Time Series Classification for Robotic Sensor Pavement Detection: Shapelet, PSO Feature Selection, and Frequency Domain Fusion

随着工业物联网和数字化技术的快速发展，时间序列数据在智能制造、设备监测和机器人控制等领域中的重要性日益凸显，其数据的高维度和复杂性给分类任务带来了巨大挑战。传统的时间序列分类方法通常侧重于全局或局部特征的单一提取，容易忽视多维特征间的协同作用。因此，本文采用了基于Shapelet方法的关键子序列提取，并结合粒子群优化(PSO)算法对生成的候选特征进行特征选择，并进一步引入频域特征增强模型的分类能力。实验评估在SonyAIBORobotSurface1和SonyAIBORobotSurface2数据集上进行，结果显示，基于Shapelet和PSO特征选择的频域融合方法(SP-FD)在分类准确率上均达到90%以上的同时验证了模型的有效性，显著优于其他五个基准模型。With the rapid advancement of industrial Internet of things and digital technologies, time-series data has become increasingly critical in fields such as intelligent manufacturing, equipment monitoring, and robotic control. However, the high-dimensionality and complexity of such data present significant challenges for classification tasks. Traditional time-series classification methods typically focus on single extraction of global or local features, thereby easily overlooking the synergistic effects among multi-dimensional features. To address this, this paper employs Shapelet-based key subsequence extraction, integrates the Particle Swarm Optimization (PSO) algorithm for feature selection on generated candidate features, and further introduces frequency-domain features to enhance the model’s classification capability. Experimental evaluations were conducted on the SonyAIBORobotSurface1 and SonyAIBORobotSurface2 datasets. The results demonstrate that the frequency-domain fusion method based on Shapelet and PSO feature selection (SP-FD) achieves classification accuracies exceeding 90%, verifying the model’s effectiveness and significantly outperforming five other benchmark models.