Fault detection method and system for electric vertical take-off and landing aircraft

Abstract

The invention provides a fault detection method and system for an electric vertical take-off and landing aircraft, and relates to the technical field of aircraft fault detection. The method comprises the steps of performing multi-source data fusion and adaptive fault feature extraction, and constructing an initial fault data set; constructing a generative adversarial network based on the initial fault data set, and forming an expanded fault data set; according to the expanded fault data set, constructing a fault propagation model to simulate a dynamic propagation path of a fault between components in the electric vertical take-off and landing aircraft, and cooperatively optimizing a fault identification threshold by combining a federated learning framework; and constructing a hypergraph structure, processing a fault propagation path in the hypergraph structure by using hypergraph correlation analysis, and outputting a fault diagnosis result in combination with the optimized fault recognition threshold. According to the method, the fault detection precision, the early warning capability and the system self-adaptive evolution capability can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of aircraft fault detection technology, and specifically to a fault detection method and system for electric vertical takeoff and landing aircraft. Background Technology

[0002] Electric Vertical Take-off and Landing (eVTOL) aircraft, as a core carrier of urban air mobility, have a high degree of system integration and tight coupling across multiple electromechanical, hydraulic, and thermal domains. During operation, even a minor fault in a single component can trigger a cascading effect through a complex network of interactions, ultimately leading to system-wide functional failure. However, existing fault detection schemes have significant limitations: First, they mostly rely on single-parameter threshold judgments or shallow machine learning models, making it difficult to extract effective fault precursor features from high-dimensional, nonlinear, and dynamically changing flight data; second, since eVTOL is still in the early stages of commercialization, the number of real fault samples available for training is extremely scarce, severely limiting the generalization performance of data-driven models; finally, existing methods generally lack the ability to explicitly model the internal fault propagation paths of the system, failing to achieve early warning and accurate fault tracing.

[0003] Therefore, there is an urgent need in this field for a new technology solution that can comprehensively utilize a small amount of real data and prior knowledge, and accurately simulate fault propagation dynamics, thereby achieving early, accurate, and adaptive fault detection. Summary of the Invention

[0004] To address the technical problems of existing fault detection schemes that rely on single parameters or shallow models, making it difficult to extract effective features, the scarcity of real fault samples limiting the generalization performance of models, and the lack of explicit modeling capabilities for fault propagation paths, thus hindering early warning and accurate source tracing, this invention provides a fault detection method and system for electric vertical takeoff and landing aircraft. The core of this invention lies in constructing a closed-loop technical framework that extends from data augmentation to propagation simulation and then to intelligent diagnosis. Through multi-level information fusion and reasoning, it enables the accurate capture and location of complex faults.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A fault detection method for an electric vertical takeoff and landing aircraft, the method comprising: Multi-source data fusion and adaptive fault feature extraction are performed to construct an initial fault dataset; Based on the initial fault dataset, a generative adversarial network is constructed to form an expanded fault dataset; Based on the expanded fault dataset, a fault propagation model is constructed to simulate the dynamic propagation path of faults among various components in an electric vertical takeoff and landing (eVTOL) aircraft, and the fault identification threshold is collaboratively optimized using a federated learning framework. A hypergraph structure is constructed, and the fault propagation path in the hypergraph structure is processed using hypergraph association analysis. Combined with the optimized fault identification threshold, the fault diagnosis result is output.

[0006] On the other hand, the present invention also provides a fault detection system for an electric vertical takeoff and landing aircraft. The system includes a memory for storing computer program instructions and a processor for executing the program instructions. When the computer program instructions are executed by the processor, the system is triggered to execute the aforementioned fault detection method for an electric vertical takeoff and landing aircraft.

[0007] Compared with the prior art, the beneficial effects of the present invention are: 1) Significantly improved detection accuracy and early warning capability: By combining "Gaussian process regression + RBF neural network", the system can sensitively capture the characteristics of weak nonlinear faults. Combined with the time-series deduction of the fault propagation model, it can significantly extend the fault warning window period and buy valuable time for safety decision-making.

[0008] 2) Significantly enhanced generalization ability for unknown faults and complex working conditions: By utilizing generative adversarial networks and domain adaptation technology, a wide-coverage and high-quality "virtual fault library" was constructed, enabling the diagnostic model to be exposed to and learn diverse fault modes during the training phase, thus maintaining a high recognition rate when facing rare or complex faults in the real world.

[0009] 3) Achieved accurate fault tracing and impact analysis: By using cellular automata and hypergraph models, fault detection is expanded from "points" to "lines" and "surfaces". It can not only locate the source of the fault, but also clearly depict its propagation path and scope of impact, which greatly improves the interpretability of diagnostic results and operation and maintenance efficiency.

[0010] 4) The system possesses self-learning and adaptive evolution capabilities: The introduction of the federated learning framework enables the entire fault detection system to continuously evolve as fleet operating experience accumulates. Fault identification thresholds and model parameters can be dynamically adjusted, thereby continuously adapting to hardware aging, environmental changes, and emerging operating modes, ensuring the robustness and reliability of the system throughout its entire lifecycle.

[0011] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0012] Figure 1 This is a flowchart of a fault detection method for an electric vertical takeoff and landing aircraft according to the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings, so as to more clearly understand the purpose, features and advantages of this invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of this invention, but are only for illustrating the essential spirit of the technical solutions of this invention. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0014] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0015] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0016] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0017] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0018] The implementation details of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0019] This invention proposes a fault detection method and system for electric vertical takeoff and landing (EVTOL) aircraft, aiming to effectively solve the aforementioned technical problems. The flowchart of the fault detection method for EVTOL aircraft is shown below. Figure 1 As shown, the method specifically includes: S1. Perform multi-source data fusion and adaptive fault feature extraction to construct an initial fault dataset.

[0020] By deploying a multi-source sensor network in key components of the eVTOL, such as the power system, flight control system, and energy system, raw data is collected synchronously. This raw data includes time-series data such as vibration, temperature, current, voltage, pressure, and attitude angle. Subsequently, the raw data undergoes alignment, filtering, denoising, and standardization preprocessing to eliminate dimensional differences and improve the signal-to-noise ratio.

[0021] The Gaussian Process Regression (GPR) algorithm, a nonparametric regression method based on Gaussian processes, is used to probabilistically model the behavior of electric vertical takeoff and landing (EVTOL) aircraft under normal conditions based on historical health data. GPR can provide predicted values ​​and their uncertainty ranges, thereby constructing a dynamic health model with confidence intervals.

[0022] Real-time operational data is input into the health model to calculate its predicted residuals. If the residual sequence continuously exceeds the dynamic health boundary set based on the uncertainty of the health model, the sequence is marked as deviation data, indicating that the system has an anomaly.

[0023] A Radial Basis Function Neural Network (RBFNN) is pre-trained with off-dimensional data as input. Through its local receptive field and nonlinear mapping capabilities, the RBFNN maps high-dimensional off-dimensional data to a low-dimensional fault feature space and outputs a preliminary probability that the off-dimensional data belongs to various known faults (such as motor overheating, sensor drift, etc.). The RBFNN is a feedforward neural network based on radial basis functions, possessing local receptive fields and nonlinear mapping capabilities, enabling it to map high-dimensional input data to a low-dimensional feature space. It is commonly used for tasks such as pattern recognition, classification, and regression.

[0024] The off-target data that has been classified by RBFNN are associated with their fault type labels to construct a structured initial fault dataset.

[0025] S2. Based on the initial fault dataset, construct a generative adversarial network to form an expanded fault dataset.

[0026] Using the real initial fault dataset obtained in step S1 as the training set, a conditional generative adversarial network (GAN) is constructed. The generator G takes random noise and fault type labels as input and learns to generate virtual fault data with realistic temporal characteristics; the discriminator D tries to distinguish between real data and generated data, and the two evolve together through adversarial training.

[0027] Domain adaptation techniques are introduced into the training of GANs. By adding a domain classifier to the feature layer of the discriminator D and using a gradient reversal layer, the data generated by the generator G is forced to not only approximate the real data in terms of overall distribution, but also to make its deep features indistinguishable from the real data in the feature space, thereby achieving effective alignment between the virtual fault domain and the real fault domain.

[0028] Quality assessment and dataset fusion: The generated virtual fault data is quantitatively evaluated using metrics such as structural similarity index and dynamic time warping distance. Quality-compliant virtual fault data is added to the initial fault dataset to form an expanded fault dataset, providing ample data support for subsequent model training.

[0029] S3. Based on the expanded fault dataset, construct a fault propagation model to simulate the dynamic propagation path of faults among various components in an electric vertical takeoff and landing aircraft, and combine it with a federated learning framework to collaboratively optimize the fault identification threshold.

[0030] Specifically, based on the expanded fault dataset, the entire eVTOL is divided into several functional components (such as battery cells, motors, ESCs, flight control computers, etc.), and a directed topology graph describing the connection relationships between components is constructed according to the physical connections, energy flow and information flow relationships.

[0031] Each component in the directed topology graph is mapped to a cell in the cellular automaton model. Three states are defined for each cell: healthy (H), alert (A), and faulty (F). Simultaneously, state transition rules based on neighborhood states are defined; for example, if a cell's neighboring cell is in faulty state F, then that cell has a certain probability of entering alert state A in the next time step.

[0032] Dynamic simulation of fault propagation is performed based on the constructed cellular automata model. To more realistically reflect the physical process, a time-gating function is introduced. g ( t ): Among them, P ij Basic propagation probability; Δt is the time difference; τ ij Minimum propagation delay represents the shortest time required for a fault to propagate from node i to node j, used to simulate the inherent delay of signal transmission or physical effects; β γ is the sensitivity adjustment parameter; γ is the time decay coefficient.

[0033] This time-gating function can dynamically adjust the timing and intensity of fault state propagation, thereby constructing a fault propagation graph that includes time information.

[0034] Next, the fault identification threshold (a threshold used to define whether a component is operating normally) is collaboratively optimized using a fault propagation model and a federated learning framework: Local fault detection models are deployed on multiple eVTOL nodes. Each node uses local data to calculate model updates and uploads the encrypted updates to the central server via a secure aggregation protocol. The server aggregates global information, updates model parameters (including the optimal fault identification threshold for each component), and then distributes the new model to each node. Through this federated learning paradigm, continuous collaborative optimization of the fault identification threshold is achieved while ensuring data privacy, enabling it to adapt to differences in different nodes and operating conditions. This collaborative optimization mechanism for the fault identification threshold can match fluctuations in operating conditions in real time, giving the fault diagnosis results both strong correlation path tracing capabilities and robustness across multiple scenarios.

[0035] S4. Construct a hypergraph structure and use hypergraph association analysis to process the fault propagation path in the hypergraph structure. Combine the optimized fault identification threshold to output the fault diagnosis result.

[0036] The system components of eVTOL are abstracted as supernodes. Going beyond the traditional pairwise connection method in graph theory, based on rules such as functional coupling (e.g., multiple sensors jointly participate in a control law), fault co-occurrence (certain faults often occur simultaneously in historical data), and information interaction (sharing the same data bus), multiple (two or more) strongly related supernodes are connected by hyperedges to construct a hypergraph structure.

[0037] We analyze fault propagation paths in hypergraph structures using gated attention propagation networks. A multilayer perceptron is used to encode the state and attributes of each hypernode, obtaining its feature vector. h Additive attention and bidirectional gating mechanisms are employed to calculate the correlation between supernodes m and n. S mn : S mn For relevance score; h m h n W represents the eigenvectors of supernodes m and n. a and b a These represent the weight matrix and bias vector of the additive attention layer, respectively. v W represents the projection vector from the attention space onto real numbers. g1 W g2 b g σ represents the weight matrix and bias vector of the bidirectional gating mechanism; σ is the Sigmoid activation function.

[0038] Finally, using the attention weight matrix, the features of all supernodes connected by each superedge are summed in a weighted manner, and the propagation intensity of the fault is dynamically adjusted in different superedges by combining the state information of the supernodes with the propagation environment.

[0039] Combining the weighted propagation strength obtained from the above steps, the real-time status of each supernode, the path information provided by the fault propagation model, and the optimized fault identification threshold, a comprehensive inference module makes a decision. This comprehensive inference module outputs a final fault diagnosis report, which includes: fault type, location information, severity level, confidence level, and recommended response strategy.

[0040] Based on the above-mentioned fault detection method for electric vertical takeoff and landing aircraft, this invention provides a specific embodiment for early thermal runaway fault detection in a battery management system.

[0041] 1) Scenario and Data: This embodiment is for a high-voltage battery pack for eVTOL. Sensors collect the voltage and temperature of each battery cell, as well as the total current of the battery pack.

[0042] 2) Fault extraction is performed based on step S1 of the above method. The health model (GPR) predicts the voltage and temperature range of each cell based on current and historical health data.

[0043] The system detected that the measured temperature of a certain unit was consistently slightly higher than the predicted upper limit. Although it did not reach the traditional fixed threshold alarm line, it was identified as "deviation data".

[0044] The probability of RBFNN classifying this deviation data as "early stage of increased internal resistance" is 78%.

[0045] 3) Data augmentation based on step S2 of the above method: The GAN model generates virtual fault data under different ambient temperatures and different discharge rates based on such "increased internal resistance" samples, which enriches the training set.

[0046] 4) Based on step S3 of the above method, perform propagation simulation and threshold optimization: In the fault propagation model, the state of a single cell changes from "healthy H" to "warning A".

[0047] The time-gating function begins calculating the probability of this warning state propagating to adjacent cells and the battery management unit. Due to the inertia of thermal propagation, the time delay τ... ij It was set to a few seconds.

[0048] Based on the experience of other aircraft under similar operating conditions, the federated learning framework dynamically lowered the temperature warning threshold for this scenario, enabling the system to trigger warnings earlier.

[0049] 5) Perform intelligent diagnosis based on step S4 of the above method: The hypergraph structure connects a certain unit, coolant pump, and cooling fan, etc., through a "thermal management hyperedge".

[0050] The gated attention network calculates the correlation between a specific component and the coolant pump under the current high ambient temperature. S mn Extremely high.

[0051] Based on all the information, the system output a diagnosis minutes before thermal runaway occurred: "A certain battery cell has an early increase in internal resistance and overheating risk. It is recommended to immediately strengthen cooling and prepare for landing," and automatically increased the power of the cooling system. This successfully averted a potentially serious thermal runaway accident.

[0052] The present invention also provides a fault detection system for an electric vertical takeoff and landing (EVTOL) aircraft. The system includes a memory for storing computer program instructions and a processor for executing the program instructions. When the computer program instructions are executed by the processor, the system is triggered to execute the aforementioned fault detection method for EVTOL aircraft.

[0053] This invention proposes a fault detection method and system for electric vertical takeoff and landing (EVTOL) aircraft. An initial fault dataset is constructed through multi-source data fusion and adaptive fault feature extraction. Generative adversarial networks are used to expand the dataset to enhance the model's generalization ability. A fault propagation model is built to simulate the dynamic propagation path of faults, and a federated learning framework is combined to optimize the fault identification threshold, achieving early warning and accurate fault tracing. Furthermore, complex fault propagation paths are processed through hypergraph structures and association analysis, outputting an accurate diagnostic report containing information such as fault type, location, and severity level. This significantly improves fault detection accuracy, early warning capability, and system adaptive evolution capability.

[0054] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention. Such modifications or substitutions should all fall within the scope of the invention, or any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A fault detection method for an electric vertical takeoff and landing aircraft, characterized in that, The method includes: Multi-source data fusion and adaptive fault feature extraction are performed to construct an initial fault dataset; Based on the initial fault dataset, a generative adversarial network is constructed to form an expanded fault dataset; Based on the expanded fault dataset, a fault propagation model is constructed to simulate the dynamic propagation path of faults among various components in an electric vertical takeoff and landing (eVTOL) aircraft, and the fault identification threshold is collaboratively optimized using a federated learning framework. A hypergraph structure is constructed, and the fault propagation path in the hypergraph structure is processed using hypergraph association analysis. Combined with the optimized fault identification threshold, the fault diagnosis result is output.

2. The method according to claim 1, characterized in that, The process of multi-source data fusion and adaptive fault feature extraction to construct an initial fault dataset specifically includes: By deploying a multi-source sensor network at key parts of eVTOL, raw data is collected synchronously, and the raw data is preprocessed by alignment, filtering, denoising and standardization. By using the Gaussian process regression (GPR) algorithm, a dynamic health model with confidence intervals is constructed by providing predicted values ​​and their uncertainty range. Real-time running data is input into the health model, and the prediction residuals of the health model are calculated. If the residual sequence continuously exceeds the dynamic health boundary set based on the uncertainty of the health model, the residual sequence is marked as deviation data, indicating that the system has an anomaly. The deviation data is input into a pre-trained radial basis function neural network (RBFNN). Through its local receptive field and nonlinear mapping capability, the RBFNN maps the high-dimensional deviation data to a low-dimensional fault feature space and outputs the preliminary probability of its belonging to various known faults. The off-target data that has been classified by RBFNN are associated with the fault type labels to construct a structured initial fault dataset.

3. The method according to claim 1, characterized in that, The key components of the eVTOL include the power system, flight control system, and energy system; the raw data includes vibration, temperature, current, voltage, pressure, and attitude angle.

4. The method according to claim 3, characterized in that, The process of constructing a generative adversarial network based on the initial fault dataset to form an expanded fault dataset specifically includes: Using the initial fault dataset as the training set, a conditional generative adversarial network (GAN) is constructed. The generator G takes random noise and fault type labels as input and learns to generate virtual fault data with realistic temporal characteristics. The discriminator D distinguishes between real data and generated data, and the two evolve together through adversarial training. In the training of GAN, a domain adaptation technique is introduced: by adding a domain classifier to the feature layer of the discriminator D and using a gradient reversal layer, the data generated by the generator G is forced to not only approximate the real data in the overall distribution, but also to make its deep features indistinguishable from the real data in the feature space, thereby achieving effective alignment of the virtual fault domain to the real fault domain. Quality assessment and dataset fusion: The generated virtual fault data is quantitatively evaluated using structural similarity index and dynamic time warping distance index; qualified virtual fault data is added to the initial fault dataset to form an expanded fault dataset.

5. The method according to claim 4, characterized in that, The process of constructing a fault propagation model based on the expanded fault dataset to simulate the dynamic propagation path of faults among various components in an electric vertical takeoff and landing aircraft specifically includes: Based on the expanded fault dataset, the entire eVTOL is divided into several functional components, and a directed topology graph describing the connection relationship between the components is constructed according to the physical connection, energy flow and information flow relationship. Each component in the directed topology graph is mapped to a cell in the cellular automaton model, and three states are defined for each cell: healthy (H), warning (A), and fault (F). At the same time, state transition rules based on neighborhood states are defined. Based on the constructed cellular automaton model, dynamic simulation of fault propagation is performed.

6. The method according to claim 5, characterized in that, The constructed cellular automata model is used for dynamic simulation of fault propagation. To more realistically reflect the physical process, a time-gating function is introduced. g ( t ): Among them, P ij Basic propagation probability; Δt is the time difference; τ ij Minimum propagation delay represents the shortest time required for a fault to propagate from node i to node j, used to simulate the inherent delay of signal transmission or physical effects; β γ is the sensitivity adjustment parameter; γ is the time decay coefficient; This time-gating function can dynamically adjust the timing and intensity of fault state propagation, and construct a fault propagation graph that includes time information.

7. The method according to claim 6, characterized in that, The method of collaboratively optimizing the fault identification threshold using a federated learning framework specifically includes: A local fault detection model is deployed on multiple eVTOL nodes. Each node uses local data to calculate model updates and uploads the encrypted updates to the central server via a secure aggregation protocol. The server aggregates global information, updates model parameters, and then distributes the new model to each node. Through this federated learning paradigm, continuous collaborative optimization of fault identification thresholds is achieved while ensuring data privacy.

8. The method according to claim 7, characterized in that, The construction of the hypergraph structure, and the processing of fault propagation paths within the hypergraph structure using hypergraph association analysis, specifically includes: The system components of eVTOL are abstracted as super nodes. Based on the rules of functional coupling, fault co-occurrence, and information interaction, two or more strongly related super nodes are connected by hyper edges to construct a hypergraph structure. We analyze fault propagation paths in hypergraph structures using gated attention propagation networks; and encode the state and attributes of each hypernode using a multilayer perceptron to obtain feature vectors. h Furthermore, additive attention and bidirectional gating mechanisms are employed to calculate the correlation degree between supernodes m and n. S mn : in, S mn For relevance score; h m h n W represents the eigenvectors of supernodes m and n. a and b a These represent the weight matrix and bias vector of the additive attention layer, respectively. v W represents the projection vector from the attention space onto real numbers. g1 W g2 b g Here, represents the weight matrix and bias vector of the bidirectional gating mechanism; σ is the Sigmoid activation function. Finally, using the attention weight matrix, the features of all supernodes connected by each superedge are summed in a weighted manner, and the propagation intensity of the fault is dynamically adjusted in different superedges by combining the state information of the supernodes with the propagation environment.

9. The method according to claim 8, characterized in that, The process of combining the optimized fault identification threshold to output fault diagnosis results specifically includes: The comprehensive reasoning module makes a decision by combining the weighted propagation strength, the real-time status of each supernode, the path information provided by the fault propagation model, and the optimized fault identification threshold. The comprehensive reasoning module outputs the final fault diagnosis report, which includes: fault type, location information, severity level, confidence level, and recommended coping strategies.

10. A fault detection system for an electric vertical takeoff and landing (EVTOL) aircraft, the system comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein, When the computer program instructions are executed by the processor, the system is triggered to execute the fault detection method for an electric vertical take-off and landing aircraft as described in any one of claims 1 to 9.

Images