【作者】 战卫侠；

【导师】 谭继文；

【作者基本信息】 青岛理工大学 ， 机械设计及理论， 2013， 博士

【摘要】 钢丝绳作为工程承载的关键构件应用广泛，在其长期作业过程中受环境腐蚀、不确定性交变载荷、机械冲击、磨损等的影响，会出现诸如断丝、磨损、锈蚀等损伤，日积月累导致强度降低甚至突然断裂，轻则造成生产停顿、重则机毁人亡。因此，对钢丝绳的损伤进行检测和准确识别，避免断绳事故的发生，具有重要的社会意义和巨大的经济效益。由于钢丝绳结构的复杂性、损伤的多样性导致损伤产生的磁场变化及其特征与钢丝绳损伤之间的本质关系异常复杂，且涉及多学科领域，因此至今仍未形成钢丝绳损伤检测的完善成型的理论和有效方法，所研发的仪器对钢丝绳损伤诊断的准确率和重复率不高、且不具通用性。本文以危害最大、最常见的损伤形式——断丝损伤为研究对象，分别在钢丝绳励磁器及检测器设计、断丝损伤信号的采集与处理、特征提取及断丝损伤定量识别等方面进行了理论与试验研究。分析了钢丝绳的磁化方式与磁场特征、研究了励磁回路的分析计算方法、磁性材料的优选原则及励磁装置结构对磁场的影响规律，并设计了励磁装置；研究了基于霍尔元件的钢丝绳断丝漏磁检测原理及检测器设计、检测信号预处理及信号采集方法；搭建了钢丝绳断丝检测试验台，并通过试验对理论分析和设计结果进行了验证。研究了基于小波系数尺度间的相关性改进阈值的信号去噪方法，即根据噪声随小波分解层数的增加、稠密度减少、幅值降低、而钢丝绳断丝信号与之相反的特性，提出了特征系数的概念及改进阈值新算法，该特征系数可表征信号的性质。钢丝绳断丝信号其值较小，而噪声信号其值较大，因此采用该特征系数在断丝信号处减小阈值、在噪声处增大阈值。基于该改进阈值对小波分解系数进行处理，即可实现逐点自适应改变阈值大小，达到更好地保留断丝信号，最大限度的去除噪声信号的目的。同时，研究了基于表征钢丝绳断丝信号局部特征的自相关系数逐点选择最佳预测器和更新器的降噪方法，实现了在保留有用断丝损伤信号的前提下对检测信号的有效降噪。通过对钢丝绳断丝损伤检测信号的模拟及实际检测信号的降噪处理试验，验证了上述方法降噪效果的优越性。分析了钢丝绳断丝损伤检测信号的时域、时频域特征及其对断丝损伤的区分度；引入信息熵技术，通过理论分析和试验，研究了检测信号的功率谱熵及重心频率对钢丝绳断丝位置、断口宽度及断丝根数的区分性能，结果表明，功率谱熵及重心频率组成的二维信息可有效区分钢丝绳的断丝位置，而对断口宽度及断丝根数的区分度较差。在对钢丝绳断丝损伤检测信号多域特征研究的基础上，进一步提出了以由检测信号的频域特征（功率谱熵及重心频率）、时频域特征（小波能量）和时域特征（峰值、波形下面积、波宽）组成的混合特征向量作为输入的钢丝绳断丝损伤定量识别方法，通过可分离性判据和试验证明，与采用时域与时频域特征向量相比，采用混合特征向量对钢丝绳断丝损伤进行识别，具有更好的可分离性。因此，该混合特征向量可作为钢丝绳断丝损伤定量识别的有效特征输入。研究了钢丝绳断丝损伤定量识别BP神经网络模型、RBF神经网络模型及支持向量机模型的构建依据和方法，以混合特征向量作为输入，分别构建了基于模式匹配的钢丝绳断丝损伤定量识别的BP神经网络模型、基于函数逼近理论的钢丝绳断丝损伤定量识别的RBF神经网络模型、支持向量机模型，并采用相同的训练样本集和测试样本集对所构建的模型进行训练及测试，对比研究了三种模型对断丝损伤定量识别的性能。结果表明，BP神经网络模型对于钢丝绳断丝损伤识别，具有较强的泛化能力；RBF神经网络模型在只有内层或外层断丝时比BP神经网络稍优；支持向量机模型泛化性能优越，对小样本条件下的钢丝绳断丝损伤的识别效果好，为解决小样本条件下的钢丝绳断丝损伤定量识别问题提供了有效的途径。在对钢丝绳断丝损伤定量识别BP神经网络模型、RBF神经网络模型及支持向量机模型进行理论分析和试验研究的基础上，提出了基于D-S证据理论的钢丝绳断丝损伤多模型融合决策识别方法。研究了多模型融合决策识别模型构建、识别框架构造、证据体基本概率分配函数以及证据体联合规则和决策规则确定的理论依据和方法，构建了基于D-S证据理论的钢丝绳断丝损伤多模型融合决策识别系统，并进行了试验。结果表明，多模型融合决策识别系统对钢丝绳断丝损伤定量识别结果的准确率和可靠性与单一模型的识别结果相比，明显提高，是钢丝绳断丝损伤定量识别的有效新方法。更多还原

【Abstract】 The wire rope is widely used in engineering as a key load-bearing component.During a long-term service, by the effect of enviromental corrosion, uncertainalternating load, mechanical impact and wear etc., damages such as wire breaking,wear, rust will happen, which cumulatively decreases the strength or even lead tofracture of the wire rope. Consequently it may result in a standstill of the machine or afatal crash. Therefore, the detection and accurate identification of the damagecondition of the wire rope in order to avoid the wire breaking accident is of socialsignificance and huge ecomonical benefit.Because of the intricate structure of the wire ropes, the complicated relationbetween the diverse damages and the variation and feature of the magnetic fieldleading to the damages, and multidisciplinary fields involved, by far there is nocomplet theory and effective method formed on the detection of the wire rope damage.Moreover, the instruments developed are not univerisal and usually giveinreproducible and inaccurate results.In this thesis, the most commn damage, wire breaking, is studied theoretically andexperimentally by the design of exciter and detector, signal collectionn and processing,feature extraction and quantitative recognition,etc..Firstly, by analyzing the mode of magnetization and magnetic field characteristicsof wire rope, the analysis and calculation method of excitation circuit, optimumselecting principle of magnetic materials and the law of the exciter structureinfluencing on magnetic field are studied and a excitation device is designed. Thedetection principle of leakage magnetic field based on the hall element, the detectordesign, the method of signal preprocessing and acquisition are investigated. A test rigfor the detection of wire breaking is setted up and the theoretical analysis and designare verified through the experiments.Secondly, the signal denoising method by improving the threshold based on thecorrelation of multi-dimension wavelet coefficients is studied. According to thecharacteristics that the density and the amplitude of noise are decreased with the increase of the layer of wavelet decomposition, but wire breaking signal is opposite,the concept of characteristic coefficient has been put forward. The characteristiccoefficient can characterize the nature of signal, which of wire breaking signal is small,but the one of noise signal is big, the characteristic coefficient is adopted to decreasethe threshold in the wire breaking signal, while increase the threshold in the noise.With the processing of wavelet decomposition coefficients based on the improvedthreshold, the adaptive change of threshold point by point can be realized, so thepurpose of better retaining wire breaking signal and removing noise signal is reached.At the same time, the noise reduction method is studied which chooses the bestpredictor and updater point by point based on the autocorrelation coefficientcharacterizing the local characteristics of wire breaking signal. This method achievedthe effective noise reduction for signal detection on the premise of remaining usefulbroken wire signal. The superiority of the above methods’s noise reduction effect isverified through the experiment of denoising processing of simulating signal and theactual detection signal.Thirdly, the time domain and time-frequency domain characteristics of thedetection signal and the dipartite degree of wire breaking are analyzed. The divisionalperformance of signal power spectral entropy and centroid frequency to wire breakingposition, fracture width and the number of broken wires is studied through theoreticalanalysis and experiment. The results show that the two-dimensional information madeup of power spectral entropy and centroid frequency can effectively distinguish theposition of broken wires, but has a poor differentiation to the fracture width and thenumber of broken wires. Based on the study of the multi-domain characteristics ofdetection signal when the wire rope appears wire breaking, the quantitativeidentification method for broken wire is furtherly put forward which takes the mixedfeature vector consisting of the frequency domain characteristics (power spectralentropy and centroid frequency), time-frequency domain characteristic(wavelet energy)and time domain features (peak value, area of waveform and wave width) as input. Theseparability criterion and experiment prove that using mixed feature vector for brokenwire identification has better separability than using time domain and time-frequency domain feature vector. Therefore, the mixed feature vector can be used as the effectivecharacteristics input of the quantitative identification when wire rope appears brokenwires.At last, the basis and method of building quantitative identification models ofBP neural network, RBF neural network and support vector machine (SVM) for brokenwire are studied. Taking the mixed feature vector as input, the BP neural networkmodel based on pattern matching, the RBF neural network model based on the functionapproximation theory and the support vector machine model are respectivelyestablished and trained by the same data. Then the performance of the three trainedmodels is compared using the same test sample sets to test the model. Results showthat BP neural network model can be used for damage identification of wire rope, RBFneural network model is slightly better than BP neural network, support vectormachine model has superior generalization performance and identification effect underthe condition of small sample of broken wire. SVM provides an effective way forsolving the problem of broken wire’s quantitative identification under the condition ofsmall sample.On the basis of theoretical analysis and experimental research on the BP neuralnetwork model, RBF neural network model and support vector machine model forbroken wire’s quantitative identification, the multi-model fusion decision-makingrecognition method is proposed for the broken wire based on Dempster-Shaferevidential theory. The theory basis and method are researched including the building ofmulti-model fusion recognition model, the construction of frame of discernment,determination of basic probability assignment function of evidence and establishmentof evidential combination algorithm and decision rule. Then a multi-model fusiondecision-making recognition system for broken wire based on D-S evidence theory isconstructed and the experiments are carried on. The results show that the accuracy andthe reliability of quantitative recognition results for broken wire with multi-modelfusion decision-making recognition system, compared with the single modelidentification, are significantly improved. It is a new effective method for brokenwire’s quantitative identification.更多还原