HVAC engineer trying to figure out how to apply ML algorithm to detect equipment failure

AI Questions
, ,
1

Hi, I am an HVAC engineer by profession. During my course of work I found that there is a strong need for understanding the design and specification of chillers behind the performance enhancement using data analytics. I wish to create a model using Supervised Machine learning and Deep learning for generation of fault detection before equipment failure and energy conservation.Those interested can share your thoughts.

Currently pursuing MS in Data Science & Machine Learning from Woolf Malta. Any suggestions/guidance would be appreciated to reach my goal or further study.

Regards
Tameem islam

2 Likes
2

Do you have access to a set of data that you could use to train a predictive model?

3 Likes
3

I know nothing about HVAC but my first question/thought would be: define your variables. What data do you have and can you measure ? And if there is a factor you feel is important, yet not presently easy to measure, define/come up with a means that data can be collected. Defining your variables will also help with model selection.

4 Likes
4

To the best of my understanding, what you ask about is called predictive maintenance and has been an area of study for quite a long time - at least a decade. Give the interweb a search and you will find many IT (eg IBM) and industrial automation (eg Siemans AG) companies very active in this field.

3 Likes
5

@Tameem_Islam I also wish I knew better how this applied to ML, yet I do not. Though the PID (Proportional–integral–derivative) algorithm could be useful for your use case and is maybe something to look up.

2 Likes
6

The PID is commonly used to form a closed-loop control law.
I do not see how it applies to failure detection.

2 Likes
7

Thank you so much for taking time, the variables am considering are energy consumption, indoor air quality(IAQ), temperature, humidity, airflow.

2 Likes
8

Thank you so much ai_curious for taking interest in environmental study. #carbonfootprint

1 Like
9

I have a set of data taken up from kaggle but needs pre processing of data to make it ready for test ready process.

1 Like
10

@TMosh I mean that is the whole idea-- The hardware fails when it breaks this loop.

1 Like
11

They keys then are:

  • having a very accurate model of the physical plant.
  • defining what ‘the loop is broken’ means in a measurable and unambiguous form.
  • detecting the ‘broken loop’ state.
1 Like
12

Yes, as mentioned I do not have first hand experience applying a PID application to ML, though I do have experience in other more standard situations. That said a quick Google search I did after posting suggests others out there are thinking about this too–

1 Like
13

However, the whole idea of predictive maintenance is to avoid hardware failure altogether. PID, even PID + ML for control, is a different problem space.

1 Like
14

I can’t post links, but search for “Amazon predicts HVAC.” It’ll bring up a blog post on how Amazon did this with Carrier. I’m not promoting Amazon but the article lays out the basics that you can use to get started.

3 Likes
15

Interesting they are using transformers.

1 Like
16

I work specifically in this area. You would need to sensor that measure the air quliaty in terms of PM, CO2, O3 etc. There are off-the-shelf products that measure these. One such example is this one from delta
Feel free to reach me out at brikesh987@gmail.com to know more details.

1 Like
17

Hello Tameem,
Try to collect as many sensor data as you can, including input voltages, or voltages inside the circuitry, temperature, humidity, airflow, airflow temperature wrt desired temperature or ambient temperature, etc…

Then you can try either one of the following methods:

  • Multivariate gaussian distribution
  • Autoencoders
  • RNNs

I have also heard that some people resorting to forests but I do not have concrete information.

Cheers.