时间序列的模式识别是数字信号处理领域的重要研究方向之一，其中大量研究工作都基于神经网络。近年来神经网络因在深度学习上的突破而快速发展，为时间序列的模式识别领域带来了新的思路和方法。模式识别在周界安防领域的一些应用也有望通过新的方法做出优化和改进。本篇论文围绕基于神经网络的用于周界安防的时间序列模式识别展开了一系列研究，重点在于利用深度学习进行时间序列的特征学习和人类行为识别、神经网络自适应周界环境、利用分级识别和多节点融合提升识别准确率排除冗余识别结果、融入了新算法的物联网周界安防架构。本文的研究为基于神经网络的用于周界安防的时间序列模式识别研究及应用提供了一定的理论支撑。 首先，本文研究了惯性传感器的时间序列的特征学习和目标识别技术，提出了多通道长短期记忆网络算法。传统的时间序列识别算法非常依赖于人工设定的特征，特征选择直接影响到了识别的效果。本文提出了通道合并长短期记忆网络算法，通过长短期记忆网络实现了特征学习，取代了人工挑选特征的过程。为了优化网络结构降低了网络规模，结合惯性传感器自身特性，提出了多通道长短期记忆网络算法。在无需人工选择特征的情况下，新算法分类准确率高于传统分类算法，说明新算法能学习到有效的特征，且优于人工挑选的特征。多通道长短期记忆网络算法的参数个数也能极大降低。 其次，针对神经网络自适应周界环境的问题提出了情景感知多防区自适应算法。周界安防中应用神经网络时的一个难点是：周界环境复杂，这使得在理想的实验环境中训练出的神经网络参数在投入到现实环境后会出现性能下降的现象。情景感知多防区自适应算法结合了物联网情景感知及基于神经网络的自适应技术，在数据融合时采用了环境信息，利用适配数据对周界不同片区的神经网络参数进行自适应调整，使得周界的不同片区可以获得更适于自身环境的神经网络参数。实验通过变换防区环境验证了算法能自适应不同环境的周界防区。 之后，为了进一步提升对周界目标的识别性能，提出了分级识别融合算法。机器学习的众多研究都证明通过合理的组合多种分类器可提升总体识别性能，通过融合多个相同或不同类的传感器的数据也可提升识别性能。多级识别算法通过前端探测设备中的初级识别，排除掉了大量无活动时的背景数据，降低了上传的数据量，提升了总体的识别准确率，相对于传统方式，可以简化前端探测设备的硬件，降低功耗。多节点融合算法通过融合地理位置相邻的节点的识别结果，减少了相邻的节点的重复识别结果，提升了准确率。实验中使用了分级识别算法的优化后的算法的累计数据传输量极大降低，多节点融合算法能有效去除冗余识别结果。 最后，为了在物联网周界安防系统中应用以上新算法及对传统架构进行五方面的优化，本文提出了一套周界安防系统参考架构。参考架构分为4层：数据层、感知层、服务层、应用层，共包含27个模块。数据层进行多级识别算法的初级识别，并获取环境数据，感知层使用分类器(例如多通道长短期记忆网络算法)进行第二级识别，然后使用多节点融合算法合并识别结果，并进行自适应算法进行适配训练。应用层收集用户的人工识别结果添加到适配数据中。周界安防系统参考架构在满足传统周界安防系统需求的情况下，也同时融入了本文新提出的算法。本文的实验研究中的软件系统均参照了此系统实现。 关键词：神经网络，周界安防，时间序列，模式识别

 Pattern recognition of time series is one of the main subjects of digital signal processing, relies heavily on neural networks. In recent years, neural network technology has been developing rapidly due to the breakthrough of deep learning and keeps updating pattern recognition of time series with new methods which are helpful to applications of pattern recognition on perimeter surveillance domain. This dissertation researches neural network based time series pattern recognition applied to perimeter surveillance domain, focuses on time series representation learning and classification of human activities, perimeter environment adaptation algorithm of neural network, optimization based on multi-level recognition and multi-node fusion, inertial sensor perimeter security system reference architecture. Our research provides theoretical support for the pattern recognition technology of the time series of the inertial sensor and application in the perimeter surveillance domain. First of all, this dissertation studies time series representation learning and classification of inertial sensors, and proposes multi-channel long short term algorithm. Conventional algorithm of time series classification relies heavily on human-designed features which is important to the performance of algorithm. Channel combined long short term algorithm is proposed which learns representation by long short term memory network in place of human feature selecting. Multi-channel long short term algorithm is proposed by optimizing channel combined long short term algorithm in order to reduce network scale. These two new algorithms are more accuracy then conventional algorithm, which implies that features learned by new algorithm rival human selected features. Multi-channel long short term algorithm also have a smaller network scale then channel combined long short term algorithm. After that, a context-aware multi-section adaptation algorithm is proposed to make neural network adapt to perimeter environment. One of the difficulties in applying neural networks in perimeter security is that the perimeter environment is complex which reduces performances of neural network whose parameters are trained in the ideal experimental environment instead of the real environment. Context-aware multi-section adaptation algorithm is based on context-aware and neural network adaptation. Environmental information is fused, and adaptation data is used to adjust the parameters of neural network for different areas of the perimeter, so neural network parameters for different areas of the perimeter are more suitable for their own environment. Experiments show that the algorithm can adapt to different environment perimeter areas by changing the environments. Next , we propose a multi-level recognition multi-node fusion algorithm to improve recognition performance. Many researches prove that total performance can be improved by combine different classifiers and/or fuse data of multi sensors. Multi-level recognition algorithm excludes sensor data of no activities by first level recognition, so that uploaded data is reduced and accuracy is improved and front-end hardware can be simplified. Multi-node fusion algorithm fuses recognition data of nearby sensors to reduce redundant recognition results and improves accuracy. Experiments show that multi-level recognition algorithm reduces accumulated transferred data, multi-node fusion algorithm reduces redundant recognition results efficiently. Finally, a new perimeter surveillance system reference framework is proposed which contains these new algorithms and is optimized based on conventional framework. The new framework has 4 layers: data layer, sensing layer, service layer, application layer, and 27 modules. First recognition of multi-level recognition algorithm is deployed in data layer to retrieve environment data. Second level of multi-level recognition algorithm is deployed in sensing layer to do the job of classification then the third level fuses results. Context-aware multi-section adaptation algorithm is deployed in sensing layer to do adaptation training. In application layer user confirmed results are added to adaptation data. Our software system used in experiments accords with the framework. Key Words: Neural Network, Perimeter Surveillance, Time Series, Pattern Recognition