Hi,
I'm Arya!

❮ ❯

Digital Signal Conditioning: 2nd-Order Butterworth Filter Design and Performance Analysis

Implemented a digital LPF using a 2nd-Order IIR Butterworth. To validate the filter's performance,a sample ambient audio with chirp signal of (10 kHz - 24 kHz) was adopted to simulate wideband noise interference.

Butterworth architecture was selected to maximise the natural characteristic of flat passband which ensures signal of interest reimains undistorted. Opting for a digital implementation was driven by design flexibility and scalability, through rigorous parameter tuning and iteration.

To validate the filter's performance, Signal-to-Noise Ratio (SNR) analysis was opted. It achieved a 13.35 dB improvement, transitioning from -7.976 dB to a clear +5.372 dB, demonstrating the filter's ability in suppressing noisy band interference.

❮ ❯

Predicting CO2 Emission from an Offshore Gas Turbine Using Deep Learning: A Dense Neural Network Approach

Following the 3-model comparison (Linear Regression, Decision Tree, XGBoost), a Dense Neural Network was introduced to investigate wether deep learning could further improve predictive accuracy on the CO2 emission task. The 3 layer architecture (64 → 32 → 16 units, ELU activation), achieved R-Squared = 0.981 and RMSE = 0.494. L1 regularisation was applied accross all hidden layers to suppress redundant feature weights - directly addressing the severe multicolinearity identified in the sensor data - and early stopping to halt training and peak validation performance.

Future work targets edge deployment on low-cost microcontrollers such as the ESP32-P4, eliminating cloud dependency for real-time inference.

❮ ❯

A CO2 Emission Prediction Machine-Learning-driven Model: XGBoost, Decision Tree, and Linear Regression Accuracy Comparison

Explored 3 different Machine Learning models - XGBoost, Decision Tree, and Linear regression - for CO2 emission prediction from an offshore gas turbinewere (50,371 samples, 9 features). XGBoost and Decision Tree outperformed Linear Regression by capturing non-linear relationships, achieving RMSE of 0.027(R-Squared = 0.999) and 0.122 respectively, compared to Linear Regression's higher error.

Linear Regression fails in this case due to multicolinearity (r ≥ 0.95 across most feature pairs), producing high-variance predictions that generalise poorly.

❮ ❯

Non-Invasive Blood Pressure Device: Pneumatic Control & Oscillometric Signal Processing

Led the hardware assembly, schematics, and was responsible for the hardware stack- from the pneumatic circuit design to gates using transistor for automated control based from pressure transducers reading. The device successfully approximated diastolic and systolic blood pressure within a tolerance of ± 13 mmHg. Mean Arterial Pressure (MAP) was extracted via peak envelope detection; derived systolic and diastolic via emperical amplitude ratios.

Pipeline:

Cuff presurisation to ± 180 mmHg → Holding phase to minimise noise and stabilise readings → Deflation phase to deflate cuff pressure to ± 30 mmHg → Calculation phase of MAP, Systolic Blood Pressure (SBP), and Diastolic Blood Pressure (DBP).