基于深度残差概率随机森林的时间序列分类方法

时间序列数据广泛存在于工业、医疗等应用领域,由于其时序相关性强、特征空间维度大,使得传统的时间序列分类方法普遍存在精度不足和需要复杂特征工程等问题.充分考虑深度神经网络在处理复杂时序数据上的优越性以及决策树方法拟合数据能力强的优势,提出一种基于残差网络和概率决策树的端到端统一深度学习模型.该模型利用残差网络从原始时间序列中提取高级特征,为了更好地建立时序数据特征与类别标签间的映射关系,将概率决策树融入至残差网络的分类层.同时,设计随机子空间的集成策略,缓解由于残差网络的深层结构产生的过度拟合现象,并给出联合优化模型分裂参数和预测参数的迭代优化方案.在大量的基准数据集和工业案例上进行实验和分析,实验结果表明,所提出模型的分类性能优于传统方法与其他深度学习方法,且可有效提高残差网络的泛化能力.

Time series classification using deep residual probability random forests

Time series data widely exists in industrial, medical and other application fields. Due to its strong temporal correlation and large feature space dimension, traditional time series classification methods generally have problems of insufficient accuracy and complex feature engineering. This paper proposes an end-to-end unified deep learning model based on residual networks and probability decision trees by fully considering the superiority of deep neural networks in dealing with complex time series data and the strong ability of a decision tree method to fit data. This model uses a residual network to extract advanced features from original time series. In order to better establish the mapping relationship between features and labels, probability decision trees are integrated into the classification layer of the residual network. Meanwhile, the integration strategy of random subspace is designed to alleviate the over-fitting phenomenon caused by the deep structure of the residual network. We also give the iterative optimization scheme to jointly optimize model''s split parameters and prediction parameters. Experiments and analysis on a large number of benchmark datasets and an industrial case show that the classification performance of the proposed model is better than that of traditional methods and other deep learning methods, and the generalization ability of the residual network can be improved effectively.