A method for aircraft engine failure mode recognition and life prediction

Abstract

This invention relates to a method for aircraft engine failure mode identification and life prediction based on a multi-modal deep neural network (DNN). The method includes: constructing a DNN-based failure mode identification and life prediction model, using sensor measurements as input, and outputting predicted failure modes and response variables; training the failure mode identification and life prediction model using the EM backpropagation algorithm, which includes an E-step and an M-step. The E-step iteratively updates the model parameters by repeating the E-step and M-step until the loss function converges, completing the training. Specifically, the E-step calculates the posterior probability distribution of the probability vectors of components under different failure modes to determine the log-likelihood expectation; the M-step constructs the loss function based on the log-likelihood expectation and updates the model parameters based on the backpropagation algorithm; and making predictions based on the trained failure mode identification and life prediction model. Compared with existing technologies, this invention has advantages such as high identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aircraft engine failure modeling, and particularly to a method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network. BACKGROUND

[0002] The stability of manufacturing systems is of great significance to the entire production process, and its unexpected failure often leads to serious consequences such as logistics interruption, production downtime, equipment safety threat, and customer dissatisfaction. Predictive and health management (PHM) of manufacturing systems plays an important role in ensuring the stability of manufacturing systems and has been widely used in various fields including aircraft engines, batteries, and semiconductors. In health management, failure mode identification and remaining useful life (RUL) prediction based on sensor signal data is an important topic.

[0003] Unlike traditional regression problems that directly predict from sensor measurements, for multi-mode regression and mode identification, the profile or mode of sensor measurements is highly dependent on mode information, and the mapping from sensor measurements to the variable of interest (VoI) will exhibit various characteristics under different operating modes. In order to accurately depict the mapping from sensor measurements to VoI, the diversity of the mapping under multiple operating modes should be fully evaluated, and the specific operating mode of the component should be effectively identified based on sensor measurements. In the practical process of some industries (such as aircraft engine sensors), the mapping from sensor measurements to RUL exhibits different characteristics under different failure modes. This fact promotes the application of multi-mode regression and mode identification aimed at learning the mapping from engine sensor measurements to RUL under multiple failure modes, and inferring the failure mode of the engine by fully analyzing the sensor measurements.

[0004] Multi-modal regression and pattern recognition still face the following challenges: first, most existing methods directly predict VoI, ignoring the influence of various patterns. But in fact, since sensor measurements exhibit different patterns under different operating modes, the mapping from sensor measurements to VoI is highly related to pattern information; second, some recent studies implement VoI prediction and pattern recognition as independent tasks, thus failing to utilize the interaction among patterns, sensor measurements and VoI; third, only a few studies on joint implementation of multi-modal regression and pattern recognition through supervised learning models requiring operating mode of components can be used for model training, but this training method cannot be applied to scenarios where operating mode labels are not provided in advance. Multi-modal regression and pattern recognition are actually to build a regression and classification integrated model. In addition, when training such a model, assuming that VoI data are available as labels of model input and operating mode of components is unknown in advance, the regression term is a supervised learning problem, while the pattern recognition term is an unsupervised learning problem, thus obtaining a supervised and unsupervised integrated problem, and no model training method for the problem is currently proposed. SUMMARY

[0005] The purpose of the present application is to provide a kind of based on multi-mode deep neural network aircraft engine failure mode recognition and life prediction method.

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

[0007] A kind of based on multi-mode deep neural network aircraft engine failure mode recognition and life prediction method, comprising the following steps:

[0008] Construct the failure mode recognition and life prediction model based on DNN, the model with the sensor measurement value of a component of aircraft engine as input, estimate the probability that the component belongs to each failure mode, realize the mapping of sensor measurement value to response variable under each failure mode, to output the predicted failure mode and response variable, wherein the response variable is predicted based on multi-modal regression, and the response variable includes RUL, there is a multi-modal relationship between sensor measurement value and response variable, and the failure mode is predicted based on pattern recognition;

[0009] Train the failure mode recognition and life prediction model based on EM back propagation algorithm, the EM back propagation algorithm includes E step and M step, repeat E step and M step to iteratively update model parameters, until the loss function converges, complete training, wherein the E step determines the log-likelihood expectation by calculating the posterior distribution probability of the probability vector of the component under different failure modes, the M step constructs a loss function based on the log-likelihood expectation, and updates the model parameters based on the back propagation algorithm;

[0010] Sensor measurements of an aircraft engine component are obtained and a failure mode and response variable of the component are predicted based on a trained failure mode identification and life prediction model.

[0011] The failure mode identification and life prediction model is constituted by wherein,

[0012]

[0013]

[0014]

[0015] x is an input; z is a probability vector of the component in different failure modes, which is used to capture the component mode information; θ, α k , β k are model parameters; represents a DNN; y (k) is a response variable in failure mode k; u (k) is an intrinsic variable prediction value in failure mode k; and are encoding and decoding processes about x, respectively, K is the number of failure modes; x (k) is x in failure mode k reconstructed by back propagation.

[0016] The construction process of the model training set is: the parameter set of the model is γ = {θ, α1, …, α K , β1, …, β K}, and there are L components for training, for component l, l = 1, …, L, the sensor measurements x l are collected as the input of the model, and the actual values of y l and x l are used as the labels of the model input, and the model training set is represented as D = {(x1, y1), …, (x L , y L )}.

[0017] The E step obtains the posterior distribution P(z l |D, γ) of the latent variable z l in the known training set D, where z l represents the probability vector of component l in different failure modes, z l = [z l1 , …, z lk , …, z lK ], if component l belongs to mode k, z lk = 1 is defined, otherwise z lk=0, for each z lk For k = 1, ..., K, we have:

[0018]

[0019] Based on the model parameter set Y estimated in the i-th iteration (i) According to its posterior distribution P(z) lk |D,Υ (i) The estimated result is...

[0020]

[0021] The log-likelihood expectation is based on the log-likelihood function and the latent variable z. l It is determined that the log-likelihood function is:

[0022]

[0023] in,

[0024]

[0025] P(x) l ,y l |z l ,γ,x l Decomposed into P(x) l |y l ,z l ,γ,x l ) and P(y l |z l ,Υ,c l The product of ) and according to y l By x l The mapping yields P(x) l |y l ,z l ,Υ,x l )=P(x l |z l ,Υ,x l ),therefore:

[0026] P(x l ,y l |z l ,γ,x l )=P(x l |z l ,Υ,x l )P(y l |z l ,γ,x l )

[0027] in,

[0028]

[0029] and

[0030]

[0031] where, and obeys normal distribution, i.e. where x and y denote the covariance matrix of error terms, respectively.

[0032] The log-likelihood expectation of the i-th iteration is:

[0033]

[0034] where C is a constant value independent of model parameters.

[0035] The loss function is:

[0036]

[0037] where,

[0038]

[0039]

[0040] and

[0041]

[0042] The M-step updates model parameters by maximizing the log-likelihood expectation, i.e. (i+1) = argmax γ Q(γ; γ (i) ), which is equivalent to updating model parameters by minimizing the loss function.

[0043] For given D = {(x1, y1), …, (x L , y L )} and The updating of model parameters based on the back propagation algorithm is:

[0044]

[0045]

[0046]

[0047] where ε is the learning rate.

[0048] Sensor measurement value x of a component to be predicted * and the trained failure mode identification and life prediction model, according to Failure mode prediction is performed:

[0049]

[0050] Wherein, k * is the prediction result of the failure mode, and the highest probability is taken as the point estimator prediction result;

[0051] According to the trained model parameter set γ, P(y|z * , x * , γ) and P(z * |x * , γ) are obtained respectively, and the product of the two is P(y|x * , γ):

[0052]

[0053] Wherein, z * represents the probability vector of the component to be predicted under different failure modes, and the predicted value y * of the response variable is the expectation of y under the condition of P(y|x * , γ):

[0054] y * = E P(y|x*,γ) (y) = ∫yP(y|x * , γ)dy

[0055] If the above formula has no analytical solution, then y * is approximated based on the Monte Carlo Markov chain using sample mean value:

[0056]

[0057] Wherein, R is the number of random samples, and y r is the y value extracted from P(y|x * , γ) according to the sample.

[0058] Compared with the prior art, the present application has the following beneficial effects:

[0059] ​(1) Since the mapping from sensor measurements to VoI is highly dependent on pattern information, this invention focuses on multimodal regression and pattern recognition to predict VoI in multiple modes and identify specific patterns of aircraft engines from sensor measurements. The proposed DNN is a joint learning model that can perform VoI prediction in multiple modes and pattern recognition of given sensor measurements as a collaborative task, effectively combining the patterns and predicted VoI, and can cleverly capture the complex relationship between sensor measurements, VoI and operating modes.

[0060] (2) This invention innovatively proposes an EM backpropagation algorithm, which uses deep learning to solve the learning problem with latent variables in the scenario where no running mode label is provided in advance, and realizes the training of DNN model, providing a certain reference for the integrated training problem of supervised and unsupervised training.

[0061] (3) The present invention takes into account the influence of different failure modes when performing lifetime prediction, resulting in high prediction accuracy and more accurate prediction results. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method of the present invention;

[0063] Figure 2 The selected sensor signals of a historical aircraft engine that failed due to two different failure modes are shown in the example.

[0064] Figure 3 The embodiment shows the identification accuracy of fault modes at different actual RUL levels obtained by using the method described in this invention.

[0065] Figure 4 This is a schematic diagram illustrating the relative prediction error between the method described in this invention and the actual RUL. Detailed Implementation

[0066] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0067] This embodiment provides a method for aircraft engine failure mode identification and life prediction based on a multi-modal deep neural network, developing a novel deep learning method for multimodal regression and mode recognition. Specifically, a DNN-based regression and classification ensemble framework is established. This DNN framework takes sensor measurements of a component as model input and estimates the probability of the component belonging to each mode, thereby realizing the mapping from sensor measurements to VoI (Voice of Interest) in each mode. Finally, a hybrid structure combining regressions from multiple modes is obtained to output the predicted VoI and the identified mode of the component. In this embodiment, VoI includes RUL (Rounded Entities). Training this DNN is equivalent to solving a supervised and unsupervised ensemble problem. Assuming VoI data is available and the operating mode of the component is not provided, regression in each mode is a supervised learning problem, while estimating the probability of the component belonging to each mode is an unsupervised learning problem. This embodiment develops a backpropagation algorithm based on Expectation Maximization (EM) for model training. The mode of the component is set as a latent variable, and its estimate is obtained in the Expectation (E) step. Then, in the Maximization (M) step, a loss function is built based on the probability of complete data to update the model parameters. The specific implementation steps are as follows:

[0068] 1) Construct a failure mode identification and lifetime prediction model based on DNN

[0069] 1-1) Deep Neural Network (DNN) Model

[0070] This invention introduces the DNN model commonly used in deep learning. A DNN consists of an input layer, multiple hidden layers, and an output layer. x is the model input vector, y is the model output vector, and the formula for a deep neural network with J hidden layers is as follows:

[0071]

[0072] h j is the value of hidden layer j, where j = 1, ..., J, and h0 = x is located in the input layer. j and b j These are the weight matrix and bias matrix of hidden layer j, respectively. It is an activation function, such as the commonly used rectified linear unit function (ReLU).

[0073] If a deep neural network is used to solve a regression problem, the output layer is typically constructed from a linear function, in the following form:

[0074] y = h J V+d

[0075] Where V and d are the weight matrix and bias matrix of the output layer, respectively, and y is the response variable after regression. If a deep neural network is used to solve a classification problem, the output layer is constructed using the softmax function, as follows:

[0076]

[0077] Where y = [y (1) ,…y (k) ,…,y (K) ], where y is the probability vector of any category it belongs to, y (k) This represents the probability of classifying something as k.

[0078] The parameter set of the DNN model is represented as α={W1,b1,…,W J ,b J For ease of representation, let's name the DNN f and represent it as y = f. α (x). The model input x, model output y, model parameter set α, and model name f can be set according to the actual situation.

[0079] 2) DNN-based failure mode identification and lifetime prediction model

[0080] Consider a common scenario in manufacturing systems: engineers use a set of variables to describe the state of a component system. Interesting, but in reality, the only data that can be obtained is sensor signals. The objective of this invention is to provide a model for studying the theoretical mapping from x to y, thereby predicting y from a given x. x is an input vector representing sensor measurements, and y is a response variable representing a VoI vector. The components operate in multiple modes, and the multimodal characteristics of the sensor measurements x can reflect these modes. In practice, various operating conditions lead to a multimodal relationship between x and y.

[0081] To obtain the multimodal relationship between components x and y, this invention proposes an auxiliary variable vector z to capture component mode information. Assume there are K modes, each with an index k. Since the sensor measurement x can reflect the multimodal characteristics of the component, given x, the DNN can... To identify the mode in which a component is in. It is actually a multi-class classification model. Let z represent a DNN, θ represent the model parameters, and z be the model output, a vector representing the probabilities of a component being in different modes. In z = [z...] (1) ,…,z (k) ,…,z (K) In the middle, z (k) It is the probability that the component is in mode k, which is denoted as for ease of understanding.

[0082] To investigate how sensor measurements x affect and generate VoIy under arbitrary mode k, this invention establishes a DNN network as follows: The form, where k = 1, ..., K, It is a model with parameter α k The regression model, y (k) This is the VoI prediction vector under pattern k. However, since the nominal dimension usually satisfies D... y >D x The intrinsic dimension D of y y It will also be less than the dimension D of x. x Therefore, the transformation from x to y leads to information loss. Consequently, the response variable y may not contain sufficient information to reflect the multimodal characteristics of the component. To address this issue, an intrinsic variable u is introduced, carrying multimodal features from x that the response variable y cannot capture as information inherent to the model. This allows for the representation of multimodal features under any mode k. Updated to Where k = 1, ..., K, u (k) This represents the predicted value of the intrinsic variable under pattern k. For simplicity, it will be referred to as...

[0083] To ensure that the intrinsic variable u contains sufficient information about the multimodal features, this invention introduces a backward process. To reconstruct x (k = 1, ..., K) under any pattern k. This backpropagation process is also a regression model of a DNN, with model parameters β. k Intrinsic variable u (k) and the predicted value x measured by the sensor (k) These are the model input and output for any mode k, respectively. and These can be viewed as encoding and decoding processes related to x, respectively.

[0084] The formulas for the multimodal regression and model recognition neural network proposed in this invention are summarized as follows:

[0085]

[0086]

[0087]

[0088] x is the input; z is the probability vector of the component in different failure modes, used to capture component mode information; θ, α k β k These are model parameters; Represents DNN; y (k) The response variable under failure mode k; u (k) These are the predicted values ​​of the intrinsic variables under failure mode k; and These are the encoding and decoding processes for x, respectively. K is the number of failure modes; x (k) x is the failure mode k reconstructed using backpropagation.

[0089] 2) Training the failure mode identification and lifetime prediction model based on the EM backpropagation algorithm

[0090] In multimodal regression and pattern recognition, the patterns of components may be unknown and need to be learned from sensor measurement results x, which is also a challenge faced in model training. For example, when predicting the RUL of an aircraft engine, the model training does not provide historical engine failure modes; only sensor measurement results and RUL are available. Therefore, when predicting the RUL of an engine, the engine's failure modes should be identified simultaneously. Under these conditions, the problem to be solved is the integration of classification and regression, as well as the integration of unsupervised and supervised learning. Although the traditional backpropagation algorithm is widely applicable to supervised learning, it cannot be used to train the network proposed in this invention. To solve these problems, this invention provides a training method based on the EM backpropagation algorithm for model training, the pseudocode of which is as follows:

[0091]

[0092] The process of constructing the model training set is as follows: the parameter set of the model is γ={θ,α1,…,α K ,β1,…,β K Let there be L components used for training. For component l, l = 1, ..., L, sensor measurements x are collected respectively. l As input to the model, and y l and x l The actual value is used as the label for the model input. Here, considering... It can be viewed as an encoder-decoder process for reconstructing x, x l This can then be used as model input and the labels for the model input. The model training set is represented as D = {(x1, y1), ..., (x... L ,y L )}.

[0093] This embodiment first discusses the construction process of the log-likelihood function.

[0094] Since model training cannot directly obtain the pattern information of components, the variable z, which is related to the pattern data of component l, will be used instead.l If we set it as a latent variable, then the log-likelihood function of the complete data is:

[0095]

[0096] in,

[0097]

[0098] P(x) l ,y l |z l ,Υ,x l Decomposed into P(x) l |y l ,z l ,γ,x l ) and P(y l |z l ,Υ,x l The product of ) and according to y l By x l The mapping yields P(x) l |y l ,z l ,Υ,x l )=P(x l |z l ,γ,x l ),therefore:

[0099] P(x l ,y l |z l ,γ,x l )=P(x l |z l ,γ,x l )P(y l |z l ,γ,x l )

[0100] in,

[0101]

[0102] and

[0103]

[0104] In deep learning used for regression, it is assumed that... and It follows a normal distribution, that is... Where Σ x and Σ yLet Σ represent the covariance matrix of the error term. In most existing deep learning literature, the covariance matrix of the error term is often fixed and ignored (Tian et al., 2020; Kim and Liu, 2021; Wang et al., 2021). Therefore, in this embodiment, Σ x and Σ y Setting it as an identity matrix is ​​a common approach.

[0105] However, due to z l It is an unknown latent variable, and P(x) cannot be directly calculated. l ,y l |z l ,γ,x l Therefore, this invention proposes the EM backpropagation algorithm to implement model training.

[0106] The EM backpropagation algorithm includes an E step and an M step. The E step and M step are repeated to iteratively update the model parameters until the loss function converges and the training is completed. The E step determines the log-likelihood expectation by calculating the posterior distribution probability of the probability vector of the component in different failure modes. The M step constructs the loss function based on the log-likelihood expectation and updates the model parameters based on the backpropagation algorithm.

[0107] 2-1) Step E

[0108] The E step obtains the latent variable z given the training set D. l The posterior distribution P(z) l |D,Υ), where z l z represents the probability vector of component l under different failure modes. i =[z l1 ,…,z lk ,…,z lK If component l belongs to pattern k, then define z. lk =1, otherwise z lk =0, for each z lk For k = 1, ..., K, we have:

[0109]

[0110] Based on the model parameter set γ estimated in the i-th iteration (i) According to its posterior distribution P(z) lk |D,γ (i) The estimated result is...

[0111]

[0112] The model parameters at the first iteration, i.e., γ when i = 0 (i) The setup is as follows: Clustering methods (such as K-means, AGNES) are used to estimate the labels of corresponding failure modes based on sensor measurement data. Then, the estimated mode parameters are used as labels, and model parameters are obtained by training a neural network for pattern recognition and another neural network for prediction in each mode.

[0113] Log-likelihood expectation is based on the log-likelihood function and the latent variable z. l Determined, based on the posterior distribution P(z) at the i-th iteration. l |D,γ (i) The log-likelihood expectation of the i-th iteration is:

[0114]

[0115] Here, C is a constant value that is independent of the model parameters.

[0116] 2-2) M step

[0117] The M-step updates the model parameters, i.e., γ, by maximizing the log-likelihood expectation. (i+1) =argmax γ Q(γ;γ (i) To achieve this goal, the neural network proposed in this invention is designed with Q(γ,γ) in the i-th iteration. (i) Heuristic loss function The above objective is equivalent to updating the model parameters by minimizing the loss function.

[0118] The loss function is:

[0119]

[0120] in,

[0121]

[0122]

[0123] and

[0124]

[0125] For a given D = {(x1,y1),…,(x L ,y L )}and The model parameters updated based on the backpropagation algorithm (Chauvin and Rumelhart, 2013) are as follows:

[0126]

[0127]

[0128]

[0129] Where ε is the learning rate.

[0130] Repeat steps E and M to update the model parameters until the model converges. The convergence criterion is RMSE. i+1 %δ, where RMSE i+1 It is γ (i+1) and γ (i) The mean squared error is given by δ, which is a pre-specified convergence threshold. The proposed model training algorithm has been proven to have local convergence properties.

[0131] 3) Obtain sensor measurements of aircraft engine components and predict the failure modes and response variables of the components based on the trained failure mode recognition and life prediction model.

[0132] For a new component to be identified, given its sensor measurement data x * The goal is to predict VoI y using multimodal regression. * And identify the failure mode k of the component. * .

[0133] Among them, according to Perform failure mode prediction:

[0134]

[0135] Where, k * The prediction results of failure modes are as follows: The highest probability is used as the prediction of the point estimator.

[0136] By identifying the failure mode k * The specific characteristics of components in this mode can be determined, which is beneficial for the operation of the components. For example, if a degraded engine is identified as being in a fault mode, then customized maintenance can be performed once the engine fails.

[0137] To predict VoI y using multimodal regression * We need to obtain its probability density function P(y|x). * ,Υ), based on the given x * The parameter set Υ of the trained model can be used to obtain P(y|z) respectively. * ,x * ,Υ) and P(z) * |x * The product of the two is P(y|x). *,γ):

[0138]

[0139] Among them, z * This represents the probability vector of the component to be predicted under different failure modes, and the predicted value y of the response variable. * That is, given P(y|x) * Expected value of y under the condition of ,γ):

[0140] y * =E P(y|x*,Υ) (y)=∫yP(y|x * ,γ)dy

[0141] If the above equation has no analytical solution, then y * Based on Monte Carlo Markov chains, the following approximation is obtained using sample means:

[0142]

[0143] Where R is the number of random samples, y r From P(y|x) * The y-values ​​are obtained from the sample in ,Υ).

[0144] This embodiment presents a case study of a failure dataset for an aero-gas turbine engine. This dataset was generated by the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) developed by NASA and is used to simulate the degradation process of a turbofan engine.

[0145] During the service life of an aircraft engine, 21 sensor signals are available for monitoring its failure state. The collected failure dataset includes data from 100 historical aircraft engines and 100 currently operating aircraft engines. Sensor signals for historical aircraft engines were collected prior to failure. For currently operating aircraft engines, sensor signals were collected at some point before failure, providing actual RUL (Real-Use Limiting) data. Furthermore, each currently operating aircraft engine is affected by one of two potential failure modes: HPC-induced failure or engine fan-induced failure. Actual failure modes of currently operating aircraft engines are provided for model validation.

[0146] Based on the criterion that sensor signals exhibit a consistent increasing (or decreasing) trend across all historical aircraft engines under each failure mode, this embodiment selects six sensors: T24, T30, T50, P30, Ps30, and Phi. The selected sensor signals from both historical and currently operating aircraft engines are then preprocessed using logarithmic transformation and z-score normalization. The measurements from these selected sensors are used as input to the aircraft engine failure mode identification and life prediction model. RUL data is preprocessed using 0-1 normalization, which divides the actual RUL value by a fixed time length T. max . Figure 2 This paper presents selected sensor signals from a historical aircraft engine that failed due to two different failure modes, clearly demonstrating that sensor signals are often correlated and that some sensors are more sensitive to certain types of failure modes (i.e., P30 and phi). This example uses the sensor measurement "phi" and the corresponding RUL function under two different failure modes. The degradation trends of "phi" are clearly opposite in the two failure modes, and the mapping from sensor measurements to RUL exhibits different characteristics in different failure modes. In this case, a multimodal regression model should be considered to better predict RUL and identify failure modes from the sensor measurements of the aircraft engine.

[0147] This embodiment uses historical aircraft engine data to train the proposed network model. Inspired by data augmentation, a large number of training samples are generated from each historical aircraft engine data set by sliding time windows of varying widths across the entire failure data as observation points. The RUL value is obtained directly by subtracting the last observation time from the recorded failure time or by applying a piecewise linear RUL function. Finally, more than 20,000 training samples were obtained.

[0148] For model testing, this embodiment uses each aircraft engine in use as a test sample, then predicts the RUL (Range Limit Up) of the aircraft engine in use and identifies the failure modes. The performance of the proposed model is verified by comparing the predicted RUL values ​​and identified failure modes with the actual values ​​of the aircraft engines in use.

[0149] This embodiment sets two metrics to evaluate model performance. To predict the VoI variable of interest, this embodiment considers the relative RUL prediction error as a metric. The relative RUL prediction error of the new component is the predicted RUL value y. * and reality The absolute difference between them divided by the actual failure time τ * As shown below:

[0150]

[0151] For pattern recognition, this embodiment uses recognition accuracy as an indicator to measure the percentage of components whose patterns are correctly recognized.

[0152] Table 1 lists the hyperparameters selected in this embodiment. The proposed network model is then trained using historical aircraft engine data. For aircraft engines currently in use, this embodiment evaluates model performance by comparing the predicted RUL and recognition patterns of the aircraft engine with their corresponding actual values.

[0153] Table 1. Hyperparameter values ​​of the network proposed in the actual case study.

[0154]

[0155]

[0156] The prediction results are obtained based on the method described above. Figure 3 This paper demonstrates the accuracy of the fault mode identification method described in this invention at different actual RUL levels. For the actual RUL levels, "20, 40, ..., 120" refer to aircraft engines in use with actual RULs less than or equal to 20, 40, ..., 120, and "+∞" refers to all aircraft engines in use. When the actual RUL is small (e.g., less than 40), meaning that sufficient observation time points for the aircraft engines in use are available, this method exhibits high pattern recognition accuracy. When the actual RUL is less than 80, this method can identify fault modes well. As the actual RUL increases, the accuracy of this method gradually decreases, but it still maintains a relatively high recognition accuracy compared to existing technologies, demonstrating the effectiveness of this method for pattern recognition, even under conditions of insufficient sensor measurements.

[0157] also, Figure 4 The prediction errors of the model described in this invention were compared at different actual RUL levels. The points and bars in each level of the actual RUL represent the mean and standard error of the RUL prediction error, respectively. Here, the standard error is the standard deviation of the RUL prediction error divided by the square root of the number of aircraft engines. When the actual RUL is less than 80, the method described in this invention performs well in RUL prediction because the observation time points of the aircraft engines in use are relatively sufficient to provide effective information. As the actual RUL increases, the superiority of the method described in this invention becomes apparent.

[0158] Table 2 shows a comparison of the recognition accuracy and relative prediction error between the method described in this invention and existing methods. The existing methods include:

[0159] Comparison Method 1: Use k-means to cluster data with unknown patterns, then use KNN to train a classification model, and then use a neural network method to obtain RUL prediction results;

[0160] Comparison Method 2: Use AGNES to cluster data with unknown patterns, train a classification model using SVM, and then use a neural network method to obtain RUL prediction results.

[0161] Table 2 shows the model performance of the proposed method.

[0162]

[0163] Based on the above comparison results, it can be seen that the method described in this invention has significant advantages in both failure mode identification accuracy and prediction error.

[0164] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for aircraft engine failure mode identification and life prediction based on multi-mode deep neural networks, characterized in that, Includes the following steps: A failure mode recognition and life prediction model based on DNN is constructed. The model takes the sensor measurement values ​​of a component of an aircraft engine as input, estimates the probability of the component belonging to each failure mode, realizes the mapping from the sensor measurement values ​​to the response variable under each failure mode, and outputs the predicted failure mode and response variable. The response variable is predicted based on multimodal regression and includes RUL. There is a multimodal relationship between the sensor measurement values ​​and the response variable. The failure mode is predicted based on pattern recognition. The failure mode identification and lifetime prediction model is trained based on the EM backpropagation algorithm. The EM backpropagation algorithm includes an E step and an M step. The model parameters are iteratively updated by repeating the E step and the M step until the loss function converges and the training is completed. The E step determines the log-likelihood expectation by calculating the posterior distribution probability of the probability vector of the component under different failure modes. The M step constructs the loss function based on the log-likelihood expectation and updates the model parameters based on the backpropagation algorithm. Acquire sensor measurements of aircraft engine components and predict the failure modes and response variables of the components based on a trained failure mode recognition and life prediction model; The failure mode identification and lifetime prediction model is composed of , , Composition, among which, ， For input; This is a probability vector of a component in different failure modes, used to capture component mode information; , , These are model parameters; Represents a DNN; In failure mode The response variable is as follows; Failure mode Predicted values ​​of intrinsic variables under the following conditions; and They are respectively about The encoding and decoding process, , ; K The number of failure modes; Failure modes reconstructed using backpropagation k Below .

2. The method for aircraft engine failure mode identification and life prediction based on multi-mode deep neural networks according to claim 1, characterized in that, The process of constructing the model training set is as follows: the model's parameter set is... Let the components used for training be a total of For components , Collect sensor measurements respectively As input to the model, and and If the actual values ​​are used as labels for the model input, then the model training set is represented as... .

3. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 2, characterized in that, The E step obtains latent variables given the training set D. posterior distribution ,in, Representation Component The probability vector of under different failure modes If the component Belongs to pattern Then define ,otherwise For each ,have: According to the i The model parameter set estimated in the next iteration Based on its posterior distribution Estimated : 。 4. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 3, characterized in that, The log-likelihood expectation is based on the log-likelihood function and latent variables. It is determined that the log-likelihood function is: in, , Will Decomposed into and The product of, and according to Depend on The mapping yields, then ,therefore: in, and in, and It follows a normal distribution, that is , ,in and Let represent the covariance matrices of the error terms, respectively.

5. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 4, characterized in that, No. i The log-likelihood expectation of the next iteration is: in, It is a constant value that is independent of the model parameters.

6. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 5, characterized in that, The loss function is: in, and 。 7. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 6, wherein the M-step updates the model parameters by maximizing the log-likelihood expectation, i.e. This is equivalent to updating the model parameters by minimizing the loss function.

8. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 7, characterized in that, For a given and The model parameters are updated based on the backpropagation algorithm as follows: in, It is the learning rate.

9. The method for aircraft engine failure mode identification and life prediction based on a multi-mode deep neural network according to claim 8, characterized in that, Given sensor measurements of the component to be predicted And the trained failure mode recognition and lifetime prediction model, based on Perform failure mode prediction: in, The prediction results of failure modes are as follows: The highest probability is used as the prediction of the point estimator; Based on the trained model parameter set They were obtained respectively Multiplying the two together gives : . in, Represents the probability vector of the component to be predicted under different failure modes, and the predicted value of the response variable. That is, given under conditions Expectations: If the above equation has no analytical solution, then Based on Monte Carlo Markov chains, the following approximation is obtained using sample means: in, The number of random samples. From The middle is obtained from the sample extraction value.

Images