基于统计特征的工程器械声识别算法研究

发布时间：2018-02-27 18:35

本文关键词： 地下管线工程器械声音统计特征 SVM ELM　出处：《杭州电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：地下管线系统是城市重要的基础设施,承载着供水排水、能源输送、电力供应、通信传输等各方面的任务,为城市生活秩序提供了基本条件。然而在城镇化进程中,城市的建设施工活动日益增多,施工使用的各类设备对地下管线的安全造成了严重的威胁。由于管线埋于地下,且线路分布广泛,对于工程器械的外力破坏很难进行及时预防和保护,目前急需有效的智能监控系统。经过实地考察发现,经常造成地下管线挖断事故的器械主要有手持电镐、液压破碎锤、挖掘机和切割机这几类。本文研究了这几类破坏性工程器械的工作方式,深入分析了其声音信号的特性,在此基础上提出了一种基于统计参数的特征提取方法,并使用支持向量机(SVM)和极限学习机(ELM)分别作为智能分类器,构建了一套工程器械的声音识别算法。该算法可用于地下管线监控系统,能够全天候远程工作,有效地识别破坏性施工器械,检测到管线被挖断的潜在威胁,从而提出预警,识别算法可以使监控系统更加智能化,更有效率地协助管线维护单位的工作。本文的主要工作和成果如下:1.采用传声器阵列采集工程器械的实际施工声音,包括多种声源距离和多种背景环境,用于声音样本库的建立和识别测试实验的实施。2.在特征提取方面,研究了梅尔频率倒谱系数(MFCC)和线性预测倒谱频率(LPCC)特征,并针对其在识别应用领域的局限性,提出了一种统计特征提取方法。该方法选取并计算短时帧的统计参数,从时域和频域两个方面描述声音信号。通过信号参数的变化范围,设置了其中相关的阈值。3.在识别分类方面,通过对SVM和ELM算法的研究,采用这两种学习算法作为智能分类器,进行样本特征的训练建模和测试信号的识别分类。4.对统计特征的时域和频域参数分别进行了实验,对比了两者的识别性能,并分别分析了主要作用。5.对统计特征进行了识别性能实验,并使用MFCC和LPCC特征设计了对比实验。统计特征在SVM和ELM两种分类器下,都准确识别施工器械,在挖掘机的识别方面具有显著的优势;而且使用近距离样本可以识别远距离的声音;对白噪声干扰下的信号,也能保证较小的特征变化范围。
[Abstract]:The underground pipeline system is an important infrastructure of the city, carrying the tasks of water supply and drainage, energy transmission, electric power supply, communication transmission and so on, which provides the basic conditions for the order of urban life. However, in the process of urbanization, With the increasing construction activities in cities, the safety of underground pipelines is seriously threatened by all kinds of equipment used in the construction. Because the pipelines are buried underground, and the lines are widely distributed, It is difficult to prevent and protect the external force damage of engineering instruments in time. At present, an effective intelligent monitoring system is urgently needed. Through field inspection, it is found that the main instruments that often cause accidents of underground pipeline breakage include hand-held electric picks and hydraulic hammer. In this paper, the working methods of these kinds of destructive engineering instruments are studied, and the characteristics of their sound signals are analyzed in depth. On this basis, a feature extraction method based on statistical parameters is proposed. Using support vector machine (SVM) and extreme learning machine (ELM) as intelligent classifiers, a set of sound recognition algorithm for engineering instruments is constructed. The algorithm can be used in underground pipeline monitoring system and can work remotely all the time. It can effectively identify destructive construction equipment, detect the potential threat of pipeline breakage, and put forward early warning. The recognition algorithm can make the monitoring system more intelligent. The main work and results of this paper are as follows: 1.According to the microphone array to capture the actual construction sound of engineering instruments, including various sound source distances and background environments, In the aspect of feature extraction, Mel frequency cepstrum coefficients (MFCCs) and linear predictive cepstrum frequency (LPCCC) features are studied, and their limitations in recognition applications are pointed out. In this paper, a statistical feature extraction method is proposed, which selects and calculates the statistical parameters of the short time frame, and describes the sound signal in time domain and frequency domain. In recognition of classification, through the research of SVM and ELM algorithm, the two learning algorithms are used as intelligent classifier. Training modeling of sample feature and recognition classification of test signal. 4. The time domain and frequency domain parameters of statistical feature are tested, and the recognition performance of them is compared. The main function of statistical feature is analyzed. 5. The performance experiment of statistical feature recognition is carried out, and the contrast experiment is designed by using MFCC and LPCC features. The statistical feature is used in SVM and ELM classifier to identify construction equipment accurately. It has significant advantages in excavator recognition, and the use of close-range samples can recognize long-distance sound, and the signal with white noise interference can also ensure a small range of feature changes.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3;TU990.3

【参考文献】