融合多种机制的交通时序数据异常检测模型研究
2025-08-09张培培刘佳奇
中图分类号:TP183 文献标识码:A DOI:10.7535/hbkd.2025yx03003
Research on anomaly detection model for traffic time series data integrating multiple mechanisms
ZHANG Peipei,LIU Jiaqi
(SchoolofEconomicsandManagement,NorthChinaUniversityofScienceandTechnology,Tangshan,HebeiO63oo,China)
Abstract:Toenhancetheanomalyrecognitionabilityof trafictimeseriesdata,ahybrid modelwasconstructed.Firstly,the multi-headatention,residualsandprobabiliticsparseself-atentionwerecombined toformaglobalfeaturerecognition(GFR) module,enhancing theability whilereducingcomputationalcomplexity.Secondly,the dilated convolutionandmulti-scale convolution werecombined to formalocalfeaturerecognition(LFR)module,furtheroptimizing local featureextraction. Thirdly,the FreeRunning training strategy wasused to improve model robustness.Fourthly,the modules and training strategy werecombined withLSTM,whiletheresultof theself-atention mechanismreplaced theLSTM input gate,soas to optimize long sequence memoryabilityand reducecomputing complexity.Finally,a multivariate Gausian distribution probabilityfunction wasused to discriminateanomalies.Theresults show that ading each moduleon the basis of LSTM significantly improvesthemodel'spredictionandanomalydetectionability;Comparedwiththegeneralhybridmodel Transformer-Bi-LSTM,the proposed model hasstronger prediction abilityandlowercomputational complexity.The proposed modelperformsefectivelyinrecognizingboth globalandlocalanomalies in trafictime seriesdata,which provides reference forimproving the operational efficiency and safety of the traffic system.
Keywords:computer neural networks;long sequence time-series data;anomaly detection; attention mechanism;LSTM
在交通领域,交通时序数据异常检测至关重要。这类数据包括交通流量、速度和占有率等信息,其异常检测对保障交通安全、优化资源和提升交通效率意义重大。深度学习模型(如RNN、LSTM和Transformer)被广泛用于异常检测[1],通过学习数据模式和长期依赖关系,比较预测值与实际值,来判断异常[2]。然而,交通数据动态性强、复杂度高,受多种因素干扰[3],需识别全局和局部特征异常[4]。1)全局特征检测问题:传统RNN难以处理长期依赖问题,LSTM虽有改进,但对极长序列的处理仍有限制[5],Transformer虽能捕捉远距离依赖[6],但计算复杂度高,不利于超长序列处理[7]。2)局部特征检测问题:卷积层虽能提取局部特征[8],但因卷积核感受野固定,难以捕捉多尺度特征[9]。3)混合模型问题:如 Transformer-Bi-LSTM混合模型虽能提升预测能力[10],但在局部特征提取、鲁棒性和计算效率方面仍有不足。针对上述问题,本文提出一种混合模型,旨在优化交通时序数据的异常检测能力。
1 模型设计
交通时序数据的复杂性使得交通数据的预测极具挑战性,单一方法难以满足高效、精准的预测需求。为此,本文提出的交通时序数据预测模型融合了三大核心模块:1)全局特征识别(global feature recognition,GFR)模块,通过集成 Transformer的多头注意力、残差和概率稀疏自注意力机制,在精准提取全局特征的同时降低计算复杂度;2)局部特征识别(local feature recognition,LFR)模块,结合膨胀卷积和多尺度卷积技术,灵活捕捉交通数据中的局部细节特征,优化局部特征提取效果;3)FreeRunning 训练策略,增强模型训练的鲁棒性。……
