Semi-supervised aero-engine remaining service life prediction method based on serialized variational auto-encoder

Abstract

The invention discloses a semi-supervised aero-engine residual service life prediction method based on a serialized variational auto-encoder, belongs to the subject crossing field of aero-engine residual service life prediction, artificial intelligence and semi-supervised learning combination, and particularly relates to a semi-supervised aero-engine residual service life prediction method. The invention aims to solve the problems of low prediction precision, poor interpretability and strong uncertainty of the residual service life of the aero-engine under the semi-supervised learning condition. The method comprises the following steps: constructing a semi-supervised serialized variational auto-encoder network based on a self-attention mechanism-serialized variational auto-encoder network under a supervised condition and an unsupervised condition; a trained semi-supervised serialized variational auto-encoder network is obtained; and preprocessing online data acquired by an aero-engine sensing device, inputting the preprocessed online data into the trained semi-supervised serialized variational auto-encoder network, and predicting the remaining service life of the aero-engine.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of combining aircraft engine remaining useful life prediction with artificial intelligence and semi-supervised learning, and specifically relates to a semi-supervised aircraft engine remaining useful life prediction method. Background Art

[0002] As the development of intelligent aircraft operation and maintenance systems accelerates, the reliability of aircraft engines, the core of aircraft power systems, directly determines aviation safety and operational economics. High-pressure compressor blades in aircraft engines are subjected to long-term high-temperature, high-speed, and alternating loads, leading to the initiation and propagation of microcracks. Furthermore, components in the hot end of the engine's combustion chamber inevitably experience creep deformation under extreme thermal gradients. These progressive degradation phenomena make remaining useful life (RUL) prediction a key technology for implementing condition-based maintenance.

[0003] Currently, methods for predicting the remaining useful life of aircraft engines can be broadly divided into two types: mechanism-based modeling and data-driven methods. Mechanism-based modeling methods rely on physical laws and domain knowledge to construct mathematical models and predict remaining useful life by analyzing the evolution of engine structural characteristics and material properties. These methods require precise physical parameters and constitutive relationships, but their prediction accuracy is easily affected by simplified model assumptions in complex actual service environments (such as dynamic environmental disturbances and unsteady load fluctuations), making it difficult to fully reflect degradation behavior under the influence of multiple coupled factors. Data-driven methods, on the other hand, use big data analysis techniques to automatically extract degradation features from multi-source sensor data and establish a correlation model between monitoring signals and remaining useful life. With the development of big data technology, these methods have demonstrated powerful nonlinear modeling capabilities, can adaptively process time series data under complex operating conditions, and significantly improve prediction adaptability in practical aerospace engineering.

[0004] Acquiring high-confidence degradation data and remaining useful life (RUS) labels in the lifecycle health management of aircraft engines presents a significant challenge. As high-value, precision power plants, critical aircraft engine components (such as high-pressure compressor rotor blades) must operate continuously under extreme thermal and mechanical coupling environments. This results in high costs for collecting data for full-lifecycle monitoring. Acquiring complete degradation data for a single engine requires tens of thousands of flight hours, and accelerated degradation testing is difficult due to airworthiness safety regulations. Traditional data-driven approaches rely on massive amounts of labeled samples, which are out of touch with the actual operation and maintenance scenarios of aircraft engines. Although semi-supervised learning methods attempt to integrate unlabeled vibration spectra and airpath parameter data, under extremely small sample sizes (labeling rates <5%), the prediction results of existing deep neural networks still fail to meet aircraft engine safety requirements. Furthermore, traditional aircraft engine RUS prediction methods based on deep neural networks require further research in terms of interpretability and uncertainty. Due to the black-box nature of deep neural networks, their hidden layer features lack an explicit correlation with the evolutionary trends of aircraft engine degradation data, resulting in a need for improved RUS prediction results. Furthermore, the degradation process of aircraft engines is significantly affected by random factors (such as fluctuations in the blockage rate of cooling film holes and gust loads), and traditional point estimation prediction methods are unable to quantify the confidence interval of the remaining useful life. Therefore, it is of great significance to study interpretable aircraft engine remaining useful life prediction methods from the perspective of semi-supervised learning and to achieve uncertainty measurement of the degradation process. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of low accuracy, poor interpretability and strong uncertainty in the prediction of the remaining useful life of aircraft engines under semi-supervised learning conditions, and to propose a semi-supervised aircraft engine remaining useful life prediction method based on a serialized variational autoencoder.

[0006] The specific process of the semi-supervised aircraft engine remaining useful life prediction method based on serialized variational autoencoder is as follows:

[0007] Step 1: Collect multi-source sensor monitoring data of aircraft engines under various service conditions, pre-process the data, and obtain a training sample set;

[0008] Step 2: Build a supervised loss function J based on the self-attention mechanism-serialized variational autoencoder network s ;

[0009] Step 3: Build a supervised self-attention mechanism-sequential variational autoencoder network;

[0010] Step 4: Build the loss function J based on the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions u ;

[0011] Step 5: Build an unsupervised self-attention mechanism-sequential variational autoencoder network;

[0012] Step 6. Loss function J of the self-attention mechanism-sequential variational autoencoder network under supervised conditions built in step 2 s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network;

[0013] Step 7. Based on the supervised self-attention mechanism-serialized variational autoencoder network and the unsupervised self-attention mechanism-serialized variational autoencoder network, a semi-supervised serialized variational autoencoder network is constructed; the semi-supervised serialized variational autoencoder network is trained based on the training sample set obtained in step 1 until the loss function J of the semi-supervised serialized variational autoencoder network converges, thereby obtaining a trained semi-supervised serialized variational autoencoder network;

[0014] Step 8: After the online data collected by the aircraft engine sensor device undergoes data preprocessing in step 1, it is input into the semi-supervised sequential variational autoencoder network trained in step 7. The trained semi-supervised sequential variational autoencoder network realizes the online prediction of the remaining service life of the aircraft engine.

[0015] The beneficial effects of the present invention are:

[0016] (1) Aiming at the low accuracy of remaining useful life prediction of traditional semi-supervised learning methods, the present invention targets the multi-scale correlation characteristics of time series degradation features. The present invention designs a serialized time series modeling method, introduces a self-attention mechanism to dynamically focus on key degradation features in different time windows, and combines a bidirectional gated time series feature extractor to synchronously capture the implicit correlation patterns in the forward / reverse degradation trends, effectively enhancing the time series model representation capability of the aircraft engine degradation process and significantly optimizing the remaining useful life prediction accuracy.

[0017] (2) From the perspective of probabilistic deep generative networks, this paper proposes a hybrid modeling approach that integrates the variational Bayesian inference framework with dynamic system state estimation theory. By modeling the probabilistic manifold distribution of degradation trajectories in latent variable space, confidence interval assessment results for the remaining useful life prediction value are generated.

[0018] (3) Based on the hybrid supervision paradigm, the present invention constructs a semi-supervised architecture that collaboratively optimizes self-attention weight allocation and serialized variational inference, and proposes a semi-supervised network based on the attention mechanism-serialized variational autoencoder, which can simultaneously utilize labeled and unlabeled data samples to learn valuable feature representations from a large amount of unlabeled degradation data, and can achieve high-precision remaining useful life prediction even when the annotation ratio is low.

[0019] (4) In order to solve the problem of poor interpretability of traditional semi-supervised learning methods, the present invention proposes a modeling method based on the combination of self-attention mechanism and serialized variational reasoning. The self-attention layer is used to dynamically assign the importance weights of feature variables in different time windows, and the contribution of feature variables at each moment to the evolution of aircraft engine degradation state and life prediction is clarified; at the same time, combined with the serialized latent variable modeling method, a degradation trajectory probability manifold with clear physical interpretation is constructed through variational reasoning, which explicitly reveals how the latent variable sequence drives the evolution trend of the aircraft engine degradation process, significantly enhancing the transparency and interpretability of the model.

[0020] (5) To address the problems of strong uncertainty and difficulty in quantification in existing methods, the present invention adopts a probabilistic deep generative network framework, models the probability distribution of degradation trajectories in the latent variable space through a serialized variational autoencoder, explicitly predicts the probability distribution parameters (mean and variance) of the remaining useful life, and generates a confidence interval for the remaining useful life prediction value, thereby effectively quantifying the model prediction uncertainty caused by service environment fluctuations, data acquisition errors and random disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a workflow diagram of the present invention; Figure 2 This is the structure diagram of the self-attention mechanism of the present invention; Figure 3a This is a result diagram of the remaining service life prediction value of the present invention under working condition FD002; Figure 3b This is a result diagram of the actual value of the remaining service life of the present invention under working condition FD004; Figure 4 The actual value, predicted value and corresponding confidence interval result diagram of the remaining useful life of the aircraft turbofan engine No. 24 under the FD001 working condition; Figure 5 This is the phase space trajectory diagram of the degradation process of the aircraft turbofan engine numbered 24 under the FD001 working condition. DETAILED DESCRIPTION

[0022] Specific embodiment 1: The specific process of the semi-supervised aircraft engine remaining service life prediction method based on the serialized variational autoencoder in this embodiment is as follows:

[0023] Step 1: Collect multi-source sensor monitoring data of aircraft engines under various service conditions, pre-process the data, and obtain a training sample set;

[0024] Step 2: Build a supervised loss function J based on the self-attention mechanism-serialized variational autoencoder network s ;

[0025] Step 3: Build a supervised self-attention mechanism-sequential variational autoencoder network;

[0026] Step 4: Build the loss function J based on the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions u ;

[0027] Step 5: Build an unsupervised self-attention mechanism-sequential variational autoencoder network;

[0028] Step 6. Loss function J of the self-attention mechanism-sequential variational autoencoder network under supervised conditions built in step 2 s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network;

[0029] Step 7. Based on the supervised self-attention mechanism-serialized variational autoencoder network and the unsupervised self-attention mechanism-serialized variational autoencoder network, a semi-supervised serialized variational autoencoder network is constructed; the semi-supervised serialized variational autoencoder network is trained based on the training sample set obtained in step 1 until the loss function J of the semi-supervised serialized variational autoencoder network converges, thereby obtaining a trained semi-supervised serialized variational autoencoder network;

[0030] Step 8: After the online data collected by the aircraft engine sensor device undergoes data preprocessing in step 1, it is input into the semi-supervised sequential variational autoencoder network trained in step 7, and the trained semi-supervised sequential variational autoencoder network performs online prediction of the aircraft engine's remaining useful life;

[0031] Evaluate the remaining useful life prediction performance.

[0032] It should be noted that steps 1 to 8 belong to the offline network training phase, and step 9 belongs to the online network prediction phase.

[0033] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that, in step 1, multi-source sensor monitoring data of the aircraft engine under various service conditions is collected, and the data is preprocessed to obtain a training sample set; the specific process is as follows:

[0034] Step 11: Collect multi-source sensor monitoring data of the aircraft engine under various service conditions;

[0035] Step 1 and 2: Given that some sensor channels exhibit approximately constant values ​​throughout the engine's lifecycle, a low-variance feature filtering strategy is used to remove redundant dimensions from the multi-source sensor monitoring data of aircraft engines under various service conditions collected in step 1, significantly improving data characterization efficiency. The specific process is as follows:

[0036] Calculate the variance of each sensor channel data of the aircraft engine throughout its life cycle. If the variance of a sensor channel data is less than a set threshold (such as 0.05), it is determined to be a low-variance feature and the corresponding sensor channel data is removed.

[0037] Taking the public C-MAPSS dataset used in this invention as an example, each sampling moment of a single aircraft engine contains 21 sensor channel data, corresponding to various physical quantities such as temperature, pressure, speed, flow, etc., and the original data dimension is 21 dimensions.

[0038] Step 13: To address the gradient imbalance problem caused by dimensional differences among multiple sensors, perform maximum-minimum normalization on the sensor monitoring data output from Steps 1 and 2, eliminating redundant dimensions, to ensure consistency in the numerical distribution of different physical quantities (such as temperature gradient, vibration acceleration, and fuel flow), thereby optimizing model training stability.

[0039] Step 14: The present invention proposes a dynamic time series segmentation mechanism, which uses a sliding window function to segment and reorganize the continuous monitoring data normalized in step 13 to obtain a feature sequence to capture the time series correlation characteristics of the degradation data and provide a structured input feature sequence for subsequent time series modeling.

[0040] The obtained feature sequence is used as the training sample set.

[0041] Other steps and parameters are the same as those in the first embodiment.

[0042] Specific embodiment three: This embodiment is different from specific embodiment one or two in that the loss function J based on the self-attention mechanism-serialized variational autoencoder network under supervised conditions is constructed in step two. s ; The specific process is:

[0043] The essence of the data-driven remaining useful life prediction method is to construct a mapping relationship between the feature variable x and the life label y. As an important type of data-driven method, the core goal of the probabilistic generation model is to maximize the conditional probability density p(y|x). For labeled training data, considering the time dependence of degradation data, it is recombined in the form of a sliding time window and then input into the self-attention mechanism-serialized variational autoencoder network. Step 2 considers maximizing p(y in the form of a serialized time window. 1:L |x 1:L ), where L is the time window length, specifically expressed as minimizing the loss function L in formula (1):

[0044] L=-logp(y 1:L |x 1:L )=-log∫p(y 1:L ,φ 1:L |x 1:L )dφ 1:L (1)

[0045] Among them, y 1:L represents the remaining useful life tag sequence of length 1 to L, x 1:L represents a characteristic variable with a length of 1 to L, p(y 1:L |x 1:L ) represents x 1:L y under condition 1:L The probability density distribution, p(y 1:L ,φ 1:L |x 1:L ) represents x 1:L y under condition 1:L and φ 1:L The joint probability density distribution of 1:L Represents a sequence of latent variables in a probabilistic generative model;

[0046] The present invention introduces an importance distribution p(φ 1:L |y 1:L ,x 1:L ) describes the latent variable space φ 1:L , formula (1) is written as follows:

[0047]

[0048] Among them, p θ (φ 1:L |y 1:L ,x 1:L ) represents y 1:L and x 1:L The hidden variable sequence φ under the condition 1:L The probability density distribution of Indicates pθ (φ 1:L |y 1:L ,x 1:L ) under the mathematical expectation value;

[0049] On the basis of formula (2), based on the Jensen inequality of convex functions (which means that the result of the convex function on the weighted sum of probabilities is not greater than the result of each function after the weighted probabilities. In deep generative models, Jensen's inequality is often used to derive variational lower bounds, converting the difficulty of direct optimization of the log-likelihood into a computable lower bound, thereby facilitating model training and inference.), there exists f[E(x)]≤E[f(x)], so the following expression is obtained:

[0050]

[0051] Where E(x) represents the mathematical expectation of x, f[E(x)] represents the mathematical expectation of x as a convex function of the independent variable, f(x) represents the convex function of x as the independent variable, and E[f(x)] represents the mathematical expectation of the convex function f(x) of x as the independent variable; x represents the characteristic variable;

[0052] Let the inequality on the right side of (3) be J, and the derivation process is the same as formula (4):

[0053]

[0054] Among them, y l Represents the remaining service life label at the lth moment, y 1:l-1 represents the remaining useful life tag sequence of length 1 to l-1, φ 1:l-1 represents a latent variable sequence of length 1 to l-1, represents y 1:l-1 、φ 1:l-1 、x 1:L The probability density distribution of the remaining useful life label at the lth moment under the joint condition, φ l represents the latent variable at the lth moment, Represents φ 1:l-1 、y 1:l-1 、x 1:L The probability density distribution of the latent variable at the lth moment under the joint condition, pθ(φ l |φ 1:l-1 ,y 1:L ,x 1:L ) represents the conditional probability, l represents the lth moment, L represents the total length of the time window, represents pθ(φ l |φ 1:l-1 ,y 1:L ,x 1:L) is the mathematical expectation calculation formula under the conditions of , J represents the value on the right side of formula (3);

[0055] Formula (4) can be further decomposed into the form described in formula (5):

[0056]

[0057] in, Indicates p θ (φ 1:l-1 |y 1:L ,x 1:L ) under the condition of mathematical expectation calculation formula;

[0058] Indicates p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) under the condition of mathematical expectation calculation formula;

[0059] Combined with the mathematical properties of KL divergence (the mathematical properties of KL divergence are reflected in its non-negativity: that is, the KL divergence between any two probability distributions is always greater than or equal to zero, and is zero only when the two distributions are completely consistent. This property ensures that the optimization problem with KL divergence as the target always has a lower bound, which helps to achieve effective training and convergence guarantee of the probability model), formula (5) is expressed in the form described by formula (6):

[0060]

[0061] in, represents the probability distribution p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) and probability distribution The KL divergence between

[0062] J s Represents the loss function of the supervised self-attention mechanism-sequential variational autoencoder network;

[0063] From formula (6), we can see that the loss function J of the entire deep neural network is s It can be split into a negative logarithmic term and a KL divergence distance term, represented by J1 and J2 respectively as follows:

[0064]

[0065] Among them, J1 represents the negative logarithmic loss function value, and J2 represents the KL divergence distance loss function value;

[0066] The above derivation process is crucial for guiding the construction of deep neural network models. Therefore, the present invention first designs the loss function of the serialized variational autoencoder. It can be clearly seen that the construction of the probabilistic deep generative network specifically includes three sub-networks, namely the inference sub-network, the encoding sub-network and the decoding sub-network; in order to better describe the uncertainty in the remaining service life prediction process, combined with the properties of the probabilistic deep generative network, the output of each sub-network corresponds to a Gaussian distribution, and the corresponding mean and variance are output. In formula (6), p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ), As shown in formulas (9)-(11) respectively:

[0067] Prior / Encoding Subnetwork:

[0068] Posterior / inference subnetwork:

[0069] Decoding subnetwork:

[0070] Represent the mean and variance of the prior sub-network respectively; Represent the mean and variance of the posterior sub-network respectively; represent the mean and variance of the decoding sub-network respectively; N(,) represents Gaussian distribution.

[0071] It is worth mentioning that this paper explores the correspondence between a Bayesian filtering method and the state-space model in control theory. Specifically, the posterior subnetwork corresponds to the state transition equation in the state-space model, and the decoding subnetwork corresponds to the measurement equation in the state-space model.

[0072] For labeled training data, the present invention takes into account the time dependence of degraded data and recombines it in the form of a sliding time window before inputting it into a self-attention mechanism-sequential variational autoencoder network. The conditional probability density function is maximized in the form of a sequential time window. Based on the Jensen inequality and Bayesian formula of convex functions, combined with the mathematical properties of KL divergence, the loss function of the entire deep neural network is split into a negative logarithmic term and a KL divergence distance term. The present invention first designs the loss function of the sequential variational autoencoder and divides the probabilistic deep generative network into an inference subnetwork, an encoding subnetwork, and a decoding subnetwork.

[0073] Other steps and parameters are the same as those in the first or second embodiment.

[0074] Specific embodiment 4: This embodiment differs from any one of specific embodiments 1 to 3 in that, in step 3, a supervised self-attention mechanism-serialized variational autoencoder network is constructed;

[0075] The self-attention mechanism-serialized variational autoencoder network under supervised conditions inputs the feature variable x at the lth moment in the time window 1:L ;The output is formula (17), (18), (19);

[0076] The specific process is:

[0077] This step targets the characteristic variable x input by the network 1:L and the corresponding remaining useful life label y 1:L First, a bidirectional gated recurrent unit based on the self-attention mechanism is designed for encoding. The reason for using the self-attention mechanism network is to focus on the key parts of the input time window data by designing the self-attention layer, thereby improving the sensitivity to key time points. Figure 2 It is a structural diagram of the self-attention mechanism of the present invention;

[0078] Step 3.1. Obtain the characteristic variable x at the lth moment in the time window l With the characteristic variable x l Importance of α l The mapping relationship is shown in expression (12):

[0079]

[0080] Among them, x l represents the characteristic variable at the lth moment (the data collected by the sensors in the aircraft engine in step 14), w l and b l are the weight value and bias value at the lth moment respectively;

[0081] Step 32: Based on the characteristic variable x at the lth moment in the time window obtained in step 31 l and importance α l , the output of the forward self-attention mechanism layer is obtained by matrix multiplication The expression is shown in formula (13):

[0082]

[0083] in, is the characteristic variable of the first moment output from the self-attention mechanism layer, is the characteristic variable of the second moment output by the self-attention mechanism layer, is the characteristic variable of the lth moment output from the self-attention mechanism layer, is the feature variable at the Lth moment output by the self-attention mechanism layer; α1 is the importance of the feature variable x1, α2 is the importance of the feature variable x1, α l is the characteristic variable x l The importance of α L is the characteristic variable x L Importance; x1 is the characteristic variable at the first moment, x2 is the characteristic variable at the second moment, x l is the characteristic variable at the lth moment, x L is the characteristic variable at the Lth moment;

[0084] Step 3. The output of the forward self-attention mechanism layer Input to the forward gated recurrent unit network, the forward gated recurrent unit network GRU outputs the feature encoding matrix As shown in formula (14):

[0085] The gated recurrent unit (GRU) is a variant of a recurrent neural network that features two gating structures: an update gate and a reset gate. These structures help control the flow of temporal information. The reset gate controls the degree to which new input information is combined with past information. The update gate controls the degree to which the network unit retains past information.

[0086]

[0087] in, Indicates that the output of the forward self-attention mechanism layer Input to the gated recurrent unit network; is the feature encoding matrix output by the forward gated recurrent unit network; GRU represents the gated recurrent unit network;

[0088] Step 3 and 4: In order to provide more comprehensive time series information so that the model can better mine the implicit information of time window data, the present invention constructs a backward self-attention mechanism gated recurrent unit network;

[0089] Will According to the characteristic variables from the last moment The characteristic variable at the first moment The sequential input of the backward gated recurrent unit network GRU, the backward gated recurrent unit network GRU outputs the feature encoding matrix;

[0090] The forward and backward directions of the forward gated recurrent unit network GRU and the backward gated recurrent unit network GRU are for Whether the characteristic variables are described from front to back or from back to front;

[0091] Step 35: Combine the feature coding matrix output from step 34 and the feature coding matrix output from step 33 formula (14) Perform splicing;

[0092] Step 36: Combine the concatenated feature encoding matrix and importance level β j The output of the backward self-attention mechanism layer is obtained by matrix multiplication; where the importance degree β j Expressed in the form described by formula (15):

[0093]

[0094] Among them, β j represents the importance of the concatenated encoding matrix; d represents the hidden layer dimension of the forward gated recurrent unit output; w j and b j Respectively represent the corresponding weight value and bias value of the j-th dimension of the splicing matrix;

[0095] express and The j-th dimension of the concatenated matrix;

[0096] Step 37: Based on step 36, obtain the characteristic variables (h1 and are all feature variables) feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism As shown in formula (16):

[0097]

[0098] in, Represents the feature encoding matrix; h1 represents the feature encoding matrix output by the forward gated recurrent unit network The first dimension component of h2 represents the feature encoding matrix output by the forward gated recurrent unit network The second dimension component, h d Represents the feature encoding matrix of the forward gated recurrent unit network output d-th dimension component of ; β1 represents the first dimension component of the importance of the coding matrix after splicing, β2 represents the second dimension component of the importance of the coding matrix after splicing, β d The d-th dimension component representing the importance of the concatenated coding matrix, β d+1 The d+1th dimension component representing the importance of the concatenated coding matrix, β d+2 The d+2th dimension component representing the importance of the concatenated coding matrix, β d+L The d+Lth dimension component representing the importance of the concatenated coding matrix; Represents the feature variable at the first moment output by the self-attention mechanism layer, Represents the feature variable at the second moment output by the self-attention mechanism layer, Represents the feature variable at the Lth moment output by the self-attention mechanism layer;

[0099] Step 38: Remaining service life label y 1:L Obtain the label encoding matrix through steps 31, 32, and 33.

[0100] Step 39: Remaining service life label y 1:L Obtain the label encoding matrix through steps 31, 32, 33, 34, 35, 36, and 37

[0101] and They correspond to the label encoding under the prior and posterior conditions respectively. Under the prior condition, y 1:l The encoding only considers the sequence before the time window l, and the sequence at time l and after time l needs to be masked;

[0102] and Corresponding to the feature encoding matrix under the prior and posterior conditions, respectively, under the prior condition x 1:l The encoding only considers the sequence before the time window l, and the sequence at time l and after time l needs to be masked;

[0103] The encoding under the prior condition only considers the labels before the time window l, and the labels of the subsequent moments need to be masked. This means that when encoding the features or labels of the current moment (the lth moment), the prior encoding network can only use the current and previous label information, and cannot use the labels of future moments in advance.

[0104] Therefore, the labels corresponding to future moments (i.e. l+1, l+2... until the end of the time window) need to be "masked" during encoding, that is, the information of these future labels is blocked in the input so that the network cannot use future data for inference, so as to ensure the temporal causal consistency of the model and prevent "information leakage".

[0105] Step 30: Obtain the feature encoding matrix and label encoding matrix On the basis of , the inference model is described based on the Gaussian distribution network GussianNet, which is expressed as described in formula (17):

[0106]

[0107] in, represents the mean of the posterior sub-network, Represents the variance of the posterior subnetwork; GussianNet (1) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents the feature encoding matrix, represents the label encoding matrix;

[0108] Based on the feature encoding matrix and label encoding matrix Substitute into formula (17) to obtain the mean of the posterior sub-network and variance

[0109] right and Perform Gaussian distribution sampling to obtain the latent variable sequence φ 1:l ; The process is:

[0110] Sample random noise ε from the standard normal distribution N(0,1) l , and then calculate the latent variables

[0111] Among them, φ 1:l represents a latent variable sequence of length 1 to l, φ l represents the latent variable sequence at the lth moment;

[0112] Perform the above operation at each moment in the time window l to obtain the latent variable sequence φ 1:l (used in formula 19);

[0113] GussianNet is based on a multi-layer perceptron, which uses a fully connected layer to map the encoding matrix to the mean and standard deviation of the latent variable estimates.

[0114] Step 31: Posterior / inference subnetwork based on formula (10) and the mean of formula (17) variance Get the conditional probability p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L );

[0115] Step 32: To achieve the forward prediction of the implicit variables as described in formula (10), similar to the state transition equation model in state space theory;

[0116] Design a Gaussian distribution network to realize the latent variable sequence φ1:l-1 , feature encoding matrix and label encoding matrix and the mean of the prior subnetwork and variance The mapping relationship of , and the mean of the prior sub-network is obtained and variance It is expressed as follows:

[0117]

[0118] Among them, GussianNet (2) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP;

[0119] The prior subnetwork of expression (9) The mean of

[0120] The prior subnetwork of expression (9) variance;

[0121] Based on the latent variable sequence φ 1:l-1 , feature encoding matrix and label encoding matrix Substitute into formula (18) to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence φ l , and then obtain the new latent variable sequence φ 1:l .

[0122] Based on the latent variable sequence φ 1:l , feature encoding matrix and label encoding matrix Substitute into formula (18) to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence φ l+1 , and then obtain a new latent variable sequence φ 1:l+1 ;

[0123] Repeat until the mean of the prior sub-network is obtained and variance

[0124] Step 33: Prior sub-network based on formula (9) and the mean in formula (18) and variance get

[0125] calculate and the conditional probability p obtained in step 31 θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ), and calculate the KL divergence distance loss function value based on the KL divergence

[0126] Step 34: Design a Gaussian distribution network to implement the latent variable sequence φ 1:l , feature encoding matrix and label encoding matrix and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows:

[0127]

[0128] in, represents the mean, Represents variance, GussianNet (3) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; φ 1:l represents a latent variable sequence of length 1 to l;

[0129] Based on the feature encoding matrix Label encoding matrix and φ 1:1 , put it into formula (19) to obtain the mean of the posterior sub-network and variance

[0130] Based on the feature encoding matrix Label encoding matrix and φ 1:2 , put it into formula (19) to obtain the mean of the posterior sub-network and variance

[0131] Until,

[0132] Based on the feature encoding matrix Label encoding matrix and φ 1:L , put it into formula (19) to obtain the mean of the posterior sub-network and variance

[0133] Step 35: Decoding subnetwork based on formula (11) and the mean in formula (19) and standard deviation Obtain the posterior probability conditional on

[0134] Step 36: Based on the posterior probability condition Calculate the negative logarithmic loss function value

[0135] Step 37: Add the negative logarithmic loss function value J1 calculated in step 36 and the KL divergence distance loss function value J2 calculated in step 33 to obtain the loss function J of the entire self-attention mechanism-serialized variational autoencoder network in formula (6): s , for parameter update of the entire deep neural network.

[0136] The other steps and parameters are the same as those in the first to third embodiments.

[0137] Specific embodiment 5: The difference between this embodiment and one of the specific embodiments 1 to 4 is that the loss function J based on the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions is constructed in step 4. u ; The specific process is:

[0138] Loss function of self-attention mechanism-sequential variational autoencoder network under unsupervised conditionsJ u It is expressed as described in formula (23):

[0139]

[0140] in, express The mathematical expectation calculation formula under the conditions is: express The mathematical expectation calculation formula under the conditions, D KL Represents the KL divergence calculation formula;

[0141] The unsupervised self-attention-based sequential variational autoencoder network has roughly the same network form as the supervised deep network. Unlike the non-causal modeling form in supervised learning, the unsupervised model pays more attention to the data generation mechanism, attempts to sample from the feature distribution of degraded data, simulates the data generation process, reconstructs the features of degraded data, and improves the utilization value of unlabeled degraded data. On this basis, for unlabeled training data, the encoding subnetwork, inference subnetwork, and decoding subnetwork need to be reorganized into the form described by formulas (20)-(22);

[0142] In formula (23) As shown in formulas (20)-(22) respectively:

[0143] Prior / Encoding Subnetwork:

[0144] Posterior / inference subnetwork:

[0145] Decoding subnetwork:

[0146] in, express and Joint unsupervised The probability density function of represents the latent variable at the lth moment under unsupervised conditions; represents a latent variable sequence of length 1 to l-1 under unsupervised conditions; Represents an input feature sequence of length 1 to l-1 under unsupervised conditions; express Joint unsupervised The probability density distribution of represents a latent variable sequence of length 1 to L under unsupervised conditions; express Joint unsupervised The probability density distribution of ; N(,) represents Gaussian distribution; Represent the mean and variance of the prior sub-network respectively; Represent the mean and variance of the posterior sub-network respectively; Represent the mean and variance of the decoding sub-network respectively;

[0147] In order to distinguish it from the parameter representation under supervised conditions, the feature variables and latent variables under unsupervised conditions are uniformly expressed using and To express;

[0148] It's important to note that the principles of the prior and posterior subnetworks are the same as those of supervised learning networks, except that the remaining useful life label is no longer used as the encoded input of the network, but is replaced by the feature variable. Furthermore, the output of the decoding subnetwork under unsupervised conditions is also replaced by the feature variable itself.

[0149] The other steps and parameters are the same as those in the first to fourth embodiments.

[0150] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that in step 5, a self-attention mechanism-serialized variational autoencoder network is constructed under unsupervised conditions; the specific process is as follows:

[0151] The self-attention mechanism-serialized variational autoencoder network under unsupervised conditions inputs the time window data x of length L. 1:L ;The output is formula (20), (21), (22);

[0152] Step 5.1. Obtain the characteristic variable x at the lth moment in the time window l With the characteristic variable x l Importance of α l The mapping relationship is shown as follows:

[0153]

[0154] Among them, x l represents the characteristic variable at the lth moment (the data collected by the sensors in the aircraft engine in step 14), w l and b l are the weight value and bias value at the lth moment respectively;

[0155] Step 52: Based on the characteristic variable x at the lth moment in the time window obtained in step 51 l and importance α l , the output of the forward self-attention mechanism layer is obtained by matrix multiplication The expression is as follows:

[0156]

[0157] in, is the feature variable of the first moment output from the self-attention mechanism layer, is the characteristic variable of the second moment output by the self-attention mechanism layer, is the characteristic variable of the lth moment output from the self-attention mechanism layer, is the feature variable at the Lth moment output by the self-attention mechanism layer; α1 is the importance of the feature variable x1, α2 is the importance of the feature variable x1, α l is the characteristic variable x l The importance of α L is the characteristic variable x L Importance; x1 is the characteristic variable at the first moment, x2 is the characteristic variable at the second moment, x l is the characteristic variable at the lth moment, x L is the characteristic variable at the Lth moment;

[0158] Step 53: The output of the forward self-attention mechanism layer Input to the forward gated recurrent unit network, the forward gated recurrent unit network GRU outputs the feature encoding matrix As shown in the following formula:

[0159] The gated recurrent unit (GRU) is a variant of a recurrent neural network that features two gating structures: an update gate and a reset gate. These structures help control the flow of temporal information. The reset gate controls the degree to which new input information is combined with past information. The update gate controls the degree to which the network unit retains past information.

[0160]

[0161] in, Indicates that the output of the forward self-attention mechanism layer Input to the gated recurrent unit network; is the feature encoding matrix output by the forward gated recurrent unit network; GRU represents the gated recurrent unit network;

[0162] In order to provide more comprehensive time series information so that the model can better mine the implicit information of time window data, the present invention constructs a backward self-attention mechanism gated recurrent unit network;

[0163] Step 54: According to the characteristic variables from the last moment The characteristic variable at the first moment The sequential input of the backward gated recurrent unit network GRU, the backward gated recurrent unit network GRU outputs the feature encoding matrix;

[0164] Step 55: Combine the feature coding matrix output from step 54 and the feature coding matrix output from step 523 Perform splicing;

[0165] Step 5 and 6: Combine the concatenated feature encoding matrix and importance level β j Obtain the output of the backward self-attention mechanism layer through matrix multiplication;

[0166] Among them, the importance degree β j It is expressed in the form described by the following formula:

[0167]

[0168] Among them, β j represents the importance of the concatenated encoding matrix; d represents the hidden layer dimension of the forward gated recurrent unit output; w j and b jRespectively represent the corresponding weight value and bias value of the j-th dimension of the splicing matrix; express and The j-th dimension of the concatenated matrix;

[0169] Step 57: Based on step 56, obtain the characteristic variables (h1 and are all feature variables) feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism As shown below:

[0170]

[0171] in, Represents the feature encoding matrix; h1 represents the feature encoding matrix output by the forward gated recurrent unit network The first dimension component of h2 represents the feature encoding matrix output by the forward gated recurrent unit network The second dimension component, h d Represents the feature encoding matrix of the forward gated recurrent unit network output d-th dimension component of ; β1 represents the first dimension component of the importance of the coding matrix after splicing, β2 represents the second dimension component of the importance of the coding matrix after splicing, β d The d-th dimension component representing the importance of the concatenated coding matrix,

[0172] β d+1 The d+1th dimension component representing the importance of the concatenated coding matrix, β d+2 The d+2th dimension component representing the importance of the concatenated coding matrix, β d+L The d+Lth dimension component representing the importance of the concatenated coding matrix; Represents the feature variable at the first moment output by the self-attention mechanism layer, Represents the feature variable at the second moment output by the self-attention mechanism layer, Represents the feature variable at the Lth moment output by the self-attention mechanism layer;

[0173] Step 58: Obtaining the feature encoding matrix On this basis, the inference model is described based on the Gaussian distribution network GussianNet, which is expressed as follows:

[0174]

[0175] in, represents the mean of the posterior sub-network, represents the variance of the posterior sub-network; Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents the feature encoding matrix;

[0176] Based on the feature encoding matrix Substitute into formula (29) to obtain the mean of the posterior sub-network and variance

[0177] pass and Perform Gaussian distribution sampling to obtain the latent variable sequence The process is:

[0178] Sample random noise ε from the standard normal distribution N(0,1) l-1 , and then calculate the latent variables

[0179] in, represents a latent variable sequence of length 1 to l-1, represents the latent variable sequence at the l-1th moment;

[0180] Perform the above operation at each moment with a length of 1 to l-1 in the time window L to obtain the hidden variable sequence (used in formula (31));

[0181] GussianNet is based on a multi-layer perceptron, which uses a fully connected layer to map the encoding matrix to the mean and standard deviation of the latent variable estimates.

[0182] Step 59: Based on and the mean of the above formula variance Get conditional probability

[0183] Step 50: To achieve forward prediction of implicit variables, similar to the state transition equation model in state space theory;

[0184] Designing Gaussian distribution network to realize latent variable sequence Feature variable sequence and the mean of the prior subnetwork and variance The mapping relationship of , and the mean of the prior sub-network is obtained and variance Expressed in the form described below:

[0185]

[0186] in, Represents a Gaussian distribution network, which is a multi-layer perceptron MLP;

[0187] The prior subnetwork of expression (20) The mean of

[0188] The prior subnetwork of expression (20) variance;

[0189] Based on latent variable sequence Input feature sequence Substitute the above formula to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence Then obtain a new latent variable sequence

[0190] Based on latent variable sequence Input feature sequence Substitute the above formula to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence Then obtain a new latent variable sequence

[0191] Repeat until the mean of the prior sub-network is obtained and variance

[0192] Step 51: Based on and mean variance Get conditional probability

[0193] Step 52: Design a Gaussian distribution network to implement latent variable sequence Feature variables and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows:

[0194]

[0195] in, represents the mean, represents the variance, Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents a latent variable sequence of length 1 to l-1;

[0196] Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance

[0197] Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance

[0198] Until,

[0199] Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance

[0200] Step 53: Based on and mean variance Obtain the posterior probability conditional on

[0201] Step 54: Based on Calculate the loss function J of the self-attention mechanism-sequential variational autoencoder network under unsupervised conditions u ;

[0202]

[0203] The other steps and parameters are the same as those in the first to fifth embodiments.

[0204] Specific embodiment seven: This embodiment differs from any one of the specific embodiments one to six in that the loss function J based on the self-attention mechanism-serialized variational autoencoder network under the supervised condition built in step two in step six is s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network; the specific process is:

[0205] Given that aircraft engines are required to operate in a healthy state, acquiring degradation data and corresponding remaining useful life labels throughout their entire lifecycle is typically expensive and time-consuming. However, traditional data-driven approaches require massive amounts of aircraft engine degradation samples for supervised model training, making it difficult to fully utilize unlabeled data. Therefore, this paper investigates a method for remaining useful life prediction from a semi-supervised perspective, utilizing both labeled and unlabeled data samples.

[0206] For labeled training data and unlabeled training data The labeled training data is trained using the content in steps 2 and 3.

[0207] Loss function J based on the self-attention mechanism-sequential variational autoencoder network under supervised conditions s And the loss function J based on the self-attention mechanism-sequential variational autoencoder network under unsupervised conditions u , construct the loss function of the self-attention mechanism-serialized variational autoencoder network under semi-supervised conditions; as shown in formula (33):

[0208]

[0209] The distributions described by the supervised prior subnetwork and the posterior subnetwork in formula (21) are consistent, which means that the unsupervised model and the supervised model use the same structural modeling conditional distribution, and network parameters can be shared between the two tasks. In this sense, a large amount of unlabeled degradation data can be well utilized, thereby improving the generalization performance and efficiency of the model.

[0210] The other steps and parameters are the same as those in the first to sixth embodiments.

[0211] Specific embodiment eight: This embodiment differs from one of specific embodiments one to seven in that, in step seven, a semi-supervised sequential variational autoencoder network is constructed based on a self-attention mechanism-sequential variational autoencoder network under supervised conditions and a self-attention mechanism-sequential variational autoencoder network under unsupervised conditions; the semi-supervised sequential variational autoencoder network is trained based on the training sample set obtained in step one until the loss function J of the semi-supervised sequential variational autoencoder network converges, thereby obtaining a trained semi-supervised sequential variational autoencoder network; the specific process is:

[0212] The input of the self-attention mechanism-sequential variational autoencoder network under semi-supervised conditions is the feature variable x from the 1st moment to the Lth moment in the time window. 1:L ;

[0213] The characteristic variable x at the lth moment in the time window l After step 3, the prior / encoding sub-network, the posterior / inference sub-network, and the decoding sub-network of step 3 output probability distributions respectively;

[0214] The characteristic variable x at the lth moment in the time window l After step 5, the prior / encoding sub-network, the posterior / inference sub-network, and the decoding sub-network of step 5 respectively output probability distributions;

[0215] Repeat step 7 until the loss function of the semi-supervised self-attention mechanism-serialized variational autoencoder network converges, and a trained semi-supervised serialized variational autoencoder network is obtained.

[0216] The other steps and parameters are the same as those in the first to seventh embodiments.

[0217] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that in step 8, the online data collected by the aircraft engine sensor device is pre-processed in step 1 and then input into the semi-supervised sequential variational autoencoder network trained in step 7. The trained semi-supervised sequential variational autoencoder network then performs online prediction of the aircraft engine's remaining useful life. The specific process is as follows:

[0218] The online data collected by the aircraft engine sensor device is preprocessed in step 1. The preprocessed data is processed through steps 31, 32, 33, 34, 35, 36, and 37 to obtain the feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism.

[0219] In obtaining the feature encoding matrix Based on the Gaussian distribution network GussianNet, the inference model is described and the mean of the posterior sub-network is obtained. and the variance of the posterior subnetwork The expression is:

[0220]

[0221] in, represents the mean of the posterior sub-network, represents the variance of the posterior sub-network;

[0222] GussianNet (1) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP;

[0223] represents the feature encoding matrix;

[0224] Based on the feature encoding matrix Substitute into formula (17) to obtain the mean of the posterior sub-network and variance

[0225] right and Perform Gaussian distribution sampling to obtain the latent variable sequence φ 1:l ; The process is:

[0226] Sample random noise ε from the standard normal distribution N(0,1) l , and then calculate the latent variables

[0227] Among them, φ 1:l represents a latent variable sequence of length 1 to l, φ l represents the latent variable sequence at the lth moment;

[0228] Perform the above operation at each moment in the time window l to obtain the latent variable sequence φ 1:l ;

[0229] Design a Gaussian distribution network to realize the latent variable sequence φ 1:l , feature encoding matrix and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows:

[0230]

[0231] in, represents the mean, Represents variance, GussianNet (3) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; φ 1:l represents a latent variable sequence of length 1 to l;

[0232] Based on the feature encoding matrix and φ 1:1 , put it into formula (35) to obtain the mean of the posterior sub-network and variance

[0233] Based on the feature encoding matrix and φ 1:2 , put it into formula (35) to obtain the mean of the posterior sub-network and variance

[0234] Until,

[0235] Based on the feature encoding matrix and φ1:L , put it into formula (35) to obtain the mean of the posterior sub-network and variance

[0236] Based on the decoding sub-network and the mean in formula (35) and standard deviation Obtain the posterior probability conditional on

[0237] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0238] Evaluation of the remaining useful life prediction effect: The root mean square error (RMSE) is used to evaluate the proposed remaining useful life prediction effect. The expression of the root mean square error is as follows:

[0239]

[0240] Among them, N is the number of samples of test data, i is the serial number of the sample, RUL pi and RUL ti are the predicted value and true value of the remaining useful life of the i-th sample, respectively. The smaller the value of the root mean square error, the better the online prediction effect of the remaining useful life proposed by the present invention.

[0241] Example:

[0242] The following examples are used to verify the beneficial effects of the present invention:

[0243] The present invention uses the aircraft engine performance degradation simulation data set (CMAPSS) to verify the effectiveness of the remaining service life prediction model proposed in the present invention. This data set reproduces the complete degradation trajectory of the engine from the initial healthy state to functional failure through multi-dimensional working condition coupling simulation technology (including the dynamic combination of different thrust loads and environmental parameters), and each simulation unit corresponds to a complete flight mission cycle. The data set contains the full life cycle monitoring records of multiple turbofan engines, covering the dynamic monitoring parameters of 26-dimensional sensor channels (such as combustion chamber outlet temperature, compressor outlet pressure, fuel flow, etc.), and comprehensively characterizes the time-varying laws of the key performance indicators of aircraft engines. The CMAPSS data set for each working condition is divided into a training set and a test set: the training set contains multi-source sensor data continuously sampled throughout the entire life cycle of the engine (covering the initial healthy to complete failure stage), and the test set selects degradation process fragments of some operating stages, and the remaining service life prediction value at the end time is used as the evaluation benchmark for the model prediction performance. The detailed parameter configuration is shown in Table 1.

[0244] Table 1 Description of the CMAPSS aeroengine dataset

[0245]

[0246] The implementation steps of the present invention are as follows:

[0247] Step 1, collect multi-source sensor monitoring data of aircraft engines under various service conditions, pre-process the data, and obtain a training sample set: in view of the characteristics of some sensor channels showing approximately constant values ​​throughout the life cycle of the engine, a low-variance feature filtering strategy is used to eliminate redundant dimensions, significantly improving data characterization efficiency. The present invention eliminates the sensor data with constant values ​​of 7 types numbered 1, 5, 6, 10, 16, 18, and 19. In order to solve the gradient imbalance problem caused by the dimensional differences of multiple sensors, the sensor monitoring data with redundant dimensions eliminated is normalized to the maximum-minimum value 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. In addition, the present invention proposes a dynamic time series segmentation mechanism, which uses a sliding window function to segment and reorganize the normalized continuous monitoring signal to obtain a feature sequence to capture the time series correlation characteristics of the degradation data and provide a structured input feature sequence for subsequent time series modeling.

[0248] Step 2: Build a supervised loss function J based on the self-attention mechanism-serialized variational autoencoder network s :For labeled training data, the present invention takes into account the time dependence of the degraded data, recombines it in the form of a sliding time window and inputs it into a self-attention mechanism-serialized variational autoencoder network, maximizes the conditional probability density function in the form of a serialized time window, and based on the Jensen inequality of convex functions and the Bayesian formula, combined with the mathematical properties of KL divergence, splits the loss function of the entire deep neural network into a negative logarithmic term and a KL divergence distance term. The present invention first designs the loss function of the serialized variational autoencoder and divides the probabilistic deep generative network into an inference subnetwork, an encoding subnetwork, and a decoding subnetwork.

[0249] Step 3: Build a supervised self-attention mechanism-sequential variational autoencoder network: The present invention designs a self-attention mechanism-based bidirectional gated recurrent unit for encoding the feature variables and corresponding remaining useful life labels of the network input. The reason for using the self-attention mechanism network is to focus on the key parts of the input time window data by designing the self-attention layer. Based on the obtained feature encoding matrix and label encoding matrix, an inference model under posterior conditions can be constructed. The present invention constructs a Gaussian distribution network based on GussianNet to describe the corresponding inference model. The mapping relationship from the encoding matrix to the estimated mean and standard deviation of the latent variable is realized through a fully connected layer. The Gaussian distribution network is designed to realize the mapping relationship between the latent variable sequence, feature encoding matrix, label encoding matrix and the latent variable mean and standard deviation. Based on this, the KL divergence is calculated and the loss function term is constructed. The loss function of the entire self-attention mechanism-sequential variational autoencoder network is thus obtained, which is used to update the parameters of the entire deep neural network.

[0250] Step 4: Build the loss function J based on the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions u ;

[0251] Step 5: Build an unsupervised self-attention mechanism-sequential variational autoencoder network;

[0252] Step 6: Loss function J of the self-attention mechanism-sequential variational autoencoder network under supervised conditions built in step 2 s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network;

[0253] Step 7. Based on the supervised self-attention mechanism-serialized variational autoencoder network and the unsupervised self-attention mechanism-serialized variational autoencoder network, a semi-supervised serialized variational autoencoder network is constructed; the semi-supervised serialized variational autoencoder network is trained based on the training sample set obtained in step 1 until the loss function J of the semi-supervised serialized variational autoencoder network converges, thereby obtaining a trained semi-supervised serialized variational autoencoder network; the loss function of the semi-supervised serialized variational autoencoder network training is the mean square error loss function, the optimization algorithm is the Adam optimization algorithm, the learning rate is 0.001, and the network training process is carried out in a hardware environment of 1 GPU (GTX 4060 graphics card).

[0254] Step 8: After the online data collected by the aircraft engine sensor device undergoes data preprocessing in step 1, it is input into the semi-supervised sequential variational autoencoder network trained in step 7. The trained semi-supervised sequential variational autoencoder network realizes the online prediction of the remaining service life of the aircraft engine.

[0255] Step 9: Evaluate the remaining useful life prediction effect: Use the root mean square error to evaluate the online remaining useful life prediction effect of the method proposed in the present invention.

[0256] Considering that the degradation data in working conditions FD002 and FD004 are relatively complex, Figure 3a 、 3b The predicted values ​​and true values ​​of the proposed method under the operating conditions FD002 and FD004 are shown respectively. Table 2 shows the comparison of the root mean square error results of the present invention and a variety of existing aircraft engine remaining service life prediction methods. In addition, for the semi-supervised conditions, for each operating condition, the present invention randomly retains training data at different percentages (2%, 5%, 10%, 50% and 100%) to evaluate the remaining service life prediction performance of the deep network model under different amounts of training data availability. The root mean square error experimental results under the entire semi-supervised conditions are summarized in Table 2. From Figure 3a 、 3b As can be seen from Table 2, the technical solution proposed in the present invention can achieve high-precision prediction of the remaining useful life of aircraft engines. The experimental results of remaining useful life prediction under semi-supervised conditions with different percentages of labeled data are shown in Table 3. By comparing the model performance under different percentages of supervised data, when the labeled data accounts for only 2%, that is, when only a small proportion of labeled data is available, the semi-supervised aircraft engine remaining useful life prediction method based on serialized variational autoencoders proposed in the present invention can still achieve good prediction performance.

[0257] Table 2 Comparison of the results of the present invention and various existing aircraft engine remaining service life prediction methods

[0258]

[0259] Table 3 Experimental results of remaining useful life prediction under semi-supervised conditions with different percentages of labeled data

[0260]

[0261] Taking the test data set FD001 as an example, the present invention shows the true value, predicted value and corresponding confidence interval of the remaining service life of the aircraft turbofan engine numbered 24 (the present invention shows it with a 95% confidence interval), as shown in the figure. Figure 4 As shown in , this is used to measure the uncertainty of the degradation model. Figure 4As can be seen from the graph, for larger values ​​of remaining useful life, the confidence intervals for the predictions are wider, indicating that the model is more cautious in its predictions. On the other hand, for smaller values ​​of remaining useful life, the confidence intervals for the predictions are narrower. This suggests that the remaining useful life predictions will become more accurate over time. Figure 5 It shows the phase space trajectory of the degradation process of the aircraft turbofan engine No. 24 under the FD001 working condition, thereby showing the evolution process of the implicit variables in the entire degradation process and improving the interpretability of the neural network.

[0262] The present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A semi-supervised aircraft engine remaining useful life prediction method based on a sequential variational autoencoder, characterized by: The specific process of the method is: Step 1: Collect multi-source sensor monitoring data of aircraft engines under various service conditions, pre-process the data, and obtain a training sample set; Step 2: Build a supervised loss function J based on the self-attention mechanism-serialized variational autoencoder network s ; Step 3: Build a supervised self-attention mechanism-sequential variational autoencoder network; Step 4: Build the loss function J based on the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions u ; Step 5: Build an unsupervised self-attention mechanism-sequential variational autoencoder network; Step 6. Loss function J of the self-attention mechanism-sequential variational autoencoder network under supervised conditions built in step 2 s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network; Step 7. Based on the supervised self-attention mechanism-serialized variational autoencoder network and the unsupervised self-attention mechanism-serialized variational autoencoder network, a semi-supervised serialized variational autoencoder network is constructed; the semi-supervised serialized variational autoencoder network is trained based on the training sample set obtained in step 1 until the loss function J of the semi-supervised serialized variational autoencoder network converges, thereby obtaining a trained semi-supervised serialized variational autoencoder network; Step 8: After the online data collected by the aircraft engine sensor device undergoes data preprocessing in step 1, it is input into the semi-supervised sequential variational autoencoder network trained in step 7. The trained semi-supervised sequential variational autoencoder network realizes the online prediction of the remaining service life of the aircraft engine.

2. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 1 is characterized in that: In the step 1, multi-source sensor monitoring data of the aircraft engine under various service conditions is collected, and the data is pre-processed to obtain a training sample set; The specific process is: Step 11: Collecting aircraft engine sensor monitoring data; Step 12: Use a low-variance feature filtering strategy to remove redundant dimensions from the aircraft engine monitoring data collected in step 11. The specific process is as follows: Calculate the variance of each sensor channel data of the aircraft engine throughout its life cycle. If the variance of a sensor channel data is less than the set threshold, it is determined to be a low-variance feature and the corresponding sensor channel data is removed. Step 13: performing maximum-minimum normalization processing on the sensor monitoring data output from step 12 after removing redundant dimensions; Step 14: Use the sliding window function to segmentally reorganize the continuous monitoring data normalized in step 13 to obtain a feature sequence; The obtained feature sequence is used as the training sample set.

3. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 2 is characterized in that: The loss function J of the self-attention mechanism-serialized variational autoencoder network under supervised conditions is constructed in step 2. s ; The specific process is: For labeled training data, minimize the loss function L in formula (1): L=-logp(y 1:L |x 1:L )=-log∫p(y 1:L ,f 1:L |x 1:L )dφ 1:L (1) Among them, y 1:L represents the remaining useful life tag sequence of length 1 to L, x 1:L represents a characteristic variable with a length of 1 to L, p(y 1:L |x 1:L ) represents x 1:L y under condition 1:L The probability density distribution, p(y 1:L ,φ 1:L |x 1:L ) represents x 1:L y under condition 1:L and φ 1:L The joint probability density distribution of 1:L Represents a sequence of latent variables in a probabilistic generative model; Introducing an importance distribution p(φ 1:L |y 1:L ,x 1:L ), formula (1) is written as follows: Among them, pθ(φ 1:L |y 1:L ,x 1:L ) represents y 1:L and x 1:L The hidden variable sequence φ under the condition 1:L The probability density distribution of Indicates p θ (θ 1:L |y 1:L ,x 1:L ) under the mathematical expectation value; Based on formula (2), there exists f[E(x)]≤E[f(x)], so we have the following expression: Where E(x) represents the mathematical expectation of x, f[E(x] represents the mathematical expectation of x as a convex function of the independent variable, f(x) represents the convex function of x as the independent variable, and E[f(x)] represents the mathematical expectation of the convex function f(x) with x as the independent variable; x represents the characteristic variable; Let the inequality on the right side of (3) be J, and the derivation process is the same as formula (4): Among them, y l Represents the remaining service life label at the lth moment, y 1:l-1 represents the remaining useful life tag sequence of length 1 to l-1, φ 1:l-1 represents a latent variable sequence of length 1 to l-1, represents y 1:l-1 、φ 1:l-1 、x 1:L The probability density distribution of the remaining useful life label at the lth moment under the joint condition, φ l represents the latent variable at the lth moment, Represents φ 1:l-1 、y 1:l-1 、x 1:L The probability density distribution of the latent variable at the lth moment under the joint condition, p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) represents the conditional probability, l represents the lth moment, L represents the total length of the time window, Indicates p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) is the mathematical expectation calculation formula under the conditions of , J represents the value on the right side of formula (3); Formula (4) can be further decomposed into the form described in formula (5): in, Indicates p θ (φ 1:l-1 |y 1:L ,x 1:L ) under the condition of mathematical expectation calculation formula; Indicates p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) under the condition of mathematical expectation calculation formula; Expressing formula (5) into the form described by formula (6): in, represents the probability distribution p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ) and probability distribution The KL divergence between J s Represents the loss function of the supervised self-attention mechanism-sequential variational autoencoder network; From formula (6), we can see that the loss function J of the entire deep neural network is s It can be split into a negative logarithmic term and a KL divergence distance term, represented by J1 and J2 respectively as follows: Among them, J1 represents the negative logarithmic loss function value, and J2 represents the KL divergence distance loss function value; In formula (6) p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ), As shown in formulas (9)-(11) respectively: Prior / Encoding Subnetwork: Posterior / inference subnetwork: Decoding subnetwork: Represent the mean and variance of the prior sub-network respectively; Represent the mean and variance of the posterior sub-network respectively; Represent the mean and variance of the decoding sub-network respectively; N(,) represents Gaussian distribution.

4. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 3 is characterized by: In step 3, a supervised self-attention mechanism-sequential variational autoencoder network is constructed. The specific process is as follows: Step 3.

1. Obtain the characteristic variable x at the lth moment in the time window l With the characteristic variable x l Importance of α l The mapping relationship is shown in expression (12): Among them, x l represents the characteristic variable at the lth moment, w l and b l are the weight value and bias value at the lth moment respectively; Step 32: Based on the characteristic variable x at the lth moment in the time window obtained in step 31 l and importance α l , the output of the forward self-attention mechanism layer is obtained by matrix multiplication The expression is shown in formula (13): in, is the feature variable of the first moment output from the self-attention mechanism layer, is the characteristic variable of the second moment output from the self-attention mechanism layer, is the characteristic variable of the lth moment output from the self-attention mechanism layer, is the feature variable at the Lth moment output by the self-attention mechanism layer; α1 is the importance of feature variable x1, α2 is the importance of feature variable x1, α l is the characteristic variable x l The importance of α L is the characteristic variable x L the importance of x1 is the characteristic variable at the first moment, x2 is the characteristic variable at the second moment, and x l is the characteristic variable at the lth moment, x L is the characteristic variable at the Lth moment; Step 3. The output of the forward self-attention mechanism layer Input to the forward gated recurrent unit network, the forward gated recurrent unit network GRU outputs the feature encoding matrix As shown in formula (14): in, Indicates that the output of the forward self-attention mechanism layer Input to the gated recurrent unit network; is the feature encoding matrix output by the forward gated recurrent unit network; GRU represents the gated recurrent unit network; Steps 3 and 4 Will According to the characteristic variables from the last moment The characteristic variable at the first moment The sequential input of the backward gated recurrent unit network GRU, the backward gated recurrent unit network GRU outputs the feature encoding matrix; Step 35: Combine the feature coding matrix output from step 34 and the feature coding matrix output from step 33 Perform splicing; Step 36: Combine the concatenated feature encoding matrix and importance level β j Obtain the output of the backward self-attention mechanism layer through matrix multiplication; Among them, the importance degree β j Expressed in the form described by formula (15): Among them, β j represents the importance of the concatenated encoding matrix; d represents the hidden layer dimension of the forward gated recurrent unit output; w j and b j Respectively represent the corresponding weight value and bias value of the j-th dimension of the splicing matrix; express and The j-th dimension of the concatenated matrix; Step 37: Based on step 36, obtain the feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism for the feature variable As shown in formula (16): in, represents the feature encoding matrix; h1 represents the feature encoding matrix output by the forward gated recurrent unit network The first dimension component of h2 represents the feature encoding matrix output by the forward gated recurrent unit network The second dimension component, h d Represents the feature encoding matrix of the forward gated recurrent unit network output The d-th dimension component of ; β1 represents the first dimension component of the importance of the coding matrix after splicing, β2 represents the second dimension component of the importance of the coding matrix after splicing, and β d The d-th dimension component representing the importance of the concatenated coding matrix, β d+1 The d+1th dimension component representing the importance of the concatenated coding matrix, β d+2 The d+2th dimension component representing the importance of the concatenated coding matrix, β d+L The d+Lth dimension component representing the importance of the concatenated coding matrix; Represents the feature variable at the first moment output by the self-attention mechanism layer, Represents the feature variable at the second moment output by the self-attention mechanism layer, Represents the feature variable at the Lth moment output by the self-attention mechanism layer; Step 38: Remaining service life label y 1:L Obtain the label encoding matrix through steps 31, 32, and 33. Step 39: Remaining service life label y 1:L Obtain the label encoding matrix through steps 31, 32, 33, 34, 35, 36, and 37 and They correspond to the label encoding under the prior and posterior conditions respectively. Under the prior condition, y 1:l The encoding only considers the sequence before the time window l, and the sequence at time l and after time l needs to be masked; and Corresponding to the feature encoding matrix under the prior and posterior conditions, respectively, under the prior condition x 1:l The encoding only considers the sequence before the time window l, and the sequence at time l and after time l needs to be masked; Step 30 In obtaining the feature encoding matrix and label encoding matrix On the basis of , the inference model is described based on the Gaussian distribution network GussianNet, which is expressed as described in formula (17): in, represents the mean of the posterior sub-network, represents the variance of the posterior sub-network; GussianNet (1) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents the feature encoding matrix, represents the label encoding matrix; Based on the feature encoding matrix and label encoding matrix Substitute into formula (17) to obtain the mean of the posterior sub-network and variance right and Perform Gaussian distribution sampling to obtain the latent variable sequence φ 1:l ; The process is: Sample random noise ε from the standard normal distribution N(0,1) l , and then calculate the latent variables Among them, φ 1:l represents a latent variable sequence of length 1 to l, φ l represents the latent variable sequence at the lth moment; Perform the above operation at each moment in the time window l to obtain the latent variable sequence φ 1:l ; Step 31: Posterior / inference subnetwork based on formula (10) and the mean of formula (17) variance Get the conditional probability p θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ); Step 32: Design a Gaussian distribution network to realize the latent variable sequence φ 1:l-1 , feature encoding matrix and label encoding matrix and the mean of the prior subnetwork and variance The mapping relationship of , and the mean of the prior sub-network is obtained and variance It is expressed as follows: Among them, GussianNet (2) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; The prior subnetwork of expression (9) The mean of The prior subnetwork of expression (9) variance; Based on the latent variable sequence φ 1:l-1 , feature encoding matrix and label encoding matrix Substitute into formula (18) to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence φ l , and then obtain a new latent variable sequence φ 1:l . Based on the latent variable sequence φ 1:l , feature encoding matrix and label encoding matrix Substitute into formula (18) to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence φ l+1 , and then obtain a new latent variable sequence φ 1:l+1 ; Repeat until the mean of the prior sub-network is obtained and variance Step 33: The prior sub-network based on formula (9) and the mean in formula (18) and variance get calculate and the conditional probability p obtained in step 31 θ (φ l |φ 1:l-1 ,y 1:L ,x 1:L ), and calculate the KL divergence distance loss function value based on the KL divergence Step 34: Design a Gaussian distribution network to realize the latent variable sequence φ 1:l , feature encoding matrix and label encoding matrix and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows: in, represents the mean, Represents variance, GussianNet (3) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; φ 1:l represents a latent variable sequence of length 1 to l; Based on the feature encoding matrix Label encoding matrix and φ 1:1 , put it into formula (19) to obtain the mean of the posterior sub-network and variance Based on the feature encoding matrix Label encoding matrix and φ 1:2 , put it into formula (19) to obtain the mean of the posterior sub-network and variance Until, Based on the feature encoding matrix Label encoding matrix and φ 1:L , put it into formula (19) to obtain the mean of the posterior sub-network and variance Step 35 Decoding subnetwork based on formula (11) and the mean in formula (19) and standard deviation Obtain the posterior probability conditional on Step 36 Based on the posterior probability Calculate the negative logarithmic loss function value Step 37 Add the negative logarithmic loss function value J1 calculated in step 36 and the KL divergence distance loss function value J2 calculated in step 33 to obtain the loss function J of the entire self-attention mechanism-serialized variational autoencoder network. s .

5. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 4 is characterized in that: The loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions is constructed in step 4. u ; The specific process is: Loss function of self-attention mechanism-sequential variational autoencoder network under unsupervised conditionsJ u It is expressed as described in formula (23): in, express The mathematical expectation calculation formula under the conditions is: express The mathematical expectation calculation formula under the conditions, D KL Represents the KL divergence calculation formula; In formula (23) As shown in formulas (20)-(22) respectively: Prior / Encoding Subnetwork: Posterior / inference subnetwork: Decoding subnetwork: in, express and Joint unsupervised The probability density function of represents the latent variable at the lth moment under unsupervised conditions; represents a latent variable sequence of length 1 to l-1 under unsupervised conditions; Represents an input feature sequence of length 1 to l-1 under unsupervised conditions; express Joint unsupervised The probability density distribution of represents a latent variable sequence of length 1 to L under unsupervised conditions; express Joint unsupervised The probability density distribution of N(,) represents Gaussian distribution; Represent the mean and variance of the prior sub-network respectively; Represent the mean and variance of the posterior sub-network respectively; denote the mean and variance of the decoding sub-network respectively.

6. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 5 is characterized by: In step 5, a self-attention mechanism-sequential variational autoencoder network is constructed under unsupervised conditions; the specific process is as follows: Step 5.

1. Obtain the characteristic variable x at the lth moment in the time window l With the characteristic variable x l Importance of α l The mapping relationship is shown as follows: Among them, x l represents the characteristic variable at the lth moment, w l and b l are the weight value and bias value at the lth moment respectively; Step 52: Based on the characteristic variable x at the lth moment in the time window obtained in step 51 l and importance α l , the output of the forward self-attention mechanism layer is obtained by matrix multiplication The expression is as follows: in, is the feature variable of the first moment output from the self-attention mechanism layer, is the characteristic variable of the second moment output from the self-attention mechanism layer, is the characteristic variable of the lth moment output from the self-attention mechanism layer, is the feature variable at the Lth moment output by the self-attention mechanism layer; α1 is the importance of feature variable x1, α2 is the importance of feature variable x1, α l is the characteristic variable x l The importance of α L is the characteristic variable x L the importance of x1 is the characteristic variable at the first moment, x2 is the characteristic variable at the second moment, and x l is the characteristic variable at the lth moment, x L is the characteristic variable at the Lth moment; Step 53: The output of the forward self-attention mechanism layer Input to the forward gated recurrent unit network, the forward gated recurrent unit network GRU outputs the feature encoding matrix As shown in the following formula: in, Indicates that the output of the forward self-attention mechanism layer Input to the gated recurrent unit network; is the feature encoding matrix output by the forward gated recurrent unit network; GRU represents the gated recurrent unit network; Step 54: According to the characteristic variables from the last moment The characteristic variable at the first moment The sequential input of the backward gated recurrent unit network GRU, the backward gated recurrent unit network GRU outputs the feature encoding matrix; Step 55: Combine the feature coding matrix output from step 54 and the feature coding matrix output from step 523 Perform splicing; Step 5 and 6: Combine the concatenated feature encoding matrix and importance level β j Obtain the output of the backward self-attention mechanism layer through matrix multiplication; Among them, the importance degree β j It is expressed in the form described by the following formula: Among them, β j represents the importance of the concatenated encoding matrix; d represents the hidden layer dimension of the forward gated recurrent unit output; w j and b j Respectively represent the corresponding weight value and bias value of the j-th dimension of the splicing matrix; express and The j-th dimension of the concatenated matrix; Step 57: Based on step 56, obtain the feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism for the feature variable As shown below: in, represents the feature encoding matrix; h1 represents the feature encoding matrix output by the forward gated recurrent unit network The first dimension component of h2 represents the feature encoding matrix output by the forward gated recurrent unit network The second dimension component, h d Represents the feature encoding matrix of the forward gated recurrent unit network output The d-th dimension component of ; β1 represents the first dimension component of the importance of the coding matrix after splicing, β2 represents the second dimension component of the importance of the coding matrix after splicing, and β d The d-th dimension component representing the importance of the concatenated coding matrix, β d+1 The d+1th dimension component representing the importance of the concatenated coding matrix, β d+2 The d+2th dimension component representing the importance of the concatenated coding matrix, β d+L The d+Lth dimension component representing the importance of the concatenated coding matrix; Represents the feature variable at the first moment output by the self-attention mechanism layer, Represents the feature variable at the second moment output by the self-attention mechanism layer, Represents the feature variable at the Lth moment output by the self-attention mechanism layer; Step 58: Obtaining the feature encoding matrix On this basis, the inference model is described based on the Gaussian distribution network GussianNet, which is expressed as follows: in, represents the mean of the posterior sub-network, represents the variance of the posterior sub-network; Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents the feature encoding matrix; Based on the feature encoding matrix Substitute into formula (29) to obtain the mean of the posterior sub-network and variance pass and Perform Gaussian distribution sampling to obtain the latent variable sequence The process is: Sample random noise ε from the standard normal distribution N(0,1) l-1 , and then calculate the latent variables in, represents a latent variable sequence of length 1 to l-1, represents the latent variable sequence at the l-1th moment; Perform the above operation at each moment with a length of 1 to l-1 in the time window L to obtain the hidden variable sequence Step 59 based on and the mean of the above formula variance Get conditional probability Step 50: Design a Gaussian distribution network to implement latent variable sequence Feature variable sequence and the mean of the prior subnetwork and variance The mapping relationship of , and the mean of the prior sub-network is obtained and variance Expressed in the form described below: in, Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; The prior subnetwork of expression (20) The mean of The prior subnetwork of expression (20) variance; Based on latent variable sequence Input feature sequence Substitute the above formula to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence Then obtain a new latent variable sequence Based on latent variable sequence Input feature sequence Substitute the above formula to obtain the mean of the prior sub-network and variance Prior-based sub-network mean and variance Perform Gaussian distribution sampling to obtain the latent variable sequence Then obtain a new latent variable sequence Repeat until the mean of the prior sub-network is obtained and variance Step 51 based on and mean variance Get conditional probability Step 52: Designing Gaussian distribution network to realize latent variable sequence Feature variables and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows: in, represents the mean, represents the variance, Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents a latent variable sequence of length 1 to l-1; Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance Until, Based on feature variables and Substitute the above formula to obtain the mean of the posterior sub-network and variance Step 53: based on and mean variance Obtain the posterior probability conditional on Step 54: based on Calculate the loss function J of the self-attention mechanism-sequential variational autoencoder network under unsupervised conditions u ; 7. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 6 is characterized in that: The loss function J of the self-attention mechanism-sequential variational autoencoder network under the supervised condition built in step 2 in step 6 is s And the loss function J of the self-attention mechanism-serialized variational autoencoder network under unsupervised conditions built in step 4 u , build the loss function J of the semi-supervised sequential variational autoencoder network; the specific process is: Loss function J based on the self-attention mechanism-sequential variational autoencoder network under supervised conditions s And the loss function J based on the self-attention mechanism-sequential variational autoencoder network under unsupervised conditions u , construct the loss function of the self-attention mechanism-serialized variational autoencoder network under semi-supervised conditions; as shown in formula (33):

8. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 7 is characterized in that: In the step 7, a semi-supervised sequential variational autoencoder network is constructed based on the self-attention mechanism-sequential variational autoencoder network under supervised conditions and the self-attention mechanism-sequential variational autoencoder network under unsupervised conditions; the semi-supervised sequential variational autoencoder network is trained based on the training sample set obtained in step 1 until the loss function J of the semi-supervised sequential variational autoencoder network converges, thereby obtaining a trained semi-supervised sequential variational autoencoder network; the specific process is: The input of the self-attention mechanism-sequential variational autoencoder network under semi-supervised conditions is the feature variable x from the first moment to the Lth moment in the time window. 1:L ; The characteristic variable x at the lth moment in the time window l After step 3, the prior / encoding sub-network, the posterior / inference sub-network, and the decoding sub-network of step 3 output probability distributions respectively; The characteristic variable x at the lth moment in the time window l After step 5, the prior / encoding sub-network, the posterior / inference sub-network, and the decoding sub-network of step 5 respectively output probability distributions; Repeat step 7 until the loss function of the semi-supervised self-attention mechanism-serialized variational autoencoder network converges, and a trained semi-supervised serialized variational autoencoder network is obtained.

9. The method for semi-supervised aircraft engine remaining useful life prediction based on sequential variational autoencoders according to claim 8, characterized in that: In step eight, the online data collected by the aircraft engine sensor device is pre-processed in step one and then input into the semi-supervised sequential variational autoencoder network trained in step seven. The trained semi-supervised sequential variational autoencoder network realizes online prediction of the aircraft engine's remaining useful life. The specific process is as follows: The online data collected by the aircraft engine sensor device is preprocessed in step 1. The preprocessed data is processed through steps 31, 32, 33, 34, 35, 36, and 37 to obtain the feature encoding matrix of the bidirectional gated recurrent unit based on the self-attention mechanism. In obtaining the feature encoding matrix Based on the Gaussian distribution network GussianNet, the inference model is described and the mean of the posterior sub-network is obtained. and the variance of the posterior subnetwork The expression is: in, represents the mean of the posterior sub-network, represents the variance of the posterior sub-network; GussianNet (1) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; represents the feature encoding matrix; Based on the feature encoding matrix Substitute into formula (17) to obtain the mean of the posterior sub-network and variance right and Perform Gaussian distribution sampling to obtain the latent variable sequence φ 1:l ; The process is: Sample random noise ε from the standard normal distribution N(0,1) l , and then calculate the latent variables Among them, φ 1:l represents a latent variable sequence of length 1 to l, φ l represents the latent variable sequence at the lth moment; Perform the above operation at each moment in the time window l to obtain the latent variable sequence φ 1:l ; Design a Gaussian distribution network to realize the latent variable sequence φ 1:l , feature encoding matrix and the mean of the decoding subnetwork and variance The mapping relationship is expressed as follows: in, represents the mean, Represents variance, GussianNet (3) Represents a Gaussian distribution network, which is a multi-layer perceptron MLP; φ 1:l represents a latent variable sequence of length 1 to l; Based on the feature encoding matrix and φ 1:1 , put it into formula (35) to obtain the mean of the posterior sub-network and variance Based on the feature encoding matrix and φ 1:2 , put it into formula (35) to obtain the mean of the posterior sub-network and variance Until, Based on the feature encoding matrix and φ 1:L , put it into formula (35) to obtain the mean of the posterior sub-network and variance Based on the decoding sub-network and the mean in formula (35) and standard deviation Obtain the posterior probability conditional on

Images