This week’s progress:
This week I continued working on the SVM and tried to research more in depth into the alternative methods. I also wanted to figure out the logistics of training as this will be one of the main things that we would be moving forward with.
Adithi has obtained several large HVAC ducts from the trade school. These are galvanized steel square ducts approximately 1 ft × 1 ft in cross-section and 5–6 ft in length. These ducts will serve as the primary physical samples for generating the acoustic dataset used to train the model. For training, the goal is to generate a dataset with clear, controlled categories of structural degradation so that the classifier can learn the acoustic signatures associated with different defect types. Rather than introducing damage randomly, the plan is to structure the dataset around several levels of defect severity.
Last week I did quite a bit of research looking into how we could structure the corrosion aspect of testing. However, during our meeting with the professor, we got the suggestion that it would be better to obtain HVAC ducts that were already corroded and focus more on the cracks simulation on the clean ducts that we obtained. So this week I tried looking more into the specifics on training for the HVAC ducts that we have:
For each duct sample, we will first collect baseline recordings with no defects present. These recordings will capture the acoustic and vibration characteristics of a healthy duct under consistent excitation conditions. The baseline data will act as the reference class for the model.
Next, we will introduce controlled mechanical defects in stages. Based on the discussion from the last meeting, the idea is to introduce multiple cracks of varying lengths and depths so that we can capture how the acoustic response changes with increasing structural damage. For example:
- Minor damage condition: introduce ~3 small longitudinal cracks (≈0.5–1 mm width and ~1–2 cm length).
- Moderate damage condition: introduce ~6 cracks (1–3 mm width and ~3–5 cm length).
- Severe damage condition: introduce ~10 cracks or slots, including several deeper partial-depth cuts (up to ~25–50% wall thickness).
These cracks will be distributed across the duct surface rather than concentrated in a single location so that the model learns generalized damage signatures instead of position-specific effects. At each stage of damage, we will collect a new set of acoustic and vibration recordings. This staged approach allows the dataset to capture progressive degradation rather than only binary healthy vs. damaged states. Ideally, each condition will include multiple recordings (e.g., 20–50 samples) to account for variability and improve model robustness. This will also let us get away with gathering more data from a limited number of HVAC ducts.
Overall, the training dataset will therefore consist of three types of samples:
- Healthy ducts (baseline recordings)
- Ducts with controlled mechanical defects
- Naturally corroded ducts
Collecting data across these categories should provide enough variation for training the initial SVM model and later experimenting with more advanced models if needed.
Here is my progress for the week (Continued research on training, and my improved SVM model): https://docs.google.com/document/d/1FbG87Cg-6bPmjf2CVkg7BYE6hkrSWeNEl6PgoUjYYN8/edit?usp=sharing
Things to do for next week:
Next week is spring break, so I will be traveling. However, I will still try to assist Adithi remotely with parts of the embedded system integration while she is in the lab. The main goals for next week will be:
- Start collecting baseline acoustic and vibration data from intact ducts.
- Work with Adithi on the embedded systems control logic for synchronized sensor data collection.
- Begin organizing the dataset structure so that recordings from each defect stage are labeled and stored consistently for training.
The main thing will be to get adequate training data, contact people for already corroded HVAC ducts, and work on embedded logic.