中文摘要：

信道错误和冲突是导致无线网络中分组丢失和错误的两类主要因素．有效识别分组丢失和错误的原因是实现高性能无线传感器网络协议的基础．然而传感器节点资源的有限性和无线通信环境的复杂性给该问题的研究带来挑战．如何提高分组丢失和错误原因识别方法的准确率，同时保持低的识别开销和易于在节点上实现是该文研究的重点．基于监督学习理论提出一种轻量级的、准确实时的分组丢失和错误原因识别方法EasiPLED．Easi—PLED通过大量实地场景统计实验对分组接收情况进行分析，提取基于RSSI、LQI和F-BER组成的特征向量作为监督学习模型的输入，并实现了一种低开销、控制和数据分组驱动相结合的F-BER计算方法．为了降低噪声、射频硬件本身和高度动态变化的通信环境对特征值计算的影响，作者提出了一种基于误差滤波器的特征值预测方法．通过3种监督学习方法对二元和多类EasiPLED分类模型进行离线训练和检验，结果表明EasiPLED获得至少79．8％的准确率．最后基于EasiPLED的概率轮询协议验证了EasiPLED的识别性能．实验结果表明，与已有最新方法相比，该文方法最大可以将轮询成功的概率提高43．5％．

英文摘要：

It is well known that there are two kinds of causes, namely channel-errors and colli- sions, which lead to high probability of packet losses and errors in wireless networks. The ability of discriminating the above two causes provides many opportunities for implementing high effi- cient networking protocols in wireless sensor networks （WSNs）. However, the limited resources of sensor nodes and the highly complex communication environment pose great challenges to cop- ing with the above problem. This paper focuses on how to improve the accuracy of discriminating the causes of packet losses and errors with low overhead and the simplicity of implementation on sensor nodes. Based on supervised learning theory, we propose a light-weighted discriminator, named EasiPLED, to differentiate the root causes of packet losses and errors with high accuracy and timeliness. EasiPLED investigates the F-BER patterns of error packets and the statistic char- acteristics of received packets＇ RSSI and LQI in different environments through extensive indoor experimental studies on packet reception. EasiPLED extracts the input features for supervised learning model based on F-BER, RSSI and LQI, and implements a low-overhead F-BER estima- tion method by combining the control and data-driven mechanisms together. To mitigate the effect of noises, hardware limitations and highly dynamic communication environment on the esti- mation of feature values, the paper presents an adaptive feature estimator based on error-based filter. We model and test the EasiPLED model through three widely used supervised learning methods. The testing results show that EasiPLED can achieves at least 79.8% of accuracy. Finally, we apply the EasiPLED to the probabilistic polling protocol to evaluate its performance. Experimental results show that EasiPLED yields a promotion of the probability of successful polling by up to 43.5% when compared to the recent method.