| 1,IntroductionOptoelectronic Coupled Devices (OCDs) have been widely used in navigation ,satellite communications and other fields of the high reliability requirements . So itis a subject of great significance on theory and application to study on a screeningmethod to estimate the quality of devices. The internal noise of OCDs is a sensitiveindicator of quality and reliability of devices. The method of noise analysis hasadvantages such as speediness, convenience, nodestruction and having definite judgement, and so on. At present, the method of noise power spectrum is widelyused to measure and analyze the lowfrequency noise.However, the method of noise power spectrum has several disadvantages .First, it takes a relative long time to analyze the noise signals and the measuringfrequency is not sufficiently low. And reducing measuring frequency will costmore time? Second, Fourier transform can only analyze stationary signals and italso doesn't reflect any time information? Third, some characteristics of the noisewill show more obvious in the time domain such as burst noise and the fractalcharacteristic of flicker noise (1/f noise). Therefore, international researchers studyextensively on a new fieldthe study on the timedomain analysis of the noise atpresent. If found suitable method, the timedomain analysis can greatly reduce theanalyzing time . So if that, it is able to get integrated research on the characteristicsof the noise in both the frequency and time domain.In view of the current disadvantages of the screening method for the reliabilityof OCDs, this paper has developed a set of test system for the lowfrequencynoise of OCDs based on the dynamic signal analyzerAgilent35670Aand the existing reliability screening method. And it has analyzed the components of the lowfrequencynoise and a new method which is Independent Component Analysis(ICA) and used as the reliability screening method in the time domain.2,The technology of Independent Component AnalysisIndependent component analysis (ICA) is a new approach which is developedin the field of signal processing. Just as its name implies , it divides a signal intoseveral independent components. If the signal itself is a mixture of severalindependent sources, it just naturally wants to break down these sources. Based on the principle , it is impossible to observe only a single channel for thedecomposition . So it needs the synchronous observation of the multichannelsignals whose sources are mixed by different proportions.Because the mechanism of producing the basic noises of the semiconductor isdifferent, the basic noises are statistically independent. The different OCDs containthe noises that are mixed by the basic noises in different proportions. That meansthe noise signal is linearly composed of the basic noises by a certain proportion.Thus according to the principle of ICA, we can consider a noise signal as theobserved signal, and the basic noises as the original signals. Then the observedsignal is linearly mixed by the original signals in a certain proportion. We getmultipathsignals through multichannelly measuring the noise of OCDs. Then weuse ICA to analyze the signals and estimate the original signals although the mixedproportion is unknown.2.1 KurtosisKurtosis or fourthorder cumulants is the method to judge the randomvariables are whether or not nonGaussian. It's an important parameter whichdescribes the characteristics of the probability density function of random variablesand it's the method to measure the peaks of the distribution of timeseries. Kurtosismust be estimated by the samples that are measured. Because it's sensitive to theoutliers, kurtosis is not a robust nonGaussianmeasurement method and this paperjust uses this property of kurtosis.According to a large number of experimental data, the general kurtosis valueof the Gaussian white noise floats between 2.6 and 3.6 while the kurtosis value offlicker noise (1/f noise) floats between 2.3 and 4.0. So the range of kurtosis (peakpeakvalue) of 1/f noise is slightly wider than the white noise's. According to thecharacteristic of kurtosis, the waveform of the OCD that is contained burst noisewill present the obvious outliers which kurtosis is sensitive to. So the value ofkurtosis of the waveform which is contained burst noise is obviously higher thanothers and the range of kurtosis is the widest in the three basic noises'such as thewhite noise's, 1/f noise's and the burst noise's.2.2 EntropyIn the Information Entropy, the entropy of the Gaussian variable s is thebiggest, which is to say the Gaussian distribution has the most information and isthe most random. Thus in the separated signals at the ICA, the signal which has the largest entropy is the Gaussian white noise. However, it doesn't confirm whichvalue of the entropies of the 1/f noise and burst noise is the larger one, so weshould use kurtosis to determine the two other separated signals.3,Results of research3.1 The characteristics of 1/f noise in ICAAccording to kurtosis and entropy, 1/f noise reflects the followingcharacteristics : the value of the kurtosis floats between 2.3 and 4.0 and its range ofthe kurtosis (peakpeakvalue) is slightly wider than the white noise's. Its entropy isless than the white noise's. Based on these characteristics, we use ICA to separate1/f noise from the noise signals.3.2 The classification rules in the time domain of the reliability ofOCDsThis paper uses ICA to analyze the noise signals of OCDs in the time domain.And after research, we draw the classification rules of the reliability of OCDs inthe time domain as follows:1. Devices of Class I:①The range of the kurtosis (peakpeakvalue)of theobserved signals is slightly small and the average value is about 3.0?②Afteranalyzing by ICA, there is no abnormal noise (such as burst noise, etc.) which isseparated from the observed signals?③The kurtosis values of the separatedsignals are about 3.0 and their difference is very small(less than 0.1).2. Devices of Class III:①The maximum value of kurtosis of the observedsignals is very large(more than 10), or the maximum and minimum value are bothvery small (less than 2), and the range of the kurtosis is wider than devices of ClassI's?②After analyzing by ICA, there is the abnormal noise (such as burst noise,etc.) which is separated from the observed signals?③The separated signal whoseentropy is largest is the white noise. Between the two remaining separated signals,the one whose kurtosis is obviously larger is the burst noise and the one whosekurtosis is close to 3.0 is the 1/f noise.3. Devices of Class II: The devices except Class I and Class III.Thereinto, the reliability of devices of Class III is the worst, which has theburst noise and the 1/f noise whose magnitude is very large. This class of devicesstrictly never allowed to be used in the practical applications? the reliability of Class II is worse, which has the 1/f noise whose magnitude is smaller than ClassIII's. The reliability of Class I is the best. Generally, the devices of Class I andClass II can be used normally.3.3 The test system for the noise of OCDs based on the ICA andAgilent 35670AAccording to the screening method of the reliability, we develop a set of testsystem for the noise. It comes into the functions which the existing systems have,such as realtimedisplay of data, frequencydomainmeasurements and analysis,document preservation, generating databases and reports and automatically printingand so on. The new system effectively improves the integrity of the system and thelevel of automation . There are the following characteristics of the system:1. The system uses the dynamic signal analyzerAgilent35670A to measurethe noise of OCDs. Its measuring frequency is lower and its bandwidth offrequency is wider than the traditional test systems'.2. The system uses the General Purpose Interface BusGPIBto connect thecomputer with Agilent 35670A and to control the analyzer with the computer. Itcan make the system to have high stability .3. The system makes the most of the powerful function of processing datawhich LabVIEW has. LabVIEW achieves the functions of data processing, deviceclassification and report generation.4. The system breaks through the limitations of the timedomainmeasurements of the traditional test systems and uses ICA to analyze the timedomaindata. According to the kurtosis and entropy which are the parameters in thetime domain, we draw the classification rules of the reliability of OCDs in the timedomain.4,ConclusionsAccording to the disadvantages of the traditional systems, this paper focuseson the study on the measurement in the time domain. A new timedomainanalyzing methodIndependentComponent Analysis (ICA) is applied to the newsystem. This paper analyzes the characteristics of 1/f noise with the two parametersof kurtosis and entropy. According to these characteristics, 1/f noise is separatedfrom the noise signals using ICA. And we draw the classification rules of thereliability of OCDs in the time domain according to the method of ICA, which fills the gaps in this field. The paper develops a new test system for the lowfrequencynoise of OCDs based on ICA and Agilent 35670A. The system not only inherits allthe functions of the existing systems, but also adds some new functions, whichmakes the new system more integrated and flexibler. |
PDF Full Text Request