Analysis of some hobbyist oscillators

[dnessett]

For about a year, I have been creating a test setup to investigate "hobbyist oscillators". By this I mean oscillators that hobbyists might use in their projects and also oscillators that don't receive alot of attention from the professionals (and therefore warrant investigation by hobbyists). Candidates for study are retired oscillators bought on ebay (examples: used Morion MV89As and FEI-5650a rubidium oscillators), new oscillators bought on ebay that have little or no published specs (example: low cost GPSDOs, cheap XO can oscillators), and other devices that are not normally studied.

My objective is to create information about these oscillators so hobbyists can determine which are best for their particular projects. The journey to create the test setup was long and arduous and is not fully complete. However, I have enough test equipment in place to begin to look at these oscillators and provide some information about them.

The test setup is documented in this EEVblog topic. It is a record not only of the evolution of the test setup, but also what I learned while buiilding it. Few will be interested in reading the whole topic (which as of this writing, 6/4/2019, comprises 12 EEVblog pages). Those who want just the punchline should read: post 1; post 2, which is partially superceded by post 3; post 4 and post 5, which provide some of the mathematical background behind the test setup (just algebra is required to understand them); post 6, which describes in detail the necessary steps to use the test setup; and post 7, which provides test data showing that the signal from the test setup does not go below the noise floor of the spectrum analyzer used (a PicoScope 4262).

Reading this material is not a prerequisite for understanding or discussing the results of the oscillator tests. These references are provided only for those who might be interested.

The first oscillator I decided to study was the Morion MV89A, a low phase-noise oscillator that I picked up from ebay (actually I bought 3 of them). I built an enclosure for these, which is not remarkable in any way, but for the sake of completeness, is shown in figure 1.



Figure 1 - Enclosure for MV89A.

Note that there is a switch that selects "Ref" or "Adj". When in the "Ref" position, an internal reference voltage (2.5V) is connected to the frequency adjust pin of the oscillator, keeping the output frequency at 10 MHz. When in the "Adj" position, the frequency adjust pin is connectd to the BNC connector shown at the top of the image. This allows the adjustment of the oscillator's frequency so it can be used as a reference oscillator when measuring phase noise of a second oscillator. This feature is not used in the experiments documented in this post.

Also note in the image that the date on the oscillator is 05/06, which means it is 13 years old. This is mentioned for later discussion.

To test this oscillator, I used the frequency discriminator configuration of an HP11729C (see post 3). To ensure the signal exiting the HP11729C did not go below the noise floor of the spectrum analyzer used in the test setup, I ran an experiment comparing this signal with the SA noise floor. The results of this experiment are documented in post 7. The punchline is: the signal stayed above the noise floor.

I ran some experiments analyzing the phase noise of the MV89A. Before collecting data I let the oscillator and the rest of the test equipment warm up for over an hour. I monitored the frequency of the oscillator using an HP5335A frequency counter. Each test took about 25 minutes to run and during that time the oscillator had a stable frequency of 9,999,999.98 Hz. For those who have read the test setup description (others can ignore the rest of this sentence), the HP11729C input from the oscillator was 2.87 dBm, the IF output was 11.3 dBm and the output of the Delay Complex was 2.41 dBm.

The properties of the low-frequency spectrum analyzer (a PicoScope 4262) used for the experiment are shown in figure 2. Note that the gate time of the collected spectrum is 52.43 seconds. Consequently, the results documented here are for short-term phase noise, not for longer periods such as would be studied to characterize the oscillator's use in, for example, long-term time keeping applications.



Figure 2 - Properties of the Spectrum used for the test (on a PicoScope 4262)

One important parameter missing from figure 2 is the averaging scheme used (this is given in a different window on the PicoScope software). Specifically, it should be noted that 30 segments were power averaged in order to create the spectrum.

After correcting the data according to the steps described in the test setup posts, I plotted the phase noise measured, which is shown in figure 3.



Figure 3 - MV89A measured phase noise

The plot reveals two surprises. First, the phase noise at various offsets was significantly above that given in the (MV89A specification ). In particular, the specification asserts that (for a 5MHz oscillator) the MV89A has the following phase noise characteristics (dBc/Hz) - 1 Hz: -105; 10 Hz: -130 ; 100 Hz: -145 ; 1 KHz: -150 ; and 10 KHz: -155. The corresponding characteristics from the plot are: - 1 Hz: -4.27; 10 Hz: -28.52 ; 100 Hz: -46.80 ; 1 KHz: -65.15. This represents a difference of almost -100 dB. This is discussed below. (for an explanation of the strike through text see this post)

Second, there were significant spurs in the phase noise spectrum below 500 Hz. This is also discussed below.

Beginning with the differences between the published and measured phase noise specs, at first I suspected my test setup equipment or its procedures were either faulty or I was not following the latter precisely. This is still a possibility, but I decided to do a "back of the envelope" check on the results by connecting the MV89A to my Siglent 3032X and look at the spectrum in the immediate vicinity of 10 MHz. This should give a ball-park check on the phase noise data I generated from the test setup. The Siglent was configured according to the following parameters:

• Start Frequency: 9.995 MHz
• Stop Frequency: 10.005 MHz
• RBW : 10 Hz
• Detect type" Average Power"
• Sweep time: auto
• Sweep: Single
• Sweep Mode: FFT

I captured the spectrum results in a CSV file and moved it to my analysis computer. Using Octave, I deleted the spectrum to the left-hand side of the Carrier and then normalized the power in each spectrum bin by subtracting the carrier power (in dBm) from all other bins (this yielded dBc units) and then subtracted 10 dB from each bin to convert the values to dBc/Hz (remember that each bin was 10 Hz wide and that to get power values normalized to 1 Hz, it is necessary to divide by 10 or in the log world to subtract 10 dB). The results are shown in figure 4.



Figure 4 - Phase Noise from Frequency Discriminator versus side bands on Siglen 3032X

It is evident that the two plots converge around 5 KHz. It is important to keep in mind that the Siglent data was collected from a sweep of about 1-2 seconds, so averaging does not occur over the same time period as the frequency discriminator experiments. Figure 5 shows the comparison in the offset frequency range 0-120 Hz.



Figure 5 - Data in Figure 4 zoomed in to 0 - 120 Hz

Two conclusions arise from these two images. First, it is plausible that the Frequency Disciminator analysis of phase noise is correct (but see below for a discussion why the data may be corrupted by instrument noise floors). The convergence at 5 KHz supports such a conclusion. Second, below ~40 Hz the Frequency Discriminator results are questionable. This agrees with the conclusion from the noise floor analysis in post 7 and with the property that the frequency discriminator configuration is not suitable for analyzing phase noise close to the carrier.

How is the discrepancy between the values of the Frequency Disciminator analysis of phase noise and those of the published specs explained? After all, -100 dB is an enormous difference, representing a ratio of 10^-10. Several possible factors may be involved:

• It isn't clear from the spec what was the length of time over which was collected the data leading to the spec. As stated above, the time period used for this experiment was 52.43 seconds. 30 data segments with this gate time were collected and power averaged to generate the spectrum. If the spec data is based on much shorter or longer segment gate times, then it is possible some of the disagreement is explained by this fact. (NB: the spec also does not indicate whether averaging was used or, if so, how many segments were averaged.)
• The published phase noise specs for the MV89A are for a 5 MHz oscillator. The MV89A is a 10 MHz oscillator, which suggests it is frequency doubling 5 MHz. This would degrade the phase noise characteristics of the source oscillator. However, doubling the frequency should impose a degradation of about -6 dB (in theory, although in practice it may be more). See this article.
• One factor is the age of the MV89A - 13 years. One would not expect the phase noise characteristics to degrade by -100 dB in this time period, but some degradation wouldn't be surprising.
• An important consideration is the noise floor characteristics of the HP11729C and the Siglent 3032X. In the latter case, the Siglent 3032X datasheet specifies phase noise of <-95dBc/Hz, <-98dBc/Hz typical at 10 KHz offset for a 1 GHz carrier. So, the comparison of the Frequency Discriminator result with the Siglent data may not be valid as any phase noise <-100 dBc/Hz may be in the phase noise floor of the Siglent (there is no spec data for lower offset or carrier frequencies). Also, the HP11729C has the following residual noise specs (dBc/Hz @ <1.28 GHz carrier): 10 Hz: -115; 100 Hz: -126 Hz; 1KHz: -135; 10 KHz: -145. This is much better than the Siglent, but still above the specs for the MV89A (5Mhz). Note: these noise figures are for residual noise, not phase noise.
  
  The spec data isn't sufficiently precise or complete to demonstrate that the results of the experiment are corrupted by noise floor limits. The offset values generated by the experiment are significantly above the residual noise spec values, so it isn't clear the experiment hit the HP11729C residual noise floor, in which case the difference between the MV89A phase noise spec and the results remain unexplained.

The bottom line is the results are inconclusive. More work is required to determine the correct phase noise spectrum for a 13 year-old MV89A. If noise floor problems are the cause, then it will not be possible to resolve this issue using the Frequency Discriminator configuration of the HP11729C.

The second surprise from the experimental results were the spurs observed in the phase noise spectrum. Figure 6 shows the spectrum zoomed in to 0-500 Hz.



Figure 6 - Phase noise spectrum zoomed to 0-500 Hz.

From the octave data, I have inserted the frequency and power values of the spurs into the spectrum. They occur at 60 Hz, 180 Hz and 300 Hz. They obviously represent power line 60 Hz noise modulating the carrier. The fact that it is the odd harmonics of 60 Hz that appear in the spectrum may have an explanation, but I am not aware of it (perhaps someone with more extensive knowledge of modulation theory would know.)

A likely candidate for the source of the 60 Hz contamination is the supply powering the oscilllator. This was a Rigol DP832, which is a triple linear power supply. In the case of the MV89A, 12 volts drives the oscillator. I connected the output of the DP832 @ 12V to the PicoScope input, using AC coupling, which produced the spectrum shown in Figure 7.



Figure 7 - DP832 12V power supply spectrum.

At the top of each spur in this plot is its offset frequency (this is different than figure 6, which labels each spur by its power value). What is somewhat suprising is lack of correlation with the spurs in the phase noise plot. In particular:

• There is no significant spur at 180 Hz, rather there is a major spur at 168 Hz.
• There is a huge spur at 250 Hz, for which there is no corresponding spur in the phase noise spectrum.
• The spur at 300 Hz is evident, but there also are spurs at 328 Hz and 332 Hz  that have no corresponding spur in the phase noise spectrum.
• There are spurs at 41 Hz and 82 Hz that have no correspondents in the phase noise spectrum.

So, the power supply may not be the source of the power line noise that modulates the carrier, resulting in the spurs. This is something that requires further investigation.

Preliminary Conclusions

The experimental results presented in this post seem to raise more questions than they answer. Here is summarized the preliminary conclusions derived from the results.

• The phase noise of the MV89A published in the specs could not be reproduced. This may have occured because there were noise floor limitations that prevented an accurate measurement. However, the published noise floor specs of the measurement instruments do not supply any obvious support for this explanation. The residual noise floor of the HP11729C was considerably below that of the measured (and corrected) noise floor of the MV89A dervied by the experiment. The noise floor of the spectrum analyzer (a PicoScope 4262) was sufficiently below the signal output of the frequency discriminator to suggest it was not an influence. Why the measured noise floor is almost 100 dB above the published specs has not been explained; although this magnitude of difference strongly suggests something isn't right.
• The spurs in the phase noise measurements at 60 Hz and its odd harmonics could not be explained by corruption of the power supply voltage. However, one conclusion is these spurs would introduce significant frequency fluctuations at those offsets. So, when designing an oscillator circuit for use in an application for which phase noise is an issue, a designer should pay particular attention to low frequency noise to ensure it doesn't introduce deterministic adulteration of the oscillator output.
• The results of this experiment focuses on short-term phase noise (on the order of a minute). They should not be considered valid in applications for which phase noise over longer periods is important. Study of longer-term phase noise is a future objective.
• It is well-known that a frequency discriminator cannot analyze phase noise close to the carrier. The results of this experiment suggest that for offset frequencies below 40 Hz, the phase noise results cannot be considered valid.

[SoundTech-LG]

Not only are these MV89A oscillators on Ebay retired, there are also many fakes being sold. Labels produced that look similar to Morion labels. China seems adept at this. The first one I got seemed to be the genuine article. I left it to burn-in for over a year. Disconnected it from power, let it cool. It never powered up again. The second one I bought was certainly a fake. Not even the same mechanical dimensions. It also had poor measured specs. So, not sure what is currently being offered there, but certainly the buyer should beware.

[bob91343]

I think you are having a lot of fun with this.  (Please correct spelling, the word is hobbyist.)

The absence of even harmonics of the spurious 60 Hz frequncy implies that the source is symmetrical.  A symmetrical wave (positive excursion being a mirror image of negative excursion) has no even harmonics.  The presence of even harmonics indicaates a difference between the two polarities.

Your oddball frequency spurious signals may be internally generated within the test equipment.  Spectrum analyzers are notorious for this.  It takes an experienced eye to differentiate between internally generated signals and those that are contained within the input signal.

I am by no means an expert in these matters so I will stop guessing.

[dnessett]

(Please correct spelling, the word is hobbyist.)

Done. I hope I found them all.

Thanks,

Dan

[dnessett]

Not only are these MV89A oscillators on Ebay retired, there are also many fakes being sold. Labels produced that look similar to Morion labels. China seems adept at this. The first one I got seemed to be the genuine article. I left it to burn-in for over a year. Disconnected it from power, let it cool. It never powered up again. The second one I bought was certainly a fake. Not even the same mechanical dimensions. It also had poor measured specs. So, not sure what is currently being offered there, but certainly the buyer should beware.

Any tell-tale external properties that would identify a fake?

[awallin]

FWIW I got 3 MV89A's from ebay in around 2016 and here are some old measurements, two MV89As against eachother, and also against a H-maser
http://www.anderswallin.net/2016/09/morion-mv89a-ocxos/

I may repeat this at some point if I get inspired. The distribution-amplifiers we've made consistently measure at below -160dBc/Hz for a 10MHz carrier, so in principle the measurement-system is good at least to -160...170 dBc/Hz @ 10MHz - but in practice achieving this with independent OCXO or H-maser sources has been difficult.

IMO the way to get accurate numbers on hobby budget is SDR: https://arxiv.org/abs/1605.03505
the low numbers -150dBc/Hz and lower will require two ADCs per channel (4 in total) and cross-correlation - not many hobbyists have demonstrated this yet, see e.g. http://www.aholme.co.uk/PhaseNoise/Main.htm

[dnessett]

FWIW I got 3 MV89A's from ebay in around 2016 and here are some old measurements, two MV89As against eachother, and also against a H-maser
http://www.anderswallin.net/2016/09/morion-mv89a-ocxos/

I may repeat this at some point if I get inspired. The distribution-amplifiers we've made consistently measure at below -160dBc/Hz for a 10MHz carrier, so in principle the measurement-system is good at least to -160...170 dBc/Hz @ 10MHz - but in practice achieving this with independent OCXO or H-maser sources has been difficult.

Thanks for the info, awallin. I am reading the paper you linked, but it is going to take some time, as I have my hands full with other tasks at the moment (not all electronics related). So, I wanted to get back to you with some questions before then.

Do you happen to remember what corrections you applied to the raw data to get the MV89A phase noise? For example, if you used an FFT SA, did you adjust for processing gain? Or, if you used a swept-tuned SA, did you correct for effective noise bandwidth? Just asking, as one possibility is I am not properly applying the necessary corrections or am applying erroneous corrections to my data. Of course, I am using a delay-line frequency discriminator, whereas you are using (in HP terminology) a phase detector. So, the corrections would be different. You mention that measuring down to -160dBc/Hz for 10 MHz is difficult. What problems did you have and how did you overcome them? (for an explanation of the strike through text see this post)

IMO the way to get accurate numbers on hobby budget is SDR: https://arxiv.org/abs/1605.03505
the low numbers -150dBc/Hz and lower will require two ADCs per channel (4 in total) and cross-correlation - not many hobbyists have demonstrated this yet, see e.g. http://www.aholme.co.uk/PhaseNoise/Main.htm

As I said, I am reading the paper and have not finished it. It appears that the SDR is implemented on an FPGA. While I imagine the FPGAs are commercially available, is the configuration data also published and free to use? Or is the FPGA with the configuration commercially available?

Once again, thanks for the response.

[jpb]

As I said, I am reading the paper and have not finished it. It appears that the SDR is implemented on an FPGA. While I imagine the FPGAs are commercially available, is the configuration data also published and free to use? Or is the FPGA with the configuration commercially available?

I've just read the papers (first pass through - I'll go back to them when I have more time), very interesting.

In the second paper the FPGA he uses is on a demo board as are the ADCs (on a second demo board) though he had to physically cut the corner out to get the two to directly connect.
I think that he provides links to the verilog(?) code as well as other code on his web site but I've not checked this out in detail.
He does use specifically designed power splitter boards but I imagine you could use something like the LTC6597 demo boards instead as they add very little phase noise. The oscillator is also homebrew, I think, but not specifically for this project.

[dnessett]

IMO the way to get accurate numbers on hobby budget is SDR: https://arxiv.org/abs/1605.03505
the low numbers -150dBc/Hz and lower will require two ADCs per channel (4 in total) and cross-correlation - not many hobbyists have demonstrated this yet, see e.g. http://www.aholme.co.uk/PhaseNoise/Main.htm

Hi awallin,

I finally had time to finish reading the paper on SDR based phase noise measurement that you referenced. Here are some comments:

• As a concept it seems like an interesting and fruitful approach. But, with only the paper for guidance, there are practical difficulties, which are mentioned in the next bullets.
• The SDR used for the body of the results reported in the paper is the Ettus USRP N210. New, this costs $1,943. I did find a used unit on ebay, but its cost was still $1,690. In addition, at least one external ADC is required, which in the paper is identified as a TI ADS62P44, which on mouser costs $83.15 in units of 1. This is just the raw ADC, which will need supporting electronics, power supply, etc. In addition, other parts, such as PLL, reference oscillator, etc. are required. So, as a rough estimate, the single channel Software-defined radio phase noise measurement system will cost in the vicinity of ~ $2,500. This is not within my budget, but perhaps others might have the excess recreational financial resources to build this test setup.
• The post processing of the data output by the SDR test setup is only sketched out. For example, the quantity z(t_k) shown as the output of the single channel SDR is the input signal downconverted  to a lower frequency. This signal is post-processed according to the block diagram shown in figure 1(d). However, there are insufficient details to actually implement this post-processing step. Furthermore, there is no mention of the corrections required for backing out processing gain for an FFT spectrum analyzer, compenstating for equivalent noise bandwidth due to the window employed to reduce spectral leakage, etc.  (for an explanation of the strike through text see this post) In other words, this is an academic paper, which assumes the reader either isn't interested in the practical details or can figure them out.
  

So, while this avenue of approach to analyzing phase noise seems promising, a hobbyist, such as I, cannot follow it until more details of the experimental setup are available.

[dnessett]

I decided to test the other 2 Morion MV89A oscillators I bought from ebay to see if the results might shed some light on the issues I documented in this post. The MV89A that generated the results in that post was actually the third I built an enclosure for. To keep the results of the other two distinct, I have decided to label it MV89A-3 and the other two MV89A-1 and MV89A-2.

After warming up MV89A-1 for an hour, its frequency stablized at about 10,000,003 Hz. This was not consistent with my original testing, which observed all oscillators to run to within a few hundredths of a Hz of 10 MHz. To correct the 3 Hz discrepancy, I used the adjust circuit built into each enclosure to bring MV89A-1 down to 10,000,000.05 Hz. This required applying 0.02 V to the Freq-Cont X-Osc input of the adjust circuit.

For those familiar with the test setup procedure, input power to the HP11729C was 2.78 dBm, IF output from it was 11.28 dBm and output from the delay complex was 2.44 dBm.

When I warmed up MV89A-2 for an hour, the carrier frequency wouldn't stabilize. In particular the HP 5335A frequency counter flucuated randomly between 9,999,998.32 and 9,999,998.88. This was after bringing the oscillator center frequency up from 9,999,996.xx by setting the frequency adjust circuit input, Freq-Cont X-Osc to 4.92V. Because of the frequency instabilities in the MV89A-2, I decided not to compare it against the other two MV89As.

Figure 1 plots the raw output of the Frequency Discriminator for MV89A-1 and MV89A-3 against the PicoScope noise floor. As with the previous figure comparing MV89A raw output, the removal of processing gain and combining bins to achieve a width of 1 Hz were the only corrections applied to each data set. Thus, as with the other figure, the blue and green plots do not represent the phase noise of these oscillators.



Figure 1 - MV89A-1 raw output from Frequency Discriminator (green) versus MV89A-3 raw output (blue) versus PicoScope Noise Floor.

The first characteristic that stands out is MV89A-1 raw output is fairly flat at all frequencies except those very close to the carrier, whereas MV89A-3 raw output traces an arc starting at around 2.5 KHz down to the carrier frequency. The MV89A-1 raw output still does not hit the noise floor of the PicoScope 4262 until very close to the carrier frequency, but it is closer.

The flatness of the MV89A-1 raw data raised the possibility that it is hitting the noise floor of the Frequency Discriminator. A typical Frequency Discriminator noise floor is given on page 1-8 of the HP11729C manual (in dBc/Hz) 1 Hz: -54; 10 Hz: -84; 100 Hz: -104; 1 KHz: -116; 10 KHz: -139. In order to compare these values to the MV89A data, the latter must be converted to phase noise, which is shown in figure 2.



Figure 2 - MV89A-1 phase noise (green) versus MV89A-3 phase noise (blue). (for an explanation of the strike through text see this post)

A quick glance at the figure indicates that both plots remain above this "typical" Frequency Discriminator noise floor. However, whether the configuration of the test setup yields a "typical noise floor" is an open question (since the manual doesn't specify how to measure this quantity).

Conclusions

The strange "hump" in the raw data of MV89A-3 suggests a problem with that oscillator. MV89A-2 also has properties (unstable warmed up carrier frequency) that suggests it too has abnormalities. MV89A-1 wouldn't oscillate at the spec'd frequency without adjustment. Furthermore, the measured phase noise of MV89A-1 and MV89A-3 is sigificantly above that specified for a new MV89A.

What can a hobbyist conclude from the analysis so far? Perhaps it is that using a 13 year-old oscillator that is removed from operating equipment (for reasons not given) is a chancy proposition. The failure of MV89A-2 to reach a stable carrier frequency indicates internal problems. That MV89A-1 was off by 3 Hz from 10 MHz is also an indication of internal problems. There are other reports of MV89A failures, which suggest it is not a good candidate for use in hobbyist projects.

[texaspyro]

Check the MV89 output levels.  They have a known failure mode when the output level drops to very low levels.  Caused by an coupling capacitor with bad solder joints.

[dnessett]

Check the MV89 output levels.  They have a known failure mode when the output level drops to very low levels.  Caused by an coupling capacitor with bad solder joints.

Thanks for the info. However, I measured the input power level for all three oscillators and they were all greater than 3 dBm (.359 V_RMS or .9 V_P-P). I know this because I had to attenuate the input to 3 dBm or lower to ensure the HP11729C didn't go into compression.

[dnessett]

Having run experiments for a (ostensibly) low phase noise oscillator, I decided to examine the other end of the phase noise spectrum and evaluate a generic XO can oscillator. I bought 10 of these for $1.39 each from Jameco. These are low end devices without temperature or any other type of envirommental compensation. However, they do contain sufficient electronics so no external passive devices are necessary to their operation.

There is surprisingly little published data on these devices (at least on the internet). I could find no datasheets for any of the devides shipped by Jameco in fulfilment of the order. So, examining them may provide others with selection data that does not exist or exists in hard to find places.

Since these oscillators generate a square wave at 10 MHz, I put a 10 MHz low pass filter (LPF) on their output so it was suitable for the Frequency Discriminator input (see figure 1). The device tested was warmed up for 1/2 hour before testing. Since these devices have no oven or other temperature compensating electronics, "warming them up" is something of a misnomer. However, as seen from the data below, there was some linear changes to their output frequency in respect to time. What caused these changes is unclear.



Figure 1 - Connections to generic can XO showing the 10 MHz LPF (silver cylinder to the right of the DIP socket)

The output of the LPF fed the directional coupler shown in figure 1 of this post.

The particular device I tested was labeled "TMJ" on the top of the can (see figure 2)



Figure 2 - Top of XO can identifying the device as TMJ

I could find no information on this device on the internet, but did find a pin diagram (see page 16) for generic XO cans that I used to connect power, ground and RF output lines. Again, to ensure the outupt of the frequency discriminator did not hit the noise floor of the PicoScope 4262, I plotted the raw output of the frequency discriminator compensated only for processing gain and a bin width of 1 Hz. This is shown in figure 3, which shows not only the TMJ raw output compared with the PicoScope noise floor, but also the raw output of one of the MV89As (MV89A-1).



Figure 3 - TMJ and MV89A-1 raw output versus PicoScope noise floor

As is evident from the figure, the TMJ raw output stayed above the PicoScope noise floor and in addition was above the raw output of the MV89A-1. What is interesting is the raw output isn't much above that of the MV89A-1, suggesting again that the used MV89As I purchased were defective.

Figure 4 plots the phase noise of both the TMJ and the MV89A-1. Again, it is striking that while the MV89A-1 has slightly better phase noise than the TMJ, it isn't much better.



Figure 4 - TMJ and MV89A-1 phase noise (for an explanation of the strike through text see this post)

One characteristic of the TMJ that stands out is the much greater number of and magnitude of spurs on its phase noise spectrum. Figure 5 shows these spurs in greater detail between 0 and 1 KHz.



Figure 5 - TMJ and MV89A-1 phase noise between 0 and 1 KHz

The top of each spur is labeled with its power and the x-axis shows where in the spectrum each spur resides. It is clear that these spurs are the result of 60 Hz modulating the carrier. What is different between a similar plot for the MV89As is not only are the odd harmonics of 60 Hz appearing in the spectrum, so are the even harmonics. This suggests the power supply driving the XO can (5 Volts) is not filtered as well by the device as the power supply driving the MV89As (also 5 Volts). The appearance of even harmonics suggests (according to the comments by Bob9134) that the 60 Hz ripple in the power supply is not symmetrical between the positive and negative voltage swings.

I also used the statistics capability of the HP5335A to calculate the mean and standard deviation of the output frequency over a sample period. This period is defined by the number of samples collected (100) and the gate time of the HP5335A (1 sec). So samples are over 100 seconds, one sample per second.

I recorded the mean frequency over 5 sample periods. The values were:

• 9,999,941.18 Hz
• 9,999,941.15 Hz
• 9,999,941.16 Hz
• 9,999,941.01 Hz
• 9,999,940.95 Hz

It is evident that even after 1/2 hour warmup the device is still not stabilized around an obvious carrier frequency.

I also recorded the standard deviation of a different set of samples with the same sample parameters (the HP5335A does not support collecting both the mean and standard deviation of a sample set). Again, I recorded the standard deviation over 5 sample periods, yielding the following values:

• 123*10^-3 Hz
• 34.7*10^-3 Hz
• 42.9*10^-3 Hz
• 92.1*10^-3 Hz
• 54.5*10^-3 Hz

If it is assumed the frequency centers of the oscillator are normally distributed around the true (as opposed to the sample) mean, this implies 68% of the samples were within plus or minus the measured standard deviation. For example, if the mean was 9,999,940.50 and standard deviation 90 mHz, then 68% of the sampled means were within the interval, [9,999,939.60 Hz, 9,999,941.40 Hz]. Since that implies 32% of the samples were outside these bounds, it is clear that XO can oscillators are not very stable in terms of their center frequency.

[tomato]

Your test setup is a big antenna.  It's likely that the peaks in your data have nothing to do with the oscillator's inherent phase noise.

[dnessett]

Your test setup is a big antenna.  It's likely that the peaks in your data have nothing to do with the oscillator's inherent phase noise.

I had thought of that, but had previously run an experiment to demonstrate it is not the problem. However, it is totally appropriate for you to raise this issue, since I did not mention this possibility in my post nor provide any evidence to support the conclusion.

The only part of the "TMJ test setup" that is susceptible to acting like an antenna is that part comprising the wires and grabber connections to the DIP (including the DIP socket itself). All other parts are connected with RG-58 and therefore properly shielded. To test whether the "exposed parts" were picking up 60 Hz eminations, I put the susceptible part of the test setup in a makeshift Faraday bag. First I wrapped the exposed wiring in plastic wrap to protect it from shorting (figure 1).



Figure 1 - Exposed parts of the TMJ setup wrapped in plastic wrap.

Then I put the plastic wrap on a piece of aluminum foil and folded it over to shield the exposed electronics (figure 2).



Figure 2 - Plastic wrap covered with aluminum foil

I then ran the same test as before. Figure 4 shows the phase noise results.



Figure 3 - Phase noise of TMJ in Faraday bag

Comparing Figure 3 with Figure 5 of my last post



it is apparent that putting the exposed parts of the "TMJ test setup" in a Faraday bag sometimes decreases the spur power (e.g., at 60 and 120 Hz) and sometimes increases the spur power (e.g., at 180 and 300 Hz). In any case, the differences are not great, all less than 10 dB.

So, I conclude that it is not parts of the test setup picking up EMI from the surrounding environment that is causing the 60 Hz and harmonic spurs.

[dnessett]

There were two other brands of XO can in the batch sent me by Jameco. Since I could find no information on the manufacturer of these devices, I decided to call the first (figure 1) TShield, and the second (figure 2) CQ.



Figure 1 - TShield XO can



Figure 2 - CQ XO can

Figure 3 shows a comparison of the phase noise of the TMJ-1, TShield-1 and CQ-1 oscillators (the "1" simply differentiates between these pieces and others of the same brand in the batch Jameco sent me.)



Figure 3 - TMJ-1 versus TShield-1 versus CQ-1 Phase Noise

The image reveals that the best phase noise is exhibited by the TMJ-1, followed by the CQ-1 and then the TShield-1. However, the differences are slight, being no more than 3-5 dB. (for an explanation of the strike through text see this post)

The image also shows that the CQ-1 has very prominent spurs at 1 KHz, 2 KHz and 4 KHz. Additional spurs of note in the spectrum of the CQ-1 XO can occur between 2-3 KHz.

For those who are interested (and have read the test setup material) here are the power measurements for the Frequency Discriminator. TShield-1: In - 2.66 dBm; IF - 11.36 dBm; Out - 2.32 dBm. CQ-1: In - 2.91 dBm; IF - 11.44 dBm; Out - 2.31 dBm.

I also measured the mean frequency and standard deviation of (100) samples captured by the HP5335A. I combine all three XO can data sets in the following table.

TMJ-1	TShield-1	CQ-1
Mean Frequency	Mean Frequency	Mean Frequency
9,999,941.18 Hz	10,000,049.51 Hz	9,999,811.90 Hz
9,999,941.15 Hz	10,000,048.81 Hz	9,999,812.77 Hz
9,999,941.16 Hz	10,000,048.68 Hz	9,999,813.40 Hz
9,999,941.01 Hz	10,000,048.63 Hz	9,999,813.31 Hz
9,999,940.95 Hz	10,000,048.59 Hz	9,999,813.41 Hz
Std Dev	Std Dev	Std Dev
123*10-3 Hz	41.1*10-3 Hz	235*10-3 Hz
34.7*10-3 Hz	29.5*10-3 Hz	56.7*10-3 Hz
42.9*10-3 Hz	36.3*10-3 Hz	489o*10-3 Hz
92.1*10-3 Hz	25.6*10-3 Hz	617*10-3 Hz
54.5*10-3 Hz	55.7*10-3 Hz	305*10-3 Hz

Figure 4 - Carrier Mean Frequency and Standard Deviation for the TMJ-1, TShield-1 and CQ-1 XO cans.

There are some interesting characteristics displayed in this data. The data is shown in the order I captured it. So, the second Mean Frequency value is that captured immediately after the first Mean Frequency value. This is also true for the standard deviation, although there doesn't seem to be much time dependent structure in that data.

Notice that the Mean Frequency for TMJ-1 and TShield-1 (roughly) decreases over time. This may be due to temperature changes during the measurement period. However, CQ-1 Mean Frequency (roughly) increases over time. The data for the TMJ-1 was collected on a different day than that for the TShield-1 and CQ-1. On the other hand the data for the TShield-1 and CQ-1 were collected on the same day during an afternoon session. The temperature of the room did not decrease during this session (if anything it increased, although my house is air conditioned, so its temperature is fairly stable)

I am not expert in temperature effects on crystals, so it is possible that the same temperature variation might have caused one device to increase in Mean Frequency and the other to decrease. I don't know. Since I did not make a record of the temperature during the test period, there is really no way to determine why the Mean Frequency of one device decreased and the other increased.

Another interesting property is found in the third entry under CQ-1 standard deviation. Note the "o" after the numver (i.e, 489o*10^-3 Hz). This is not a typo. This is what the HP5335A displayed. I couldn't find in the manual an explanation for this, but I suspect the "o" means overflow. This is probably due to the algorithm used to compute standard deviation (see second answer in this post).

Another interesting thing to notice in the standard deviation data is the CQ-1 displays much higher values than the other two devices. That and the already noted behavior of increasing in Mean Frequency suggests this device is constructed differently than the other two. However, it isn't possible to say all CQ branded devices would display these characteristics. (Unfortunately, the batch sent to me by Jameco only has two CQ oscillators, which is insufficient to draw any conclusions with respect to this brand.)

The Mean Frequency and Standard Deviation figures were calculated by the HP5335A over an interval with period 1 min 40 seconds. To see such variation in the Mean Frequency of these devices during such a short period indicates they are not suited for even short-term time-keeping tasks. For example, there are many physics experiments in high school that last only a couple of minutes. You don't need a Rubidium oscillator for such experiments. On the other hand, these cheap oscillators are not candidates for such use either. Perhaps the one use they are suited for is as clocks driving digital hardware that has no real-time timing function (e.g., an electronic calculator)

[edigi]

I've read this thread with great interest. The reason is that I've built a hobbyist GPS disciplined frequency counter with 10 digit/s precision.

Basically I use the kind of architecture described in this paper
http://n1.taur.dk/permanent/frequencymeasurement.pdf
but I use for measuring the fraction TI TDC7200s (time to digital converter) instead of the analogues stuff. In my use it has a resolution of roughly 60ps so 10 digit/s can be achieved relatively effortlessly (and cheap as well, the whole thing is below 100 USD).

Naturally any frequency counter needs a good reference. Since in my case the reference is compensated using the GPS 1pps signal it does not have to be incredible precise (and doesn't need any kind of long term stability either). However it should have small phase noise (and good short term stability as well).
The big question is actually how small and how much does it matter if statistics are used?
Since even significantly more expensive (and probably more complex to use) TDC chips don't gain too much, my main target is improving precision via statistics. So basically going for a statistical interpolating reciprocal counter.

In fact I've tried to make statics of roughly 100 time stamps in the MCU I use (which brings at best 1 extra digit) but with even a low end and cheap FPGA statistics of 10000 or more timestamps is not entirely out of range (I haven't gone that direction yet). That would mean a 12 digit/s counter (due to FGPA may not fit to 100 USD budget but that's not my most important concern).

Measuring back its own reference of the current counter shows little fluctuation even with 12 digits (not round because it's real GPS compensated value), so it should be OK.
However even with 10 digits I've noticed that measuring non reference TCXO means some fluctuation in the 10th digit. I heavily suspect that it's the phase noise (both reference and the TCXO as source)...

What do you think?
Maybe in the other thread you've mentioned about the cyclostationary nature of the oscillators. Do you have concrete results about this e.g. period etc?

(Note: The older screenshot is from the time I was using 80 MHz reference, since that I've changed to the standard 10 MHz, precision is not changed though as it's determined by the TDC.)

[dnessett]

edigi,

I'm sorry it took so long for me to reply. For some reason the notifications from this topic wound up in my junk mail box and it was only by chance that I found it.

Your project is interesting, especially keeping the costs so low. However, we have different objectives. You are focused on long-term time-keeping. On the other hand, I am interested in evaluating oscillators for a wide variety of uses. Long-term time-keeping is one of those uses; although nothing I have done so far addresses this. But, there are others, such as short-term time-keeping (over periods of several minutes), hobbyist doppler radar (e.g., to keep track of model drones), hobbyist spread spectrum communications (which requires precise chirp signals), and so forth.

I would encourage you to document your project in more detail. For example, provide details of your implementation, how you obtained the measurements you cite, etc. This should be the subject of a new topic in the metrology category.

[dnessett]

Given my experience with the Morion MV89A oscillators, I decided to buy a new low-phase noise oscillator to see if the delay line Frequency Discriminator would properly measure its phase noise or if I was running up against limitations in the Frequency Discriminator configuration of the HP11729C. The oscillator I chose was a Wenzel HF ONYZ IV, specifically, part number 501-22578-04, which has these typical phase noise properties:

10 Hz	100 Hz	1 KHz	10 KHz	100 KHz
-135 dBc/Hz	-160 dBc/Hz	-163 dBc/Hz	-165 dBc/Hz	-165 dBc/Hz

These phase noise figures are well below those of the Morion MV89As. I set up the Frequency Discriminator with the following signal strength values: 1) In - 3.23 dBm; 2) IF - 11.24 dBm; and 3) 2.05 dBm. Surprisingly, the Wenzel output frequency measured by my HP5335 was 9,999,996.24 (with the mHz figure varying between .25 and .23). Given the Wenzel is a brand new device, I suspect the HP3553 has gone out of calibration.

Figure 1 shows the output of the Frequency Discriminator versus the PicoScope noise floor.



Figure 1 - Wenzel Frequency Discrimintor Output versus PicoScope Noise Floor

As for the other oscillators, the Wenzel Frequency Discriminator output is well above this noise floor.

Figure 2 shows the calculated Phase noise of the Wenzel verus the MV89A-1 oscillator.



Figure 2 - Phase Noise Measurements of Wenzel and MV89A Oscillators Using the Frequency Discriminator Configuration of the HP11729C

The two plots are virtually identical (except for the spurs related to 60 Hz harmonics). As the phase noise values for the Wenzel is significantly above the values quoted in its specs, this is strong evidence that both the MV89A-1 and Wenzel are hitting the noise floor (sensitivity floor?) of the Frequency Discriminator. (for an explanation of the strike through text see this post)

So, it seems this configuration isn't very useful for measuring phase noise accurately, at least at low offset frequencies (a property attested by a number of experienced engineers). This means I need to figure out how to set up the HP11729C in its Phase Detector configuration and run some experiments on oscillators using it. That is the next goal in this investigation.

[dnessett]

The next step in analyzing the phase noise of hobbyiest oscillators is to use the phase detector mode of the HP11729C to run experiments. I have documented both the mechanical architecture and the test set up procedure in separate messages on this forum. Here is presented the first results using these to run and analyze experiments.

Bliley NV47A1282 Experimental Results

The first oscillator I chose to examine is a used Bliley NV47A1282 OCXO purchased on ebay. In fact I purchased 3 of these and analyzed them together. Its specification is found here.

I originally had intended to use the phase detector set up to run experiments on the generic XO can oscillators I purchased from Jameco. However, none of these oscillators had carrier frequencies close enough to 10 MHz to fulfill the requirement necessary to use the HP11729C phase detector documented in the links cited above. Specifically, the quadrature phase lock loop in the HP11729C requires the carrier frequency to be within +/- 10 HZ of 10 MHz. None of the generic XO can oscillators met this requirment.

The test set up procedure requires setting power levels and measuring various characteristics before making the final phase noise measurement. The reader is invited to read the test set up procedure to properly understand these. Here is presented the requisite values without explanation for each of the three oscillators tested, which are identified as Bliley NV47A1281-1, NV47A1281-2, and NV47A1281-3.

Bliley NV47A1282-1

Ref Osc Frequency: 10,000,002 Hz

Ref Osc in: 0.73 dBm
DUT Osc in: 2.46 dBm
DG1022 in: -39.25 dBm
Delta_SB_Cal: -39.98 dBm
P_cal: -46.81 dBm
-Delta_SB_Cal-P_cal = 6.63 dBm

Bliley NV47A1282-2

Ref Osc Frequency: 9,999,999 Hz

Ref Osc in: 0.73 dBm
DUT Osc in: 2.47 dBm
DG1022 in: -39.31 dBm
Delta_SB_Cal: -40.04 dBm
P_cal: -46.64 dBm
-Delta_SB_Cal-P_cal = 6.60 dBm

Bliley NV47A1282-3

Ref Osc Frequency: 10,000,003 Hz

Ref Osc in: 0.61 dBm
DUT Osc in: 2.44 dBm
DG1022 in: -39.39 dBm
Delta_SB_Cal: -40 dBm
P_cal: -46.58 dBm
-Delta_SB_Cal-P_cal = 6.58 dBm

The parameters used on the PicoScope 4262 are:

Sample interval: 5 us
Sample rate: 200 kS/s
No. samples: 2,097,148
No. bins: 1,048,576
Window: Blackman-Harris
Bin width: 95.37 mHz
Time gate: 10.49 sec
Number of segments used for averaging: 30
Span: 0-100KHz
Internal 200 KHz filter: On

I ran tests for LNA gain settings of 1000X, 100X and 10X. However, I only present the results for a gain setting of 1000X. I have only briefly looked at the difference between the phase noise spectrum generated using 1000X and 100X gain. There was no apparent difference. However, this is an as yet unexplored facet of the experimental results.

Figures 1-3 show the phase noise spectra for the three oscillators tested.



Figure 1 - Phase Noise Spectrum for NV47A1282-1



Figure 2 - Phase Noise Spectrum for NV47A1282-2



Figure 3 - Phase Noise Spectrum for NV47A1282-3

I have inserted rough slope lines in each plot to indicate potential power law behavior in the spectra. In figures 1 and 3 the red slope line indicates the rough angle of the spectra between 1 and 100 Hz. The black slope line indicates the rough angle of the spectra from 100 Hz to 100 KHz.

Concentrating on figures 1 and 3, the most prominent characteristic is the large number of spurs in the spectra. These are especially powerful and numerous from 10 to 100 Hz. I have no idea what causes these, but there is regularity in their frequency difference, which is examined below.

Figure 2 is significantly different than the other two figures. First, there is a completely unexpected positve power law slope between roughly 33 and 185 Hz. I have no idea what causes this completely unexpected power dip in the phase noise spectrum (when examined from 185 to 33 Hz).

The power law slope of the figure 2 spectrum above 185 Hz is roughly the same as that in figure 3, but in terms of spectral power the line in figure 2 runs from about -135 to -145, while the line in figure 3 runs from about -125 to -138. This contrasts with the black line slope in figure 1, which runs from about -131 to -135.

Another difference between figures 1/3 and figure 2 is there are considerably less spurs in the spectrum of the latter. The phase noise is much more "noise-like".

One interesting, if not prominent, characteristic of the figure 3 plot is the little bump up in the power of the spectrum around 50 to 100 KHz. There is another bump in figure 2 between roughly 70 and 100 KHz. There is no corresponding bump in figure 1. I have no explanation for these.

Figure 4 zooms into the spectrum of the NV47A1282-3 using a span of 0 to 10 KHz.



Figure 4 - Phase Noise Spectrum from 0 to 10 KHz for the NV47A1282-3

This spectrum is very similar to that of the NV47A1282-1. I have plotted several additional power slope lines in this figure. The solid red and black lines represent an approximation of the floor of the spectrum from (respectively) 0 - 100Hz and 100Hz - 10KHz. The dotted lines are extrapolations of the phase noise figures given in the specification (see above). As can be seen, the experimental results suggest a degradation in phase noise performance of 5 dB (from 10Hz to about 100 Hz) and 10 dB (from 1KHz to 10 KHz). Between 100Hz and 1 KHz the degradation widens from 5 to 10 dB. Whether this degradation is due to aging or simply to overoptimistic marketing bias in the specification cannot be determined. I tried to find a new NV47A1282 to buy and test, but they are no longer made, so comparing the experimental results of testing both a new and used oscillator is not possible.

Also plotted in figure 4 is a 1/f power law line with slope -1. Comparing this with the experimental power law line from 10-100 Hz shows the latter represents flicker phase noise in the spectrum. However, the experimental power law line from 100Hz - 10 KHz does not have zero slope, which is what is expected for white noise. In fact it is somewhat negative indicating a slightly pink spectrum. This is yet another example of the validity of the old aphorism, "In theory, theory and practice are the same; but in practice they are different."

I decided to examine the spurs in the spectrum of NV47A1282-3 from 10-100Hz in more detail. Figure 5 shows them using a log-linear plot before the bins are summed to produce bin widths of 1 Hz (i.e., the bin width of the data used for the plot is 95.37 mHz).



Figure 5 - Spurs between 10-100Hz (log-linear plot)

There is obvious regularity in the frequency of these, which becomes apparent by removing the non-spur data and tabulating the results. The following list shows the distance in Hz between the spur maximum values (the spurs spanned multiple bins) from 10Hz to around 60 Hz.

• 3.624000
• 3.529000
• 3.624000
• 3.624000
• 3.528000
• 3.624000
• 3.529000
• 3.528000
• 3.624000
• 3.529000
• 3.624000
• 3.529000
• 3.623000
• 3.529000
• 3.529000
• 3.624000
• 3.528000
• 3.624000

There is a striking pattern in the distance between these spurs. It is either 3.62 or 3.53 rounding up to the hundredth Hz. Noticing that 3.62 Hz ~= 3.53Hz + .095 Hz suggests the difference in distance is related to the bin width of the spectrum, which subsequently suggests the distance is effectively the same for all spurs. I have no explanation why these spurs are separated by ~3.6 Hz.

Discussion

The experimental results of the Bliley NV47A1282 tests suggest a number of conclusions, including:

• Used oscillators purchased on ebay may not behave according to the specs given for new devices. The spectrum of the NV47A1282-2 is significantly different than that for the NV47A1282-1 and NV47A1282-3. Since the latter two oscillators have similar spectra, I suspect the NV47A1282-2 is faulty. This conclusion is also supported by the unexpected positve power slope in the NV47A1282-2 spectrum between 33 and 185 Hz. The spurs in the spectra of the NV47A1282-1 and NV47A1282-3 between 10 and 100 Hz may represent faulty operation due to aging.
• The spurs between 10 and 100 Hz in the spectra of the NV47A1282-1 and NV47A1282-3 are of significant power. At the lower frequencies they reach approximately -75 dBc/Hz, which is well above the power levels given in the specs.
• Focusing on the NV47A1282-1 and NV47A1282-3, The power law slope below 100 Hz appears to represent flicker phase noise, while the power law slope above 100 Hz is white noise slightly adulterated by the addition of low-level pink noise.
• The noise floor in the spectrm of the NV47A1282-1 and NV47A1282-3 is within 5-10 dB of the specs. Whether this difference is due to aging or overzealous marketing influence on the specs is unknown.
• The regularity in the frequency difference between the spurs in the 10 and 100 Hz range is a mystery. I could come up with no explanation for this spectral property.

Conclusions

Drawing general conclusions from the results of testing one brand of oscillator is dangerous. Prudence would suggest limiting the applicability of the results to it. However, some tenative observations are possible, if only as hypotheses to examine when other oscillator brands are tested.

• The difference between the spectra of the NV47A1282-2 and the NV47A1282-1/3 suggests the necessity of purchasing multiple used units to ensure the acquisition of a working unit. This raises the question: when is it economically advantageous to purchase used units as opposed to purchasing a new unit with phase noise specs in the same range? The savings due to the reduced used unit cost may be illusionary, since it is necessary to purchase more than one used unit to make up for the liklihood of unit failure.
• Used units also may not deliver the performance quoted in the new unit specs. While it is not clear what causes the difference between the NV47A1282 spec and the measured phase noise of the used units, it is clear that the spurs in the range 10-100Hz for the NV47A1282-1 and NV47A1282-3 put their phase noise performance well out of spec in that frequency span. This another reason why purchasing used units may not provide the performance necessary for a particular application.

[iMo]

Did you try to measure those DIL14 oscillators with, say, 100nF ceramic soldered between socket's pins 14 and 7 (Vcc, GND)?

[dnessett]

Did you try to measure those DIL14 oscillators with, say, 100nF ceramic soldered between socket's pins 14 and 7 (Vcc, GND)?

No. What do you expect that will achieve?

Dan

[dnessett]

I am stuck in a conundrum. When analyzing the HP11729C correction procedures for its phase detector configuration, I concluded that backing out PicoScope FFT processing gain is not only not required, but in fact leads to erroneous results.

This is evident when considering how processing gain arises in the first place. When measuring the power spectrum of a signal corrupted by noise (more accurately, white noise), the total noise power over the complete spectrum is constant. Dividing the power spectrum into bins distributes this total noise power across each bin, which effectively yields bin noise power by dividing the total noise power by the number of bins.

Comparing the power of an FFT spectrum using N bins with that using M bins, where M > N, means that each bin noise power value for the "M bin" spectrum will be less than the bin noise power value for the "N bin" spectrum. Since non-stochastic signals fall within a single bin in both the "M bin" spectrum and the "N bin" spectrum, the signal-to-noise ratio is larger for the "M bin" spectrum than for the "N bin" spectrum.

However, when measuring noise processes (such as phase noise), the spectrum values are normalized to a 1 HZ noise bandwidth (e.g., dBm/Hz). Consequently, the normalized values for a "N bin" spectrum are equal to the normalized values for a "M bin" spectrum. No adjustment for processing gain is required.

I figured this out when creating the correction algorithm for the HP11729C phase detector configuration. The results presented using this configuration are correct.

However, I didn't understand this when I created the correction algorithm for the Frequency Discriminator configuration. Reading the results presented for the experiments using this configuration shows that I knew something was wrong, but did not know what. I hypothesized it had something to do with the noise floor of the Frequency Discriminator versus the noise floor of the PicoScope. This conjecture is incorrect. The anomolous results are due to backing out procesing gain; a correction that gives erroneous results.

This brings me to the conundrum. When I figured out that I should not correct for processing gain, I decided to go back and correct the phase noise data in those posts that document experiments using the Frequency Discriminator configuration. I still have the raw uncorrected data and so recomputing the corrected data is straight forward.

But, I then reread the thread concerned with the Frequency Discriminator results. Unfortunately, there are replies by others that reference these results. If I change my messages reporting these results, the context of these replies becomes corrupted and so it may look like the respondents were confused.

So, I decided I cannot destructively change my messages. Instead, I will use the strike through feature of the markup language to indicate where the results I reported are erroneous. I will also leave a forward pointer to this message attached to the strike through text to explain why that text is incorrect. On the other hand, I can update (and have updated) the Frequency Discriminator correction algorithm documented in the thread (see this post). So, going forward, anyone who decides to use the Frequency Discriminator configuration that is documented here, will have access to the correct procedure.

That leaves the problem of correcting the Frequency Discriminator configuration results. This configuration was used in 3 cases: 1) for the MV89A oscillators, 2) for the generic XO can oscillators and 3) for the Wenzel HF ONYZ IV oscillator. It is pretty clear that the generic XO can oscillators are unsuitable for use in any application that requires a stable carrier output frequency. So, I don't think anything is served by providing the corrected phase noise results for this class of oscillator. Unfortunately, I did not retain the raw data for the Wenzel oscillator. Furthermore, I have torn down the Frequency Discriminator configuration, so I would have to reconstitute it to get the raw data for the Wenzel. I have decided that the effort to do so is not justified by the value of the results.

That leaves the MV89A oscillator data. I have decided to regenerate the Frequency Disciminator configuration results when I test these oscillators using the Phase Detector configuration and compare the two. Hopefully, this will show the relative merit of each configuration.

[dnessett]

I have not posted to the EEVBlog forums for over 3 months. This was the result not of a lack of interest, but of some developments in my personal life, which I will briefly summarize. I have experienced 2 deaths in my family, one sudden and the other beginning in February and lasting several months. These and their aftermath have taken prority over my hobbyist activities.

In the near future I will be moving out of state, which will mean disassembling my electronics lab and moving it along with the rest of my household goods. It is likely that I will not have the opportunity to reassemble this lab and use it to continue my investigations on oscillator phase noise for 6-12 months. So, I thought it prudent to provide the parital results I have developed, even though they are incomplete and may raise more questions than provide answers.

I spent the first few months of 2020 trying to eliminate spurs due 60 Hz and its harmonics. The discussion of this took place on the HP-Agilent-Keysight-equipment@groups.io mail list. For those interested, I posted a message concerning this activity on the EEVBlog forum here.

After concluding that the elimination of these spurs was impractical, I decided to modify the procedures I used to measure phase noise. These are documented in the above referenced post. In summary I started using a faraday cage to isolate the oscillator under test, used batteries to power both the test oscillator and reference oscillator, and post processed the data from the HP 11729C using a moving average to eliminate the 60Hz harmonic spurs.

I applied these new procedures to a brand new Connor_Winfield OH-200 10 MHz very low phase oscillator. My hope was the result would demonstrate the new procedures generated a phase noise plot that follows the published manufacture phase noise specifications closely. Regretably, this did not happen.

Figure 1 shows the results of using the new procedure on the Connor_Winfield oscillator with an external LNA amplifying the 11729C output by 60 dB. The 60 dB amplification effect is removed during post processing. Its use is to raise the output of the HP11729C sufficiently to clear the noise floor of the PicoScope 4262.


Figure 1 - Phase noise of the Connor_Winfield OH-200 10 MHz oscillator generated by the new procedures.

In this plot, the blue trace represents the unsmoothed phase noise plot, while the superimposed black trace is the moving average of the unsmoothed data. When I first looked at the plot, it seemed reasonable enough. But then I compared it against the published spec.

Figure 2 shows the spec phase noise data published by the manufacturer (solid red line) superimposed on the phase noise plot I got from the 11729C. As you can see, the published data has a very different shape and values than what I got.


Figure 2 - Phase noise of the Connor_Winfield OH-200 10 MHz oscillator with superimposed plot (red) of the published spec.

I think I understand why the data I obtained seems to be floored at ~ -128 dBc/Hz. The residual phase noise of the 11729C is 10Hz: -115; 100 Hz:-126; 1 KHz: -135; 10 KHz -145. So, the bottoming at around -128 dBc/Hz is probably due to the residual noise floor of the 11729C (given that my 11729C is used and probably at least 13 years old, I would not be surprised if its noise floor has degraded somewhat).

However, below ~100 Hz, the spec phase noise rises significantly and at around 30 Hz exceeds what I measured. I can find no explanation for this. My measurements seem to suggest that the Connor-Winfield is better than its specs.

Unfortunately, at this point my family situation intervened and I was unable to formulate a hypothesis for the unexpected low offset frequency phase noise and test it. As indicated above, I will be unable to explore possible explanations for a considerable amount of time. So, I thought providing my incomplete results might stimulate others to speculate on the causes of the low offset frequency anomalies.

[sorin]

It has been 4 years since your last post. I hope you are well!