Deep learning-based electric aircraft power system fault prediction method and system

Abstract

The application relates to the technical field of electric aircraft power system fault prediction, and proposes an electric aircraft power system fault prediction method and system based on deep learning, which comprises the following steps: collecting multi-source sensor time series data and performing pretreatment to obtain a fault feature time series matrix and a battery electrochemical parameter time series matrix; based on the fault feature time series matrix, a fault causal prediction network is constructed by using a deep learning algorithm optimized based on FPGA, and a fault causal relationship matrix is output; based on the battery electrochemical parameter time series matrix and the fault causal relationship matrix, multi-scale degradation prediction is performed by using a physically constrained neural network, and a power system comprehensive health index is output; a fault prediction strategy is generated according to the fault causal relationship matrix and the power system comprehensive health index, and predictive maintenance and fault early warning are performed on the electric aircraft power system based on the fault prediction strategy. The application realizes efficient, accurate and collaborative prediction and early warning of power system faults.

Description

Technical Field

[0001] This invention relates to the field of electric aircraft power system fault prediction technology, and in particular to a method and system for electric aircraft power system fault prediction based on deep learning. Background Technology

[0002] Fault prediction for electric aircraft power systems is a technology that integrates multi-source sensor data and electrochemical parameters from the battery-motor-propeller system to achieve early fault identification and predictive maintenance. It has the advantages of high prediction accuracy, low maintenance cost, and strong system reliability. It can improve the operational safety of electric aircraft through real-time monitoring and intelligent analysis, and is the core development direction of intelligent operation and maintenance technology for electric aircraft.

[0003] In existing technologies, when handling fault monitoring and prediction in the power system of electric aircraft, sensor data acquisition and fault prediction analysis operate independently. Sensor monitoring mainly relies on a single data source for anomaly detection, while fault prediction primarily uses traditional threshold judgments for early warning and control. When early signs of battery or motor degradation appear during flight and require early warning, although multi-source sensors can collect abnormal signals such as current, voltage, and temperature, they lack in-depth correlation analysis. Although fault prediction algorithms have learning capabilities, they operate independently without effectively integrating with electrochemical parameters and hardware-in-the-loop test data. This results in delayed power system fault warnings and insufficient prediction accuracy. Sensor anomaly detection and deep learning prediction are executed independently, making it impossible to dynamically adjust the prediction strategy based on the causal relationship of fault propagation. This leads to delayed warnings of critical flight safety risks and poor predictive maintenance effectiveness. Summary of the Invention

[0004] In view of this, this invention proposes a deep learning-based method and system for predicting faults in the power system of electric aircraft. This method addresses the problem that when early signs of degradation in the battery or motor of an electric aircraft are observed during flight and early warning is needed, although multi-source sensors can collect abnormal signals such as current, voltage, and temperature, they lack deep correlation analysis. Although the fault prediction algorithm has learning capabilities, it operates independently without effectively integrating with electrochemical parameters and hardware-in-the-loop test data. This results in delayed warning times and insufficient prediction accuracy for power system faults. Furthermore, sensor anomaly detection and deep learning prediction are executed independently, making it impossible to dynamically adjust the prediction strategy based on the causal relationship of fault propagation. This leads to delayed warnings of critical flight safety risks and poor predictive maintenance performance.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides a method for predicting faults in the power system of an electric aircraft based on deep learning, comprising the following steps: Collect and preprocess time-series data from multiple sources of sensors in the electric aircraft power system to obtain fault characteristic time-series matrix and battery electrochemical parameter time-series matrix. Based on the fault feature time series matrix, a fault causal prediction network is constructed using a deep learning algorithm optimized for FPGA, and the fault causal relationship matrix is ​​output. Based on the battery electrochemical parameter time matrix and fault causality matrix, a physical constraint neural network is used to perform multi-scale degradation prediction and output a comprehensive health index of the power system. A fault prediction strategy is generated based on the fault causal relationship matrix and the comprehensive health index of the power system. Based on the fault prediction strategy, predictive maintenance and fault early warning are performed on the power system of the electric aircraft.

[0006] In some embodiments, the step of constructing a fault causal prediction network based on the fault feature time-series matrix and using an FPGA-optimized deep learning algorithm to output a fault causal relationship matrix includes: Based on the fault feature time series matrix, a long short-term memory fault prediction network is constructed. The weights of the long short-term memory fault prediction network are causally mapped according to the fault propagation causal chain of electric aircraft. Dynamic weight pruning is performed in combination with hardware-in-the-loop test fault injection feedback to obtain the pruned long short-term memory fault prediction network. A demand prediction model is established based on the severity level of the fault. The pruned long short-term memory fault prediction network is then subjected to adaptive quantization processing to obtain a quantized optimized fault prediction network. The fault feature timing matrix is ​​adaptively compensated by FPGA load prediction, and the compensated fault feature timing matrix is ​​input into the quantized optimized fault prediction network for fault causality analysis, and the fault causality matrix is ​​output.

[0007] In some embodiments, the step of constructing a long short-term memory (LSTM) fault prediction network based on the fault feature time-series matrix, performing causal mapping on the LSM fault prediction network weights according to the causal chain of electric aircraft fault propagation, and performing dynamic weight pruning in conjunction with fault injection feedback from hardware-in-the-loop testing to obtain a pruned LSM fault prediction network includes: Based on the fault propagation path of battery-motor-propeller in the power system of electric aircraft, a fault causal chain graph is constructed, and the fault causal chain graph is mapped into a weight importance scoring matrix of long short time memory fault prediction network. The real-time fault injection data of the hardware-in-the-loop test system is used as a feedback signal to dynamically adjust the weight retention threshold in the weight importance scoring matrix. The Long Short-Term Memory (LSTM) Fault Prediction Network is pruned based on the adjusted weight retention threshold to obtain the pruned LSM Fault Prediction Network.

[0008] In some embodiments, the step of establishing a demand prediction model based on the fault severity level and adaptively quantizing the pruned long short-term memory fault prediction network to obtain a quantized optimized fault prediction network includes: Establish a demand prediction model for FPGA computing resources, storage resources, and communication bandwidth based on the severity level of the fault type; Based on the demand prediction model, determine the quantization bit allocation strategy for different layers in the pruned long short-term memory fault prediction network; The weights and activation values ​​of the pruned Long Short-Term Memory (LSTM) fault prediction network are adaptively quantized according to the quantization bit allocation strategy to obtain a quantized optimized fault prediction network.

[0009] In some embodiments, the adaptive timing compensation of the fault feature timing matrix through FPGA load prediction, the input of the compensated fault feature timing matrix into the quantized optimized fault prediction network for fault causality analysis, and the output of the fault causality matrix include: A load prediction model is built based on the current load status and processing capacity of the FPGA to predict the processing delay time of multi-sensor data; Based on the processing delay time, time-series alignment compensation is performed on the different sensor data in the fault feature time-series matrix to obtain the compensated fault feature time-series matrix. The compensated fault feature time series matrix is ​​input into the quantized optimized fault prediction network for forward inference, and the fault causal relationship matrix is ​​output.

[0010] In some embodiments, the step of using a physically constrained neural network to perform multi-scale degradation prediction based on the battery electrochemical parameter time series matrix and the fault causality matrix, and outputting a comprehensive health index of the power system, includes: Based on the battery electrochemical parameter time series matrix, a temperature-adaptive physical information neural network is constructed by real-time parameter identification of electrochemical impedance spectroscopy. The network structure is then optimized by combining the fault causal relationship matrix to obtain the temperature-adaptive physical information neural network model. Based on the temperature-adaptive physical information neural network model, a sparse attention mechanism based on the physical mechanism of low-temperature batteries is used to predict degradation at multiple time scales, and a battery cell health status prediction matrix is ​​obtained. Based on the battery cell health status prediction matrix, a dynamic graph neural network with thermal conduction physical constraints is used to collaboratively predict the battery pack consistency degradation and output the power system comprehensive health index.

[0011] In some embodiments, the step of constructing a temperature-adaptive physical information neural network based on the battery electrochemical parameter time series matrix, through real-time parameter identification of electrochemical impedance spectroscopy, and optimizing the network structure by combining the fault causal relationship matrix to obtain a temperature-adaptive physical information neural network model includes: Real-time electrochemical parameter identification is performed based on the impedance spectrum data in the battery electrochemical parameter time series matrix to obtain a temperature response parameter matrix, which includes battery internal resistance, diffusion impedance and charge transfer impedance. The temperature response parameter matrix is ​​used as a physical constraint to construct the loss function of a temperature-adaptive physical information neural network. Based on the fault causal relationship matrix, the weight allocation of the hidden layer structure of the temperature adaptive physical information neural network is optimized to obtain the temperature adaptive physical information neural network model.

[0012] In some embodiments, the temperature-adaptive physical information neural network model employs a sparse attention mechanism based on the physical mechanism of low-temperature batteries to perform multi-timescale degradation prediction, obtaining a battery cell health state prediction matrix, including: Based on the electrochemical reaction kinetics mechanism of lithium batteries under low temperature conditions, a sparse attention weight matrix is ​​constructed and integrated into the temperature adaptive physical information neural network model. The temperature-adaptive physical information neural network model is used to perform time-series prediction by a multi-time-scale sliding window to obtain short-term battery degradation trend vectors, medium-term battery degradation trend vectors, and long-term battery degradation trend vectors, respectively. By integrating the short-term battery degradation trend vector, the medium-term battery degradation trend vector, and the long-term battery degradation trend vector, a battery cell health status prediction matrix is ​​obtained.

[0013] In some embodiments, the step of using a dynamic graph neural network based on the battery cell health state prediction matrix and thermal conduction physical constraints to collaboratively predict battery pack consistency degradation and output a comprehensive power system health index includes: A dynamic graph structure is constructed based on the thermal conductivity physical characteristics of the battery pack, and the health status prediction matrix of the individual battery cells is mapped to the feature matrix of graph nodes. A graph convolutional neural network constrained by thermal conduction is used to aggregate neighbor node information of the graph node feature matrix to obtain a battery pack consistency degradation prediction matrix. Based on the battery pack consistency degradation prediction matrix, the overall health status of the battery pack is calculated, and the comprehensive health index of the power system is output.

[0014] On the other hand, the present invention also provides a deep learning-based fault prediction system for electric aircraft power systems, the system comprising: The data acquisition and preprocessing module is used to acquire and preprocess time-series data from multiple sources of sensors in the electric aircraft's power system to obtain a fault characteristic time-series matrix and a battery electrochemical parameter time-series matrix. The fault causal prediction module is used to construct a fault causal prediction network based on the fault feature time series matrix and an FPGA-optimized deep learning algorithm, and output a fault causal relationship matrix. The multi-scale degradation prediction module is used to perform multi-scale degradation prediction based on the battery electrochemical parameter time series matrix and the fault causality matrix, using a physically constrained neural network, and outputs a comprehensive health index of the power system. The fault prediction decision module is used to generate a fault prediction strategy based on the fault causal relationship matrix and the comprehensive health index of the power system, and to perform predictive maintenance and fault early warning for the electric aircraft power system based on the fault prediction strategy.

[0015] The fault prediction method and system for electric aircraft power systems based on deep learning of the present invention have the following advantages over the prior art: (1) This application collects multi-source sensor time-series data and hardware-in-the-loop test data of the electric aircraft power system, and constructs a fault feature time-series matrix and a battery electrochemical parameter time-series matrix. Based on the FPGA-optimized deep learning algorithm, a fault causal prediction network is established to capture the fault propagation causal relationship. Combined with the temperature-adaptive physical information neural network, multi-scale degradation prediction is performed to generate a comprehensive health index of the power system. Through the correlation analysis between the fault causal relationship matrix and the health index, the fault risk is evaluated and a prediction strategy is generated. Thus, the multi-dimensional sensor abnormal signals and electrochemical parameters are deeply integrated, realizing the intelligent mapping between the fault symptoms of the electric aircraft power system and the predictive maintenance decision, and realizing the efficient, accurate, and collaborative prediction and early warning of power system faults. (2) This application constructs a long short-term memory fault prediction network based on fault propagation causal chain, and performs dynamic weight pruning in combination with hardware-in-the-loop test feedback. It establishes a fault severity level-driven demand prediction model to perform adaptive quantization processing on the network. Subsequently, it performs time-series alignment compensation on multi-source sensor data through FPGA load prediction model. Using the quantized and optimized fault prediction network, it models the complex causal relationship between fault features, extracts deep fault propagation patterns to complete causal relationship analysis, and outputs a fault causal relationship matrix. This deeply couples the fault propagation path of the electric aircraft power system with the deep learning network architecture, realizes efficient mapping of fault feature time-series data to fault causal relationship, and significantly improves the accuracy and real-time performance of fault causal prediction. (3) This application identifies impedance spectrum parameters in real time based on the time series matrix of battery electrochemical parameters, constructs a temperature-adaptive physical information neural network, optimizes the network structure by combining the fault causal relationship matrix, and generates a temperature-adaptive physical information neural network model; then, it uses a sparse attention mechanism based on the physical mechanism of low-temperature batteries to predict degradation at multiple time scales, uses a dynamic graph neural network with thermal conduction physical constraints to model the complex thermal conduction coupling relationship between battery cells, extracts deep degradation collaborative mode to complete consistent prediction, and outputs the comprehensive health index of the power system, thereby deeply integrating the physical mechanism of battery electrochemicals with the deep neural network architecture, realizing the multi-scale nonlinear mapping from battery electrochemical parameters to the health state of the power system, and significantly improving the accuracy, timeliness and temperature adaptability of the degradation prediction of the electric aircraft power system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a deep learning-based method for predicting faults in the power system of an electric aircraft, according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Currently, fault prediction for electric aircraft propulsion systems often relies on extensive testing and engineers' experience for fault identification. The optimal prediction parameters vary significantly across different battery systems and operating conditions. Identifying each fault mode requires multiple experimental verifications, which is time-consuming and costly, resulting in long prediction cycles and poor early warning stability. Existing empirical models for novel propulsion systems suffer from a lack of rapid adaptability, insufficient accuracy, and difficulty in accurately modeling the mapping relationship between complex electrochemical characteristics and fault propagation.

[0020] In related technologies, threshold-based fault prediction systems cannot effectively handle multidimensional nonlinear relationships, leading to prediction results that deviate from actual requirements. Furthermore, traditional algorithms have limited responsiveness to changes in the dynamic flight environment, making it difficult to meet the high-precision fault prediction requirements of advanced electric aircraft.

[0021] Based on this, in some embodiments, please refer to Figure 1 This invention provides a deep learning-based method for predicting faults in the power system of electric aircraft, comprising the following steps: S1 collects and preprocesses multi-source sensor time-series data of the electric aircraft's power system to obtain the fault characteristic time-series matrix and the battery electrochemical parameter time-series matrix.

[0022] It is understandable that step S1 involves collecting real-time operational data through various sensors deployed at key locations in the electric aircraft's power system, while simultaneously collecting fault injection feedback data using a hardware-in-the-loop test system and impedance spectrum data using an impedance spectrometer. The collected raw data undergoes preprocessing operations such as time-series alignment, outlier handling, and data standardization to ultimately construct a time-series matrix containing fault characteristics and a time-series matrix of battery electrochemical parameters.

[0023] S2, Based on the fault feature time series matrix, a fault causal prediction network is constructed using a deep learning algorithm optimized for FPGA, and the fault causal relationship matrix is ​​output.

[0024] It is understandable that step S2 uses the fault feature time series matrix as input, constructs a long short-term memory fault prediction network and performs weight mapping by combining the fault propagation causal chain of electric aircraft, uses hardware-in-the-loop test feedback for dynamic weight pruning, and then establishes a demand prediction model through fault severity level to perform adaptive quantization processing on the network. Finally, after time series compensation through FPGA load prediction, the fault causal relationship matrix is ​​output, realizing the accurate identification of fault propagation path.

[0025] S3. Based on the battery electrochemical parameter time series matrix and fault causal relationship matrix, a physical constraint neural network is used to perform multi-scale degradation prediction and output the comprehensive health index of the power system.

[0026] It is understandable that step S3 identifies impedance spectrum parameters in real time based on the battery electrochemical parameter time series matrix, constructs a temperature-adaptive physical information neural network model, uses a sparse attention mechanism based on the low-temperature battery physical mechanism to predict degradation at multiple time scales, and then uses a dynamic graph neural network with thermal conduction physical constraints to predict the consistent degradation of the battery pack. Finally, it outputs a comprehensive health index of the power system that comprehensively considers the health status of individual battery cells and the overall performance of the system.

[0027] S4. A fault prediction strategy is generated based on the fault causal relationship matrix and the comprehensive health index of the power system. Based on the fault prediction strategy, predictive maintenance and fault warning are performed on the power system of the electric aircraft.

[0028] It is understood that step S4 constructs a fault risk assessment model and generates a fault prediction strategy by analyzing the correlation between the fault causal relationship matrix and the comprehensive health index of the power system. The fault prediction strategy includes fault type, occurrence probability and maintenance time window. Then, based on the preset maintenance operation library and time scheduling algorithm, specific maintenance decision instructions are generated to realize proactive predictive maintenance and real-time fault early warning of the electric aircraft power system.

[0029] This embodiment proposes a deep learning-based method for predicting faults in electric aircraft propulsion systems. It achieves end-to-end mapping from multi-source sensor data to fault prediction strategies, deeply integrating the complex fault propagation mechanism of electric aircraft propulsion systems with deep learning technology, and significantly improving the accuracy, real-time performance, and reliability of fault prediction.

[0030] In some embodiments, based on the above embodiments, step S1 includes: By collecting timing data from multi-source sensors and hardware-in-the-loop test systems deployed at key parts of the electric aircraft's power system, multi-source sensor timing data and hardware-in-the-loop test data are obtained.

[0031] It is understood that the multi-source sensors in this embodiment include, but are not limited to, temperature sensors, voltage sensors, current sensors, vibration sensors, speed sensors, and impedance spectrometers, which are respectively arranged in key parts such as motors, electronic controllers, battery packs, and propellers to achieve comprehensive monitoring of the operating status of the power system.

[0032] For example, impedance spectrum time-series data of the battery pack is acquired using an impedance spectrometer; fault injection feedback time-series data is acquired using a hardware-in-the-loop test system to obtain hardware-in-the-loop test data; temperature time-series data of the motor, electronic controller, and battery pack are acquired using a temperature sensor; voltage and current time-series data of the battery pack and motor are acquired using a voltage sensor and a current sensor; vibration time-series data of the motor and propeller are acquired using a vibration sensor; and speed time-series data of the motor are acquired using a speed sensor. The temperature time-series data, voltage time-series data, current time-series data, vibration time-series data, speed time-series data, and impedance spectrum time-series data are then combined to obtain multi-source sensor time-series data.

[0033] Among them, the impedance spectrum time series data reflects the frequency domain information of the internal electrochemical impedance characteristics of the battery, including key parameters such as DC internal resistance, ohmic impedance and polarization impedance; the hardware-in-the-loop test data is obtained by injecting preset fault modes into the test environment, including the response characteristics of various fault types such as sensor faults, actuator faults and system cascade faults.

[0034] The time-series data from the multi-source sensors and the hardware-in-the-loop test data are cleaned, feature extracted, and normalized to obtain the fault feature time-series matrix and the battery electrochemical parameter time-series matrix, respectively.

[0035] It is understood that the data preprocessing process in this embodiment adopts a multi-level cleaning strategy to ensure the quality and consistency of the input data.

[0036] For example, the multi-source sensor time-series data and hardware-in-the-loop test data are cleaned by performing outlier detection and missing value imputation to obtain cleaned time-series data; statistical features and time-frequency domain features are extracted from the cleaned time-series data to obtain a feature matrix; the feature matrix is ​​normalized to obtain a normalized feature matrix; fault-related features and hardware-in-the-loop test fault injection feedback features are selected from the normalized feature matrix according to fault detection requirements to obtain a fault feature time-series matrix; electrochemical-related features and impedance spectrum features are selected from the normalized feature matrix according to battery health status assessment requirements to obtain a battery electrochemical parameter time-series matrix.

[0037] Among them, outlier detection adopts a statistical distribution-based approach. The strategy combines rule-based and time-series correlation-based local anomaly detection methods; missing value imputation uses a combination of linear interpolation and forward imputation; statistical features include indicators describing data distribution characteristics such as mean, standard deviation, kurtosis, and skewness; time-frequency domain features are extracted through Fast Fourier Transform and Wavelet Transform, including frequency domain energy distribution and time-frequency domain local features; normalization processing uses the min-max normalization method to map feature values ​​to... Within the specified range, the influence of features of different magnitudes on model training is eliminated.

[0038] By establishing a complete data acquisition and preprocessing system, step S1 realizes the transformation from raw sensor signals to standardized feature matrices, laying a data foundation for fault prediction. Step S1 combines multi-source heterogeneous sensor data with hardware-in-the-loop test data, and through systematic data cleaning and feature engineering methods, generates two types of specialized time-series feature matrices for fault detection and battery health assessment, effectively improving data quality and feature representation capabilities.

[0039] In some embodiments, based on the above embodiments, step S2 includes: Based on the fault feature time series matrix, a long short-term memory (LSTM) fault prediction network is constructed. The weights of the LSTM fault prediction network are causally mapped according to the fault propagation causal chain of the electric aircraft. Dynamic weight pruning is performed in combination with hardware-in-the-loop test fault injection feedback to obtain the pruned LSTM fault prediction network.

[0040] For example, the Long Short-Term Memory (LSTM) fault prediction network in this embodiment adopts an encoder-decoder architecture. The encoder part includes multi-layer LSTM units for extracting temporal patterns of fault features; the decoder part maps to fault type classification results through fully connected layers. The causal mapping process transforms the fault propagation causal chain into weight importance constraints, guiding the parameter learning direction during network training.

[0041] The encoder output calculation formula of the Long Short-Term Memory Fault Prediction Network is as follows: ; in, For the encoder at time The hidden state output vector; This is the hidden state weight matrix; The input gate weight matrix; For a moment The fault feature input vector; For circular connection of weight matrices; For a moment The hidden state vector; This is the hidden state bias vector; for sigmoid Activation function; For a moment The output of the forget gate; The function is a hyperbolic tangent activation function. This embodiment controls the retention rate of historical information through a forget gate, achieving adaptive extraction of temporal patterns of fault features. A demand prediction model is established based on the fault severity level, and the pruned Long Short-Term Memory (LSTM) fault prediction network is adaptively quantized to obtain a quantized optimized fault prediction network.

[0042] It is understood that the demand prediction model in this embodiment dynamically adjusts the network quantization strategy according to the differentiated requirements of computing resources for faults of different severity levels. High-severity faults use higher precision quantization to ensure prediction accuracy, while low-severity faults use lower precision quantization to improve computing efficiency.

[0043] The fault feature timing matrix is ​​adaptively compensated by FPGA load prediction, and the compensated fault feature timing matrix is ​​input into the quantized optimized fault prediction network for fault causality analysis, and the fault causality matrix is ​​output.

[0044] It is understandable that FPGA load prediction estimates data processing latency based on the current processing task queue and hardware resource usage, and eliminates time differences between data from different sensors through timing alignment algorithms to ensure that the data input to the network has good timing consistency.

[0045] In some embodiments, based on the above embodiments, the step of constructing a long short-term memory (LSTM) fault prediction network based on the fault feature time-series matrix, performing causal mapping on the LSM fault prediction network weights according to the causal chain of electric aircraft fault propagation, and performing dynamic weight pruning in conjunction with fault injection feedback from hardware-in-the-loop testing to obtain a pruned LSM fault prediction network includes: Based on the fault propagation path of the battery-motor-propeller in the power system of an electric aircraft, a fault causal chain graph is constructed, and the fault causal chain graph is mapped into a weight importance scoring matrix of a long short-term memory fault prediction network.

[0046] It is understood that the fault causal chain graph in this embodiment uses a directed graph structure to represent the causal relationship between different fault types. The nodes in the graph represent fault types, the weights of the edges represent the strength and probability of fault propagation, and the weight importance scoring matrix calculates the contribution of each connection weight to the fault prediction task through graph theory algorithms.

[0047] For example, the formula for calculating the weighted importance score is: ; in, Assign an importance score to the network connectivity weight; This represents the propagation probability of the corresponding edge in the fault causal chain graph; The degree of a graph node is the number of edges connected to that node. The centrality coefficient of a node reflects its criticality in the fault propagation network; This is the propagation weighting coefficient, used to adjust the contribution of the fault propagation probability. The centrality weight coefficient is used to adjust the degree of influence of node importance. This embodiment balances the influence of nodes with high numbers using a logarithmic function, while combining centrality analysis to ensure that the weights of critical faulty nodes are appropriately preserved.

[0048] The real-time fault injection data from the hardware-in-the-loop test system is used as a feedback signal to dynamically adjust the weight retention threshold in the weight importance scoring matrix.

[0049] It is understandable that hardware-in-the-loop testing systems obtain real fault response data by injecting preset fault modes into actual hardware platforms. The feedback signals include fault type, severity, and propagation delay, which are used to correct the accuracy of weighted importance scores.

[0050] For example, the formula for calculating the weight retention threshold is: ; in, Retain the threshold for the dynamically adjusted weights; The threshold for retaining basic weights; The causal strength coefficients between nodes in the fault causal chain graph; This is the standardized value of the hardware-in-the-loop test feedback factor; This is the causality attenuation coefficient, used to control the degree of influence of causality strength on the threshold. This is the feedback oscillation amplitude coefficient, used to adjust the sensitivity of hardware-in-the-loop test feedback; The feedback frequency parameter reflects the update frequency of hardware-in-the-loop test data. This embodiment introduces an exponential decay mechanism to model the propagation characteristics of fault causality in the weighted retention threshold, while employing a sine function to capture the periodic variation of hardware-in-the-loop test feedback, thus achieving dynamic threshold adjustment based on the dual constraints of causal chain strength and test feedback.

[0051] The Long Short-Term Memory (LSTM) Fault Prediction Network is pruned based on the adjusted weight retention threshold to obtain the pruned LSM Fault Prediction Network.

[0052] It is understood that the weight pruning operation in this embodiment removes connections with importance scores below a threshold using a structured pruning method, while maintaining the network's main predictive ability. The pruned network has fewer parameters and faster inference speed.

[0053] In some embodiments, based on the above embodiments, the step of establishing a demand prediction model based on the fault severity level and performing adaptive quantization processing on the pruned long short-term memory fault prediction network to obtain a quantized optimized fault prediction network includes: Establish a demand forecasting model for FPGA computing resources, storage resources, and communication bandwidth based on the severity level of the fault type.

[0054] It is understood that the demand forecasting model in this embodiment establishes a mapping relationship between resource consumption and severity level based on historical fault handling data, and uses multiple regression analysis to quantify the degree of demand for various types of resources for faults of different severity, providing a decision-making basis for quantification strategies.

[0055] For example, the calculation formula of the resource demand forecasting model is: ; in, Forecast resource demand; The demand for basic resources; This is a standardized value for the severity level of the fault, with a range of [value range missing]. ; This is the severity amplification factor, used to adjust the intensity of the impact of severity on resource demand; It is a non-linear index, reflecting the non-linear relationship between severity and resource demand; This is the concurrency factor; This represents the number of faults currently being processed concurrently. This embodiment uses an exponential function to model the non-linear relationship between severity and resource requirements, and considers the additional resource requirements of processing multiple faults simultaneously through the concurrent processing term.

[0056] Based on the demand prediction model, the quantization bit allocation strategy for different layers in the pruned long short-term memory fault prediction network is determined.

[0057] It is understood that the quantization bit allocation in this embodiment adopts a hierarchical adaptation strategy, allocating more bits to layers with high prediction accuracy requirements and fewer bits to layers that are sensitive to computational complexity, and determining the quantization accuracy requirements of each layer through sensitivity analysis.

[0058] For example, the formula for calculating the quantization bit allocation is: ; in, The number of bits allocated to the network layer; The maximum number of available bits; This is the sensitivity index of the network layer, calculated through gradient analysis; This is the bit compression factor, which controls the aggressiveness of quantization; This is a sensitivity amplification factor, used to adjust the effect of sensitivity on bit allocation; This is the floor function. This embodiment uses a logarithmic function to achieve a non-linear mapping from sensitivity to bit count, allocating more bits to the high-sensitivity layer to ensure prediction accuracy, and allocating fewer bits to the low-sensitivity layer to improve computational efficiency.

[0059] The weights and activation values ​​of the pruned Long Short-Term Memory (LSTM) fault prediction network are adaptively quantized according to the quantization bit allocation strategy to obtain a quantized optimized fault prediction network.

[0060] It is understood that the adaptive quantization processing in this embodiment includes two stages: weight quantization and activation value quantization. The weight quantization uses a non-uniform quantization method to maintain the accuracy of important weights, and the activation value quantization uses a dynamic quantization range adjustment strategy to adapt to the distribution characteristics of different input data.

[0061] For example, the formula for calculating the weight quantization is: ; in, These are the quantized weight values; These are the original weight values; This is the quantization step size; It is a symbolic function; This is a rounding function; the quantization step size The formula for calculation is: ;in, and These are the maximum and minimum values ​​of the weights, respectively. The number of bits allocated to the network layer. This embodiment uses a symbol-amplitude separation quantization scheme to achieve accurate amplitude quantization while preserving the weight symbol information.

[0062] In some embodiments, based on the above embodiments, the step of adaptively compensating the fault feature timing matrix through FPGA load prediction, inputting the compensated fault feature timing matrix into the quantized optimized fault prediction network for fault causality analysis, and outputting a fault causality matrix includes...

[0063] A load prediction model is built based on the current load status and processing capacity of the FPGA to predict the processing delay time of multi-sensor data.

[0064] It is understood that the load prediction model in this embodiment adopts a time series prediction method, combining indicators such as FPGA resource utilization, task queue length and data throughput, and uses the Kalman filter algorithm to achieve accurate prediction of processing delay time.

[0065] For example, the formula for predicting the processing delay time is: ; in, This is a predicted value for the multi-sensor data processing delay time; This represents the current FPGA task queue length, i.e., the number of tasks. Standard processing time for a single task; This is a task complexity weighting factor, reflecting the computational complexity of different fault prediction tasks; This represents the theoretical maximum throughput of the FPGA. The current FPGA resource utilization rate includes the combined utilization rate of logic units, memory blocks, and DSP resources. Basic system overhead time; This is the severity amplification factor, used to adjust the impact of fault severity on processing priority; This represents the standardized value for the fault severity level. In this embodiment, FPGA resource utilization is used as a denominator correction factor to reflect the nonlinear impact of hardware resource contention on processing performance. Simultaneously, an exponential function is used to model the amplification effect of fault severity on system overhead, achieving adaptive latency prediction based on hardware status and fault characteristics.

[0066] Based on the processing delay time, time-series alignment compensation is performed on the different sensor data in the fault feature time-series matrix to obtain the compensated fault feature time-series matrix.

[0067] It is understood that the timing alignment compensation in this embodiment adopts an adaptive interpolation algorithm to perform timestamp correction based on the delay differences of each sensor data, so as to ensure that all sensor data remain consistent in the time dimension.

[0068] For example, the interpolation formula for the timing alignment compensation is: ; in, For a moment Sensor data values ​​after timing alignment compensation; For a moment The raw sensor data values, For a moment The raw sensor data values; This is the predicted processing delay time for the sensor; The data sampling interval; This represents the integer number of sampling periods corresponding to the delay time. This embodiment uses a linear interpolation method to compensate for delays in sampling intervals that are not integer multiples, ensuring precise alignment of data from different sensors on the time axis.

[0069] The compensated fault feature time series matrix is ​​input into the quantized optimized fault prediction network for forward inference, and the fault causal relationship matrix is ​​output.

[0070] It can be understood that the forward reasoning process in this embodiment includes three stages: feature encoding, temporal modeling, and causal relationship parsing. The final output fault causal relationship matrix includes the correlation strength and propagation probability information between each fault type.

[0071] By constructing a deep learning-based fault prediction framework optimized for FPGA, step S2 realizes a complete process from fault feature extraction to causal relationship analysis. Step S2 integrates fault propagation causal chain knowledge into the neural network weight design, combines a dynamic weight pruning mechanism based on hardware-in-the-loop test feedback, and an adaptive quantization strategy based on fault severity, thereby improving the performance and prediction accuracy of the fault prediction network in resource-constrained environments.

[0072] In some embodiments, based on the above embodiments, step S3 includes: Based on the battery electrochemical parameter time series matrix, a temperature-adaptive physical information neural network is constructed through real-time parameter identification of electrochemical impedance spectroscopy. The network structure is then optimized by combining the fault causal relationship matrix to obtain the temperature-adaptive physical information neural network model.

[0073] It is understood that the temperature-adaptive physical information neural network in this embodiment embeds electrochemical physical mechanisms as prior knowledge into the network architecture, and ensures that the network output conforms to the laws of electrochemical reaction kinetics through a physical constraint loss function, thereby achieving a deep integration of data-driven and physical mechanisms.

[0074] Based on the temperature-adaptive physical information neural network model, a sparse attention mechanism based on the physical mechanism of low-temperature batteries is used to predict degradation at multiple time scales, thereby obtaining a prediction matrix for the health status of individual battery cells.

[0075] It is understood that the sparse attention mechanism in this embodiment selectively focuses on the most critical temporal features for degradation prediction based on the limited lithium-ion diffusion and the changes in electrochemical reaction kinetics under low temperature conditions, thus avoiding the computational complexity problem of traditional full attention mechanisms in long sequence prediction.

[0076] Based on the battery cell health status prediction matrix, a dynamic graph neural network with thermal conduction physical constraints is used to collaboratively predict the battery pack consistency degradation and output the power system comprehensive health index.

[0077] It is understood that the dynamic graph neural network in this embodiment captures the differentiated degradation patterns caused by uneven temperature distribution inside the battery pack by modeling the thermal coupling relationship and electrical connection topology between battery cells, thus achieving an organic unity from cell prediction to system-level health assessment.

[0078] In some embodiments, based on the above embodiments, the step of constructing a temperature-adaptive physical information neural network based on the battery electrochemical parameter time series matrix, through real-time parameter identification of electrochemical impedance spectroscopy, and optimizing the network structure by combining the fault causal relationship matrix to obtain a temperature-adaptive physical information neural network model includes: Real-time electrochemical parameter identification is performed based on the impedance spectrum data in the battery electrochemical parameter time series matrix to obtain a temperature response parameter matrix, which includes battery internal resistance, diffusion impedance, and charge transfer impedance.

[0079] For example, the calculation formula for the real-time electrochemical parameter identification is: ; in, For frequency The electrochemical impedance value obtained from the identification; These are the battery's internal resistance parameters; These are the charge transfer resistance parameters; The charge transfer time constant; The index is for constant-phase elements, reflecting the non-ideal nature of the electrode interface; The Weber coefficient characterizes the diffusion impedance properties; The unit is the imaginary unit. This embodiment achieves real-time identification of electrochemical equivalent circuit parameters through nonlinear least-squares fitting of multi-frequency impedance measurement data. It can be understood that the real-time parameter identification in this embodiment employs a recursive least-squares method based on Kalman filtering. By fitting the Nyquist plot characteristics of the impedance spectrum, key parameters reflecting the internal electrochemical processes of the battery are extracted. The temperature response characteristics are modeled using the Arrhenius equation to depict the parameter variations at different temperatures.

[0080] For example, the formula for calculating the temperature response parameter is: ; in, For temperature Electrochemical parameter values ​​under the following conditions; Reference temperature The baseline value of the parameters is as follows; It is the activation energy of the reaction; Boltzmann's constant has a value of . ; Absolute temperature; Reference absolute temperature; This is an exponential function. This embodiment establishes an exponential relationship between temperature and electrochemical parameters based on the Arrhenius equation, achieving temperature-adaptive dynamic parameter adjustment. The temperature response parameter matrix is ​​used as a physical constraint to construct the loss function of a temperature-adaptive physical information neural network.

[0081] It can be understood that the physical constraint loss function in this embodiment includes two parts: a data fitting term and a physical consistency term. The data fitting term measures the deviation between the network output and the measured data, while the physical consistency term ensures that the network prediction results meet the theoretical constraints of the electrochemical impedance model. The weighted fusion achieves a balance between data-driven and mechanism-driven approaches.

[0082] For example, the formula for calculating the physical constraint loss function is as follows: ; in, The total constraint loss function for the temperature-adaptive physical information neural network; To standardize the fitting error of empirical data; To standardize the error of physical constraint violation; To constrain violations of strength indicators; The empirical loss weighting coefficient; These are the basic weighting coefficients for physical constraints; To constrain the penalty intensity parameter. This embodiment... and All are standardized to ensure dimensional consistency; exponential functions Used to nonlinearly amplify the penalty for violating physical constraints; when When (no violation), the exponent term is 1, degenerating into a standard linear combination; when As the value increases, the physical constraint terms grow exponentially, reflecting a strong penalty for violating the laws of physics.

[0083] Based on the fault causal relationship matrix, the weight allocation of the hidden layer structure of the temperature adaptive physical information neural network is optimized to obtain the temperature adaptive physical information neural network model.

[0084] It is understood that the weight allocation optimization in this embodiment dynamically adjusts the activation weights of the corresponding hidden layer neurons in the network according to the degree of influence of different fault types on battery degradation in the fault causal relationship matrix, so that the network pays more attention to the electrochemical characteristics related to key faults.

[0085] For example, the formula for calculating the hidden layer weight allocation is: ; in, This is the adjusted hidden layer weight matrix; This is the original hidden layer weight matrix; The causal influence strength extracted from the fault causality matrix; This is the weight adjustment magnitude coefficient, which controls the strength of the weight correction based on causality. These are nonlinear transformation coefficients used to adjust the sensitivity to causal effects; The hyperbolic tangent function ensures that the weight adjustments are within a reasonable range. This embodiment uses the saturation characteristic of the hyperbolic tangent function to prevent excessive weight adjustments, thus achieving stable network structure optimization.

[0086] In some embodiments, based on the above embodiments, the step of using a sparse attention mechanism based on the low-temperature battery physical mechanism to perform multi-timescale degradation prediction based on the temperature-adaptive physical information neural network model to obtain a battery cell health state prediction matrix includes: Based on the electrochemical reaction kinetics mechanism of lithium batteries under low temperature conditions, a sparse attention weight matrix is ​​constructed and integrated into the temperature-adaptive physical information neural network model.

[0087] It is understood that the sparse attention weight matrix in this embodiment uses a gating mechanism to selectively activate key time-series features related to temperature, based on the physical mechanism characteristics of reduced lithium-ion diffusion coefficient and increased charge transfer resistance under low-temperature conditions, thereby achieving efficient allocation of computing resources.

[0088] For example, the sparse attention weight matrix is ​​calculated as follows: ; in, This is a sparse attention weight matrix; and These are the query matrix and the key matrix, respectively, composed of features extracted by a temperature-adaptive physical information neural network. is the dimension of the key matrix, used for normalizing the attention score; This is a sparse mask matrix constructed based on the physical mechanism of low-temperature batteries, used to mask non-critical features; This represents element-wise multiplication. It is a linear rectification activation function; This is the normalized exponential activation function.

[0089] The sparse mask matrix The formula for calculation is: ; in, The nth in the mask matrix Line number Column elements; For temperature The lithium-ion diffusion coefficient at the specified depth; Features and characteristics Importance coefficient of transmission between; This is the sparsity threshold parameter, used to control the sparsity of the attention weights. The diffusion coefficient is calculated as follows: ;in, The pre-exponential factor of the diffusion coefficient; It is the diffusion activation energy; Boltzmann's constant has a value of . ; It is an exponential function. This embodiment models the inhibitory effect of low-temperature environment on ion transport using a temperature-dependent diffusion coefficient.

[0090] The temperature-adaptive physical information neural network model is used to perform time-series predictions through a multi-time-scale sliding window to obtain short-term battery degradation trend vectors, medium-term battery degradation trend vectors, and long-term battery degradation trend vectors.

[0091] It is understood that the multi-timescale prediction in this embodiment adopts a sliding window strategy with different window lengths. Short-term prediction focuses on instantaneous changes at the hour level, medium-term prediction captures cyclical degradation patterns at the day level, and long-term prediction models monthly capacity decay trends.

[0092] By integrating the short-term battery degradation trend vector, the medium-term battery degradation trend vector, and the long-term battery degradation trend vector, a battery cell health status prediction matrix is ​​obtained.

[0093] It is understood that the fusion strategy in this embodiment adopts a time-scale-aware weighted average method, assigning corresponding weights to the trend vectors of each scale according to the different prediction time spans, and obtaining a prediction matrix that comprehensively reflects the evolution law of battery health status.

[0094] In some embodiments, based on the above embodiments, the step of using a dynamic graph neural network with thermal conduction physical constraints to collaboratively predict battery pack consistency degradation based on the battery cell health status prediction matrix and outputting a comprehensive power system health index includes: A dynamic graph structure is constructed based on the thermal conductivity physical characteristics of the battery pack, and the health status prediction matrix of the individual battery cells is mapped to the feature matrix of the graph nodes.

[0095] It is understood that the dynamic graph structure in this embodiment establishes the connection relationship between nodes based on the physical layout and heat conduction path of the battery pack. The adjacency matrix elements of the graph represent the thermal coupling strength between adjacent battery cells, and the node features include the health status, temperature distribution and electrical parameters of the battery cells.

[0096] A graph convolutional neural network with thermal conduction physical constraints is used to aggregate neighbor node information of the graph node feature matrix to obtain a battery pack consistency degradation prediction matrix.

[0097] It is understood that the message passing process of the graph convolutional neural network in this embodiment follows the physical constraints of the heat conduction equation, ensuring that information aggregation conforms to the heat propagation law inside the battery pack, and capturing the complex thermoelectric coupling relationship inside the battery pack through multi-layer graph convolution operations.

[0098] Based on the battery pack consistency degradation prediction matrix, the overall health status of the battery pack is calculated, and the comprehensive health index of the power system is output.

[0099] It is understood that the overall health status of the battery pack in this embodiment is obtained by weighted fusion of the health scores of each individual battery cell. The weighting considers the importance of the battery in the circuit topology and its impact on system performance. The comprehensive health index is output using a standardized scoring system. The health assessment results.

[0100] By constructing a multi-scale degradation prediction framework that integrates electrochemical physical mechanisms and deep learning technology, step S3 achieves a comprehensive health status assessment from individual battery cells to the entire battery pack system. Step S3 organically combines electrochemical impedance spectroscopy parameter identification, temperature-adaptive physical information neural networks, sparse attention mechanisms, and thermal conduction constraint graph neural networks to form a data-driven prediction method guided by physical mechanisms, which significantly improves the accuracy and physical interpretability of battery health status prediction.

[0101] In some embodiments, based on the above embodiments, step S4 includes: Based on the aforementioned fault causal relationship matrix and the comprehensive health index of the power system, a fault risk assessment model is constructed, and a fault prediction strategy is output.

[0102] It is understood that the fault risk assessment model in this embodiment adopts a hierarchical assessment architecture. First, a risk propagation network is constructed based on the correlation strength and propagation path between different fault types in the fault causal relationship matrix. Then, the comprehensive health index of the power system is combined as a quantitative indicator of the overall health status of the system. Finally, the probability of occurrence and the degree of impact of various faults are calculated.

[0103] For example, a correlation analysis is performed between the fault weight coefficients in the fault causal relationship matrix and the comprehensive health index of the power system to establish a fault severity level classification standard. Based on the fault severity level classification standard, a risk quantification assessment is performed on potential fault types, and a fault prediction strategy is output. The fault prediction strategy includes fault type, probability of occurrence, and maintenance time window.

[0104] The correlation analysis employs a combination of Pearson correlation coefficient and mutual information theory to quantify the linear and nonlinear correlation between the fault weight coefficient and the comprehensive health index. The fault severity level classification standard is established based on historical fault data statistics, including four levels: minor, moderate, severe, and fatal, with each level corresponding to a different risk threshold range. The risk quantification assessment adopts the fuzzy comprehensive evaluation method, which transforms the qualitative description of the fault impact into a quantitative risk score. The probability of occurrence is calculated using Bayesian inference, and the maintenance time window is determined based on the fault development trend and remaining useful life prediction.

[0105] Based on the fault prediction strategy, maintenance decision instructions are generated to perform predictive maintenance and fault warnings on the electric aircraft power system.

[0106] It is understood that the maintenance decision instruction generation process in this embodiment includes three sub-stages: maintenance strategy matching, resource scheduling optimization, and execution timing planning. By establishing a complete maintenance decision knowledge base and intelligent scheduling algorithm, the automatic conversion from fault prediction to specific maintenance operations is realized.

[0107] For example, the fault type and occurrence probability in the fault prediction strategy are analyzed, a corresponding maintenance plan is obtained by matching it with a preset maintenance operation library, a maintenance plan scheduling is formulated according to the maintenance time window, a maintenance decision instruction is generated, the maintenance decision instruction includes maintenance type, maintenance component and execution time, and a fault warning signal is sent to the aircraft control system.

[0108] The pre-set maintenance operation library includes standardized maintenance procedures for different fault types, covering various maintenance types such as inspection, replacement, repair, and calibration. Each maintenance operation corresponds to specific operation steps, required tools, and estimated man-hours. The maintenance plan scheduling adopts a constraint satisfaction problem-solving method, comprehensively considering constraints such as maintenance resource availability, flight mission arrangement, and maintenance urgency to generate the optimal maintenance time schedule. The fault warning signal includes fault type identification, urgency level, expected occurrence time, and suggested handling measures, which are sent to the aircraft control system via CAN bus or Ethernet interface.

[0109] By constructing a complete closed loop for fault prediction decision-making, step S4 achieves fully automated management of the entire process from fault prediction to maintenance execution. Step S4 establishes a risk assessment model based on causal relationship analysis and health index fusion, as well as a maintenance decision generation mechanism based on knowledge base matching and constraint optimization, effectively improving the practicality and operability of fault prediction.

[0110] In some embodiments, the present invention also provides a deep learning-based fault prediction system for electric aircraft power systems, the system comprising: The data acquisition and preprocessing module is used to acquire and preprocess time-series data from multiple sources of sensors in the electric aircraft's power system to obtain a fault characteristic time-series matrix and a battery electrochemical parameter time-series matrix. The fault causal prediction module is used to construct a fault causal prediction network based on the fault feature time series matrix and an FPGA-optimized deep learning algorithm, and output a fault causal relationship matrix. The multi-scale degradation prediction module is used to perform multi-scale degradation prediction based on the battery electrochemical parameter time series matrix and the fault causality matrix, using a physically constrained neural network, and outputs a comprehensive health index of the power system. The fault prediction decision module is used to generate a fault prediction strategy based on the fault causal relationship matrix and the comprehensive health index of the power system, and to perform predictive maintenance and fault early warning for the electric aircraft power system based on the fault prediction strategy.

[0111] It is understood that the deep learning-based electric aircraft power system fault prediction system of this embodiment realizes the end-to-end mapping from multi-source sensor data to fault prediction strategy, deeply integrates the complex fault propagation mechanism of electric aircraft power system with deep learning technology, and significantly improves the accuracy, real-time performance and reliability of fault prediction.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting faults in the power system of an electric aircraft based on deep learning, characterized in that, Includes the following steps: Collect and preprocess time-series data from multiple sources of sensors in the electric aircraft power system to obtain fault characteristic time-series matrix and battery electrochemical parameter time-series matrix. Based on the fault feature time series matrix, a fault causal prediction network is constructed using a deep learning algorithm optimized for FPGA, and the fault causal relationship matrix is ​​output. Based on the battery electrochemical parameter time matrix and fault causality matrix, a physical constraint neural network is used to perform multi-scale degradation prediction and output a comprehensive health index of the power system. A fault prediction strategy is generated based on the fault causal relationship matrix and the comprehensive health index of the power system. Based on the fault prediction strategy, predictive maintenance and fault early warning are performed on the power system of the electric aircraft.

2. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 1, characterized in that, The fault causal prediction network, constructed using a deep learning algorithm optimized for FPGA based on the fault feature time series matrix, outputs a fault causal relationship matrix, including: Based on the fault feature time series matrix, a long short-term memory fault prediction network is constructed. The weights of the long short-term memory fault prediction network are causally mapped according to the fault propagation causal chain of electric aircraft. Dynamic weight pruning is performed in combination with hardware-in-the-loop test fault injection feedback to obtain the pruned long short-term memory fault prediction network. A demand prediction model is established based on the severity level of the fault. The pruned long short-term memory fault prediction network is then subjected to adaptive quantization processing to obtain a quantized optimized fault prediction network. The fault feature timing matrix is ​​adaptively compensated by FPGA load prediction, and the compensated fault feature timing matrix is ​​input into the quantized optimized fault prediction network for fault causality analysis, and the fault causality matrix is ​​output.

3. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 2, characterized in that, The long short-term memory (LSTM) fault prediction network is constructed based on the fault feature time-series matrix. The weights of the LSM fault prediction network are causally mapped according to the fault propagation causal chain of the electric aircraft. Dynamic weight pruning is then performed in conjunction with fault injection feedback from hardware-in-the-loop testing, resulting in a pruned LSM fault prediction network, including: Based on the fault propagation path of battery-motor-propeller in the power system of electric aircraft, a fault causal chain graph is constructed, and the fault causal chain graph is mapped into a weight importance scoring matrix of long short time memory fault prediction network. The real-time fault injection data of the hardware-in-the-loop test system is used as a feedback signal to dynamically adjust the weight retention threshold in the weight importance scoring matrix. The Long Short-Term Memory (LSTM) Fault Prediction Network is pruned based on the adjusted weight retention threshold to obtain the pruned LSM Fault Prediction Network.

4. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 2, characterized in that, The requirement prediction model based on fault severity level is used to adaptively quantize the pruned long short-term memory fault prediction network to obtain a quantized optimized fault prediction network, including: Establish a demand prediction model for FPGA computing resources, storage resources, and communication bandwidth based on the severity level of the fault type; Based on the demand prediction model, determine the quantization bit allocation strategy for different layers in the pruned long short-term memory fault prediction network; The weights and activation values ​​of the pruned Long Short-Term Memory (LSTM) fault prediction network are adaptively quantized according to the quantization bit allocation strategy to obtain a quantized optimized fault prediction network.

5. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 2, characterized in that, The step involves adaptively compensating the fault feature timing matrix using FPGA load prediction, inputting the compensated fault feature timing matrix into the quantized optimized fault prediction network for fault causality analysis, and outputting a fault causality matrix, including: A load prediction model is built based on the current load status and processing capacity of the FPGA to predict the processing delay time of multi-sensor data; Based on the processing delay time, time-series alignment compensation is performed on the different sensor data in the fault feature time-series matrix to obtain the compensated fault feature time-series matrix. The compensated fault feature time series matrix is ​​input into the quantized optimized fault prediction network for forward inference, and the fault causal relationship matrix is ​​output.

6. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 1, characterized in that, Based on the battery electrochemical parameter time-series matrix and the fault causality matrix, a physically constrained neural network is used to perform multi-scale degradation prediction, outputting a comprehensive health index of the power system, including: Based on the battery electrochemical parameter time series matrix, a temperature-adaptive physical information neural network is constructed by real-time parameter identification of electrochemical impedance spectroscopy. The network structure is then optimized by combining the fault causal relationship matrix to obtain the temperature-adaptive physical information neural network model. Based on the temperature-adaptive physical information neural network model, a sparse attention mechanism based on the physical mechanism of low-temperature batteries is used to predict degradation at multiple time scales, and a battery cell health status prediction matrix is ​​obtained. Based on the battery cell health status prediction matrix, a dynamic graph neural network with thermal conduction physical constraints is used to collaboratively predict the battery pack consistency degradation and output the power system comprehensive health index.

7. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 6, characterized in that, The process involves constructing a temperature-adaptive physical information neural network based on the battery electrochemical parameter time-series matrix, identifying real-time parameters through electrochemical impedance spectroscopy, and optimizing the network structure using the fault causality matrix to obtain the temperature-adaptive physical information neural network model, including: Real-time electrochemical parameter identification is performed based on the impedance spectrum data in the battery electrochemical parameter time series matrix to obtain a temperature response parameter matrix, which includes battery internal resistance, diffusion impedance and charge transfer impedance. The temperature response parameter matrix is ​​used as a physical constraint to construct the loss function of a temperature-adaptive physical information neural network. Based on the fault causal relationship matrix, the weight allocation of the hidden layer structure of the temperature adaptive physical information neural network is optimized to obtain the temperature adaptive physical information neural network model.

8. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 6, characterized in that, The temperature-adaptive physical information neural network model employs a sparse attention mechanism based on the physical mechanism of low-temperature batteries to perform multi-timescale degradation prediction, obtaining a battery cell health status prediction matrix, including: Based on the electrochemical reaction kinetics mechanism of lithium batteries under low temperature conditions, a sparse attention weight matrix is ​​constructed and integrated into the temperature adaptive physical information neural network model. The temperature-adaptive physical information neural network model is used to perform time-series prediction by a multi-time-scale sliding window to obtain short-term battery degradation trend vectors, medium-term battery degradation trend vectors, and long-term battery degradation trend vectors, respectively. By integrating the short-term battery degradation trend vector, the medium-term battery degradation trend vector, and the long-term battery degradation trend vector, a battery cell health status prediction matrix is ​​obtained.

9. The method for predicting faults in the power system of an electric aircraft based on deep learning as described in claim 6, characterized in that, The method, based on the battery cell health status prediction matrix, employs a dynamic graph neural network constrained by thermal conduction to collaboratively predict battery pack consistency degradation, outputting a comprehensive power system health index, including: A dynamic graph structure is constructed based on the thermal conductivity physical characteristics of the battery pack, and the health status prediction matrix of the individual battery cells is mapped to the feature matrix of graph nodes. A graph convolutional neural network constrained by thermal conduction is used to aggregate neighbor node information of the graph node feature matrix to obtain a battery pack consistency degradation prediction matrix. Based on the battery pack consistency degradation prediction matrix, the overall health status of the battery pack is calculated, and the comprehensive health index of the power system is output.

10. A deep learning-based fault prediction system for electric aircraft propulsion systems, used to execute the deep learning-based fault prediction method for electric aircraft propulsion systems as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire and preprocess time-series data from multiple sources of sensors in the electric aircraft's power system to obtain a fault characteristic time-series matrix and a battery electrochemical parameter time-series matrix. The fault causal prediction module is used to construct a fault causal prediction network based on the fault feature time series matrix and an FPGA-optimized deep learning algorithm, and output a fault causal relationship matrix. The multi-scale degradation prediction module is used to perform multi-scale degradation prediction based on the battery electrochemical parameter time series matrix and the fault causality matrix, using a physically constrained neural network, and outputs a comprehensive health index of the power system. The fault prediction decision module is used to generate a fault prediction strategy based on the fault causal relationship matrix and the comprehensive health index of the power system, and to perform predictive maintenance and fault early warning for the electric aircraft power system based on the fault prediction strategy.

Images