【作者】 王娜；

【导师】 刘大同； 王本宽；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2024， 硕士

【摘要】 无人机在飞行过程中产生了大量数据,利用这些数据可以评估无人机是否偏离预期状态,发现潜在安全隐患,便于决策调整。基于预测模型可以对运行状态进行评估,方法通过对比模型预测输出与观测值的偏差实现状态评估。但是忽略了多运行工况背景下数据分布变化以及未考虑模型输出数据不确定性,由此导致的预测模型精度不足以及状态评估结果可信性低。基于此,本课题提出多运行工况下无人机状态自适应评估建模方法,主要研究内容如下:首先,分析多运行工况下复杂数据特征提取与运行状态自适应有效评估的需求,论证课题各部分研究确立思路。通过对无人机状态参数有效预测实现状态监测,域自适应方法用来解决多运行工况下数据特征复杂以及新工况出现导致的预测模型精度下降问题,使用蒙特卡洛Dropout方法为模型引入不确定性估计能力。基于状态监测结果获取实际数据偏离值,同时考虑模型不确定性设计多阈值运行状态评估量化机制,实现无人机运行状态的有效评估,由此确立多运行工况下无人机状态评估建模总体方案。其次,针对多运行工况下无人机特征分布变化导致状态参数预测单一模型精度不足的问题,提出基于多运行工况数据簇划分的数据处理方法,将原始多架次数据重构成多个新的数据集。新数据集内部工况特征更为相似,模型更易提取数据特征。通过1DCNN算法基于多个新数据集构建多运行状态下的无人机参数预测模型库。基于多架次真实飞行数据开展实验验证,进一步开展对比实验验证所提方法在无人机运行状态监测上的性能。然后,针对目标无人机数据工况变化以及新工况出现导致的模型精度低、拟合能力差的问题,提出基于跨域相似性度量与域自适应方法的模型匹配与更新方法。首先提出跨域数据相似性度量指标,计算历史数据集(源域无人机数据)和待测数据集(目标无人机数据)的相似度。然后基于度量结果为目标数据匹配最优模型,获取的最佳模型在目标域数据上自适应更新以实现运行状态参数的准确预测。最后在多个无人机数据集上开展实验验证,设置对比试验以验证跨域模型匹配与自适应更新方法有效性。最后,针对未考虑多工况数据变化引入的不确定性导致的无人机状态评估方法可信性低以及单一阈值评估信息有限的问题,提出多工况下无人机运行状态自适应量化评估方法。通过蒙特卡洛Dropout方法为模型引入不确定性量化估计能力,获取模型优化后的稳定预测输出及其不确定性量化估计结果。设计考虑不确定性的运行状态评估机制,对输出不确定性估计值以及观测数据偏离值进行综合打分,获取无人机状态量化评估分数。最后开展的实验验证以验证所提无人机运行状态评估方法的有效性。更多还原

【Abstract】 The unmanned aerial vehicle(UAV)generates a large amount of data reflecting its operational state during flight.Utilizing this data enables the assessment of whether a UAV deviates from its expected state,thereby identifying potential safety hazards and facilitating decision-making adjustments.Evaluation of operational status can be achieved through predictive modeling,assessing deviations between model predictions and observed values.However,this approach overlooks changes in data distribution under multiple operational conditions and fails to consider the uncertainty of model output data,resulting in inadequate predictive model accuracy and low confidence in state assessment results.To address these challenges,this study proposes a method for adaptive evaluation modeling of UAV states under multiple operational conditions.The main research contents are as follows:Firstly,the study analyzes the requirements for extracting complex data features and adaptively evaluating operational states under multiple operational conditions,thereby establishing the research directions for each part of the study.Effective prediction of UAV state parameters enables state monitoring,while domain adaptation methods address issues such as data complexity and declining predictive model accuracy under multiple operational conditions.The Monte Carlo Dropout method is employed to introduce uncertainty estimation capabilities to the model.Based on monitoring results and considering model uncertainty,a multi-threshold operational state assessment quantification mechanism is designed to effectively evaluate UAV operational states,thereby establishing an overall scheme for UAV state evaluation under multiple operational conditions.Secondly,to address the issue of insufficient predictive model accuracy due to UAV feature distribution changes under multiple operational conditions,a data processing method based on clustering of data clusters is proposed.This method reconstructs the original multi-flight data into multiple new datasets,where the internal operational characteristics are more similar,making it easier for models to extract data features.Using the 1DCNN algorithm,a library of UAV parameter prediction models under multiple operational states is constructed based on these new datasets.Experimental verification is conducted using multiple real flight datasets,further comparing the performance of the proposed methods in UAV state monitoring.Next,to tackle the problem of low model accuracy and poor fitting capability due to changes in UAV data conditions and the emergence of new operational conditions,a model matching and updating method based on cross-domain similarity measurement and domain adaptation is proposed.Initially,a cross-domain data similarity measurement indicator is proposed to calculate the similarity between historical datasets(source domain UAV data)and test datasets(target domain UAV data).Then,based on the measurement results,the optimal model is matched to the target data,and the best model is adaptively updated on the target domain data to achieve accurate prediction of operational state parameters.Finally,experiments are conducted using multiple UAV datasets,setting up comparative experiments to validate the effectiveness of cross-domain model matching and model adaptive updating methods.Finally,to address the low reliability of UAV state assessment methods due to the uncertainty introduced by ignoring changes in data under multiple conditions and the limited information provided by single-threshold assessment,a method for adaptive quantification assessment of UAV operational states under multiple conditions is proposed.By introducing uncertainty quantification capabilities to the model through the Monte Carlo Dropout method,stable predictive outputs and uncertainty quantification results are obtained after model optimization.A comprehensive scoring mechanism considering uncertainty estimation values and observed data deviations is designed for operational state assessment.Experimental verification is conducted to validate the effectiveness of the proposed UAV operational state assessment method.更多还原