Unmanned aerial vehicle battery endurance flight capability prediction system

Abstract

The invention relates to the technical field of unmanned aerial vehicles, and discloses an unmanned aerial vehicle battery endurance flight capability prediction system, which comprises a multi-dimensional data acquisition module, a feature mapping module, a prediction module, an optimization module and a feedback optimization module, and can be additionally provided with an early warning module. The multi-dimensional data acquisition module acquires battery data and cleans the battery data to generate standardized data; the feature mapping module maps the data to a feature space, and generates a feature sequence cluster containing a multi-dimensional association relationship by using a time sequence segmentation algorithm; the prediction module divides prediction intervals based on a support vector machine algorithm and extracts prediction indexes; the optimization module generates an endurance prediction strategy by predicting and optimizing the network model; and the feedback optimization module performs multi-source data fusion optimization and outputs a prediction instruction. The early warning module can associate the prediction instruction with the battery health degree, output a grading early warning signal and trigger a response mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs), and in particular to a UAV battery life flight capability prediction system. Background Art

[0002] With the widespread application of drone technology in aerial photography, logistics, agriculture, inspection, and other fields, drone endurance has become a key factor limiting its application scope and operational efficiency. A drone's endurance is directly dependent on battery performance, which is affected by a variety of complex factors, including the battery's state (voltage, current, temperature, remaining capacity, etc.), flight parameters (flight speed, altitude, attitude, payload, etc.), and environmental factors (temperature, wind speed, air pressure, etc.). Existing technologies for predicting drone battery endurance have numerous shortcomings, making them difficult to meet practical application requirements.

[0003] Traditional range prediction methods typically analyze data based on a single dimension, such as the battery's remaining capacity or a simple discharge curve. These methods ignore the multi-dimensional relationships between battery status, flight parameters, and environmental factors, resulting in low prediction accuracy. For example, the actual battery discharge characteristics vary significantly at different flight speeds and ambient temperatures, making it impossible to accurately predict range under current operating conditions based solely on historical capacity data.

[0004] Existing prediction models often lack adaptive adjustment capabilities and struggle to cope with the complex and changing conditions experienced during drone flight. During flight, the battery's state of health gradually degrades with the number of charge and discharge cycles, and flight missions and environmental conditions can change in real time. The fixed parameters and prediction rules of traditional models are unable to adapt to these changes, leading to increasing prediction errors.

[0005] Inadequate data preprocessing and feature extraction also affect prediction accuracy. In actual flight, collected battery data may contain noise, missing data, and duplication. Without effective cleaning and preprocessing, this can lead to biases in subsequent analysis. Furthermore, extracting features with temporal and multidimensional correlations from massive amounts of time series data is a challenge facing existing technologies. Traditional feature extraction methods typically employ simple statistical analysis or fixed-window processing, failing to fully tap into the deep features and dynamic patterns within the data.

[0006] The existing system lacks an effective optimization and feedback mechanism. The prediction model struggles to continuously optimize based on actual feedback data during operation, preventing the formation of a closed-loop prediction and optimization system. Consequently, the system's prediction performance cannot continuously improve as data accumulates. Furthermore, the application of prediction results is relatively limited, lacking effective integration with battery maintenance and flight adjustment strategies, making it impossible to proactively manage and optimize the drone's endurance.

[0007] Existing UAV battery life flight capability prediction technologies have problems such as insufficient multi-dimensional data correlation analysis, poor model adaptability, imprecise data preprocessing, insufficient feature extraction, and imperfect optimization feedback mechanism. There is an urgent need for a prediction system that can comprehensively consider multi-dimensional factors and has adaptive adjustment and optimization capabilities to improve prediction accuracy and reliability and meet the application needs of UAVs in complex scenarios. Summary of the Invention

[0008] The purpose of the present invention is to provide a UAV battery life flight capability prediction system to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a UAV battery life flight capability prediction system, the system comprising:

[0010] A multi-dimensional data acquisition module is used to obtain a battery data set within a preset time range during the flight of the UAV, and pre-process the battery data set according to a preset data cleaning strategy to generate standardized battery data;

[0011] A feature mapping module, configured to map the standardized battery data into a feature space and generate a feature sequence cluster based on a time series segmentation algorithm, wherein the feature sequence cluster includes a multi-dimensional correlation between battery status, flight parameters, and environmental factors;

[0012] A prediction module is used to perform adaptive parameter adjustment on the feature sequence cluster according to preset prediction rules, divide the prediction intervals by a support vector machine algorithm, and extract a set of prediction indicators within each interval;

[0013] An optimization module is used to construct a decision unit including a prediction optimization network model, and iteratively update the weights of the prediction optimization network model using the prediction indicator set to generate a range prediction strategy;

[0014] The feedback optimization module is used to perform multi-source data fusion optimization on the endurance prediction strategy according to a preset sensitivity analysis model, and output the UAV battery endurance flight capability prediction instruction.

[0015] Preferably, the generating of feature sequence clusters based on a time series segmentation algorithm includes:

[0016] Extracting timestamps, battery IDs, and operating parameters from the standardized battery data to construct a time series matrix;

[0017] Using a window sliding segmentation algorithm to segment the time series matrix to generate a set of time series sub-segments;

[0018] Using a support vector machine-based classification model to perform pattern recognition on the time series sub-segment set and eliminate abnormal data points;

[0019] The remaining sub-segments are integrated at multiple levels through an aggregation algorithm to generate feature sequence clusters with temporal correlation.

[0020] Preferably, the adaptive parameter adjustment includes:

[0021] Based on the historical battery data distribution in the forecast interval, calculate the mean and standard deviation of key forecast indicators in each interval;

[0022] Based on the moving average algorithm, the mean and standard deviation are smoothed and corrected to generate a dynamic parameter benchmark;

[0023] The rule engine is used to adjust the interval boundary of the dynamic parameter benchmark to generate an adaptive parameter interval.

[0024] Preferably, the support vector machine algorithm divides the prediction interval, including:

[0025] The objective function is defined as a weighted balance between battery consumption and flight stability within the prediction interval;

[0026] The constraints are set as flight path continuity, battery capacity limit and environmental condition tolerance;

[0027] The objective function is solved by the gradient descent method, and the optimal prediction interval partitioning scheme is output.

[0028] Preferably, the prediction optimization network model performs iterative weight updating, including:

[0029] Inputting the prediction indicator set into the input layer of the prediction optimization network model, and extracting feature vectors using a recurrent neural network;

[0030] sorting the feature vectors by weight using a priority allocation mechanism to generate a parameter priority sequence;

[0031] The back propagation algorithm is combined with the particle swarm optimization algorithm to update the network weights and the decision parameters of the decision unit.

[0032] Preferably, the data cleaning strategy includes:

[0033] Identifying duplicate data segments in the battery data set and removing duplicates based on a timestamp alignment rule;

[0034] Detecting missing data points and interpolating the missing data points using linear interpolation;

[0035] The interpolated data were normalized to generate standardized battery data with a mean of zero and a variance of one.

[0036] Preferably, the sensitivity analysis model includes:

[0037] Construct a multivariate sensitivity calculation framework based on variance decomposition method to quantify the sensitivity of the predicted instructions to various operating parameters;

[0038] Generate parameter perturbation sample sets through random sampling and calculate the contribution of each parameter;

[0039] Parameters with contributions higher than the preset threshold are selected as core variables for optimized prediction.

[0040] Preferably, the multi-source data fusion optimization includes:

[0041] Integrate battery sensor data, flight log data, and weather forecast data to build a heterogeneous data fusion matrix;

[0042] Using principal component analysis algorithm to perform dimensionality reduction processing on the heterogeneous data fusion matrix and extract the main eigenvector;

[0043] The main feature vector is input into the feedback optimization module to generate an optimized prediction instruction.

[0044] Preferably, the system further comprises:

[0045] Constructing an early warning module including a battery health assessment function, performing correlation analysis between the prediction instructions and the battery health, and outputting a graded early warning signal;

[0046] The preset response mechanism is triggered according to the graded warning signal to generate battery maintenance or flight adjustment instructions.

[0047] Preferably, the generation of the graded warning signal includes:

[0048] Calculating a real-time health score based on the battery health assessment function;

[0049] Classifying the health scores through a threshold comparison algorithm to generate early warning signals of different levels;

[0050] The event triggering mechanism is used to map the warning signal to the corresponding response instruction.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The multi-dimensional data acquisition module uses a preset data cleaning strategy to deduplicate, interpolate, and normalize the raw battery data, effectively resolving data noise, missing data, and inconsistency issues, generating standardized battery data and providing a high-quality data foundation for subsequent analysis, avoiding prediction biases caused by data quality issues. The feature mapping module maps standardized data to feature space and uses a time series segmentation algorithm to construct a feature sequence cluster containing multi-dimensional correlations between battery status, flight parameters, and environmental factors. By extracting timestamps, battery IDs, and operating parameters to construct a time series matrix, combined with window sliding segmentation, support vector machine classification, and aggregation algorithms, multi-level feature extraction of the data is achieved, fully tapping the hidden temporal correlations and multi-dimensional interaction patterns in the data, laying the foundation for accurate predictions.

[0053] The prediction module divides prediction intervals based on a support vector machine algorithm. By defining a weighted balance objective function between battery consumption and flight stability, and incorporating constraints such as flight path continuity and battery capacity limitations, it uses gradient descent to find the optimal interval partitioning scheme, achieving dynamic range prediction of flight endurance. Furthermore, through an adaptive parameter adjustment mechanism, the mean and standard deviation of key indicators are calculated based on historical data distribution. This is combined with a moving average algorithm, smoothing corrections, and rule engine adjustments to generate dynamic parameter baselines and adaptive parameter intervals. This allows the prediction model to adapt in real time to changes in battery status and flight conditions, improving prediction accuracy and robustness.

[0054] The prediction optimization network model constructed by the optimization module extracts feature vectors through a recurrent neural network, uses a priority allocation mechanism to sort weights, and combines backpropagation and particle swarm optimization algorithms to iteratively update weights, achieving dynamic optimization of decision-making unit parameters. This deep learning-based optimization mechanism continuously learns complex patterns in data, continuously improving prediction performance as data accumulates, forming a closed-loop optimization system.

[0055] The feedback optimization module uses sensitivity analysis models and multi-source data fusion optimization to quantify the sensitivity of various operating parameters to prediction results, screen core variables, and integrate battery sensor data, flight log data, and weather forecast data for dimensionality reduction and extraction of principal eigenvectors, further improving the accuracy and reliability of prediction instructions. The fusion of multi-source data enables the system to comprehensively consider real-time environmental changes and flight status, enhancing its adaptability to complex scenarios.

[0056] The early warning module further expands the system's application value. By constructing a battery health assessment function, it correlates prediction commands with battery health, generates graded early warning signals, and triggers corresponding response mechanisms, enabling real-time monitoring and proactive management of battery status. This not only provides early warning of potential battery failures and prevents flight accidents, but also optimizes drone operation strategies by generating battery maintenance or flight adjustment commands, extending battery life and improving flight safety and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a working principle diagram of the UAV battery endurance flight capability prediction system of the present invention;

[0058] Figure 2 This is the design diagram of the feature mapping module;

[0059] Figure 3 This is a diagram showing the working principle of the data cleaning strategy;

[0060] Figure 4 This is a working principle diagram of the sensitivity analysis model. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] See also Figures 1-4 The present invention relates to a UAV battery life flight capability prediction system, which includes: a multi-dimensional data acquisition module, a feature mapping module, a prediction module, an optimization module and a feedback optimization module. The specific implementation steps are as follows:

[0063] Multi-dimensional data acquisition module: This module acquires a battery data set within a preset time range during the drone's flight. The data covers basic parameters such as battery voltage, current, temperature, and remaining capacity, as well as related data such as flight duration, flight speed, and flight altitude. After acquiring the data, a preset data cleaning strategy is implemented: first, duplicate data segments in the battery data set are identified and deduplication is performed based on timestamp alignment rules; second, missing data points are detected and interpolated using linear interpolation. For example, if battery voltage data at a certain moment is missing, it is filled in by linear fitting using the voltage values at adjacent time points; finally, the interpolated data is normalized, and a standardization algorithm is used to ensure that the data meets the distribution characteristics of zero mean and one variance, generating standardized battery data.

[0064] The feature mapping module maps standardized battery data into a feature space. The specific process involves extracting the timestamp, battery ID, and operating parameters (such as voltage, current, and flight altitude) from the standardized battery data. This module constructs a time series matrix with the timestamp sequence as the row dimension and the various data parameters as the column dimension. A window sliding segmentation algorithm is used to segment the time series matrix, using a fixed-length sliding window (e.g., 5 minutes) with a fixed step size (e.g., 1 minute) to generate a set of time series subsegments. A support vector machine-based classification model is used to perform pattern recognition on the set of time series subsegments. The trained classification model identifies and removes anomalous data points (e.g., those with voltage fluctuations outside the normal range). The remaining subsegments are then aggregated using a multi-level aggregation algorithm (e.g., hierarchical clustering) to form clusters of temporally correlated feature sequences based on chronological order and data correlation. These clusters contain multi-dimensional relationships between battery status, flight parameters, and environmental factors (e.g., ambient temperature and wind speed).

[0065] The prediction module adaptively adjusts parameters for feature sequence clusters based on pre-set prediction rules. First, based on the historical battery data distribution within the prediction interval, the mean and standard deviation of key prediction indicators (such as battery consumption rate and flight stability indicators) are calculated for each interval. A moving average algorithm is used to smooth the mean and standard deviation, for example, using a five-point moving average to eliminate short-term fluctuations, generating a dynamic parameter benchmark. A rule engine is used to adjust the interval boundaries of the dynamic parameter benchmark, and adaptive parameter intervals are generated based on real-time data features. The prediction interval is divided using a support vector machine algorithm. The objective function is defined as a weighted balance between battery consumption and flight stability within the prediction interval, with weights set based on actual flight requirements (e.g., a higher weight is given to battery consumption when flight endurance is prioritized). Constraints are set as flight path continuity, battery capacity limits, and environmental tolerances (e.g., ambient temperature must not fall below -10°C or exceed 40°C). A gradient descent algorithm is used to solve the objective function, iteratively optimize the parameters, output the optimal prediction interval partitioning scheme, and extract a set of prediction indicators within each interval (e.g., predicted remaining battery capacity and flight time).

[0066] Optimization Module: Constructs a decision-making unit (DMU) containing a predictive optimization network model, which utilizes a recurrent neural network (RNN) architecture. A set of prediction indicators is fed into the DMU's input layer. The hidden layer nodes of the RNN are used to extract feature vectors, capturing long-term dependencies in the time series. Feature vectors are weighted and sorted using a priority allocation mechanism. A parameter priority sequence is generated based on the impact of each feature on the range prediction. For example, remaining battery capacity takes precedence over flight speed. A backpropagation algorithm combined with a particle swarm optimization algorithm is used to update network weights. The backpropagation algorithm calculates the prediction error and propagates it layer by layer to update the weights. The particle swarm optimization algorithm simulates the foraging behavior of bird flocks to globally search for optimal weights, thereby updating the DMU's decision parameters and generating a range prediction strategy.

[0067] The feedback optimization module optimizes the range prediction strategy using multi-source data fusion based on a preset sensitivity analysis model. This sensitivity analysis model constructs a multivariate sensitivity calculation framework based on variance decomposition to quantify the sensitivity of the prediction command to various operating parameters (such as battery voltage and ambient wind speed). A set of parameter perturbation samples is generated through random sampling, and the contribution of each parameter is calculated. For example, a large number of parameter samples are generated through Monte Carlo simulation, and the impact of each parameter change on the prediction result is analyzed. Parameters with contributions exceeding a preset threshold are selected as core variables for optimization and prediction. During the multi-source data fusion optimization process, battery sensor data, flight log data, and weather forecast data are integrated to construct a heterogeneous data fusion matrix, with different data sources in the matrix representing different dimensional features. Principal component analysis is used to reduce the dimensionality of the heterogeneous data fusion matrix, extracting the principal eigenvector that explains the majority of the data variance, reducing the data dimension while retaining key information. This principal eigenvector is then fed into the feedback optimization module, where an optimization algorithm generates optimized prediction commands, ultimately outputting the prediction command for the drone's battery range flight capability.

[0068] The present invention will be further described below in conjunction with Examples 1 to 5:

[0069] Embodiment 1:

[0070] In the feature mapping module, the specific implementation method of generating feature sequence clusters based on the time series segmentation algorithm is as follows:

[0071] The timestamp, battery ID, and operating parameters are extracted from the standardized battery data output by the multidimensional data acquisition module. The timestamp identifies the time of data collection, accurate to the second or millisecond level, ensuring the continuity and accuracy of the time series. The battery ID uniquely identifies each drone battery, facilitating the tracking and analysis of historical data for each individual battery. Operating parameters include battery status parameters (such as voltage, current, remaining capacity, and temperature), flight parameters (such as altitude, speed, acceleration, and heading angle), and environmental parameters (such as ambient temperature, wind speed, and direction). These parameters are stored as structured data, for example, in a table format, with each row corresponding to the sampled data at a time point and each column corresponding to a specific parameter dimension. After extracting this data, it is arranged in chronological order to construct a time series matrix. The rows of this matrix are timestamp sequences, and the columns are the parameter dimensions, forming a two-dimensional data structure. This ensures that all parameter data at each time point has a unique position index in the matrix, facilitating subsequent segmentation and pattern recognition.

[0072] The time series matrix is segmented using a sliding window segmentation algorithm. The core of this algorithm is to set a fixed-length sliding window and step size. The window length should be set based on both the temporal correlation of the UAV flight data and computational efficiency. For example, it can be set to 5 minutes, 10 minutes, etc., depending on the speed of battery status changes and the characteristics of the flight mission in the actual application scenario. The step size is typically smaller than the window length, for example, 1 minute, 2 minutes, etc., to ensure overlap between adjacent subsegments and avoid data fragmentation. During processing, the sliding window begins at the beginning of the time series matrix and slides backwards in sequence according to the set step size. After each slide, the submatrix within the window is intercepted and defined as a time series subsegment. For example, if the time series matrix contains timestamp data from t1 to tN, the window length is L, and the step size is S, then the first subsegment is from t1 to t1+L-1, the second subsegment is from t1+S to t1+S+L-1, and so on, until the entire time series matrix is covered, generating a set of time series subsegments. Each sub-segment contains multi-dimensional parameter data within a continuous time interval, which can reflect the local characteristics of the drone's battery and flight status during that time period.

[0073] A support vector machine-based classification model performs pattern recognition on a collection of time series subsegments to eliminate anomalous data points. The support vector machine classification model is pre-trained using historically labeled data. The training data contains both normal and abnormal data subsegments. The abnormal data subsegments cover situations such as sudden battery voltage drops, sudden current changes, abnormal flight altitude fluctuations, and ambient temperature exceeding acceptable limits. During training, the model learns the distribution differences between normal and abnormal data in feature space and constructs a classification hyperplane. When processing a collection of time series subsegments, each subsegment is input into the trained support vector machine model, which outputs a determination of whether the subsegment is anomalous. Subsegments identified as anomalous require further analysis to identify the anomalous data points. For example, if the battery voltage in a subsegment suddenly drops below the minimum safety threshold for a short period of time and the flight altitude fluctuates dramatically at the same time, the relevant data points in that subsegment are considered anomalous. For subsegments containing outliers, two approaches can be used: one is to directly remove the entire outlier subsegment, which is suitable for situations where the outliers account for a large proportion of the subsegment; the other is to remove only the outliers in the subsegment and retain the normal data points, which is suitable for situations where the outliers are relatively isolated. Through these processes, outliers in the time series subsegment set are removed, improving the quality and reliability of the data.

[0074] The remaining subsegments are then aggregated multi-level using an aggregation algorithm to generate clusters of temporally correlated feature sequences. This aggregation algorithm can employ a hierarchical clustering algorithm, whose basic principle is to gradually merge similar subsegments to form higher-level clusters. During the aggregation process, the similarity between each subsegment is first calculated. This similarity can be calculated using metrics such as Euclidean distance and cosine similarity, which measure the proximity of subsegments in a multidimensional parameter space. For example, for two subsegments, the Euclidean distance between them is calculated for parameters such as voltage, current, and flight altitude. The smaller the distance, the higher the similarity. The subsegments with the highest similarity are then merged into a small cluster. The similarity between this small cluster and other subsegments or clusters is then calculated. This merging process continues until all subsegments have been aggregated into clusters at different levels. Through multi-level aggregation, temporally ordered feature sequence clusters are formed. Each cluster contains multiple related subsegments. This cluster not only reflects the local characteristics within a single subsegment, but also, through the temporal correlations between subsegments, reflects the temporal evolution and interactions of battery status, flight parameters, and environmental factors. For example, a feature sequence cluster may contain multiple continuous sub-segments, which reflect the associated changes in the battery voltage gradually decreasing, the current increasing, and the ambient temperature decreasing with increasing altitude during the drone's climbing process.

[0075] Throughout the feature sequence cluster generation process, timestamps are a key factor in ensuring the temporal relevance of data. The construction of a time series matrix and window sliding segmentation ensure the temporal orderliness and continuity of sub-segments. The support vector machine's pattern recognition eliminates abnormal fluctuations in the time series, allowing the remaining sub-segments to more realistically reflect normal flight status and battery characteristics. The hierarchical clustering algorithm's multi-level integration, based on temporal order and data similarity, aggregates sub-segments into logically related feature sequence clusters, providing subsequent prediction modules with feature data containing multi-dimensional correlations. These steps are closely linked, and through structured processing and feature extraction of standardized battery data, the conversion from raw data to multi-dimensional feature sequences with temporal relevance is achieved, laying a solid data foundation for predicting drone battery endurance and flight capabilities.

[0076] Example 2:

[0077] The adaptive parameter adjustment process in the prediction module is as follows:

[0078] Based on the distribution of historical battery data within the prediction interval, a statistical analysis is performed on key prediction indicators within each preset prediction interval. Key prediction indicators include, but are not limited to, battery consumption rate (battery capacity reduction per unit time), flight stability indicators (such as altitude fluctuation amplitude and heading angle change rate), and environmental parameter influencing factors (such as the impact of ambient temperature on battery performance). These indicators are calculated using historical flight data. For example, the battery consumption rate can be calculated by dividing the difference between the initial and final battery capacities within a prediction interval by the flight time. For each prediction interval, a certain number of historical data samples are collected (e.g., data from the past 50 or 100 missions of the same type for that interval), and the mean and standard deviation of each indicator are calculated. The mean reflects the average level of the indicator in the historical data. For example, if the mean battery consumption rate for a certain interval is X units of capacity / minute, it means that under typical flight conditions within that interval, the battery consumes X units of capacity per minute on average. The standard deviation measures the dispersion of the data. If the standard deviation is Y, the actual data fluctuates roughly within the range of X±Y around the mean X. Through the mean and standard deviation, we can initially understand the distribution characteristics of each indicator in historical data and provide a benchmark for subsequent parameter adjustments.

[0079] The mean and standard deviation are smoothed using a moving average algorithm. The core concept of the moving average algorithm is to use local averaging of time series data to eliminate short-term fluctuations and highlight long-term trends. Specifically, a moving window length is set (e.g., 5 or 10 historical data samples) and a moving average is performed on the mean and standard deviation series. For example, for the mean series M1, M2, M3, ..., Mn, a moving window length of 5 is used. The first smoothed mean is (M1+M2+M3+M4+M5) / 5, the second smoothed mean is (M2+M3+M4+M5+M6) / 5, and so on. This results in a smoothed dynamic parameter benchmark series. Similarly, the standard deviation series is subjected to the same moving average process. This process filters out abnormal fluctuations caused by accidental factors (such as brief environmental disturbances or sudden drone operations), ensuring that the parameter benchmark better reflects the long-term stable trend of the data. For example, if the mean battery consumption rate in a certain prediction interval increases abnormally in individual samples due to sudden strong winds, the impact of this abnormal value on the parameter benchmark will be weakened after smoothing through moving average, thereby obtaining a more reliable dynamic parameter benchmark.

[0080] A rule engine is used to adjust the interval boundaries of dynamic parameter benchmarks to generate adaptive parameter intervals. The rule engine is a predefined logical set containing a series of adjustment rules based on factors such as flight conditions, environmental conditions, and mission requirements. For example, when the current flight mission is detected as a long-duration mission, the rule engine automatically adjusts the interval boundaries of the battery consumption rate, appropriately relaxing the upper limit to allow for increased battery consumption and extended flight time while ensuring safety. When the ambient temperature falls below the battery's optimal operating temperature range, the rule engine tightens the interval boundaries of the flight stability indicator, requiring the drone to maintain a more stable flight attitude to reduce flight risks that may be caused by battery performance degradation at low temperatures. During the specific adjustment process, the rule engine first obtains current real-time data, including flight phases (such as takeoff, cruise, and landing), environmental parameters (such as temperature and wind speed), and mission type (such as inspection and logistics delivery). Then, based on the pre-set rule matching conditions, it determines the parameters that need to be adjusted and the adjustment range. For example, if the current ambient temperature is T°C, which is lower than the battery's optimal operating temperature (T0°C), the upper and lower bounds of the dynamic parameter benchmark will be adjusted accordingly, based on the rule: "For every 1°C drop below T0°C, the upper limit of the battery consumption rate decreases by A%, and the lower limit of the flight stability index increases by B%. This adjusted parameter range can adapt to current flight conditions and mission requirements in real time, making the prediction model's parameter settings more flexible and accurate."

[0081] The adaptive parameter adjustment process relies on statistical analysis of historical data, establishing an initial parameter baseline through the mean and standard deviation. Smoothing corrections using the moving average algorithm are crucial, eliminating interference from short-term fluctuations and making the baseline more representative. Dynamic adjustment of the rule engine is central, incorporating real-time operating conditions and mission requirements into parameter settings, enabling the predictive model's adaptability. These three steps work together to form a comprehensive process: extracting patterns from historical data, refining them within real-time data, and adjusting them based on actual needs. For example, when a drone performs inspection missions in different seasons, in winter, at low temperatures, adaptive parameter adjustments automatically reduce the expected battery drain rate while imposing higher requirements on flight stability. In summer, at high temperatures, the model appropriately relaxes the relevant parameter ranges to accommodate seasonal variations in battery performance. This adjustment mechanism ensures that the prediction module consistently delivers predictions within reasonable parameter ranges under a variety of complex and changing conditions, improving the accuracy and reliability of the drone's battery endurance predictions and providing high-quality input data for subsequent optimization and feedback optimization modules.

[0082] Throughout the entire process, parameter adjustments are based on the statistical characteristics of historical data and logical judgment of real-time operating conditions, avoiding the limitations of subjective assumptions and fixed thresholds. Through dynamically generated adaptive parameter intervals, the prediction module can more accurately capture the changing patterns of battery life in different scenarios, providing scientific decision-making support for drone flight planning and battery management. For example, when a drone is predicted to enter a specific flight range, the system can provide an early warning of battery consumption based on the adjusted parameter interval, helping ground control personnel make timely decisions on flight path adjustments or battery replacements, thereby improving drone flight safety and mission execution efficiency.

[0083] Example 3:

[0084] The specific implementation of the support vector machine algorithm to divide the prediction interval is as follows:

[0085] The objective function is defined as a weighted balance between battery consumption and flight stability within the prediction interval. The objective function needs to comprehensively consider the dual requirements of battery life and flight stability during the flight of the drone, and the degree of emphasis on both is reflected by setting different weight coefficients. Let the objective function be:

[0086]

[0087] in, Represents the objective function value, which is used to measure the overall performance of the prediction interval; It is a battery consumption indicator that reflects the degree of battery capacity consumption within the prediction interval and can be quantified by the reduction in battery capacity per unit time or the total consumed capacity; It is a flight stability index used to evaluate the stability of the UAV's flight state within the prediction interval. It can be calculated comprehensively through parameters such as the flight altitude fluctuation amplitude and the heading angle change rate. is the weight coefficient, and its value range is ,when When it approaches 1, it indicates that the system is more concerned with minimizing battery consumption and is suitable for long-duration missions; when When it approaches 0, it indicates that the system pays more attention to flight stability and is suitable for tasks with high requirements on flight attitude (such as precision photography, environmental monitoring, etc.).

[0088] Set constraints to ensure the rationality and feasibility of the prediction interval division. Constraints include the following three aspects:

[0089] Flight path continuity constraint: The flight path of the UAV in adjacent prediction intervals must transition smoothly to avoid discontinuities such as sudden turns and altitude changes. Specifically, the starting and ending flight parameters (such as position coordinates, flight speed, heading angle, etc.) of adjacent intervals must meet the preset continuity threshold. For example, the heading angle change in adjacent intervals must not exceed The flight altitude change rate shall not exceed Meters / second.

[0090] Battery capacity constraint: The battery consumption within each prediction interval must not exceed the safe range of the battery's current remaining capacity, that is, the battery's remaining capacity at the end of the prediction interval Need to meet ,in The minimum safe capacity of the battery, set by the battery manufacturer or the drone system, is used to ensure that the drone has sufficient power to complete subsequent necessary operations (such as return, landing, etc.).

[0091] Environmental condition tolerance range constraints: Environmental parameters within the prediction interval (such as ambient temperature , wind speed The ambient temperature must be within the tolerance range of the drone and battery. ,in and They are the minimum and maximum temperatures for normal battery operation; the wind speed must meet ,in The maximum wind speed at which the drone can fly stably.

[0092] After defining the objective function and constraints, the objective function is solved using the gradient descent method to obtain the optimal prediction interval partitioning scheme. The gradient descent method is an iterative optimization algorithm. Its core idea is to gradually approach the minimum (or maximum) value of the objective function by calculating the gradient of the objective function and updating the model parameters in the opposite direction of the gradient. The specific implementation steps are as follows:

[0093] Parameter initialization: Set the initial prediction interval division scheme, for example, evenly divide the entire flight time into several equal-length intervals, and initialize the parameters of each interval (such as start time, end time, initial value of battery consumption index, initial value of flight stability index, etc.).

[0094] Gradient calculation: For each prediction interval, calculate the objective function The gradient of interval partitioning parameters (such as interval length, interval start time, etc.). For example, the gradient It indicates the rate at which the objective function value changes with the interval length, reflecting the impact of adjusting the interval length on the overall performance.

[0095] Parameter update: According to the gradient calculation results, according to the preset learning rate (A small positive number used to control the step size of each parameter update) Update the interval partition parameter. The update formula is:

[0096]

[0097] By continuously adjusting the parameters, the objective function value is gradually reduced (or increased, depending on the optimization goal).

[0098] Convergence judgment: Repeat the above gradient calculation and parameter update steps until the change in the objective function value is less than the preset convergence threshold , or the maximum number of iterations is reached, the algorithm is considered to have converged, and the current prediction interval partitioning scheme is output as the optimal solution.

[0099] In the solution process, it is necessary to deal with the coupling relationship between the objective function and the constraints. For the prediction interval division scheme that violates the constraints, the constraints can be integrated into the objective function through the penalty function method. For example, if the remaining battery capacity in a certain interval is Less than , then add a penalty term to the objective function (in is the penalty coefficient (a large positive number), which increases the objective function value of solutions that violate the constraints, automatically eliminating them from the optimization process. This method effectively ensures that the final prediction interval partitioning solution meets all preset constraints, improving the reliability and practicality of the prediction results.

[0100] The entire process of dividing the prediction interval by the support vector machine algorithm is essentially to find an optimal solution in the multi-dimensional parameter space. This solution not only meets the weighted balance goal of battery consumption and flight stability, but also complies with the constraints such as flight path continuity, battery capacity limit and environmental condition tolerance range. For example, when a drone performs a logistics delivery task, if the task priority is to extend the flight time to cover a longer delivery distance, the weight coefficient can be set to If set to a higher value (such as 0.8), after optimization by gradient descent, the system will tend to divide the prediction interval with lower battery consumption, and appropriately relax the flight stability requirements within the scope allowed by the constraints. On the contrary, if the mission requires the drone to maintain stable flight in complex terrain to ensure the safety of the cargo, then Setting it to a lower value (such as 0.2) will make the system focus more on optimizing flight stability indicators and divide the prediction range for a more stable flight attitude.

[0101] Through this prediction interval partitioning method based on support vector machines and gradient descent, the system can dynamically generate optimal prediction intervals based on different flight missions, environmental conditions, and drone status, providing a scientific basis for subsequent extraction of prediction indicator sets within each interval. Battery consumption and flight stability within each prediction interval are optimized to the optimal balance under current conditions, making predictions of drone battery life and flight capabilities more accurate and providing strong technical support for drone flight control and mission planning. In practical applications, this method can be programmed to achieve automated calculations, combined with real-time collected flight data and environmental parameters, to quickly generate prediction interval partitioning schemes adapted to different scenarios, significantly improving the intelligence level and environmental adaptability of drone systems.

[0102] Embodiment 4:

[0103] The process of iterative weight update of the prediction optimization network model is as follows:

[0104] The set of prediction indicators output by the prediction module is input into the input layer of the prediction optimization network model. This set of prediction indicators contains multi-dimensional data, such as remaining battery capacity, battery consumption rate, flight time, flight altitude, flight speed, ambient temperature, and wind speed for each prediction interval. These indicators are structured in numerical form, for example, forming a vector of length M, with each element corresponding to a specific indicator. For example, for a drone inspection mission, the input layer might receive a vector containing the following indicators: [remaining battery capacity 35%, battery consumption rate in the current interval 0.8 mAh / s, remaining flight time 45 minutes, flight altitude 200 meters, horizontal flight speed 5 m / s, ambient temperature 22°C, wind speed 3 m / s]. The number of nodes in the input layer matches the number of prediction indicators. Each node receives a single indicator data point, which is normalized (e.g., scaled to the range [0, 1]) before being passed to the hidden layer.

[0105] The prediction and optimization network model uses a recurrent neural network (RNN) architecture. The hidden layer contains multiple neurons, each of which receives the output from the input layer and the previous hidden layer through weighted connections. The characteristic of RNN is its ability to capture long-term dependencies in time series data, making it suitable for processing endurance prediction problems with time series characteristics. For example, when a drone is in a high-power flight state (such as frequent climbs) for multiple consecutive prediction intervals, the hidden layer neurons can retain the previous battery consumption pattern through a memory mechanism, thereby more accurately predicting the endurance of subsequent intervals. The hidden layer neurons extract features from the input data and output a feature vector after nonlinear transformation through an activation function (such as tanh or ReLU). This vector contains multi-dimensional correlation features of battery status, flight parameters, and environmental factors.

[0106] A priority allocation mechanism weights the feature vectors to generate a parameter priority sequence. This priority allocation mechanism is designed based on the degree of influence of each feature on the range prediction. For example, remaining battery capacity directly determines whether the drone can complete subsequent missions and therefore takes precedence over flight speed. Ambient temperature, by affecting the battery's internal resistance and thus its consumption rate, takes precedence over flight altitude. In specific implementations, the initial priority weights for each feature can be determined in advance based on domain knowledge or historical data statistics. For example, a weight of 0.3 for remaining battery capacity, 0.25 for battery consumption rate, 0.2 for ambient temperature, 0.15 for flight time, and 0.1 for the sum of all other parameter weights. In real-time calculations, these weights are dynamically adjusted based on the current flight conditions. For example, if the ambient temperature approaches the critical battery operating temperature, the ambient temperature weight is automatically increased to 0.3, and the flight altitude weight is automatically decreased to 0.05. The priority allocation mechanism is implemented using a weight matrix. The feature vector output by the hidden layer is multiplied by the weight matrix to generate a prioritized parameter sequence, highlighting the dominant role of key features in range prediction.

[0107] The network weights are updated using a backpropagation algorithm combined with a particle swarm optimization algorithm. The core of the backpropagation algorithm is to calculate the prediction error and propagate it layer by layer to adjust the weights. The specific process is as follows: First, the feature vector is input into the output layer, which generates a range prediction strategy (such as predicted range, recommended flight speed range, etc.) based on the current weights. The prediction results are then compared with the actual observed data (such as the actual remaining battery capacity after the flight) to calculate loss functions such as the mean squared error. Next, the gradient of each weight with respect to the loss function is calculated from the output layer back to the hidden and input layers. The direction of the gradient indicates the degree to which the weight adjustment affects the error reduction. Finally, the weights are updated based on the gradient direction and the preset learning rate. For example, if the gradient of a weight is positive, it means that increasing the weight will reduce the loss function value, so the weight value is increased along the gradient direction.

[0108] The particle swarm optimization (PSO) algorithm is used to improve the global search capability of weight updates. This algorithm simulates the foraging behavior of a flock of birds, treating each weight combination as a particle in the search space, with particles having position (weight value) and velocity (weight update step size) attributes. Each particle adjusts its movement direction and velocity based on its own historical optimal position and the optimal position of the swarm, thereby finding the optimal weight combination in the search space. For example, when the particle swarm is initialized, it is randomly distributed in the weight space. Each particle calculates the fitness value (such as the inverse of the loss function) of its current position, and the position of the particle with the highest fitness is regarded as the swarm optimal position. Particles update their speed and position using a formula. The speed formula includes a self-awareness term (based on its own historical optimality) and a social awareness term (based on the swarm optimality), while the position formula adjusts the particle coordinates based on the speed. This iterative search by the particle swarm prevents the backpropagation algorithm from getting stuck in local optima, improving the efficiency and accuracy of weight updates.

[0109] In practice, backpropagation and particle swarm optimization are executed alternately. First, the backpropagation algorithm performs several local weight adjustments to bring the model closer to the optimal solution within the current neighborhood. Then, the particle swarm optimization algorithm is activated to search for the global optimal solution within a larger weight space. After updating the particle positions, the algorithm switches back to the backpropagation algorithm for fine-tuning. This combined approach leverages the rapid convergence of the backpropagation algorithm and the global search capabilities of the particle swarm optimization algorithm to avoid local optimality, thereby improving the generalization ability of the predictive optimization network model.

[0110] Taking a drone's agricultural plant protection mission as an example, the input metrics for the prediction and optimization network model might include: the current remaining battery capacity of 40%, a flight speed of 3 m / s in the plant protection operation area (lower than normal inspection speed), an additional battery consumption rate of 1.2 mAh / s due to the power consumption of the spraying equipment, an ambient temperature of 28°C (close to the battery's high temperature threshold), and a wind speed of 4 m / s in the operation area (affecting drone stability). The hidden layer extracts feature vectors using an RNN, identifying high-power equipment operation and high temperatures as the primary factors affecting flight endurance. A priority allocation mechanism automatically increases the weights for battery consumption and ambient temperature. The backpropagation algorithm adjusts the weights between the hidden and output layers based on the error between the actual remaining battery capacity at the end of the operation (e.g., 25%) and the predicted value (e.g., 28%), mitigating prediction bias caused by overestimating the impact of high temperatures. The particle swarm optimization algorithm searches the weight space and finds that adjusting the weight for ambient temperature from 0.2 to 0.25 and the weight for flight speed from 0.1 to 0.08 significantly reduces the overall prediction error, thus selecting this set of weights as the new optimal solution.

[0111] The entire iterative weight update process is a closed loop of dynamic optimization: the input layer receives real-time prediction indicators, the hidden layer extracts time series features through RNN, a priority allocation mechanism highlights key parameters, and the backpropagation algorithm and particle swarm optimization algorithm collaboratively adjust network weights, allowing the model to continuously adapt to new flight conditions and data distributions. Through this mechanism, the predictive optimization network model can gradually learn the impact of various parameters on battery life in different scenarios, generating a more accurate battery life prediction strategy. For example, after performing multiple nighttime flight missions in low-temperature environments, the model automatically increases the weight of the ambient temperature parameter and adjusts the prediction logic for the battery consumption rate to more accurately reflect the attenuation effect of low temperatures on battery performance.

[0112] This process requires no human intervention and is entirely data-driven, making it suitable for the complex and ever-changing demands of drone flight scenarios. The frequency of weight iterations can be set based on computing resources and mission requirements, for example, once per second for real-time flight control and once per minute for mission planning. This ensures that the predictive optimization network model can promptly respond to environmental changes and battery status fluctuations, providing continuous optimization decision support for drone endurance prediction.

[0113] Example 5:

[0114] The sensitivity analysis model and multi-source data fusion optimization in the feedback optimization module are implemented as follows:

[0115] The sensitivity analysis model is used to quantify the sensitivity of predicted instructions to various operating parameters in order to identify key factors influencing the prediction of drone battery endurance and flight capabilities. This model constructs a multivariate sensitivity calculation framework based on the variance decomposition method. Its core idea is to decompose the total variance of the predicted instructions into variance components independently contributed by each operating parameter, and then determine the importance of each parameter by comparing the size of each component. For example, in a certain flight prediction, the predicted instruction was "38 minutes of remaining flight time." The variance decomposition method analysis found that the battery voltage parameter contributed 35% of the variance, the ambient temperature contributed 28%, the flight speed contributed 20%, and the other parameters contributed a total of 17%. This indicates that battery voltage and ambient temperature are the core parameters affecting the prediction result.

[0116] The specific steps for implementing sensitivity analysis are as follows: First, a set of parameter perturbation samples is generated through random sampling. The sampling range covers the normal fluctuation range of each operating parameter. For example, battery voltage can be randomly sampled within ±10% of the nominal voltage, ambient temperature can be sampled uniformly between -10°C and 40°C, and flight speed can be randomly selected between 30% and 80% of the drone's maximum speed. For a system with N operating parameters, M samples (e.g., M = 1000) are generated. Each sample is an N-dimensional vector representing a set of parameter combinations. Each sample is then input into the prediction model to calculate the corresponding predicted command value (such as the endurance prediction value or the battery consumption rate prediction value). Next, the contribution of each parameter is calculated: that is, the proportion of the predicted command variance caused by the parameter change to the total variance. The contribution is calculated based on statistical methods. For example, for each parameter, the variance of the predicted command is calculated by fixing the other parameters at their mean and varying only the value of that parameter. The change in the predicted command variance is then calculated. The ratio of this change to the total variance is the contribution of that parameter. Finally, parameters with contribution higher than the preset threshold (such as 20%) are selected as core variables for optimization prediction. These variables will be processed in the subsequent multi-source data fusion optimization.

[0117] Taking the urban logistics delivery scenario of drones as an example, operating parameters include battery voltage, current, remaining capacity, flight altitude, flight speed, ambient temperature, wind speed, and delivery payload weight. After randomly sampling a sample set, calculations revealed that the remaining battery capacity (32%), ambient temperature (25%), and delivery payload weight (23%) all contributed over 20%. Therefore, these three parameters were identified as core variables for optimized prediction, while flight altitude (8%) and wind speed (5%) were treated as secondary variables.

[0118] The multi-source data fusion optimization process integrates battery sensor data, flight log data, and weather forecast data to improve the accuracy and reliability of prediction instructions. A heterogeneous data fusion matrix is constructed. Battery sensor data includes high-frequency data (e.g., sampled once per second) such as voltage, current, temperature, and remaining capacity, collected in real time. Flight log data records medium-frequency data (e.g., recorded once per minute) such as flight path, waypoint coordinates, and operational instructions. Weather forecast data provides pre-acquired regional weather information, including low-frequency data (e.g., updated hourly) such as temperature, wind speed, and precipitation probability. These data are aligned by timestamp and stored in a matrix format, with each row corresponding to a time point and each column corresponding to a data source feature (e.g., battery voltage, waypoint altitude, predicted wind speed, etc.). For example, a row of data might read: [Battery voltage 3.8V, current 2.5A, remaining capacity 55%, flight altitude 150 meters, current waypoint coordinates (X1, Y1), predicted wind speed 5m / s in 1 hour, temperature 25°C].

[0119] Principal component analysis (PCA) is used to reduce the dimensionality of heterogeneous data fusion matrices. The PCA algorithm maps high-dimensional data to a low-dimensional space through linear transformations while preserving the data's primary variance. Specifically, the covariance matrix of the data matrix is calculated, its eigenvalues and eigenvectors are solved, and the top K eigenvectors with the largest eigenvalues are selected as principal components (e.g., K = 3 or 4). These principal components can explain more than 80% of the variance in the original data. For example, in a heterogeneous data matrix containing 10 original features, PCA can extract three principal components. The first principal component might comprehensively reflect the battery status (voltage, remaining capacity), the second principal component reflects flight intensity (altitude, speed), and the third principal component reflects environmental conditions (temperature, wind speed). The original data is projected onto these three principal components to obtain principal eigenvectors, which have a much lower dimensionality than the original data, thus reducing computational complexity while preserving key information.

[0120] The main eigenvector is input into the feedback optimization module to generate optimized prediction instructions. The feedback optimization module uses the main eigenvector as input based on a machine learning model (such as a neural network or regression model) and outputs adjusted prediction instructions. For example, in the above-mentioned urban logistics and distribution scenario, the main eigenvector contains comprehensive information on battery status, flight intensity, and environmental conditions. The feedback optimization module has learned from historical data that when the remaining battery capacity is less than 60% and the ambient temperature is above 28°C, the actual flight time is an average of 8% shorter than the initial prediction value. Therefore, when the current main eigenvector shows a remaining battery capacity of 55% and an ambient temperature of 29°C, the module automatically adjusts the initial predicted flight time of 38 minutes to 35 minutes and outputs the optimized prediction instructions.

[0121] The entire process uses sensitivity analysis models to identify key parameters, avoiding computational redundancy caused by processing large amounts of secondary data. Multi-source data fusion, by integrating information from different dimensions, overcomes the limitations of a single data source. For example, in mountainous flight scenarios, relying solely on battery sensor data may not be able to predict the additional flight power consumption caused by complex terrain. However, the waypoint altitude changes in flight log data and sudden changes in wind speed in weather forecast data can be captured through principal component analysis after fusion, thereby correcting the deviation of the initial prediction.

[0122] The integration of sensitivity analysis and multi-source data fusion also enables dynamic optimization strategies. For example, when a drone enters a mountainous area from a plain, the sensitivity analysis model recalculates the contribution of each parameter in real time. If it finds a significant increase in the contribution of the flight altitude change rate and wind speed (for example, from 10% to 25% and 20% respectively), the multi-source data fusion matrix will automatically increase the sampling frequency of these two parameters and assign them higher weights in the principal component analysis, enabling the feedback optimization module to more accurately adapt to the impact of terrain changes on flight range.

[0123] Through this entire process, the feedback optimization module achieves multi-source data-driven optimization of the endurance prediction strategy, ensuring that the prediction instructions can comprehensively consider the real-time battery status, flight history patterns, and environmental change trends. This optimization mechanism is not only applicable to conventional flight scenarios, but also can respond to emergencies (such as sudden weather changes and equipment abnormalities). By promptly integrating new data and adjusting the prediction model, it provides a strong guarantee for the safe flight of drones and the successful execution of missions. For example, when the weather forecast data is updated to indicate that strong winds will occur in the area ahead, the system can quickly increase the contribution of the wind speed parameter to 30%, and strengthen its role in the prediction through principal component analysis, issuing an early warning of reduced endurance and reminding ground control personnel to adjust the flight path or add an alternate landing point.

[0124] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0125] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A UAV battery life flight capability prediction system, characterized by: The system comprises: A multi-dimensional data acquisition module is used to obtain a battery data set within a preset time range during the flight of the UAV, and pre-process the battery data set according to a preset data cleaning strategy to generate standardized battery data; A feature mapping module, configured to map the standardized battery data into a feature space and generate a feature sequence cluster based on a time series segmentation algorithm, wherein the feature sequence cluster includes a multi-dimensional correlation between battery status, flight parameters, and environmental factors; A prediction module is used to perform adaptive parameter adjustment on the feature sequence cluster according to preset prediction rules, divide the prediction intervals by a support vector machine algorithm, and extract a set of prediction indicators within each interval; An optimization module is used to construct a decision unit including a prediction optimization network model, and iteratively update the weights of the prediction optimization network model using the prediction indicator set to generate a range prediction strategy; The feedback optimization module is used to perform multi-source data fusion optimization on the endurance prediction strategy according to a preset sensitivity analysis model, and output the UAV battery endurance flight capability prediction instruction.

2. The UAV battery life flight capability prediction system according to claim 1, characterized in that: The generating of feature sequence clusters based on the time series segmentation algorithm includes: Extracting timestamps, battery IDs, and operating parameters from the standardized battery data to construct a time series matrix; Using a window sliding segmentation algorithm to segment the time series matrix to generate a set of time series sub-segments; Using a support vector machine-based classification model to perform pattern recognition on the time series sub-segment set and eliminate abnormal data points; The remaining sub-segments are integrated at multiple levels through an aggregation algorithm to generate feature sequence clusters with temporal correlation.

3. The UAV battery life flight capability prediction system according to claim 2, characterized in that: The adaptive parameter adjustment includes: Based on the historical battery data distribution in the forecast interval, calculate the mean and standard deviation of key forecast indicators in each interval; Based on the moving average algorithm, the mean and standard deviation are smoothed and corrected to generate a dynamic parameter benchmark; The rule engine is used to adjust the interval boundary of the dynamic parameter benchmark to generate an adaptive parameter interval.

4. The UAV battery life flight capability prediction system according to claim 1, characterized in that: The support vector machine algorithm divides the prediction interval, including: The objective function is defined as a weighted balance between battery consumption and flight stability within the prediction interval; The constraints are set as flight path continuity, battery capacity limit and environmental condition tolerance; The objective function is solved by the gradient descent method, and the optimal prediction interval partitioning scheme is output.

5. The UAV battery life flight capability prediction system according to claim 1, characterized in that: The prediction optimization network model performs iterative weight updating, including: Inputting the prediction indicator set into the input layer of the prediction optimization network model, and extracting feature vectors using a recurrent neural network; sorting the feature vectors by weight using a priority allocation mechanism to generate a parameter priority sequence; The back propagation algorithm is combined with the particle swarm optimization algorithm to update the network weights and the decision parameters of the decision unit.

6. The UAV battery life flight capability prediction system according to claim 1, characterized in that: The data cleaning strategy includes: Identifying duplicate data segments in the battery data set and removing duplicates based on a timestamp alignment rule; Detecting missing data points and interpolating the missing data points using linear interpolation; The interpolated data were normalized to generate standardized battery data with a mean of zero and a variance of one.

7. The UAV battery life flight capability prediction system according to claim 1, characterized in that: The sensitivity analysis model includes: Construct a multivariate sensitivity calculation framework based on variance decomposition method to quantify the sensitivity of the predicted instructions to various operating parameters; Generate parameter perturbation sample sets through random sampling and calculate the contribution of each parameter; Parameters with contributions higher than the preset threshold are selected as core variables for optimized prediction.

8. The UAV battery life flight capability prediction system according to claim 1, wherein: The multi-source data fusion optimization includes: Integrate battery sensor data, flight log data, and weather forecast data to build a heterogeneous data fusion matrix; Using principal component analysis algorithm to perform dimensionality reduction processing on the heterogeneous data fusion matrix and extract the main eigenvector; The main feature vector is input into the feedback optimization module to generate an optimized prediction instruction.

9. The UAV battery endurance flight capability prediction system according to any one of claims 1 to 8, wherein: The system further comprises: Constructing an early warning module including a battery health assessment function, performing correlation analysis between the prediction instructions and the battery health, and outputting a graded early warning signal; The preset response mechanism is triggered according to the graded warning signal to generate battery maintenance or flight adjustment instructions.

10. The UAV battery life flight capability prediction system according to claim 9, characterized in that: The generation of the graded warning signal includes: Calculating a real-time health score based on the battery health assessment function; Classifying the health scores through a threshold comparison algorithm to generate early warning signals of different levels; The event triggering mechanism is used to map the warning signal to the corresponding response instruction.

Images