Aero-engine residual life uncertainty quantification method under online prediction scene

Abstract

The invention relates to the technical field of aero-engines, in particular to an aero-engine residual life uncertainty quantification method under an online prediction scene, and the method comprises the steps: obtaining a training data set of an aero-engine in a test run and operation process; on the basis of a diffusion model, according to the monitoring data and in combination with the representation mapping function, obtaining a de-noising network model; collecting a test sample with random missing and sensor fault data in the online prediction scene; according to the test sample, obtaining a filling sample through the optimized denoising network model; combining the filling sample with the test sample to obtain a filled test sample; and constructing an RUL prediction interval for the filled test sample according to the trained RUL prediction model and the weighted empirical distribution function through conformal prediction. The method aims at accurately evaluating the comprehensive performance of the aero-engine and improving the reliability and robustness of a prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engines, and in particular to a method for quantifying uncertainty of the remaining life of an aero-engine in an online prediction scenario. Background Art

[0002] The reliability of aircraft engines is the basis for ensuring aviation safety, improving operational efficiency and achieving economic feasibility. Due to the high complexity of aircraft engine design and operation, it is crucial to ensure its comprehensive performance. With the continuous advancement of aviation technology, the need to ensure the continuous and safe operation of engines is becoming increasingly urgent. In this context, Prognostics and Health Management (PHM) has gradually become a key strategy. PHM aims to monitor, predict and manage the health status and comprehensive performance of aircraft engines in real time by combining sensor monitoring, data acquisition, comprehensive performance control optimization and maintenance decision models. Through active management, PHM not only improves the safety and reliability of the system, but also optimizes maintenance plans, reduces downtime, and significantly reduces operating costs. In the performance control optimization and comprehensive performance evaluation of aircraft engines, the remaining useful life (RUL) refers to the time interval from the current moment to the expected occurrence of a failure. As one of the core indicators of PHM, RUL is an important parameter for quantitatively evaluating the changes in the comprehensive performance of engines, and provides an important basis for the performance control and evaluation of aircraft engines and the formulation of preventive maintenance decisions.

[0003] In aircraft engine RUL prediction, uncertainty quantification is an important method to evaluate the reliability and robustness of prediction results. It can not only provide prediction intervals to help decision makers comprehensively evaluate the comprehensive performance and potential risks of aircraft engines, but also optimize maintenance strategies and enhance the robustness of prediction models. By quantifying the uncertainty of prediction results, decision makers can clearly understand the possible range of prediction errors, thereby achieving more accurate performance control and optimization, and making more scientific and reliable decisions.

[0004] In recent years, deep learning methods have demonstrated strong performance in aircraft engine RUL prediction, and can automatically extract deep degradation features from sensor monitoring data. In the existing technology, deep learning methods are used to achieve prediction through deterministic neural networks, mainly providing point estimates of RUL prediction. Although these methods have made significant progress in prediction performance, they have failed to effectively quantify the uncertainty of the prediction results, resulting in the inability to accurately evaluate the comprehensive performance of aircraft engines in complex and changing actual operating environments.

[0005] In addition, in actual online prediction scenarios, especially in aircraft engines operating under harsh working conditions, sensor monitoring data often has missing data. There are two main types of missing data: one is random data missing, and the other is complete data missing due to sensor failure. Random data missing is usually caused by temporary sensor failure, network transmission problems, or data storage failure. This missing mode is usually intermittent and scattered. In contrast, complete missing due to sensor failure is more destructive because it causes the entire segment of data from one or more sensors to be unavailable. Such missing data not only seriously affects the integrity of the data, but may also lead to incorrect assessments of the comprehensive performance and health of aircraft engines. The problems of random missing and partial sensor failure in online prediction scenarios significantly affect the performance of existing RUL prediction methods based on deep learning. Summary of the invention

[0006] In order to solve the above technical problems, the present invention mainly aims at the problem that the existing deep learning methods cannot effectively quantify the uncertainty of the prediction results in the RUL prediction of aircraft engines, and the negative impact of random missing and partial sensor failure data on the performance of existing methods in online prediction scenarios. The present invention provides a method for quantifying the uncertainty of the remaining life of aircraft engines in online prediction scenarios, which aims to accurately evaluate the comprehensive performance of aircraft engines and improve the reliability and robustness of prediction results.

[0007] The purpose of the present invention is to provide a method for quantifying uncertainty of the remaining life of an aircraft engine in an online prediction scenario, comprising: Acquire a training data set of an aircraft engine during a test run and operation process, wherein the training data set includes monitoring data and a remaining life label corresponding to the monitoring data; Characterize the monitoring data through a pre-training strategy according to the monitoring data to obtain a representation mapping function; Based on the diffusion model, according to the monitoring data and combined with the characterization mapping function, a denoising network model is obtained; Divide the training data set into a training subset and a calibration subset; The RUL prediction model is trained according to the training subset to obtain a trained RUL prediction model; A mapping network that maps monitoring data to latent space is constructed through the calibration subset; Collect test samples with random missing and sensor failure data in online prediction scenarios; Obtain a filled sample according to the test sample through the optimized denoising network model; combine the filled sample with the test sample to obtain a filled test sample; The samples in the calibration subset and the padded test samples are mapped to the latent space through the mapping network to obtain the weight of each sample in the calibration subset; Construct a conformity score based on the absolute residual of each sample in the calibration subset, and weight the conformity scores by the weights of different samples to define a weighted empirical distribution function; For the filled test samples, the RUL prediction interval is constructed by conformal prediction based on the trained RUL prediction model and the weighted empirical distribution function.

[0008] Preferably, the characterization mapping function is obtained according to the following steps: The monitoring data is masked and used as input, and the pre-trained model is trained until convergence to obtain a trained pre-trained model; wherein the pre-trained model includes input encoding, Transformer encoder and linear layer in sequence; the linear layer is removed from the trained pre-trained model to obtain the representation mapping function.

[0009] Preferably, the denoising network model is obtained according to the following steps: The monitoring data is divided into two parts: filling targets and observation values; Encoding the observed value through a representation mapping function to obtain a first conditional information representation; Based on the diffusion model, Gaussian noise is added to the filling target to generate noisy data; and based on the noisy data and the first conditional information representation as the input of the denoising network model, the parameters of the denoising network model are optimized through back propagation and gradient descent algorithms to obtain the optimized denoising network model.

[0010] Preferably, the filling sample acquisition process includes: An initial Gaussian noise sample is generated according to each missing value position in the test sample. Based on the optimized denoising network model, the noise is gradually removed through the reverse generation process to obtain the filling sample.

[0011] Preferably, the filling sample acquisition process further includes: Encoding the test sample through a representation mapping function to obtain a second conditional information representation; Each step in the reverse generation takes the second conditional information representation as the input of the optimized denoising network model.

[0012] Preferably, the step of combining the filled sample with the test sample includes: filling the filled sample into the position of the missing value corresponding to the test sample to obtain the filled test sample.

[0013] Preferably, the absolute residual of each sample in the calibration subset comprises: Each sample in the calibration subset is passed through the trained RUL prediction model to obtain the prediction result; Obtain the absolute residual for each sample in the calibration subset based on the actual result and the predicted result for each sample in the calibration subset.

[0014] Preferably, the defining a mapping network for mapping the monitoring data to the latent space comprises: defining a triplet network model consisting of three mapping networks sharing weights; The calibration subset is used to construct training samples including anchor samples, positive samples and negative samples to train the triplet network model and obtain a trained mapping network.

[0015] Preferably, when constructing the training samples from the calibration subset to include anchor samples, positive samples and negative samples, it includes: selecting an anchor sample from the calibration subset; selecting multiple neighbors with the closest absolute residual distance to the anchor sample as positive samples, and the remaining samples as negative samples.

[0016] Preferably, the RUL prediction model consists of an LSTM layer, a Dropout layer, a GELU activation function and a linear layer.

[0017] The present invention has at least the following beneficial effects: The present invention provides a method for quantifying the uncertainty of the remaining life of an aircraft engine in an online prediction scenario. The method effectively captures the characterization of aircraft engine sensor monitoring data in an online prediction scenario through a pre-training strategy. Based on the strategy, a self-supervised learning task is designed to train a diffusion model to fit the distribution of randomly missing and partially faulty sensor data, thereby achieving efficient and reliable data filling. The method can significantly improve the integrity of the data, ensure accurate RUL prediction in the case of missing data, and enhance the robustness of the comprehensive performance evaluation of aircraft engines.

[0018] The present invention introduces a residual pattern recognition method based on deep metric learning, and weights the conformal prediction conformity score according to the distance of the input data in the latent space, so as to construct a more accurate and reliable aircraft engine RUL prediction interval and achieve an accurate evaluation of the comprehensive performance of the aircraft engine. In addition, the construction process of the prediction interval does not rely on specific data distribution or model assumptions, and can flexibly adapt to the existing RUL prediction model based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the technical solution of the present invention.

[0020] Figure 2 This is a diagram of the pre-training model architecture of the present invention.

[0021] Figure 3 This is a diagram of the denoising network model architecture of the present invention.

[0022] Figure 4 Schematic diagram of the self-supervised training of the diffusion model of the present invention.

[0023] Figure 5 This is a diagram of the RUL prediction model architecture of the present invention.

[0024] Figure 6 This is a diagram of the triplet network model architecture of the present invention.

[0025] Figure 7 This is a diagram showing the RUL prediction interval effect of the test engine under data missing scenario 1 in subset FD001 of the present invention.

[0026] Figure 8 This is a diagram showing the RUL prediction interval effect of the test engine under data missing scenario 2 in subset FD001 of the present invention.

[0027] Fig. 9 This is a diagram showing the RUL prediction interval effect of the test engine under data missing scenario 3 in subset FD001 of the present invention.

[0028] Fig.10 This is a diagram showing the RUL prediction interval effect of the test engine under data missing scenario 4 in subset FD001 of the present invention. DETAILED DESCRIPTION

[0029] In order to illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description in conjunction with embodiments.

[0030] The present invention aims to solve the problem that existing deep learning methods cannot effectively quantify the uncertainty of prediction results in aircraft engine RUL prediction, as well as the negative impact of random missing and partial sensor failure data on the performance of existing methods in online prediction scenarios. The purpose of the present invention is to provide a method for quantifying the uncertainty of aircraft engine remaining life in online prediction scenarios, which aims to accurately evaluate the comprehensive performance of aircraft engines and improve the reliability and robustness of prediction results.

[0031] See also Figure 1As shown in the figure, a method for quantifying the uncertainty of the remaining life of an aircraft engine in an online prediction scenario mainly includes (1) representation learning based on a pre-training strategy, which obtains the representation of sensor monitoring data through the pre-training strategy and provides effective feature input for subsequent distribution fitting tasks; (2) distribution fitting based on a diffusion model, that is, training the diffusion model by constructing a self-supervised learning task to fit the distribution of randomly missing and partially faulty sensor data in the online prediction scenario; (3) residual pattern recognition based on deep metric learning, which mainly identifies the residual pattern in RUL prediction through deep metric learning and maps the input data with similar residuals to the latent space. , making these data close to each other in the space, while input data with dissimilar residuals are far away from each other in the space; (4) Construction of RUL prediction interval based on conformal prediction; (5) Online prediction: For aviation engine test samples with random missing and partial sensor failure data in the online prediction scenario, data filling is performed through distribution fitting method to obtain complete sensor monitoring data; according to the distance of the input data in the latent space, the conformal prediction compliance score is weighted to construct the RUL prediction interval of the test sample, accurately quantify the uncertainty of the prediction result, and realize accurate evaluation of the comprehensive performance of the aviation engine.

[0032] In order to achieve the above purpose, a method for quantifying the uncertainty of the remaining life of an aircraft engine in an online prediction scenario includes: S1. Acquire a training data set of an aircraft engine during test and operation, wherein the training data set includes monitoring data and a remaining life label corresponding to the monitoring data; Specifically, monitoring data of the aircraft engine during the test run and operation is obtained; the monitoring data is preprocessed to obtain a training data set; wherein the training data set includes the monitoring data and the remaining life label corresponding to the monitoring data; In the aircraft engine system, first correctly install and connect key sensors such as temperature, pressure, and speed. Next, configure the sensor's data transmission method, sampling frequency, and data storage format to ensure that the sensor can work normally and collect data stably. Through these sensors, the monitoring data of the aircraft engine is continuously recorded, and the collected data is effectively stored. Furthermore, the Min-Max normalization method is used to normalize the collected sensor monitoring data to eliminate the influence of different dimensions and ensure that various types of data can be compared and analyzed on a unified scale. At the same time, the sliding window technology is used to accurately capture the data correlation between adjacent time points, and the sensor monitoring data is converted into the required sample size to construct a training data set. .in, express N Sensor monitoring data of each engine running to failure, Representation and The corresponding remaining life label.

[0033] No. The monitoring data of an engine is expressed as ,in Indicates The total number of monitoring time steps for an engine from operation to failure, express Middle The sampled data of time steps, K is the total number of monitoring sensors.

[0034] In this embodiment, the training data set is complete and does not contain any missing data. represents an aircraft engine test sample in an online prediction scenario. This sample has two types of data missing: random missing and partial sensor failure. Assume that the proportion of random missing data is The proportion of sensor failure is .

[0035] The present invention aims to train the data set Construct a prediction model and apply it to aircraft engine test samples with random missing and partial sensor failure data in online prediction scenarios Quantify the uncertainty of RUL prediction.

[0036] S2. Characterize the monitoring data through a pre-training strategy according to the monitoring data to obtain a representation mapping function; The characterization mapping function is obtained according to the following steps: The monitoring data is masked and used as input, and the pre-trained model is trained until convergence to obtain a trained pre-trained model; wherein the pre-trained model includes input encoding, Transformer encoder and linear layer in sequence; the linear layer is removed from the trained pre-trained model to obtain the representation mapping function.

[0037] In this embodiment, the sensor monitoring data Perform mask processing to obtain masked data , as the input of the pre-trained model, the specific steps of mask processing are as follows: First, the mask ratios of random missing and sensor failure are set as and ; Furthermore, for sensor monitoring data Each sample in , generate random missing mask matrix and sensor fault mask matrix ,in L express The total number of monitoring time steps, K Indicates the total number of monitoring sensors, mask value , In the mask matrix In the mask value Bernoulli distribution Generate. In the mask matrix In, randomly select sensors, assuming that the index set of these sensors is D For all j belong D The sensor of the sensor, the mask value of all time steps of the sensor are all set to 0; otherwise, the mask value Set to 1. Also consider the random missing mask matrix and sensor fault mask matrix , generate the comprehensive mask matrix ,in Represents the matrix and Perform logical AND operations on the corresponding position elements; Furthermore, the sensor monitoring data With the integrated mask matrix Multiply the corresponding position elements of to obtain the monitoring data after mask processing ,Right now ,in Indicates the multiplication of elements at corresponding positions. Masked data Will be used as input to the pre-trained model.

[0038] See also Figure 2 As shown in Figure 1, the architecture of the pre-trained model mainly includes three parts: input encoding, Transformer encoder and linear layer. First, the masked sensor monitoring data is As the input of the model, it undergoes input encoding. The input encoding is done by 1×8 and K ×1 convolutional layer to extract local temporal features and correlations between variables. M A 1×8 temporal convolution filter is applied to the input data to extract the temporal pattern. indivual K ×1 spatial convolution filter performs a secondary convolution on the output of the temporal convolution to capture the correlation between variables in the sensor data, constructing a dimension of Then, The layer Transformer encoder captures the complex spatiotemporal dependencies in the data through a multi-head attention mechanism, a feedforward neural network, and a normalization layer. Finally, the encoded representation Pass the linear layer to generate the prediction result of the masked value .

[0039] The optimization goal of the pre-trained model is to minimize the predicted value and the true value The mean square error between The definition is as follows:

[0040] in, N Indicates sensor monitoring data The total number of samples in express The set of positions of all masked values ​​in , Representing a collection Then, the model parameters are optimized by back propagation and gradient descent algorithm so that the loss function minimize.

[0041] After the model pre-training is completed, remove the linear layer and construct the representation mapping function , used to obtain sample representation ,in The dimension of sample representation.

[0042] S3, based on the diffusion model, according to the monitoring data, and combined with the characterization mapping function, a denoising network model is obtained; The denoising network model is obtained according to the following steps: The monitoring data is divided into two parts: filling targets and observation values; Encoding the observed value through a representation mapping function to obtain a first conditional information representation; Based on the diffusion model, Gaussian noise is added to the filling target to generate noisy data; and based on the noisy data and the first conditional information representation as the input of the denoising network model, the parameters of the denoising network model are optimized through back propagation and gradient descent algorithms to obtain the optimized denoising network model.

[0043] It should be noted that since it is impossible to obtain the true labels of missing data in the online prediction scenario, this embodiment constructs a self-supervised learning task to train the diffusion model through the complete sensor monitoring data in the training data set to fit the distribution of randomly missing and partially faulty sensor data, thereby achieving effective data filling. The diffusion model is a generative model that aims to learn the high-dimensional and complex distribution of data and generate new samples. Its core concept is to gradually transform data into noise through the forward diffusion process, and then reconstruct the data distribution from the noise through the reverse generation process and generate new samples.

[0044] In this embodiment, Represents a complete monitoring data sample in the training data set and divides it into two parts: observations and fill target The data partitioning method is similar to the mask processing method described in S2, by generating a comprehensive mask matrix , we can determine the observed value , fill target ,in Represents the multiplication of elements at corresponding positions. The training goal of the diffusion model is to fit the filling target The distribution of is assumed to be .

[0045] In the forward diffusion process, Gaussian noise is gradually Add to fill target In the process, the T It can be represented by a Markov chain with time steps:

[0046] in, In From time step 1 to T Markov transition of . Noise data It also follows a Gaussian distribution:

[0047] in, and denote the mean and covariance of Gaussian distribution respectively; , , is an adjustable hyperparameter that controls the level of noise added during each diffusion process. T After diffusion, The distribution of will eventually be perturbed into a standard Gaussian distribution .

[0048] In the reverse generation process, from the standard Gaussian distribution Sampling in , which is then progressively denoised to reconstruct the filled target In order to make full use of the observed values The information is represented by the mapping function constructed in S2 Encode it to obtain the first conditional information representation Then, As the generation condition information, it is introduced into the reverse generation process. The reverse generation process can be expressed as the following Markov chain:

[0049] in, In Represents the time step arrive The Markov transition It is a parameterized model that gradually removes Gaussian noise. This parameterized model also obeys Gaussian distribution:

[0050] Its mean and covariance They are as follows:

[0051]

[0052] in, ; ; ; .

[0053] is a denoising network, a neural network model that represents , time step t and the noisy data at that time step To predict Gaussian noise .

[0054] See also Figure 3 As shown in the figure, the overall architecture of the denoising network model is shown, in which the specific dimensions of each output tensor are marked. Figure 3 middle, B represents the batch size of training samples, C represents the number of residual channels of the network, K represents the total number of monitoring sensors (spatial dimension), L represents the total time steps (time dimension). Specifically, the denoising network consists of Residual Channel C The core of each residual layer is the time-space attention mechanism, which introduces two Transformer encoders with different encoding directions to process the dimensions of and The tensor input is used to capture the temporal and spatial dependencies in multivariate sensor monitoring data. In addition, in order to achieve efficient feature extraction, a 1×1 convolution layer is further introduced in each residual layer, and the nonlinear representation ability of the network is enhanced through a gated activation layer. For residual layer 1, its main input is the noisy data , the input is processed by a 1×1 convolutional layer and a GELU activation function and then input into the residual layer 1. For each subsequent residual layer, the output of the previous residual layer is used and combined with the residual connection of the previous residual layer as its main input. In addition, each residual layer also receives two additional inputs, namely the diffusion time step t Information and conditional information representation , these additional inputs are processed through a linear layer and a GELU activation function. For each diffusion time step , using a dimension of The encoding vector as the embedding of the current time step. The calculation method of this encoding vector can be expressed as , ,in Finally, the skip connections of all residual layers are added together, and then processed through a 1×1 convolution layer, a normalization layer, and a GELU activation function to generate the final noise prediction result.

[0055] Furthermore, the diffusion model is trained by solving the following optimization problem. The optimization goal is to make the denoising network The prediction result is as close as possible to the actual Gaussian noise added , and optimize the parameters of the denoising network model through back propagation and gradient descent algorithms:

[0056] See also Figure 4 As shown in the figure, the self-supervised training process of the diffusion model is summarized as follows: First, the complete monitoring data in the training set Divide into observations and fill target , and through the characterization mapping function constructed in S2 For the observed value Encode to obtain the first conditional information representation Next, according to the given noise level , the Gaussian noise Add to fill target Generate noisy data ,in, , At this time, the noise data , First condition information representation and time step t Constructing the denoising network The input is used to predict the time step t Gaussian noise added Finally, minimize the loss function , optimize the parameters of the denoising network model through back propagation and gradient descent algorithm , that is, obtain the optimized denoising network model.

[0057] S4, dividing the training data set into a training subset and a calibration subset; The RUL prediction model is trained according to the training subset to obtain a trained RUL prediction model; A mapping network that maps monitoring data to latent space is constructed through the calibration subset; The defining of a mapping network for mapping monitoring data to a latent space includes: defining a triplet network model consisting of three mapping networks sharing weights; The calibration subset is used to construct training samples including anchor samples, positive samples and negative samples to train the triplet network model and obtain a trained mapping network.

[0058] When constructing the training samples of the calibration subset including anchor samples, positive samples and negative samples, the method includes: selecting an anchor sample from the calibration subset; selecting multiple neighbors with the closest absolute residual distance to the anchor sample as positive samples, and the remaining samples as negative samples.

[0059] The RUL prediction model consists of an LSTM layer, a Dropout layer, a GELU activation function, and a linear layer. Figure 5 shown.

[0060] The present invention identifies residual patterns in RUL prediction through deep metric learning, maps input data with similar residuals to a latent space so that these data are close to each other in the space, while input data with dissimilar residuals are far away from each other in the space.

[0061] In this embodiment, it will include N training data of samples Divided into two subsets: Training subset and calibration subset ,in .

[0062] pass Training a RUL prediction model , the model takes as input the feature space Mapping to output space , and solve the RUL prediction task by minimizing the mean square error loss function. It should be emphasized that It can be based on the existing RUL prediction model based on deep learning methods. It is used for deep metric learning to identify residual patterns for RUL prediction and for conformal prediction to construct prediction intervals.

[0063] Complete the training Application , get the prediction results , and calculate the absolute residual ,in In addition, remove The last linear layer is used to construct the embedding function ,in is the embedding dimension.

[0064] Furthermore, through the continuous triple construction method, training sample pairs are constructed, including anchor samples , positive sample and negative samples :for Given a small batch training sample set in , select an anchor sample , and select its absolute residual label The nearest k neighbors as the positive sample set The remaining samples are used as negative sample sets Then, all possible sample pairs are constructed through the positive sample set and the negative sample set , and compare it with the anchor sample Combining to form triples , to ensure that the model observes as diverse triplets as possible during training.

[0065] Deep metric learning maps sensor monitoring data to latent space through a triplet network model. The triplet network model consists of three sub-networks with shared weights. Composition, see Figure 6 As shown. Each sub-network receives anchor samples separately , positive sample and negative samples As input, it is passed through LSTM layer and Dropout layer for feature extraction. The extracted features are then embedded with the corresponding samples , and The features are further processed through linear layers and Dropout layers, and nonlinear information is introduced through the GELU activation function to map the input data to the latent space and generate sample representations in the latent space. , and .

[0066] Furthermore, the triplet network model is trained using the Log-ratio loss function. is defined as follows:

[0067] in, represents the Euclidean distance, , and Represent the absolute residuals of anchor samples, positive samples and negative samples respectively. After the training of the triplet network model is completed, the mapping network is obtained , the input data sample points are mapped to a latent space. In this latent space, the Euclidean distance between points reflects the similarity of their absolute residuals, thus achieving effective residual pattern recognition.

[0068] S5, collecting test samples with random missing and sensor failure data in the online prediction scenario; Obtain a filled sample according to the test sample through the optimized denoising network model; combine the filled sample with the test sample to obtain a filled test sample; The filling sample acquisition process includes: An initial Gaussian noise sample is generated according to each missing value position in the test sample. Based on the optimized denoising network model, the noise is gradually removed through the reverse generation process to obtain the filling sample.

[0069] The filling sample acquisition process also includes: Encoding the test sample through a representation mapping function to obtain a second conditional information representation; Each step in the reverse generation takes the second conditional information representation as the input of the optimized denoising network model.

[0070] The combining of the filling sample with the test sample includes: filling the filling sample into the position of the missing value corresponding to the test sample to obtain the filled test sample.

[0071] The present invention fills data for aviation engine test samples with random missing and partial sensor failure data in an online prediction scenario through a distribution fitting method to obtain complete sensor monitoring data.

[0072] In this embodiment, the test samples with random missing and sensor failure data in the online prediction scenario are As an observed value , through the representation mapping function constructed in S2 Encode it to get the second condition information representation .

[0073] against Generate an initial Gaussian noise sample for each missing value position in , and gradually remove the noise through the reverse generation process. Each step of the generation process represents the second condition information As a conditional input, an estimate of the missing value is generated, which can be expressed as:

[0074] in, ; ; ; ; ; is a given noise level.

[0075] This is the denoising network model trained in S3. After multiple iterations, the final filling result is obtained. By combining and , obtain complete aviation engine sensor monitoring data in online prediction scenarios .

[0076] S6, mapping the samples in the calibration subset and the padded test samples to the latent space through the mapping network, and obtaining the weight of each sample in the calibration subset; Construct a conformity score based on the absolute residual of each sample in the calibration subset, and weight the conformity scores by the weights of different samples to define a weighted empirical distribution function; The absolute residual for each sample in the calibration subset includes: Each sample in the calibration subset is passed through the trained RUL prediction model to obtain the prediction result; Obtain the absolute residual for each sample in the calibration subset based on the actual result and the predicted result for each sample in the calibration subset.

[0077] In this embodiment, for the calibration subset Samples in and filled test samples in online prediction scenarios , according to the mapping network obtained in S4 , map these samples into the latent space and calculate the weight of each sample in the calibration subset :

[0078] in, , Denotes the Euclidean distance. Weight Reflects the calibration sample With test samples The similarity between the absolute residuals of RUL predictions. The larger the value, the more similar the absolute residuals are, which means that the corresponding calibration sample is more important for generating the prediction interval of the test sample; conversely, the larger the difference in the absolute residuals, the less reference value the corresponding calibration sample has for generating the prediction interval of the test sample. In other words, if the test sample Calibration samples that are closer in residual pattern to some samples in the calibration subset are assigned higher weights.

[0079] According to the calibration subset The absolute residual set of samples in constructs the conformity score , and by weights of different samples Weight the conformity scores and define the weighted empirical distribution function :

[0080] in, ,when The value is 1 when , otherwise it is 0. The residuals of are unknown, so Used in the definition of To cover the possible residual values ​​of the new sample, that is , .

[0081] S7. For the filled test samples, conformal prediction is performed to construct the RUL prediction interval according to the trained RUL prediction model and the weighted empirical distribution function.

[0082] In this embodiment, the conformality score of the conformal prediction is weighted according to the distance of the input data in the latent space, so as to construct the RUL prediction interval of the test sample and accurately quantify the uncertainty of the prediction result to achieve an accurate evaluation of the comprehensive performance of the aircraft engine.

[0083] For the padded test samples in the online prediction scenario , RUL prediction interval constructed by conformal prediction It can be expressed as:

[0084] in, Indicates that through the training subset Trained RUL prediction model; represents the weighted empirical distribution function of Quantile, is the confidence level parameter.

[0085] The prediction interval construction method constructs the prediction interval by dividing the data set and calculating the quantile of the residual, without relying on specific data distribution or model assumptions, so it can flexibly adapt to the existing RUL prediction model based on deep learning. By analyzing the RUL prediction interval, the comprehensive performance of the aircraft engine can be accurately evaluated.

[0086] In order to further illustrate the uncertainty quantification method of the remaining life of an aircraft engine in an online prediction scenario provided by the present invention, it is described in conjunction with a specific embodiment.

[0087] In this embodiment, the C-MAPSS data set of aircraft engine sensor monitoring provided by NASA is used as a benchmark to verify the technical solution of the present invention. The C-MAPSS data set simulates the operation process of aircraft engines under different operating conditions and failure modes, and includes four subsets, as shown in Table 1.

[0088] Table 1 Detailed information of the C-MAPSS dataset

[0089] Each subset contains training engine and test engine data. This embodiment builds a prediction model through the training engine data in the subset and quantifies the uncertainty of RUL prediction for the test engine in the online prediction scenario. In the online prediction scenario, in order to more comprehensively evaluate the prediction effect of the model under different data missing modes of random missing and sensor failure, four data missing scenarios are set for the test engine in the C-MAPSS data set, which are respectively set as scenario 1: random missing ratio , Sensor failure ratio ; Case 2: Random missing proportion , Sensor failure ratio ; Scenario 3: Random missing proportion , Sensor failure ratio ; Case 4: Random missing proportion , Sensor failure ratio When quantifying the uncertainty of RUL prediction, the present invention sets the confidence level , which means that the constructed prediction interval is required to meet 90% coverage.

[0090] It should be noted that the present invention evaluates the performance of the model in RUL prediction through the root mean square error (RMSE) and the scoring function (Score). The smaller the RMSE and Score values, the better the RUL prediction effect of the model. In addition, the present invention also evaluates the performance of the model in uncertainty quantification in RUL prediction through the prediction interval coverage (PICP), the prediction interval standardized average width (PINAW) and the coverage width criterion (CWC). Specifically, the larger the PICP value, the higher the coverage of the prediction interval and the more reliable the model is in terms of uncertainty quantification; the smaller the PINAW value, the narrower the average width of the prediction interval and the more accurate the model is in terms of quantifying uncertainty; CWC is an indicator for comprehensively evaluating the quality of the prediction interval. The larger the CWC value, the higher the quality of the prediction interval, and the model can more effectively balance the width and coverage of the prediction interval, thereby improving the accuracy and reliability of the prediction interval.

[0091] Table 2 shows the effect of the method proposed in this embodiment on the RUL prediction and uncertainty quantification of the test engine under different data missing situations.

[0092] Table 2 RUL prediction and uncertainty quantification effects under different data missing scenarios in the C-MAPSS dataset

[0093] As can be seen from Table 2, as the proportion of missing data increases, the RUL prediction effect of the model remains stable overall, and there is no significant decline in prediction performance as the proportion of missing data increases. At the same time, despite the missing data, the model can still maintain a high PICP value. In particular, when the proportion of missing data is low, its PICP performance is close to that of complete data, indicating that the model can still reliably quantify the uncertainty of the prediction results and ensure the coverage of the prediction interval when data is missing. In terms of PINAW and CWC, as the proportion of missing data increases, the width of the prediction interval always remains within a reasonable range. In addition, the CWC value always maintains a high level, indicating that the model performs well in balancing the width and coverage of the prediction interval and can generate high-quality prediction intervals.

[0094] In general, the method proposed in the present invention demonstrates excellent robustness and high precision when processing randomly missing and partially faulty sensor data in online prediction scenarios, ensures the coverage and width of the prediction interval in the uncertainty quantification process, and demonstrates its significant advantages in accuracy and reliability.

[0095] In order to more intuitively demonstrate the effect of the proposed method in RUL prediction and uncertainty quantification, the RUL prediction intervals of all test engines under different data missing conditions are visualized in the subset FD001. The results are shown in Figure 2. Figures 7 to 10 As shown in the figure, the horizontal axis represents the serial number of the test engine, which is arranged in ascending order according to the actual RUL value; the vertical axis is the RUL value. The curve in the figure represents the actual RUL value of the engine, the square represents the predicted value, and the vertical lines on both sides of the square represent the corresponding prediction interval. Figure 7 It can be seen that most of the true RUL values ​​are successfully covered by the prediction interval, indicating that the model has high reliability in uncertainty quantification. At the same time, the model shows significant adaptability: for points with smaller prediction errors, the prediction interval is narrower, reflecting the accuracy of the model under high confidence; for points with larger prediction errors, the prediction interval is relatively wide to better cover the actual value. This adaptability highlights the advantages of the method proposed in this invention in uncertainty quantification: when the model is more certain about the prediction results, the prediction interval will shrink; when the uncertainty is higher, the interval will expand appropriately to ensure that the true value can be covered. Figure 8 , Fig. 9 and Fig.10 It can be seen that as the proportion of missing data increases, the model still maintains the adaptive adjustment of the prediction interval. For points with smaller prediction errors, the prediction interval is still narrow, while for points with larger errors, the prediction interval is wider.

[0096] In summary, when there is less data missing, the method proposed in the present invention can provide accurate and reliable prediction intervals; and as the proportion of data missing increases, although the prediction interval widens and the coverage rate decreases slightly, the model still shows high reliability overall. This shows that the method proposed in the present invention can effectively balance the accuracy and reliability of the prediction interval under different degrees of data missing, showing good practicality and robustness.

[0097] Based on accurate and reliable RUL prediction intervals, a comprehensive and scientific evaluation of the comprehensive performance of aircraft engines can be carried out. When the RUL prediction value is large and the prediction interval is narrow, it indicates that the expected life of the engine is long and the uncertainty of the prediction is low. In this case, it can be judged that its comprehensive performance is good and its health status is stable. For example, in Figure 7In the data missing scenario 1, for the engine No. 87 of the subset FD001, the RUL prediction value is 120, and the prediction interval is [112, 129]. Since the RUL prediction value is high and the confidence interval is narrow, it means that the health status of the engine is relatively stable, the operating reliability is strong, and the overall performance is good. However, when the RUL prediction value is large but the prediction interval is wide, although the expected life of the engine is long, due to the high prediction uncertainty, it is impossible to make a clear assessment of its overall performance. In this case, we should be vigilant about the health status of the engine and plan the maintenance window reasonably. For example, in Figure 7 In the example, the RUL prediction value of engine No. 11 is 41, and the prediction interval is [13, 68]. Although its expected life seems sufficient, the large uncertainty indicates that the uncertainty of its operating status is high. Therefore, it is necessary to focus on monitoring this engine and formulate a maintenance strategy in advance to reduce the operating risk. For engines with small RUL prediction values ​​and narrow prediction intervals, it can be clearly determined that their overall performance is poor, their health status has entered an abnormal stage, and maintenance needs to be arranged immediately. For example, Figure 7 The engine with serial number 6 has a predicted RUL value of 8 and a prediction interval of [3, 13]. Since its expected life is significantly insufficient and the prediction uncertainty is low, it can be determined that its operating capacity can no longer meet normal requirements. Therefore, maintenance measures should be implemented as soon as possible to prevent the occurrence of potential failures and ensure the safety and reliability of system operation. Through the above analysis, it can be seen that the comprehensive performance evaluation of aircraft engines based on the RUL prediction interval can not only provide decision makers with accurate health status judgment, but also effectively identify potential risks and optimize maintenance strategies, thereby improving the scientificity and execution efficiency of health management and maintenance decisions.

[0098] The method proposed in this paper performs well in processing randomly missing and partially faulty sensor data in online prediction scenarios, demonstrating high accuracy and strong robustness in uncertainty quantification in aircraft engine RUL prediction. In addition, this method is distribution-independent and model-independent, and can flexibly adapt to existing deep learning-based RUL prediction models, providing accurate and reliable prediction intervals, thereby achieving accurate evaluation of the comprehensive performance of aircraft engines and providing important support for health management and maintenance decisions.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario, characterized in that: include: Acquire a training data set of an aircraft engine during a test run and operation process, wherein the training data set includes monitoring data and a remaining life label corresponding to the monitoring data; Characterize the monitoring data through a pre-training strategy according to the monitoring data to obtain a representation mapping function; Based on the diffusion model, according to the monitoring data and combined with the characterization mapping function, a denoising network model is obtained; Divide the training data set into a training subset and a calibration subset; The RUL prediction model is trained according to the training subset to obtain a trained RUL prediction model; A mapping network that maps monitoring data to latent space is constructed through the calibration subset; Collect test samples with random missing and sensor failure data in online prediction scenarios; Obtain a filled sample according to the test sample through the optimized denoising network model; combine the filled sample with the test sample to obtain a filled test sample; The samples in the calibration subset and the padded test samples are mapped to the latent space through the mapping network to obtain the weight of each sample in the calibration subset; Construct a conformity score based on the absolute residual of each sample in the calibration subset, and weight the conformity scores by the weights of different samples to define a weighted empirical distribution function; For the filled test samples, the RUL prediction interval is constructed by conformal prediction based on the trained RUL prediction model and the weighted empirical distribution function.

2. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 1 is characterized in that: The characterization mapping function is obtained according to the following steps: The monitoring data is masked and used as input, and the pre-trained model is trained until convergence to obtain a trained pre-trained model; wherein the pre-trained model includes input encoding, Transformer encoder and linear layer in sequence; the linear layer is removed from the trained pre-trained model to obtain the representation mapping function.

3. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 1 is characterized in that: The denoising network model is obtained according to the following steps: The monitoring data is divided into two parts: filling targets and observation values; Encoding the observed value through a representation mapping function to obtain a first conditional information representation; Based on the diffusion model, Gaussian noise is added to the filling target to generate noisy data; The noisy data and the first conditional information representation are used as the input of the denoising network model, and the parameters of the denoising network model are optimized through back propagation and gradient descent algorithms to obtain the optimized denoising network model.

4. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 1, characterized in that: The filling sample acquisition process includes: An initial Gaussian noise sample is generated according to each missing value position in the test sample. Based on the optimized denoising network model, the noise is gradually removed through the reverse generation process to obtain the filling sample.

5. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 4 is characterized in that: The filling sample acquisition process also includes: Encoding the test sample through a representation mapping function to obtain a second conditional information representation; Each step in the reverse generation takes the second conditional information representation as the input of the optimized denoising network model.

6. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 4 is characterized in that: The combining of the filling sample with the test sample includes: filling the filling sample into the position of the missing value corresponding to the test sample to obtain the filled test sample.

7. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 1, characterized in that: The absolute residual for each sample in the calibration subset includes: Each sample in the calibration subset is passed through the trained RUL prediction model to obtain the prediction result; Obtain the absolute residual for each sample in the calibration subset based on the actual result and the predicted result for each sample in the calibration subset.

8. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 7 is characterized in that: The defining of a mapping network for mapping monitoring data to a latent space includes: defining a triplet network model consisting of three mapping networks sharing weights; The calibration subset is used to construct training samples including anchor samples, positive samples and negative samples to train the triplet network model and obtain a trained mapping network.

9. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 8, characterized in that: When constructing the training samples of the calibration subset to include anchor samples, positive samples and negative samples, the method includes: selecting an anchor sample from the calibration subset; selecting multiple neighbors with the closest absolute residual distance to the anchor sample as positive samples, and the remaining samples as negative samples.

10. The method for quantifying uncertainty of remaining life of an aircraft engine in an online prediction scenario according to claim 1, characterized in that: The RUL prediction model consists of an LSTM layer, a Dropout layer, a GELU activation function and a linear layer.

Images