Accomplishments
This week, our team pivoted from sourcing decommissioned HVAC duct sections from local institutions (including Carnegie Mellon University, University of Pittsburgh, Trade Institute of Pittsburgh, and Rosedale Technical College) to independently acquiring and modifying HVAC units for controlled corrosion testing. After receiving initial positive responses from facility managers and administrators, we determined that building and corroding our own duct infrastructure would give us greater experimental control and enable systematic data collection. This shift allows us to simulate realistic degradation patterns while ensuring reproducibility for robotic inspection validation.
In parallel, I conducted a deep literature review on acoustic based leak and corrosion detection systems to inform our sensing and modeling strategy. I analyzed a recent CNN based leak detection study that used adapted ResNet18, VGG16, and AlexNet architectures trained on spectrogram representations of leak audio. The paper introduced a dual dataset “acoustic mixup” approach to synthetically combine clean lab leak sounds with industrial background noise, along with targeted frequency windowing (11–20 kHz) to suppress irrelevant noise. Results demonstrated over 95% accuracy on vent leaks and strong performance on tube leaks, significantly outperforming shallow CNN baselines, highlighting the importance of network depth for extracting subtle acoustic features in noisy environments.
I also examined a corrosion severity classification framework using Acoustic Emission signals, which applied Wavelet Packet Transform and Fast Fourier Transform feature extraction followed by a Linear Support Vector Classifier. The model achieved 99% accuracy in classifying corrosion severity stages, outperforming multiple classical ML baselines. Key insights for our project include the effectiveness of multiresolution time frequency decomposition, statistical feature engineering (RMS, kurtosis, mean frequency), and the feasibility of real time, non destructive structural health monitoring. These findings are directly shaping our experimental design and model selection as we integrate acoustic sensing into our robotic HVAC inspection system.
Progress Report:
https://docs.google.com/document/d/15zzXubsDVOVUDyqlZ87QFHDfqAO0mv9-JDg05TYcSkg/edit?usp=sharing
Schedule
I am currently on schedule, based on this week’s literature review, I now have a structured understanding of potential model architectures. I will continue reviewing related work to evaluate alternative modeling strategies and signal-processing pipelines so that, once we generate our own training dataset, we can make better informed decisions about architecture selection, feature extraction methods, and evaluation metrics.
In the coming week, I plan to focus on our experimental methodology for inducing controlled corrosion in HVAC duct sections and establishing a structured data collection process. This includes defining corrosion stages, selecting sensing modalities, and outlining repeatable testing protocols. In parallel, I will assist Adithi with robot setup and the hardware–software integration pipeline to ensure that our sensing, data acquisition, and robotic inspection systems operate cohesively.