Method for predicting residual life of equipment under multiple failure modes based on small sample learning

Abstract

The present application relates to a kind of based on small sample learning's equipment residual life prediction method under multiple failure mode, the method includes the following steps: S1, the sensor signal of equipment, failure mode and residual service life are collected;S2, construct reasoning network, obtain failure mode probability distribution and residual service life probability distribution based on bayesian model agnostic meta-learning model;S3, loss function is constructed based on probability distribution, and model parameters are estimated and updated;S4, identify the failure mode of new running equipment and predict its residual service life.Compared with prior art, the present application has the advantages of simultaneously realizing failure mode probability distribution and residual life prediction of probability distribution, high prediction accuracy in small sample situation, strong generalization ability and the like.

Description

Technical Field

[0001] This invention relates to the field of equipment remaining life prediction technology, and in particular to a method for predicting equipment remaining life under multiple failure modes based on few-sample learning. Background Technology

[0002] Remaining Useful Life (RUL) prediction plays a crucial role in prognostics and health management (PHM). It can prevent unexpected machine failures, thereby further preventing serious consequences, including production downtime, equipment safety issues, and logistical disruptions. PHM extracts information about the degradation characteristics of operating equipment by collecting signals from various types of sensors and uses this information to predict the equipment's RUL.

[0003] In recent years, deep learning has provided a powerful tool for RUL prediction due to its flexibility and universality. This class of methods includes Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and their variants, Long Short-Term Memory (LSTMs). All of these methods can be used to predict RUL under a single failure mode. However, in many industrial applications, such as aircraft engines, batteries, and semiconductor devices, multiple failure modes may exist, and devices exhibit different degradation patterns under different failure modes, resulting in different mapping characteristics from sensor data to RUL. In these common industrial scenarios, realizing failure mode identification and predicting the RUL of equipment under multiple failure modes is of great significance for improving the stability and safety of equipment operation.

[0004] Existing deep learning-based methods perform failure mode identification (FMO) and right-to-close (RUL) prediction independently, without considering the interaction between the two. A more critical issue is that, due to the complexity of the network structure, existing deep learning-based multi-mode identification and RUL prediction models require large amounts of data to train. However, due to various limitations, the scale of sensor data collected in industrial applications is often finite. For example, labeling a sufficient number of failure modes is costly, and mounting sensors on certain specialized mechanical equipment presents technical difficulties. Furthermore, existing deep learning models for multi-mode identification and RUL prediction exhibit poor generalization ability when sensor data is insufficient, limiting the application and development of deep learning in the industrial field.

[0005] In summary, the challenges and research gaps in predicting the remaining useful life of equipment under multiple failure modes are as follows: First, current small-sample models only focus on classification or regression problems, lacking models that can simultaneously achieve failure mode identification and remaining useful life prediction. Second, current small-sample diagnostic models can only be used for single failure modes; however, in engineering applications, equipment has multiple failure modes, and different modes exhibit different degradation patterns. Third, there is still a gap in research on failure mode identification and remaining useful life prediction for new tasks with limited data and small sample sizes. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning. This method utilizes a small amount of multi-sensor data collected during equipment operation to simultaneously identify the failure modes and predict the remaining life of the equipment.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for predicting the remaining life of equipment under multiple failure modes based on few-shot learning, the method comprising the following steps:

[0009] S1. Collect sensor signals, failure modes, and remaining service life of the equipment;

[0010] S2. Construct an inference network and obtain the failure mode probability distribution and remaining useful life probability distribution based on the Bayesian model agnostic meta-learning model.

[0011] S3. Construct a loss function based on the probability distribution, and estimate and update the model parameters;

[0012] S4. Identify the failure modes of new operating equipment and predict its remaining useful life.

[0013] Furthermore, the inference network includes a prototype network and a regression sub-network, used to take the features extracted from the sensor signals as the input x of the Bayesian model agnostic meta-learning model, and output the probability distribution vector z of K failure modes and the predicted remaining useful life probability distribution vector y. The Bayesian model agnostic meta-learning model is trained through multiple historical tasks, and then used to identify and predict new tasks. The historical tasks consist of a small sample dataset containing a support set and a query set. The support set contains label values ​​for training the model. The query set is used to obtain the model output values ​​and compare them with the corresponding label values ​​to evaluate the model performance.

[0014] Furthermore, the acquisition of the failure mode probability distribution specifically refers to:

[0015] S2011. The prototype network uses a DNN to map the input x to an embedding space, in which the prototype computation expression of the support set is:

[0016]

[0017] in, The embedding function Θ represents the probability distribution of failure modes. C c is the parameter set of the embedded function. k (k = 1, ..., K) is the mean vector of the embedding function of all support points under failure mode k. This indicates the set of subtasks that support the centralized failure mode k, n. k Indicates support set The number of tasks in failure mode k;

[0018] S2012. The prototype network uses the softmax function to measure the distance between x in the query set and the prototype of each failure mode in the embedding space to generate the probability distribution of each failure mode. For failure mode k, its probability distribution z (k) The formula for calculation is:

[0019]

[0020] in, A similarity metric vector is represented using exponential Eulerian distance, i.e.:

[0021]

[0022] S2013, Constructing the Failure Mode Probability Distribution Vector In the formula, z = [z (1) ,…,z (k) ,…,z (K) ], A simplified notation for the calculation expression of the failure mode probability distribution vector.

[0023] Furthermore, the specific method for obtaining the probability distribution of remaining useful life is as follows:

[0024] S2021. Establish a regression subnetwork to realize the candidate remaining useful life y of input x to failure mode k (k=1,…,K) under failure mode k. (k) The mapping, that is:

[0025]

[0026] in, Indicates the regression subnetwork, For the parameter set of the regression subnetwork;

[0027] S2022, Given a support set The formula for calculating the mixed probability distribution of remaining useful life corresponding to input x in the query set under K failure modes is as follows:

[0028]

[0029]

[0030] Where, Θ={Θ C ,Θ R};

[0031] S2023, Based on parameter set Use its support set Update the prediction model parameters for each historical task τ to obtain the posterior probability distribution of the task. in This represents the set of task posterior distribution parameters for the regressive subnetwork;

[0032] S2024, the meta-update process is based on the task posterior distribution parameter set. Obtain the mixed probability distribution of remaining useful life for x in the query set, and apply Empirical Bayes to the hierarchical probability distribution (i.e.,

[0033] S2025. Insert Bayesian integrals to update the mixed probability distribution of remaining lifetime for all tasks under K failure modes, as shown in the following formula:

[0034]

[0035] in, It is a collection of historical tasks;

[0036] S2026. The Monte Carlo approximation is used to calculate the mixed probability distribution of remaining useful life:

[0037]

[0038]

[0039] in, From The random samples generated in the sample, and R random samples are sampled. The distribution is Gaussian in the DNN formula, i.e.:

[0040]

[0041] σ 2 σ is the variance of the noise term in the observed value y. 2 =1.

[0042] Furthermore, the loss function constructed based on the probability distribution in S3 is as follows:

[0043] S3011. Based on the probability distribution of failure mode k, obtain the likelihood function L for failure mode k identification. c :

[0044]

[0045] in, Represents the query set. This represents the actual failure mode vector of length K.

[0046] S3012, Loss function for identifying failure mode k Designed as the likelihood function L c The negative logarithm;

[0047]

[0048] S3013. Based on the mixed probability distribution of remaining useful life, obtain the likelihood function L for predicting remaining useful life under multiple failure modes. R :

[0049]

[0050] in, This represents the actual remaining useful life value corresponding to input x;

[0051] S3014, The remaining useful life prediction loss function under multiple failure modes. Designed as the likelihood function L R The negative logarithm:

[0052]

[0053] in, Let τ represent the loss function for predicting the remaining lifetime of task τ;

[0054] S3015. Obtain the total loss function based on all data in the support set and query set. And update the parameter set Θ of the embedded function accordingly. C Parameter set of the regression subnetwork

[0055] Furthermore, the parameter set Θ of the updated embedding function C The formula is as follows:

[0056]

[0057] in, For parameter set Θ C Based on the total loss function The partial derivative of ε, where ε represents the learning rate.

[0058] Furthermore, the parameter set of the updated regression subnetwork... The update is performed using the Stan variational gradient descent method, as shown in the following formula:

[0059]

[0060]

[0061] in, Let ε represent the average gradient of all particles, and ε represent the learning rate. Let the loss function be the one corresponding to predicting the remaining lifetime of task τ. express right The partial derivative of , where v represents a positive definite kernel.

[0062] Furthermore, S4 specifically refers to:

[0063] S4011. Identify the failure mode with the highest probability of occurrence as the new failure mode of the operating equipment, that is:

[0064]

[0065] in, This represents the probability of K failure modes;

[0066] S4012. Obtain the probability density function for remaining useful life prediction through the mixed probability distribution of remaining useful life:

[0067]

[0068] S4013, Predict the remaining useful life of new operating equipment. * As the expectation of y given P(y|x,Θ), that is: y * =E P(y|x,0) (y)=∫yP(y|x,Θ)dy;

[0069] S4014. Predict remaining useful life y using numerical methods * Approximately the sample mean:

[0070]

[0071] S4015. Based on the formula in S4014, complete the prediction of the remaining service life of the new operating equipment.

[0072] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.

[0073] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] I. This invention proposes a multi-mode Bayesian model agnostic meta-learning model (MBMAML), which simultaneously predicts the failure mode probability distribution and the remaining lifetime probability distribution based on an inference network. This solves the uncertainty problem introduced by insufficient samples, resulting in higher prediction accuracy and more accurate prediction results.

[0076] Second, the present invention has two objectives in constructing a loss function integration model based on the likelihood function of a mixed probability distribution, which makes up for the shortcomings of existing methods in separating failure mode identification and RUL prediction.

[0077] Third, the prediction model of this invention uses task posterior to characterize the uncertainty of model parameters on different tasks, and updates the model parameters without updating them, thereby improving the model's generalization ability on new tasks. This enables failure mode identification and remaining life prediction for new equipment with only a small amount of data under small sample conditions. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0079] Figure 2 This is a framework diagram of the inference network of the present invention;

[0080] Figure 3 This is the feature extraction result of each sensor of a device in use in an embodiment of the present invention, wherein the first row and the second row are the first and second derivatives of the failure state;

[0081] Figure 4 To assess the accuracy of the model and benchmark method described in the embodiments of the present invention in identifying failure modes at different actual RUL levels;

[0082] Figure 5 The RUL prediction errors of the model and benchmark method described in the embodiments of the present invention are under different actual RUL levels. Detailed Implementation

[0083] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0084] Example

[0085] like Figure 1 As shown, this embodiment uses a failure dataset of an aircraft gas turbine engine as an example to provide a method for predicting the remaining service life (RUL) of equipment under multiple failure modes based on few-shot learning. This failure dataset was generated on the Commercial Modular Aero-Propulsion System Simulator (C-MAPSS) developed by NASA (National Aeronautics and Space Administration), simulating the failure process of a turbofan engine. The dataset has four sub-datasets; this embodiment uses the FD003 sub-dataset for a case study. In FD003, the engine has two possible failure modes (K=2): failure caused by the high-pressure compressor and failure caused by the engine fan. The failure state of the aircraft engine is monitored by signals from 21 sensors. The dataset includes 100 sets of historical data and 100 sets of online data. The historical data consists of multi-sensor monitoring data collected from engine startup to failure, while the test data includes multi-sensor data collected from engine startup to a point before failure and the actual RUL. By adopting the technical solution provided by this invention, it helps to solve the problem of model overfitting in data-driven modeling caused by insufficient aircraft engine failure data, and realizes aircraft engine failure mode diagnosis and remaining service life prediction, effectively improving the economy of maintenance and the stability and safety of flight. The method includes the following steps:

[0086] S1. Collect sensor signals, failure modes, and remaining service life of the equipment, taking an aircraft gas turbine engine as an example.

[0087] S2. Construct an inference network and obtain the failure mode probability distribution and RUL (remaining useful life) probability distribution based on MBMAML (Bayesian Model Agnostic Meta-Learning Model).

[0088] The MBMAML model introduces an embedding function based on a DNN (Deep Neural Network). The embedding function is a non-linear mapping of the input, consisting of an input layer, multiple hidden layers, and an output layer. Let the model input vector be denoted as x, and the model output vector as u. The calculation expression for the embedding function of the J hidden layers is as follows:

[0089]

[0090] u=h J V+d

[0091] Where h j (j = 1, ..., J) are the values ​​of the j-th hidden layer, and h0 = x, parameter W j and b jLet V and d be the weights and biases of the j-th hidden layer, respectively, and let V and d be the weights and biases of the output layer, respectively. It is an activation function (e.g., a linear rectified function). For simplicity, the parameter set of the embedding function is represented as α = {W1, b1, ..., W...}. J ,b J Let the embedding function be f, then u = f. α (x). The input x, output u, parameter set α, and model name f of the embedding function can be designed according to the actual situation.

[0092] like Figure 2 As shown, the inference network includes a prototype network and a regression sub-network, which are used to take the features extracted from the sensor signals as the input x of the Bayesian model agnostic meta-learning model, and output the probability distribution vector z of K failure modes and the predicted remaining useful life probability distribution vector y.

[0093] The failure mode identification and RUL prediction for each aircraft engine are denoted as task τ. Considering that the engines operate under identical environmental conditions, the MBMAML model assumes that the data for each task is sampled from a common task distribution, and that the sampled tasks are independent and identically distributed, sharing the statistical regularities of the task distribution. The MBMAML model... Training on historical tasks, among which It is a collection of historical tasks, and it addresses new tasks. * (i.e., a new engine) to identify and predict.

[0094] The historical task τ is based on a small sample dataset. The dataset consists of a support set. and query set (Right now The support set is similar to the training set in a traditional DNN; it contains label values ​​used to train the model. The query set is similar to the validation set in a traditional DNN; it is used to retrieve the model's output values ​​and compare them with the corresponding label values ​​to evaluate model performance. Then, the dataset for all historical tasks is represented as... and These are the support set and query set of all historical data, i.e.

[0095] Considering the significant noise in sensor data, relying solely on raw sensor data is insufficient to identify the failure mode of an aircraft engine. The MBMAML model extracts features from the sensor data as its input. Given that engines exhibit different failure rates under different failure modes, the extracted features are set as the failure rate at the previous observation time point (i.e., the first and second derivatives of the failure state). Furthermore, this model establishes an enhanced feature extraction method within the mixed-effects model framework, sharing information across all engines through the posterior distribution of random effects parameters to enhance the feature extraction process.

[0096] The enhanced feature extraction process is as follows: First, the signal of engine sensor i (i = 1, ..., I) at time point t is denoted as... Based on the hybrid effects model of Lu and Meeker, the MBMAML model captures the engine failure state according to sensor signal i:

[0097]

[0098] Where ψ(t) is a basis function for time t, ε i This is the corresponding noise term, assumed to follow a Gaussian distribution. Estimating the random effects parameter Γ using the least squares method i It can be obtained Among them Ψ i and These are ψ(t) from the start to the last observation point and The matrix formed. Given the failure state g at time t. i (t)=ψ T (t)Γ i The original extracted features were designed as first and second derivatives standardized by decimal scaling.

[0099] However, due to the limited sensor data from a single task, the original extracted features are unreliable. To address this issue, the MBMAML model considers learning shared information from all historical task data to enhance feature extraction. Taking a specific aircraft engine l as an example, the mixed-effects model assumes a prior distribution Γ i ~G i And using the sensor signals of engine l Update posterior distribution Where p(Γ) l,i (From the prior distribution G) i The enhancement features of group information were considered. For g l,i The expectation of (t) is calculated as follows:

[0100]

[0101] Since the above equation has no analytical expression, an approximate value is calculated using the sample mean of historical tasks:

[0102]

[0103] In the formula, N τ This represents the number of historical tasks, where d is the number of tasks to ensure. The constant.

[0104]

[0105] In the formula, T l The noise variance is expressed as: (This represents the time point of the last observation.)

[0106]

[0107] The MBMAML model sets the extracted feature x to the previous observation time T. l Enhanced features extracted from all sensor signals, i.e.

[0108] like Figure 3 The image shows an example of the feature extraction results from various sensors of a currently operating aircraft engine in this embodiment. The first and second rows represent the first and second derivatives of the failure state. Figure 3 The two different points represent sensor data and extracted features under two failure modes: HPC (High Pressure Compressor) and FAN (Fan). Clearly, the extracted features have a strong relationship with RUL (Rate Utility Rate). Under different failure modes, the mapping from extracted features (such as P30 and Phi) to RUL exhibits distinct characteristics. The extracted feature x captures the failure rate of the sensor data and eliminates the influence of noise, making it more effective for multi-failure mode identification and prediction.

[0109] Failure mode identification of aircraft engines is a small-sample classification problem, and overfitting is a common issue due to limited available data. To address this problem, the model learns an embedding function of the input x to construct a prototype network. The specific steps for obtaining the failure mode probability distribution are as follows:

[0110] S2011. The prototype network uses a DNN to map the input x into an embedding space, in which embedding points cluster around the prototype of each failure mode, and the prototype of each mode (referred to as the class prototype) is set as the average of its support set in the embedding space. Then, the prototype network performs classification on the embedding points in the query set by finding the nearest class prototype, in which the prototype calculation expression of the support set is:

[0111]

[0112] in, The embedding function Θ represents the probability distribution of failure modes. C For the parameter set of the embedded function, c k (k = 1, ..., K) is the mean vector of the embedding function of all support points under failure mode k. This indicates the set of subtasks that support the centralized failure mode k, n. k Indicates support set The number of tasks in failure mode k;

[0113] S2012. The prototype network uses a softmax function to measure the distance between x in the query set and the prototype of each failure mode in the embedding space to generate the probability distribution for each failure mode. For failure mode k, its probability distribution z... (k) The formula for calculation is:

[0114]

[0115] in, A similarity metric vector is represented using exponential Eulerian distance, i.e.:

[0116]

[0117] S2013, Constructing the probability distribution vector of the failure model In the formula, z = [z （1) ,…,z (k) ,…,z (K) ], A simplified notation for the calculation expression of the failure mode probability distribution vector.

[0118] RUL prediction under multiple failure modes is a problem of small-sample classification and regression fusion. Training complex models (such as DNNs) with insufficient samples introduces significant uncertainty, leading to overfitting. To address this issue, this invention proposes the MBMAML model for RUL prediction. Due to the high cost of aircraft engines, historical mission samples can only be obtained from a limited number of engines, and each mission sample contains only one or a few data points. Furthermore, although all missions are sampled from the same mission distribution and share statistical properties, different estimated parameters will be generated when learning each mission, and the limited available data will cause overfitting for each mission. The MBMAML model addresses these problems by estimating shared mission information parameters and proposes a mission posterior and meta-update process to mitigate overfitting and improve model accuracy. The mission posterior captures the uncertainty of the model parameters from the historical mission τ, and based on this, the meta-update process uses data from the query set to perform meta-learning of the model parameters.

[0119] The specific method for obtaining the probability distribution of remaining useful life is as follows:

[0120] S2021. When performing RUL prediction, since aircraft engines exhibit different failure modes under different failure modes, the MBMAML model models RUL under multiple failure modes to achieve accurate RUL prediction. The MBMAML model establishes a regression subnetwork to realize the candidate remaining useful life y from input x to k (k=1,…,K) under failure mode k. (k) The mapping, that is:

[0121]

[0122] in, Indicates the regression subnetwork, For the parameter set of the regression subnetwork;

[0123] S2022, Given a support set The formula for calculating the mixed probability distribution of the remaining useful life of x in the query set under K failure modes is as follows:

[0124]

[0125]

[0126] Where, Θ={Θ C ,Θ R The K regression subnetworks combine the prototype network. The failure mode probability helps to provide more comprehensive information to improve the accuracy of RUL prediction;

[0127] S2023, Task Post-hoc Process Based on Parameter Set Use its support set Update the prediction model parameters for each historical task τ to obtain the posterior probability distribution of the task. in This represents the set of task posterior distribution parameters for the regressive subnetwork;

[0128] S2024, the meta-update process is based on the task posterior distribution parameter set. Obtain the mixed probability distribution of remaining useful life for x in the query set, and apply Empirical Bayes to the hierarchical probability distribution (i.e.,

[0129] S2025. Since the model assumes that the sampling tasks are identical and independently distributed, a Bayesian integral is inserted to update the mixed probability distribution of the remaining lifetime of all tasks under the K failure modes, as shown in the following formula:

[0130]

[0131] in, It is a collection of historical tasks; this hybrid probability distribution supports the probability quantification of instances based on a small amount of data, and mitigates the large amount of uncertainty caused by limited data through the task posterior and meta-updates of empirical Bayesian formulas and model parameters; compared with point estimation or approximations using simple Gaussian distributions, such as DNNs and Bayesian neural networks (BNNs), the MBMAML model provides a more flexible approach through its task-training posterior distribution. It captures the uncertainty of model parameters without assuming any specific approximate distribution; based on the task-training posterior distribution for all tasks, it performs meta-learning on the model parameters to further extract common task information.

[0132] S2026. Considering that the integral in S2025 is difficult to solve analytically, the Monte Carlo approximation is used to calculate the mixed probability distribution of the remaining useful life:

[0133]

[0134]

[0135] in, From The random samples generated in the sample, and R random samples are sampled. The distribution is Gaussian in the DNN formula, i.e.:

[0136]

[0137] σ 2σ is the variance of the noise term in the observed value y. In most existing literature on deep learning and FSL (few-shot learning), the variance of the noise term is fixed, therefore σ is set... 2 =1.

[0138] S3. Construct a loss function based on the probability distribution, and estimate and update the model parameters.

[0139] After constructing a network of query points based on the support set through probabilistic modeling, the MBMAML model uses data from both the support set and the query set to train the network, denoting the labeled query set samples as...

[0140] The loss function constructed based on the probability distribution in S3 is as follows:

[0141] S3011. Based on the probability distribution of failure mode k, obtain the likelihood function L for failure mode k identification. c :

[0142]

[0143] in, Represents the query set. This represents the actual failure mode vector of length K.

[0144] S3012, Loss function for identifying failure mode k Designed as the likelihood function L c The negative logarithm;

[0145]

[0146] S3013. Based on the mixed probability distribution of remaining useful life, obtain the likelihood function L for predicting remaining useful life under multiple failure modes. R :

[0147]

[0148] in, This represents the actual remaining useful life value corresponding to input x;

[0149] S3014, The remaining useful life prediction loss function under multiple failure modes. Designed as the likelihood function L R The negative logarithm:

[0150]

[0151] in, Let τ represent the loss function for predicting the remaining lifetime of task τ;

[0152] S3015. Obtain the total loss function based on all data in the support set and query set. And update the parameter set Θ of the embedded function accordingly. C Parameter set of the regression subnetwork

[0153] The parameter set Θ of the update embedding function C The formula is as follows:

[0154]

[0155] in, For parameter set Θ C Based on the total loss function The partial derivatives, where ε represents the learning rate;

[0156] Parameter set Θ C Based on the total loss function The expression for calculating the partial derivative is as follows:

[0157]

[0158] in, To simplify the formula, within the framework of the prototype network, The calculation expression is as follows:

[0159]

[0160] in, Calculated by DNN.

[0161] The parameter set of the updated regression subnetwork The Stan variational gradient descent method is used for updating, considering that the MBMAML model is based on The MBMAML model uses Bayes' theorem for probabilistic modeling and employs a non-parametric variational inference method: Stein variational gradient descent (SVGD) for updating. Unlike traditional variational inference, which restricts a series of approximate distributions to a tractable parameter distribution, SVGD does not restrict the parameter distribution. It integrates the advantages of Monte Carlo and variational inference, and converges faster than Monte Carlo due to its deterministic update rules. Specifically, SVGD maintains model parameters for R instances. Named as particles, the particle set is represented as follows:

[0162]

[0163] Utilize support set data in each iteration Task τ(τ∈ ) task posterior value As The loss function is calculated using the following formula:

[0164]

[0165]

[0166] in, Let represent the average gradient of all particles, and ε represent the learning rate. Let the loss function be the one corresponding to predicting the remaining lifetime of task τ. express right The partial derivatives are calculated using a DNN; A repulsive force is applied between particles to ensure they do not shrink to a point, where v is a positive definite nucleus;

[0167] The MBMAML model is based on the ensemble learning of model parameters from the prototype network and the regression sub-network. Traditional multi-objective deep learning models ignore the relationship between the two objectives, that is, the model parameters for failure mode recognition and RUL prediction are learned independently as classification and regression problems, respectively. The MBMAML model proposed in this invention associates the two objectives together through a mixture of probability distributions and enhances the capabilities of the proposed model by jointly optimizing the ensemble loss function. The parameter estimation algorithm of the MBMAML model is shown in Table 1.

[0168] Table 1 MBMAML Parameter Estimation Algorithm

[0169]

[0170]

[0171] S4. Identify the failure modes of a new operational aircraft engine and predict its remaining service life. The new operational aircraft engine (new mission) contains only unlabeled input x. The goal of the MBMAML model is to identify the failure modes k of the new mission. * And predict its remaining useful life under multiple failure modes, wherein S4 specifically includes:

[0172] S4011. For failure mode identification, the output of the prototype network in the MBMAML model is the probability of K failure modes. The failure mode with the highest probability of occurrence is identified as the new failure mode of the operating equipment, that is:

[0173]

[0174] S4012. For RUL prediction of a new task under multiple failure modes, since only unlabeled data is available, it is impossible to update specific tasks. Therefore, the MBMAML model is constructed by using meta-parameters. The model is used to handle new tasks, and the probability density function for predicting the remaining useful life is obtained through the mixture probability distribution of the remaining useful life:

[0175]

[0176] S4013, Predicting the remaining service life of newly operational aircraft engines. * As the expectation of y given P(y|x,Θ), that is: y * =E P(y|x,Θ) (y)=∫yP(y|x,Θ)dy;

[0177] S4014. Predict remaining useful life y using numerical methods * Approximately the sample mean:

[0178]

[0179] S4015. Based on the formula in S4014, complete the prediction of the remaining service life of the new operating aircraft engine.

[0180] To evaluate the effectiveness of the MBMAML model, this invention uses accuracy, precision, recall, and F1 score as indicators for failure mode identification, and mean absolute error (MAE) and RUL prediction error as indicators for RUL prediction. MAE is the arithmetic mean of the absolute differences between the predicted RUL and the actual RUL of the in-service unit. The RUL prediction error for online aircraft engines ∈ * The definition of y is the RUL predicted value. * Compared with actual RUL value The absolute difference between them divided by the actual runtime T * ,Right now:

[0181]

[0182] The smaller the prediction error, the better the model performance.

[0183] like Figure 4As shown, the proposed method is compared with several benchmark methods to evaluate its performance. These benchmark methods include: (1) traditional machine learning methods, where failure mode identification and RUL prediction are performed separately. KNN (Nearest Neighbor), DT (Autoregressive Model), SVM (Support Vector Machine), LR (Logistic Regression), BC (Branch Cutting), and NN (Neural Network) are used for failure mode identification. After identifying the failure mode, NN is used for RUL prediction under known failure modes to achieve a nonlinear mapping between extracted features and RUL; (2) Joint BNN (Joint Bayesian Neural Network) method, which achieves both failure mode identification and RUL prediction by characterizing the relationship between failure modes and RUL, but does not consider the impact of small sample size; (3) existing typical FSL models: Prototype&MAML (Model Agnostic Meta-Learning Model Based on Prototype Network) and Prototype&BMAML (Bayesian Model Agnostic Meta-Learning Model Based on Prototype Network). Prototype&MAML identifies failure modes through the prototype network and predicts RUL independently through MAML. Prototype&BMAML identifies failure modes through a prototype network and independently predicts RUL through BMAML, with its loss function set to independently learn the model parameters of both the prototype network and BMAML. In contrast, the proposed model employs an interactive loss function that integrates failure mode identification into the RUL prediction process for more accurate model parameter estimation.

[0184] Figure 4 The accuracy of the proposed method and the baseline method in failure mode identification is shown at different actual RUL levels. Actual RUL levels "20, 40, ..., 120" refer to operating units with actual RUL less than or equal to 20, 40, ..., 120, and level "All" refers to all operating units. When the actual RUL is small (e.g., less than 40), i.e., when there are enough observation time points for the operating units, the accuracy of other methods, except for KNN and DT, is relatively high. As the actual RUL increases, the accuracy of all methods gradually decreases, but the proposed method still outperforms the baseline method, indicating that the model can effectively identify failure modes.

[0185] Figure 5The RUL prediction errors of the stated model and benchmark methods were compared under real-world RUL levels at different levels. At each level, the points and error bars represent the mean and standard deviation of the RUL prediction error, respectively. The standard deviation is the standard deviation of the RUL prediction error divided by the square root of the number of tasks in the test set. Compared with existing machine learning models (i.e., KNN+NN, DT+NN, SVM+NN, LR+NN, BC+NN, and Joint BNN), the FSL models (i.e., Prototype&MAML, Prototype&BMAML, and the present model) have smaller RUL prediction errors at all levels. At RUL level 20, the RUL prediction error of the BMAML model of the present invention is slightly larger than that of Prototype&MAML and Prototype&BMAML because at this level, failure mode identification errors negatively impact the RUL prediction of the present method. At all other levels, the prediction error of the BMAML model is the smallest. Furthermore, the standard deviation of the BMAML model is much smaller than that of conventional machine learning methods, indicating its good robustness.

[0186] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.

[0187] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0188] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting the remaining life of a device under multiple failure modes based on small sample learning, characterized in that, The method includes the following steps: S1. Collect sensor signals, failure modes, and remaining service life of the equipment; S2, construct a reasoning network, and obtain a failure mode probability distribution and a remaining useful life probability distribution based on a Bayesian model-agnostic meta-learning model; the reasoning network comprises a prototype network and a regression sub-network, and is configured to take features extracted from a sensor signal as an input x of the Bayesian model-agnostic meta-learning model, and simultaneously output a failure mode probability distribution vector z and a predicted remaining useful life probability distribution vector y. a failure mode probability distribution vector z and a predicted remaining useful life probability distribution vector y. The specific steps to obtain the failure mode probability distribution are as follows: S2011, The prototype network uses a DNN to process the input. Mapped into the embedding space, within that embedding space; S2012, the prototype network uses the softmax function to measure the query set. The distance to the prototype of each failure mode in the embedding space is used to generate the probability distribution of each failure mode, for each failure mode. ; S2013, Construct the failure mode probability distribution vector; The specific steps to obtain the probability distribution of remaining useful life are as follows: S2021. Establish a regression subnetwork to realize failure modes. Enter below To failure mode , Candidate remaining useful life The mapping, that is: ， in, Indicates the regression subnetwork, For the parameter set of the regression subnetwork; S2022, Given a support set ,get Input in the query set under various failure modes The formula for calculating the mixed probability distribution of the corresponding remaining useful life is as follows: in, ; The parameter set of the embedding function for the failure mode probability distribution; Let be the failure mode probability distribution vector, where , , A simplified notation for the calculation expression of the failure mode probability distribution vector; S2023, Based on parameter set , using its support set Update each historical mission The prediction model parameters are used to obtain the posterior probability distribution of the task. ,in This represents the set of task posterior distribution parameters for the regressive subnetwork; S2024, the meta-update process is based on the task posterior distribution parameter set. Get query set The remaining useful life mixture probability distribution is used to apply empirical Bayesian methods to the hierarchical probability distribution. ); S2025, Insert Bayesian integral to update all tasks. The mixed probability distribution of remaining useful life under various failure modes is given by the following formula: in, It is a collection of historical tasks; S2026. The Monte Carlo approximation is used to calculate the mixed probability distribution of remaining useful life: in, From Random samples generated in the process, and sampling A random sample, The distribution is Gaussian in the DNN formula, i.e.: It is the variance of the noise term in the observed value y. ; S3. Construct a loss function based on the probability distribution, and estimate and update the model parameters; S4. Identify the failure modes of new operating equipment and predict its remaining useful life.

2. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 1, characterized in that, The Bayesian agnostic meta-learning model is trained on multiple historical tasks and then used to identify and predict new tasks. The historical tasks consist of a small sample dataset containing a support set and a query set. The support set contains label values ​​used to train the model. The query set is used to obtain the model output values ​​and compare them with the corresponding label values ​​to evaluate the model performance.

3. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 2, characterized in that, The prototype computation expression for the support set is: in, An embedding function representing the probability distribution of failure modes. This is the parameter set of the embedded function. Failure mode , The mean vector of the embedding function of all support points. This indicates support for centralized failure mode. The sub-task set, Indicates support set Failure Mode The number of tasks; Probability distribution of each failure mode The formula for calculation is: in, A similarity metric vector is represented using exponential Eulerian distance, i.e.: Failure Mode Probability Distribution Vector In the formula , , A simplified notation for the calculation expression of the failure mode probability distribution vector.

4. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 1, characterized in that, The loss function constructed based on the probability distribution in S3 is as follows: S3011, Based on Failure Mode k The probability distribution is used to obtain the failure modes. k Likelihood function of identification : in, Represents the query set. Indicates length is The actual failure mode vector, ; S3012, Failure Mode k Loss function for identification Designed as a likelihood function The negative logarithm; S3013. Based on the mixed probability distribution of remaining useful life, obtain the likelihood function for predicting remaining useful life under multiple failure modes. : in, Indicates input The corresponding actual remaining useful life value; S3014, The remaining useful life prediction loss function under multiple failure modes. Designed as a likelihood function The negative logarithm: in, Indicates task The remaining useful life prediction loss function; S3015. Obtain the total loss function based on all data in the support set and query set. And update the parameter set of the embedded function accordingly. Parameter set of the regression subnetwork .

5. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 4, characterized in that, The parameter set of the update embedding function The formula is as follows: in, For parameter set Based on the total loss function The partial derivatives, This represents the learning rate.

6. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 5, characterized in that, The parameter set of the updated regression subnetwork The update is performed using the Stan variational gradient descent method, as shown in the following formula: in, This represents the average gradient of all particles. Indicates the learning rate. For the task The loss function corresponding to the prediction of remaining useful life. express right The partial derivatives, This represents a positive definite kernel.

7. The method for predicting the remaining life of equipment under multiple failure modes based on few-sample learning according to claim 6, characterized in that, Specifically, S4 is: S4011. Identify the failure mode with the highest probability of occurrence as the new failure mode of the operating equipment, that is: in, express The probability of a failure mode; S4012. Obtain the probability density function for remaining useful life prediction through the mixed probability distribution of remaining useful life: S4013. Predicting the remaining service life of new operating equipment As given The expected value of time y is: ; S4014. Predicting remaining useful life using numerical methods Approximately the sample mean: S4015. Based on the formula in S4014, complete the prediction of the remaining service life of the new operating equipment.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images