A method for identifying degradation of an eVTOL electric propulsion system and related devices

Abstract

The application discloses a degradation identification method of an eVTOL electric propulsion system and related equipment, and relates to the technical field of equipment maintenance. The method comprises the following steps: acquiring motor operation data and battery management data of each electric propulsion unit of an electric vertical take-off and landing aircraft (eVTOL) and flight stage data of the eVTOL; performing coding fusion processing on the motor operation data, the battery management data and the flight stage data to determine a fusion coding data sequence of each electric propulsion unit; performing feature extraction on the fusion coding data sequence of each electric propulsion unit based on a graph structure through a neural network model to generate health state data of each electric propulsion unit at each time step; and determining a remaining life prediction result of each electric propulsion unit based on the health state data. The application builds a coupling relationship between units through a graph structure modeling unit, extracts space-time features by using a neural network, simultaneously identifies degradation behaviors of multiple units, and is convenient for management of a distributed electric propulsion system.

Description

Technical Field

[0001] This application relates to the field of equipment maintenance technology, and in particular to a degradation identification method and related equipment for an eVTOL electric propulsion system. Background Technology

[0002] In the distributed electric propulsion system of eVTOL (Electric Vertical Take-off and Landing) aircraft, multiple electric propulsion units are coupled together through electrical topology, aerodynamic layout, and heat transfer paths. Changes in the degradation state of one unit can affect the degradation process of other units due to load redistribution or environmental interactions. However, existing degradation prediction methods typically treat each electric propulsion unit as an independent object, performing state estimation and lifetime prediction solely based on the operational data of each unit, neglecting the coupled degradation relationships between electric propulsion units. Therefore, a degradation identification method for eVTOL electric propulsion systems is urgently needed to address the aforementioned technical problems. Summary of the Invention

[0003] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key and essential technical features of the claimed technical solutions, nor is it intended to determine the scope of protection of the claimed technical solutions.

[0004] In a first aspect, this application proposes a degradation identification method for an eVTOL electric propulsion system, the eVTOL electric propulsion system comprising multiple electric propulsion units, the method comprising: Acquire motor operation data and battery management data for each electric propulsion unit on the electric vertical takeoff and landing (eVTOL) aircraft, as well as flight phase data of the eVTOL; The motor operation data, the battery management data, and the flight phase data are encoded and fused to determine the fused coded data sequence for each electric propulsion unit. The fused coded data sequence includes the fused coded data for each electric propulsion unit at each time step. Based on the graph structure, features are extracted from the fused coded data sequence of each electric propulsion unit using a neural network model to generate health status data of each electric propulsion unit at each time step. Based on the health status data, the remaining lifetime prediction result for each electric propulsion unit is determined.

[0005] In some implementations, the neural network model includes a temporal attention mechanism and a graph attention mechanism. Based on a graph structure, the neural network model extracts features from the fused encoded data sequences of each electric propulsion unit to generate health status data for each electric propulsion unit at each time step, including: For each electric propulsion unit, the temporal dependency between the fused coded data at each time step in the fused coded data sequence of the electric propulsion unit is calculated through the time attention mechanism to determine the temporal enhancement feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the spatial dependency between the temporal augmentation feature data of the electric propulsion unit at the current time step and the temporal augmentation feature data of other electric propulsion units in the graph structure at the same time step is calculated through the graph attention mechanism, thereby generating the spatial aggregated feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the temporal enhancement feature data and the spatial aggregation feature data are fused to generate health status data for each electric propulsion unit at each time step.

[0006] In some embodiments, before extracting features from the fused coded data sequences of each electric propulsion unit using a neural network model based on a graph structure to generate health status data for each electric propulsion unit at each time step, the method further includes: Using each electric propulsion unit as a node, edges are constructed between the electric propulsion units to obtain a graph structure; For any two nodes in the graph structure corresponding to electric propulsion units, the spatial distance coefficient, aerodynamic interference coefficient, and electrical topology coefficient between the two electric propulsion units are weighted and summed. The summation result is determined as the edge weight of the corresponding edge in the graph structure. The spatial distance coefficient represents the geometric distance between the two electric propulsion units on the aircraft structure, the aerodynamic interference coefficient represents the degree of aerodynamic influence of the rotor airflow of one electric propulsion unit on the other electric propulsion unit, and the electrical topology coefficient represents the degree of power supply correlation between the two electric propulsion units in the power distribution network.

[0007] In some implementations, it also includes: Based on the flight phase data at the current time step, determine the phase adjustment coefficient; Multiply the aerodynamic interference coefficient by the stage adjustment coefficient to obtain the corrected aerodynamic interference coefficient; The spatial distance coefficient between the two electric propulsion units, the corrected aerodynamic interference coefficient, and the electrical topology coefficient are weighted and summed to obtain the corrected edge weights, and the edge weights of the corresponding edges in the graph structure are updated to the corrected edge weights.

[0008] In some implementations, the motor operation data, the battery management data, and the flight phase data are encoded and fused to determine the fused encoded data sequence for each electric propulsion unit, including: The motor operation data, the battery management data, and the flight phase data are respectively subjected to time-series encoding processing to generate motor feature sequences, battery feature sequences, and phase feature sequences; The motor feature sequence, the battery feature sequence, and the stage feature sequence are subjected to cross-modal fusion processing to generate a fused coded data sequence for each electric propulsion unit.

[0009] In some embodiments, the step of performing time-series encoding processing on the motor operation data, the battery management data, and the flight phase data to generate motor feature sequences, battery feature sequences, and phase feature sequences includes: The motor operation data is subjected to temporal convolution compression processing to generate a motor feature sequence; The battery management data is subjected to time interpolation to generate a battery feature sequence; The flight phase data is embedded and encoded, and time-series filling is performed to generate a phase feature sequence. The motor feature sequence, the battery feature sequence, and the phase feature sequence have the same number of time steps, and the fused encoded data at the same index position correspond to the same timestamp.

[0010] In some implementations, determining the remaining lifetime prediction result for each electric propulsion unit based on the health status data includes: For each electric propulsion unit, extract the health status data corresponding to the last time step from the health status data of the electric propulsion unit at each time step, wherein the last time step is the time step with the largest timestamp among all time steps; For each electric propulsion unit, the health status data is input into the remaining lifetime prediction network corresponding to the electric propulsion unit, and the remaining lifetime prediction result of the electric propulsion unit is output.

[0011] Secondly, this application proposes a degradation identification device for an eVTOL electric propulsion system, wherein the eVTOL electric propulsion system includes multiple electric propulsion units, and the device includes: The data acquisition module is used to acquire motor operation data and battery management data of each electric propulsion unit on the electric vertical takeoff and landing (eVTOL) aircraft, as well as flight phase data of the eVTOL. The feature generation module is used to perform encoding and fusion processing on the motor operation data, the battery management data and the flight phase data to determine the fusion encoded data sequence of each electric propulsion unit. The fusion encoded data sequence includes the fusion encoded data of each electric propulsion unit at each time step. The state construction module is used to extract features from the fused encoded data sequence of each electric propulsion unit based on a graph structure and through a neural network model, and generate health status data of each electric propulsion unit at each time step. The lifespan prediction module is used to determine the remaining lifespan prediction result for each electric propulsion unit based on the health status data.

[0012] Thirdly, this application proposes an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the degradation identification method for the eVTOL electric propulsion system described in any of the first aspects.

[0013] Fourthly, this application also proposes a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the degradation identification method for the eVTOL electric propulsion system of any one of the first aspects.

[0014] In summary, the degradation identification method for eVTOL electric propulsion systems provided in this application acquires motor operation data, battery management data, and eVTOL flight phase data for each electric propulsion unit. By encoding and fusing these three types of data, the fused encoded data for each electric propulsion unit at each time step is determined, unifying multi-source data with different sampling frequencies to the same temporal resolution. Based on a graph structure, features are extracted from the fused encoded data sequences of each unit to generate health status data for each unit at each time step. This graph structure explicitly models the coupling relationships between units, such as spatial distance, aerodynamic interference, and electrical connections, enabling the feature extraction process to simultaneously capture the long-range dependence of degradation in the time dimension and the mutual influence relationships between units in the spatial dimension, resulting in a health status representation that integrates spatiotemporal information. Based on this health status data, the remaining lifetime prediction result for each electric propulsion unit is determined, quantifying the lifetime uncertainty of a single unit. Since the health status data incorporates multi-source heterogeneous information and the coupling degradation relationships between units, the prediction result not only reflects the degradation trend of the unit itself but also indirectly reflects the influence of other units on the remaining lifetime of this unit, improving the accuracy of the remaining lifetime prediction. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic flowchart of a degradation identification method for an eVTOL electric propulsion system provided in this application embodiment; Figure 2 A schematic diagram of a degradation identification device for an eVTOL electric propulsion system provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a degradation identification electronic device for an eVTOL electric propulsion system provided in an embodiment of this application. Detailed Implementation

[0016] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to those explicitly listed, but may also include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0017] This application relates to a degradation identification method for eVTOL electric propulsion systems, primarily applicable to distributed electric propulsion systems in electric vertical take-off and landing (eVTOL) aircraft. In this system, multiple electric propulsion units are coupled together through electrical topology, aerodynamic layout, and heat transfer paths. Changes in the degradation state of one unit can affect the degradation process of other units due to load redistribution or environmental interactions. However, existing degradation prediction methods typically treat each electric propulsion unit as an independent object, performing state estimation and lifetime prediction solely based on the unit's own operational data, neglecting the coupling degradation relationships between units. Therefore, this application proposes a degradation identification method that simultaneously models the temporal characteristics of unit degradation and the spatial coupling relationships between units, applicable to maintenance decision-making and health management scenarios for eVTOL aircraft.

[0018] To clearly describe the technical solution of this application, the meanings of specific terms in this application will be explained below.

[0019] The electric propulsion unit in this application refers to the smallest independent functional unit that provides power in an eVTOL aircraft. Each unit includes a drive motor and its associated power battery pack, and multiple units are arranged in a distributed manner around the aircraft fuselage.

[0020] The motor operating data in this application refers to high-frequency sensing data collected from the drive motor side, including but not limited to current waveforms, vibration spectra, and temperature sequences. The sampling frequency is usually not less than 100Hz, and it is used to reflect the real-time operating conditions and early signs of degradation of the motor.

[0021] The battery management data in this application refers to low-frequency state data obtained from the battery management system, including health state estimates, state of charge values, and individual cell voltage ranges. The sampling interval is typically on the order of seconds or minutes, and is used to characterize the long-term degradation trend of the battery.

[0022] The flight phase data in this application refers to discrete labels describing the current flight mission phase of eVTOL, including ground standby, vertical takeoff, transition flight, cruise flight, and vertical landing, which are used to provide operational context information.

[0023] The fused coded data sequence in this application refers to the data vector of a unified dimension obtained at each time step after the original multi-source data (motor operation data, battery management data, flight phase data) has been coded and fused, and the sequence is formed by arranging them in chronological order.

[0024] The graph structure in this application is a data model that abstracts multiple electric propulsion units into a graph. Each electric propulsion unit corresponds to a node in the graph. The coupling relationships between units, such as spatial distance, aerodynamic interference, and electrical connection, correspond to the edges between nodes. Each edge is assigned a weight value, and the larger the weight, the stronger the mutual influence between the two units.

[0025] The spatial distance coefficient of this application characterizes the geometric distance between two electric propulsion units on the aircraft structure; the smaller the distance, the stronger the coupling. The aerodynamic interference coefficient characterizes the degree of aerodynamic influence of the rotor airflow of one electric propulsion unit on the other electric propulsion unit, and is usually obtained through computational fluid dynamics simulation or wind tunnel test. The electrical topology coefficient characterizes the degree of power supply association between the two electric propulsion units in the power distribution network; units powered by the same bus have a higher coefficient.

[0026] The phase adjustment coefficient in this application is a scalar preset based on the current flight phase (such as hovering, transition flight, cruise) and is used to correct the aerodynamic interference coefficient to reflect the dynamic changes in coupling strength under different operating conditions.

[0027] The temporal convolutional compression in this application refers to downsampling and feature extraction of high-frequency motor data using a temporal convolutional network to align the output sequence with the temporal resolution of low-frequency data; temporal interpolation refers to linear or spline interpolation of low-frequency battery management data to improve the sampling rate; embedding encoding refers to mapping discrete flight phase labels into vectors; temporal padding refers to forward padding or linear interpolation of the embedded phase vectors according to the target temporal resolution to make them have the same number of time steps as other sequences.

[0028] The temporal attention mechanism of this application is used to calculate the dependencies between data at different historical time steps of the same electric propulsion unit, thereby capturing the long-range temporal characteristics of the degradation process; the graph attention mechanism relies on the graph structure to calculate the spatial dependencies between different electric propulsion units at the same time step, realizing the weighted aggregation of information from adjacent units.

[0029] The health status data in this application refers to the deep representation of the degradation trend of the electric propulsion unit itself and its influence by the coupling of adjacent units after feature extraction of the fused encoded data sequence by a neural network model (e.g., a spatiotemporal graph attention network containing time attention mechanism and graph attention mechanism).

[0030] The remaining lifetime prediction result of this application refers to the remaining operating time of the electric propulsion unit from the current moment until it reaches the failure threshold, which is predicted by regression network based on health status data. It is usually output in the form of a probability distribution.

[0031] Please see Figure 1 This is a schematic flowchart of a degradation identification method for an eVTOL electric propulsion system provided in an embodiment of this application. The eVTOL electric propulsion system includes multiple electric propulsion units, specifically including: S110. Acquire motor operation data and battery management data of each electric propulsion unit on the electric vertical takeoff and landing (eVTOL) aircraft, as well as flight phase data of the eVTOL. For example, motor operation data is high-frequency sensor data collected from the drive motor side, reflecting the motor's real-time dynamic operating conditions; battery management data is low-frequency state data obtained from the battery management system, used to characterize the battery's long-term health trend; and flight phase data are discrete labels describing the flight mission phases.

[0032] S120. Encode and fuse the motor operation data, battery management data and flight phase data to determine the fused coded data sequence for each electric propulsion unit. The fused coded data sequence includes the fused coded data corresponding to each electric propulsion unit at each time step. For example, the acquired motor operation data, battery management data, and flight phase data are encoded and fused to unify these original data with different sampling frequencies and physical meanings to the same time resolution, and generate fused encoded data for each electric propulsion unit at each time step. The fused encoded data sequence is obtained by arranging the data in chronological order, and this sequence is used as the input for feature extraction.

[0033] S130. Based on the graph structure, feature extraction is performed on the fused encoded data sequence of each electric propulsion unit through a neural network model to generate health status data of each electric propulsion unit at each time step. For example, based on a pre-built graph structure, which uses each electric propulsion unit as a node and the coupling relationships between units such as spatial distance, aerodynamic interference, and electrical connection as edge weights, a neural network model with time attention and graph attention mechanisms is used to simultaneously extract features in the time and spatial dimensions of the fused encoded data sequence of each electric propulsion unit, and output the health status data of each electric propulsion unit at each time step. This health status data includes both the degradation trend information of the unit itself and the influence transmitted by adjacent units through coupling relationships.

[0034] S140. Based on health status data, determine the predicted remaining lifetime of each electric propulsion unit.

[0035] For example, based on the health status data of each electric propulsion unit at each time step generated in step S130, the health status data of each electric propulsion unit at the last time step is extracted, and the data is input into the corresponding remaining lifetime prediction network to output the remaining lifetime prediction result of each electric propulsion unit, which is expressed as the expected value or probability distribution of the remaining lifetime.

[0036] In summary, the degradation identification method for eVTOL electric propulsion systems provided in this application acquires motor operation data, battery management data, and eVTOL flight phase data for each electric propulsion unit. By encoding and fusing these three types of data, the fused encoded data for each electric propulsion unit at each time step is determined, unifying multi-source data with different sampling frequencies to the same temporal resolution. Based on a graph structure, feature extraction is performed on the fused encoded data sequences of each unit to generate health status data for each unit at each time step. This graph structure explicitly models the coupling relationships between units, such as spatial distance, aerodynamic interference, and electrical connections, enabling the feature extraction process to simultaneously capture the long-range dependence of degradation in the time dimension and the mutual influence between units in the spatial dimension, resulting in a health status representation that integrates spatiotemporal information. Based on this health status data, the remaining lifetime prediction result for each electric propulsion unit is determined, quantifying the lifetime uncertainty of a single unit. Since the health status data incorporates multi-source heterogeneous information and the coupling degradation relationships between units, the prediction result not only reflects the degradation trend of the unit itself but also indirectly reflects the influence of other units on the remaining lifetime of this unit, improving the accuracy of the remaining lifetime prediction.

[0037] In some instances, motor operation data, battery management data, and flight phase data are encoded and fused to determine the fused coded data sequence for each electric propulsion unit, including: The motor operation data, battery management data, and flight phase data are time-series encoded to generate motor feature sequences, battery feature sequences, and phase feature sequences. Cross-modal fusion processing is performed on the motor feature sequence, battery feature sequence, and stage feature sequence to generate a fused coded data sequence for each electric propulsion unit, including: Temporal convolution compression is performed on the motor operation data to generate motor feature sequences; Perform time interpolation on battery management data to generate battery feature sequences; The flight phase data is embedded and encoded, and time-series filling is performed to generate phase feature sequences. The motor feature sequence, battery feature sequence, and phase feature sequence have the same number of time steps, and the fused encoded data at the same index position correspond to the same timestamp.

[0038] For example, temporal convolutional compression is performed on motor operating data. The motor operating data is a high-frequency sampling sequence containing instantaneous information such as current waveforms and vibration spectra, with sampling frequencies typically exceeding 100Hz. Temporal convolutional compression employs a multi-layer dilated convolutional network, progressively expanding the receptive field to compress high-frequency data within each fixed time window into low-dimensional feature data, reducing temporal resolution while preserving key degradation information. This compression process outputs a motor feature sequence, where each time step corresponds to a single motor feature data point. Each motor feature data point is a numerical array of a preset dimension, with its time step size aligned with the target resolution of subsequent processing.

[0039] Time interpolation is performed on the battery management data. This data, sourced from the battery management system, includes health status estimates and state of charge (SOC) values. Its sampling interval is typically on the order of seconds or minutes, resulting in a significantly lower temporal resolution than the motor operating data. The time interpolation process uses a linear interpolation method, inserting intermediate time steps between adjacent sampling points to upscale the original low-frequency sequence to the same temporal resolution as the motor feature sequence. The resulting battery feature sequence outputs a sequence where each time step corresponds to a single battery feature data point. Each battery feature data point is a numerical array of predefined dimensions, with timestamps at the same index position consistent with the motor feature sequence.

[0040] Embedding encoding and temporal padding are performed on the flight phase data. The flight phase data consists of discrete labels, including mission phase identifiers such as ground standby, vertical takeoff, transition flight, cruise flight, and vertical landing. Embedding encoding maps each discrete label to embedded data of a fixed dimension, with the dimension of the embedded data preset according to the model design. Since the discrete labels only change at phase transition times, temporal padding pads each embedded data forward to all time steps during the duration of that phase, ensuring that each time step has corresponding phase feature data. The output after padding is a phase feature sequence, where each time step corresponds to one phase feature data point. Each phase feature data point is a numerical array of a preset dimension, and this sequence has the same number of time steps as the motor feature sequence and battery feature sequence.

[0041] Cross-modal fusion processing is performed on three feature sequences. This process takes the motor feature data, battery feature data, and stage feature data of the same electric propulsion unit at the same time step as input. A cross-modal attention mechanism is used to calculate the attention weights among these three data points. The three data points are then weighted and summed according to their weights to obtain the fused encoded data for that time step. The fused encoded data is a numerical array of a pre-defined dimension, determined by the output mapping layer in the cross-modal attention mechanism. The fused encoded data from all time steps are arranged chronologically to form a fused encoded data sequence for each electric propulsion unit. Each element in the sequence corresponds to the fused encoded data of one time step, and elements at the same index position correspond to the same timestamp, achieving a unified representation of multi-source heterogeneous data in both the temporal and feature dimensions.

[0042] In summary, this application's embodiments reduce the dimensionality of the massive high-frequency signal to a time resolution matching that of the low-frequency data by performing temporal convolutional compression on the high-frequency motor operation data, thus preserving early degradation information in the motor operation data. Temporal interpolation is performed on the low-frequency battery management data to fill sampling gaps, ensuring that each time step has an estimate of the battery health status and avoiding data loss in temporal modeling. Embedded encoding and temporal padding are applied to flight phase labels, converting discrete operating conditions into continuous feature data and covering all time steps with padding, enabling the model to perceive flight phase information at each time point. A cross-modal attention mechanism is used to fuse the three types of data, adjusting the weights of motor data, battery data, and phase data in the fusion process, so that the fused encoded data sequence highlights the information source of the current time step according to the actual data distribution.

[0043] It should be noted that the fused encoded data generated by each electric propulsion unit at each time step is a numerical array with a preset dimension. This dimension is determined by the number of channels in the output layer of the cross-modal attention mechanism. This output layer is a fully connected layer with a number of neurons set to a positive integer, such as 64 or 128 dimensions. Before cross-modal fusion processing, temporal convolutional compression compresses the motor operation data into motor feature data of the same dimension, temporal interpolation interpolates the battery management data into battery feature data of the same dimension, and embedding encoding encodes the flight phase data into phase feature data of the same dimension. That is, the dimensions of the three types of feature data are consistent with this preset dimension. The cross-modal attention mechanism performs weighted fusion of the three feature data of the same dimension at the same time step, outputting fused encoded data of the same dimension. Therefore, the fused encoded data sequence of each electric propulsion unit consists of fused encoded data from multiple time steps, and each fused encoded data is a numerical array of this fixed dimension. This dimension matches the input layer dimension of the neural network model, ensuring that the fused encoded data sequence serves as the input to the neural network model.

[0044] In some instances, before generating health status data for each electric propulsion unit at each time step by extracting features from the fused encoded data sequences of each electric propulsion unit using a neural network model based on a graph structure, the method further includes: Using each electric propulsion unit as a node, edges are constructed between the electric propulsion units to obtain a graph structure; For any two nodes in the graph structure, the spatial distance coefficient, aerodynamic interference coefficient, and electrical topology coefficient between the two electric propulsion units are weighted and summed. The summation result is determined as the edge weight of the corresponding edge in the graph structure. Among them, the spatial distance coefficient represents the geometric distance between the two electric propulsion units on the aircraft structure, the aerodynamic interference coefficient represents the degree of aerodynamic influence of the rotor airflow of one electric propulsion unit on the other electric propulsion unit, and the electrical topology coefficient represents the degree of power supply association between the two electric propulsion units in the power distribution network.

[0045] For example, a graph structure is constructed before feature extraction based on the graph structure. Each electric propulsion unit on the aircraft is taken as a node. Since the electric propulsion units are distributed around the fuselage, there may be mutual influences between any two units in terms of spatial position, aerodynamic airflow, and electrical connection. Therefore, an edge is constructed between each electric propulsion unit, that is, the initial graph structure is constructed in a fully connected manner, with each edge connecting two nodes, to obtain the basic graph structure.

[0046] For any two nodes in the graph structure corresponding to electric propulsion units, determine the spatial distance coefficient between the two units. The spatial distance coefficient is calculated based on the three-dimensional structural model of the aircraft, specifically the Euclidean distance between the two electric propulsion units at their installation positions on the fuselage. The closer the distance, the larger the coefficient value, and vice versa. This coefficient is used to quantify the physical coupling strength between the two units due to their spatial proximity, such as heat transfer and vibration transmission.

[0047] For two identical electric propulsion units, determine their aerodynamic interference coefficient and electrical topology coefficient. The aerodynamic interference coefficient, obtained based on computational fluid dynamics simulations or wind tunnel test data, characterizes the aerodynamic influence of the airflow generated by the rotor rotation of one electric propulsion unit on the other. The aerodynamic interference coefficient is higher when the downstream unit is located in the wake region of the upstream unit. The electrical topology coefficient is determined based on the schematic diagram of the aircraft's power distribution network. If both electric propulsion units are powered by the same bus, the electrical topology coefficient is higher; otherwise, it is lower or zero. This coefficient reflects the correlation between the two units in terms of load distribution and voltage fluctuations.

[0048] It should be noted that computational fluid dynamics (CFD) simulation or wind tunnel testing methods are used to determine the aerodynamic interference coefficient. Specifically, a three-dimensional geometric model of the aircraft is established, including the position, rotational speed, and blade parameters of each rotor. The airflow field distribution around the aircraft is solved using CFD simulation to obtain the airflow velocity vector at the location of each electric propulsion unit. For any two electric propulsion units, the deflection angle and velocity attenuation ratio of the airflow from the rotor wake of the upstream unit on the downstream unit are defined; this ratio serves as the initial value for the aerodynamic interference coefficient. If wind tunnel testing is used, a scaled-down model of the aircraft is placed in a wind tunnel, and the changes in airflow velocity and direction at the downstream unit are measured under different combinations of rotor rotational speeds. Data is collected using a sensor array, and the relationship curve between interference intensity and unit spacing and rotor rotational speed is fitted. The aerodynamic interference coefficient obtained using the above method is a value between 0 and 1; a larger value indicates a more significant aerodynamic influence of the upstream unit on the downstream unit. This coefficient can be pre-calculated and stored as a lookup table, which is retrieved during aircraft operation based on the current rotor rotational speed and flight stage.

[0049] When determining the electrical topology coefficient, the schematic diagram of the aircraft's power distribution network is analyzed. The electrical topology coefficient reflects the degree of electrical connection between the two electric propulsion units in the power supply circuit. Specifically, it is determined by obtaining a single-line diagram of the power distribution network and identifying the bus nodes, distribution branches, and shared protection devices connected to each electric propulsion unit. If the two electric propulsion units are connected to the same DC bus and there is no isolation diode or contactor disconnection between them, the electrical topology coefficient is set to a first preset value, such as 1.0; if the two units are connected to different buses but interconnected via a bidirectional DC-DC converter or tie switch, the electrical topology coefficient is set to a second preset value, such as 0.5; if the two units have no direct electrical connection path in the power distribution network, the electrical topology coefficient is set to a third preset value, such as 0.1. The electrical topology coefficient is a dimensionless value used to characterize the load coupling and voltage coupling strength between the two units due to electrical connection. This coefficient can be determined during the aircraft's electrical architecture design phase.

[0050] After obtaining the three coefficients mentioned above, the spatial distance coefficient, aerodynamic interference coefficient, and electrical topology coefficient are multiplied by their respective preset weighting factors. The three products are then summed, and the sum is used as the edge weight between the two nodes. During the weighted summation process, the magnitudes of the three weighting factors are preset based on aircraft design parameters and operational experience. For example, the weighting factor for the spatial distance coefficient can be set to 0.3, the weighting factor for the aerodynamic interference coefficient to 0.5, and the weighting factor for the electrical topology coefficient to 0.2. The specific values ​​are adjusted according to the actual application scenario. This weighted summation operation is performed on each pair of nodes in the graph structure, assigning a corresponding edge weight value to each edge.

[0051] In summary, this embodiment uses each electric propulsion unit as a node and constructs edges between all units, forming a fully connected graph structure. This allows for feature extraction of the coupling relationship between any two electric propulsion units, avoiding the omission of weak coupling relationships due to preset connection thresholds. By weighted summation of spatial distance coefficients, aerodynamic interference coefficients, and electrical topology coefficients, different coupling factors are unified to the same numerical scale. Each coefficient represents a specific mutual influence mechanism, facilitating adjustments based on the actual aircraft configuration. The edge weights are determined through weighted summation, preserving the contribution proportions of various coupling factors while controlling the importance of different coupling factors through weighting factors, making the graph structure adaptable to the design characteristics of different aircraft models. The graph structure with edge weights serves as the basis for spatial dependencies in the neural network model, ensuring that each node selectively aggregates information from adjacent nodes based on actual coupling strength during feature extraction, improving the accuracy of spatial feature aggregation.

[0052] In some instances, it also includes: Determine the phase adjustment coefficient based on the flight phase data at the current time step; Multiply the aerodynamic interference coefficient by the stage adjustment coefficient to obtain the corrected aerodynamic interference coefficient; The spatial distance coefficient, the corrected aerodynamic interference coefficient, and the electrical topology coefficient between the two electric propulsion units are weighted and summed to obtain the corrected edge weights. The edge weights of the corresponding edges in the graph structure are then updated to the corrected edge weights.

[0053] For example, during actual aircraft operation, the aerodynamic interference intensity among the electric propulsion units varies significantly across different flight phases, necessitating correction of the aerodynamic interference coefficient. Specifically, during the hovering phase, each rotor generates a vertically downward airflow that interacts with the fuselage and wings, resulting in aerodynamic interference at a baseline level. During the transition phase, the rotors tilt from a vertical to a horizontal attitude, and the airflow beneath them is influenced by the forward flow, forming complex vortices. The wakes of each rotor interact, leading to significantly higher aerodynamic interference intensity than during hovering. During the cruise phase, lift is primarily provided by the wings, and the rotors are unloaded, resulting in relatively less airflow interference. To reflect this phase-dependent variation, a phase adjustment coefficient is introduced.

[0054] Based on the flight phase data acquired at the current time step, the phase adjustment coefficient corresponding to that flight phase is determined. The phase adjustment coefficient is a preset scalar value, with each flight phase corresponding to an independent adjustment coefficient. The adjustment coefficient for the hovering phase is set to 1.0, the adjustment coefficient for the transition phase is set to a value greater than 1.0, and the adjustment coefficient for the cruise phase is set to a value less than 1.0. Vertical takeoff and vertical landing phases are set with corresponding values ​​based on actual airflow characteristics. These preset values ​​can be pre-calibrated using computational fluid dynamics simulations or flight test data and stored in configuration files of the aircraft's onboard systems or ground maintenance systems. During runtime, the corresponding phase adjustment coefficient is read from the configuration file based on the flight phase label of the current time step.

[0055] The corrected aerodynamic interference coefficient is obtained by multiplying the initial aerodynamic interference coefficient by the stage adjustment coefficient. The initial aerodynamic interference coefficient is a static value determined based on simulation or wind tunnel tests during the aircraft design phase. After multiplying by the stage adjustment coefficient, the corrected aerodynamic interference coefficient is amplified during the transition phase, reduced during the cruise phase, and remains unchanged during the hovering phase. This achieves the adjustment of the aerodynamic interference coefficient with the flight phase, ensuring that the side weight calculation accurately reflects the actual coupling strength under the current operating conditions.

[0056] The spatial distance coefficient between the two electric propulsion units, the corrected aerodynamic interference coefficient, and the electrical topology coefficient are weighted and summed to obtain the corrected edge weights. The edge weights of the corresponding edges in the graph structure are then updated to the corrected edge weights. The weighting factors used in the weighted summation are consistent with those used in the initial edge weight calculation. The corrected edge weights vary with the flight phase: the baseline weights are used during hovering, they increase during the transition phase due to enhanced aerodynamic interference, and they decrease during the cruise phase due to weakened aerodynamic interference. This allows the graph structure to adapt to the changing coupling characteristics of the aircraft under different operating conditions, improving the accuracy of feature extraction.

[0057] In summary, this application's embodiments introduce a phase adjustment coefficient, allowing the aerodynamic interference coefficient to adjust with changes in flight phase, thus avoiding the limitations of static aerodynamic interference coefficients that cannot adapt to different flight conditions. The corrected aerodynamic interference coefficient is re-weighted and summed with the spatial distance coefficient and electrical topology coefficient, ensuring that the edge weights reflect the influence of the three types of coupling factors under different flight phases. At each time step, the edge weights of the graph structure are updated based on flight phase data, enabling the neural network model to use a graph structure that matches the current operating condition when extracting features, improving the accuracy of spatial feature aggregation, especially in scenarios with frequent flight phase transitions, allowing for timely adjustment of coupling relationship modeling.

[0058] In some instances, the neural network model includes temporal attention and graph attention mechanisms. Based on the graph structure, the neural network model extracts features from the fused encoded data sequences of each electric propulsion unit to generate health status data for each electric propulsion unit at each time step, including: For each electric propulsion unit, the temporal dependency between the fused coded data at each time step in the fused coded data sequence of the electric propulsion unit is calculated through a time attention mechanism to determine the temporal enhancement feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the spatial dependency between the temporal augmentation feature data of the electric propulsion unit at the current time step and the temporal augmentation feature data of other electric propulsion units in the graph structure at the same time step is calculated through the graph attention mechanism, thereby generating the spatial aggregated feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the temporal enhanced feature data and spatial aggregated feature data are fused to generate health status data for each electric propulsion unit at each time step.

[0059] For example, after constructing the fused encoded data sequence, the neural network model extracts spatiotemporal features from the data of each electric propulsion unit. The neural network model includes a temporal attention mechanism and a graph attention mechanism. For each electric propulsion unit, the fused encoded data sequence is input into the temporal attention mechanism. The temporal attention mechanism calculates the attention weights between the current time step and each historical time step for the unit's own time series data. Specifically, for each time step, the temporal attention mechanism generates a query vector and uses the fused encoded data of historical time steps as key-value pairs. It calculates the similarity score between the current time step and each historical time step using a dot product or additive model, and then converts it into attention weights using a normalized exponential function. The attention weights are then weighted and summed with the fused encoded data of historical time steps to obtain the temporal enhancement feature data of the current time step. This process ensures that the output of the current time step not only depends on the current input but also selectively integrates information related to the current degradation state from historical time steps, thereby discovering the long-range dependencies of the electric propulsion unit's own degradation process in the temporal dimension.

[0060] After obtaining the temporal augmentation feature data of each electric propulsion unit at each time step, the neural network model models the spatial dependencies between different electric propulsion units within the same time step. For each time step, the temporal augmentation feature data of all electric propulsion units at that time step are used as input to the graph attention mechanism. The graph attention mechanism relies on a pre-constructed graph structure in which each node represents an electric propulsion unit, and the edge weights between nodes are determined by factors such as spatial distance, aerodynamic disturbances, and electrical connections. For a target node at the current time step, the graph attention mechanism calculates the attention coefficient between the temporal augmentation feature data of that node and the temporal augmentation feature data of all neighboring nodes in the graph structure.

[0061] The attention coefficient is calculated as follows: First, the feature data of the target node and its neighboring nodes are linearly transformed by a learnable weight matrix. Then, the two transformed vectors are concatenated or subjected to a dot product to obtain the feature similarity score. Next, the similarity score is multiplied by the edge weight of the corresponding edge in the graph structure to obtain the weighted similarity. Finally, the normalized exponential function is used to normalize all neighboring nodes of the same node to obtain the attention weight of each neighboring node to the target node.

[0062] The attention weight is weighted and summed with the temporal enhancement feature data of neighboring nodes to obtain the spatial aggregated feature data of the target node at the current time step. This spatial aggregated feature data integrates information from neighboring nodes, and the fusion strength is determined by the edge weights of the graph structure and the feature similarity between nodes, thus realizing the modeling of the coupling relationship between units.

[0063] For each electric propulsion unit, the temporal enhancement feature data and spatial aggregation feature data are fused to generate health status data for each electric propulsion unit at each time step, using the following implementation method: In one implementation, the neural network model employs a single-layer spatiotemporal feature extraction structure. For each electric propulsion unit, a temporal attention mechanism is used to process the fused encoded data sequence of that unit, calculating the attention weights between the current time step and each historical time step within a preset time window. The temporal enhancement feature data of that unit at each time step is obtained through weighted summation. For each time step, the temporal enhancement feature data of all electric propulsion units at the same time step are used as input. A graph attention mechanism is used, relying on a pre-constructed graph structure, to calculate the attention coefficients between each unit and its neighboring units. The temporal enhancement feature data of neighboring units are weighted and aggregated based on these attention coefficients to obtain the spatial aggregated feature data of that unit at that time step. The temporal enhancement feature data and spatial aggregated feature data of the same unit at the same time step are then fused. The fusion method employs feature concatenation, feature addition, or gated fusion mechanisms. The fused data serves as the health status data of that unit at that time step.

[0064] In another implementation, the neural network model employs a multi-layer spatiotemporal feature extraction structure, wherein the temporal attention mechanism is executed only once, and the graph attention mechanism is stacked in multiple layers, as detailed below: A temporal attention mechanism is used to process the fused coded data sequence of each electric propulsion unit. Attention weights between the current time step and each historical time step within a preset time window are calculated, and weighted summation is performed to obtain the temporal augmentation feature data of each unit at each time step. This temporal augmentation feature data serves as the input to a multi-layer graph attention mechanism. For each time step, the temporal augmentation feature data of all electric propulsion units at that time step are used as the input to the first graph attention layer. In the first graph attention layer, attention coefficients between each unit and all neighboring units are calculated based on a pre-constructed graph structure. The temporal augmentation feature data of neighboring units are then weighted and aggregated according to these attention coefficients, outputting the first-layer spatial aggregated feature data for each unit.

[0065] The spatial aggregation feature data of the first layer is used as input to the second graph attention layer. In the second graph attention layer, the attention coefficient between each unit and its neighboring units is calculated again using the same graph structure. At this point, the input features of each unit already contain the first-order neighbor information after aggregation in the first layer. Therefore, in the spatial aggregation feature data of the second layer output after the second graph attention calculation, each unit indirectly incorporates the information of neighboring units with a distance of two hops. This process is repeated, stacking a preset number of graph attention layers as needed. With each additional layer, the neighbor range perceived by each unit expands outward by one hop.

[0066] After stacking multiple layers of graph attention, the spatial aggregation feature data output from the last graph attention layer is fused with the temporal enhancement feature data output from the temporal attention mechanism. The fusion method employs feature concatenation, feature addition, or gated fusion mechanisms. The fused data serves as the health status data for that unit at that time step. This health status data includes both the unit's own temporal degradation information and the coupling effects of all units within the multi-hop neighborhood transmitted through the multi-layer graph attention, thereby capturing the indirect coupling relationships between distant units in the system.

[0067] It should be noted that the above spatiotemporal feature extraction process can be implemented using a Spatio-Temporal Graph Attention Network (ST-GAT) as the basic model architecture. This model consists of multiple stacked spatiotemporal attention modules, each containing a temporal attention layer and a graph attention layer. The temporal attention layer employs a multi-head attention mechanism, performing self-attention calculation on the fused encoded data sequence of each electric propulsion unit in the temporal dimension. It maps the input sequence to three different representation spaces using learnable query, key, and value matrices, then calculates the association weights between each pair of time steps, outputting temporally enhanced feature data. The graph attention layer takes the temporally enhanced feature data of all units at the same time step as input, using the edge weights in the graph structure as bias terms in the calculation of attention coefficients. It first calculates the feature similarity between nodes, then multiplies it by the edge weights to obtain a weighted similarity. After normalization, it aggregates the features of adjacent nodes, outputting spatially aggregated feature data.

[0068] Residual connections and layer normalization can be added between the temporal attention layer and the graph attention layer to enhance the training process. The entire spatiotemporal graph attention network can be stacked with 3 to 6 spatiotemporal attention modules, with the feature dimension of the output of each layer remaining consistent. The health status data output from the last layer is input into the remaining lifespan prediction network. This model learns spatiotemporal dependencies end-to-end, and the multi-head attention mechanism supports parallel computation, resulting in higher training efficiency compared to recurrent neural networks. It is suitable for degradation identification tasks in multi-unit, long-term sequences of eVTOL electric propulsion systems.

[0069] The above model structure alternately stacks temporal attention layers and graph attention layers, with each module executing temporal attention first and then graph attention. The first layer of temporal attention extracts primary temporal features, while the first layer of graph attention aggregates adjacent spatial information. The second layer of temporal attention extracts temporal dependencies again from features already containing first-order neighbor information, enabling the model to capture the influence of adjacent historical states on the current node. The second layer of graph attention further expands the spatial scope. This architecture is designed because, in certain coupled scenarios, the degradation history of adjacent units can have a delayed impact on the current unit (e.g., if an adjacent unit experienced an overload some time ago, its thermal effects only affect the current unit after being slowly transmitted through the cooling system). This necessitates re-modeling the temporal process after each layer of graph attention. The above model structure is suitable for systems with complex spatiotemporal coupling, but the computational overhead increases accordingly.

[0070] In summary, this application's embodiments extract the long-range dependencies of each electric propulsion unit's degradation process over time using a temporal attention mechanism, improving the accuracy of temporal feature extraction. By using a graph attention mechanism to spatially aggregate features from different units within the same time step using a graph structure, the health status data of each unit is fused with coupling information from adjacent units, achieving modeling of spatial dependencies between units. The health status data is generated by fusing temporal enhancement features with spatial aggregation features, ensuring that this data simultaneously includes the unit's own degradation trajectory and coupling effects. Using a single-layer spatiotemporal extraction structure results in low computational cost and directly captures the coupling relationships between adjacent units, making it suitable for scenarios with high real-time requirements. Using a multi-layer spatiotemporal extraction structure, by stacking multiple graph attention layers, the health status data of each unit is fused layer by layer with multi-hop adjacent information, thereby capturing the indirect effects between indirectly coupled units in the system and improving the accuracy of degradation identification under complex coupled networks.

[0071] In some instances, the remaining lifetime prediction for each electric propulsion unit is determined based on health status data, including: For each electric propulsion unit, extract the health status data corresponding to the last time step from the health status data of the electric propulsion unit at each time step, where the last time step is the time step with the largest timestamp among all time steps; For each electric propulsion unit, health status data is input into the remaining lifetime prediction network corresponding to the electric propulsion unit, and the remaining lifetime prediction result of the electric propulsion unit is output.

[0072] For example, after obtaining the health status data of each electric propulsion unit at each time step, the current state representation is extracted from these time-series data for remaining lifetime prediction. Since the time steps of the health status data sequence are arranged in ascending order of timestamps, the last time step corresponds to the most recent sampling time and contains the latest degradation information. Therefore, for each electric propulsion unit, the health status data corresponding to the last time step is extracted from the unit's health status data sequence as the degradation state representation of that unit at the current moment. This health status data, through the previous encoding fusion and spatiotemporal feature extraction processes, integrates the unit's own temporal degradation trend and the coupling effects of adjacent units, reflecting the current health level.

[0073] The extracted health status data from the last time step is input into the remaining lifetime prediction network corresponding to the electric propulsion unit. The remaining lifetime prediction network can employ a multilayer perceptron structure, containing an input layer, one or more hidden layers, and an output layer. The number of neurons in the input layer matches the dimensionality of the health status data. The hidden layers use non-linear activation functions to enhance the network's expressive power. The output layer is set to one or two neurons depending on the requirements of the prediction task. When the output is a single neuron, the network directly outputs a point estimate of the remaining lifetime, i.e., the remaining operating time from the current moment until the unit reaches the failure threshold, expressed in flight hours or cycles. When the output is two neurons, the network outputs the mean and variance of the remaining lifetime probability distribution, respectively. The mean represents the expected remaining lifetime, and the variance represents the prediction uncertainty. In this case, a negative log-likelihood loss is used for training.

[0074] The remaining lifetime prediction network can be configured in two ways. One approach is to set up an independent remaining lifetime prediction network for each electric propulsion unit, meaning each unit has its own multilayer perceptron parameters. The other approach is for all electric propulsion units to share the same remaining lifetime prediction network. In this case, the network learns the mapping patterns for different units based on the differences in the distribution of data among the units in the input health status data. Using independent networks better adapts to the individual degradation patterns of each unit due to differences in installation location and operating conditions; using a shared network requires fewer parameters and more training data, but has stronger generalization ability. Regardless of the configuration, training is performed using supervised learning, with the actual remaining lifetime recorded in historical operating data as the label. After training, the health status data of the last time step is input into the network, and the remaining lifetime prediction result for that unit is calculated through forward propagation. This result can be used for maintenance early warning, spare parts scheduling, or maintenance decisions.

[0075] After obtaining the remaining lifetime prediction results for each electric propulsion unit, the overall health status of the eVTOL electric propulsion system can be assessed. First, the remaining lifetime prediction results for all units are compared, and the minimum value is taken as the system's shortest remaining lifetime. This value reflects the remaining operating time corresponding to the unit most likely to reach the failure threshold first. Second, the average remaining lifetime prediction results for all units are calculated as the overall average remaining lifetime of the system. Furthermore, the standard deviation or range between the remaining lifetime prediction results of each unit is calculated to assess the balance of degradation within the system. If the differences in remaining lifetime among units are too large, it indicates uneven load distribution or significant coupling effects, and the system faces the risk of a chain reaction caused by localized accelerated degradation. When the remaining lifetime prediction result of at least one unit is lower than a preset emergency threshold, the system is determined to be in a high-risk state, triggering a system-level maintenance warning. When the remaining lifetime prediction results of all units are higher than a preset safety threshold and the difference in remaining lifetime is less than a preset difference threshold, the overall health status of the system is considered good. If the minimum remaining lifetime is higher than the emergency threshold but the difference in remaining lifetime exceeds the difference threshold, the system is determined to be in a medium-risk state, and it is recommended to prioritize checking units with abnormal load distribution and coupling relationships. The comprehensive assessment results can be used to formulate overall aircraft maintenance plans, flight mission scheduling, or load redistribution strategies.

[0076] Please see Figure 2 This is a schematic diagram of a degradation identification device 20 for an eVTOL electric propulsion system provided in an embodiment of this application. The eVTOL electric propulsion system includes multiple electric propulsion units, and the device includes: The data acquisition module 21 is used to acquire motor operation data and battery management data of each electric propulsion unit on the electric vertical takeoff and landing aircraft (eVTOL), as well as flight phase data of the eVTOL. The feature generation module 22 is used to encode and fuse motor operation data, battery management data and flight phase data to determine the fused coded data sequence of each electric propulsion unit. The fused coded data sequence includes the fused coded data of each electric propulsion unit at each time step. The state construction module 23 is used to extract features from the fused encoded data sequence of each electric propulsion unit based on a graph structure and through a neural network model, and generate health status data of each electric propulsion unit at each time step. The life prediction module 24 is used to determine the remaining life prediction result of each electric propulsion unit based on health status data.

[0077] In some embodiments, the feature generation module 22 further includes: The timing coding unit is used to perform timing coding processing on motor operation data, battery management data and flight phase data respectively, to generate motor feature sequence, battery feature sequence and phase feature sequence; The multimodal fusion unit is used to perform cross-modal fusion processing on the motor feature sequence, battery feature sequence and stage feature sequence to generate the fused coded data sequence for each electric propulsion unit.

[0078] In some embodiments, the timing coding unit further includes: The convolution processing subunit is used to perform temporal convolution compression processing on the motor operation data to generate a motor feature sequence. The interpolation processing subunit is used to perform time interpolation processing on battery management data to generate battery feature sequences. The encoding processing subunit is used to perform embedding encoding and time-series padding processing on the flight phase data to generate phase feature sequences. Among them, the motor feature sequence, battery feature sequence and phase feature sequence have the same number of time steps, and the fused encoded data at the same index position correspond to the same timestamp.

[0079] In some embodiments, the state construction module 23 further includes: The temporal feature unit is used to calculate the temporal dependency between the fused coded data at each time step in the fused coded data sequence of each electric propulsion unit through a temporal attention mechanism, and to determine the temporal enhancement feature data of the electric propulsion unit at each time step. Spatial feature units are used to calculate the spatial dependency between the temporal augmentation feature data of each electric propulsion unit at the current time step and the temporal augmentation feature data of other electric propulsion units in the graph structure at the same time step through a graph attention mechanism, thereby generating spatial aggregated feature data of the electric propulsion unit at each time step. The fusion feature unit is used to fuse temporal enhanced feature data and spatial aggregated feature data for each electric propulsion unit to generate health status data for each electric propulsion unit at each time step.

[0080] In some embodiments, the degradation identification device 20 for the eVTOL electric propulsion system further includes: The graph structure construction module is used to construct edges between electric propulsion units, with each electric propulsion unit as a node, to obtain a graph structure. The graph structure weighting module is used to perform a weighted summation of the spatial distance coefficient, aerodynamic interference coefficient, and electrical topology coefficient between any two electric propulsion units corresponding to any two nodes in the graph structure. The summation result is used to determine the edge weight of the corresponding edge in the graph structure. Among them, the spatial distance coefficient represents the geometric distance between the two electric propulsion units on the aircraft structure, the aerodynamic interference coefficient represents the degree of aerodynamic influence of the rotor airflow of one electric propulsion unit on the other electric propulsion unit, and the electrical topology coefficient represents the degree of power supply association between the two electric propulsion units in the power distribution network.

[0081] In some embodiments, the degradation identification device 20 for the eVTOL electric propulsion system further includes: The phase coefficient adjustment module is used to determine the phase adjustment coefficient based on the flight phase data of the current time step. The aerodynamic interference correction module is used to multiply the aerodynamic interference coefficient by the stage adjustment coefficient to obtain the corrected aerodynamic interference coefficient; The edge weight correction module is used to perform a weighted summation of the spatial distance coefficient, the corrected aerodynamic interference coefficient, and the electrical topology coefficient between two electric propulsion units to obtain the corrected edge weight, and update the edge weight of the corresponding edge in the graph structure to the corrected edge weight.

[0082] In some embodiments, the lifetime prediction module 24 further includes: The health status extraction unit is used to extract the health status data corresponding to the last time step from the health status data of each electric propulsion unit at each time step. The last time step is the time step with the largest timestamp among all time steps. The lifetime prediction unit is used to input health status data into the remaining lifetime prediction network corresponding to each electric propulsion unit and output the remaining lifetime prediction result of the electric propulsion unit.

[0083] Please see Figure 3 This application also provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and running on the processor. When the processor 320 executes the computer program 311, it implements the steps of the degradation identification method for the eVTOL electric propulsion system.

[0084] Since the electronic device described in this embodiment is the device used to implement the degradation identification device of the eVTOL electric propulsion system in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in this application embodiment. Therefore, how the electronic device implements the method in this application embodiment will not be described in detail here. Any device used by those skilled in the art to implement the method in this application embodiment is within the scope of protection of this application.

[0085] In practice, when the computer program 311 is executed by the processor, it can implement any of the embodiments corresponding to the first aspect.

[0086] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are loaded and executed by a processor, they implement the steps of the degradation identification method for the eVTOL electric propulsion system described above.

[0087] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0088] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

[0089] Although preferred embodiments have been described in this specification, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications beyond the scope of this specification.

[0090] Obviously, those skilled in the art can make various modifications to this specification without departing from its spirit and scope. Therefore, this specification also intends to include any modifications that fall within the scope of the claims and their equivalents.

Claims

1. A degradation identification method for an eVTOL electric propulsion system, characterized in that, The eVTOL electric propulsion system includes multiple electric propulsion units, and the method includes: Acquire motor operation data and battery management data for each electric propulsion unit on the electric vertical takeoff and landing (eVTOL) aircraft, as well as flight phase data of the eVTOL; The motor operation data, the battery management data, and the flight phase data are encoded and fused to determine the fused coded data sequence for each electric propulsion unit. The fused coded data sequence includes the fused coded data for each electric propulsion unit at each time step. Based on the graph structure, features are extracted from the fused coded data sequence of each electric propulsion unit using a neural network model to generate health status data of each electric propulsion unit at each time step. Based on the health status data, the remaining lifetime prediction result for each electric propulsion unit is determined.

2. The method according to claim 1, characterized in that, The neural network model includes a temporal attention mechanism and a graph attention mechanism. Based on a graph structure, the neural network model extracts features from the fused encoded data sequences of each electric propulsion unit to generate health status data for each electric propulsion unit at each time step, including: For each electric propulsion unit, the temporal dependency between the fused coded data at each time step in the fused coded data sequence of the electric propulsion unit is calculated through the time attention mechanism to determine the temporal enhancement feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the spatial dependency between the temporal augmentation feature data of the electric propulsion unit at the current time step and the temporal augmentation feature data of other electric propulsion units in the graph structure at the same time step is calculated through the graph attention mechanism, thereby generating the spatial aggregated feature data of the electric propulsion unit at each time step; For each electric propulsion unit, the temporal enhancement feature data and the spatial aggregation feature data are fused to generate health status data for each electric propulsion unit at each time step.

3. The method according to claim 1, characterized in that, Before generating health status data for each electric propulsion unit at each time step by extracting features from the fused encoded data sequence of each electric propulsion unit using a neural network model based on a graph structure, the method further includes: Using each electric propulsion unit as a node, edges are constructed between the electric propulsion units to obtain a graph structure; For any two nodes in the graph structure corresponding to electric propulsion units, the spatial distance coefficient, aerodynamic interference coefficient, and electrical topology coefficient between the two electric propulsion units are weighted and summed. The summation result is determined as the edge weight of the corresponding edge in the graph structure. The spatial distance coefficient represents the geometric distance between the two electric propulsion units on the aircraft structure, the aerodynamic interference coefficient represents the degree of aerodynamic influence of the rotor airflow of one electric propulsion unit on the other electric propulsion unit, and the electrical topology coefficient represents the degree of power supply correlation between the two electric propulsion units in the power distribution network.

4. The method according to claim 3, characterized in that, Also includes: Based on the flight phase data at the current time step, determine the phase adjustment coefficient; Multiply the aerodynamic interference coefficient by the stage adjustment coefficient to obtain the corrected aerodynamic interference coefficient; The spatial distance coefficient between the two electric propulsion units, the corrected aerodynamic interference coefficient, and the electrical topology coefficient are weighted and summed to obtain the corrected edge weights, and the edge weights of the corresponding edges in the graph structure are updated to the corrected edge weights.

5. The method according to claim 1, characterized in that, The process of encoding and fusing the motor operation data, the battery management data, and the flight phase data to determine the fused encoded data sequence for each electric propulsion unit includes: The motor operation data, the battery management data, and the flight phase data are respectively subjected to time-series encoding processing to generate motor feature sequences, battery feature sequences, and phase feature sequences; The motor feature sequence, the battery feature sequence, and the stage feature sequence are subjected to cross-modal fusion processing to generate a fused coded data sequence for each electric propulsion unit.

6. The method according to claim 5, characterized in that, The step of performing time-series encoding on the motor operation data, the battery management data, and the flight phase data to generate motor feature sequences, battery feature sequences, and phase feature sequences includes: The motor operation data is subjected to temporal convolution compression processing to generate a motor feature sequence; The battery management data is subjected to time interpolation to generate a battery feature sequence; The flight phase data is embedded and encoded, and time-series filling is performed to generate a phase feature sequence. The motor feature sequence, the battery feature sequence, and the phase feature sequence have the same number of time steps, and the fused encoded data at the same index position correspond to the same timestamp.

7. The method according to claim 1, characterized in that, The determination of the remaining lifetime prediction result for each electric propulsion unit based on the health status data includes: For each electric propulsion unit, extract the health status data corresponding to the last time step from the health status data of the electric propulsion unit at each time step, wherein the last time step is the time step with the largest timestamp among all time steps; For each electric propulsion unit, the health status data is input into the remaining lifetime prediction network corresponding to the electric propulsion unit, and the remaining lifetime prediction result of the electric propulsion unit is output.

8. A degradation identification device for an eVTOL electric propulsion system, characterized in that, The eVTOL electric propulsion system includes multiple electric propulsion units, and the device includes: The data acquisition module is used to acquire motor operation data and battery management data of each electric propulsion unit on the electric vertical takeoff and landing (eVTOL) aircraft, as well as flight phase data of the eVTOL. The feature generation module is used to perform encoding and fusion processing on the motor operation data, the battery management data and the flight phase data to determine the fusion encoded data sequence of each electric propulsion unit. The fusion encoded data sequence includes the fusion encoded data of each electric propulsion unit at each time step. The state construction module is used to extract features from the fused encoded data sequence of each electric propulsion unit based on a graph structure and through a neural network model, and generate health status data of each electric propulsion unit at each time step. The lifespan prediction module is used to determine the remaining lifespan prediction result for each electric propulsion unit based on the health status data.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program stored in the memory, implements the steps of the degradation identification method for the eVTOL electric propulsion system as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the degradation identification method for the eVTOL electric propulsion system as described in any one of claims 1 to 7.

Images