FIRE Schools: Future-ready Innovation and Research Excellence

Developed a polygon-based, event-specific framework for wildfire prediction using VIIRS active fire detection and ERA5-Land meteorology data. Addressed data skewness, temporal alignment, and event delineation challenges for real-time operational use.

Built a deep learning system using ResNet-inspired CNN architecture to detect sleep apnea from snoring audio, providing a low-cost alternative to traditional polysomnography testing methods.

Built a machine learning-based fault detection system that analyzes truck engine sounds to identify mechanical issues. The project uses MFCC feature extraction, data augmentation (Gaussian and salt & pepper noise), and a hierarchical classification pipeline (Binary + LSTM) to detect and diagnose faults such as air leaks and oil cap removal.

Developed a novel energy consumption forecasting model using sinusoidal temporal encoding to better capture cyclic patterns in time-series data. Combined advanced feature engineering (rolling statistics, lag features) with ensemble models such as XGBoost and LightGBM to improve prediction accuracy in smart grid systems.

Created a machine learning system to classify baby cries into categories such as hunger, discomfort, belly pain, and tiredness. Combined two datasets to enhance generalizability and applied mel-spectrogram feature extraction with noise reduction techniques.

Designed a vehicle price prediction system using ensemble regression models enhanced with SHAP explainability. Combined Kaggle datasets (Russian car sales) and applied preprocessing steps including inflation adjustment, outlier removal, and label encoding to improve prediction accuracy and transparency.