The first multi-layer perceptron is used to output a predicted value of the state variable of the aeroengine according to the feature set of the aeroengine. The second multi-layer perceptron is used to simulate the degradation process of the aeroengine according to the feature set of the aeroengine and the partial differential equation of the state variable, and output the degradation rate of the aeroengine. During the training process of the physics-informed neural network, the degradation rate output by the second multi-layer perceptron is used to optimize the first multi-layer perceptron, and the trained first multi-layer perceptron is used as the aeroengine state prediction model.
[0026] Step 103: Input the feature set to be predicted into the aero-engine state prediction model, and output the predicted value of the state variable.
[0027] Step 104: Calculate the remaining life of the aero-engine based on the predicted value of the state variable.
[0028] Among them, in the process of training the physics-informed neural network in Step 102, during the training process of the physics-informed neural network, the first multi-layer perceptron is trained with the feature set as the input and the state variable as the output, and the first multi-layer perceptron is optimized by using the degradation rate output by the second multi-layer perceptron. The trained first multi-layer perceptron is used as the aero-engine state prediction model.
[0029] Optimizing the first multi-layer perceptron by using the degradation rate output by the second multi-layer perceptron specifically includes: calculating the predicted value of the degradation rate according to the predicted value of the state variable output by the first multi-layer perceptron; optimizing the first multi-layer perceptron according to the loss between the degradation rate output by the second multi-layer perceptron and the predicted value of the degradation rate.
[0030] In this application, the aero-engine state prediction model is obtained by training the physics-informed neural network. The degradation rate of the aero-engine is simulated by the physics-informed neural network based on the partial differential equation of the feature set and the state variable. Among them, the partial differential equation based on the state variable can more realistically simulate the degradation process of the engine, improve the fitting ability of the physics-informed features, and make up for the generalization ability of the prediction simulation, thereby improving the prediction accuracy of the degradation rate.
[0031] Among them, in Step 104, the predicted value of the state variable is differentiated with respect to time to obtain the degradation rate. The remaining life of the aero-engine is determined according to the degradation rate.
[0032] In an exemplary embodiment, Step 101 specifically includes: obtaining the aero-engine data set; using the locally weighted regression (Lowess) feature extraction algorithm to extract the feature data from the sample data in the aero-engine data set, and the extracted feature data constitutes the aero-engine training set.
[0033] The locally weighted regression feature extraction algorithm analyzes the dataset and easily finds that the volatility of almost all sensor data is very strong, and such data input is quite difficult for a multi-layer perceptron to fit. Among the commonly used data smoothing algorithms, locally weighted regression can better solve the problem of data volatility and still have a good fitting effect in the case of many outliers. The implementation idea of locally weighted regression is as follows: for any data point in the sample, take this data point as the center, intercept a section of data with a length of len before and after it, and use the weight function W to perform weighted linear regression on the data. The central value of this regression line is (x, ŷ), where ŷ is the corresponding value of the fitted curve and x is the value of the data point. For an object with N data points, N weighted regression lines can be made, and the connection of the central values of each regression line is the locally weighted regression curve of this data object. For the data points at both ends, the data points at both ends and the data points with the closest data of length len use the same sequence data for weighted linear regression.
[0034] Steps for implementing the locally weighted regression code: 1. Calculate the regression weights for each data point in the span, where the span is the sequence data with a length of len. The weights are given by the following tricubic function where x is the predicted value associated with the response value to be smoothed, x i is the nearest neighbor of x defined by the span, d(x) is the distance along the abscissa from x to the farthest predicted value within the span, and w i is the weight of x i . The weights have the following characteristics: the data point to be smoothed has the largest weight and has the greatest impact on the fitting; the weights of the data points outside the span are zero and have no impact on the fitting. 2. Perform weighted linear least squares regression. For locally weighted regression, the regression uses a first-degree polynomial. 3. The smoothed value is given by the weighted regression at the predicted value of the corresponding point.
[0035] Both the first multi-layer perceptron and the second multi-layer perceptron are neural networks including at least one hidden layer and with a depth of at least two layers. Both the first multi-layer perceptron and the second multi-layer perceptron introduce one or more hidden layers between the input layer and the output layer of a single-layer neural network. The neurons therein are called hidden units, and its output layer is the hidden layer variable, as Figure 2 shown, where h1, h2, h3, h4, and h5 are all hidden layer variables, and o1, o2, and o3 are all output layer variables.
[0036] The aeroengine is a turbofan engine. For a turbofan engine, there are many data values that affect the operating state, which makes it very difficult to construct a complete model with high transparency. For the commonly used CMAPSS dataset of aeroengines, some of the sensor data it provides cannot accurately construct a turbofan engine state degradation model based on physical equations. Therefore, using a neural network to predict the engine state degradation becomes a more practical method.
[0037] The degradation state of the aeroengine can be reflected in the data of multiple sensors, such as the rotational speed matching of the inner and outer rotors, the outlet pressure of the high-pressure compressor, etc. These usually contain potential data describing the engine degradation trajectory. If only the data of a single sensor is used to estimate the degradation state, such as the outlet temperature of the high-pressure compressor, then for the entire engine, the state data in the turbine and combustion chamber is discarded. As is well known, the states at positions such as the turbine, combustion chamber, and nozzle also profoundly affect the engine state. Therefore, when using a neural network to predict the aeroengine state, a function equation with multi-variable input and state-variable output should be established: T = f(t, x).
[0038] Among them, T is the state variable, and f(t, x) represents the first multi-layer perceptron. t is the time, and x represents the feature set of each eigenvalue at time t. The features in the feature set include the rotational speed matching degree of the inner and outer rotors, the matching degree of the total temperature at the outlet of the high-pressure compressor and the outlet of the low-pressure compressor, the total temperature at the outlet of the low-pressure compressor, the total temperature at the outlet of the high-pressure compressor, the total temperature at the outlet of the low-pressure turbine, the total pressure at the outlet of the high-pressure compressor, the physical rotational speed of the fan, the physical rotational speed of the core engine, the static pressure at the outlet of the high-pressure compressor, the ratio of the fuel flow rate to the static pressure at the outlet of the high-pressure compressor, the corrected rotational speed of the fan, the corrected rotational speed of the core engine, the bypass ratio, the extraction steam enthalpy, the cold air flow rate of the high-pressure turbine, and the cold air flow rate of the low-pressure turbine, a total of 16 parameter variables.
[0039] For a typical degradation process, the degradation rate can be set to follow Equation X: Here, g is a non-linear equation used to predict the instantaneous degradation rate after receiving t, x, and T. And Equation X is an explicit partial differential equation.
[0040] Since it is difficult to obtain the explicit equation of g, this application needs to approximate the input and output of g by other methods, or obtain the explicit equation by means of decomposing g. This application uses a deep hidden physical model to fit the internal information of g, so as to realize the approximate fitting of this multi-input and single-output partial differential equation by designing a neural network.
[0041] The multi-input and single-output partial differential equation is:
[0042] In Equation X, the state variable T is differentiated with respect to the time dimension t Derivative of the state variable with respect to the feature set x Take and as inputs to the neural network. Among them, the and of the input g are obtained by using automatic differentiation for the equation of the previous prediction T, and derivatives of different orders are input into the neural network so that the network can better learn the degradation process. However, considering the issues of data volume and computing power at the same time, since there are only 17 dimensions, only the first-order derivative is considered here, and the second-order and higher-order derivatives are not considered.
[0043] From the above description, a physical information state equation can be defined for the aero-engine:
[0044]
[0045] Among them, P(t,x) is the physical information state equation, and in equation X represents the derivative of the state prediction value with respect to the time dimension t obtained by using automatic differentiation when the neural network predicts the state value. is the partial differential equation represented by equation X, and here a neural network is still used for approximate fitting.
[0046] For the state variable T, since the data is discrete data, the discrete parameter k is used to replace t in equation X, and then equation X is used to define it:
[0047] RUL total represents the total length of the current state sequence data of the current engine, and also represents the total remaining useful life prediction (Remaining Useful Life, RUL) of the engine since the start of recording. k represents the sequence value of the time series data, that is, the position of the corresponding state in the sequence. At this time, for this physical information neural network, its corresponding partial differential equation loss function can be defined by equation X:
[0048] In the equation, n represents the number of sequences, and k is the sequence value of the corresponding data of a single sequence. Similarly, the error loss of the data also needs to be defined, which is defined by equation X:
[0049] Since the data is normalized during processing, it is necessary to reduce the possibility that the predicted value is greater than the specified range [0,1], so the boundary loss is defined:
[0050] The second multi-layer perceptron is expressed as:
[0051] Among them, represents the degradation rate at time t, x represents the feature set at time t, and T represents the state variable (label data) at time t. represents the predicted value of the state variable at time t.
[0052] During the training process of the physical information neural network, the loss function includes partial differential equation loss, data loss, and boundary loss.
[0053] The partial differential equation loss is expressed as:
[0054] The data loss is expressed as:
[0055] The boundary loss is expressed as:
[0056] Among them, L PDE is the partial differential equation loss, L data is the data loss, L edge is the boundary loss, P(k j , x j ) is the physical information state equation, k j represents the discrete time of the j-th sequence, x j represents the feature set of the j-th sequence, n represents the number of sequences, T j represents the state variable of the j-th sequence. represents T j 's predicted value, and λ is a constant. λ is set to 1000 so that the predicted value of the neural network will not appear in the range of [0, 1] under normal circumstances. The partial differential equation loss is the loss between the degradation rate output by the second multi-layer perceptron and the predicted value of the degradation rate.
[0057] In step 103, the feature set to be predicted is the feature set of the target aero-engine. The feature set of the target aero-engine is input into the aero-engine state prediction model, and the predicted value of the state variable of the target aero-engine is output.
[0058] The model proposed in this application is based on the MLP network and has achieved good results under the guidance of feature extraction and the physical information loss function. Among them, in the first dataset (FD001), the model proposed in this application has less prediction error, second only to the state-of-the-art model. The reason is that the proposed model involves two neural networks and the loss of relevant physical information equations, combined with the first-order derivative of the network prediction value with respect to time, which can better capture the degradation relationship between data. In the fourth dataset (FD004), its Mean-Square Error (MSE) is relatively large, which is related to the internal physical information of the data. Through the analysis of the dataset, it can be seen that the first three datasets only represent the degradation process of a single fault mode, while the fourth dataset contains two fault states, and the complexity of its data is higher than that of the first two datasets. Therefore, it is understandable that the accuracy decreases. This can be verified by the training results of other models on the fourth dataset.
[0059] Through the error analysis in different processes, it is found that the prediction effect of the model proposed in this application is relatively average in the initial stage of engine degradation, but the data in the middle and late stages of degradation is better than that of the Dual-task Long Short-Term Memory (Dual-task LSTM) model. In the actual operation process of the engine, the middle and late stages of degradation are the key objects of maintenance work, and the accuracy of degradation prediction has important guiding significance for the use of the engine. In addition to MSE, the second degradation prediction evaluation criterion proposed by the Dual-task LSTM model is also used, that is, the prediction accuracy rate of the degradation stage. In this indicator, the model proposed in this application has a ten-percentage-point higher accuracy rate than the Dual-task LSTM model. This shows that although the aero-engine state prediction model in this application has defects in accurately predicting the state value, in the degradation stage prediction with a certain degree of tolerance, the model in this application has achieved very excellent prediction results. At the same time, similar to the MSE prediction, due to the data complexity of FD004, the model in this application also has a slight decrease in the prediction accuracy rate of the degradation stage.
[0060] Compared with conventional neural network models, the model proposed in this application has a higher improvement effect, and there is a great improvement compared with the benchmark model MLP, but there is still a certain gap with the Dual-task LSTM model. As shown in Table 1.
[0061] Table 1 Comparison results of the model proposed in this application and existing models
[0062]
[0063]
[0064] Although it is lacking in the overall error, the model of this application has better accuracy in the later stage of prediction, as shown in Table 2. RUL_N represents the prediction error when the remaining life is predicted to be N. The value range of N is [5, 10, 20, 50, 80, 120, all], and N is all, which represents the initial life of the aircraft engine.
[0065] Table 2 Comparison of prediction errors of this application model and the Dual-task LSTM model on specific data sequences
[0066]
[0067] In order to better utilize the historical characteristics of aircraft engines in physical information, this application designs an aircraft engine stability margin and local weighted regression feature extraction algorithm, so that a single data can not only show the status at this moment, but also measure the stability of the engine operation before and after, getting rid of the difficulty of data being unable to provide historical data for the neural network.
[0068] To better track engine degradation, this application uses a neural network to simulate the nonlinear partial differential equation of engine degradation based on the state degradation equation. The PDE equation simulated by the neural network is used to calculate losses, guiding the training and regression of the engine state prediction neural network, achieving good prediction results.
[0069] This application compares the training effects of the engine state prediction neural network under the guidance of three loss functions, and proves that the three loss functions have different guiding training effects in the physical information neural network architecture, and all have positive guiding training significance.
[0070] The present application was tested on the general aviation engine dataset CMAPSS and the test results proved that the present application method has certain advancement and good degradation stage accuracy.
[0071] Based on the same inventive concept, embodiments of the present application also provide an aircraft engine state and degradation rate prediction device for implementing the aforementioned aircraft engine state and degradation rate prediction method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the aircraft engine state and degradation rate prediction device provided below can be found in the above-mentioned limitations of the aircraft engine state and degradation rate prediction method and will not be further elaborated here.
[0072] In an exemplary embodiment, the present application provides an aircraft engine state and degradation rate prediction device comprising:
[0073] A training set acquisition module for acquiring an aero-engine training set; each sample in the aero-engine training set includes input data and label data, the input data is a feature set of the aero-engine, the label data is a state variable corresponding to the input data, the state variable represents the performance of the aero-engine within a preset usage period, and the smaller the value of the state variable, the worse the performance.
[0074] A prediction model training module for training a physics-informed neural network using the aero-engine training set to obtain an aero-engine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence, the first multi-layer perceptron is used to output a predicted value of the state variable of the aero-engine according to the feature set of the aero-engine; the second multi-layer perceptron is used to simulate the degradation process of the aero-engine according to the feature set of the aero-engine and the partial differential equation of the state variable, and output the degradation rate of the aero-engine; during the training process of the physics-informed neural network, the degradation rate output by the second multi-layer perceptron is used to optimize the first multi-layer perceptron, and the trained first multi-layer perceptron is used as the aero-engine state prediction model.
[0075] A state variable prediction module for inputting the feature set to be predicted into the aero-engine state prediction model and outputting a predicted value of the state variable.
[0076] A remaining life calculation module for calculating the remaining life of the aero-engine according to the predicted value of the state variable.
[0077] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.
[0078] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.
[0079] Specific examples are used in this article to elaborate on the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.
Claims
1. A method for predicting the state and degradation rate of an aero-engine, characterized in that, The method for predicting the state and degradation rate of an aeroengine includes: Obtaining an aeroengine training set; each sample in the aeroengine training set includes input data and label data, the input data is a feature set of the aeroengine, the label data is the state variable corresponding to the input data, the state variable represents the performance of the aeroengine within a preset usage period, and the smaller the value of the state variable, the worse the performance; Training a physics-informed neural network using the aeroengine training set to obtain an aeroengine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence, the first multi-layer perceptron is used to output a predicted value of the state variable of the aeroengine according to the feature set of the aeroengine; the second multi-layer perceptron is used to simulate the degradation process of the aeroengine according to the feature set of the aeroengine and the partial differential equation of the state variable, and output the degradation rate of the aeroengine; during the training process of the physics-informed neural network, the degradation rate output by the second multi-layer perceptron is used to optimize the first multi-layer perceptron, and the trained first multi-layer perceptron is used as the aeroengine state prediction model; Inputting the feature set to be predicted into the aeroengine state prediction model, and outputting a predicted value of the state variable; Calculating the remaining life of the aeroengine according to the predicted value of the state variable.
2. The method for predicting the state and degradation rate of an aero-engine according to claim 1, characterized in that During the training process of the physics-informed neural network, using the degradation rate output by the second multi-layer perceptron to optimize the first multi-layer perceptron, specifically including: During the training process of the physics-informed neural network, calculating a predicted value of the degradation rate according to the predicted value of the state variable output by the first multi-layer perceptron; Optimizing the first multi-layer perceptron according to the loss between the degradation rate output by the second multi-layer perceptron and the predicted value of the degradation rate.
3. The method for predicting the state and degradation rate of an aeroengine according to claim 2, wherein The loss function during the training process of the physics-informed neural network includes partial differential equation loss, data loss, and boundary loss; The partial differential equation loss is expressed as: The data loss is expressed as: The boundary loss is expressed as: Among them, L PDE is the partial differential equation loss, L data is the data loss, L edge is the boundary loss, P(k j , x j ) is the physical information state equation, k j represents the discrete moment of the j-th sequence, x j represents the feature set of the j-th sequence, n represents the number of sequences, T j represents the state variable of the j-th sequence, represents the predicted value of T j , is the predicted value of the state variable, λ is a constant, and the partial differential equation loss is the loss between the degradation rate output by the second multi-layer perceptron and the predicted value of the degradation rate.
4. The method for predicting the state and degradation rate of an aero-engine according to claim 1, wherein Obtaining the aeroengine training set, specifically including: Obtaining an aeroengine data set; Using a locally weighted regression feature extraction algorithm to extract feature data from the sample data in the aeroengine data set, and the extracted feature data constitutes the aeroengine training set.
5. The method for predicting the state and degradation rate of an aero-engine according to claim 1, wherein Both the first multi-layer perceptron and the second multi-layer perceptron are neural networks including at least one hidden layer and having a depth of at least two layers.
6. The method for predicting the state and degradation rate of an aero-engine according to claim 1, wherein The aeroengine is a turbofan engine.
7. The method for predicting the state and degradation rate of an aeroengine according to claim 1, characterized in that The features in the feature set include the matching degree of the internal and external rotor speeds, the matching degree of the total temperature at the outlet of the high-pressure compressor and the outlet of the low-pressure compressor, the total temperature at the outlet of the low-pressure compressor, the total temperature at the outlet of the high-pressure compressor, the total temperature at the outlet of the low-pressure turbine, the total pressure at the outlet of the high-pressure compressor, the physical speed of the fan, the physical speed of the core engine, the static pressure at the outlet of the high-pressure compressor, the ratio of the fuel flow to the static pressure at the outlet of the high-pressure compressor, the fan corrected speed, the core engine corrected speed, the bypass ratio, the extraction steam enthalpy, the cooling air flow of the high-pressure turbine, and the cooling air flow of the low-pressure turbine.
8. An aero-engine state and degradation rate prediction device, characterized in that The device for predicting the state and degradation rate of the aeroengine includes: A training set acquisition module for acquiring an aero-engine training set; each sample in the aero-engine training set includes input data and label data, the input data being a feature set of the aero-engine, and the label data being a state variable corresponding to the input data, the state variable representing the performance of the aero-engine within a preset usage period, and the smaller the value of the state variable, the worse the performance; A prediction model training module for training a physics-informed neural network using the aero-engine training set to obtain an aero-engine state prediction model; the physics-informed neural network includes a first multi-layer perceptron and a second multi-layer perceptron connected in sequence, the first multi-layer perceptron being used to output a predicted value of the state variable of the aero-engine according to the feature set of the aero-engine; the second multi-layer perceptron being used to simulate the degradation process of the aero-engine according to the feature set of the aero-engine and the partial differential equation of the state variable, and output the degradation rate of the aero-engine; during the training process of the physics-informed neural network, the degradation rate output by the second multi-layer perceptron is used to optimize the first multi-layer perceptron, and the trained first multi-layer perceptron is used as the aero-engine state prediction model; A state variable prediction module for inputting the feature set to be predicted into the aero-engine state prediction model and outputting a predicted value of the state variable; A remaining life calculation module for calculating the remaining life of the aero-engine according to the predicted value of the state variable.