基于深度森林的无线传感器网络故障分类算法

Fault classification algorithm for wireless sensor networks based on deep forest

摘要:

针对无线传感器网络（WSN）节点容易出现故障从而导致网络瘫痪的问题，提出了一种基于改进的深度森林的无线传感器网络故障分类方法。深度森林是基于森林的集成学习方法，其输入是多维特征向量，特征向量将由多粒度扫描和级联森林这两个主要组成部分进行处理，多粒度扫描通过处理数据之间的关系来增强数据表示的能力，级联森林用于分类或预测。针对级联森林部分随着层数的增加可能造成的维数问题进行优化后，将该算法用于故障分类可以提高故障诊断的精确度。在仿真验证阶段，将本算法与深度神经网络（DNN）和支持向量机（SVM）算法进行对比。结果显示，本算法可以准确的识别出不同的故障类型，并且在损坏故障和电源故障的识别达到了最高精度，综合平均精度在98.4%。对偏移故障、漂移故障和通信故障的识别略低于卷积神经网络（CNN）算法，但综合训练时间、参数调节来看，该算法更能满足实际工程的需要。

Abstract:

Aiming at the problem that wireless sensor network (WSN) nodes are prone to failure, a fault classification method for WSN based on improved deep forest is proposed. Deep forest is an integrated learning method based on forest. Its input is multidimensional feature vector, which is processed by the two main components of multi-granularity scan and cascade forest. The multi-granularity scan enhances the ability of data representation by processing the relationship between data, and the cascade forest is used for classification or prediction. After optimizing the dimension problem caused by the increase of layers in cascaded forest, the algorithm is applied to fault classification to improve the accuracy of fault diagnosis. At the simulation stage, the proposed algorithm was compared with deep neural network (DNN) and support vector machine (SVM) algorithms. The results show that this algorithm can accurately identify different fault types, and the identification of damage fault and power fault has reached the highest accuracy, the comprehensive average accuracy of 98.4%. The identification of offset fault, drift fault and communication fault is slightly lower than that of convolutional neural network (CNN) algorithm, but the algorithm can better meet the needs of practical engineering in terms of comprehensive training time and parameter adjustment.