基于机器学习的直升机飞行状态识别技术研究

发布时间：2018-01-19 11:18

本文关键词： 二叉树SVM 随机森林飞行状态识别线性相关性小样本　出处：《南昌航空大学》2017年硕士论文　论文类型：学位论文

【摘要】：直升机飞行在不同状态时,有寿件和动部件的损伤程度不同。因此,正确识别飞行状态对直升机关键部件的寿命预测及故障诊断,具有重要的意义。在实际中,用于直升机飞行状态训练的样本一般为小样本,而传统神经网络方法,在训练样本较少时,识别率欠佳。针对这个问题,本文采用二叉树SVM和随机森林方法,研究了直升机飞行状态识别技术,旨在提高识别率和识别速度,可为我国直升机健康和使用监测系统(HUMS)研制提供核心的技术。本文主要研究工作和成果如下:(1)研究并实现直升机飞行状态识别预处理。主要包括数据预处理、敏感参数提取和状态预分类。数据预处理首先采用去野点、限幅及中值滤波对飞行数据进行去噪;然后,利用最小二乘法,拟合得到飞行参数的变化率,作为新的飞行参数;去噪实验验证了本文方法的有效性。敏感参数提取是根据直升机操纵特性和飞行参数线性相关性进行的,并通过真实飞行参数数据进行了验证。状态预分类是利用选择的敏感参数,将35种某型直升机飞行状态预分为10小类,通过真实飞行参数数据验证了预分类方法的有效性。(2)提出并实现基于二叉树SVM的直升机飞行状态识别。在状态识别预处理的基础上,首先,对每个小类进行二叉树SVM分类器的设计;然后,利用粒子群算法和遗传算法对二叉树SVM进行参数寻优,从而提高了识别率;最后,分别对每一个二叉树SVM分类器进行样本训练,并将已训练好的网络模型,用于直升机飞行状态识别。以某型直升机真实飞行数据作为实验数据,并将本方法与SVM方法和RBF神经网络方法进行了对比实验,结果表明,在小样本训练情况下,二叉树SVM对直升机飞行状态识别率有明显的提高。但是,该方法的识别速度不快,针对此问题,利用随机森林具有小样本情况下泛化能力较强和训练收敛速度快的特征,进一步开展了飞行状态识别方法研究。(3)提出并实现基于随机森林的直升机飞行状态识别。在状态识别预处理的基础上,首先,设计每个小类的随机森林分类器;然后,利用分类回归树,构建随机森林,并对每一个随机森林分类器进行样本训练;最后,将已训练好的网络模型,用于识别直升机飞行状态。以直升机真实飞行数据作为实验数据,并将随机森林方法与二叉树SVM方法和RBF神经网络方法进行对比实验,结果表明,在小样本训练情况下,随机森林的识别速度明显优于二叉树SVM和RBF神经网络,同时该方法的识别率与二叉树SVM方法相近,明显高于RBF神经网络方法。
[Abstract]:When the helicopter is flying in different states, the damage degree of the parts with longevity and moving parts is different. Therefore, it is of great significance to correctly identify the flight state for the life prediction and fault diagnosis of the key parts of the helicopter. The samples used for helicopter flight state training are usually small samples, but the traditional neural network method has poor recognition rate when the training samples are small. In this paper, binary tree SVM and stochastic forest method are used to study the helicopter flight status recognition technology, which aims to improve the recognition rate and speed. It can provide the core technology for the development of Chinese helicopter health and use monitoring system (HUMS). The main work and results of this paper are as follows: 1). Research and implementation of helicopter flight status recognition preprocessing, including data preprocessing. First, the de-field point, amplitude limit and median filter are used for de-noising flight data. Then, by using the least square method, the change rate of flight parameters is obtained as a new flight parameter. The effectiveness of the proposed method is verified by denoising experiments. The sensitive parameters are extracted according to the linear correlation between helicopter control characteristics and flight parameters. It is verified by the real flight parameter data. The state pre-classification is to pre-classify 35 kinds of helicopter flight state into 10 subcategories by using the selected sensitive parameters. The validity of the pre-classification method is verified by the real flight parameter data. (2) the helicopter flight state recognition based on binary tree SVM is proposed and realized. First of all, based on the pre-processing of state recognition. A binary tree SVM classifier is designed for each subclass. Then, the particle swarm optimization algorithm and genetic algorithm are used to optimize the parameters of binary tree SVM, which improves the recognition rate. Finally, each binary tree SVM classifier is trained and the trained network model is used for helicopter flight status recognition. The actual flight data of a certain helicopter is used as experimental data. The method is compared with the SVM method and the RBF neural network method. The results show that in the case of small sample training. Binary tree SVM can improve the recognition rate of helicopter flight state obviously. However, the speed of this method is not fast. The random forest has the characteristics of strong generalization ability and fast training convergence rate in the case of small samples. Further research on the flight state recognition method is carried out. (3) the helicopter flight state recognition based on random forest is proposed and realized. First of all, on the basis of the pre-processing of state recognition. Design a random forest classifier for each subclass; Then, the random forest is constructed by using the classification and regression tree, and each random forest classifier is trained by sample. Finally, the trained network model is used to identify the flight state of the helicopter, and the real flight data of the helicopter is taken as the experimental data. Compared with the binary tree SVM method and the RBF neural network method, the random forest method is compared. The results show that in the case of small sample training. The recognition speed of random forest is obviously better than that of binary tree SVM and RBF neural network, and the recognition rate of this method is similar to that of binary tree SVM method, and is obviously higher than that of RBF neural network method.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：V275.1;V267;TP181

【参考文献】