Unmanned aerial vehicle electric appliance state evaluation and fault prediction method and device based on ensemble learning, and medium
By combining wavelet denoising processing and LightGBM regression model with Bayesian optimization algorithm, the noise interference and hyperparameter selection problems of UAV electrical systems are solved, achieving high-precision state assessment and fault prediction, and improving the reliability and fault early warning capability of UAV electrical systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 四川腾盾科技有限公司
- Filing Date
- 2025-12-23
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to effectively handle high-frequency pulse noise in UAV electrical systems, resulting in the suppression of early fault characteristics. Shallow models exhibit poor adaptability, and hyperparameter selection is time-consuming and prone to getting trapped in local optima, making it difficult to meet the timeliness requirements for online UAV maintenance.
By employing wavelet denoising combined with the LightGBM regression model, and using a Bayesian optimization algorithm to search for the optimal combination in the hyperparameter space, along with a Gaussian process surrogate model for intelligent tuning, an ensemble learning method is constructed to achieve electrical condition assessment and fault prediction for unmanned aerial vehicles (UAVs).
It improves the accuracy of UAV electrical system status assessment and the timeliness of fault early warning, effectively identifies abnormal states and locates fault sources, thus enhancing the accuracy and efficiency of the model.
Smart Images
Figure CN122017385A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) health monitoring technology, and more specifically, to a method, device, and medium for assessing the electrical condition and predicting faults of UAVs based on ensemble learning. Background Technology
[0002] With the widespread application of drones in power line inspection, logistics transportation, and emergency rescue, the reliability of their electrical systems has become a core element in ensuring flight safety. Batteries, motors, and electronic control components operate under high-altitude, strong electromagnetic, violent maneuvering, and wide-temperature environments for extended periods, making them highly susceptible to early failures such as insulation aging, abnormal commutation, and overheating. However, the voltage, current, and temperature signals collected by drone onboard sensors are often accompanied by high-frequency pulses and random drift. Traditional low-pass filtering or moving average methods, while suppressing noise, can also weaken transient fault pulses, causing early fault characteristics to be masked and subsequent diagnostic models to miss detections.
[0003] In terms of condition assessment and fault prediction, engineering practices commonly employ threshold over-limit alarms or shallow models such as support vector machines and random forests. These methods are poorly adapted to high-dimensional, sparse, and nonlinear UAV electrical data, and rely on manual feature engineering, resulting in a sharp decline in model accuracy with varying operating conditions. In recent years, Gradient Boosting Tree (GBDT) and its variant LightGBM have demonstrated advantages in high accuracy and speed for high-dimensional regression tasks; however, hyperparameter selection still relies on grid search or empirical tuning, which is computationally time-consuming and prone to getting trapped in local optima, making it difficult to meet the timeliness requirements of online UAV maintenance.
[0004] Bayesian optimization, guided by a Gaussian process surrogate model, can significantly reduce the number of evaluations and has been used in XGBoost and neural network hyperparameter tuning. However, the combination of Bayesian optimization with LightGBM and its systematic optimization for small-sample, high-noise UAV electrical data has not yet been reported in the literature. Furthermore, existing studies do not consider the correlation between electrical signal frequency bands and fault mechanisms, and the denoising process easily loses weak fault components, leading to insufficient recall in subsequent models. Therefore, a systematic solution integrating the entire chain of "feature-preserving denoising - hyperparameter optimization - regression prediction" is urgently needed to improve the accuracy of UAV electrical system status assessment and the timeliness of fault early warning. Summary of the Invention
[0005] The present invention aims to solve at least one of the aforementioned technical problems existing in the prior art.
[0006] To this end, the first aspect of the present invention provides a method for assessing the electrical condition of unmanned aerial vehicles (UAVs) and predicting faults based on ensemble learning.
[0007] A second aspect of the present invention provides a computer device.
[0008] A third aspect of the present invention provides a computer-readable storage medium.
[0009] This invention provides a method for assessing the electrical condition and predicting faults of unmanned aerial vehicles (UAVs) based on ensemble learning, comprising: Collect operational data of the UAV's electrical system; the operational data of the UAV's electrical system includes operational parameter data and operational status data; The operating data of the UAV's electrical system are subjected to wavelet denoising processing to obtain denoised data; The denoised data is input into a pre-constructed LGBM regression model, which outputs the predicted values of each continuous parameter of the UAV electrical system. The LGBM regression model is obtained by retraining after searching for the optimal hyperparameter combination in the hyperparameter space using a Bayesian optimization algorithm. Calculate the residual between the predicted value and the actual value at the same moment. If the residual exceeds a preset threshold, it is determined that there is an anomaly in the electrical system at that moment.
[0010] The UAV electrical condition assessment and fault prediction method based on ensemble learning according to the above-described technical solution of the present invention may further have the following additional technical features: In the above technical solution, the operating parameter data includes one or more of the following: battery voltage, battery current, motor three-phase current, motor speed, motor temperature, controller temperature, battery SOC, and battery SOH. The operational status data includes one or more of the following: flight attitude angle, flight altitude, GPS signal strength, load weight, fault code words, and system self-test flags.
[0011] In the above technical solution, the wavelet denoising processing of the operating data of the UAV electrical system includes: The Daubechies wavelet basis function is used to perform J-level wavelet decomposition on the original signal, i.e. the running data, to obtain the approximation coefficients and detail coefficients of each level. The noise standard deviation is calculated based on the detail coefficients of the Jth layer, and the noise set threshold is determined based on the noise standard deviation. Soft thresholding is used to shrink and adjust the signal coefficients based on a threshold set for noise. The signal is reconstructed based on the processed detail coefficients and approximation coefficients to obtain the denoised data of the original signal.
[0012] In the above technical solution, the LGBM regression model is constructed using the LightGBM framework, with the objective function being the mean squared error and the base learner being a gradient boosting tree; the model output is the estimated value of the continuous parameters of the UAV electrical system at the next time step; the expression of the LGBM regression model is:
[0013] in, This represents the LGBM regression model, i.e., the strong learner; x represents the input feature. This represents the initial model; in regression tasks, it represents the mean of the training data. This represents the m-th weak decision tree; The weight of the m-th tree is represented by ; M represents the number of weak decision trees, i.e., the M iterations of the LGBM regression model. The objective function of the LGBM regression model is expressed as:
[0014] in, The mean squared error function is represented by y; the target value is represented by y. This represents the predicted value.
[0015] In the above technical solution, the LGBM regression model is trained through the following iterative method: In the m-th iteration, the negative gradient of the current model with respect to each sample is first calculated, serving as the training objective for the new tree, i.e., the pseudo-residual, which includes:
[0016] in, This represents the negative gradient of the i-th sample in the m-th round, i.e., the pseudo residual; This represents the target value corresponding to the i-th sample; This represents the i-th input feature, i.e., the i-th input sample; This represents the strong learner after the (m-1)th iteration; Solve for the new tree with the goal of minimizing the tree fitting error. Objective function:
[0017] Where n represents the total number of training samples; The predictions from the new tree are superimposed onto the current model, including:
[0018] in, This represents the learning rate.
[0019] In the above technical solution, the step of searching for the optimal hyperparameter combination in the hyperparameter space using a Bayesian optimization algorithm includes: Define the hyperparameter vector to be optimized; The objective function of the Bayesian optimization algorithm is treated as a black box function. A surrogate model of the objective function of the Bayesian optimization algorithm is established using Gaussian process, and the prediction distribution of the surrogate model is obtained. The proxy model is iteratively updated by the acquisition function until the algorithm converges and the optimal hyperparameter combination is output. The optimal hyperparameters are then incorporated into the LGBM regression model training process.
[0020] In the above technical solution, the step of establishing a surrogate model of the objective function of the Bayesian optimization algorithm using a Gaussian process and obtaining the predicted distribution of the surrogate model includes: Assume the objective function of the Bayesian optimization algorithm follows a Gaussian process prior, expressed as:
[0021] in, Denotes the objective function of the Bayesian optimization algorithm; m( ) is the mean function; It is a kernel function used to measure two parameter points. and Similarity; For unobserved parameters Based on the observed data D, the unobserved parameters are predicted using a Gaussian process. The posterior distribution includes:
[0022] in, Indicates unobserved parameters The objective function; D represents the observed data; Indicates to The optimal estimate; This indicates the uncertainty of the forecast.
[0023] In the above technical solution, the step of iteratively updating the proxy model through the acquisition function includes: Based on the posterior distribution of the unobserved parameters, the TPE acquisition function is used to select the hyperparameters for the next iteration; The TPE acquisition function is expressed as follows:
[0024] in, Indicates the TPE acquisition function; This represents the probability density of a good sample, i.e., the objective function value is greater than the current optimal value of the objective function. The probability density of the samples; This represents the probability density of bad samples, i.e., the objective function value is no greater than the current optimal value of the objective function. The probability density of the samples; choose The point with the largest value is used as the hyperparameter for the next iteration.
[0025] The present invention provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded and executed by the processor, it implements the method for assessing the electrical condition of unmanned aerial vehicles and predicting faults based on ensemble learning as described in any of the above technical solutions.
[0026] The present invention provides a computer-readable storage medium storing a program that, when loaded by a processor, implements the method for assessing the electrical condition of unmanned aerial vehicles and predicting faults based on ensemble learning as described in any of the above technical solutions.
[0027] In summary, due to the adoption of the above-mentioned technical features, the beneficial effects of the present invention are: This invention has significant advantages in processing UAV electrical data, and can effectively solve problems such as low regression accuracy, difficulty in status monitoring, and delayed fault early warning caused by the complex structure, variable operating environment, and scarcity of fault data of UAV electrical systems.
[0028] Specifically, due to issues such as acquisition accuracy and environmental influences, the data collected from UAV flight data often contains noisy data. This noisy data is not helpful for analyzing the electrical health status of the UAV, but rather hinders the analysis of normal data. This invention uses wavelet denoising to eliminate the influence of noisy data on the real data and restore the true information change trend.
[0029] To address the challenges posed by the complex structure, unstable operating environment, high dimensionality, and sparse nature of UAV electrical data, which makes conventional regression prediction models inadequate for modeling and analyzing UAV electrical data, this invention employs the LGBM algorithm, which has significant advantages in processing high-dimensional and sparse data. By using multiple weighted weak decision trees as the final prediction result, accuracy is guaranteed. After LGBM learns the health patterns of UAV electrical data from historical data, it can identify abnormal states of the UAV electrical system if the prediction results differ significantly from the actual UAV data. Furthermore, the source of the fault can be identified based on the degree of deviation of the variables.
[0030] Furthermore, the selection of hyperparameters in LGBM has a significant impact on the model's accuracy. Traditional methods such as grid search or parameter combination matching are not only inefficient but also prone to missing the optimal parameter combination. In contrast, the Bayesian optimization algorithm used in this invention can perform intelligent search in the hyperparameter space of LGBM, and select and combine hyperparameters in a directional manner, which greatly optimizes the LGBM modeling and optimization time, thereby improving model accuracy and efficiency.
[0031] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description
[0032] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an embodiment of the present invention: an unmanned aerial vehicle (UAV) electrical condition assessment and fault prediction method based on ensemble learning. Detailed Implementation
[0033] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0035] The following reference Figure 1 This describes a method, device, and medium for assessing the electrical condition of unmanned aerial vehicles (UAVs) and predicting faults based on ensemble learning, provided by some embodiments of the present invention.
[0036] Some embodiments of this application provide a method for assessing the electrical condition of unmanned aerial vehicles (UAVs) and predicting faults based on ensemble learning.
[0037] like Figure 1 As shown, the first embodiment of the present invention proposes a method for assessing the electrical condition of a UAV and predicting faults based on ensemble learning, including the following steps S1 to S4.
[0038] S1. Collect operating data of the UAV electrical system; the operating data of the UAV electrical system includes operating parameter data and operating status data.
[0039] Specifically, the operating parameter data includes one or more of the following: battery voltage, battery current, motor three-phase current, motor speed, motor temperature, controller temperature, battery SOC, and battery SOH. The operational status data includes one or more of the following: flight attitude angle, flight altitude, GPS signal strength, load weight, fault code words, and system self-test flags.
[0040] Understandably, the more comprehensive the operational data collection, the more accurate the final prediction result will be. In a specific implementation, the operational data includes all the aforementioned operational parameter data and operational status data.
[0041] S2. Perform wavelet denoising processing on the operating data of the UAV electrical system to obtain denoised data.
[0042] Wavelet denoising is a signal denoising technique based on wavelet transform. Due to its excellent localization characteristics in the time-frequency domain, it has been widely used in various fields such as image processing, speech signal processing, and seismic data processing. The core idea of wavelet denoising is to use wavelet transform to distinguish between the different manifestations of signal and noise. The wavelet denoising algorithm and its core process mainly include wavelet decomposition of the signal, thresholding of the wavelet coefficients, and inverse wavelet transform reconstruction of the signal.
[0043] In some embodiments, step S2 includes the following steps: S21. Select the Daubechies wavelet basis function to perform J-level wavelet decomposition on the original signal, i.e. the running data, to obtain the approximation coefficients and detail coefficients of each level.
[0044] Step S21 corresponds to the wavelet decomposition of the signal. In this step, it is necessary to select an appropriate wavelet function and decomposition level to perform wavelet transform on the noisy signal and obtain wavelet coefficients at different scales (frequency) and locations, including approximation coefficients (reflecting the low-frequency components of the signal) and detail coefficients (reflecting the high-frequency components of the signal, where noise is mostly concentrated).
[0045] In this embodiment, the wavelet basis function used is Daubechies (abbreviated as db4). The db4 wavelet function belongs to the orthogonal wavelet basis functions. The db4 wavelet has four vanishing moments, which means it can approximate functions with third-order continuous derivatives well while maintaining compact support properties. Since the db4 function expression does not have a simple explicit form, it is usually based on the recursive relationship between the scaling function and the wavelet function, as shown below:
[0046]
[0047]
[0048] in, The scaling function, also called the parent wavelet, is the foundation of wavelet analysis and is used to characterize the low-frequency approximation components of a signal. t is a wavelet function, also called the mother wavelet, used to characterize the high-frequency details of a signal; t is a time variable, representing the time-domain coordinates of the signal. The shape of the scaling function is determined by the low-pass filter coefficients (length is 2N, N=4 in db4, so the filter coefficient index n=0,1,2,3); For high-pass filter coefficients (and) (One-to-one correspondence) determines the shape of the wavelet coefficients.
[0049] Based on the above, the process of obtaining the approximation coefficients and detail coefficients of each layer can be expressed as:
[0050]
[0051] in, Represents the approximation coefficients for the j-th layer, j=1,2,3,...,J; This represents the detail coefficient of the j-th layer, where k is the sequence element index; S22. Calculate the noise standard deviation based on the detail coefficients of the Jth layer, and determine the noise setting threshold based on the noise standard deviation.
[0052] Specifically, in db4, the threshold is first set based on the noise level:
[0053] in, The standard deviation of the noise (the noise is mainly concentrated in the highest frequency detail coefficients of the wavelet decomposition, usually the last level detail coefficients). Since the wavelet coefficients of the noise approximately follow a Gaussian distribution with a mean of 0, we have:
[0054] in, This is the median function; 0.6745 is the 0.75 quantile of the standard normal distribution, which can be used to convert the median to the standard deviation.
[0055] S23. Soft thresholding is used to shrink and adjust the signal coefficients according to the noise threshold.
[0056] Specifically, in db4, noise processing and signal preservation are mainly achieved through threshold functions, which are divided into hard thresholding and soft thresholding. This embodiment uses the soft thresholding scheme, that is, by "shrinking" the preserved signal coefficients to make the result smoother, as shown below:
[0057]
[0058] Where d represents the coefficient element. ; It is a symbolic function.
[0059] S24. Based on the processed detail coefficients and approximation coefficients, the signal is reconstructed to obtain the denoised data of the original signal.
[0060] Specifically, the formula for the inverse transform reconstructed signal is as follows:
[0061] In the above formula, This represents the reconstructed signal, i.e., the denoised data. Since the scaling function and wavelet function satisfy the orthogonal characteristic, the components will not interfere with each other during the inverse transform, and the data structure can be accurately restored.
[0062] S3. Input the denoised data into the pre-constructed LGBM regression model and output the predicted values of each continuous parameter of the UAV electrical system; wherein, the LGBM regression model is obtained by retraining after searching for the optimal hyperparameter combination in the hyperparameter space using the Bayesian optimization algorithm.
[0063] Specifically, LGBM (Light Gradient Boosting Machine) regression is a high-efficiency ensemble learning algorithm, improved from the Gradient Boosting Tree (GBDT) framework, characterized by fast training speed, low memory consumption, and high fitting accuracy. The core idea of LGBM originates from Gradient Boosting Tree (GBDT), which iteratively integrates multiple weak decision trees to gradually optimize the model, ultimately forming a strong learner.
[0064] In some embodiments, the LGBM regression model is constructed using the LightGBM framework, with the objective function being the mean squared error and the base learner being a gradient boosting tree; the model output is an estimate of the continuous parameters of the UAV electrical system at the next time step. If the training set data is ,in These are input features. The target value is LGBM, which is a weighted result of multiple weak decision trees. The mathematical expression of the LGBM regression model is as follows:
[0065] in, This represents the LGBM regression model, i.e., the strong learner; x represents the input feature. This represents the initial model; in regression tasks, it represents the mean of the training data. This represents the m-th weak decision tree; The weight (learning rate) of the m-th tree is represented; M represents the number of weak decision trees, i.e., the M iterations of the LGBM regression model; The core of gradient boosting is to train a new tree in each iteration to fit the residual gradient of the current model, gradually reducing the loss. In regression tasks, mean squared error (MSE) is used to define the difference between the model's predictions and the actual results. Specifically, the objective function of the LGBM regression model is expressed as:
[0066] in, The mean squared error function is represented by y; the target value is represented by y. This represents the predicted value.
[0067] In some embodiments, the LGBM regression model is trained through the following iterative method: In the m-th iteration, the negative gradient of the current model with respect to each sample is first calculated, serving as the training objective for the new tree, i.e., the pseudo-residual, which includes:
[0068] in, This represents the negative gradient of the i-th sample in the m-th round, i.e., the pseudo residual; This represents the target value corresponding to the i-th sample; This represents the i-th input feature, i.e., the i-th input sample; This represents the strong learner after the (m-1)th iteration; The above expression is the formula for calculating pseudo residuals, when the loss function... Without MSE, the true residuals cannot be calculated: To ensure that the fitted residuals fit all loss functions, gradient boosting introduces the idea of pseudo-residuals = negative gradients. When the loss function is MSE, the pseudo-residuals and the true residuals are consistent.
[0069] Solve for the new tree with the goal of minimizing the tree fitting error. Objective function:
[0070] Where n represents the total number of training samples; The predictions from the new tree are superimposed onto the current model, including:
[0071] in, The learning rate can be determined by minimizing the loss function, or simplified to a fixed learning rate.
[0072] Furthermore, the Bayesian optimization-based LGBM regression prediction algorithm determines the hyperparameter space for Bayesian optimization based on the LGBM hyperparameter type and previous optimization experience. Then, it uses the Bayesian optimization algorithm to find the optimal combination of hyperparameters, and then retrains the LGBM model to obtain the regression prediction model for each continuous parameter of the UAV. The residual difference between the true value and the predicted value is used as the criterion for judging the state and fault of the UAV, and the fault factors are located through the residual performance.
[0073] Specifically, Bayesian optimization is an efficient hyperparameter optimization method based on probabilistic models. Its core idea is Bayes' theorem and probabilistic modeling. It achieves efficient optimization of complex objective functions by "intelligently exploring" the parameter space rather than blindly traversing it. The core of Bayesian optimization is to use a probabilistic model to describe the uncertainty of the objective function f(θ) and continuously update the model based on new observations, ultimately guiding parameter selection.
[0074] In some embodiments, the search for the optimal hyperparameter combination in the hyperparameter space using the Bayesian optimization algorithm includes the following steps S31-S33.
[0075] S31. Define the hyperparameter vector to be optimized; in a specific embodiment, the hyperparameter vector θ to be optimized is defined as a vector containing hyperparameters such as the number of trees, the number of leaf nodes, the learning rate, the regularization coefficient, and the row sampling ratio, with each parameter point corresponding to a hyperparameter combination.
[0076] S32. Treat the objective function of the Bayesian optimization algorithm as a black box function, use Gaussian processes to establish a surrogate model of the objective function of the Bayesian optimization algorithm, and obtain the prediction distribution of the surrogate model.
[0077] In some embodiments, step S32 includes: Assume the objective function of the Bayesian optimization algorithm follows a Gaussian process prior, expressed as:
[0078] in, Denotes the objective function of the Bayesian optimization algorithm; m( ) is the mean function; It is a kernel function used to measure two parameter points. and Similarity; For unobserved parameters Based on the observed data D, the unobserved parameters are predicted using a Gaussian process. The posterior distribution includes:
[0079] in, Indicates unobserved parameters The objective function; D represents the observed data. ; Indicates to The optimal estimate; This indicates the uncertainty of the forecast.
[0080] S33. Iteratively update the surrogate model using the acquisition function until the algorithm converges and outputs the optimal hyperparameter combination.
[0081] After obtaining the predicted distribution of the surrogate model, the next parameter to be evaluated is selected through the acquisition function. The main purpose of the acquisition function is to balance exploration and exploitation. Exploration prioritizes parameters with higher predicted mean values from the surrogate model; Exploitation prioritizes parameters with higher prediction variance in the surrogate model.
[0082] Commonly used sampling functions include Expectation Improvement (EI), Probability Improvement (PI), Upper Confidence Bound (UCB), and Information Entropy Search (MSE). In this embodiment, the sampling function is TPE (Tree-structured Parzen Estimator), whose core principle is to guide sampling by partitioning the parameter space and constructing a probability model.
[0083] The core of TPE is to distinguish between "good" and "bad" samples using two probability density functions (PDFs), and to define the acquisition function based on their ratio, assuming that the value of the objective function for all historical trials is... The top few samples are considered good samples, and the rest are bad samples. TPE selects the next sampling point by maximizing the ratio of the two density functions.
[0084] In one specific embodiment, step S33 includes: Based on the posterior distribution of the unobserved parameters, the TPE acquisition function is used to select the hyperparameters for the next iteration; The TPE acquisition function is expressed as follows:
[0085] in, Indicates the TPE acquisition function; This represents the probability density of a good sample, i.e., the objective function value is greater than the current optimal value of the objective function. The probability density of the samples; This represents the probability density of bad samples, i.e., the objective function value is no greater than the current optimal value of the objective function. The probability density of the samples; The larger the value, the stronger the value at that point. The probability of achieving a "good result" is relatively higher than that of a "bad result," therefore, choosing... The point with the largest value is used as the hyperparameter for the next iteration, until the improvement meets the threshold requirement or the maximum number of iterations is reached, at which point the optimal hyperparameter combination is output.
[0086] S34. Introduce the optimal hyperparameters into the LGBM regression model training process.
[0087] S4. Calculate the residual between the predicted value and the actual value at the same time. If the residual exceeds the preset threshold, it is determined that there is an abnormality in the electrical system at that time.
[0088] It is understandable that in step S4, the actual operating data (actual value) collected in real time at the current moment (or a certain moment) can be compared with the predicted value for the current moment output by the LGBM regression model. If the residual exceeds the preset threshold, it is determined that there is an abnormality in the electrical system at that moment. Furthermore, the electrical subsystem causing the abnormality can be located based on the magnitude and direction of the residual, thereby achieving fault prediction.
[0089] In some embodiments, locating the electrical subsystem causing the anomaly can be achieved by establishing a residual vector, normalizing the residual vector, calculating the contribution of each subsystem, identifying the subsystem with the largest contribution as the fault source, outputting the fault type and suggested maintenance measures, and thus realizing fault prediction.
[0090] Other embodiments of the present invention provide a computer device including a processor and a memory, wherein the memory stores a computer program that, when loaded and executed by the processor, implements the UAV electrical condition assessment and fault prediction method based on ensemble learning as described in any of the above embodiments.
[0091] Other embodiments of the present invention provide a computer-readable storage medium storing a program that, when loaded by a processor, implements the UAV electrical condition assessment and fault prediction method based on ensemble learning as described in any of the above embodiments.
[0092] In this specification, the illustrative expressions of the terms used do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0093] Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention shall be included within the scope of protection of this invention.
Claims
1. A method for electrical condition assessment and fault prediction of unmanned aerial vehicles (UAVs) based on ensemble learning, characterized in that, include: Collect operational data of the UAV's electrical system; the operational data of the UAV's electrical system includes operational parameter data and operational status data; The operating data of the UAV's electrical system are subjected to wavelet denoising processing to obtain denoised data; The denoised data is input into a pre-constructed LGBM regression model, which outputs the predicted values of each continuous parameter of the UAV electrical system. The LGBM regression model is obtained by retraining after searching for the optimal hyperparameter combination in the hyperparameter space using a Bayesian optimization algorithm. Calculate the residual between the predicted value and the actual value at the same moment. If the residual exceeds a preset threshold, it is determined that there is an anomaly in the electrical system at that moment.
2. The method for assessing the electrical condition and predicting faults of unmanned aerial vehicles based on ensemble learning according to claim 1, characterized in that, The operating parameter data includes one or more of the following: battery voltage, battery current, motor three-phase current, motor speed, motor temperature, controller temperature, battery SOC, and battery SOH. The operational status data includes one or more of the following: flight attitude angle, flight altitude, GPS signal strength, load weight, fault code words, and system self-test flags.
3. The method for assessing the electrical condition and predicting faults of unmanned aerial vehicles based on ensemble learning according to claim 1, characterized in that, The wavelet denoising process for the operating data of the UAV's electrical system includes: The Daubechies wavelet basis function is selected to perform J-level wavelet decomposition on the original signal, i.e. the running data, to obtain the approximation coefficients and detail coefficients of each level. The noise standard deviation is calculated based on the detail coefficients of the Jth layer, and the noise set threshold is determined based on the noise standard deviation. Soft thresholding is used to shrink and adjust the signal coefficients based on a threshold set for noise. The signal is reconstructed based on the processed detail coefficients and approximation coefficients to obtain the denoised data of the original signal.
4. The method for assessing the electrical condition of unmanned aerial vehicles and predicting faults based on ensemble learning according to claim 1, characterized in that, The LGBM regression model is constructed using the LightGBM framework, with the objective function being the mean squared error and the base learner being a gradient boosting tree. The model output is the estimated value of the continuous parameters of the UAV electrical system at the next time step. The expression of the LGBM regression model is as follows: in, This represents the LGBM regression model, i.e., the strong learner; x represents the input feature. This represents the initial model; in regression tasks, it represents the mean of the training data. This represents the m-th weak decision tree; The weight of the m-th tree is represented by ; M represents the number of weak decision trees, i.e., the M iterations of the LGBM regression model. The objective function of the LGBM regression model is expressed as: in, The mean squared error function is represented by y; the target value is represented by y. This represents the predicted value.
5. The method for assessing the electrical condition and predicting faults of unmanned aerial vehicles based on ensemble learning according to claim 4, characterized in that, The LGBM regression model is trained using the following iterative method: In the m-th iteration, the negative gradient of the current model with respect to each sample is first calculated, serving as the training objective for the new tree, i.e., the pseudo-residual, which includes: in, This represents the negative gradient of the i-th sample in the m-th round, i.e., the pseudo residual; This represents the target value corresponding to the i-th sample; This represents the i-th input feature, i.e., the i-th input sample; This represents the strong learner after the (m-1)th iteration; Solve for the new tree with the goal of minimizing the tree fitting error. Objective function: Where n represents the total number of training samples; The predictions from the new tree are superimposed onto the current model, including: in, This represents the learning rate.
6. The method for assessing the electrical condition of unmanned aerial vehicles and predicting faults based on ensemble learning according to claim 1, characterized in that, The process of searching for the optimal hyperparameter combination in the hyperparameter space using a Bayesian optimization algorithm includes: Define the hyperparameter vector to be optimized; The objective function of the Bayesian optimization algorithm is treated as a black box function. A surrogate model of the objective function of the Bayesian optimization algorithm is established using Gaussian process, and the prediction distribution of the surrogate model is obtained. The proxy model is iteratively updated by collecting data through a function until the algorithm converges and outputs the optimal combination of hyperparameters. The optimal hyperparameters are then incorporated into the LGBM regression model training process.
7. The method for assessing the electrical condition of unmanned aerial vehicles and predicting faults based on ensemble learning according to claim 6, characterized in that, The process of establishing a surrogate model for the objective function of the Bayesian optimization algorithm using Gaussian processes and obtaining the predicted distribution of the surrogate model includes: Assume the objective function of the Bayesian optimization algorithm follows a Gaussian process prior, expressed as: in, Denotes the objective function of the Bayesian optimization algorithm; m( ) is the mean function; It is a kernel function used to measure the difference between two parameter points. and Similarity; For unobserved parameters Based on the observed data D, the unobserved parameters are predicted using a Gaussian process. The posterior distribution includes: in, Indicates unobserved parameters The objective function; D represents the observed data; Indicates to The optimal estimate; This indicates the uncertainty of the forecast.
8. The method for assessing the electrical condition and predicting faults of unmanned aerial vehicles based on ensemble learning according to claim 7, characterized in that, The iterative update of the proxy model through the acquisition function includes: Based on the posterior distribution of the unobserved parameters, the TPE acquisition function is used to select the hyperparameters for the next iteration; The TPE acquisition function is expressed as follows: in, Indicates the TPE acquisition function; This represents the probability density of a good sample, i.e., the objective function value is greater than the current optimal value of the objective function. The probability density of the samples; This represents the probability density of bad samples, i.e., the objective function value is no greater than the current optimal value of the objective function. The probability density of the samples; choose The largest point is used as the hyperparameter for the next iteration.
9. A computer device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded and executed by the processor, implements the UAV electrical condition assessment and fault prediction method based on ensemble learning as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The system stores a program that, when loaded by a processor, implements the method for assessing the electrical condition of a drone and predicting faults based on ensemble learning as described in any one of claims 1 to 8.