Aircraft engine residual life prediction method under multiple working conditions

Abstract

The application relates to a kind of methods for predicting the residual life of aircraft engine under multiple working conditions based on adaptive interactive learning, comprising: obtaining the multivariate sensor data of aircraft engine and carrying out signal preprocessing and signal selection; adaptive interactive learning model is constructed and trained using historical multivariate sensor data, the adaptive interactive learning model is based on interactive recurrent neural network construction, including interconnected fusion model and failure process model, while realizing the construction of health index HI of engine and failure process modeling by fusing multivariate sensor data; the preprocessed multivariate sensor data is input into the trained adaptive interactive learning model, and the health index HI of aircraft engine is obtained; based on the health index HI of aircraft engine and the parameters of adaptive interactive learning model, the distribution result of the residual service life of aircraft engine is obtained, and the residual life prediction of aircraft engine is realized. Compared with the prior art, the application has the advantages of high prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aircraft engine failure process modeling and remaining life prediction, and particularly relates to a remaining life prediction method for aircraft engines under multiple working conditions based on adaptive interactive learning. BACKGROUND

[0002] As the "heart" of an aircraft, the reliability and stability of an aircraft engine are important guarantees for the safety of aircraft flight. In the process of engine operation, online state monitoring and prediction of the remaining useful life (RUL) are performed, and a reasonable maintenance plan is arranged according to the prediction results, so that sufficient life margin is left to ensure the safety of flight and the economy of maintenance.

[0003] In recent years, aircraft engine failure process modeling and RUL prediction based on deep learning are one of the focuses and hotspots of domestic and foreign enterprises and institutions. This kind of method can be divided into two categories. The first category takes the measured values or extracted features of multi-sensor data as the input of the neural network model, and takes the RUL as the output of the model. This kind of model includes deep belief network (DBN), convolutional neural network (CNN) and recurrent neural network (RNN). Although the neural network designed by this kind of method can directly predict the RUL, it does not consider the degradation mechanism of the aircraft engine, and the interpretability is poor. The second category of model first fuses multi-sensor data to construct a health index (HI), and then models the degradation process of the aircraft engine according to the constructed health index. The construction of the health index can be completed by a deep learning model, such as an artificial neural network, a long short-term memory network (LSTM), etc. This kind of model can visualize the engine failure process and monitor the real-time health status of the running aircraft engine. However, this kind of method still has the following three problems.

[0004] First, existing methods assume that the aircraft engine operates in a single working condition, while in actual operation, the pressure, temperature and rotation speed in the working environment are constantly changing, and the statistical characteristics of the collected sensor data will change with the working environment, which is not conducive to accurately estimating the actual failure state. Second, most existing neural network models based on HI separate HI construction and failure process modeling into two independent parts, ignoring the inherent relationship between the two parts. In addition, the generalization ability of the fusion model and the degradation model is also important. The fusion model not only needs to effectively fuse multi-sensor data to construct HI, but also needs to fuse the corresponding degradation state from the multi-sensor signal. Similarly, the degradation model cannot only model the degradation process based on HI, but also needs to model the degradation process based on each sensor signal. Third, compared with supervised learning in deep learning, the construction of HI is more challenging because the actual health state of the aircraft engine cannot be directly observed from the sensor data. SUMMARY

[0005] The purpose of the present application is to provide a multi-working-condition aircraft engine residual life prediction method based on adaptive interactive learning, which fully considers the influence of multi-working-condition on the degradation state of the aircraft engine, uses the multi-sensor signal data collected during the operation of the aircraft engine, realizes the failure process modeling and residual life prediction of the aircraft engine under multi-working-condition, and improves the prediction accuracy.

[0006] The purpose of the present application can be realized by the following technical solutions:

[0007] A multi-working-condition aircraft engine residual life prediction method based on adaptive interactive learning, comprising the following steps:

[0008] Obtain multi-sensor data of the aircraft engine and perform signal preprocessing and signal selection;

[0009] Construct an adaptive interactive learning model and train it using historical multi-sensor data, wherein the adaptive interactive learning model is constructed based on an interactive recurrent neural network and includes a fusion model and a failure process model connected to each other, and simultaneously realizes the fusion of multi-sensor data to construct the health index HI of the engine and the failure process modeling;

[0010] Input the multi-sensor data into the trained adaptive interactive learning model to obtain the health index HI of the aircraft engine;

[0011] Based on the health index HI of the aircraft engine and the parameters of the adaptive interactive learning model, obtain the distribution result of the residual service life of the aircraft engine, and realize the residual life prediction of the aircraft engine.

[0012] The health index HI of the engine y l(t) is defined as the potential failure state of the engine plus a corresponding white noise term, whose expression is:

[0013]

[0014] wherein, is a white noise term, is the potential failure state corresponding to the health index HI, is synthetically described by a common feature function and individual random effect parameters, whose expression is:

[0015]

[0016] wherein ψ(t) is a common feature function describing the evolution of failure states of all aircraft engines, is an individual random effect parameter describing the engine l;

[0017] multivariate sensor data signal x l (t) is defined as its corresponding failure state g l (t) plus a corresponding white noise term ε l (t), whose expression is:

[0018] x l (t) = g l (t) + ε l (t)

[0019] The failure process model is also used to describe the evolution of failure states of the sensor data, whose expression is:

[0020] g l (t) = ψ(t)Γ l

[0021] wherein, represents the individual random feature of the multivariate sensor data of the engine l.

[0022] The ψ(·) function used by the failure process model is the hidden layer value of a long short-term memory network with time t as input.

[0023] The η(·) function used by the fusion model is a recurrent neural network, set as the parameter set of the RNN, and the expression of the RNN is simplified as y t = η(x t , ω);

[0024] The fusion model is used to fuse the multivariate sensor data to construct the health index HI, and to fuse the failure states corresponding to the multivariate sensor data to construct the fused failure state.

[0025] The expression of the HI is as follows:

[0026] y l (t) = η(x l (t), ω)

[0027] where η(·) is the function form of RNN, x l (t) is the multi-sensor signal data of engine l at time t, and ω is the RNN network parameter.

[0028] The fusion failure state is constructed as follows:

[0029]

[0030] where g l (t) is the failure state corresponding to the multi-sensor data signal.

[0031] The total loss function calculation expression of the adaptive interactive learning model is as follows:

[0032]

[0033] where λ1, λ2, λ3, and λ4 are weight coefficients, which determine the influence of each attribute on the constructed HI; The loss function corresponding to the optimized HI characteristics includes the variance of the HI at the failure time, the monotonicity, and the curvature; The loss function for minimizing the domain bias caused by multiple working conditions by using domain adaptation; Corresponding to the fitting error between the degradation state and the sensor signal, respectively, and the health index HI.

[0034] The loss function for optimizing the characteristics of the HI is:

[0035]

[0036]

[0037]

[0038] where y l (t) is the health index HI of the engine, is the failure time of the engine l, L is the number of aircraft engines in the historical data, m is a preconfigured failure threshold, τ1 is the size of the moving window corresponding to the monotonicity, d l (t) = y l (t-τ1) - y l (t), and τ2 is the size of the moving window corresponding to the optimized curvature.

[0039] The loss function for minimizing the domain shift caused by the field adaptation multi-conditions is:

[0040]

[0041] In the formula, is the condition k at time t t is the number of observation points in condition i matrix, is the condition k at time t t is the number of observation points in condition j matrix, P is the number of condition types, L is the number of aircraft engines in the historical data, d(·,·) is a measure function of the domain shift, and the CORAL algorithm is used, and its calculation expression is:

[0042]

[0043]

[0044]

[0045] where n i is the number of observation points in condition i, n j is the number of observation points in condition j, 1 T is a column vector of all 1s, d is the dimension of the feature layer, is the Frobenius norm.

[0046] The loss function for minimizing the fitting error between the degradation state and the sensor signal, the health index HI, is:

[0047]

[0048]

[0049] where, is the failure time of engine l, n l is the number of engine sensor data observation points, L is the number of aircraft engines in the historical data, is the fused failure state, y l (t) is the health index HI of the engine, is the latent failure state corresponding to the health index HI, g l (t) is the failure state corresponding to the multi-sensor data signal, x l (t) is the multi-sensor data signal.

[0050] The training of the adaptive interactive learning model is based on the loss function to iteratively train the interactive RNN network to estimate the model parameters, specifically including the following steps:

[0051] Step 1: Initialize the health index HIy of engine l in historical data as and set k = 1;

[0052] Step 2: In the kth iteration, given the sensor data x l (t) and the health index HIy l (t), train the adaptive interactive learning model, and the loss function is simplified as:

[0053]

[0054] to update the degradation parameters Γ l , and ψ(t), and obtain the corresponding degradation state g l (t) and

[0055] Step 3: Given the sensor data x l (t) and the corresponding degradation state g l (t), train the adaptive interactive learning model, and the loss function is simplified as:

[0056]

[0057] to update the fusion model parameters ω, and fuse the multi-sensor data and the corresponding degradation state to update HIy l (t) and the fusion degradation state

[0058] Step 4: Calculate the distance between and , i.e.:

[0059]

[0060] Based on the distance between and , judge whether the convergence is reached, when the distance between and is close to zero, i.e. ∈ is the pre-configured iteration threshold, then stop iteration; otherwise, repeat steps 2-4.

[0061] The calculation expression of the distribution result of the remaining useful life of the aircraft engine is:

[0062]

[0063]

[0064] wherein, ξq is the remaining useful life of the in-service aircraft engine q, m is the failure threshold of the aircraft engine, y q is the health index HI of the in-service aircraft engine q, ψ(·) is the in-service aircraft engine q about time t and the common feature function, which is estimated from historical aircraft engine failure data, Γ q is the individual random effect parameter of the in-service aircraft engine q, and its prior distribution is estimated from historical aircraft engine failure data Γ l is estimated, then:

[0065]

[0066] In the formula, Φ(·) is the cumulative distribution function of the standard normal distribution.

[0067] Compared with the prior art, the present application has the following beneficial effects:

[0068] 1) The present application helps to accurately describe the health state and failure process of the aircraft engine under multiple working conditions, and realizes accurate prediction of the remaining life of the aircraft engine, which can guarantee the economy of engine maintenance and the reliability and stability of aircraft flight.

[0069] 2) The present application uses the field adaptive method to reduce the field deviation of the aircraft engine sensor data under multiple working conditions, and reduces the influence of working condition changes on failure state estimation.

[0070] 3) The present application uses an interactive RNN model to simultaneously realize the construction of the health index HI and the modeling of the failure process, which enhances the generalization ability of the model.

[0071] 4) The present application uses fusion of unsupervised learning and supervised learning to realize optimization of multiple characteristics of the health index HI, which helps to accurately estimate the failure state.

[0072] 5) The present application optimizes the characteristics of HI according to the degradation process mechanism, such as failure time, monotonicity of degradation signal, curvature and range, and fuses multiple sensor data to construct HI, which overcomes the problem that the actual health state of the aircraft engine cannot be directly observed from the sensor data, thereby overcoming the problem that HI cannot be constructed. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 is the flow chart of the method of the present application;

[0074] Figure 2 is the sensor failure signal data, health index HI and corresponding failure state of the aircraft engine in the embodiment of the present application, wherein the sensor failure signal data and the constructed health index HI are represented by points, and the corresponding failure state is represented by a line;

[0075] Figure 3 The figure shows the prediction error of the adaptive interactive learning model in different RUL states in the embodiment of the present application, wherein the bar and the corresponding error box represent the average value and the standard error of the prediction error, respectively. DETAILED DESCRIPTION

[0076] The present application will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives a detailed implementation and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0077] The present embodiment provides a method for predicting the remaining useful life of an aircraft engine under multiple working conditions based on adaptive interactive learning, as shown in the figure, comprising the following steps: Figure 1

[0078] 1) Obtain the multi-element sensor data of the aircraft engine and perform signal preprocessing and signal selection.

[0079] The signal preprocessing is to first perform log transformation on the input data and then perform z-score standardization.

[0080] The criterion for sensor signal selection is whether the sensor signal has a clear monotonic increasing or decreasing trend during the entire operation of the engine, and if so, it is selected as the input of the subsequent model.

[0081] The present embodiment takes the failure data set of an aircraft gas turbine engine as an example. The failure data set of the aircraft gas turbine engine used in the present embodiment is generated by a commercial modular aero-propulsion system simulation (C-MAPSS), which is developed by NASA and used to simulate the failure process of an aircraft gas turbine engine. Ten kinds of sensor failure signals in the data set are selected to monitor the failure state of the aircraft engine under six working conditions and a single fault model. The data set contains 260 historical observation data of aircraft engines (i.e., L = 260), specifically including 53759 observation data (i.e., N = 53759). The historical observation data of the 260 aircraft engines includes data from the start of operation to the occurrence of failure of the aircraft engine, while the data of the 259 aircraft engines in use includes data from the start of operation to a certain time point of the aircraft engine, as well as the corresponding RUL true data. The present embodiment uses the historical observation data of the 260 aircraft engines to train the model, uses the data of the 259 aircraft engines in use to predict the RUL, and verifies the performance of the model of the present application by comparing the predicted RUL value with the corresponding RUL true data.

[0082] ​​2) Constructing an adaptive interactive learning model based on interactive recurrent neural network (RNN) and training it with historical multi-sensor data, which includes a fusion model and a failure process model connected with each other, and realizing the fusion of multi-sensor data to construct the health index HI of the engine and the failure process modeling.

[0083] 21) Sensor data failure process model of aircraft engine

[0084] Data x collected from each sensor of the aircraft engine l (t) contains relevant information about its failure state, and by removing the measurement noise ε l (t), the failure state corresponding to the sensor signal can be extracted, that is:

[0085] x l (t) = g l (t) + ε l (t) (1)

[0086] In the formula, represents the data vector collected from M sensors of unit l at each time point t, represents the failure state corresponding to the sensor data, assuming that the noise term of the sensor data obeys Gaussian distribution

[0087] The evolution of the failure state corresponding to the sensor data over time can be described by a failure process model, considering that the initial failure state and failure rate of each engine are different due to different initial states, loads and working conditions, and a common characteristic function and individual random effect parameters should be used for comprehensive description, that is:

[0088] g l (t) = ψ(t)Γ l (2)

[0089] In the formula, ψ(t) is a common characteristic function describing the evolution of the failure state of all aircraft engines, represents the individual random characteristics of the multi-sensor data of engine l.

[0090] 22) Health index HI of aircraft engine and its failure process model

[0091] The health index HI of the aircraft engine reflects the potential failure state of its evolution over time. The health index HI y l (t) of the lth engine at time t is defined as its potential failure state plus the corresponding white noise term that is:

[0092]

[0093] wherein, is a white noise term, is the potential failure state corresponding to the health index HI. It is assumed that each aircraft engine is independent of each other, and the noise term of the aircraft engine l at time t obeys a Gaussian distribution

[0094] is synthetically described by the common feature function and individual random effect parameters, and the calculation expression is:

[0095]

[0096] wherein, ψ(t) is the common feature function describing the evolution of all aircraft engine failure states, is the individual random effect parameter describing the engine l.

[0097] 23) An adaptive interactive learning model based on recurrent neural network RNN

[0098] The η(·) function adopted by the fusion model is a recurrent neural network, and is set is the parameter set of the RNN, and the expression of the RNN is simplified as y t = η(x t , ω).

[0099] The fusion model is used to fuse multi-sensor data to construct a health index HI, and fuse the failure states corresponding to the multi-sensor data to construct a fusion failure state, wherein,

[0100] The construction expression of the fusion multi-sensor data to construct the health index HI is as follows:

[0101] y l (t) = η(x l (t), ω) (5)

[0102] wherein, η(·) is the function form of the RNN, x l (t) is the multi-sensor signal data of the engine l at time t, and ω is the RNN network parameter;

[0103] The construction of the fusion failure state is as follows:

[0104]

[0105] wherein, g l (t) is the failure state corresponding to the multi-sensor data signal.

[0106] The ψ(·) function adopted by the failure process model is the hidden layer value of a long short-term memory network (LSTM) with time t as input, that is, the input of the long short-term memory neural network is time t, and the value of the hidden layer is ψ(t).

[0107] 24) Loss function

[0108] Since the actual degradation state cannot be directly obtained from sensor data, the construction of HI under multiple working conditions is a comprehensive problem that combines supervised and unsupervised learning. In order to effectively construct HI, the present application considers three aspects when designing the loss function. First, optimize the characteristics of HI, including: variance of HI at failure time, monotonicity and curvature; second, minimize the domain shift caused by multiple working conditions using domain adaptive, so that the degradation model is not affected by multiple working conditions; third, minimize the fitting error between degradation state and sensor signal, health index HI.

[0109] 241) Loss function for optimizing the characteristics of HI

[0110] A, minimize the variance of HI at failure time

[0111] Failure time τ of the aircraft engine l defined as the degradation state of HI the time exceeding the pre-set failure threshold m, that is Therefore, the HI at the failure time should also be around the failure threshold m, that is, the variance of the HI at the failure time of the aircraft engine should be minimized to accurately infer the actual degradation state. Therefore, its loss function is set as:

[0112]

[0113] wherein, is the failure time of engine l; L is the number of aircraft engines in the historical data, and the sensor data of these engines is collected before failure occurs.

[0114] B, monotonicity of HI

[0115] If the failure model HI reaches the failure threshold at a certain time in the future, then the trend of HI should be monotonic. In order to maximize the monotonicity of HI, a reward mechanism is proposed to maximize the HI increment of each time interval (t-1, t):

[0116]

[0117] This mechanism helps to maximize the range of HI, that is

[0118] However, it should be noted that due to measurement noise, the increment of HI between adjacent time steps is not always positive. To address this issue, this invention employs a moving window method to relax the constraints. The moving window is defined as containing y l (t) is the time series at τ consecutive time points, i.e., y l (t-τ+1),y l (t-τ+2),…,y l (t). Then, assume the y-coordinates of two adjacent moving windows are... l The sum of (t) is monotonic, that is It should be a positive number. Therefore, based on maximizing the monotonicity of HI using a moving window, it is set as follows:

[0119]

[0120] Where τ1 is the size of the moving window.

[0121] Furthermore, to avoid cancellation caused by linear addition, the model employs an exponential function to maximize monotonicity, and its loss function is... as follows:

[0122]

[0123] The curvature of C and HI

[0124] The degradation rate of an aircraft engine increases with time t, so the increment of HI also increases with time, similar to the monotonicity of HI. A moving window is used to optimize the curvature characteristics of HI, and the corresponding loss function is... for:

[0125]

[0126] d l (t)=y l (t-τ1)-y l (t) is the increment of HI in the time interval (t-τ1,t), and τ2 is the size of the moving window corresponding to the optimized curvature.

[0127] 242) Employ a loss function that adaptively minimizes the domain offset caused by multiple operating conditions.

[0128] The model of this invention employs domain adaptation to minimize the neighborhood shift of HI under multiple operating conditions. Existing neighborhood adaptation methods often use higher layers of neural networks to reduce neighborhood shift; therefore, the model of this invention uses neighborhood adaptation to minimize the last hidden layer of the RNN. The domain offset, i.e.:

[0129]

[0130] wherein, is the working condition k at time t t is the i-th matrix, is the working condition k at time t t is the j-th matrix, p is the number of working condition types, L is the number of aircraft engines in the historical data, d(·,·) is a measure function of field deviation, which adopts the CORAL (Correlation Alignment) algorithm, and the calculation expression is:

[0131]

[0132] wherein, n i is the number of observation points under working condition i, n j is the number of observation points under working condition j, 1 T is a column vector with all 1s, d is the dimension of the feature layer, is the Frobenius norm.

[0133] 243) minimizing the loss function of fitting error between the degradation state and the sensor signal, the health index HI

[0134] The fusion model can fuse not only the multi-element sensor data but also the degradation state corresponding to the sensor, and the fusion result is is the actual degradation state estimated by the fusion model, which should be close to the value of HI, so the loss function is:

[0135]

[0136] The degradation model can simultaneously fit the sensor data and the HI obtained by the fusion model, and the corresponding loss function is:

[0137]

[0138] wherein, is the failure time of engine l, n l is the number of observation points of engine sensor data.

[0139] In summary, the total loss function calculation expression of the adaptive interactive learning model is as follows:

[0140]

[0141] wherein, λ1, λ2, λ3 and λ4 are weight coefficients, which determine the influence of each attribute on the constructed HI.

[0142] 25) parameter estimation

[0143] The present application proposes an interactive training parameter estimation algorithm, and the pseudo code in Table 1 illustrates the calculation process of the algorithm. In the k = 1, 2, 3, … iterations, the adaptive interactive AAIL model is trained based on the simplified loss function respectively.

[0144] Table 1 Pseudo code of interactive training algorithm

[0145]

[0146] Specifically, the steps of training the interactive RNN network based on the loss function to estimate the model parameters are as follows:

[0147] Step 1: initialize the health index HI of the engine l in the historical data as and set k = 1, where, obtained from the method in the existing literature;

[0148] Step 2: in the kth iteration, given the sensor data x l (t) and the health index HIy l (t), the adaptive interactive learning model is trained, and the loss function is simplified as:

[0149]

[0150] to update the degradation parameters Γ l , and ψ(t), and according to the multi-sensor data and the health index HI, the corresponding degradation states g l (t) and

[0151] Step 3: given the sensor data x l (t) and the corresponding degradation state g l (t), the adaptive interactive learning model is trained, and the loss function is simplified as:

[0152]

[0153] to update the fusion model parameters ω, and to fuse the multi-sensor data and the corresponding degradation state, respectively update HIy l (t) and the fusion degradation state

[0154] Step 4: calculate the distance between and , that is:

[0155]

[0156] based on and The distance between them determines whether convergence has been achieved. and The distance between them is close to zero, that is Then stop the iteration, where ∈ is the iteration threshold, which is set to a relatively small positive value in this embodiment; otherwise, repeat steps 2-4.

[0157] This embodiment uses cross-validation to determine the hyperparameters and tuning parameters in the model, and uses historical observation data from 260 aircraft engines to estimate the AAIL model parameters ψ(t),Γ. l , And ω. Then, the online sensor data of each aircraft engine is fused to construct the corresponding HI and obtain its failure state distribution.

[0158] 3) Input the multi-sensor data into the trained adaptive interactive learning model to obtain the aircraft engine health index HI.

[0159] Figure 2 This paper presents sensor failure signal data, HI (High-Intensity Detection) values, and corresponding failure states for a certain aircraft engine. The sensor failure signal data and the constructed HI are represented by points, and the corresponding failure states are represented by lines. The results show that the constructed HI reflects a more pronounced failure state trend than any other sensor signal.

[0160] 4) Based on the aircraft engine health index HI and the parameters of the adaptive interactive learning model, the distribution results of the remaining service life of the aircraft engine are obtained, and the remaining service life of the aircraft engine is predicted.

[0161] After training the model using historical data of multi-factor sensor failure signals from L aircraft engines, this invention predicts the RUL (Range Limiting) of an aircraft engine currently in use. The multi-factor sensor failure signals of the aircraft engine q currently in use are considered at time points... All of the preceding values ​​are observable, where n q This represents the number of observable multi-sensor failure signal data for the aircraft engine q currently in use. To predict the RUL of the aircraft engine q currently in use, we first base it on the formula: y l (t)=η(x l (t),ω) yields the time point t q The preceding HI. Since the aircraft engine has not yet failed, given that RUL is a positive number (i.e., ξ... q >0), to obtain its RULξ q The cumulative distribution function is as follows:

[0162]

[0163] in,

[0164]

[0165] m represents the threshold for aircraft engine failure. Given y q , The posterior distribution is Where Γ q The prior distribution Γ q ~G(·) can pass through all Γ l The values ​​of l = 1, ..., L are estimated. Due to the noise term ε... q,t ~N(0,σ 2 Therefore, according to Y q =Ψ q Γ q +ε q get in Assumption The prior distribution is a multivariate Gaussian distribution. but posterior distribution It is also a multivariate Gaussian distribution, i.e., Γ q |Y q ~N(μ) q ,∑ q ),in, but:

[0166]

[0167] In the formula, Φ(·) is the cumulative distribution function of the standard normal distribution.

[0168] RUL prediction error ∈ of aircraft engine q in use q Defined as predicted RUL and RU real data ξ q The absolute difference between them divided by the actual time τ when the failure occurred q :

[0169]

[0170] The smaller the prediction error, the better the model performance. Figure 3 The diagram shows the prediction errors of the model under different RUL states. For each RUL state, the bars and corresponding error bins represent the mean and standard deviation of the prediction error under that RUL state, respectively. For example, "20" represents the mean and standard deviation of the RUL prediction errors for all currently used aircraft engines in a state where the actual RUL data is less than or equal to 20, and "positive infinity" represents the mean and standard deviation of the RUL prediction errors for all currently used aircraft engines. Figure 3As shown, the model of the application can achieve good aircraft engine RUL prediction in each RUL state, and the superiority of the model of the application is more significant when the aircraft engine is close to failure (for example, the RUL real data is less than 60), which has important practical significance for avoiding sudden failure of the aircraft engine.

[0171] The preferred embodiments of the application are described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations without departing from the concept of the application. Therefore, any technical solutions obtained by logical analysis, reasoning or limited experiments according to the concept of the application on the basis of the prior art should be within the protection scope of the claims.

Claims

1. A method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning, characterized in that, Includes the following steps: Acquire multi-sensor data of the aircraft engine and perform signal preprocessing and signal selection; An adaptive interactive learning model is constructed and trained using historical multi-sensor data. The adaptive interactive learning model is based on an interactive recurrent neural network and includes an interconnected fusion model and a failure process model. It also realizes the construction of the engine health index HI and failure process modeling by fusing multi-sensor data. By inputting multi-sensor data into the trained adaptive interactive learning model, the aircraft engine health index HI is obtained. Based on the aircraft engine health index HI and the parameters of the adaptive interactive learning model, the distribution results of the remaining service life of the aircraft engine are obtained, and the remaining service life of the aircraft engine is predicted. The engine's health index HI Defined as the potential failure state of the engine plus the corresponding white noise term, its calculation expression is: In the formula, For white noise, This represents the potential failure state corresponding to the health index HI. Described by a combination of common characteristic functions and individual random effects parameters, its calculation expression is as follows: In the formula, It is a common characteristic function describing the evolution of failure states of all aircraft engines. It describes the engine Individual random effects parameters; Multi-sensor data signals Define its corresponding failure state Add the corresponding white noise term Its expression is: The failure process model is also used to describe the evolution of the failure state of sensor data, and its expression is: In the formula, Represents engine Individual random characteristics of multi-sensor data; The fusion model adopts The function is a recurrent neural network, set as follows: Given the parameter set of an RNN, the expression of an RNN is simplified to... ; The fusion model is used to fuse multi-sensor data to construct a health index HI, and to fuse the failure states corresponding to the multi-sensor data to construct a fused failure state, wherein... The construction expression for HI is as follows: In the formula, It is a functional form of RNN. It is an engine exist Multi-sensor signal data at any given time, These are RNN network parameters; The fusion failure state The structure is as follows: In the formula, This refers to the failure state corresponding to the data signals from multiple sensors.

2. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 1, characterized in that, The failure process model adopted The function takes time as input. Hidden layer values ​​of a Long Short-Term Memory (LSTM) network.

3. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 1, characterized in that, The total loss function of the adaptive interactive learning model is calculated as follows: In the formula, and These are weighting coefficients that determine the impact of each attribute on the constructed HI; , , The loss function corresponding to the characteristics of the optimized HI, wherein the characteristics of the optimized HI include: the variance, monotonicity and curvature of HI at the failure time; The loss function is used to minimize the domain offset caused by multiple operating conditions through domain adaptation; These correspond to minimizing the fitting error between the degraded state and the sensor signal, and the health index HI, respectively.

4. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 3, characterized in that, The loss function for optimizing the characteristics of HI , , for: In the formula, The engine's health index is HI. It is an engine The moment of failure, It refers to the number of aircraft engines in historical data. For the pre-configured failure threshold, It optimizes the size of the moving window corresponding to monotonicity. , It optimizes the size of the moving window corresponding to the curvature.

5. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 3, characterized in that, The loss function that employs domain adaptive minimization to minimize the domain offset caused by multiple operating conditions is described. for: In the formula, yes Working conditions at any time for of matrix, yes Working conditions at any time for of matrix, It refers to the number of operating conditions. It refers to the number of aircraft engines in historical data. This is the metric function for the neighborhood offset, using the CORAL algorithm. Its calculation expression is: in, For working conditions The number of observation points below For working conditions The number of observation points below It is a column vector consisting entirely of 1s. For the dimension of the feature layer, It is the Frobenius norm.

6. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 3, characterized in that, The loss function that minimizes the fitting error between the degradation state and the sensor signal, and the health index HI, is: in, It is an engine The moment of failure, It is the number of engine sensor data observation points. It refers to the number of aircraft engines in historical data. It is a fusion failure state. The engine's health index is HI. This refers to the potential failure state corresponding to the health index HI. It refers to the failure state corresponding to the data signals from multiple sensors. It is a multi-sensor data signal.

7. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 3, characterized in that, The training of the adaptive interactive learning model is based on iteratively training an interactive RNN network using a loss function to estimate model parameters, specifically including the following steps: Step 1: Extract engine data from historical data The Health Index (HI) is initialized as follows: , , and set ; Step 2: In the In each iteration, given sensor data Health Index (HI) The loss function for training the adaptive interactive learning model simplifies to: To update degradation parameters , and Based on multi-sensor data and the Health Index (HI), the corresponding degradation states are obtained. and ; Step 3: Given sensor data and the corresponding degradation state The loss function for training the adaptive interactive learning model simplifies to: To update the fusion model parameters It integrates multi-sensor data and corresponding degradation status to update HI separately. and fusion degradation state ; Step 4: Calculation and The distance between them, that is: based on and The distance between them determines whether convergence has been achieved. and The distance between them is close to zero, that is , If the iteration threshold is the pre-configured threshold, then stop the iteration; otherwise, repeat steps 2-4.

8. The method for predicting the remaining life of an aircraft engine under multiple operating conditions based on adaptive interactive learning according to claim 1, characterized in that, The calculation expression for the distribution of the remaining service life of the aircraft engine is as follows: In the formula, For aircraft engines in use The remaining service life, This refers to the failure threshold of the aircraft engine. For aircraft engines in use The health index HI, For aircraft engines in use Regarding time The common characteristic function is estimated from historical aircraft engine failure data. For aircraft engines in use The individual random effects parameter, whose prior distribution is derived from historical aircraft engine failure data. If we estimate that: In the formula, It is the cumulative distribution function of the standard normal distribution.

Images