Unmanned aerial vehicle health prediction method based on multi-source time sequence signal collaborative analysis driving

Abstract

The invention discloses an unmanned aerial vehicle health prediction method based on multi-source time sequence signal collaborative analysis driving, and belongs to the technical field of unmanned aerial vehicle health management, the unmanned aerial vehicle health prediction method comprises the following steps: S1, carrying out individualized frequency domain health baseline migration for the unmanned aerial vehicle health prediction method, and carrying out frequency domain health baseline migration; and health assessment is carried out on the individualized frequency domain health baseline migration structure. According to the unmanned aerial vehicle health prediction method based on multi-source time sequence signal collaborative analysis driving, multi-source heterogeneous time sequence data, such as an acquisition structure, battery voltage and current, signal strength and temperature, generated in the flight process of an unmanned aerial vehicle are acquired, and the multi-source heterogeneous time sequence data are acquired for six core dimensions of batteries, structures, communication, navigation, hardware and power; a set of refined health degree evaluation model fusing time domain and frequency domain analysis is respectively designed, each dimension is calculated by a plurality of bottom layer indexes through a segmentation scoring or weighting formula method to obtain a five-score system score, and finally, the five-score system score is integrated into a visual multi-dimensional health scoring radar map.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) health management technology, specifically to a UAV health prediction method driven by multi-source time-series signal collaborative analysis. Background Technology

[0002] Current drone health monitoring methods integrate drone sensor data or flight logs and other parameters with multi-dimensional time-series information to construct a deep spatiotemporal correlation model, thereby achieving accurate prediction of the degradation trend of key components. However, in practice, existing drone health monitoring methods are inconvenient for multi-source data collection and preprocessing, which affects the accuracy of the data.

[0003] To overcome the aforementioned deficiencies, existing technology (Chinese patent application CN202311369886.2, application date 20231023) provides a health management system and method for electric aircraft engines. This system comprises a sensing unit, a data transmission unit, a cloud data storage and processing unit, a status monitoring unit, a status prediction unit, a health check unit, a decision-making unit, and a precision control unit. It also utilizes a deep network perception model, addressing the lack of health management and precision control in existing UAV electric aircraft engines. This achieves the goal of fault detection and diagnosis for UAV electric aircraft engines, offering the beneficial effects of comprehensive health management, high intelligence, and high control precision. Furthermore, existing technology (Chinese patent application CN202411285072.5, application date 20240913) also provides this technology. (Patent Application) A method and system for monitoring drone batteries in an intelligent hangar, which inputs the real-time charge-discharge count of the battery into a preset empirical degradation model to obtain an initial value for battery health assessment; inputs the real-time operating condition information of the drone battery into a trained error compensation model to obtain a battery health error assessment value; and superimposes the battery health error assessment value with the initial value for battery health assessment to obtain a battery health assessment value. The error compensation model can effectively supplement the uncertainty in the battery degradation process, enhancing the model's adaptability and prediction accuracy. And the prior art (Chinese patent application CN202111109826.8, application date 20210918) A method for assessing drone health operation based on digital twins, which uses digital twin technology to effectively achieve synergy between safe drone operation and cost reduction. Combined with 5G technology, it effectively improves the drone's communication capabilities. Using finite element method technology or a digital twin model of the drone built in the Simulink environment, the drone in the physical environment is realistically displayed in the virtual environment. The constructed drone visualization and control platform can quickly and intuitively display the drone's health operation status, so as to make timely health assessments and reduce corresponding losses. By interacting with drone data in the virtual and real environments, the system can accurately and efficiently display the real status of drones in the virtual environment, providing support for the development, operation, and maintenance of drones. Although existing technologies can predict drone health, current drone health monitoring relies heavily on time-domain threshold alarms or data-driven black-box models during operation. These models are limited to single-parameter threshold alarms and lack systematic and predictive comprehensive assessments, making it difficult to capture early performance degradation. The latter model carries the risk of "illusion" and is difficult to explain.

[0004] To address the aforementioned issues, there is an urgent need for innovative designs based on existing drone health prediction methods. Summary of the Invention

[0005] The purpose of this invention is to provide a UAV health prediction method driven by multi-source time-series signal collaborative analysis, in order to solve the problems mentioned in the background art that current UAV health monitoring mostly relies on time-domain threshold alarms or data-driven black box models, and is limited to single parameter threshold alarms, lacking systematic and predictive comprehensive evaluation, making it difficult to capture early performance degradation, and the latter has the risk of "illusion" and is difficult to explain.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a UAV health prediction method driven by multi-source time-series signal collaborative analysis, the UAV health prediction method comprising the following steps: S1: For the UAV health prediction method, individualized frequency domain health baseline migration is performed, and the structure of the individualized frequency domain health baseline migration is evaluated for health, and baseline establishment is controlled, and real-time monitoring and migration calculation are performed, as well as a health score for the UAV; S2: For the UAV health prediction method, a dynamically weighted cross-dimensional correlation matrix is ​​constructed for fault prediction, and the overall system state is detected as stable through the construction, dynamic learning and adjustment of the correlation matrix in the dynamically weighted cross-dimensional correlation matrix and the calculation of the comprehensive health entropy value; S3: For the UAV health prediction method, a hybrid approach of physical model and data-driven methods is used to predict the remaining useful life, and the trend is anchored according to the physical model, and the deviation of the data-driven approach is corrected; S4: For the UAV health prediction method, example data and output are provided, and different dimensional states are displayed in conjunction with radar charts.

[0007] Preferably, the baseline is established after the new aircraft or major overhaul, and calibration flights are performed in a standard environment to collect high-frequency vibration signals of the structure (IMU) (sampling rate ≥1kHz); high-resolution Fourier transform (FFT) is performed on the signals to generate a standard frequency domain energy distribution spectrum (F0) of the UAV in a healthy state, and the spectrum is unique.

[0008] Preferably, after each flight, the real-time monitoring and migration calculation also generates a frequency domain energy distribution map (Ft) for that flight; calculates the migration degree (D) between the current map and the healthy baseline, and uses spectral correlation density (SCD) or Wasserstein distance to quantify the difference between the two distributions, where the index is extremely sensitive to small frequency band energy changes, and the calculation formula is: Dt=SCD(F0,Ft).

[0009] Preferably, the smaller the health score migration degree (D), the higher the health score. The structural health score Sstructure is calculated based on the migration degree: Sstructure=5*exp(-λ*Dt) (where λ is the attenuation coefficient, calibrated based on historical data). This method can detect early imbalances or micro-structural cracks that are completely undetectable by traditional time-domain indicators (such as total vibration).

[0010] Preferably, the correlation matrix construction defines a 6x6 health status correlation matrix M, with elements M ij This represents the weight of the influence of the health status of the i-th dimension on the j-th dimension (e.g., the weight of the influence of dynamic system anomalies on structural health); the initial weights are determined by domain expert knowledge and historical failure data.

[0011] Preferably, the weights of the dynamically learned and adjusted matrix are not fixed. The system continuously monitors and uses Granger causality tests or transition entropy analysis to dynamically update the weight matrix M. For example, if historical data repeatedly shows that "battery voltage drops sharply" (i) is followed by frequent occurrences of "navigation module restart" (j), the system will automatically increase M. ij The weight.

[0012] Preferably, the comprehensive health entropy value (H) calculation introduces the concept of "health entropy" to assess the overall disorder of the system. First, the independent scores Si of each dimension are calculated, and then corrected by the correlation matrix M to obtain the comprehensive health entropy value H. The calculation formula is: H=Σ(Pi*log(Pi)), where Pi=(ΣjMji*Sj) / ΣSj. The lower the entropy value H, the more stable and healthy the overall state of the system. An increase in entropy value indicates that the system is in disorder and the risk of failure is increased.

[0013] Preferably, for components with well-defined degradation models, such as batteries, the physical model and data-driven approach no longer simply use long short-term memory networks for prediction. Instead, a physical model is used to anchor the trend: an empirical degradation model based on an electrochemical model (such as a bi-exponential model) is used to predict the macroscopic degradation curve, and data-driven correction of biases is used: a Kalman filter is used to use real-time monitored health indicators (such as internal resistance growth) as observations to correct the prediction results of the physical model in real time. This method utilizes the stability of the physical model and incorporates feedback from real-time data, making the prediction results more reliable and interpretable.

[0014] Preferably, the example data and the output data show that after this flight, the power system health score Spower dropped sharply from 4.5 to 3.0 (due to increased motor torque fluctuations), while the structural system health score Sstructure remained at 4.8 (time-domain indicators were normal). System analysis: Querying the dynamic correlation matrix, it was found that Mpower->structure (the weight of the power's influence on the structure) was very high, at 0.8. The system immediately calculated that the abnormal fluctuations in the power system, even if they did not immediately cause excessive structural vibration, would continue to cause latent fatigue damage to the boom.

[0015] Preferably, the example data and the output in the output include a warning message: "Abnormal fluctuations in the power system have been detected, which are expected to cause the structural health to drop to the warning level (<2.5 points) within the next 15 take-off and landing cycles"; and a maintenance recommendation: "It is recommended to immediately check the balance of the motor and propeller, and to perform non-destructive testing on the critical connection structure."

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] 1. This UAV health prediction method based on multi-source time-series signal collaborative analysis collects multi-source heterogeneous time-series data generated during UAV flight (such as data acquisition structure, battery voltage and current, signal strength, temperature, etc.). For six core dimensions—battery, structure, communication, navigation, hardware, and power—a set of refined health assessment models integrating time-domain and frequency-domain analysis is designed. Each dimension is calculated into a five-point score by several underlying indicators through segmented scoring or weighted formula method, and finally integrated into an intuitive multi-dimensional health score radar chart.

[0018] 2. This UAV health prediction method based on multi-source time-series signal collaborative analysis uses historical big data analysis to determine scoring parameters, realizing the transformation from single fault alarm to system health trend prediction. It can comprehensively, quantitatively, and proactively assess the overall health status of UAVs, providing core decision-making basis for predictive maintenance, significantly improving the operational safety and reliability of UAVs, and pioneering two core mechanisms: "individualized frequency domain health baseline" and "cross-dimensional fault propagation map".

[0019] 3. This UAV health prediction method, driven by multi-source time-series signal collaborative analysis, establishes a unique high-frequency vibration frequency domain energy distribution baseline for each UAV and monitors its migration in real time, thereby achieving the perception of subtle changes in structural health. Furthermore, by constructing a dynamically weighted cross-dimensional correlation matrix, it quantifies the coupling relationship of health status between different systems (such as power and structure, battery and temperature), thereby predicting chain reaction failures caused by the failure of a single component. Through comprehensive evaluation of multi-dimensional health entropy values, it outputs predictive maintenance suggestions that are not only accurate but also highly interpretable, achieving a leap from "post-event alarm" to "pre-event prediction" and from "single-point monitoring" to "system-level insight". Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of the UAV health prediction method driven by multi-source time-series signal collaborative analysis according to the present invention;

[0021] Figure 2 This is a schematic diagram showing the detailed steps of the UAV health prediction method driven by multi-source time-series signal collaborative analysis of the present invention.

[0022] Figure 3 This is a schematic diagram of the data structure for the battery health dimension scoring rules of this invention;

[0023] Figure 4 This is a schematic diagram of the data structure for the health dimension scoring rules of this invention;

[0024] Figure 5 This is a schematic diagram of the multi-dimensional fusion and health radar chart structure of the present invention;

[0025] Figure 6 This is a schematic diagram of an example data structure of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figures 1-6 This invention provides a technical solution: a method for predicting the health of unmanned aerial vehicles (UAVs) based on multi-source time-series signal collaborative analysis, the method comprising the following steps:

[0028] Step 1: For the UAV health prediction method, perform individualized frequency domain health baseline migration, conduct health assessment on the structure of the individualized frequency domain health baseline migration, establish the control baseline, perform real-time monitoring and migration calculation, and score the health of the UAV.

[0029] Step 2: For the UAV health prediction method, a dynamic weighted cross-dimensional correlation matrix will be constructed for fault prediction. The overall system status will be detected by constructing the correlation matrix in the dynamic weighted cross-dimensional correlation matrix, dynamically learning and adjusting it, and calculating the comprehensive health entropy value.

[0030] Step 3: For the UAV health prediction method, a combination of physical model and data-driven approach will be used to predict the remaining useful life, anchor the trend based on the physical model, and correct the deviation of the data-driven approach.

[0031] Step 4: For the aforementioned UAV health prediction method, example data and output will be provided, and radar charts will be used to display different dimensions of the state (e.g., Figure 1 and Figure 2 (As shown).

[0032] The baseline is established after the new aircraft is put into operation or after major overhaul. Calibration flights are conducted under standard conditions to collect high-frequency vibration signals from the structure (IMU) (sampling rate ≥1kHz). The signals are then subjected to high-resolution Fourier transform (FFT) to generate a standard frequency domain energy distribution spectrum (F0) of the UAV in a healthy state. This spectrum is its unique "health fingerprint".

[0033] Real-time monitoring and migration calculation also generate the frequency domain energy distribution map (Ft) of the current flight after each flight; calculate the migration degree (D) between the current map and the healthy baseline, and use spectral correlation density (SCD) or Wasserstein distance to quantify the difference between the two distributions. The index is extremely sensitive to small frequency band energy changes. The calculation formula is: Dt=SCD(F0,Ft).

[0034] The smaller the health score mobility (D), the higher the health score. The structural health score Sstructure is calculated based on the mobility score: Sstructure=5*exp(-λ*Dt) (where λ is the attenuation coefficient, calibrated based on historical data). This method can detect early imbalances or micro-structural cracks that are completely undetectable by traditional time-domain indicators (such as total vibration).

[0035] The association matrix is ​​constructed by defining a 6x6 health status association matrix M, with elements M... ij This represents the weight of the influence of the health status of the i-th dimension on the j-th dimension (e.g., the weight of the influence of dynamic system anomalies on structural health); the initial weights are determined by domain expert knowledge and historical failure data.

[0036] The matrix weights for dynamic learning and adjustment are not fixed. The system continuously monitors and uses Granger causality tests or transition entropy analysis to dynamically update the weight matrix M. For example, if historical data repeatedly shows "battery voltage drop" (i) followed by frequent "navigation module restart" (j), the system will automatically increase M. ij The weight.

[0037] The calculation of the comprehensive health entropy (H) introduces the concept of "health entropy" to assess the overall disorder of the system. First, the independent scores Si of each dimension are calculated, and then the scores are corrected by the correlation matrix M to obtain the comprehensive health entropy H. The calculation formula is: H=Σ(Pi*log(Pi)), where Pi=(ΣjMji*Sj) / ΣSj. The lower the entropy value H, the more stable and healthy the overall state of the system. An increase in entropy value indicates that the system is in disorder and the risk of failure is increased.

[0038] For components with well-defined degradation models, such as batteries, physical models and data-driven approaches no longer rely solely on long short-term memory networks for prediction. Instead, they employ physical models to anchor trends: using empirical degradation models based on electrochemical models (such as bi-exponential models) to predict macroscopic degradation curves, and using data-driven correction of biases: using Kalman filters to take real-time monitored health indicators (such as internal resistance growth) as observations to correct the prediction results of the physical model in real time. This method utilizes the stability of the physical model and incorporates feedback from real-time data, resulting in more reliable and interpretable prediction results.

[0039] The example data and output data show that after this flight, the power system health score Spower dropped sharply from 4.5 to 3.0 (due to increased motor torque fluctuations), while the structural system health score Sstructure remained at 4.8 (time-domain indicators are normal). System analysis: Querying the dynamic correlation matrix revealed that Mpower->structure (the weight of the power system's influence on the structure) was very high, at 0.8. The system immediately calculated that the abnormal fluctuations in the power system, even if they do not immediately cause excessive structural vibration, will continue to cause latent fatigue damage to the boom.

[0040] The example data and output include the warning message: "Abnormal fluctuations in the power system have been detected, which are expected to cause the structural health to drop to the warning level (<2.5 points) within the next 15 takeoff and landing cycles"; and maintenance recommendations: "It is recommended to immediately check the balance of the motor and propeller, and to perform non-destructive testing on the critical connection structures."

[0041] working methods

[0042] Multi-source data acquisition and preprocessing

[0043] Data source:

[0044] Battery: Voltage, current, temperature, internal resistance, and number of charge / discharge cycles.

[0045] Structure: Triaxial accelerometer (vibration standard deviation, kurtosis), gyroscope (angular velocity standard deviation, kurtosis).

[0046] Communication: Signal strength (RSSI), packet loss rate, communication delay.

[0047] Navigation: GPS positioning accuracy (HDOP), number of satellites, and standard deviation of positioning drift.

[0048] Hardware: CPU temperature, memory usage, and memory read / write error rate.

[0049] Power: motor speed, ESC output current, thrust efficiency.

[0050] Sampling frequency:

[0051] High-frequency signals (such as vibration): 100Hz

[0052] Intermediate frequency signals (such as voltage and current): 10Hz

[0053] Low-frequency signals (such as temperature, communication quality): 1Hz

[0054] Preprocessing:

[0055] Denoising: Wavelet transform (Daubechies5 wavelet) is used to denoise the vibration and current signals.

[0056] Normalization: Use the standard deviation standardization method (Z-score standardization method) to bring all indicators to the same dimension.

[0057] Missing value handling: use linear interpolation or fill with values ​​from previous or next frame.

[0058] Health Indicator Calculation and Scoring Rules

[0059] The health score for each dimension is based on multiple sub-indicators and calculated through a rule engine, without relying on machine learning models, ensuring interpretability and stability.

[0060] Battery health dimensions

[0061] Indicators: Voltage smoothness (discharge curve variance), maximum voltage drop, internal resistance growth rate, Scoring rules: Each indicator is scored on a five-point scale (e.g., ...). Figure 3 As shown in the figure, the lowest score is taken as the battery health score;

[0062] Structural health dimension

[0063] Time-domain metrics: vibration standard deviation, vibration kurtosis, angular velocity standard deviation, angular velocity kurtosis

[0064] Frequency domain metrics: Perform a Fourier transform (FFT) on the vibration signal to calculate the energy percentage (Eband) of a specific frequency band (e.g., 100-200Hz).

[0065] Scoring rules: Each indicator is scored individually, and the lowest score among all indicators is taken (e.g., ...). Figure 4 (as shown)

[0066] Communication health dimensions

[0067] Weighted scoring method is used:

[0068] Scomm=0.5*ScoreRSSI+0.3*(1-ScoreLoss)+0.2*(1-ScoreDelay)

[0069] The scoring rules for each sub-index are similar (e.g., signal strength > -70dBm is 5 points, < -90dBm is 1 point).

[0070] Other dimensions

[0071] A similar method is used to set a threshold table for scoring, and finally take the lowest score or weighted score according to the dimension attribute;

[0072] Multi-dimensional fusion and health radar chart generation

[0073] The scores for each dimension are aggregated into a six-dimensional health vector, which is visualized using a radar chart to intuitively display health status (e.g., ...). Figure 5 (As shown).

[0074] Trend Forecasting and Maintenance Recommendations

[0075] Trend analysis: Calculate the score slope for each dimension based on historical data (such as the battery score trend over the past 10 flights).

[0076] Warning rules:

[0077] If any dimension score is ≤2, an immediate alert will be issued;

[0078] If the score drops for three consecutive times, a "performance degradation" message will be displayed.

[0079] By combining changes in scores across multiple dimensions (such as a decrease in battery score and a decrease in temperature score), potential failures can be predicted.

[0080] Example data and output

[0081] Example data (a single flight segment) (e.g.) Figure 6 (as shown)

[0082] Health score output:

[0083] Battery health score = min(3,3,4) = 3 (passing grade)

[0084] Structural health score = min(2,2) = 2 (early warning)

[0085] Communication health score = 0.5*3 + 0.3*(1-0.2) + 0.2*(1-0.1) = 3.2 (rounded to 3)

[0086] Hardware health score = 3 (high CPU temperature)

[0087] Radar chart and maintenance recommendations:

[0088] Radar chart: The structure dimension scored the lowest (2 points), and the battery and hardware dimensions scored relatively low (3 points).

[0089] Maintenance recommendations:

[0090] "Structural health warning: It is recommended to check the motor mounting screws and the balance of the blades."

[0091] "Battery performance degradation: Deep charge-discharge calibration is recommended."

[0092] "CPU temperature is too high: We recommend cleaning the heat dissipation holes."

[0093] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the health of unmanned aerial vehicles (UAVs) based on collaborative analysis of multi-source time-series signals, characterized in that, The drone health prediction method includes the following steps: S1: For the UAV health prediction method, perform individualized frequency domain health baseline migration, conduct health assessment on the structure of individualized frequency domain health baseline migration, control baseline establishment, perform real-time monitoring and migration calculation, and score the health of the UAV. S2: For the aforementioned UAV health prediction method, a dynamic weighted cross-dimensional correlation matrix will be constructed for fault prediction. Furthermore, the overall system status will be detected as stable through the construction, dynamic learning and adjustment of the correlation matrix in the dynamic weighted cross-dimensional correlation matrix, and the calculation of the comprehensive health entropy value. S3: The UAV health prediction method will combine physical model and data-driven approaches to predict the remaining useful life, anchor trends based on the physical model, and correct deviations for the data-driven approach. S4: For the aforementioned UAV health prediction method, example data and output will be provided, and different dimensions of the status will be displayed in conjunction with radar charts.

2. The method for UAV health prediction based on multi-source time-series signal collaborative analysis as described in claim 1, characterized in that: The baseline is established after the aircraft is new or overhauled, and calibration flights are conducted under standard conditions to collect high-frequency vibration signals of the structure with a sampling rate ≥1kHz. The signals are then subjected to high-resolution Fourier transform to generate a standard frequency domain energy distribution spectrum F0 of the UAV in a healthy state, which is unique.

3. The method for UAV health prediction based on multi-source time-series signal collaborative analysis according to claim 2, characterized in that: The real-time monitoring and migration calculation also generates a frequency domain energy distribution map Ft for each flight; the migration degree D between the current map and the healthy baseline is calculated, and the spectral correlation density SCD is used to quantify the difference between the two distributions. The index is extremely sensitive to small frequency band energy changes. The calculation formula is: Dt=SCD(F0,Ft).

4. The method for UAV health prediction based on multi-source time-series signal collaborative analysis as described in claim 3, characterized in that: The smaller the health score mobility D, the higher the health score. The structural health score Sstructure is calculated based on the mobility: Sstructure = 5 * exp(-λ * Dt). Where λ is the attenuation coefficient, calibrated based on historical data, this method can detect early imbalances or microstructural cracks that are completely undetectable by traditional time-domain indicators.

5. The method for UAV health prediction based on multi-source time-series signal collaborative analysis according to claim 1, characterized in that: The correlation matrix construction defines a 6x6 health status correlation matrix M, with elements M... ij This represents the weight of the influence of the health status in the i-th dimension on the j-th dimension; The initial weights are determined by domain expert knowledge and historical failure data.

6. The method for UAV health prediction based on multi-source time-series signal collaborative analysis according to claim 1, characterized in that: The weights of the dynamically learned and adjusted matrix are not fixed. The system continuously monitors and uses Granger causality tests or transition entropy analysis to dynamically update the weight matrix M. For example, if historical data shows multiple instances of sudden battery voltage drops (denoted as i) and subsequent frequent navigation module restarts (denoted as j), the system will automatically increase M. ij The weight.

7. The UAV health prediction method based on multi-source time-series signal collaborative analysis as described in claim 6, characterized in that: The comprehensive health entropy value H is calculated by introducing the concept of health entropy to assess the overall disorder of the system. First, the independent scores Si of each dimension are calculated, and then corrected by the correlation matrix M to obtain the comprehensive health entropy value H. The calculation formula is: H=Σ(Pi*log(Pi)). Where Pi = (ΣjMji*Sj) / ΣSj; the lower the entropy value H, the more stable and healthy the overall system state; an increase in entropy value indicates that the system is in chaos and the risk of failure is increased.

8. The UAV health prediction method based on multi-source time-series signal collaborative analysis according to claim 1, characterized in that: For components with well-defined degradation models, such as batteries, the physical model and data-driven approach no longer simply use long short-term memory networks for prediction. Instead, a physical model is used to anchor the trend: an empirical degradation model based on an electrochemical model is used to predict the macroscopic degradation curve, and data-driven bias correction is used: a Kalman filter is used to use real-time monitored health indicators as observations to correct the prediction results of the physical model in real time.

9. The method for predicting UAV health based on multi-source time-series signal collaborative analysis according to claim 1, characterized in that: The example data and outputs include an early warning message: abnormal fluctuations in the power system are detected, and it is expected that the structural health will decline to the warning level within the next 15 takeoff and landing cycles, at which point maintenance will be recommended.

Images