Abstract: In view of the problem of low recognition rate due to insufficient training samples, a flight condition recognition method based on random forest is proposed. Firstly, the flight data undergo preprocessing by outlier removal, clipping, and smoothing and the flight condition is classified into eight categories according to the characteristic parameters. Secondly, taking advantage of high recognition rate characteristics of random forest, a random forest classifier is designed for each category. Finally, every random forest classifier is trained with training samples, and all flight conditions of the helicopter are identified by the trained random forest classifier. Taking a helicopter real flight data as the experimental data, actual flight experiments show that, compared with the RBF neural network method and SVM method, the proposed method can significantly improve recognition rate and recognition speed is also improved under a small sample condition. It provides a reference for helicopter life prediction.

Key words: random forest, flight condition recognition, preprocessing, characteristic parameters, small sample

摘要： 针对直升机飞行状态识别训练样本数据少而导致识别率不高的问题，提出一种基于随机森林的直升机飞行状态识别方法。首先利用去野点、限幅、平滑处理对飞行数据进行预处理，并根据特征参数将飞行状态分为8个小类；然后利用随机森林识别率较高的特点，对每一小类进行随机森林分类器设计；最后利用训练样本训练每个随机森林分类器，并将训练好的随机森林分类器识别直升机全起落飞行状态。以某型直升机实飞数据作为实验数据，将该方法与RBF神经网络法和SVM法进行对比实验，结果表明在小样本情况下该方法识别率有明显提高，识别速度也有所提高，可为直升机寿命预测提供依据。

关键词: 随机森林, 飞行状态识别, 预处理, 特征参数, 小样本