基于宽度学习的微多普勒目标分类

Micro-Doppler Based Target Classification Using a Broad Learning System

通过提取目标的多普勒信息,可以实现对不同类型目标的分类。传统的机器学习方法需要人工从回波中提取浅层的特征,特征判别性不强,影响分类性能。深度神经网络能够提取到识别对象中更深层次的信息特征,但是因其网络结构复杂,不仅需要设置过多的超参数,并且需要多次迭代,使得训练时间长而饱受诟病。宽度学习不仅有着很好的逼近能力,使得分类效果并不低于深度神经网络,并且由于只有输入层和输出层,不需要进行迭代求取层与层之间的连接权重,所以宽度学习的训练速度非常快速,极大地减少了训练时间。因此,考虑将宽度学习算法应用到雷达的目标分类识别,探索其可行性。

Classification of different types of targets can be achieved by extracting the Doppler information of the targets.Traditional machine learning methods require manual extraction of shallow features from the echoes,and the features are not discriminative,which negatively affects the classification performance.Deep neural networks can extract deeper target features,but they suffer from long training time due to their complex network structure which requires not only setting too many hyper-parameters,but also multiple iterations.Broad learning system has a good approximation ability,resulting in a classification performance not lower than that of deep neural networks.In addition,broad learning system has only the input and output layers,so that there is no need to iterate to find the connection weights between layers,which greatly reduces the training time.Therefore,the feasibility of applying the broad learning system to target classification of radar is explored in this paper.