Data-driven HVAC Diagnostic Tool
 

Update (2020): This site will no longer be maintained.

Contacts: 978494543.qq54@gmail.com, Github Profile

Why use it?

Accurate fault diagnosing reduces cost and increases profit. Troubleshooting HVAC systems requires not only a thorough understanding of the HVAC system, but also relies heavily on the experiences of the technician. This is an easy-to-use diagnostic helper that gives the probabilities of each fault based on your inputs.

At our company, the tool is especially useful to new recruits. Rather than consulting senior techs for difficult cases, our apprentices can access this tool on their phone and give accurate diagnoses.

This is a system in development and requires more data collection for more accurate results.

How does it work?

A Neural Network learns from real life diagnoses that our senior technicians collected. Each data entry contains field measurements as well as the appropriate fix applied to that system. If an applied fix puts the system back to normal, we can reasonably deduce that the fault of the system based on the fix.

For example:

Fix Fault
Added refrigerant System undercharged
Replaced air filters Low indoor flow
Cleaned coil Dirty coil
Replaced filter drier Restriction

There may be cases of multiple fixes applied. These cases are especially valuable since they allow my neural network to learn to find multiple faults in one diagnosis.

The Architecture

All neural networks in this tool are feed-forward network.

The main network yields probabilities of each fault from ambient temperature, saturation temperatures, line temperatures and amperage ratio.

In case amperage cannot be measured, the amperage network predicts the amperage ratio from all other inputs. The network is trained with data containing actual measured amperage ratio and relevant field measurements from which the amperage ratio may be predicted.

In case of refrigerant undercharge, the network predicts the amount of refrigerant needed. The network is trained with data containing temperature and pressure measurements as well as the actual weight of refrigerant added.