【作者】 周勃；

【导师】 陈长征；

【作者基本信息】 沈阳工业大学 ， 机械设计及理论， 2008， 博士

【摘要】 空调机组通常多台安装在同一空间内,各类信号之间互相干扰很难提取准确真实的信息,采用盲信号分离技术可以从纷繁的数据中提取有用信息,其优势在于无需掌握信号产生和传播的先验知识。本文在故障源数量未知和源信号未知等条件下,着重探讨了适于空调机组故障特征提取的盲分离算法和模型,以制冷机组和冷却塔为研究对象,通过分离故障振动源和噪声源两个途径提供多元化的诊断参数和更丰富的故障信息,能够解决实际运行的空调机组难以提取信号特征的难题,既提高诊断的效率和准确度,又为设备群故障诊断提供解决途径。本文比较了改进的二阶统计量方法和传统的JADE算法,具有时序结构的源信号因其有不同的自相关函数或有非零时序相关数降低了对统计独立性的限制条件,能够较快地实现收敛,同时也能在噪声能量较大的情况下实现分离。在信噪比低于20dB的情况下改进算法性能指标小得多,而在信噪比高于20dB以后,两种算法趋于一致。改进了盲解卷算法以消除噪声影响,无监督的盲分离处理过程和有监督的噪声消除同时进行,通过相似系数来选取适合含噪信号盲解卷的非线性函数来最大化输出信号的广义能量。选择延迟系数综合考虑分离时间和收敛性能,随着延迟系数的增大,收敛误差函数值减小,收敛性能增强,当延迟系数增大到一定程度后,收敛性能的提高就不再显著,同时分离时间会相应的增长。通过实验模拟了两个故障源和多故障源振动信号混合的情况,分别应用ICA算法、Bussgang算法和改进盲解卷算法提取出典型故障信号特征,验证了不同混合模型和迭代算法对振动信号分离结果会产生很大影响。比较后发现改进盲解卷算法的分离精度最高,说明卷积混合模型适合大型空调机组的振动诊断,这是由于大型空调机组的信号传播路径的不同,因而导致同一时刻的观测信号是源信号在不同时刻数据的叠加。最后,应用改进盲解卷算法提取了JZKA31.5型螺杆机组的磨损、气流撞击和啮合不良的振动故障特征,盲提取的白化预处理和对角化造成振动信号的幅值有所变化,但如果用信号波形表征故障特征,可不考虑幅值比例对诊断结果的影响。在分析振动特性、频率特性和声辐射特性三者联系的基础上进行风机声学诊断实验,建立基于非线性混合的声学诊断模型。实验发现,风机转速越大,负载越大,低频部分的幅值差异就越明显。本文以冷却塔声学诊断为实例,初步测试噪声特性作为先验知识,将干扰噪声本身作为一个声源,基于非线性RBFN分离网络从观测信号中提取独立的声源信号特征来识别冷却塔的主要故障类型,并与基于卷积混合和BP模型的分离结果比较。仿真实验中发现采用四阶累积量估计方法比二阶累积量估计误差大,随着信号源个数增加,需要估计四阶以上的高阶累积量,导致算法性能变差,计算量也大幅度增加。因此基于二阶统计量的改进算法适用于声信号的非线性盲分离。研究复杂的大空间背景环境中设备群振动噪声信号混合交叉的盲分离,从迹的概念出发验证算法的稳定性,将无用信号认为是干扰噪声分离出来,只提取期望的随机信号,再根据独立性测度关系依次提取最显著的故障特征,大大简化了计算过程。经过改进后的自然梯度算法仍满足正交约束,而且不依赖学习速率。本文对某会所地下空调机房实施振声诊断,在多台热泵机组和水泵设备集合的情况下通过空调机房噪声频谱的非线性盲分离确定了主要的故障源为螺杆压缩机,应用盲解卷算法提取不对中和碰磨故障振动信号特征,实现了大空间设备群的振声诊断。更多还原

【Abstract】 It is very difficult to extract accurate signals when the air conditioning units are usually installed together because all kinds of complicate signals disturb mutually. Through the blind signal separation technology the useful information can be acquired from the complex data because it does not need the massive samples and priori knowledge of producing and dissemination of signals. In this paper, the mixing models and algorithm are emphatically discussed which are suitable for the fault feature extraction of air-conditioning unit on the condition that the ambient noise, the fault source and the prior knowledge are unknown. The multiplex diagnosis parameters and the richer fault information are supplied through the vibration source and the noise source for the study on the refrigeration unit and the cooling tower. We can solve the he difficult problem of the signal characteristics extraction of the air conditioning units using the method which can both enhance the diagnosis accuracy and provide the solution of failure diagnosis for all kinds of equipment group in spacious situation.In this paper, the improved second-order statistics algorithm is compared by the traditional JADE algorithm because the different autocorrelation function or the non-zero time sequence correlation reduces the limiting condition of statistical independence which can realize quick convergence and also the separation in the noise energy big situation. When the signal noise ratio is lower than 20dB, the improved second-order statistics algorithm is better. But after the signal-to-noise ratio is higher than 20dB, two algorithms tended to be consistent. Blind deconvolution algorithm is improved to eliminate noise effect, and the non-surveillance’s blind separation process and surveillance noise elimination are carried on simultaneously. The nonlinear function suitable to blind deconvolution is selected by the similarity factor to maximize the output signal generalized energy. The delay factor should be chosen according to the separating time and the convergence performance. When the delay factor increases, the value of convergence error function reduces and the convergence performance would be strengthened. When the delay factor increases to the certain extent, the convergence performance is no longer enhanced remarkably and the separating time extends.The mixing situation of two breakdown sources and the multi-breakdown source are simulated through the test separately using the ICA algorithm, the Bussgang algorithm and improved blind deconvolution algorithm to extract the typical fault signal characteristics. It is indicated that the mixing model and the iterative algorithm will influence the separating results of vibration signals. After the comparison, it is discovered that the improved blind separating algorithm enhances the separation precision to be highest, and the convolution mixing model is suitable to the vibration diagnosis of the large-scale air conditioning units. This is because the different disseminating ways leads to the result that the observation signals on the identical time become the superimposition of the source signals in the different time. Finally, the vibration signals of the JZKA31.5 screw unit are collected and the breakdown characteristics of attrition, air current hit and meshing are extracted. The blind extraction processing including whitening pretreatment and diagonalization will lead to some change of the amplitude value of the vibration signal. But if the waveform is available to express the fault feature, the amplitude value proportion does not affect the diagnosis result.The acoustic diagnosis experiments of the blower are designed on the analysis vibration characteristic, the frequency characteristic and the acoustic radiation characteristic, and the acoustics diagnostic model is established based on the nonlinear mixing model. It is discovered in the tests that the rotational speed is bigger, more obvious difference of amplitude value in the low frequency part is. This article takes the cooling tower acoustics diagnosis as an example. Through testing the noise characteristic the priori knowledge is achieved initially. The independent acoustic source signals are extracted from the observation signals based on the nonlinear RBFN separation network to distinguish the major failure types from the cooling tower when the interference noise is taken an acoustic source, and the test results are compared with the separation results based on the blind deconvolution and BP model. In the simulation experiment, it is discovered that the error of estimation using the fourth-order cumulant method is bigger than the second-order cumulant method. When the number of sources increases, higher order cumulant will depress the algorithm performance and improve the computation load. Therefore, the algorithm based on the second-order statistics is better for the nonlinear separation effect of acoustic signals.The blind source separation is studied in the complex big space background environment when the signals of the equipment group are mixed and disturbed mutually. Firstly, the algorithm stability is tested from the trace concept. And then the expected signals are attained by eliminating the unwanted random signals from the interference noise. Finally the most remarkable breakdown characteristics are extracted in turn according to the independent measure relations. The computational process will be simplified greatly by this method. The improved natural gradient algorithm still is satisfied with the orthogonal restraint and did not rely on the studying rate. In this paper, vibration and acoustics diagnosis has been completed in an air conditioning room. It is determined that main fault source was from the compressor through separating the noise spectrum in the air conditioning room based on nonlinear blind separating model when several heating pumps and water pumps are installed intently. The misalignment and rubbing faults has been diagnosed by vibration signal characteristic based on the blind deconvolution algorithm, which realized the vibration and sound diagnosis in the big space for the machine group.更多还原