SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Characterization of clocks and inertial sensors, such as accelerometers and gyroscopes, typically includes Allan variance analysis. Allan variance is ubiquitous in timing and navigation communities which may appear niche compared with generalized spectral analysis. This note provides some motivation for Allan Variance for audiences more familiar with spectral analysis.
We study the extent to which knowledge of Allan variance and other finite-difference variances determines the spectrum of a random process. The variance of first differences is known to determine the spectrum. We show that, in general, the Allan variance does not. A complete description of the ambiguity is given.
A relationship was constructed between the Allan variance parameters (H sub z, H sub 1, H sub 0, H sub -1 and H sub -2) and a Kalman Filter model that would be used to estimate and predict clock phase, frequency and frequency drift. To start with the meaning of those Allan Variance parameters and how they are arrived at for a given frequency source is reviewed. Although a subset of these parameters is arrived at by measuring phase as a function of time rather than as a spectral density, they all represent phase noise spectral density coefficients, though not necessarily that of a rational spectral density. The phase noise spectral density is then transformed into a time domain covariance model which can then be used to derive the Kalman Filter model parameters. Simulation results of that covariance model are presented and compared to clock uncertainties predicted by Allan variance parameters. A two state Kalman Filter model is then derived and the significance of each state is explained.
Although the measurement of Allan variance of oscillators is well documented, there is a need for a simplified system for finding the degradation of phase noise and Allan variance step-by-step through a system. This article describes an instrumentation system for simultaneous measurement of additive phase noise and degradation in Allan variance through a transmitter system. Also included are measurements of a 20-kW X-band transmitter showing the effect of adding a pass tube regulator.
The third-difference approach to modified Allan variance (MVAR) leads to a tractable formula for a measure of MVAR estimator confidence, the equivalent degrees of freedom (edf), in the presence of power-law phase noise. The effect of estimation stride on edf is tabulated. A simple approximation for edf is given, and its errors are tabulated. A theorem allowing conservative estimates of edf in the presence of compound noise processes is given.
This study gives strategies for estimating the modified Allan variance (mvar) and formulas for computing the equivalent degrees of freedom (edf) of the estimators. A third-difference formulation of mvar leads to a tractable formula for edf in the presence of power-law phase noise. The effect of estimation stride on edf is tabulated. First-degree rational-function approximations for edf are derived.
An approximating algorithm for computing equvalent degrees of freedom of the Modified Allan Variance and its square root, the Modified Allan Deviation (MVAR and MDEV), and the Time Variance and Time Deviation (TVAR and TDEV) is presented, along with an algorithm for approximating the inverse chi-square distribution.
Report discusses formulas for estimation of Allan variances. Presents algorithms for closed-form approximations of numbers of degrees of freedom characterizing results obtained when various estimators applied to five power-law components of classical mathematical model of clock noise.
A paper at the 1992 FCS showed how to express the modified Allan variance (mvar) in terms of the third difference of the cumulative sum of time residuals. Although this reformulated definition was presented merely as a computational trick for simplifying the calculation of mvar estimates, it has since turned out to be a powerful theoretical tool for deriving the statistical quality of those estimates in terms of their equivalent degrees of freedom (edf), defined for an estimator V by edf V = 2(EV)2/(var V). Confidence intervals for mvar can then be constructed from levels of the appropriate 2 distribution.
An investigation of the Allan variance method as a possible means for characterizing fluctuations in radiometric noise diodes has been performed. The goal is to separate fluctuation components into white noise, flicker noise, and random-walk noise. The primary means is by discrete-time processing, and the study focused primarily on the digital processes involved. Noise satisfying the requirements was generated by direct convolution, fast Fourier transformation (FFT) processing in the time domain, and FFT processing in the frequency domain. Some of the numerous results obtained are presented along with the programs used in the study.
Explore the source record for details and available documents.
A drift-rate estimator constructed from four values of the cumulative sum of clock residuals is shownn to have good error performance in the presence of the five standard power-law noises.
Generally, it is possible to obtain equally spaced timing data from oscillators. The measurement of the drifts and noises affecting oscillators is then performed by using a variance (Allan variance, modified Allan variance, or time variance) or a system of several variances (multivariance method). However, in some cases, several samples, or even several sets of samples, are missing. In the case of millisecond pulsar timing data, for instance, observations are quite irregularly spaced in time. Nevertheless, since some observations are very close together (one minute) and since the timing data sequence is very long (more than ten years), information on both short-term and long-term stability is available. Unfortunately, a direct variance analysis is not possible without interpolating missing data. Different interpolation algorithms (linear interpolation, cubic spline) are used to calculate variances in order to verify that they neither lose information nor add erroneous information. A comparison of the results of the different algorithms is given. Finally, the multivariance method was adapted to the measurement sequence of the millisecond pulsar timing data: the responses of each variance of the system are calculated for each type of noise and drift, with the same missing samples as in the pulsar timing sequence. An estimation of precision, dynamics, and separability of this method is given.
Continuous-time models of oscillator phase noise x(t) usually have stationary nth differences, for some n. The covariance structure of such a model can be characterized in the time domain by the structure function: D sub n (t;gamma sub 1, gamma sub 2) = E delta (n) sub gamma sub 1 x(s+t) delta(n) sub gamma sub 2 x (s). Although formulas for the special case D sub 2 (0;gamma,gamma) (the Allan variance times 2 gamma(2)) exist for power-law spectral models, certain estimation problems require a more complete knowledge of (0). Exhibited is a much simpler function of one time variable, D(t), from which (0) can easily be obtained from the spectral density by uncomplicated integrations. Believing that D(t) is the simplest function of time that holds the same information as (0), D(t) is called the fundamental structure function. D(t) is computed for several power-law spectral models. Two examples are D(t) = K/t/(3) for random walk FM, D(t) = Kt(2) 1n/t/ for flicker FM. Then, to demonstrate its use, a BASIC program is given that computes means and variances of two Allan variance estimators, one of which incorporates a method of frequency drift estimation and removal.
A design using a medium sized computer, a fiber optic communication link, and a desk top computer to reduce data collected at a remote site for analysis of frequency standard tests is described. Data generated at the Interim Frequency Standards Test Facility is collected on magnetic tape, which is then transported to a Sigma 5 computer. An Allan variance (two sample variance) is then calculated. The Allan variance (sigma vs tau) is then written to a data file in the Sigma 5. A fiber optic communication link between the Sigma 5 and a desk top computer was designed to transmit data in the Sigma 5 file to the desk top computer. A program was written for the desk top computer that will automatically access the data file in the Sigma 5, process the data and provide a data plot.
A custom laboratory electronic timer circuit measures the durations of successive cycles of nominally highly stable input clock signals in as many as eight channels, for the purpose of statistically quantifying the small instabilities of these signals. The measurement data generated by this timer are sent to a personal computer running software that integrates the measurements to form a phase residual for each channel and uses the phase residuals to compute Allan variances for each channel. (The Allan variance is a standard statistical measure of instability of a clock signal.) Like other laboratory clock-cycle-measuring circuits, this timer utilizes an externally generated reference clock signal having a known frequency (100 MHz) much higher than the frequencies of the input clock signals (between 100 and 120 Hz). It counts the number of reference-clock cycles that occur between successive rising edges of each input clock signal of interest, thereby affording a measurement of the input clock-signal period to within the duration (10 ns) of one reference clock cycle. Unlike typical prior laboratory clock-cycle-measuring circuits, this timer does not skip some cycles of the input clock signals. The non-cycle-skipping feature is an important advantage because in applications that involve integration of measurements over long times for characterizing nominally highly stable clock signals, skipping cycles can degrade accuracy. The timer includes a field-programmable gate array that functions as a 20-bit counter running at the reference clock rate of 100 MHz. The timer also includes eight 20-bit latching circuits - one for each channel - at the output terminals of the counter. Each transition of an input signal from low to high causes the corresponding latching circuit to latch the count at that instant. Each such transition also sets a status flip-flop circuit to indicate the presence of the latched count. A microcontroller reads the values of all eight status flipflops and then reads the latched count for each channel for which the flip-flop indicates the presence of a count. Reading the count for each channel automatically causes the flipflop of that channel to be reset. The microcontroller places the counts in time order, identifies the channel number for each count, and transmits these data to the personal computer.
The Global Positioning System (GPS) Master Control Station (MCS) currently makes significant use of the Allan Variance. This two-sample variance equation has proven excellent as a handy, understandable tool, both for time domain analysis of GPS cesium frequency standards, and for fine tuning the MCS's state estimation of these atomic clocks. The Allan Variance does not explicitly converge for the nose types of alpha less than or equal to minus 3 and can be greatly affected by frequency drift. Because GPS rubidium frequency standards exhibit non-trivial aging and aging noise characteristics, the basic Allan Variance analysis must be augmented in order to (a) compensate for a dynamic frequency drift, and (b) characterize two additional noise types, specifically alpha = minus 3, and alpha = minus 4. As the GPS program progresses, we will utilize a larger percentage of rubidium frequency standards than ever before. Hence, GPS rubidium clock characterization will require more attention than ever before. The three sample variance, commonly referred to as a renormalized Hadamard Variance, is unaffected by linear frequency drift, converges for alpha is greater than minus 5, and thus has utility for modeling noise in GPS rubidium frequency standards. This paper demonstrates the potential of Hadamard Variance analysis in GPS operations, and presents an equation that relates the Hadamard Variance to the MCS's Kalman filter process noises.
Allan Deviation computations of linear frequency synthesizer systems have been reported previously using real-time simulations. Even though it takes less time compared with the actual measurement, it is still very time consuming to compute the Allan Deviation for long sample times with the desired confidence level. Also noises, such as flicker phase noise and flicker frequency noise, can not be simulated precisely. The use of frequency domain techniques can overcome these drawbacks. In this paper the system error model of a fictitious linear frequency synthesizer is developed and its performance using a Cesium (Cs) atomic frequency standard (AFS) as a reference is evaluated using frequency domain techniques. For a linear timing system, the power spectral density at the system output can be computed with known system transfer functions and known power spectral densities from the input noise sources. The resulting power spectral density can then be used to compute the Allan Variance at the system output. Sensitivities of the Allan Variance at the system output to each of its independent input noises are obtained, and they are valuable for design trade-off and trouble-shooting.