Intelligent aero-engine fault diagnosis method and system based on meta-learning

Abstract

The invention provides an aero-engine intelligent fault diagnosis method and system based on meta-learning. The aero-engine intelligent fault diagnosis method comprises the following steps that single-working-condition fault data about an aero-engine is acquired, and high-dimensional features of the single-working-condition fault data are obtained; obtaining multi-working-condition fault data about the aero-engine and obtaining high-dimensional features of the multi-working-condition fault data; training a meta-learning model by using the high-dimensional features of the single-working-condition fault data and the multi-working-condition fault data, and obtaining a trained meta-learning model; acquiring real-time operation data of the aero-engine acquired by a sensor, inputting the real-time operation data into the trained meta-learning model, and obtaining a fault diagnosis result of the aero-engine through the trained meta-learning model; the method has the beneficial effects that the diagnosis accuracy can be improved, the accuracy of a fault diagnosis model is enhanced by introducing a meta-learning technology and utilizing common characteristics of a plurality of related tasks, and particularly, efficient and accurate fault diagnosis is realized under the condition of a small amount of annotated data.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis of aircraft engines, and in particular to an intelligent fault diagnosis method and system for aircraft engines based on meta-learning. Background Art

[0002] Aircraft engines are core components of aircraft, and their reliability and safety are crucial to the operation of the entire aircraft. With the continuous development of the aviation industry, the requirements for aircraft engine fault diagnosis technology are becoming increasingly higher. The purpose of fault diagnosis is to timely discover and predict possible faults during engine operation, so as to take corresponding maintenance measures to ensure aviation safety and extend the service life of the engine.

[0003] Existing methods for aircraft engine fault diagnosis can be divided into two main categories: (1) Rule-based expert systems: Principle: These systems rely on the knowledge and experience of domain experts to determine the state of the engine through preset rules and thresholds. These rules are usually formulated based on historical fault data, expert experience, and physical models. Advantages: Under clear fault modes, expert systems can provide fast and direct diagnostic results; Disadvantages: The coverage of rule systems is limited and cannot fully cover all possible fault modes, especially when faced with complex and changing working conditions, the diagnostic accuracy is low; In addition, the development and maintenance of expert systems requires the participation of a large number of experts, which is costly, and the rules often need to be readjusted when new faults occur, which is insufficient in flexibility. (2) Data-driven methods: Principle: With the development of sensor technology and big data analysis, data-driven methods are increasingly being used in the field of fault diagnosis. Common technologies include signal processing, statistical analysis, and machine learning. These methods perform fault diagnosis by analyzing large amounts of data collected from sensors. Advantages: Data-driven methods can automatically extract features from data and are suitable for complex nonlinear problems. In particular, the application of deep learning models makes data-driven methods show great potential in fault diagnosis. Disadvantages: Traditional machine learning methods are highly dependent on large amounts of labeled data, and the training process is time-consuming and requires a lot of computing resources. In addition, although deep learning models can automatically extract features, in the case of small samples and new faults, the model's generalization ability and adaptability are poor, and it is prone to overfitting problems.

[0004] Closest prior art: The closest prior art to the present invention is an aircraft engine fault diagnosis method based on deep learning; deep learning technology has been widely used in fault diagnosis in recent years, and its main features are as follows: (1) Automatic feature extraction: Deep learning models can automatically extract complex high-dimensional features from sensor data, which shows obvious advantages when facing nonlinear and complex problems that are difficult to handle with traditional methods. (2) Large data demand: Deep learning models usually require a large amount of labeled data for training to ensure the generalization ability and accuracy of the model. However, in the actual application of aircraft engines, labeled data is often scarce, especially when new fault types appear or operating conditions change, the problem of insufficient data is more prominent. (3) Poor adaptability: When faced with new fault types or changes in data distribution, the performance of deep learning models may drop significantly, and the generalization ability is insufficient. At the same time, the training time of deep learning models is long and cannot meet real-time requirements.

[0005] In summary, the shortcomings of existing technologies are as follows: (1) Low diagnostic accuracy: Traditional rule-based expert systems and data-driven methods often have difficulty in providing sufficiently high diagnostic accuracy when faced with complex and changeable operating conditions and diverse fault modes; these systems rely on preset rules and thresholds and cannot fully cover all possible fault types, resulting in inaccurate diagnostic results. (2) Insufficient adaptability and generalization capabilities: Traditional fault diagnosis methods lack adaptability and generalization capabilities when faced with new fault types or changes in operating conditions; in particular, although deep learning models can automatically extract features, they have poor adaptability and are prone to poor performance when data distribution changes, resulting in a decrease in the model's diagnostic capabilities. (3) Strong data dependence: Deep learning and other data-driven methods have a high dependence on a large amount of labeled data; in practical applications, labeled data is often scarce, especially in systems with high reliability requirements such as aircraft engines, where it is very difficult to obtain a large amount of fault labeled data, which in turn limits the training and application of the model. (4) High development and maintenance costs: Rule-based expert systems require a large number of experts to participate in formulating rules and maintaining systems, resulting in high development and maintenance costs; deep learning models require powerful computing resources and a large amount of data for training, further increasing costs. (5) Poor real-time performance: Existing deep learning models take a long time to train and cannot be effectively applied in scenarios with high real-time requirements; in actual fault diagnosis, delays may lead to more serious consequences, making it difficult to meet the needs of rapid response. Summary of the invention

[0006] In view of the shortcomings of the prior art mentioned above, the purpose of the present invention is to provide an intelligent fault diagnosis method and system for aircraft engines based on meta-learning, so as to solve the problems of low diagnostic accuracy, insufficient adaptability and generalization ability, and strong data dependence in the prior art.

[0007] To achieve the above objectives and other related objectives, the present invention provides the following technical solutions:

[0008] An intelligent fault diagnosis method for an aircraft engine based on meta-learning, comprising the following steps: obtaining single-operating-condition fault data of an aircraft engine, and obtaining high-dimensional features of the single-operating-condition fault data through a multimodal fusion technique; training a meta-learning model using the high-dimensional features of the single-operating-condition fault data, and obtaining a preliminarily trained meta-learning model according to the training result; obtaining multi-operating-condition fault data of an aircraft engine, and obtaining high-dimensional features of the multi-operating-condition fault data through a multimodal fusion technique and multi-operating-condition data feature extraction;

[0009] The meta-learning model that has been preliminarily trained is trained using the high-dimensional features of the multi-condition fault data, and a trained meta-learning model is obtained according to the training results; real-time operation data of the aircraft engine collected by the sensor is acquired, and the real-time operation data of the aircraft engine is input into the trained meta-learning model after being extracted through multimodal fusion technology and multi-condition data features, and the fault diagnosis result of the aircraft engine is obtained through the trained meta-learning model.

[0010] An intelligent fault diagnosis system for aircraft engines based on meta-learning, comprising: a single-operating-condition training module, used to obtain single-operating-condition fault data about aircraft engines, and obtain high-dimensional features of the single-operating-condition fault data through multimodal fusion technology; use the high-dimensional features of the single-operating-condition fault data to train a meta-learning model, and obtain a preliminarily trained meta-learning model according to the training results; a multi-operating-condition training module, used to obtain multi-operating-condition fault data about aircraft engines, and obtain high-dimensional features of the multi-operating-condition fault data through multimodal fusion technology and multi-operating-condition data feature extraction; use the high-dimensional features of the multi-operating-condition fault data to train the preliminarily trained meta-learning model, and obtain a trained meta-learning model according to the training results;

[0011] The fault diagnosis module is used to obtain real-time operating data of the aircraft engine collected by sensors. The real-time operating data of the aircraft engine is input into a trained meta-learning model after multi-modal fusion technology and multi-operating condition data feature extraction, and the fault diagnosis result of the aircraft engine is obtained through the trained meta-learning model.

[0012] An electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned meta-learning-based intelligent fault diagnosis method for aircraft engines.

[0013] In one embodiment of the present invention, the high-dimensional features of the single operating condition fault data are obtained by multimodal fusion technology, including: time synchronization and alignment of data from different data sources in the single operating condition fault data; preprocessing the single operating condition fault data after time synchronization and alignment; extracting meaningful high-dimensional features from the preprocessed single operating condition fault data, and fusing the extracted high-dimensional features of each modal data, and obtaining the high-dimensional features of the single operating condition fault data according to the fusion result.

[0014] In one embodiment of the present invention, the high-dimensional features of the single operating condition fault data are used to train a meta-learning model, and a preliminary trained meta-learning model is obtained according to the training results, including: randomly selecting a task batch from multiple related task sets in the high-dimensional features of the single operating condition fault data to form multiple subtasks, wherein each task can be regarded as an independent aircraft engine fault training task; determining the initialization model parameters, and using the high-dimensional features in each task to update the initialization model parameters.

[0015] In one embodiment of the present invention, the method of determining the initialization model parameters and updating the initialization model parameters using the high-dimensional features in each task includes: in an inner loop, using the training data in the task and utilizing the gradient descent method to update the initialization model parameters one or more times, and obtaining task-specific parameters based on the one or more update results; then using the test data in the task to determine the loss of the updated model parameters; in an outer loop, for the test results of each task, calculating the total loss and updating the initialization model parameters.

[0016] In one embodiment of the present invention, the high-dimensional features of the multi-condition fault data are obtained by multimodal fusion technology and multi-condition data feature extraction, including: time synchronization and alignment of data from different data sources in the multi-condition fault data; preprocessing the multi-condition fault data after time synchronization and alignment; extracting meaningful high-dimensional features from the preprocessed multi-condition fault data, and determining the meaningful high-dimensional features in the fault data under each condition and then fusing them, and obtaining the high-dimensional features of the multi-condition fault data according to the fusion results.

[0017] In one embodiment of the present invention, the fault diagnosis result of the aircraft engine obtained by the trained meta-learning model also includes a dynamic baseline model, an adversarial transfer training module, a Markov decision module and a fault data online learning module; the dynamic baseline model, the adversarial transfer training module, the Markov decision module and the fault data online learning module are used to determine the fault type corresponding to the new fault data when new fault data is encountered during fault diagnosis through the trained meta-learning model.

[0018] In one embodiment of the present invention, after determining the fault type corresponding to the new fault data, it also includes: using the new fault data to train the trained meta-learning model; the training of the trained meta-learning model using the new fault data includes: inputting the high-dimensional features of the new fault data into the trained meta-learning model, the trained meta-learning model performs a forward propagation according to the initialization parameters, and calculates the loss value after the forward propagation; the trained meta-learning model updates the current initialization parameters through the gradient information of a small amount of new fault data.

[0019] In one embodiment of the present invention, after obtaining the fault diagnosis result of the aircraft engine through the trained meta-learning model, it also includes: evaluating the fault diagnosis result of the aircraft engine.

[0020] As described above, an intelligent fault diagnosis method and system for aircraft engines based on meta-learning of the present invention have the following beneficial effects: the present invention can improve the accuracy of diagnosis: by introducing meta-learning technology, the common features of multiple related tasks are utilized to enhance the accuracy of the fault diagnosis model, especially in the case of a small amount of labeled data, to achieve efficient and accurate fault diagnosis; the present invention can enhance adaptability and generalization ability: the meta-learning model can quickly adapt to new fault types and operating condition changes by learning the common features between multiple tasks, significantly improving the adaptability and generalization ability of the model; the present invention can reduce data dependence: the meta-learning method of the present invention can achieve effective fault diagnosis in the case of small samples, reducing the dependence on a large amount of labeled data, so that accurate fault diagnosis can be performed even in the case of data scarcity; the present invention can reduce development and maintenance costs: the present invention reduces the dependence on expert knowledge and manual rule setting through automatic feature extraction and meta-learning methods, thereby reducing the development and maintenance costs of the system; the present invention can improve diagnostic efficiency and real-time performance: the method of the present invention achieves real-time fault diagnosis through efficient feature extraction and rapid adaptability of meta-learning, can timely discover and handle faults, reduce engine downtime, and improve operating efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Shown is a flow chart of an intelligent fault diagnosis method for an aircraft engine based on meta-learning in a first embodiment of the present invention;

[0022] Figure 2 Shown is a flow chart of an intelligent fault diagnosis method for an aircraft engine based on meta-learning in a second embodiment of the present invention;

[0023] Figure 3 Shown is a flow chart of an intelligent fault diagnosis method for an aircraft engine based on meta-learning in a third embodiment of the present invention;

[0024] Figure 4 Shown is a schematic diagram of an intelligent fault diagnosis system for an aircraft engine based on meta-learning in a fourth embodiment of the present invention;

[0025] Figure 5 It is a schematic diagram of an electronic device in a fifth embodiment of the present invention;

[0026] Figure 6 Shown is a schematic diagram of the aircraft engine fault diagnosis technology route under complex multi-fault coupling and small sample data in the present invention;

[0027] Figure 7 Shown is a schematic diagram of the overall structure of the aero-engine fault diagnosis model based on meta-learning in the present invention. DETAILED DESCRIPTION

[0028] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0029] The first embodiment of the present invention relates to an intelligent fault diagnosis method for aircraft engines based on meta-learning, the process is as follows: Figure 1 As shown, the details are as follows:

[0030] Step 101, obtain single operating condition fault data about aircraft engines, and obtain high-dimensional features of the single operating condition fault data through multimodal fusion technology; use the high-dimensional features of the single operating condition fault data to train a meta-learning model, and obtain a preliminarily trained meta-learning model based on the training results.

[0031] Specifically, the specific steps of obtaining single-condition fault data about aircraft engines and obtaining high-dimensional features of the single-condition fault data through multimodal fusion technology are as follows: first, time-synchronize and align the data from different data sources in the single-condition fault data; then, pre-process the single-condition fault data after time synchronization and alignment; then, extract meaningful high-dimensional features from the pre-processed single-condition fault data, and fuse the extracted high-dimensional features of each modal data, and obtain the high-dimensional features of the single-condition fault data according to the fusion results;

[0032] The meta-learning model is trained using the high-dimensional features of single-condition fault data. The specific steps for obtaining a preliminary trained meta-learning model based on the training results are as follows: first, a task batch is randomly selected from multiple related task sets in the high-dimensional features of the single-condition fault data to form multiple subtasks; then, the initialization model parameters are determined, and the high-dimensional features in each task are used to update the initialization model parameters.

[0033] Step 102, obtaining multi-condition fault data about aircraft engines, and obtaining high-dimensional features of the multi-condition fault data through multimodal fusion technology and multi-condition data feature extraction; using the high-dimensional features of the multi-condition fault data to train the initially trained meta-learning model, and obtaining a trained meta-learning model based on the training results.

[0034] Specifically, the specific steps for obtaining multi-condition fault data about aircraft engines and obtaining high-dimensional features of the multi-condition fault data through multimodal fusion technology and multi-condition data feature extraction are: first, time-synchronize and align the data from different data sources in the multi-condition fault data; then pre-process the multi-condition fault data after time synchronization and alignment; then extract meaningful high-dimensional features from the pre-processed multi-condition fault data, and determine the meaningful high-dimensional features in the fault data under each condition and then fuse them, and obtain the high-dimensional features of the multi-condition fault data based on the fusion results.

[0035] Step 103, obtaining the real-time operating data of the aircraft engine collected by the sensor, and inputting the real-time operating data of the aircraft engine into the trained meta-learning model after extracting the features of the multi-modal fusion technology and multi-operating condition data, and obtaining the fault diagnosis result of the aircraft engine through the trained meta-learning model.

[0036] Specifically, in practical applications, new sensor data is input into the trained meta-learning model after being subjected to multimodal fusion technology and multi-operating condition data feature extraction. At this time, the meta-learning model can quickly extract fault features based on previously learned knowledge, and has the preliminary diagnosis capability for multiple fault types, and can quickly determine the fault type of new data, providing efficient and accurate fault diagnosis. The trained meta-learning model will map the input fault data features to different fault types, output the fault probability through the classification head (such as the softmax layer), and then select the highest category as the final fault type prediction result based on the output fault type probability. For details, please refer to Figure 6 .

[0037] More specifically, the fault diagnosis results of the aircraft engine obtained through the trained meta-learning model also include a dynamic baseline model, an adversarial transfer training module, a Markov decision module and a fault data online learning module; the dynamic baseline model, the adversarial transfer training module, the Markov decision module and the fault data online learning module are used to determine the fault type corresponding to the new fault data when new fault data is encountered during fault diagnosis through the trained meta-learning model.

[0038] Among them, the dynamic baseline model is used to learn the fault characteristics of the engine under multiple working conditions. The dynamic baseline model is updated in real time as the working conditions change to ensure the adaptability of the model to different working states; the dynamic baseline model is used to extract fault characteristics under a certain working condition. Adversarial transfer training module: In order to ensure that the dynamic baseline model can adapt to new working conditions and data distribution differences, adversarial transfer training is adopted; this method uses the model generated by training to perform transfer learning between the target working condition and the source working condition, so that the characteristics under the target working condition are consistent with the source working condition; the goal of adversarial transfer is to reduce the distribution difference of the feature space under different working conditions, so that the dynamic baseline model can still correctly identify faults under new working conditions; the role of adversarial transfer training: adversarial transfer training enhances the adaptability and generalization ability of the dynamic baseline model, ensuring effective diagnosis under new working conditions.

[0039] Reinforcement learning and Markov decision: Markov decision process (MDP): Fault diagnosis under multiple working conditions is a dynamic process, which can usually be modeled as a Markov decision process (MDP); in MDP, the state of the fault diagnosis system changes over time, and the model makes decisions (diagnose the fault or not) based on the current state and obtains feedback (reward or penalty); state: the current working state of the system and sensor data; action: the diagnostic decision of the model (such as determining the type of fault); transition probability: the probability of the current state and action affecting the next state; reward function: when the diagnosis is correct, the model will receive a high reward; if it is wrong, it will receive a low reward or penalty.

[0040] Reward function and KL divergence: In reinforcement learning, the reward function is an important mechanism for guiding model optimization. In order to evaluate the accuracy of the diagnosis results, the KL divergence (Kullback-Leibler Divergence) is used to measure the difference between the predicted results and the actual distribution. The smaller the KL divergence, the closer the model's diagnosis results are to the actual data distribution. The role of the reward function and KL divergence: KL divergence is used to optimize the model's diagnostic ability, so that the model can gradually tend to the optimal decision under multiple working conditions. Reward function: Reinforcement learning adjusts the reward function to encourage the model to gradually improve the accuracy of fault feature extraction under different working conditions. Through reinforcement learning and Markov decision making, the dynamic baseline model can be adjusted more quickly. Fault data online learning module: The online learning module plays an important role in multi-condition diagnosis. It updates data and models in real time to ensure that the diagnostic model can maintain efficient fault identification capabilities when the operating conditions change. Its workflow is as follows: receive data under new conditions and incrementally update the dynamic baseline model and deep autoencoder; through the online learning mechanism, the multi-condition meta-learning diagnostic model can quickly adapt and make correct diagnoses when faced with new conditions.

[0041] Furthermore, after determining the fault type corresponding to the new fault data, the method also includes: using the new fault data to train the trained meta-learning model; the specific steps of using the new fault data to train the trained meta-learning model are: first, inputting the high-dimensional features of the new fault data into the trained meta-learning model, and the trained meta-learning model performs a forward propagation according to the initialization parameters, and calculates the loss value after the forward propagation; then the trained meta-learning model updates the current initialization parameters through the gradient information of a small amount of new fault data.

[0042] In actual operation, the task is quickly adapted: Data input: the newly extracted fault feature data is input into the trained meta-learning model, which performs a forward propagation based on the initialized parameters and calculates the loss value (i.e., the difference between the fault type prediction and the true label, if there is a historical label); Gradient update: through the gradient information of a small amount of new data, the model quickly updates its parameters (usually only 1-2 updates are required) so that the model can adapt to new fault signals; The gradient update formula based on the meta-learning algorithm (such as MAML: Model-Agnostic Meta-Learning) is: Where β represents the learning rate of the meta-learner; fine-tune the optimal initialization parameter θ through the dataset in the training task to obtain the most suitable parameter θ for the new fault task * .

[0043] The second embodiment of the present invention relates to an intelligent fault diagnosis method for aircraft engines based on meta-learning. The second embodiment is a detailed discussion of the first embodiment as a whole. The main detailed discussion is: in the second embodiment of the present invention, an embodiment is clarified, and this embodiment discusses the specific process of obtaining high-dimensional features of single operating condition fault data through multimodal fusion technology.

[0044] For this implementation, please refer to Figure 2 , including the following steps, are described as follows:

[0045] Step 201 , obtaining single operating condition fault data about an aircraft engine, and performing time synchronization and alignment on data from different data sources in the single operating condition fault data.

[0046] Step 202: pre-process the single-operating-condition fault data after time synchronization and alignment.

[0047] Among them, the processing includes denoising processing, normalization processing and format standardization processing.

[0048] Step 203 , extracting meaningful high-dimensional features from the preprocessed single-operating-condition fault data, and fusing the extracted high-dimensional features of each modal data, and obtaining high-dimensional features of the single-operating-condition fault data according to the fusion result.

[0049] Specifically, the multi-source input data is processed and features are extracted through multi-modal fusion technology. The definition of multi-modal fusion technology is as follows: Multi-modal data: flight data from different sensors, component-level baseline model simulation results, historical data, expert experience and other sources; Multi-modal fusion: by fusing these heterogeneous data, collaborative analysis of data is achieved, thereby extracting features with higher information density and fault characterization capabilities.

[0050] Steps of multimodal fusion: (1) Data synchronization and alignment: First, data from different data sources need to be synchronized and aligned in time. Since different sensors have different sampling frequencies and data formats, they must be converted into unified time series data and the temporal and spatial consistency between the data must be ensured. Time alignment: Flight data and simulation data are aligned through a unified timestamp to ensure that the data is consistent at the same time. The data collected by the sensor may be interfered by noise, such as environmental noise, sensor error, etc. Therefore, it is necessary to process the data through denoising techniques such as filters to remove useless interference signals. The magnitudes of data from different sensors are different. For example, temperature and vibration signals may have large magnitude differences. Therefore, the data needs to be normalized so that the input values ​​of different physical quantities have consistent weights in the model. Format standardization: Data of different modes (temperature, pressure, vibration, etc.) have different dimensions and units, so they need to be standardized so that they can be processed in the same model.

[0051] (2) Feature extraction and representation learning: After completing data alignment, meaningful high-dimensional features are extracted from multimodal data through representation learning methods. Here, deep learning modules such as convolutional neural networks are used for feature extraction. (3) Inter-modal feature fusion: Data from different modalities are combined through feature fusion strategies, which mainly include the following two methods: 1. Early fusion: Multimodal data are directly spliced ​​into high-dimensional inputs and input into the model for joint training; this method is suitable for situations where there is strong correlation between modal data; 2. Late fusion: First, data from different modalities are independently modeled and feature extracted to obtain their respective feature representations, and then the features of different modalities are fused through weighted averaging, splicing, etc. This method is suitable for situations where there is heterogeneity and weak correlation between modal data.

[0052] (4) Data fusion and modeling: After feature extraction, the fused features are input into the trained cloud learning model for fault prediction and diagnosis; The role of data fusion: Multimodal fusion integrates data from different modes to supplement the detailed information lost in single-mode data, reduce uncertainty, and improve the robustness and generalization ability of the diagnostic model to changes in working conditions.

[0053] Step 204 , randomly select a task batch from a plurality of related task sets in the high-dimensional features of the single operating condition fault data to form a plurality of subtasks.

[0054] Step 205, determining the initialization model parameters, and using the high-dimensional features in each task to update the initialization model parameters, thereby obtaining a preliminarily trained meta-learning model.

[0055] Specifically, the training of the meta-learning model uses the high-dimensional features of the fault data extracted by multimodal fusion technology, and is trained through a gradient-based meta-learning algorithm (such as MAML: Model-Agnostic Meta-Learning); a convolutional neural network CNN is used as the core of the meta-learning model, and the parameters of the meta-learning model are continuously updated and adjusted during the meta-learning process; through meta-learning, the model can quickly adapt to different tasks without having to train from scratch each time; specifically, by training the meta-learning model on multiple related tasks, the meta-learning model can learn the common features between tasks, thereby enabling the meta-learning model to quickly adapt to new tasks.

[0056] More specifically, the meta-learning model is trained through the following steps:

[0057] (1) Task batch selection: A task batch is randomly selected from multiple related task sets in the fault sample training data to form multiple subtasks; each task can be regarded as an independent aircraft engine fault training task.

[0058] (2) Model initialization: Initialize the model parameters θ. This set of parameters is the common starting point for the model to quickly learn on various tasks. In the inner loop (Task-SpecificUpdate): For each specific task T i : Training on the task: Using task T i The training data in the task T is used to update the model parameters once or multiple times using the gradient descent method to obtain the task-specific parameters θ′; i The update formula is: Among them, θ′ represents the optimization parameter obtained after updating through this task; T i is the i-th task in the task set; α is the learning rate of the inner loop network; f θ The parameter in the inner loop is f θ The network structure; is a loss function for a specific task; Is the gradient of the loss. Test on the task: Using task T i , and calculate the loss of the updated model parameters θ′.

[0059] In the outer loop (Meta-Update): For the test results of each task, the loss is calculated and the meta-learning model parameters are updated. The core idea of ​​MAML is to update the initialization parameters θ of the model to parameters that can perform well on multiple tasks; therefore, it is expected that the optimized parameters θ′ can minimize the loss of all different tasks. The formula is as follows: In this formula, P(T) represents the distribution of the task set. Through this process, the meta-learning model gradually learns how to quickly adapt to each new task, that is, to find a good initial point θ′ so that the model can complete the adaptation through a small amount of data and fast gradient updates when facing new fault types or changes in working conditions. Therefore, the meta-learning model will continuously update parameters during the training process to ensure that it can quickly adjust and provide accurate diagnostic results when facing new fault types or changes in working conditions.

[0060] Steps 206 to 207 are similar to steps 102 to 103 in the first embodiment, and are not described in detail here. For the diagnosis process of a single working condition, please refer to Figure 7 on the left.

[0061] The third embodiment of the present invention relates to an intelligent fault diagnosis method for aircraft engines based on meta-learning. The third embodiment is a detailed discussion of the second embodiment as a whole. The main detailed discussion is: in the third embodiment of the present invention, an embodiment is clarified, and this embodiment discusses the specific process of obtaining high-dimensional features of multi-operating condition fault data through multimodal fusion technology and multi-operating condition data feature extraction.

[0062] For this implementation, please refer to Figure 3 , including the following steps, are described as follows:

[0063] Steps 301 to 305 are similar to steps 201 to 205 in the second embodiment, and are not described in detail herein.

[0064] Step 306 , obtaining multi-operating condition fault data about the aircraft engine, and performing time synchronization and alignment on data from different data sources in the multi-operating condition fault data.

[0065] Step 307: pre-process the time-synchronized and aligned multi-operating-condition fault data.

[0066] Among them, the preprocessing includes data cleaning and denoising processing and working condition identification and normalization processing.

[0067] Step 308, extract meaningful high-dimensional features from the preprocessed multi-condition fault data, determine the meaningful high-dimensional features in the fault data under each condition, and then fuse them, and obtain the high-dimensional features of the multi-condition fault data based on the fusion results; use the high-dimensional features of the multi-condition fault data to train the initially trained meta-learning model, and obtain the trained meta-learning model based on the training results.

[0068] Specifically, feature extraction of multi-condition data: multi-condition data is not only fused through the steps of multimodal fusion, but also deep autoencoders and dynamic baseline models are added to extract features from multi-condition data.

[0069] (1) Data acquisition and preprocessing: Under multiple working conditions, the engine runs in different working states, and the data collected by the sensor will show different dynamic characteristics as the working conditions change. These data include vibration signals, pressure, temperature and other multi-dimensional information. Data cleaning and denoising: Under different working conditions, the data may be affected by environmental noise and external interference, so data denoising, cleaning and other preprocessing operations are required to ensure the quality of the input data. Working condition identification and normalization: Since the characteristics of multiple working conditions are obviously different, it is first necessary to mark each working condition to distinguish the specific working state of the data. At the same time, normalization can eliminate the order of magnitude difference of sensor data under different working conditions, so that the trained meta-learning model maintains data consistency in the input stage.

[0070] (2) Feature extraction and operating condition identification: A deep autoencoder is a neural network structure used to automatically extract high-dimensional features of data. It compresses the input data into a low-dimensional representation and then decompresses and reconstructs the original data to extract the main features in the data. The specific process is as follows: Encoder: maps high-dimensional data under multiple operating conditions to a low-dimensional feature space, which helps to eliminate redundant information in the data and retain important operating condition features; Decoder: reconstructs the low-dimensional features into high-dimensional data to evaluate the accuracy of the features learned by the encoder; The role of a deep autoencoder is to eliminate the differences between complex data under multiple operating conditions, so that the model can automatically capture the common features and fault features under each operating condition, thereby improving the generalization ability of the diagnosis. There is also a dynamic baseline model used to learn the fault features of the engine under multiple operating conditions. The dynamic baseline model is updated in real time as the operating conditions change to ensure the adaptability of the model to different working states. The dynamic baseline model is used to extract fault features under a certain operating condition.

[0071] Step 309 is similar to step 207 in the third embodiment and will not be described in detail herein.

[0072] Step 310, evaluating the fault diagnosis result of the aircraft engine.

[0073] Specifically, the fault diagnosis results are output through the result output module and displayed to the user; the evaluation module verifies the diagnosis results to ensure their accuracy and reliability. The evaluation module compares the diagnosis results with known fault data and calculates indicators such as diagnostic accuracy and recall rate to evaluate model performance. For the diagnosis process of multiple working conditions, please refer to Figure 7 on the right.

[0074] Furthermore, the present invention adopts meta-learning technology for fault diagnosis, which is the core innovation that distinguishes it from traditional fault diagnosis methods. Meta-learning enables the model to quickly adapt to new tasks by training on multiple related tasks and learning the common features between tasks. In particular, in aircraft engine fault diagnosis, meta-learning can quickly adapt to new fault types and operating condition changes with a small amount of labeled data, greatly improving the accuracy and efficiency of diagnosis. In the feature extraction process, the present invention automatically extracts high-dimensional feature representations through a deep learning model, and further combines multimodal fusion technology to extract key features related to faults from different sensor data. These high-quality features are the basis for meta-learning model training and help improve the diagnostic accuracy of the model. The fault diagnosis system of the present invention is real-time. Through efficient feature extraction and rapid adaptation of the meta-learning model, it can complete the diagnosis of new faults in a short time. In addition, the system is provided with a feedback mechanism, which can promptly feed back the diagnosis results to the engine control system to assist decision-making and further ensure aviation safety.

[0075] The fourth embodiment of the present invention relates to an intelligent fault diagnosis system for aircraft engines based on meta-learning. Figure 4 ,include:

[0076] The single-operating-condition training module is used to obtain single-operating-condition fault data of aircraft engines and obtain high-dimensional features of the single-operating-condition fault data through multimodal fusion technology; the high-dimensional features of the single-operating-condition fault data are used to train the meta-learning model, and a preliminarily trained meta-learning model is obtained based on the training results;

[0077] The multi-operating condition training module is used to obtain multi-operating condition fault data about aircraft engines, and obtain high-dimensional features of the multi-operating condition fault data through multi-modal fusion technology and multi-operating condition data feature extraction; the high-dimensional features of the multi-operating condition fault data are used to train the preliminarily trained meta-learning model, and the trained meta-learning model is obtained according to the training results;

[0078] The fault diagnosis module is used to obtain the real-time operating data of the aircraft engine collected by sensors. The real-time operating data of the aircraft engine is input into the trained meta-learning model through multimodal fusion technology and multi-operating condition data feature extraction, and the fault diagnosis results of the aircraft engine are obtained through the trained meta-learning model.

[0079] Specifically, the aircraft engine fault diagnosis system of the present invention also includes the following modules: a data acquisition and preprocessing module: collects data from multiple sensors on the aircraft engine in real time, and cleans, denoises, and normalizes the data to generate high-quality input data; a feature extraction module: uses a deep learning model to extract features from the preprocessed data and automatically obtains high-dimensional feature representations; a meta-learning model training module: constructs and trains a meta-learning model to learn the common features of multiple related tasks, so that the model has the ability to quickly adapt to new fault types and changes in operating conditions; a fault diagnosis module: uses a meta-learning model to perform fault diagnosis on new data in actual applications and quickly determine the fault type; a result output and evaluation module: outputs the diagnosis results, and evaluates the accuracy and reliability of the diagnosis through comparative analysis and verification.

[0080] Furthermore, the closed-loop fault diagnosis system adopted by the present invention is composed of an online learning module and a gas path fault diagnosis module, and realizes the virtual-reality mapping between the digital engine and the real engine through information transmission; wherein, the virtual-reality mapping is to map the behaviors of the digital engine (i.e., the digital model of the engine) and the real engine (physical engine), thereby realizing the interaction and feedback between the virtual world and the real world. This process ensures the reliability and fault diagnosis accuracy of the engine under complex working conditions by transmitting real-time flight data to the digital engine and feeding back the simulation results of the digital engine to the control system of the real engine; the physical system receives the flight data and processes the flight data through multimodal fusion technology and a dynamic baseline model, thereby extracting high-quality data features; finally, fault pre-training is performed through a meta-learning method, which can predict faults of unknown labels in advance in the case of small samples, and adapt to changes in multiple working conditions by updating the dynamic baseline model.

[0081] It is not difficult to find that this embodiment is a system embodiment corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and in order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the first embodiment.

[0082] It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of the present invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other units in this embodiment.

[0083] The fifth embodiment of the present invention relates to an electronic device, see Figure 5 ,include:

[0084] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned meta-learning-based intelligent fault diagnosis method for aircraft engines.

[0085] Among them, the memory and the processor are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0086] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0087] In summary, the present invention can improve the accuracy of diagnosis: by introducing meta-learning technology and utilizing the common features of multiple related tasks, the accuracy of the fault diagnosis model is enhanced, especially in the case of a small amount of labeled data, efficient and accurate fault diagnosis can be achieved; the present invention can enhance adaptability and generalization ability: the meta-learning model can quickly adapt to new fault types and operating condition changes by learning the common features between multiple tasks, significantly improving the adaptability and generalization ability of the model; the present invention can reduce data dependence: the meta-learning method of the present invention can achieve effective fault diagnosis in the case of small samples, reducing the dependence on a large amount of labeled data, so that accurate fault diagnosis can be performed even in the case of data scarcity; the present invention can reduce development and maintenance costs: the present invention reduces the dependence on expert knowledge and manual rule setting through automatic feature extraction and meta-learning methods, thereby reducing the development and maintenance costs of the system; the present invention can improve diagnostic efficiency and real-time performance: the method of the present invention achieves real-time fault diagnosis through efficient feature extraction and rapid adaptability of meta-learning, can promptly detect and handle faults, reduce engine downtime, and improve operating efficiency.

[0088] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. All equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical ideas disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. An intelligent fault diagnosis method for aircraft engines based on meta-learning, characterized in that: The following steps are involved: Acquire single-operating-condition fault data of an aircraft engine, and obtain high-dimensional features of the single-operating-condition fault data through multimodal fusion technology; The meta-learning model is trained using the high-dimensional features of the single-operating-condition fault data, and a preliminarily trained meta-learning model is obtained according to the training results; Acquire multi-operating condition fault data about an aircraft engine, and obtain high-dimensional features of the multi-operating condition fault data through multi-modal fusion technology and multi-operating condition data feature extraction; Using the high-dimensional features of the multi-operating condition fault data to train the initially trained meta-learning model, and obtaining a trained meta-learning model according to the training results; Real-time operating data of an aircraft engine collected by a sensor is acquired. The real-time operating data of the aircraft engine is input into a trained meta-learning model after being extracted through multimodal fusion technology and multi-operating condition data features. The fault diagnosis result of the aircraft engine is obtained through the trained meta-learning model.

2. The method for intelligent fault diagnosis of an aircraft engine based on meta-learning according to claim 1, characterized in that: The high-dimensional features of the single-operating-condition fault data are obtained by multimodal fusion technology, including: Time synchronization and alignment of data from different data sources in the single operating condition fault data; Preprocess the single-condition fault data after time synchronization and alignment; The meaningful high-dimensional features are extracted from the preprocessed single-operating-condition fault data, and the high-dimensional features of each modal data extracted are fused to obtain the high-dimensional features of the single-operating-condition fault data according to the fusion results.

3. The method for intelligent fault diagnosis of an aircraft engine based on meta-learning according to claim 1, characterized in that: The high-dimensional features of the single-operating-condition fault data are used to train the meta-learning model, and a preliminarily trained meta-learning model is obtained according to the training results, including: Randomly selecting a task batch from a plurality of related task sets in the high-dimensional features of the single-condition fault data to form a plurality of subtasks, wherein each task can be regarded as an independent aircraft engine fault training task; Determine the initialization model parameters and use the high-dimensional features in each task to update the initialization model parameters.

4. The method for intelligent fault diagnosis of an aircraft engine based on meta-learning according to claim 3 is characterized in that: The step of determining the initialization model parameters and updating the initialization model parameters using the high-dimensional features in each task includes: In an inner loop, the training data in the task is used and the initialization model parameters are updated one or more times using the gradient descent method, and task-specific parameters are obtained based on the one or more update results; and then the loss of the updated model parameters is determined using the test data in the task; In the outer loop, for each task’s test result, the loss sum is calculated and the initialized model parameters are updated.

5. The method for intelligent fault diagnosis of aircraft engines based on meta-learning according to claim 1, characterized in that: The high-dimensional features of the multi-operating condition fault data are obtained by multi-modal fusion technology and multi-operating condition data feature extraction, including: Time synchronization and alignment of data from different data sources in the multi-operating condition fault data; Preprocess the multi-condition fault data after time synchronization and alignment; The meaningful high-dimensional features in the preprocessed multi-condition fault data are extracted, and the meaningful high-dimensional features in the fault data under each condition are determined and then fused, and the high-dimensional features of the multi-condition fault data are obtained according to the fusion results.

6. The method for intelligent fault diagnosis of aircraft engines based on meta-learning according to claim 1, characterized in that: The fault diagnosis result of the aircraft engine obtained by the trained meta-learning model also includes a dynamic baseline model, an adversarial transfer training module, a Markov decision module and a fault data online learning module; The dynamic baseline model, adversarial transfer training module, Markov decision module and fault data online learning module are used to determine the fault type corresponding to the new fault data when new fault data is encountered during fault diagnosis using the trained meta-learning model.

7. The method for intelligent fault diagnosis of an aircraft engine based on meta-learning according to claim 6 is characterized in that: After determining the fault type corresponding to the new fault data, the method further includes: Use new fault data to train the trained meta-learning model; The method of using new fault data to train the trained meta-learning model includes: The high-dimensional features of the new fault data are input into the trained meta-learning model. The trained meta-learning model performs a forward propagation according to the initialization parameters and calculates the loss value after the forward propagation. The trained meta-learning model updates the current initialization parameters through the gradient information of a small amount of new fault data.

8. The method for intelligent fault diagnosis of an aircraft engine based on meta-learning according to claim 1, characterized in that: After obtaining the fault diagnosis result of the aircraft engine through the trained meta-learning model, the method further includes: The fault diagnosis results of the aircraft engine are evaluated.

9. An intelligent fault diagnosis system for aircraft engines based on meta-learning, characterized by: include: A single-operating-condition training module is used to obtain single-operating-condition fault data about aircraft engines and obtain high-dimensional features of the single-operating-condition fault data through multimodal fusion technology; The meta-learning model is trained using the high-dimensional features of the single-operating-condition fault data, and a preliminarily trained meta-learning model is obtained according to the training results; A multi-operating condition training module is used to obtain multi-operating condition fault data about aircraft engines, and obtain high-dimensional features of the multi-operating condition fault data through multi-modal fusion technology and multi-operating condition data feature extraction; Using the high-dimensional features of the multi-operating condition fault data to train the initially trained meta-learning model, and obtaining a trained meta-learning model according to the training results; The fault diagnosis module is used to obtain real-time operating data of the aircraft engine collected by sensors. The real-time operating data of the aircraft engine is input into a trained meta-learning model after multi-modal fusion technology and multi-operating condition data feature extraction, and the fault diagnosis result of the aircraft engine is obtained through the trained meta-learning model.

10. An electronic device, characterized in that: include: at least one processor; as well as, A memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the intelligent fault diagnosis method for aircraft engines based on meta-learning as described in any one of claims 1 to 8.

Images