Method for predicting residual service life of aero-engine based on interpretable drift-diffusion serialization variational auto-encoder

Abstract

The invention belongs to the field of aero-engine residual service life prediction, and particularly relates to an aero-engine residual service life prediction method based on an interpretable drift-diffusion serialization variational auto-encoder. The invention aims to solve the problem of low prediction accuracy of the residual service life of the existing aero-engine. The method comprises the following steps: obtaining processed historical data of an aero-engine sensor; based on the processed historical data of the sensor and the loss function of the constructed interpretable drift diffusion serialization variational auto-encoder network, training the constructed interpretable drift diffusion serialization variational auto-encoder network model until the loss function converges; a trained interpretable drift diffusion serialization variational auto-encoder network model is obtained; and on the basis of the trained interpretable drift-diffusion serialization variational auto-encoder network model, carrying out residual service life prediction on the on-line data of the aero-engine sensor.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intersection of aero-engine remaining useful life prediction and machine learning, and the combination of explainable artificial intelligence, and particularly relates to an aero-engine remaining useful life prediction method. BACKGROUND

[0002] In the process of the intelligent operation and maintenance system of aero-engines moving towards the paradigm transformation of Industry 4.0, the thermal-mechanical fatigue life management of high-temperature alloy components and the reliability guarantee of the whole machine have become the core indicators of airworthiness certification. Taking the high-pressure turbine blade as an example, it is subjected to multiple failure risks such as high-temperature gas scouring, high / low cycle composite fatigue load, and thermal barrier coating spalling during service, so that accurate prediction of the remaining useful life (RUL) becomes a key technical breakthrough point for realizing predictive maintenance. As a forward-looking maintenance strategy, the remaining useful life prediction refers to the length of time that the aero-engine can continue to work normally in the future under the current operating conditions. The aero-engine remaining useful life prediction fuses multiple source data such as gas path parameters (such as exhaust temperature, fuel flow) and vibration spectral features to construct a quantitative mapping relationship between high-temperature component damage evolution and remaining useful life. Accurate remaining useful life prediction can realize early warning before the aero-engine fails, avoid unnecessary downtime, reduce maintenance costs, and realize predictive maintenance.

[0003] At present, the remaining useful life prediction methods of aero-engines can be generally divided into two categories: mechanism-based modeling methods and data-driven methods. The mechanism-based modeling method relies on physical laws to construct mathematical models to predict the remaining useful life by analyzing the degradation mechanism of the equipment (such as fatigue, creep, wear, etc.). This method relies on a deep understanding of material properties, structural dynamics, and environmental impact, and requires accurate physical parameters and constitutive relations. However, due to the complexity of actual working conditions and multi-physical field coupling effects, the mechanism model is often difficult to maintain high accuracy under a wide range of operating conditions, and the modeling process usually has high computational cost and limited flexibility. The data-driven method directly learns the degradation law from the equipment operating data without relying on prior physical knowledge. By analyzing sensor monitoring data (such as temperature, vibration, pressure, etc. time series signals), the data-driven method can automatically extract key features and establish a mapping relationship between them and the remaining life. With the development of artificial intelligence technology, this method has shown strong adaptability in the remaining useful life prediction of aero-engines, and can handle high-dimensional, nonlinear data, becoming a research hotspot in the remaining useful life prediction.

[0004] Data-driven deep learning models are usually considered as black-box models, and it is difficult to understand their internal working mechanism. The explainability of the model not only affects the confidence of the remaining useful life prediction result of the aero-engine, but also determines the operability and compliance of the model in practical application. Therefore, the existing deep learning-based aero-engine remaining useful life prediction method still needs more in-depth work to reveal the potential space construction and evolution of the neural network. In addition, as the mainstream way of deep learning, the deep learning method based on discriminative learning simply focuses on the mapping relationship between the degradation data and the remaining useful life, and it is difficult to learn the overall distribution of the aero-engine degradation data. How to use the generative deep learning to realize the remaining useful life prediction needs further work. Moreover, in engineering practice, the service environment of the aero-engine is complex and changeable, and the prediction result inevitably has certain uncertainty. How to incorporate uncertainty into the learning process of the neural network in the generative deep learning framework needs in-depth work. SUMMARY

[0005] The present application is to solve the problem of weak explainability of the existing aero-engine remaining useful life prediction model, complex overall distribution of degradation data, and strong uncertainty, which leads to low accuracy of aero-engine remaining useful life prediction, and proposes an aero-engine remaining useful life prediction method based on explainable drift diffusion serialized variational autoencoder.

[0006] The specific process of the aero-engine remaining useful life prediction method based on explainable drift diffusion serialized variational autoencoder is as follows:

[0007] Step one, collect sensor historical data of the aero-engine working in a service environment, process the sensor historical data to obtain processed sensor historical data; is a positive integer greater than or equal to 2;

[0008] Step two, construct an explainable drift diffusion serialized variational autoencoder network model;

[0009] Step three, construct a loss function of the explainable drift diffusion serialized variational autoencoder network;

[0010] Step four, based on the processed sensor historical data of step one and the loss function of the explainable drift diffusion serialized variational autoencoder network constructed in step three, train the explainable drift diffusion serialized variational autoencoder network model constructed in step two until the loss function converges, and obtain the trained explainable drift diffusion serialized variational autoencoder network model;

[0011] Step five, based on the trained explainable drift diffusion serialized variational autoencoder network model, the remaining useful life of the aero-engine sensor online data is predicted.

[0012] The beneficial effects of the present application are:

[0013] (1) The present application considers a dynamic serialized modeling method, and the present application proposes a deep learning method based on generation, describes the degradation mode of the aero-engine, predicts the remaining useful life of the aero-engine, and can ensure that the model effectively captures the distribution characteristics of the aero-engine degradation data, and improves the prediction accuracy of the remaining useful life of the aero-engine.

[0014] (2) The present application is from the perspective of the probability deep generation network, and the present application designs a new type of generative loss function from the perspective of Bayesian theory. On this basis, a Gaussian distribution network is designed to describe the uncertainty of the remaining useful life in the aero-engine degradation process.

[0015] (3) The present application proposes a novel neural network explainable hidden variable construction method based on drift diffusion of stochastic differential equation, which combines the dynamic representation of state and rate of change, so that the neural network model can understand and predict the evolution behavior of the aero-engine degradation data over time.

[0016] The present application first designs the loss function of the entire deep neural network, which can clearly see that the probability deep generation network specifically includes three sub-networks, namely sub-network, sub-network and sub-network. In order to better describe the uncertainty in the remaining useful life prediction process and better expand the probability operation, combined with the properties of the probability generation network, the output of each sub-network in sub-network, sub-network and corresponds to a Gaussian distribution, and outputs the corresponding mean and variance. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the workflow diagram of the present application;

[0018] Figure 2 is the structure diagram of the long short-term memory network of the present application;

[0019] Figure 3a is the result diagram of the remaining useful life prediction value and the true value of the present application on the working condition FD001;

[0020] Figure 3b is the result diagram of the remaining useful life prediction value and the true value of the present application on the working condition FD002;

[0021] Figure 3c This is a graph showing the predicted and actual remaining service life values ​​of the present invention under operating condition FD003.

[0022] Figure 3d This is a graph showing the predicted and actual remaining service life values ​​of the present invention under operating condition FD004.

[0023] Figure 4 The actual value, predicted value, and corresponding confidence interval of the remaining service life of the aircraft turbofan engine numbered 24 under the FD001 operating condition are presented.

[0024] Figure 5a This is a phase space trajectory diagram of each aero-engine under operating condition FD001 according to the present invention;

[0025] Figure 5b This is a phase space trajectory diagram of each aero-engine under operating condition FD002 according to the present invention;

[0026] Figure 5c This is a phase space trajectory diagram of each aero-engine under operating condition FD003 according to the present invention;

[0027] Figure 5d This is a phase space trajectory diagram of each aero-engine under operating condition FD004 according to the present invention. Detailed Implementation

[0028] Specific Implementation Method 1: The specific process of the aero-engine remaining service life prediction method based on interpretable drift-diffusion serialization variational autoencoder in this implementation method is as follows:

[0029] Step 1: Collect aircraft engine data Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2;

[0030] Step 2: Construct an interpretable drift-diffusion serialization variational autoencoder network model;

[0031] Step 3: Construct the loss function for the interpretable drift-diffusion serialization variational autoencoder network;

[0032] Step 4: Based on the sensor historical data processed in Step 1 and the loss function of the interpretable drift-diffusion serialization variational autoencoder network constructed in Step 3, train the interpretable drift-diffusion serialization variational autoencoder network model constructed in Step 2 until the loss function converges, and obtain the trained interpretable drift-diffusion serialization variational autoencoder network model.

[0033] Step 5: Based on the trained interpretable drift-diffusion serialization variational autoencoder network model, predict the remaining service life of the aero-engine sensors using online data.

[0034] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step one, the aero-engine data is collected... Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; the specific process is as follows:

[0035] Step 11: Collect data from the aircraft engine. Historical data of sensors operating in various service environments;

[0036] Steps 1 and 2: Addressing the characteristic that some sensor channels exhibit approximately constant values ​​throughout the engine's entire lifespan, a low-variance feature filtering strategy is employed to eliminate redundant dimensions in the multi-source sensor historical data of the aero-engine under various service conditions collected in Step 1, significantly improving data representation efficiency. The specific process is as follows:

[0037] Calculate the variance of each sensor channel data of the aero-engine over its entire life cycle. If the variance of a certain sensor channel data is less than a set threshold (e.g., 0.05), it is determined to be a low variance feature and the corresponding sensor channel data is removed to obtain the sensor historical data with redundant dimensions removed.

[0038] This invention uses the publicly available C-MAPSS dataset as an example. Each sampling time of a single aero-engine contains 21 sensor channel data, which correspond to various physical quantities such as temperature, pressure, speed, and flow rate. The original data dimension is 21.

[0039] Step 13: To address the gradient imbalance problem caused by the difference in dimensions among multiple sensors, the historical sensor data output from Step 12 (after removing redundant dimensions) is normalized by maximum and minimum values ​​to ensure the consistency of the numerical distribution of different physical quantities (such as temperature gradient, vibration acceleration, and fuel flow), thereby optimizing the stability of model training.

[0040] Step 14: This invention proposes a dynamic temporal segmentation mechanism, which uses a sliding window function to segment and reorganize the continuous historical data after normalization in Step 13 to obtain a feature sequence, so as to capture the temporal correlation characteristics of degraded data and provide a structured input feature sequence for subsequent temporal modeling.

[0041] The obtained feature sequences are used as the training sample set, which includes both degradation data and normal data of aero engines.

[0042] The other steps and parameters are the same as in Specific Implementation Method 1.

[0043] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two, an interpretable drift-diffusion serialization variational autoencoder network model is constructed;

[0044] Interpretable drift-diffusion serialization variational autoencoder network models include encoding models (bidirectional LSTM encoders) and Gaussian distribution network models ( Subnetworks and Sub-network), drift-diffusion latent variable generation model (SDE module), decoding model ( Sub-networks);

[0045] The working process of the interpretable drift-diffusion serialization variational autoencoder network model is as follows:

[0046] 1) Input the processed historical sensor data into the coding model, and the coding model outputs the feature coding result and the remaining service life tag coding result;

[0047] 2) Input the feature encoding results and remaining useful life label encoding results output by the encoding model into the Gaussian distribution network model ( Subnetworks and Sub-network), Gaussian distribution network model ( Subnetworks and The mean and variance of the initial distribution of latent variables output by the sub-network are obtained; the mean and variance of the initial distribution of latent variables output by the Gaussian distribution network model are sampled to obtain the initial latent variables. ;

[0048] 3) Initial hidden variables Input the drift-diffusion latent variable generation model (SDE module), and the drift-diffusion latent variable generation model (SDE module) outputs the complete latent variable time series, thereby explicitly characterizing the dynamic evolution of the degradation process;

[0049] 4) Input the complete latent variable time series output by the drift-diffusion latent variable generation model (SDE module), the feature encoding results output by the encoding model, and the remaining lifetime label encoding results into the decoding model ( The sub-network decodes the remaining lifetime predictions at each time step, which are used to simultaneously obtain the prediction results and uncertainty quantification, realizing an end-to-end mapping from input features to lifetime prediction.

[0050] Other steps and parameters are the same as in specific implementation method one or two.

[0051] (1) Input stage: The structure of the interpretable drift-diffusion serialization variational autoencoder uses the preprocessed time-series feature sequence of the aero-engine multi-source sensor and its corresponding remaining service life label sequence as input. (2) Encoding stage (bidirectional LSTM encoder): The feature sequence is encoded forward and backward by a bidirectional long short-term memory network, and the service life label sequence is encoded independently to obtain a high-dimensional hidden state representation that represents time dependence. (3) Gaussian network generates the initial state of the hidden variables: The feature encoding result is input into the... Sub-network, the joint encoding result of features and labels is input to Subnets, these two Gaussian distribution networks composed of multilayer sensing mechanisms output the mean and variance of the initial distribution of latent variables, and obtain the initial latent variables of the drift-diffusion process through sampling. (4) Drift-diffusion latent variable generation (SDE module): The initial latent variables are fed into the dynamic generation module based on the second-order drift-diffusion stochastic differential equation. Under the combined action of state variables, rate of change variables and enhancement variables, the complete latent variable time series is generated through iterative calculation, thereby explicitly characterizing the dynamic evolution of the degradation process. (5) Decoding stage ( Sub-network): The generated latent variable sequence is input to In this Gaussian distributed decoder, the remaining lifetime prediction and its variance at each time step are output to simultaneously obtain the prediction results and uncertainty quantification, realizing an end-to-end mapping from input features to lifetime prediction.

[0052] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: in step 1), the processed historical sensor data is input into the encoding model, and the encoding model outputs feature encoding results and remaining service life tag encoding results; the specific process is as follows:

[0053] Long Short-Term Memory Network Structure such as Figure 2 As shown, Long Short-Term Memory Network The specific formula is shown below:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] in, This represents the input vector of the Long Short-Term Memory network at time l; and These represent the Long Short-Term Memory network in... and The hidden layer state at any given time; , , , These represent the outputs of the input gate, forget gate, output gate, and memory module, respectively. This indicates the output of the memory module; and These represent the weights and bias vectors of the input gate, respectively; and These represent the weights and bias vectors of the forget gate, respectively. and These represent the weights and bias vectors of the memory module, respectively. and These represent the weights and bias vectors of the output gate, respectively. Represents the hyperbolic tangent function; This represents the sigmoid activation function; This represents the dot product operation;

[0060] Long Short-Term Memory (LSTM) networks, through gating structures such as forget gates, input gates, output gates, and memory modules, can effectively avoid problems such as gradient vanishing and gradient explosion caused by repeated multiplication of weight matrices in traditional networks.

[0061] 11) The input is the forward LSM subnetwork of the bidirectional LSM network, and the output of the forward LSM subnetwork of the bidirectional LSM network is the forward encoding result. As shown in the following formula:

[0062] (1)

[0063] in, This represents the result of the forward encoding. This represents the forward long short-term memory subnetwork of the bidirectional long short-term memory network. Indicates the time from the first moment to the second moment. Feature variables at each time point Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point; The first dimension represents the hidden layer state. The second dimension represents the hidden layer state. The first hidden layer state represents the state of the hidden layer. Dimension Indicates the time from the first moment to the second moment. Feature variables at each time point; This represents the hidden layer dimension of the feedforward long short-term memory subnetwork of the bidirectional long short-term memory network;

[0064] The whole serves as the input to the feedforward long short-term memory network;

[0065] 12) In order to better mine the implicit information of time window data and thus obtain more comprehensive time series information, this invention arranges features according to the time window data from the first... Feature variables at time 1 Feature variables up to time 1 The sequential input is processed by the backward Long Short-Term Memory (LSTM) subnetwork of the bidirectional LSM network, and the backward LSM subnetwork of the bidirectional LSM network outputs the backward encoding result. As shown in the following formula:

[0066] (2)

[0067] Where f represents the transpose operation; Represents the backward long short-term memory subnetwork of a bidirectional long short-term memory network; This represents the result of the back-encoding. Represents the characteristic variable at time 1. Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point;

[0068] 13) The Gaussian distribution network in this invention is implemented through a fully connected layer. and The mapping relationship between the mean and standard deviation estimated from the latent variables.

[0069] From the first moment to the second moment The remaining useful life tag sequence at each moment The input is the forward LSM subnetwork of the bidirectional LSM network, and the output of the remaining lifetime tag encoding result is the forward LSM subnetwork of the bidirectional LSM network. As shown in the following formula:

[0070] (3)

[0071] in, This indicates the label encoding result for the remaining useful life; Indicates the sequence Input the forward long short-term memory subnetwork of the bidirectional long short-term memory network; Indicates the time from the first moment to the second moment. The remaining useful life tag sequence at each moment. The label indicates the remaining useful life at time 1. Indicates the first The remaining useful life label at each moment; The first dimension represents the hidden layer state; The second dimension represents the hidden layer state. The first hidden layer state represents the state of the hidden layer. Dimension; This represents the hidden layer dimension of the forward long short-term memory subnetwork of the bidirectional long short-term memory network.

[0072] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0073] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step 2), the feature encoding result and the remaining lifespan tag encoding result output by the encoding model are input into the Gaussian distribution network model ( Subnetworks and Subnets), Gaussian distribution network model ( Subnetworks and The mean and variance of the initial distribution of latent variables output by the subnet are obtained; the mean and variance of the initial distribution of latent variables output by the Gaussian distribution network model are sampled to obtain the initial latent variables. The specific process is as follows:

[0074] 21) The feature encoding results output by the encoding model enter Subnetwork, The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance ; indicates as:

[0075] (4)

[0076] in, express Subnetwork; The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP).

[0077] 22) The feature encoding results output by the encoding model and remaining useful life label coding results enter Subnetwork, The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance ; indicates as:

[0078] (5)

[0079] in, express Subnetwork; The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP).

[0080] 23) Sample the mean and variance of the initial normal distribution of the latent variables to obtain the initial latent variables. ; This represents the latent variable at time 1; the specific process is as follows:

[0081] 231) Based on The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance get ; indicates as:

[0082] (6)

[0083] in, express conditional latent variables The probability density distribution; Indicates a Gaussian distribution;

[0084] 232) Based on The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance get ; indicates as:

[0085] (7)

[0086] in, express , Latent variables under joint conditions The probability density distribution;

[0087] 233), to The mean and variance of the initial normal distribution of the latent variables output by the subnetwork are randomly sampled to obtain... Initial hidden variables under conditional probability ;

[0088] right The mean and variance of the initial normal distribution of the latent variables output by the subnetwork are randomly sampled to obtain... Initial hidden variables under conditional probability .

[0089] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0090] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One through Five in that: in step 3), the initial hidden variable... The drift-diffusion latent variable generation model (SDE module) is input, and the SDE module outputs a complete latent variable time series, thus explicitly characterizing the dynamic evolution of the degradation process; the specific process is as follows:

[0091] This invention uses the estimated These serve as the initial values ​​for the drift-diffusion equation. The sequence is obtained by solving the second-order stochastic differential equation for drift-diffusion. This enables the modeling of the dynamic characteristics of aero-engine degradation data;

[0092] 31) Set the time step ; ; This represents the l-th time point;

[0093] Length of the sliding time window By time step Discretize into ;

[0094] in , ;

[0095] 32) Order ;

[0096] 33) Latent variables based on the l-th time step A three-dimensional state variable is constructed, which are state variables. State variables and enhancing variables ;

[0097] The specific process is as follows:

[0098] The latent variable at time l Input to a multilayer perceptron (MLP), the multilayer perceptron (MLP) outputs state variables. State variables and enhancing variables ;

[0099] State variables and state variables This is described by the following second-order state-space equation:

[0100] (8)

[0101] in, express The first derivative; express The first derivative; Indicates will and Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP); Indicates the first State variables among the implicit variables at each time point Indicates the first State variables among the implicit variables at each time point;

[0102] This invention employs a multilayer perceptron (MLP) to allow the association of two state variables. The reason for using second-order state-space equations in this invention is that some behaviors in aero-engine systems (such as mechanical vibrations) exhibit second-order dynamic characteristics, meaning their states depend not only on the current state but also on the rate of change of that state. For example, if... This refers to the degradation of the system (in the field of aero-engine health management and life prediction, degradation refers to the process by which the performance of the engine and its key components gradually declines and the structure gradually deteriorates under the influence of various factors such as long-term operation, thermal load and mechanical stress, combustion pollution, corrosion and wear). This refers to the system's degradation rate (in the context of aero-engine remaining service life prediction or degradation modeling, "degradation rate" refers to the speed at which the engine's health deteriorates over time, i.e., the rate at which health indicators decline or performance is lost). This will provide some interpretability for remaining service life prediction. Furthermore, there are enhancement variables among the latent variables. This is to provide an additional degree of freedom to the aforementioned two-dimensional dynamic system, allowing it to deviate from its original dimensional constraints and thus better simulate the degradation of complex systems. Furthermore, this augmenting variable can also describe the individual differences between different complex systems.

[0103] 34) Construct the drift sub-item and the diffusion sub-item respectively, represented as:

[0104] (9)

[0105] in, and These represent multilayer perceptrons (MLPs);

[0106] This indicates that the state variable State variables and enhancing variables Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP);

[0107] This indicates that the state variable State variables and enhancing variables Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP);

[0108] Indicates the drift sub-item; Indicates a diffusion sub-item;

[0109] 35) Based on drift sub-items diffusion sub-items and time step Predicting the first Latent variables at time ; for the first Latent variables at time Correction, to obtain Latent variables at time ; indicates as:

[0110] Prediction steps: (10)

[0111] Calibration steps: (11)

[0112] in, A random variable representing a standard normal distribution;

[0113] 36) Order ,based on Latent variables at time calculate Latent variables at time The specific process is the same as the prediction and correction steps.

[0114] 37) Repeat step 36) until... Obtain the Latent variables at time ;

[0115] Until the complete latent variable time series is obtained ; Indicates the time from the first moment to the second moment. The sequence of latent variables at each time step.

[0116] By continuously performing the above process, a complete sequence can be obtained. In detail, this is as follows: Starting with the initial latent variables obtained from Gaussian network sampling, the time interval is discretized into several step sizes. In each step, the current state is first used to predict the drift-diffusion equation. Then, the predicted values ​​are used to recalculate the drift and diffusion terms for correction, and random noise is injected to reflect uncertainty. This process is continued step by step until a complete sequence is generated. The sequence is used by the decoding network to output the lifetime prediction distribution at each time step.

[0117] The essence of this iterative process is to first use a simple method (such as the Euler method) to predict the value of the next time step. Then, the drift and diffusion terms are recalculated using the results of the prediction step to correct the value of the next time step and improve the accuracy of the generation process.

[0118] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0119] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: in step 4), the complete latent variable time series output by the drift-diffusion latent variable generation model (SDE module) is input into the decoding model ( The sub-network decodes the model's output remaining lifetime predictions at each time step, simultaneously obtaining the prediction results and uncertainty quantification, thus achieving an end-to-end mapping from input features to lifetime prediction; the specific process is as follows:

[0120] 41) The complete latent variable time series output by the drift-diffusion latent variable generation model (SDE module) Feature encoding results output by the encoding model and remaining useful life label coding results enter Subnetwork, Subnetwork output mean With variance ; indicates as:

[0121] (12)

[0122] in, express Subnetwork, The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP).

[0123] 42) Based on Mean of subnetwork output With variance get ; indicates as:

[0124] (13)

[0125] in, express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point;

[0126] Mean of subnetwork output The remaining lifetime predictions at each time step are output by the decoding model.

[0127] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0128] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the loss function for constructing the interpretable drift-diffusion serialization variational autoencoder network in step three is as follows:

[0129] Step 31, based on 231) obtained and 232) obtained Obtain the loss function of the KL divergence distance term. ; indicates as:

[0130] (14)

[0131] in, The loss function representing the KL divergence distance term;

[0132] Represents probability density distribution and probability density distribution KL divergence between them;

[0133] express The expected value under the given conditions;

[0134] express , Latent variables under joint conditions The probability density distribution function; the process of obtaining it is as follows:

[0135] (15)

[0136] in, express Subnetwork, The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP). Indicates will and enter Subnetwork, The output of the subnetwork;

[0137] express The mean of the subnetwork output, express The variance of the subnetwork output;

[0138] based on Mean of subnetwork output and variance get ; indicates as:

[0139] (16)

[0140] in, express , Latent variables under joint conditions The probability density distribution function;

[0141] Step 3.2: Based on the remaining useful life tag probability density distribution obtained in step 42), Loss function for obtaining the negative logarithmic term ; indicates as:

[0142] (17)

[0143] in, The loss function representing the negative logarithmic term;

[0144] Step 3: Loss Function Based on KL Divergence Distance Term Loss function of negative logarithmic term To obtain the loss function of an interpretable drift-diffusion serialization variational autoencoder network. This is used for parameter updates in deep neural networks; it is represented as:

[0145] (18)

[0146] in, This represents the loss function of an interpretable serialization variational autoencoder network.

[0147] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0148] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the loss function of the interpretable drift-diffusion serialization variational autoencoder network in step three is... The acquisition process is as follows:

[0149] Step 331, in Remaining useful life tag sequence under certain conditions The negative log-likelihood function J is expressed as:

[0150] (19)

[0151] in, Indicates the time from the first moment to the second moment. The sequence of latent variables at time 1, This represents the latent variable at time 1. Indicates the first Latent variables at each moment, Indicates the length of the sliding time window;

[0152] Indicates the time from the first moment to the second moment. The remaining useful life tag sequence at each moment. The label indicates the remaining useful life at time 1. Indicates the first The remaining useful life label at any given time. Indicates the time from the first moment to the second moment. Feature variables at each time point Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point express Under the conditions The probability density distribution, express Under the conditions and The joint probability density distribution, Represents the negative log-likelihood function;

[0153] Step 332: To better calculate the above integrals, an importance distribution function was designed. To describe the latent variable space under joint conditions. ;

[0154] Equation (19) can be transformed into equation (20), which can be expressed as:

[0155] (20)

[0156] in, Represents the importance distribution function, express The expected value under the given conditions;

[0157] Step 333: Based on the formula in Step 332, and using Jensen's inequality, we obtain a lower bound for the negative log-likelihood function J, as shown below:

[0158] (twenty one)

[0159] Step 334: Apply Bayesian chain rule to equation (21) Expanding the image yields the following form:

[0160] (twenty two)

[0161] in, express , , The first under joint conditions The probability density distribution of the remaining lifetime tag at each time point; Indicates the first The remaining useful life label at each moment; express , , The remaining lifetime tag probability density distribution at time l under joint conditions; express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point; express conditional latent variable sequence The probability density distribution; express The parameter space of latent variables under given conditions;

[0162] Step 335: Apply Bayesian chain rule to formula (22) Expanding the image yields the following form:

[0163] (twenty three)

[0164] in, express , Latent variables under joint conditions The probability density distribution; express conditional latent variable sequence The probability density distribution; express conditional latent variables The probability density distribution; express conditional latent variables The probability density distribution;

[0165] Step 336: Calculate the result in equation (23). Substituting into equation (22) in step 334, equation (22) becomes:

[0166] (twenty four)

[0167] Step 337: Apply Bayesian chain rule to formula (20) Expanding the image yields the following form:

[0168] (25)

[0169] in, express , Latent variables under joint conditions The probability density distribution;

[0170] Step 338: Let the right side of the inequality in formula (21) be . According to formulas (24)~(25), the following expressions exist:

[0171] (26)

[0172] in, The loss function representing the interpretability drift-diffusion serialization variational autoencoder;

[0173] Step 339: Combining the properties of logarithmic functions, further decompose formula (26) to obtain the form described by the following formula:

[0174] (27)

[0175] in, express The expected value under the given conditions Represents the importance distribution function;

[0176] Step 30: Combining the mathematical properties of KL divergence, formula (27) is expressed in the form described by the following equation:

[0177]

[0178] in, express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point; This represents the importance distribution function.

[0179] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0180] The design of an interpretable drift-diffusion serialization variational autoencoder is as follows: Considering a dynamic serialization modeling approach, this invention uses data within a time window as the research object, and the loss function maximizes the conditional probability density of the generative model. Furthermore, an importance distribution function is designed to describe the latent variable space under joint conditions. Using the Bayesian chain rule, combined with the properties of the logarithmic function and the mathematical properties of KL divergence, the loss function for the entire deep neural network is designed, constructing a probabilistic deep generative network, specifically comprising three sub-networks: Subnetworks, κ1 subnetwork, and y subnetwork. Combining the properties of probabilistic generative networks, the output of each subnetwork corresponds to a Gaussian distribution, with a corresponding mean and variance.

[0181] Generative learning methods ensure that deep learning models effectively capture the distribution characteristics of degraded data. The serial variational autoencoder proposed in this invention is essentially a generative learning method; its core objective in its loss function is to maximize the conditional probability density of the generative model. ,in For the system's degradation characteristic variables, The remaining useful life is used as a label. Considering the time dependence of the degraded data, it is necessary to combine the degraded data into a sliding time window and input it into the serialization variational autoencoder. Considering a dynamic serialization modeling approach, this invention uses data from a single time window as the research object, and the loss function maximizes the conditional probability density of the generative model. , where L is the length of the sliding time window.

[0182] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that: in step five, the remaining service life prediction of the aero-engine sensor online data is performed based on the trained interpretable serialization variational autoencoder network model; the specific process is as follows:

[0183] Collect online data from aero-engine sensors, process the online data from aero-engine sensors using step one, and obtain processed online data from aero-engine sensors.

[0184] The processed online data from the aero-engine sensors is input into a trained interpretable serializable variational autoencoder network model, which then outputs a predicted value for the remaining service life of the aero-engine sensors.

[0185] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0186] Evaluation of Remaining Useful Life Prediction Performance: The root mean square error (RMSE) was used to evaluate the prediction performance of the proposed adaptive remaining useful life. The expression for the root mean square error is shown below:

[0187]

[0188] Where N is the number of test data samples, i is the sample index, and RUL pi and RUL ti These are the predicted and actual values ​​of the remaining useful life for the i-th sample, respectively. A smaller root mean square error indicates a better online prediction effect of the remaining useful life proposed in this invention.

[0189] Example:

[0190] This invention uses an aero-engine degradation simulation dataset to validate the performance of the proposed remaining service life prediction model. This dataset simulates the full-life degradation process of an aero-engine under various operating conditions (including different combinations of thrust loads and environmental parameters), with each flight cycle corresponding to a complete mission cycle. The dataset covers full-dimensional monitoring data from multiple aero-engines, completely recording the changing trends of 26-channel sensor parameters (such as exhaust temperature, compressor outlet pressure, rotor speed, etc.) as the number of flight cycles increases, accurately characterizing the dynamic evolution of the aero-engine's health status. The data for each operating condition is divided into a training subset and a test subset: the training subset contains full-cycle monitoring records from the engine's initial healthy state to functional failure; the test subset extracts degradation segments from certain operational phases, using the remaining service life value at the termination time as the benchmark for model performance quantification. Specific data configuration details are shown in Table 1.

[0191]

[0192] The implementation steps of this invention are as follows:

[0193] Step 1: Collect aircraft engine data Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; the specific process is as follows:

[0194] Step 11: Collect data from the aircraft engine. Historical data of sensors operating in various service environments;

[0195] Steps 1 and 2: A low-variance feature filtering strategy is used to remove redundant dimensions from the historical aero-engine sensor data collected in Step 1. The specific process is as follows:

[0196] Calculate the variance of each sensor channel data of the aero-engine over its entire life cycle. If the variance of a certain sensor channel data is less than a set threshold, it is determined to be a low variance feature and the corresponding sensor channel data is removed to obtain the sensor historical data with redundant dimensions removed.

[0197] Step 13: Perform maximum-minimum normalization on the sensor historical data output from Step 12 after removing redundant dimensions.

[0198] Step 14: Use a sliding window function to segment and reorganize the continuous historical data after normalization in Step 13 to obtain the feature sequence;

[0199] The obtained feature sequences are used as the training sample set, which includes both degradation data and normal data of aero engines.

[0200] In step two, an interpretable drift-diffusion serialization variational autoencoder network model is constructed;

[0201] Interpretable drift-diffusion serialization variational autoencoder network models include encoding models, Gaussian distribution network models, drift-diffusion latent variable generation models, and decoding models;

[0202] The working process of the interpretable drift-diffusion serialization variational autoencoder network model is as follows:

[0203] 1) Input the processed historical sensor data into the coding model, and the coding model outputs the feature coding result and the remaining service life tag coding result;

[0204] 2) Input the feature encoding results and remaining useful life tag encoding results output by the encoding model into the Gaussian distribution network model. The Gaussian distribution network model outputs the mean and variance of the initial distribution of the latent variables. Sample the mean and variance of the initial distribution of the latent variables output by the Gaussian distribution network model to obtain the initial latent variables. ;

[0205] 3) Initial hidden variables Input the drift-diffusion latent variable generation model, and the drift-diffusion latent variable generation model outputs the complete latent variable time series;

[0206] 4) Input the complete latent variable time series output by the drift-diffusion latent variable generation model, the feature encoding results output by the encoding model, and the remaining useful life label encoding results into the decoding model. The decoding model outputs the predicted remaining useful life value at each time step.

[0207] Step 3: Construct the loss function for an interpretable drift-diffusion serialization variational autoencoder network; the specific process is as follows:

[0208] Step 3: Obtain the loss function for the KL divergence distance term. ; indicates as:

[0209]

[0210] in, The loss function representing the KL divergence distance term;

[0211] Step 3.2: Obtain the loss function for the negative logarithmic term. ; indicates as:

[0212]

[0213] in, The loss function representing the negative logarithmic term;

[0214] Step 3: Loss Function Based on KL Divergence Distance Term Loss function of negative logarithmic term To obtain the loss function of an interpretable drift-diffusion serialization variational autoencoder network. ; indicates as:

[0215] (18)

[0216] in, This represents the loss function of an interpretable serialization variational autoencoder network.

[0217] Step 4: Based on the sensor historical data processed in Step 1 and the loss function of the interpretable drift-diffusion serialization variational autoencoder network constructed in Step 3, train the interpretable drift-diffusion serialization variational autoencoder network model constructed in Step 2 until the loss function converges, and obtain the trained interpretable drift-diffusion serialization variational autoencoder network model.

[0218] Step 5: Predict the remaining service life of aero-engine sensors based on the trained interpretable drift-diffusion sequential variational autoencoder network model; the specific process is as follows:

[0219] Collect online data from aero-engine sensors, process the online data from aero-engine sensors using step one, and obtain processed online data from aero-engine sensors.

[0220] The processed online data from the aero-engine sensors is input into a trained interpretable serializable variational autoencoder network model, which then outputs a predicted value for the remaining service life of the aero-engine sensors.

[0221] Step 6: Evaluate the remaining useful life prediction effect: The root mean square error is used to evaluate the online remaining useful life prediction effect of the method proposed in this invention.

[0222] Figure 3a , 3b Tables 3c and 3d demonstrate the predicted and actual values ​​of the proposed method under four operating conditions. Table 2 shows a comparison of the remaining service life prediction results of this invention with those of several existing aero-engine remaining service life prediction methods. Figure 3a , 3b As can be seen from 3c, 3d and Table 2, the technical solution proposed in this invention can achieve high-precision prediction of the remaining service life of aero engines, and can provide real-time health status assessment and maintenance decision optimization support for operation and maintenance personnel, thereby systematically improving the intelligent operation and maintenance level of aero engines.

[0223]

[0224] This invention uses the FD001 test dataset as an example to demonstrate the actual and predicted remaining service life of aircraft turbofan engine number 24, along with the corresponding confidence intervals (this invention uses 95% confidence intervals for demonstration). Figure 4 As shown, this is used to measure the uncertainty of the degradation model. From Figure 4 As can be seen, for cases with a large remaining useful life, the predicted confidence interval is wider, and the model's prediction is more conservative. On the other hand, for cases with a small remaining useful life, the predicted confidence interval is narrower. This indicates that the remaining useful life prediction will become more accurate over time.

[0225] Compared to traditional variational autoencoders which only have two dimensions, mean and variance, the method proposed in this invention has a third dimension: it can plot a potential phase space trajectory for the degradation process of each aircraft turbofan engine and present it as a scatter plot. Figure 5a , 5b Figures 5c and 5d demonstrate the phase space trajectory of each aero-engine under four operating conditions using the proposed method. This invention combines state and rate-of-change representation methods, enabling neural network models to understand and predict the evolution of aero-engine degradation data over time to a certain extent, thereby improving the interpretability of neural networks.

[0226] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for predicting the remaining service life of aero-engines based on an interpretable drift-diffusion serialization variational autoencoder, characterized in that: The specific process of the method is as follows: Step 1: Collect aircraft engine data Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; Step 2: Construct an interpretable drift-diffusion serialization variational autoencoder network model; Step 3: Construct the loss function for the interpretable drift-diffusion serialization variational autoencoder network; Step 4: Based on the sensor historical data processed in Step 1 and the loss function of the interpretable drift-diffusion serialization variational autoencoder network constructed in Step 3, train the interpretable drift-diffusion serialization variational autoencoder network model constructed in Step 2 until the loss function converges, and obtain the trained interpretable drift-diffusion serialization variational autoencoder network model. Step 5: Based on the trained interpretable drift-diffusion serialization variational autoencoder network model, predict the remaining service life of the aero-engine sensors using online data.

2. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 1, characterized in that: In step one, the aero-engine is collected. Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; the specific process is as follows: Step 11: Collect data from the aircraft engine. Historical data of sensors operating in various service environments; Steps 1 and 2: A low-variance feature filtering strategy is used to remove redundant dimensions from the historical aero-engine sensor data collected in Step 1. The specific process is as follows: Calculate the variance of each sensor channel data of the aero-engine over its entire life cycle. If the variance of a certain sensor channel data is less than a set threshold, it is determined to be a low variance feature and the corresponding sensor channel data is removed to obtain the sensor historical data with redundant dimensions removed. Step 13: Perform maximum-minimum normalization on the sensor historical data output from Step 12 after removing redundant dimensions. Step 14: Use a sliding window function to segment and reorganize the continuous historical data after normalization in Step 13 to obtain the feature sequence; The obtained feature sequences are used as the training sample set, which includes both degradation data and normal data of aero engines.

3. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 2, characterized in that: In step two, an interpretable drift-diffusion serialization variational autoencoder network model is constructed. Interpretable drift-diffusion serialization variational autoencoder network models include encoding models, Gaussian distribution network models, drift-diffusion latent variable generation models, and decoding models; The working process of the interpretable drift-diffusion serialization variational autoencoder network model is as follows: 1) Input the processed historical sensor data into the coding model, and the coding model outputs the feature coding result and the remaining service life tag coding result; 2) Input the feature encoding results and remaining useful life tag encoding results output by the encoding model into the Gaussian distribution network model. The Gaussian distribution network model outputs the mean and variance of the initial distribution of the latent variables. Sample the mean and variance of the initial distribution of the latent variables output by the Gaussian distribution network model to obtain the initial latent variables. ; 3) Initial hidden variables Input the drift-diffusion latent variable generation model, and the drift-diffusion latent variable generation model outputs the complete latent variable time series; 4) Input the complete latent variable time series output by the drift-diffusion latent variable generation model, the feature encoding results output by the encoding model, and the remaining useful life label encoding results into the decoding model. The decoding model outputs the predicted remaining useful life value at each time step.

4. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 3, characterized in that: In step 1), the processed historical sensor data is input into the coding model, and the coding model outputs feature coding results and remaining service life tag coding results; the specific process is as follows: 11) The input is the forward LSM subnetwork of the bidirectional LSM network, and the output of the forward LSM subnetwork of the bidirectional LSM network is the forward encoding result. ; As shown in the following formula: (1) in, This represents the result of the forward encoding. This represents the forward long short-term memory subnetwork of the bidirectional long short-term memory network. Indicates the time from the first moment to the second moment. Feature variables at each time point Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point; The first dimension represents the hidden layer state. The second dimension represents the hidden layer state. The first hidden layer state represents the state of the hidden layer. Dimension Indicates the time from the first moment to the second moment. Feature variables at each time point; This represents the hidden layer dimension of the feedforward long short-term memory subnetwork of the bidirectional long short-term memory network; 12) Arrange the features according to the order from the first... Feature variables at time 1 Feature variables up to time 1 The sequential input is processed by the backward Long Short-Term Memory (LSTM) subnetwork of the bidirectional LSM network, and the backward LSM subnetwork of the bidirectional LSM network outputs the backward encoding result. As shown in the following formula: (2) Where f represents the transpose operation; Represents the backward long short-term memory subnetwork of a bidirectional long short-term memory network; This represents the result of the back-encoding. Represents the characteristic variable at time 1. Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point; 13) Transfer the time from the first moment to the second moment. The remaining useful life tag sequence at each moment The input is the forward LSM subnetwork of the bidirectional LSM network, and the output of the remaining lifetime tag encoding result is the forward LSM subnetwork of the bidirectional LSM network. As shown in the following formula: (3) in, This indicates the label encoding result for the remaining useful life; Indicates the sequence Input the forward long short-term memory subnetwork of the bidirectional long short-term memory network; Indicates the time from the first moment to the second moment. The remaining useful life tag sequence at each moment. The label indicates the remaining useful life at time 1. Indicates the first The remaining useful life label at each moment; The first dimension represents the hidden layer state; The second dimension represents the hidden layer state. The first hidden layer state represents the state of the hidden layer. Dimension; This represents the hidden layer dimension of the forward long short-term memory subnetwork of the bidirectional long short-term memory network.

5. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 4, characterized in that: In step 2), the feature encoding results and remaining useful life tag encoding results output by the encoding model are input into the Gaussian distribution network model. The Gaussian distribution network model outputs the mean and variance of the initial distribution of the latent variables. The mean and variance of the initial distribution of the latent variables output by the Gaussian distribution network model are sampled to obtain the initial latent variables. The specific process is as follows: 21) The feature encoding results output by the encoding model enter Subnetwork, The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance ; indicates as: (4) in, express Subnetwork; The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP). 22) The feature encoding results output by the encoding model and remaining useful life label coding results enter Subnetwork, The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance ; indicates as: (5) in, express Subnetwork; The sub-network is a multilayer perceptron (MLP); 23) Sample the mean and variance of the initial normal distribution of the latent variables to obtain the initial latent variables. ; This represents the latent variable at time 1; the specific process is as follows: 231) Based on The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance get ; indicates as: (6) in, express conditional latent variables The probability density distribution; Indicates a Gaussian distribution; 232) Based on The mean of the initial normal distribution of the hidden variables output by the subnetwork and variance get ; indicates as: (7) in, express , Latent variables under joint conditions The probability density distribution; 233), to The mean and variance of the initial normal distribution of the latent variables output by the subnetwork are randomly sampled to obtain... Initial hidden variables under conditional probability ; right The mean and variance of the initial normal distribution of the latent variables output by the subnetwork are randomly sampled to obtain... Initial hidden variables under conditional probability .

6. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 5, characterized in that: In step 3), the initial hidden variable will be... Input the drift-diffusion latent variable generation model, and the drift-diffusion latent variable generation model outputs a complete latent variable time series; the specific process is as follows: 31) Set the time step ; ; This represents the l-th time point; Length of the sliding time window By time step Discretize into ; in , ; 32) Order ; 33) Latent variables based on the l-th time step A three-dimensional state variable is constructed, which are state variables. State variables and enhancing variables ; The specific process is as follows: The latent variable at time l Input to a multilayer perceptron (MLP), the multilayer perceptron (MLP) outputs state variables. State variables and enhancing variables ; State variables and state variables This is described by the following second-order state-space equation: (8) in, express The first derivative; express The first derivative; Indicates will and Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP); Indicates the first State variables among the implicit variables at each time point Indicates the first State variables among the implicit variables at each time point; 34) Construct the drift sub-item and the diffusion sub-item respectively, represented as: (9) in, and These represent multilayer perceptrons (MLPs); This indicates that the state variable State variables and enhancing variables Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP); This indicates that the state variable State variables and enhancing variables Input to a multilayer perceptron (MLP), output of the multilayer perceptron (MLP); Indicates the drift sub-item; Indicates a diffusion sub-item; 35) Based on drift sub-items diffusion sub-items and time step Predicting the first Latent variables at time ; for the first Latent variables at time Correction, to obtain Latent variables at time ; indicates as: Prediction steps: (10) Calibration steps: (11) in, A random variable representing a standard normal distribution; 36) Order ,based on Latent variables at time calculate Latent variables at time ; 37) Repeat step 36) until... Obtain the Latent variables at time ; Until the complete latent variable time series is obtained ; Indicates the time from the first moment to the second moment. The sequence of latent variables at each time step.

7. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 6, characterized in that: In step 4), the complete latent variable time series output by the drift-diffusion latent variable generation model is input into the decoding model, and the decoding model outputs the predicted remaining lifetime value for each time step; the specific process is as follows: 41) The complete latent variable time series output by the drift-diffusion latent variable generation model. Feature encoding results output by the encoding model and remaining useful life label coding results enter Subnetwork, Subnetwork output mean With variance ; indicates as: (12) in, express Subnetwork, The sub-network is a Gaussian distributed network model, and the Gaussian distributed network model is a multilayer perceptron (MLP). 42) Based on Mean of subnetwork output With variance get ; indicates as: (13) in, express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point; Mean of subnetwork output The remaining lifetime predictions at each time step are output by the decoding model.

8. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 7, characterized in that: The loss function for constructing the interpretable drift-diffusion serialization variational autoencoder network in step three is as follows: Step 31, based on 231) obtained and 232) obtained Obtain the loss function of the KL divergence distance term. ; indicates as: (14) in, The loss function representing the KL divergence distance term; Represents probability density distribution and probability density distribution KL divergence between them; express The expected value under the given conditions; express , Latent variables under joint conditions The probability density distribution function; the process of obtaining it is as follows: (15) in, express Subnetwork, The sub-network is a multilayer perceptron (MLP). Indicates will and enter Subnetwork, The output of the subnetwork; express The mean of the subnetwork output, express The variance of the subnetwork output; based on Mean of subnetwork output and variance get ; indicates as: (16) in, express , Latent variables under joint conditions The probability density distribution function; Step 3.2: Based on the remaining useful life tag probability density distribution obtained in step 42), Loss function for obtaining the negative logarithmic term ; indicates as: (17) in, The loss function representing the negative logarithmic term; Step 3: Loss Function Based on KL Divergence Distance Term Loss function of negative logarithmic term To obtain the loss function of an interpretable drift-diffusion serialization variational autoencoder network. ; indicates as: (18) in, This represents the loss function of an interpretable serialization variational autoencoder network.

9. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 8, characterized in that: The loss function of the interpretable drift-diffusion serialization variational autoencoder network in step three is as follows. The acquisition process is as follows: Step 331, in Remaining useful life tag sequence under certain conditions The negative log-likelihood function J is expressed as: (19) in, Indicates the time from the first moment to the second moment. The sequence of latent variables at time 1, This represents the latent variable at time 1. Indicates the first Latent variables at each moment, Indicates the length of the sliding time window; Indicates the time from the first moment to the second moment. The remaining useful life tag sequence at each moment. The label indicates the remaining useful life at time 1. Indicates the first The remaining useful life label at any given time. Indicates the time from the first moment to the second moment. Feature variables at each time point Represents the characteristic variable at time 1. Indicates the first Feature variables at each time point express Under the conditions The probability density distribution, express Under the conditions and The joint probability density distribution, Represents the negative log-likelihood function; Step 332: The importance distribution function was designed. ; Equation (19) can be transformed into equation (20), which can be expressed as: (20) in, Represents the importance distribution function, express The expected value under the given conditions; Step 3.

3. Based on Jensen's inequality, a lower bound for the negative log-likelihood function J is obtained, as shown below: (21) Step 334: Apply Bayesian chain rule to equation (21) Expanding the image yields the following form: (22) in, express , , The first under joint conditions The probability density distribution of the remaining lifetime tag at each time point; Indicates the first The remaining useful life label at each moment; express , , The remaining lifetime tag probability density distribution at time l under joint conditions; express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point; express conditional latent variable sequence The probability density distribution; express The parameter space of latent variables under given conditions; Step 335: Apply Bayesian chain rule to formula (22) Expanding the image yields the following form: (23) in, express , Latent variables under joint conditions The probability density distribution; express conditional latent variable sequence The probability density distribution; express conditional latent variables The probability density distribution; express conditional latent variables The probability density distribution; Step 336: Calculate the result in equation (23). Substituting into equation (22) in step 334, equation (22) becomes: (24) Step 337: Apply Bayesian chain rule to formula (20) Expanding the image yields the following form: (25) in, express , Latent variables under joint conditions The probability density distribution; Step 338: Let the right side of the inequality in formula (21) be . According to formulas (24)~(25), the following expressions exist: (26) in, The loss function representing the interpretability drift-diffusion serialization variational autoencoder; Step 339: Further break down formula (26) to obtain the form described by the following formula: (27) in, express The expected value under the given conditions Represents the importance distribution function; Step 30, Formula (27) is expressed in the form described by the following formula: in, express , , Under joint conditions The probability density distribution of the remaining lifetime tag at each time point; This represents the importance distribution function.

10. The method for predicting the remaining service life of an aero-engine based on an interpretable drift-diffusion serialization variational autoencoder according to claim 9, characterized in that: In step five, the remaining service life of the aero-engine sensors is predicted based on the trained interpretable serialization variational autoencoder network model; the specific process is as follows: Collect online data from aero-engine sensors, process the online data from aero-engine sensors using step one, and obtain processed online data from aero-engine sensors. The processed online data from the aero-engine sensors is input into a trained interpretable serializable variational autoencoder network model, which then outputs a predicted value for the remaining service life of the aero-engine sensors.

Images