We are SKF Machine Analyst users. We employ the overall alarm feature to help prompt us to changes in our routine vibration surveys that warrant additional attention. We commonly employ thresholds obtained via ISO 10816-3, however are experimenting with the statistical alarm function of our SKF Database. These thresholds are obtained by sampling a series of several surveys collected at a single point, calculating the statistical standard deviation of the series, and applying thresholds of some "X" multiplier of standard deviation from the baseline.

We recognize that overall alarms are not "airtight"...that is, careful evaluation of the spectral and/or time domain characteristics are required before a maintenance recommendation is warranted. Given the limitations of overall alarms, however, what does the board think of statistical alarms? If you use them, what multipliers are appropriate for alert and danger thresholds?

Thanks in advance,
George

I have enough good data with enough similar setups that if I had the time and patience, I'd probably get valid statistical setups of my alarms.

There was a good discussion recently on CSI's AutoStat feature. Click here for the thread.

Hello George, have you read the old saying that there are three kinds of lies, stating that one of them is statistics. I cannot understand why you would rely on a much smaller comparison for your limits of warning and fault than the background of all the countries that have made the standards. If you happen to have perfect machines, you new home made limits will be tight, if your machines are relatively bad, your limits will be sloppy. Aim should be to find the trade off beween the maintenance costs and the operational value of a machine. Such considerations normally lead to getting below the lower limit of the standards recommended warning levels. Using a statistical background would most probably tend to "allow" higher levels. The standards levels are not sacred. As time goes by and you learn more about your own machines, you can adopt levels to suit your plant better, but not based on statistics on machines with levels all from green to red. Regards Arne

I don't know much about statistical alarms. I think I remember hearing there is a feature (that if you have several sister machines) let's you set the one that has highest vibration among the group with the tightest multiple of standard deviation around it's own history and the one with lowest vibration with the loosest multiple of standard deviation around it's own history. A roundabout way to include both trending of single machine and comparison to sisters in a single alarm. That makes sense to me.

Arne - I don't understand what you're saying. Are you saying that you believe an absolute limit from a standard is better than a relative limit based on trending over time? (using statistics to help the trending process?)

This message has been edited. Last edited by: electricpete,

So...I write this question just before I go to bed on Friday night, and wake up with three answers. I hope nobody figures a way to tax this kind of service. I really appreciate the attention each writer gives to each comment. This is an amazing resource.

I think I need to explain the type of statistical alarms I'm trying to employ here. I'm working with SKF, but it sounds similar to the CSI Autostat feature that Patrick cited. The process starts with at least ten surveys(recommended) of a single component's history. The software crunches numbers, and allows you to select some multiple of standard deviation to apply to two thresholds...like 2XSD for "alert" and 3XSD for "danger". These thresholds are individually calculated for each specific survey point. They are not shared between different survey points or between different machines. Each threshold would have to be recalculated after major maintenance.

Arne...I share your respect for the standards organizations. I would not accept a statistical threshold that was less stringent than our ISO recommendations.

As an example of why I'm interested in statistical thresholds...we have a vertical AC lube oil pump attached to one of our main feed pumps. It has run consistently at 0.12 ips-p for three years. The trend, recently, has shown an abrupt turn up...and is currently running at about 0.14 ips-p. The vibe frequency responsible for this change is 1X of shaft speed...and we suspect some clearances opening or compliance developing in the support structure. ISO 10816-3 allows 0.25 ips-p before exceeding the C/D limit, or a 0.07 ips-p increase above baseline before the alert. Although the current amplitude is not very high, nor is the change very large...the abrupt turn in the trend line indicates some change has occurred that may warrant attention. This is one of our critical components...and I'm not comfortable waiting to exceed the ISO limits before recommending maintenance. Granted, I could just "eyeball" a threshold that I'm comfortable with. For some reason, however, people trust me more if I tell them I've confered with one of Arne's lying statisticians.

We've tried 3XSD for the "warning" threshold, and 6XSD for the "danger" threshold. Comments?

Thank you, again, in advance,
George

Statistical calculations for vibration purposes are done in a same way as any statistical calculations are done. A random sample is processed which results in values of mean and standard deviation (SD). General statistics interprets these results as follows. With desired confidence level the mean value and SD in the population is same as in the sample. Another conclusion is that 99.7% data values of the population lies within +/- 3SD of mean.

If subject of analysis is, say, men's hair length, then everything is clear: 99.7% of men in men's population have hair length not exceeding, say, 10 inch. Is it a problem if few men in the population have longer hair? Not necessarily.

It becomes questionable if we apply same approach for setting up vibration alarms. First, what is an alarm? Even very-very low 1X vibration will eventually lead to a bearing failure. Therefore, when setting up an alarm we have to ask ourselves what degradation rate is not acceptable. That could be determined as Arne said, based on some kind of economic trade off.

Secondly, most of the time the sample for vibration statistical calculations being used in the software is not random: often limited to a group of machines operating in the same plant or, which is even worse, based on 10 readings on the same machine, etc.

Thirdly, even with a valid sample, all we can conclude is that 99.7% of machines in a population are operating with an overall vibration not exceeding, say, 0.45 IPS.

Does it mean that everything over 0.45 IPS is in condition requiring attention? Not necessarily. Those machines may be still in condition far from some kind of adopted danger level. Overall vibration leading to a catastrophic faulure in, say, 6 months may be 0.6 IPS and catastrophic faulure in 1 month - 0.9 IPS. No one knows these numbers. Therefore, in my opinion, this simplistic statistical approach is useless.

On the other hand a trend up, if available, turns on a red flag as operating conditions change and, therefore, is a reliable indicator of a developing fault.

David

This message has been edited. Last edited by: David_G,

DLI's expert system, and the practice thoughtout the US Navy is to use the average plus 2 standard deviation as an alarm level, for data taken on idnetical machines. To create the average, throw out machines with obvious problems and average the rest. Based on DLI's experience, an average based on 16 - 24 samples is as good as it gets. Less than that doesn't prove as reliable, more than that doesn't add much value.

A challenge when adding data to the average is selecting the measurements to include. Machines with faults don't get included, but the tendency is to exclude anything that exceeds the average value. This means that the average + 2 sigma value gets tighter and tighter if care is not used in selecting the data to include.

Based on my experience in at the naval shipyard and with DLI Engineering, I believe this is a realistic way to set alarm levels. A machine going over this level warrants a closer look, not an immediate shutdown or neccessarilly any repair action.

Jon
Spintelligent Labs