Reality check - AI/ML applied Machine Health vibration analysis

The I(o)T “experts” make a common mistake by stating that vibration analysis trough AI/ML tools equals delivery of asset reliability and that often leads to disappointments and often fails the business case. In this post a bit more engineering details en I did try not to get carried away by too many of it.

Vibration 101 - quick guide

Vibration analysis is already done, successfully, for decades and, unfortunately, it has nothing new under the hood other than the IoT hype is promising that by cheap sensors, cloud and AI it delivers predictive maintenance. Sorry to bring the bad news in case you did not find it out yourself already as with the knowledge that overall vibration data from most of the IoT sensors are late stage failure indicators. On top of that it takes massive streams of data and ages to train the ML model unless a large fleet of failing assets are in reach during your project. Just to remind to the fact that all rotating equipment’s have relative motion and have always some level of vibration. Even your ML/AI toolset is good enough and your data is clean to identify abnormal (anomalies) from that basic vibration data someone (humans) will need to classify the abnormal data sets of an actual failure and verification at site by a subject matter expert. Yes, vibration analysis is complex.

Don’t get to disappointed, a more advanced approach (more sophisticated sensors) and applying ML analytics on (RAW) frequency and amplitude data, the analysis in the composite waveform with compute power at the EGDE, is able to deliver the training input for the AI model and to identify the early indication between Normal and Abnormal behavior of the machine. The “early” and true “value” of the indicator is however still highly depending on the failure type and route causes. Good enough to get started till the point that you need to scale your wonderful application en the data storage will reach Terra-bites of (useless) data stored in the cloud awaiting a failure to happen.

Field deployment

Unfortunately I have more bad news, in case you did overcome the first barrier about data requirements, where I’m now assuming that you know how to train data for the AI/ML model and know that it needs the inputs from mechanical engineers that have the domain expertise of the assets and applications you target the next challenge of field deployment will arise.

The success rate of the business case is driven by easiness of deployment and the ability to scale. ..

The deployment of the actual vibration sensors (for each individual pump a minimum of 4 or more) is complicated due to:

1. Ease of access (not even to mention submersible pumps)
2. Minimum impact from external condition
3. Maximum sensitivity to abnormal conditions
4.  Minimum signal attenuation or signal loss caused by abnormality
5. Reliability of measurement

And part of the installation of the vibration sensors, if done incorrectly, it will cause dispersion of data or inaccurate measurement. Being aware of the points to Place or attach firmly, to make all mount surface adhere closely and to align vertically or horizontally to the axis of the object. And if that is solved be aware that installation of vibration sensors may differ according to measurement frequency range. Wrongly attached pickup sensors or accelerators may cause unstable measurement and incorrect data. Generally, measurement is 1/3 off resonance frequency (“S**T in = S**T out).

Taking all of the above into consideration (and there is more) it makes perfect sense that there is a massive industry and business opportunity around vibration analysis and extremely good skilled people are delivering tangible results to monitor machine health at the shop floors and critical industries. Spending up to 30K or more for a single asset vibration analysis is very reasonable in cases where unplanned production stops generate business losses of 100K to Millions or environmental spills and human health is at risk with waste water overflows and areas are flooded with unwanted watersheds.  

There is also good news, with solutions and business outcomes that are more easier to deploy (included submersible pumps) and more attractive ROI models. Delivering real time insights on actual and future machine health as input for predictive maintenance including insights to improve energy consumption and reuse the insights into other platforms and systems. But I stop here while this post was not intended to be commercial pitch. Just reach out to me in case you would like to learn more.