Self-attention model agnostic element learning method for identifying civil aircraft faults

Abstract

The invention relates to the technical field of civil aircraft fault diagnosis and health monitoring, in particular to a self-attention model agnostic learning method for identifying civil aircraft faults, which comprises the following steps: firstly, introducing a multi-dimensional perception self-attention feature extraction module into a feature extractor to improve the discrimination and generalization ability of feature expression; secondly, adopting a polynomial to adaptively regulate and control loss in optimization, and dynamically adjusting loss weights among tasks under a model agnostic meta-learning framework; according to the method, robust and rapid fault identification can be realized under various flight working conditions and fault scenes, and the adaptability of the model to data distribution difference and task diversity is remarkably improved. The method provided by the invention is clear in structure and simple and convenient to operate, and an efficient and feasible technical approach can be provided for civil aircraft fault identification under the multi-working-condition operation condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of civil aircraft fault diagnosis and health monitoring, and particularly relates to a self-attention model agnostic meta-learning method for identifying faults of a civil aircraft. BACKGROUND

[0002] An aero-engine is a core power system of a civil aircraft, and its performance and reliability directly affect flight safety and transportation efficiency. The civil aircraft engine is operated for a long time under complex flight environments, and key components such as turbine blades, compressor blades and bearings are prone to wear, cracks, eccentric wear and vibration abnormalities and other faults. Once a fault occurs, it may lead to a decrease in engine efficiency, an increase in fuel consumption, and even a serious threat to flight safety. Therefore, it is of great engineering and application value to develop an efficient, accurate and robust fault identification method for a civil aircraft engine.

[0003] Early engine fault identification methods mainly rely on traditional signal processing techniques, including Fourier transform, wavelet transform, envelope analysis and the like, and judgment is made by extracting time-frequency features such as mean, variance, energy and spectral peak value. This kind of method has the advantages of low computational cost and strong interpretability, but under the conditions of multiple working conditions and multiple tasks of a civil aircraft, the signal noise is large, the nonlinear interference is obvious, the traditional features are difficult to cover all fault modes, and the feature design relies on human experience, making it difficult to realize automation and large-scale adaptation.

[0004] With the improvement of computing power, machine learning methods have been gradually applied to civil aircraft engine fault diagnosis. For example, support vector machines, random forests and k-nearest neighbors, etc. Through feature vector training model to realize classification. This kind of method improves the recognition accuracy to a certain extent, and supports multi-classification task. However, these methods still rely on manual feature extraction, and have limited transferability for new working conditions or new tasks, requiring re-labeling data and training models, which is difficult to meet the needs of rapid identification of civil aviation under different flight conditions.

[0005] In order to overcome the limitations of manual feature extraction, deep learning has been widely applied to civil aircraft engine fault diagnosis in recent years. Convolutional neural networks can learn deep features directly from raw signals or frequency spectrum graphs, achieving automatic feature extraction; recurrent neural networks and long short-term memory networks can capture the time series dynamic information of key components of the engine. The attention mechanism further enhances the discriminability of feature representation, and by giving different weights in the spatial or channel dimension, the model can focus on key frequency bands or time points. This kind of method can achieve higher recognition accuracy under single task or single working condition, and is superior to traditional methods to a certain extent. While deep learning performs well under specific conditions, it still faces challenges in the real-world operating environment of civil aircraft. First, engine data distribution varies significantly across different flight conditions (e.g., cruise, climb, descent), leading to poor model performance under new missions or conditions, necessitating retraining. Second, civil aircraft involve multiple missions and component failure modes, making it difficult for a single deep network to cover all tasks, easily resulting in optimization conflicts between tasks and impacting overall generalization ability. Third, traditional loss functions (e.g., cross-entropy, mean squared error) lack a dynamic balancing mechanism in multi-task training, causing some tasks to dominate training while others fail to learn sufficiently, reducing overall model performance. Finally, convolutional networks exhibit bias in feature extraction under multi-condition and multi-task data, lacking true generality and robustness, affecting the model's adaptability to different conditions.

[0006] To address the aforementioned issues, those skilled in the art have proposed methods such as multi-task learning and model-agnostic meta-learning. These methods can quickly adapt to new tasks using a small number of samples, improving model transferability. However, existing methods typically suffer from problems such as non-universal feature extraction modules, lack of dynamic balance in task-to-task optimization, and insufficient method comprehensiveness, making it difficult to simultaneously achieve general feature extraction, rapid task adaptation, and adaptive optimization. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a self-attention model agnostic learning method for identifying civil aircraft faults. This method is used for civil aircraft engine fault identification and is mainly based on a multidimensional perceptual self-attention feature extraction network and a multinomial adaptive adjustment loss. By constructing a general feature extraction network and an adaptive task optimization strategy, it achieves rapid adaptation and high-precision fault identification under multiple operating conditions and tasks within the model agnostic learning framework. This effectively solves the shortcomings of existing technologies in terms of feature universality, task transferability, and multi-task optimization balance.

[0008] The first objective of this invention is to provide a self-attention model agnostic meta-learning method for identifying faults in civil aircraft, comprising: Signals generated by different components of a civil aircraft under different operating conditions are collected during the operation of the aircraft to obtain training datasets for multiple tasks. The training datasets include source task data and single target task data under different operating conditions. Fourier transforms are performed on the source task data and the target task data respectively to obtain the corresponding spectrum samples; the spectrum samples are then used to construct a diagnostic meta-task through scenario training. A feature extraction network is constructed by introducing a multidimensional perception self-attention feature extraction module. The diagnostic meta-task is input into the feature extraction network, and the multidimensional perception self-attention feature extraction module performs global perception and feature enhancement on the input signal to extract task features. A multinomial adaptive adjustment loss function is constructed, which is an improvement on the traditional cross-entropy loss. The feature extraction network and the multinomial adaptive adjustment loss function are embedded into the model-agnostic meta-learning framework to perform inner-layer task optimization and outer-layer meta-optimization, resulting in a trained meta-model. The trained meta-model is used to identify faults in civil aircraft engines in the target domain.

[0009] In one embodiment, the feature extraction process of the feature extraction network includes: The input features are mapped to query feature vectors, key feature vectors, and value feature vectors through three 1×1 convolutional layers with different numbers of channels, respectively. Take the transpose of the key feature vector and multiply it with the query feature vector to obtain the relevance matrix; The correlation matrix is ​​normalized row by row to obtain the attention feature map; Multiplying the attention feature map by the value feature vector yields the attention-enhanced features; By introducing learnable scaling parameters, the attention-enhanced features are scaled and adjusted to obtain the final features.

[0010] In one embodiment, the feature extraction network is composed of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately in sequence.

[0011] In one embodiment, the polynomial adaptive regulation loss function is:

[0012] In the formula, These are the weighting coefficients of the polynomial terms; It is the order of the polynomial; ,in, It is the one-hot encoding of the real label. It is the probability predicted by the model; It represents the number of categories.

[0013] In one embodiment, inner task optimization includes: Given an input sample and corresponding label for any diagnostic meta-task, a feature representation for that task is generated through a feature extraction network. Based on the feature representation and the multinomial adaptive adjustment loss function, the loss for the task is calculated, and the task-specific prediction function parameters are updated.

[0014] In one embodiment, outer-layer optimization includes: Collect all updated parameters of the diagnostic meta-tasks after inner-layer optimization, construct a meta-loss function, and update the global parameters of the meta-model based on the meta-loss function.

[0015] In one embodiment, when identifying civil aircraft engine faults in the target domain based on the trained meta-model, the model identifies faults in the target domain through a weighted integration of rapid adaptation of inner-layer tasks and meta-optimization of outer-layer elements, and evaluates the generalization performance of the model under multiple operating conditions and multiple tasks.

[0016] A second objective of this invention is to provide a computer program product, comprising a computer program that, when executed by a processor, implements a self-attention model agnostic learning method for identifying faults in civil aircraft.

[0017] A third objective of this invention is to provide an electronic device comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to execute a self-attention model agnostic learning method for identifying civil aircraft faults by executing the executable instructions.

[0018] The fourth objective of this invention is to provide a system for a self-attention model agnostic meta-learning method for identifying faults in civil aircraft, comprising: The data acquisition module is used to collect signals generated by different components of a civil aircraft under different operating conditions during operation, and to obtain training datasets corresponding to multiple tasks. The training datasets include source task data and single target task data under different operating conditions. Fourier transforms are performed on the source task data and target task data respectively to obtain corresponding spectrum samples. The spectrum samples are then used to construct a diagnostic meta-task through a scenario training method. The network construction module is used to construct a feature extraction network that incorporates a multidimensional perception self-attention feature extraction module. The diagnostic meta-task is input into the feature extraction network, and the multidimensional perception self-attention feature extraction module performs global perception and feature enhancement on the input signal to extract task features. A multinomial adaptive adjustment loss function is constructed, which is based on an improvement of the traditional cross-entropy loss. The network training module is used to embed the feature extraction network and the multinomial adaptive adjustment loss function into the model agnostic meta-learning framework to perform inner task optimization and outer meta-optimization to obtain the trained meta-model. The fault identification module is used to identify faults in civil aircraft engines in the target domain based on the trained meta-model.

[0019] The present invention has at least the following beneficial effects: This invention provides a self-attention model agnostic meta-learning method for identifying faults in civil aircraft engines. First, a multi-dimensional perceptual self-attention feature extraction network is constructed to perform global perception and multi-dimensional feature enhancement on data from different operating conditions and tasks, extracting general features with discriminative and generalizable properties. Then, within the model agnostic meta-learning framework, a multinomial adaptive loss is combined to dynamically weight the losses for different tasks, achieving balanced optimization and rapid transfer between tasks. Finally, a self-attention mechanism is used to enhance key features, thereby achieving high-precision and rapidly adaptable engine fault identification in multi-condition and multi-task environments. Attached Figure Description

[0020] Figure 1 The flowchart of the self-attention model agnostic meta-learning method of the present invention for fault identification of key system components of civil aircraft is shown below. Figure 2 Flowchart for constructing the diagnostic meta-task described in step two; Figure 3 Diagram of the network structure for multidimensional perception self-attention feature extraction; Figure 4 A diagram of an agnostic meta-learning algorithm for a model with embedded multinomial adaptive adjustment loss; Figure 5 The diagnostic results of the domain adversarial neural network are as follows: Figure 5 a is the confusion matrix. Figure 5 b is a visualization of T-SNE features; Figure 6 The diagnostic results of the domain adversarial neural network are as follows: Figure 6 a is the confusion matrix. Figure 6 b is a visualization of T-SNE features; Figure 7 The diagnostic results of the domain adversarial neural network are as follows: Figure 7 a is the confusion matrix. Figure 7 b is a visualization of T-SNE features; Figure 8 The diagnostic results of the domain adversarial neural network are as follows: Figure 8 a is the confusion matrix. Figure 8 b is a visualization of T-SNE features; Figure 9 The diagram shows a comparison of the identification results of the proposed method with other advanced methods. Detailed Implementation

[0021] In order to illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description is provided in conjunction with the embodiments.

[0022] This invention aims to more effectively adapt feature distribution when there are significant differences in data distribution between different tasks and working conditions. It effectively extracts general features with discriminative and generalizable properties. The feature extractor of the model introduces a multi-dimensional perceptual self-attention feature extraction network. Through a multi-scale, multi-channel self-attention mechanism, the input signal is globally perceived, feature enhanced, and multi-dimensionally modeled to obtain a stable and robust general representation, providing a high-quality feature foundation for subsequent model agnostic learning.

[0023] In order to balance the optimization process between different tasks and different working conditions and improve the model's rapid transfer capability under the model-agnostic meta-learning framework, this invention introduces a polynomial adaptive adjustment loss. By dynamically adjusting the loss weights of different tasks in a polynomial form, adaptive optimization and balance between tasks are achieved, effectively improving the model's fault identification accuracy and generalization ability under complex working conditions.

[0024] To achieve the above objectives, see Figure 1 As shown, a self-attention model agnostic meta-learning method for identifying faults in civil aircraft includes: S1. Collect signals generated by different components of civil aircraft under different operating conditions during operation to obtain training datasets corresponding to multiple tasks. The training datasets include source task data and single target task data under different operating conditions; provide input basis for subsequent feature extraction and meta-learning.

[0025] In order to extract features more efficiently and achieve rapid migration under different operating conditions and mission conditions, this invention collects signals generated by different components of civil aircraft under different operating conditions during operation.

[0026] S2, see also Figure 2 As shown, Fourier transforms are performed on the source task data and the target task data respectively to obtain the corresponding spectrum samples; the spectrum samples are then used to construct a diagnostic meta-task through scenario training; the diagnostic meta-task is fed into the subsequent feature extraction module as input to capture key frequency components and their dynamic characteristics.

[0027] S3. Construct a feature extraction network that incorporates a multidimensional perception self-attention feature extraction module, input the diagnostic meta-task into the feature extraction network, and use the multidimensional perception self-attention feature extraction module to perform global perception and feature enhancement on the input signal to extract task features. The feature extraction process of a feature extraction network includes: The input features are mapped to query feature vectors, key feature vectors, and value feature vectors through three 1×1 convolutional layers with different numbers of channels, respectively. Take the transpose of the key feature vector and multiply it with the query feature vector to obtain the relevance matrix; The correlation matrix is ​​normalized row by row to obtain the attention feature map; Multiplying the attention feature map by the value feature vector yields the attention-enhanced features; By introducing learnable scaling parameters, the attention-enhanced features are scaled and adjusted to obtain the final features.

[0028] The feature extraction network consists of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately in sequence.

[0029] The feature extraction network, which incorporates a multi-dimensional perceptual self-attention feature extraction module, is constructed. Its structure includes multiple convolutional layers and a multi-channel self-attention module, enabling global perception and feature enhancement of the input signal at multiple scales. The core of the multi-dimensional perceptual self-attention feature extraction module includes the following components: To extract attention, from the previous hidden layer Extract features from the mapping and transform it into a feature space. and time and space .in N For the characteristic location, C This represents the number of channels. The feature extraction process is as follows:

[0030]

[0031]

[0032] in, , and It is a 1×1 convolution with different numbers of channels, where Represents the features of the input. Let the weight matrix be denoted as , where the query feature vector is denoted as . Key feature vector Sum eigenvectors .

[0033] With the first Taking the features at each position as an example, after the product operation... and The output is converted into an N-dimensional vector. The result was transposed and combined with Multiplying the outputs yields the correlation matrix. As shown in the following formula:

[0034] After row-normalizing the matrix, an attention map is obtained, with each row representing a different attention method, as shown below:

[0035] In the formula, Indicates generation Location is a network for location The level of attention.

[0036] Finally, attention feature maps are applied. M By captured attention and The outputs are multiplied together to obtain the following formula:

[0037] Furthermore, introducing the attention mechanism into the model requires the introduction of an additional learnable scaling parameter γ to obtain the final result. :

[0038] A multidimensional awareness self-attention feature extraction module is embedded into the feature extractor to ensure that the model can effectively capture key fault features and enhance the model's focus on key features. The structure of the feature extractor with the multidimensional awareness self-attention feature extraction module is as follows: Figure 3 As shown, the feature extraction network consists of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately. The feature extraction network with multidimensional perceptual self-attention feature extraction modules can effectively handle the diversity of input data, enabling the model to capture significant feature information even under small sample conditions.

[0039] To achieve adaptive optimization across different tasks within a model-agnostic meta-learning framework, a multinomial adaptive adjustment loss is designed. By dynamically adjusting the loss weights for each task, the model's ability to quickly transfer between multiple scenarios and tasks is improved, as detailed below.

[0040] S4. Construct a polynomial adaptive adjustment loss function, which is an improvement on the traditional cross-entropy loss; the polynomial adaptive adjustment loss function is:

[0041] In the formula, These are the weighting coefficients of the polynomial terms; It is the order of the polynomial; ,in, It is the one-hot encoding of the real label. It is the probability predicted by the model; It represents the number of categories.

[0042] For example, a multinomial adaptive adjustment loss is proposed to improve the performance and robustness of the model in fault diagnosis tasks. The multinomial adaptive adjustment loss is a loss function designed for classification problems, especially for tasks where fault features are ambiguous or the differences between categories are small, making classification difficult. It improves upon the traditional cross-entropy loss, as shown below:

[0043] In the formula, It is the number of categories. It is the one-hot encoding of the real label. This represents the probability predicted by the model. The core idea of ​​multinomial adaptive adjustment loss is to combine multiple loss components, each targeting a specific challenge in fault diagnosis. By adding a multinomial term, the model's performance in imbalanced data and small sample scenarios is improved. Specifically, it combines multiple loss components, each targeting a specific challenge in fault diagnosis. The multinomial adaptive adjustment loss is defined as:

[0044] In the formula, These are the weighting coefficients of the polynomial terms, controlling the balance between cross-entropy loss and the polynomial terms. The order of the polynomial determines the degree of influence of the polynomial terms on the loss function. This influences the model's predicted probability for a given sample. When the value is close to 0, it means that the sample is difficult to classify correctly. At this point, the polynomial term in the polynomial adaptive loss is adjusted as follows:

[0045] This will impose a greater penalty on the sample, forcing the model to pay more attention to this difficult-to-classify sample. This helps to better capture subtle fault differences in small-sample fault diagnosis tasks and improves the model's ability to distinguish minority class faults.

[0046] By adding a polynomial term, the polynomial adaptive loss can balance the model's gradient updates for different samples, thereby optimizing the model's performance in small-sample and imbalanced fault diagnosis tasks. This strategy for dealing with hard-to-classify and noisy samples can improve the model's sensitivity to key fault features while reducing overfitting to noisy data.

[0047] S5. Embed the feature extraction network and the multinomial adaptive adjustment loss function into the model agnostic meta-learning framework, perform inner task optimization and outer meta-optimization, and obtain the trained meta-model. Inner task optimization, including: Given an input sample and corresponding label for any diagnostic meta-task, a feature representation for that task is generated through a feature extraction network. Based on the feature representation and the multinomial adaptive adjustment loss function, the loss for the task is calculated, and the task-specific prediction function parameters are updated.

[0048] Outer layer optimization includes: Collect all updated parameters of the diagnostic meta-tasks after inner-layer optimization, construct a meta-loss function, and update the global parameters of the meta-model based on the meta-loss function.

[0049] For example, see Figure 4 As shown, a multidimensional perceptual self-attention feature extraction network and a multinomial adaptive control loss are embedded into the model for agnostic learning.

[0050] In the model-agnostic meta-learning framework, each meta-task first extracts task features through a multidimensional perceptual self-attention feature extraction network. Assume the task... The input sample is The corresponding tag is The multidimensional perception self-attention feature extraction network generates feature representations through a multi-scale, multi-channel self-attention mechanism:

[0051] In the formula, Represents network parameters, For the task High-quality, general-purpose feature representations. The multidimensional perceptual self-attention feature extraction network can extract highly discriminative and generalizable features from different tasks and working conditions, providing a stable feature foundation for subsequent inner layer optimization.

[0052] In the inner optimization of each task, the features obtained in the previous step are used. Calculate the task loss and perform weighted optimization by combining it with multinomial adaptive adjustment loss:

[0053] In the formula, These are the weighting coefficients of the polynomial terms. This represents the prediction function specific to the inner task. By adaptively adjusting the loss using a polynomial, the inner optimization not only optimizes the loss for each task but also automatically balances the contributions of different tasks, providing a more balanced gradient signal for the outer meta-optimization.

[0054] After the inner task optimization is completed, the updated parameters of all tasks are used for outer meta-optimization to form a meta-loss function with MAPA features and PARL weighting:

[0055] In the formula, The inner learning rate; These are the parameters of the outer meta-model; The loss function is defined as follows: This meta-loss function dynamically adjusts the weights of each task and combines them with extracted high-quality general features to enable the meta-model to adapt quickly to different tasks, operating conditions, and data distributions.

[0056] S6. Identify engine faults in the target domain of civil aircraft based on the trained meta-model.

[0057] When identifying faults in civil aircraft engines in the target domain based on the trained meta-model, the model identifies faults in the target domain through weighted integration of rapid adaptation of inner-layer tasks and optimization of outer-layer meta-models, and evaluates the generalization performance of the model under multiple operating conditions and multiple tasks.

[0058] Reference Figure 1 As shown, the present invention can be mainly divided into two parts: the first part is a feature extraction network that introduces a multidimensional perception self-attention module; the second part is a multinomial adaptive control loss that can achieve adaptive optimization between different tasks.

[0059] Reference Figure 2 As shown, the original data of the source and target tasks are reconstructed in two dimensions and the corresponding spectral samples are obtained through Fourier transform. Then, the spectral samples are used to construct a diagnostic meta-task through scenario training.

[0060] Reference Figure 3 As shown, the multidimensional perceptual feature extraction network consists of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately. It can effectively handle the diversity of input data, enabling the model to capture significant feature information even under small sample conditions. Reference Figure 4 As shown, the polynomial adaptive control loss designed in this invention can dynamically adjust the loss weights of each task to improve the model's rapid transfer capability under multiple conditions and tasks. To further illustrate the self-attention model agnostic learning method provided by this invention, it is described in conjunction with the accompanying drawings and specific examples.

[0061] The embodiment is implemented according to the following steps: Step 1: Use sensors to collect raw signals from key systems and components of the civil aircraft. Based on the different operating conditions of the key systems of the civil aircraft, acquire raw source domain data and target domain data.

[0062] Step 2: Perform Fourier transform on the collected source task data and target task data to obtain the corresponding spectrum samples. Construct a diagnostic meta-task from these spectrum samples using a scenario-based training method. The diagnostic meta-task will serve as the input to the proposed method.

[0063] Step 3: Construct a feature extraction network incorporating a multidimensional perceptive self-attention feature extraction module. Its structure includes multiple convolutional layers and a multidimensional perceptive self-attention module, enabling global perception and feature enhancement of the input signal. The structure of the feature extractor with the multidimensional perceptive self-attention feature extraction module is as follows: Figure 3 As shown, the feature extraction network consists of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately. The feature extraction network with multidimensional perceptual self-attention feature extraction modules can effectively handle the diversity of input data, enabling the model to capture significant feature information even under small sample conditions.

[0064] Step 4: Design a polynomial adaptive adjustment loss, which is achieved by introducing polynomial terms: This involves imposing a larger penalty on difficult samples, forcing the model to pay more attention to these hard-to-classify samples. The gradient updates of the model are balanced for different diagnostic tasks.

[0065] Step 5: Embed the multidimensional perceptual self-attention feature extraction network and the multinomial adaptive control loss into the model's agnostic learning.

[0066] Step 6: Apply the trained model to the target domain fault identification task, evaluate the model's generalization performance under multiple operating conditions, and achieve high-precision fault identification in the target domain.

[0067] Example 1 Source domain data: In this embodiment, aircraft engine data collected by an airline through a full-segment fast access recorder (QAR) was used as source domain data to verify the effectiveness of the method. As shown in Table 1 below, the QAR recorded 11 types of monitoring data, including the low-pressure shaft speed of the first and second engines, the high-pressure shaft speed, vibration value, fuel flow rate, and exhaust temperature. The data sampling frequency was 1Hz.

[0068] Table 1. Description of Monitoring Parameters

[0069] Both flight mission segment datasets contain four state samples, and each state sample includes 11 monitoring parameters. Considering that the flight monitoring parameters have different dimensions and significant differences in numerical values, the Gaussian Guided Adversarial Domain Adaptive Network (GGADAN) uses the Min-max normalization method to normalize each monitoring parameter. The normalized value of each monitoring parameter is between 0 and 1, as shown in the following equation:

[0070] In the formula, x and y represent the original data points and the normalized data points of the monitoring parameters, respectively. x max and xmin These represent the maximum and minimum values ​​of each monitoring parameter, respectively.

[0071] Target Domain Data: This embodiment uses data collected by an airline across the entire flight segment to verify the effectiveness of the proposed method. The data used includes normal and four abnormal states. The four abnormal states are: right engine N1 vibration value exceeding the limit during cruise (4.1), right engine N1 vibration value consistently high during takeoff roll (1.7 / 2.1), right engine N1 vibration value showing an upward trend during the flight, and high right engine N1 vibration value occurring during cruise. It should be noted that these abnormal data were collected from different flight missions, rather than multiple abnormal types occurring within the same mission.

[0072] As shown in Table 2 below, 11 types of monitoring data were recorded, including the low-pressure shaft speed of the left and right engines, the high-pressure shaft speed, vibration value, and fuel flow rate. The sampling frequency of the data was 1Hz.

[0073] Table 2 11 monitoring parameters

[0074] Five state samples were used, each containing 11 monitoring parameters, meaning each sample consisted of 11 data points, arranged in the order shown in Table 2. Considering the different dimensions and significant differences in the numerical values ​​of the flight monitoring parameters, the proposed method employed Min-max normalization to normalize each monitoring parameter. The normalized values ​​of each parameter were then between 0 and 1, as shown in the following formula:

[0075] In the formula, x and y These are the original data points of the monitoring parameters and the data points after normalization, respectively. x max and x min These represent the maximum and minimum values ​​of each monitoring parameter, respectively.

[0076] Migration Task Construction: In this embodiment, QAR data collected by an airline across the entire flight segment was used as the source domain data, and engine anomaly data collected by the same airline across the entire flight segment was used as the target domain data to verify the effectiveness of fault diagnosis. The target domain data included five fault states: normal state, climb vibration indication fault, climb second engine bird strike induced fault, descent vibration indication fault, and descent second engine vibration value fault. The source domain data also included normal data and four fault states. The four fault states were: right engine N1 vibration value exceeding the limit for two seconds during cruise, right engine N1 vibration value being high during takeoff roll, right engine N1 vibration value showing an upward trend during the flight, and right engine N1 vibration exceeding the limit for three seconds during cruise. It should be noted that these fault data were collected from different flight missions, rather than multiple fault types occurring in the same mission. The final constructed migration task is shown in Table 3 below.

[0077] Table 3 Construction of the migration dataset

[0078] Fault diagnosis results and analysis (1) Domain adversarial neural network diagnostic results (accuracy 89.00%): For the anomaly detection task constructed in Table 3, a domain adversarial neural network diagnostic model was designed to perform the diagnostic task, and the fault diagnosis accuracy was 89.00%. Figure 5 A confusion matrix of the diagnostic results was plotted. Figure 5 As can be seen, the domain adversarial neural network diagnostic model has good detection performance for the three abnormal states: normal state, right-hand N1 exceeding the limit, and two types of high right-hand N1. However, it exhibits some classification errors for the fourth type, which shows an increasing trend in right-hand N1. Figure 5 b. By using feature embedding graphs to visualize the extracted high-dimensional fault features, it can be observed that the clustering effect of the four abnormal states—normal state, right-side N1 exceeding the limit, and two types of right-side N1 being slightly high—is obvious. However, the clustering of some sample features of faults with a rising right-side N1 trend is confused with that of faults with slightly high right-side N1.

[0079] (2) Diagnostic results of deep correlation alignment domain adaptation network (accuracy: 80.00%): For the anomaly detection task constructed in Table 3, a deep correlation alignment domain adaptive diagnostic model was designed to perform the diagnostic task, and the fault diagnosis accuracy was 80.00%. Figure 6 A confusion matrix of the diagnostic results was plotted. Figure 6As can be seen, the deep correlation alignment domain adaptive diagnostic model has good detection performance for normal state, right-hand N1 elevation, right-hand N1 trend increase, and right-hand N1 elevation abnormal state. However, it completely misclassifies the second type of right-hand N1 exceeding the limit abnormality. Figure 6 b. By using feature embedding graphs to visualize the extracted high-dimensional fault features, it can be observed that the clustering effect of the four abnormal state features, namely normal state, right-side N1 exceeding the limit, and two types of right-side N1 being slightly high, is obvious. However, the sample features of right-side N1 exceeding the limit fault and the sample features of right-side N1 being slightly high are seriously confused.

[0080] (3) Model-agnostic meta-learning network (accuracy: 94.00%) For the anomaly detection task constructed in Table 3, a model-agnostic meta-network diagnostic model was designed to perform the diagnostic task, with a fault diagnosis accuracy of 94.00%. This network uses the same "N-way K-shot" training paradigm as the prototype network to construct the task, and the number of samples for each anomaly state in both the source and target domains is 20. Figure 7 A confusion matrix of the diagnostic results was plotted. Figure 7 As can be seen, the model-agnostic meta-network diagnostic model demonstrates excellent detection performance for the three abnormal states: right-hand N1 exceeding the limit, right-hand N1 being high, and right-hand N1 showing an upward trend. However, a small number of classification errors still exist for the normal state and the fifth type, right-hand N1 being high. See also... Figure 7 As shown in b, the extracted high-dimensional fault features are reduced to visualization using the T-distribution random neighborhood embedding technique. It can be observed that the clustering effect of the three abnormal state features of right-side N1 exceeding the limit, right-side N1 being high, and right-side N1 trending upward is obvious. However, there is some confusion between the clustering of some sample features of the normal state and the sample features of the right-side N1 high fault.

[0081] (4) Diagnostic results of the self-attention model agnostic meta-learning method (accuracy: 100.00%) Based on the above analysis, it can be found that although the first two fault diagnosis algorithms based on transfer learning only have classification errors in a single anomaly category, the error rate for this type of fault is relatively high, even resulting in complete misclassification. This is because DANN and Deepcoral transfer learning algorithms improve transfer ability by reducing the distributional differences in the domain, rather than targeting a specific category. In some categories, if the difference between the source and target domains is too large, feature alignment may cause the model to fail to distinguish features in a certain category well, leading to serious classification errors. Meta-learning algorithms have classification errors in two types of anomalies, but the error rate is relatively low. Overall, the model-agnostic meta-network diagnostic model achieves a diagnostic accuracy of 94.00%, obtaining the best diagnostic results. Furthermore, we introduce the SE attention mechanism and smooth cross-entropy loss function to improve the model-agnostic meta-network from two aspects: increasing the attention to key features and improving the model's generalization ability. This enhances the model's efficiency in extracting general diagnostic meta-knowledge from anomaly samples in a small-sample, cross-domain context.

[0082] For the anomaly detection task constructed in Table 3, an improved model-agnostic diagnostic model was designed to perform the diagnostic task, achieving a fault diagnosis accuracy of 100.00%. This network also uses the "N-way K-shot" training paradigm to construct the task, with 20 samples for each anomaly state in both the source and target domains. Figure 8 A confusion matrix of the diagnostic results was plotted. Figure 8 As can be seen, the model-agnostic meta-network diagnostic model has excellent detection performance for both normal states and the four abnormal states. See also... Figure 8 As shown in b, the extracted high-dimensional fault features are reduced to visualization using the T-distribution random neighborhood embedding technique. Observation reveals that the clustering effect of the five types of state features is obvious, with the intra-class spacing decreasing and the inter-class spacing increasing.

[0083] See Figure 9 The figure shows a comparison of the recognition results of the proposed method SAMAML-PL with DANN, DCORAL, and MAML. The proposed method has higher performance than the comparison methods. Figure 9 The horizontal axis represents different target recognition tasks; the vertical axis represents the recognition accuracy, in percentages (%).

[0084] In summary, this invention discloses a self-attention model agnostic learning method for identifying faults in civil aircraft. This method first introduces a multi-dimensional perceptual self-attention feature extraction module into the feature extractor to enhance the discriminative and generalization capabilities of feature representation, thereby achieving global modeling and focused reinforcement of features under different operating conditions and fault modes. Then, in the optimization process, a multinomial adaptive loss adjustment is employed to dynamically adjust the loss weights between tasks within the model agnostic learning framework, achieving balanced optimization and rapid transfer across different tasks. This method can achieve robust and rapid fault identification under various flight conditions and fault scenarios, significantly improving the model's adaptability to differences in data distribution and task diversity. The proposed method has a clear structure and is easy to operate, providing an efficient and feasible technical approach for fault identification of civil aircraft under multi-condition operation.

[0085] This invention provides a computer program product, including a computer program that, when executed by a processor, implements a self-attention model agnostic learning method for identifying faults in civil aircraft.

[0086] This invention provides an electronic device, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to execute a self-attention model agnostic learning method for identifying civil aircraft faults by executing the executable instructions.

[0087] This invention provides a system for a self-attention model agnostic meta-learning method for identifying faults in civil aircraft, comprising: The data acquisition module is used to collect signals generated by different components of a civil aircraft under different operating conditions during operation, and to obtain training datasets corresponding to multiple tasks. The training datasets include source task data and single target task data under different operating conditions. Fourier transforms are performed on the source task data and target task data respectively to obtain corresponding spectrum samples. The spectrum samples are then used to construct a diagnostic meta-task through a scenario training method. The network construction module is used to construct a feature extraction network that incorporates a multidimensional perception self-attention feature extraction module. The diagnostic meta-task is input into the feature extraction network, and the multidimensional perception self-attention feature extraction module performs global perception and feature enhancement on the input signal to extract task features. A multinomial adaptive adjustment loss function is constructed, which is based on an improvement of the traditional cross-entropy loss. The network training module is used to embed the feature extraction network and the multinomial adaptive adjustment loss function into the model agnostic meta-learning framework to perform inner task optimization and outer meta-optimization to obtain the trained meta-model. The fault identification module is used to identify faults in civil aircraft engines in the target domain based on the trained meta-model.

[0088] 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 within the protection scope of the present invention.

Claims

1. A self-attention model agnostic meta-learning method for identifying faults in civil aircraft, characterized in that, include: Signals generated by different components of a civil aircraft under different operating conditions are collected during the operation of the aircraft to obtain training datasets for multiple tasks. The training datasets include source task data and single target task data under different operating conditions. Fourier transforms are performed on the source task data and the target task data respectively to obtain the corresponding spectrum samples; the spectrum samples are then used to construct a diagnostic meta-task through scenario training. A feature extraction network is constructed by introducing a multidimensional perception self-attention feature extraction module. The diagnostic meta-task is input into the feature extraction network, and the multidimensional perception self-attention feature extraction module performs global perception and feature enhancement on the input signal to extract task features. A multinomial adaptive adjustment loss function is constructed, which is an improvement on the traditional cross-entropy loss. The feature extraction network and the multinomial adaptive adjustment loss function are embedded into the model-agnostic meta-learning framework to perform inner-layer task optimization and outer-layer meta-optimization, resulting in a trained meta-model. The trained meta-model is used to identify faults in civil aircraft engines in the target domain.

2. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 1, characterized in that, The feature extraction process of a feature extraction network includes: The input features are mapped to query feature vectors, key feature vectors, and value feature vectors through three 1×1 convolutional layers with different numbers of channels, respectively. Take the transpose of the key feature vector and multiply it with the query feature vector to obtain the relevance matrix; The correlation matrix is ​​normalized row by row to obtain the attention feature map; Multiplying the attention feature map by the value feature vector yields the attention-enhanced features; By introducing learnable scaling parameters, the attention-enhanced features are scaled and adjusted to obtain the final features.

3. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 1, characterized in that, The feature extraction network consists of four convolutional modules and four multidimensional perceptual self-attention feature extraction modules connected alternately in sequence.

4. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 1, characterized in that, The polynomial adaptive regulation loss function is: In the formula, These are the weighting coefficients of the polynomial terms; It is the order of the polynomial; ,in, It is the one-hot encoding of the real label. It is the probability predicted by the model; It represents the number of categories.

5. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 1, characterized in that, Inner task optimization, including: Given an input sample and corresponding label for any diagnostic meta-task, a feature representation for that task is generated through a feature extraction network. Based on the feature representation and the multinomial adaptive adjustment loss function, the loss for the task is calculated, and the task-specific prediction function parameters are updated.

6. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 5, characterized in that, Outer layer optimization includes: Collect all updated parameters of the diagnostic meta-tasks after inner-layer optimization, construct a meta-loss function, and update the global parameters of the meta-model based on the meta-loss function.

7. The self-attention model agnostic meta-learning method for identifying civil aircraft faults according to claim 1, characterized in that, When identifying faults in civil aircraft engines in the target domain based on the trained meta-model, the model identifies faults in the target domain through weighted integration of rapid adaptation of inner-layer tasks and optimization of outer-layer meta-models, and evaluates the generalization performance of the model under multiple operating conditions and multiple tasks.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the self-attention model agnostic learning method for identifying civil aircraft faults as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the self-attention model agnostic learning method for identifying civil aircraft faults according to any one of claims 1 to 7 by executing the executable instructions.

10. A system for a self-attention model agnostic meta-learning method for identifying civil aircraft faults as described in claim 1, characterized in that, include: The data acquisition module is used to collect signals generated by different components of civil aircraft under different operating conditions during operation, and to obtain training datasets corresponding to multiple tasks. The training datasets include source task data and single target task data under different operating conditions. Fourier transforms are performed on the source task data and the target task data respectively to obtain the corresponding spectrum samples; the spectrum samples are then used to construct a diagnostic meta-task through scenario training. The network construction module is used to construct a feature extraction network that incorporates a multidimensional perception self-attention feature extraction module. The diagnostic meta-task is input into the feature extraction network, and the multidimensional perception self-attention feature extraction module performs global perception and feature enhancement on the input signal to extract task features. A multinomial adaptive adjustment loss function is constructed, which is an improvement on the traditional cross-entropy loss. The network training module is used to embed the feature extraction network and the multinomial adaptive adjustment loss function into the model agnostic meta-learning framework to perform inner task optimization and outer meta-optimization to obtain the trained meta-model. The fault identification module is used to identify faults in civil aircraft engines in the target domain based on the trained meta-model.

Images