Transfer Learning for Tonal Noise Prediction in VRF Units Using Thermodynamic and Vibration Signals

The Big Picture: Predicting the "Hum" of an Air Conditioner

Imagine you have a high-tech air conditioner (specifically a VRF unit) that powers a whole building. Inside, there is a twin-rotary compressor that acts like the heart of the system. When it beats, it makes a specific low-frequency "hum" (called the 2f noise).

The problem is that this hum changes its volume and tone depending on how hard the air conditioner is working. If it's freezing outside, or if the valves inside open and close differently, the "hum" changes. Trying to predict exactly how loud it will be using old-school physics formulas is like trying to guess the weather in a new city just by looking at the weather in your hometown—it often fails because the conditions are too different.

The Solution: A Smart "Translator" (Transfer Learning)

The researchers wanted to build a computer model that could predict this noise volume accurately, even when the air conditioner is running in a brand-new situation it has never seen before.

They used a technique called Transfer Learning. Think of it like this:

• The Old Way (Standard PLS): Imagine you are a chef who learned to cook perfect soup in a kitchen with a gas stove. If you suddenly have to cook in a kitchen with an electric stove, your old recipe might fail because the heat source is different.
• The New Way (Di-PLS): This is like a "Master Chef Translator." It learns the essence of what makes the soup taste good (the core ingredients and techniques) rather than just memorizing the specific stove settings. It can then take those core lessons and adapt them perfectly to the new electric stove, even if the heat distribution is totally different.

In the paper, this "Translator" is an algorithm called Domain-invariant Partial Least Squares (Di-PLS). It looks at data from known situations (the "Source") and learns how to apply that knowledge to unknown situations (the "Target") without getting confused by the differences.

The Two Ingredients: Temperature vs. Vibration

To teach their model, the researchers used two different types of "clues" (inputs):

1. Thermodynamic Signals (The "Temperature Map"): These are measurements of heat, pressure, and refrigerant flow. It's like looking at a map of the weather.
2. Vibration Signals (The "Shaking"): These are measurements of how much the metal parts of the machine are shaking. It's like feeling the floor vibrate when a heavy truck drives by.

The Experiment: The "Left-One-Out" Test

The researchers tested their model using a strategy called "Leave-One-Condition-Out."
Imagine you have 19 different scenarios (like 19 different days with different temperatures and valve settings). You train the model on 18 of them, and then you ask it to predict the noise for the 19th one it has never seen. Then you swap them around and do it again for every single scenario. This proves if the model is truly smart or just memorizing the answers.

The Results: Why "Shaking" Won

Here is what they found:

• The "Temperature Map" (Thermodynamics) Struggled: When the air conditioner was in a weird state (like when a valve was fully closed or automatically adjusting), the temperature clues became confusing. The model got lost. It was like trying to navigate a city using a map that only works on sunny days; when it rains, the map is useless. The predictions were often off by a lot.
• The "Shaking" (Vibration) Succeeded: The vibration signals were the winners. No matter how the valves moved or how the temperature changed, the physical shaking of the machine remained a reliable clue.
  • The Analogy: If you want to know how loud a drum is, you don't need to know the humidity in the room (thermodynamics); you just need to feel how hard the drumstick hits the skin (vibration). The vibration is the direct cause of the sound.
  • The Score: The vibration-based model using the "Translator" (Di-PLS) was incredibly accurate. It predicted the noise level within 3 decibels (a very small margin of error) for almost every single test case. In fact, it was so good that 99.6% of its predictions were within that tiny 3 dB range.

The "Valve Closed" Test Case

To really stress-test the model, they looked at a specific scenario where the "vapor injection" valve was completely shut off. This changed the entire internal mechanism of the compressor.

• The Old Model: Failed completely. It was like a chef trying to bake a cake with no oven. The error was huge.
• The New Model (Di-PLS): Still worked perfectly. Because it learned the fundamental relationship between how the metal shakes and the sound it makes, it didn't care that the internal "recipe" had changed. It could still "translate" the shaking into the correct noise prediction.

The Conclusion

The paper concludes that if you want to predict the noise of these complex air conditioners, don't just look at the temperature and pressure. Instead, listen to the vibrations.

By using their smart "Translator" algorithm (Di-PLS) on vibration data, they created a system that can accurately predict noise levels even when the air conditioner is running in conditions it has never experienced before. This is a big step forward for making quieter, better air conditioning systems without needing expensive, custom testing for every single weather condition.