Method for adaptive selection of sensors for aircraft engine prognostic health management

Abstract

The application relates to an airplane engine predictive health management method based on sensor adaptive selection, which comprises the following steps: acquiring airplane engine failure original data and constructing a training sample to train a deep learning model, the deep learning model performs adaptive selection on failure signal data, failure mode recognition is performed by using a sensor selection weight, meanwhile, working condition irrelevant features are extracted, a mapping relationship between the features and RUL prediction values under different failure modes is established, and finally, the RUL prediction values are obtained in combination with the failure mode recognition results; failure data of an airplane engine to be predicted is acquired, and the trained deep learning model is used to realize failure mode recognition and RUL prediction results. Compared with the prior art, the application can automatically adjust the sensor selection weight according to different working conditions and failure modes, can extract working condition irrelevant features, and is helpful for more accurately realizing combined failure mode recognition and RUL prediction.

Description

Technical Field

[0001] This invention relates to the field of predictive and health management of aircraft engines, and in particular to a predictive health management method for aircraft engines based on sensor adaptive selection. Background Technology

[0002] Predictive health management (PHM) for manufacturing systems can assist in maintenance planning and safety inspection decisions, effectively preventing unexpected equipment failures and extending equipment lifespan. Failure mode identification and remaining useful life (RUL) prediction based on sensor signals are important research topics in PHM. In complex manufacturing systems, the coupling of multiple failure modes and operating conditions poses challenges to this research. Furthermore, some sensors are not related to degradation, thus requiring sensor screening to eliminate interference from irrelevant sensors.

[0003] Supervised methods for failure mode identification (FMT) include K-nearest neighbors, support vector machines (SVM), and Naive Bayes. However, in real-world scenarios, FMT labels may be unknown. Considering the time-consuming and laborious nature of manual labeling, in practice, labels are typically assigned to a subset of samples for semi-supervised learning. Methods include Fisher Linear Discriminant Analysis (FLDA), graphical models, and manifold regularization. Methods for right-to-close (RUL) prediction can be categorized into machine learning and deep learning methods. Machine learning methods include linear regression and health index methods. Deep learning methods include deep neural networks (DNN), convolutional neural networks (CNN), and long short-term memory networks (LSTM). However, these methods primarily rely on visual inspections and subjective evaluations for sensor selection, failing to adapt to different samples, operating conditions, and failure modes. Furthermore, these FMT and RUL prediction methods do not consider scenarios involving multiple operating conditions and coupled failure modes. Finally, some existing research treats FMT and RUL prediction as two independent objectives. However, since the mapping between sensor data and RUL differs significantly across different failure modes, methods are needed to characterize the complex relationship between these two objectives and perform joint prediction.

[0004] In summary, the challenges and research gaps in failure mode identification and range-limited liability (RUL) prediction for aircraft engines under multiple operating conditions are as follows: First, how to achieve adaptive sensor screening to remove degraded and irrelevant sensor information. Existing methods rely on visual inspection and subjective evaluation, which cannot achieve automatic sensor screening. Second, different failure modes and operating conditions pose challenges to sensor screening, failure mode identification, and RUL prediction; existing research has not considered the scenario where failure modes and operating conditions are coupled. Third, for cases with unknown failure mode labels, in practice, partial samples are usually labeled for semi-supervised learning, but existing semi-supervised learning methods do not consider joint failure mode identification and RUL prediction under multiple operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a predictive health management method for aircraft engines based on sensor adaptive selection. This method utilizes multi-sensor signal data collected during aircraft engine operation that reflects its degradation state to achieve adaptive selection of sensors for in-service aircraft engines, identification of failure modes, and prediction of RUL (Recovery and Unstable Duration).

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A sensor-adaptive selection-based predictive health management method for aircraft engines includes the following steps:

[0008] S1. Acquire aircraft engine failure data, time-varying operating condition data, failure mode data and real RUL data collected by multi-dimensional sensors as raw data;

[0009] S2. Use the time window method to process the original data to obtain training samples of failed data;

[0010] S3. The deep learning model is trained based on the failure data training samples. The deep learning model adaptively selects the input aircraft engine failure data based on the sensor selection weights, and uses the failure mode recognition network constructed according to the sensor selection weights to perform failure mode recognition, extracts the condition-independent features of the input after sensor selection, establishes the mapping relationship between the condition-independent features under different failure modes and the RUL prediction value, and obtains the final RUL prediction value by combining the failure mode recognition results.

[0011] S4. Acquire the failure data of the aircraft engine to be predicted, input it into the trained deep learning model, adaptively select the sensors of the in-service aircraft engine, and output the failure mode recognition and RUL prediction results under time-varying operating conditions.

[0012] In step S2, let the aircraft engine failure data collected by the multi-dimensional sensor be represented as... Where n represents the observation time step, M represents the number of observation sensors, and the engine failure mode is represented by a one-heat vector. Where K represents the number of failure modes, the observed true RUL is The aircraft engine failure matrix after processing using the time window method is represented as follows: This includes the last C steps of observation data in x, where C represents the length of the time window. In x, t n Observational data at time step -C+c; Record the working conditions within the corresponding time window Where P represents the number of operating conditions. Indicates the original observation t n The unique heat vector of the operating condition at time step -C+c.

[0013] The sensor selection function Represented as:

[0014]

[0015] in This represents the result after selection, w = [w1, ..., w C ]∈R C×M The weighting matrix representing sensor selection, w c ∈R 1×M express The corresponding selection weight, ⊙ represents the Hadamard product.

[0016] The sensor selection weights are dynamically generated using an attention mechanism based on sensor data and subsequent targets. The weight acquisition function... as follows:

[0017]

[0018] in, Representing the hidden states of a Long Short-Term Memory (LSTM) network, deep neural networks Used for alignment And the dimension of w, θ w express The parameter set, ζ represents the ReLU activation function, and the output after activation is denoted as j = [j1, ..., j]. C ]∈R C×M The weights are then standardized to the range (0,1) using a soft-maximum function σ.

[0019]

[0020] Where, j c,m Indicate j c The m-th element in w c,m Represents sensor m at t n The selection weight corresponding to time step -C+c, w c =[w c,1 ,…,w c,M ].

[0021] The failure mode identification network selects the last item w based on the sensor selection weight matrix. C Build:

[0022]

[0023] Where z = [z1,…,z] K ]∈R 1×K Represents the failure mode probability vector. Represents a deep neural network, θ r express The parameter set.

[0024] The extraction of condition-independent features from the sensor-selected input data specifically involves: utilizing a long short-term memory network. Data after sensor selection Perform feature extraction:

[0025]

[0026] Among them, features R represents f The hidden state, d2 represents the dimension of the feature, C represents the length of the time window, θ f express The parameter set;

[0027] Meanwhile, a domain adaptation method is employed to minimize the domain bias caused by operating conditions in order to extract operating condition-independent features, and an operating condition classifier is constructed. The working condition classifier is trained using a loss function and its extracted features. As input, output the operating condition probability matrix:

[0028]

[0029] Where o = [o1,…,o] C ]∈R C×P Let O be the probability matrix of the working conditions, where O c =[o c,1 ,…,o c,P ]∈R 1×P Indicates t n The probability vector at time step -C+c, o c,p Indicates at t n The value of p for time step -C+c. Represents a deep neural network, θ d yes The parameter set is Ψ(·), which represents the gradient reversal layer, so that the value of the gradient remains unchanged during forward propagation of the neural network, while the sign of the gradient is changed during backward propagation.

[0030] The process of establishing a mapping relationship between condition-independent features under different failure modes and RUL predicted values, and combining this with the failure mode identification results to obtain the final RUL predicted value, specifically involves:

[0031] Through deep neural networks Construct features for each failure mode k and RULy k The mapping relationship, k = 1, ..., K, where K represents the number of failure modes, predicts y k The formula is as follows:

[0032]

[0033] Where θ y,k yes The parameter set;

[0034] Joint learning network The final RUL is predicted based on the RUL prediction results and failure mode probabilities across multiple failure modes:

[0035]

[0036] Where z = [z1,…,z] K ]∈R 1×K This represents the failure mode probability vector.

[0037] The total loss function of the deep learning model Represented as:

[0038]

[0039] Where, λ v , λ e , λ p and λ d This indicates the adjustment parameter; ρ = 1 indicates that the failure mode of the sample is labeled, and ρ = 0 indicates that the failure mode of the sample is not labeled. θ w ,θ f ,θ d ,θ y,1 ,…,θ y,K ,θ r These are the parameters of a deep learning model.

[0040] The function representing the prediction loss of RUL:

[0041]

[0042] in y is the actual RUL, and y is the predicted RUL;

[0043] The sensor selection weight loss function is represented as follows:

[0044]

[0045] Where ||·||2 represents the Euclidean distance, wc This represents the sensor selection weight, and C represents the length of the time window;

[0046] Let represent the failure mode recognition loss function based on semi-labeled samples, where, for samples with failure mode labels, the cross-entropy loss function is used for failure mode prediction. Right now:

[0047]

[0048] Where z k This represents the probability that the predicted failure mode belongs to k. This represents the actual failure mode; if the failure mode is k, then... Otherwise, it is 0, where K represents the number of failure modes;

[0049] For samples without failure mode labels, a loss function based on Shannon entropy is used for failure mode prediction.

[0050]

[0051] The loss function representing the extraction of features irrelevant to the operating conditions is expressed using the cross-entropy loss function for the operating condition prediction results.

[0052]

[0053] in Indicates at t n The actual operating condition at time step -C+c, if the actual operating condition is p then Otherwise, it is 0, o c,p Indicates the prediction at t n The probability that the condition at time step -C+c belongs to p, where p represents the number of conditions.

[0054] The deep learning model, during training, uses the total loss function Recorded as By minimizing This enables sensor selection, failure mode identification, and RUL prediction, while minimizing the operating condition loss function λ. d Training model parameters θ d To obtain an accurate working condition classifier And maximize the operating condition loss function λ d To train other parameters θ of the model w ,θ f ,θ y,1 ,…,θ y,K ,θ r This makes the extracted features The operating conditions cannot be accurately identified.

[0055] The training objective of the deep learning model is expressed as:

[0056]

[0057]

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

[0059] (1) This invention proposes an adaptive sensor selection technique that can automatically adjust weights for different operating conditions and failure modes. In addition, the sensor selection weights can serve as effective features for failure mode identification, thus eliminating the need for additional effort in failure mode feature extraction.

[0060] (2) This invention proposes a novel neural network to extract condition-independent features, which helps to achieve joint failure mode identification and RUL prediction more accurately.

[0061] (3) The present invention adopts a semi-supervised learning method and designs a failure mode recognition loss function based on semi-labeled samples in the overall loss function of the model, so that the proposed network can accurately identify failure modes when the failure mode labels of a large portion of the samples are unknown. Attached Figure Description

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

[0063] Figure 2 The diagram shows the sensor selection under different failure modes. (2a) shows the sensor selection result of an in-service aircraft engine under failure mode F, and (2b) shows the sensor selection result of an in-service aircraft engine under failure mode H.

[0064] Figure 3 This is a schematic diagram of the failure mode identification and RUL prediction results of the present invention, wherein (3a) shows the failure mode identification results and (3b) shows the RUL prediction results;

[0065] Figure 4 This is a diagram showing the accuracy of failure mode recognition under different training samples with labeled failure modes. Detailed Implementation

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

[0067] This embodiment provides a sensor-adaptive selection-based predictive health management method for aircraft engines, such as... Figure 1 As shown, it includes the following steps:

[0068] S1. Acquire aircraft engine failure data, time-varying operating condition data, failure mode data, and real RUL data collected by multi-dimensional sensors as raw data.

[0069] In this embodiment, the aircraft engine failure data obtained from the multi-dimensional sensor is represented as: Where n represents the observation time step, and M represents the number of observation sensors. The engine failure modes are represented by one-heat vectors. Where K represents the number of failure modes.

[0070] The true RUL of this observation is

[0071] S2. Use the time window method to process the original data to obtain training samples of failed data.

[0072] This embodiment uses the matrix obtained by applying the time window method to x. As input to the model, This includes the last C steps of observation data in x, where C represents the length of the time window. In x, t n Observational data at time step -C+c. Record the working conditions within the corresponding time window Where P represents the number of operating conditions. Indicates the original observation t N The unique heat vector of the operating condition at time step -C+c.

[0073] S3. The deep learning model is trained based on the failure data training samples. The deep learning model adaptively selects the input aircraft engine failure data based on sensor selection weights, and uses the failure mode recognition network constructed according to the sensor selection weights to perform failure mode recognition. The condition-independent features of the input after sensor selection are extracted, and the mapping relationship between the condition-independent features under different failure modes and the RUL prediction value is established. The final RUL prediction value is obtained by combining the failure mode recognition results.

[0074] (I) Sensor Selection

[0075] Sensor-level selection function in this invention The formula is as follows:

[0076]

[0077] in This represents the result after selection, w = [w1, ..., w C ]∈R C×M w represents the selection of the weight matrix. c ∈R 1×M express The corresponding selection weights are represented by ⊙, which indicates the Hadamard product. To achieve adaptive sensor selection under multiple failure modes and operating conditions, this invention employs an attention mechanism to dynamically generate weights based on sensor data and subsequent targets. Weight acquisition function. as follows:

[0078]

[0079] in Representing the hidden states of a Long Short-Term Memory (LSTM) network, deep neural networks Used for alignment The dimension of w is ζ, which represents the ReLU activation function. The output after activation is denoted as j = [j1, ..., j]. C ]∈R C×M The weights are then standardized to the range (0,1) using a soft-maximum function σ, as shown in the following formula:

[0080]

[0081] Where j c,m Indicate j c The m-th element in w c,m Represents sensor m at t n The selection weight corresponding to time step -C+c, w c =[w c,1 ,…,w c,M ]. θ w express The parameter set in.

[0082] (II) Failure Mode Identification and RUL Prediction

[0083] This invention constructs a joint learning network to simultaneously identify failure modes and predict Recovery Upgrade Limits (RUL). The degradation process of aircraft engines differs significantly under different failure modes, and the mapping between sensor data and RUL also differs. Therefore, the sensor selection weights corresponding to RUL prediction will also vary greatly under different failure modes. This invention is based on sensor selection weights w... C The failure mode recognition network is constructed using the following formula:

[0084]

[0085] Where z = [z1,…,z] K ]∈R1×K Represents the failure mode probability vector. Represents a deep neural network, θ r express The parameter set.

[0086] To achieve RUL prediction under multiple failure modes, this invention utilizes a long short-term memory network. Data after sensor selection Perform feature extraction:

[0087]

[0088] Among them, features R represents f The hidden state, d2 represents the dimension of the feature, c represents the length of the time window, θ f express The parameter set.

[0089] This invention then utilizes deep neural networks. To construct each failure mode k and RULy k The mapping relationship is given, and k = 1, ..., K. Predict y k The formula is as follows:

[0090]

[0091] Where θ y,k This is the parameter set of the model. Ultimately, the joint learning network... The final RUL is predicted based on the RUL prediction results and failure mode probabilities under multiple failure modes, as shown in the following formula:

[0092]

[0093] Failure mode identification and RUL prediction are both independent of operating conditions; therefore, the mapping between sensor data and these two targets should be insensitive to operating conditions. However, time-varying operating condition information is included in the extracted features, thus affecting the effectiveness of failure mode identification and RUL prediction. To extract operating condition-independent features, this invention employs a domain adaptation method to minimize the domain bias caused by operating conditions. Specifically, this is achieved by constructing an operating condition classifier. The working condition classifier is trained using a loss function and its extracted features. As input, output the operating condition probability matrix:

[0094]

[0095] Where o = [o1,…,o] C ]∈R C×PLet O be the probability matrix of the working conditions, where O c =[o c,1 ,…,o c,P ]∈R 1×P Indicates t n The probability vector at time step -C+c, o c,p Indicates at t n The value of p for time step -C+c. Represents a deep neural network, θ d yes The parameter set is Ψ(·), which represents the gradient reversal layer, so that the value of the gradient remains unchanged during forward propagation of the neural network, while the sign of the gradient is changed during backward propagation.

[0096] (III) Loss Function

[0097] 1) RUL prediction loss function

[0098] The function for predicting loss using RUL is denoted as The calculation formula is as follows:

[0099]

[0100] in y is the actual RUL, and y is the predicted RUL.

[0101] 2) Sensor selection weight loss function

[0102] If raw sensor data is used as input and sensor selection weights are learned without any constraints, the final sensor selection weights may contain a lot of noise and have poor interpretability. This invention uses a weight loss function to constrain the changes in weights over time, i.e.:

[0103]

[0104] Where ∥·∥2 represents Euclidean distance. It enables smooth sensor selection weights, making them more stable and interpretable.

[0105] To illustrate the effectiveness of the sensor selection, Figure 2 Examples of sensor selection under different failure modes are given in the paper. Figure (2a) shows the sensor selection results of an in-service aircraft engine under failure mode F, and Figure (2b) shows the sensor selection results of an in-service aircraft engine under failure mode H. The horizontal axis represents time, and the vertical axis represents different sensors. The screening results shown in the figures have small changes over time and are practically meaningful. At the same time, they can adapt to different failure modes, indicating that the proposed method can realize adaptive sensor selection for failure mode identification and RUL prediction under time-varying operating conditions.

[0106] 3) Failure Mode Recognition Loss Function Based on Semi-labeled Samples

[0107] This invention focuses on scenarios where only a subset of samples have failure mode labels. It utilizes a mixture of labeled and unlabeled samples for model training, considering loss functions based on labeled and unlabeled samples respectively. For samples with failure mode labels, this invention employs a cross-entropy loss function for failure mode prediction. Right now:

[0108]

[0109] Where z k This represents the probability that the predicted failure mode belongs to k. This represents the actual failure mode; if the failure mode is k, then... Otherwise, the value is 0, where K represents the number of failure modes. For samples without failure mode labels, this invention employs a loss function based on Shannon entropy for failure mode prediction. Right now:

[0110]

[0111] Ensure that the entropy of the predicted failure mode probability is minimized. For example, if the actual failure mode is k, then the predicted failure mode probability z... k The algorithm will approach 1 to minimize entropy, even if the failure mode is unlabeled during model training. Therefore, this invention achieves semi-supervised learning based on labeled and unlabeled samples. Furthermore, considering that sensor selection weights exhibit significant differences across different failure modes, even if some training samples are unlabeled, the failure mode allocator... It will also perform well. In addition, the failure mode recognition loss function also enhances the difference in sensor selection weights between samples of different failure modes, making the sensor selection weights adaptive to different failure modes.

[0112] 4) Loss function for extracting operating condition-independent features

[0113] This invention uses the cross-entropy loss function for the operating condition prediction results, namely:

[0114]

[0115] in Indicates at t n The actual operating condition at time step -C+c, if the actual operating condition is p then Otherwise, it is 0, o c,p Indicates the prediction at t n The probability that the condition at time step -C+c belongs to p, where p represents the number of conditions.

[0116] 5) The model's total loss function

[0117]

[0118] Where λ v , λ r , λ p and λ d This indicates the adjustment parameter. ρ=1 indicates that the failure mode of the sample is labeled, and ρ=0 indicates that the failure mode of the sample is not labeled.

[0119] (iv) Model Training

[0120] For the purposes of subsequent description, this invention will use the total loss function In addition to λ d (θ w ,θ f ,θ d The sum of the parts is written as Right now By minimizing This enables sensor selection, failure mode identification, and RUL prediction, while minimizing the operating condition loss function λ. d Training model parameters θ d To obtain an accurate working condition classifier And maximize the operating condition loss function λ d To train other parameters θ of the model w ,θ f ,θ y,1 ,…,θ y,K ,θ r This makes the extracted features The operating conditions cannot be accurately identified. The objective expressions for the two adversarial optimizations mentioned above are as follows:

[0121]

[0122]

[0123] These two objectives can be achieved through backpropagation, and the formula for parameter update is as follows:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] During the back propagation process, the value of λ in equation (15) is... d The objective can be achieved through (21). Furthermore, since the gradient is inverted through the gradient inversion layer, the values ​​in equations (17) and (18) are... The presence of a negative sign previously indicates that the error in classifying working conditions will gradually increase, thus realizing the relationship between λ and the condition in equation (16). d The goal is to achieve two opposing objectives through such a gradient update process, thereby enabling effective failure mode identification and RUL prediction based on condition-independent features.

[0130] S4. Acquire the failure data of the aircraft engine to be predicted, input it into the trained deep learning model, adaptively select the sensors of the in-service aircraft engine, and output the failure mode recognition and RUL prediction results under time-varying operating conditions.

[0131] By adopting the technical solution provided by this invention, adaptive and meaningful sensor selection can be achieved, and aircraft engine failure mode identification and remaining life prediction can be realized. This can assist in inspection and maintenance planning and avoid economic and social losses caused by aircraft engine failure.

[0132] To evaluate the model's performance, this example study uses a failure dataset of aircraft turbofan engines. This dataset, generated by NASA's Commercial Modular Aero-Propulsion System Simulation (C-MAPSS), simulates the failure process of turbofan engines. The dataset comprises four sub-datasets; this example study is based on the FD004 sub-dataset. The aircraft engines in this dataset exhibit six operating conditions and two failure modes. Failure modes include fan failure (F) and high-pressure compressor failure (H). The dataset includes data from 249 historical aircraft engines containing complete observations up to failure and 248 in-service aircraft engines containing incomplete observations up to a period before failure. This invention uses historical aircraft engine data, including sensor data, operating condition data, and failure mode labels, to train the proposed model. The trained model is used to identify failure modes and predict the failure tolerance level (RUL) of in-service aircraft engines, and the model is validated using actual RUL data and failure mode labels from in-service aircraft engines.

[0133] The present invention trains its model using historical aircraft engine data, enabling adaptive sensor selection for in-service aircraft engines, as well as failure mode identification and RUL prediction. The model includes an attention-based Long Short-Term Memory (LSTM) network module. This module generates adaptive attention weights by limiting the temporal variation of weights for meaningful sensor selection. The sensor selection module is integrated into a joint learning network for joint failure mode identification and RUL prediction under time-varying operating conditions. This network adversarially extracts operating condition-independent features through end-to-end training, eliminating domain differences caused by different operating conditions, and innovatively utilizes sensor selection weights as more effective features for failure mode identification. Furthermore, considering the case where only some failure mode labels are known, the present invention introduces an entropy-based failure mode probability loss function to improve the accuracy of identifying samples with unknown failure mode labels, thereby achieving semi-supervised learning.

[0134] This embodiment sets two metrics to evaluate the model's performance.

[0135] 1) For the identification of failure modes of in-service aircraft engines, failure mode identification accuracy is used as a metric to measure its efficiency. Accuracy is defined as the percentage of in-service aircraft engines that can correctly identify failure modes.

[0136] 2) For predicting the remaining life of in-service aircraft engines, the remaining life prediction error is used as the evaluation index of the model. The remaining life prediction error of in-service aircraft engines is ∈, and its calculation formula is as follows:

[0137]

[0138] Where y is the predicted RUL, It's the real RUL. This represents the actual failure time. Furthermore, this invention considers that aircraft engines rarely degrade at the start of operation; therefore, it sets training samples and the upper limit of the RUL (Range Limit of In-Service) for in-service aircraft engines. * It is 130.

[0139] Figure 3The failure mode identification (FMD) and residual lifetime (RUL) prediction results of this invention are shown. The horizontal axis of the image represents different RUL levels; "20, 40, 60, 80, 100, 120, 130" represent the FMD and RUL prediction results for all in-service aircraft engines with RUL values ​​less than or equal to 20, 40, 60, 80, 100, 120, and 130. Figure (3a) shows the FMD identification results, and the vertical axis represents the average FMD identification accuracy at each RUL level. Even with high engine RUL levels, this invention can still guarantee 97.6% accuracy in identifying failure modes. Figure (3b) shows the RUL prediction results, where the points and error bars represent the average and standard deviation of the RUL prediction error at that RUL level, respectively. This invention can achieve accurate residual lifetime prediction at different RUL levels, aiding in the development of maintenance and repair strategies and preventing aircraft engine failures.

[0140] Figure 4 The diagram illustrates the failure mode recognition accuracy of this invention under different proportions of labeled failure modes in the training samples, where the dots represent the average failure mode recognition accuracy. The diagram shows that the method of this invention only requires 2% of the samples to have labeled failure modes to complete failure mode recognition, achieving effective semi-supervised learning.

[0141] 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 aircraft engine health based on sensor adaptive selection, characterized in that, Includes the following steps: S1. Acquire aircraft engine failure data, time-varying operating condition data, failure mode data and real RUL data collected by multi-dimensional sensors as raw data; S2. Use the time window method to process the original data to obtain training samples; In S2, let the aircraft engine failure data collected by the multi-dimensional sensor be represented as: ,in Indicates the time step of the observation. The number of observation sensors is represented by the engine failure mode, which is represented by a one-heat vector. ,in The true RUL represents the number of failure modes. The aircraft engine failure matrix after processing using the time window method is expressed as follows: , include The last one Step observation data, among which Indicates the length of the time window. express middle Time-step observation data; Record the working conditions within the corresponding time window ,in Indicates the number of operating conditions. Indicates the original observation The unique heat vector of the operating condition at each time step; S3. The deep learning model is trained based on the training samples. The deep learning model adaptively selects the input aircraft engine failure data based on the sensor selection weights, and uses the failure mode recognition network constructed based on the sensor selection weights to perform failure mode recognition. The condition-independent features of the input data after sensor selection are extracted, and the mapping relationship between the condition-independent features under different failure modes and the RUL prediction value is established. The final RUL prediction value is obtained by combining the failure mode recognition results. The extraction of condition-independent features from the sensor-selected input data specifically involves: utilizing a long short-term memory network. Data after sensor selection Perform feature extraction: Among them, features express The hidden state, Dimensions representing features Indicates the length of the time window. express The parameter set; Meanwhile, a domain adaptation method is employed to minimize the domain bias caused by operating conditions in order to extract operating condition-independent features, and an operating condition classifier is constructed. The working condition classifier is trained using a loss function and its extracted features. As input, output the operating condition probability matrix: in, Let represent the probability matrix of operating conditions, where express The probability vector of the time step. Indicates in Time step condition p The value, Represents a deep neural network. yes The parameter set, This indicates a gradient reversal layer, which keeps the gradient value unchanged during forward propagation of the neural network but changes the sign of the gradient during backward propagation. S4. Acquire the failure data of the aircraft engine to be predicted, input it into the trained deep learning model, adaptively select the sensors of the in-service aircraft engine, and output the failure mode recognition and RUL prediction results under time-varying operating conditions.

2. The aircraft engine health prediction method based on sensor adaptive selection according to claim 1, characterized in that, Sensor selection function Represented as: in This indicates the result after selection. The weight matrix represents the sensor selection. express The corresponding selection weight, It represents the Hadamardi (or Hadama) stack.

3. The aircraft engine health prediction method based on sensor adaptive selection according to claim 2, characterized in that, The sensor selection weights employ an attention mechanism and are dynamically generated based on sensor data and subsequent targets. The weight acquisition function... as follows: in, Representing the hidden states of a Long Short-Term Memory (LSTM) network, deep neural networks Used for alignment and Dimensions express The parameter set, This represents the ReLU activation function, and the output after activation is denoted as... The weights are then applied using a soft maximum function. Standardize to the range (0,1): in, express The first in One element, Representative sensor exist The selection weight corresponding to the time step, w c =[w c,1 ,…,w c,M ].

4. The aircraft engine health prediction method based on sensor adaptive selection according to claim 3, characterized in that, The failure mode recognition network selects the last item of the sensor selection weight matrix. Build: in, Represents the failure mode probability vector. Represents a deep neural network. express The parameter set.

5. The aircraft engine health prediction method based on sensor adaptive selection according to claim 4, characterized in that, The process of establishing a mapping relationship between condition-independent features under different failure modes and RUL predicted values, and combining this with the failure mode identification results to obtain the final RUL predicted value, specifically involves: Through deep neural networks Build each failure mode Lower features and RULy k The mapping relationship, , Indicates the number of failure modes, prediction The formula is as follows: in yes The parameter set; Joint learning network The final RUL is predicted based on the RUL prediction results and failure mode probabilities across multiple failure modes: in, This represents the failure mode probability vector.

6. The aircraft engine health prediction method based on sensor adaptive selection according to claim 5, characterized in that, The total loss function of the deep learning model Represented as: in, , , and This indicates that the parameters are being adjusted. =1 indicates that the failure mode of the sample is tagged. =0 indicates that the failure mode of the sample has not been marked. These are deep learning model parameters; The function representing the prediction loss of RUL: in It's the real RUL. It is the predicted RUL; The sensor selection weight loss function is represented as follows: in Represents Euclidean distance. Indicates the sensor selection weight, Indicates the length of the time window; 、 Let represent the failure mode recognition loss function based on semi-labeled samples, where, for samples with failure mode labels, the cross-entropy loss function is used for failure mode prediction. ,Right now: in Indicates that the predicted failure mode belongs to The probability, This represents the actual failure mode. If the failure mode is but Otherwise, it is 0. This represents the number of failure modes; for samples without failure mode labels, a loss function based on Shannon entropy is used for failure mode prediction. : The loss function representing the extraction of features irrelevant to the operating conditions is expressed using the cross-entropy loss function for the operating condition prediction results. in Indicates in The actual working condition at the time step, if the actual working condition is but Otherwise, it is 0. Indicates the prediction in Time step condition belongs to p The probability, This indicates the number of operating conditions.

7. The aircraft engine health prediction method based on sensor adaptive selection according to claim 6, characterized in that, The deep learning model, during training, uses the total loss function Recorded as By minimizing It enables sensor selection, failure mode identification, and RUL prediction; simultaneously, by minimizing Training model parameters To obtain an accurate working condition classifier and maximize To train other parameters of the model This makes the extracted features The operating conditions cannot be accurately identified.

8. The aircraft engine health prediction method based on sensor adaptive selection according to claim 7, characterized in that, The training objective of the deep learning model is expressed as: 。

Images