Difference Between Noise and Oscillation

One thing to examine is how fast the process can actually move. If the signal goes back-and-forth faster than that, there is noise.

Slug everything with a time constant of one second, Mildred. If it happens in less than one second on a process plant, it isn't real.

To name one (among many), Partlow controller literature has good info on this, as would a text covering PID and fuzzy control.

hi.. thanks for your response.. can u give me a rough idea how it is done..

I think you meant to ask Crabtree.

The basic idea is simply to change ("bump") the setpoint and see how the process responds. You observe the speed of response and the amount and timing of any overshoots/undershoots. From there you can calculate suitable parameters for the controller.

Oscillation usually has discrete frequency components whereas noise is generally broadband. If you have feedback loops, for any frequencies where the phase shift around the loop is an integral multiple of 360 degrees and the gain is 1 or greater, you will have oscillation. If the gain is less than 1, you may see some transient oscillation after disturbances (e.g. change in setpoint). As Tornado has suggested (#5), a step response will help you understand what is going on.

ya thank you. I have my data in form of PV,SP,OP and I couldnt be able to work in the plant (cant perform step test in the plant).

The problem is the noisy data are getting flagged as oscillatory using my code. So i want to differentiate noise form oscillation using matlab code. As u r telling i have to check for the phase shift and gain at every frequencies and then i have to decide whether it is oscilation or noise is it?? I dont have much knowledge in frequency domain. Please help me whether my understanding is correct?? thanks Rixter

Perform a Fourier Transform on your data set.

is ur point is that if i take fourier transform of the data, if it has a peak then it is a oscillation???

That is a very simplified way of looking at the Fourier Transform results. Unfortunately for you if only one spike appeared then you would easily recognize in the time domain that you had a sine wave signal. More than likely you will see many spikes appearing in your bins. Now is when the fun starts.

Identify the frequency of your spikes and look for harmonic correlations between these signals. In other words a 1 kHz and a 1.5 kHz spike may not be considered a harmonic correlation unless there is also a 500 Hz spike, too. In the first case you may have to have two independent sources to produce signal you see in the time domain. In the second case a single 500 Hz source maybe acting non-linear and produce all three spikes.

Hi redfred..I couldnt be able to get your point.. can u give me in the form of algorithm which will make me understand..

Alas, there is no such thing as an algorithm for understanding.

Going from the time to frequency domain will show ALL resonate frequencies as peaks. As mentioned, garden variety noise is generally broadband, so you'll end up with an elevated noise floor that likely has some small peaks. If there is something driving the noise then you'll end up large peaks, in addition to the process oscillation frequency.

You have data, and as long as it's sampled at 2x of the process oscillation, you'll at least get that peak. However, that's the minimum, and you really want a good deal more samples.

This has been leading up to what was already suggested: how fast can the process variable change? What's the variable? What is the sensor type? Are we talking signal noise, or could something in the process be injecting noise at a similar amplitude but different frequency than the variable oscillations? That's where an FFT can be very helpful.

Low or high pass filters can also be used, what one and at what frequency, depends on the system.

Now, because of the language, I gotta ask, are we helping with homework? If not, more details would get you a great deal more help. Generalize some stuff if it's proprietary, but not to plant, PV, etc, that is way too general for any more advice.

Steve

Hi thanks a lot for the comments.

I am wrking in a company and i have the data for whch i have to find the Oscillation index. My senior hve identified the causes of oscillation index not getting calculated correctly as follows.

o Noisy Data getting flagged as oscillatory

o Incorrect analysis for loops with intermittent oscillations

o Incorrect analysis for loops with varying oscillation frequency

so i was trying to find out the solution for this problem. I am working with the pressure,flow,temperature loops. My sampling frequency is one sec.

Danger! Danger! Will Robinson! Danger! [Oops I'm dating myself here.]

This sounds to me like you have one of two dangers to your employment here. Your senior staff may have been successfully dabbling in multivariant control theory without understanding what they were doing. The other obvious possibility is that you've been tasked with correcting a system without any knowledge of control theory yourself.

If your senior staff did know control theory then they could identify the minimum signal to noise ratio for each process variable for every control loop. Suitable measurements of existing signal to noise values will confirm either a stable or meta-stable configuration. A suitable electrical or mechanical filter may now be inserted into your system to make it more stable. However, regardless of the filter being a low, high, or band pass filter more poles have now been added to your system. A new analysis should be performed.

Your comment about intermittent oscillations implies to me that several of these existing feedback and feed forward loops have poles close to the right half plane somewhere. This is the most common meta-stable condition because some component drifts into and out of a stable condition. Ideally once a system gets properly characterized one can perform a Routh-Hurwitz stability analysis to find how to keep all of the poles in the left hand plane. That is if it can be done with your control system topology. Very often a system changed by small incremental changes in design creeps into a meta-stable condition in the control topology. This cannot always be corrected but a methodical analysis will show if the desires can be realized.

Wow, this is turning into heavy material. I have been commissioning process plant for a long time and never had to get into the mathematics you are all talking about. With all these oscillating loops , sounds like your plant is not in control.

There are loops that oscillate faster than 1 second. A good example is the furnace pressure in a coal fired boiler. If I tried to control the raw process data I would not have a working plant. Put a filter on your input. If your transmitter is fairly modern you can do that in the transmitter set up. If not, then do it in the control system.

By filtering the input you will be able to increase the Gain in the controller and hopefully bring your plant under control.

One more point, do not try to achieve this by putting restrictors on the impulse line. This gets things out of phase as well as creating blockages.

KISS.

Oscillation is energy emission/vibration in frequency with amplitude.

Noise is something-unwanted energy from another source or oscillation in the same frequency and with the same or at greater amplitude.

Noise is like in comparison to courting a girl. When you and your rival exist in the same place to wait for your ideal girl for a date.

2 source of oscillation in the same frequency, will consider each a Noise, respectively. It depends which of the two you are using.

Take the simplistic approach, scan your data for cyclical, repeating patterns over a few different sampling window sizes, those could be oscillations if they appear throughout your data set. If there are no easily discernible patterns and/or they appear randomly then you have noise.

There are also statistical tests that you can run your data through that will characterize the deviations as random or periodic, but be careful with your sample window size since over the long run the average will be zero for either type of signal, unless there's a bias which drives your process into its limits.