Operational wind records are broken by outages and quality control, but forecasting studies routinely build features and split train/test sets as though the series were continuous. Holding data and model fixed and varying only the gap-handling policy, I show this inflates reported skill by up to 10.13 percentage points, enough, on one record, to turn a null result into a significant win on identical data. Decomposed the bias into three channels and showed feature-side leakage, the focus of existing work, is negligible here while the sampling side dominates. A controlled injection experiment across five independent records establishes that gap structure, not missing-data volume, governs the effect: many short interruptions produce +5.99 points against +0.03 for few long outages. Converted through a turbine power curve, the overstatement is about 1% of rated capacity per turbine in error reduction an operator would expect and not receive. 

This project implemented the autocorrelation technique for a Doppler Wind Lidar (DWL) system, which is used to process the returning laser signal and accurately measure wind velocity in the atmosphere. 

An algorithm in MATLAB for automatic melanoma segmentation in dermoscopic images: morphological hair removal, Otsu thresholding for initial masking, and dilation for boundary refinement. Validated against expert ground truth at Dice 0.938 and Jaccard 0.884.

Comparison of envelope-spectrum and time-domain feature families for rolling- element bearing fault classification on the CWRU dataset, under a leave-one-load- out protocol that tests generalisation across operating conditions. Noise and window-length sweeps showed time-domain features outperforming envelope features down to 0 dB SNR, contradicting my initial hypothesis, and identified a hard physical limit: below 256 samples, FFT resolution exceeds the separation between inner- and outer-race defect frequencies, making the two faults indistinguishable in principle. 

Blind modulation classifier using higher-order cumulants and instantaneous signal statistics with an RBF-kernel SVM. Signals are generated rather than downloaded, so the channel model is explicit and each impairment can be swept independently. Full transmit-receive chain with root-raised-cosine pulse shaping, matched filtering and symbol-rate decimation. 

Simulated a visible light communication link using ASK modulation with wavelength division multiplexing across two wavelengths, evaluating BER and Q-factor under varying ambient light conditions. Modelled a 10 Gbps link over a 2 m line-of-sight channel in OptiSystem.

This project successfully developed a microcontroller-based tracking system using a PIC16F877A. . It measures distance via an HC-SR04 Ultrasonic Sensor and detects light intensity using LDRs, which in turn controls an L293D-driven stepper motor to automatically orient the device toward the brightest light source. 

This project involved the complete design and implementation of a stable, scalable, and secure network for a multi-floor, four-department organization. . We used Cisco Packet Tracer to configure isolated subnets, VLANs, DHCP, DNS, NAT, and routing protocols on routers and switches, integrating dedicated servers and both wired/wireless access points to ensure seamless and secure inter-departmental communication. 

Compared Ziegler-Nichols tuned PID against pole-placement state feedback for antenna azimuth positioning. State feedback reduced overshoot from 67% to 0.15% and settling time from 2.97 s to 0.38 s on the nominal plant. Frequency-domain analysis showed the ZN loop is conditionally stable with only 15.4° phase margin: because both plant and controller contain integrators, phase begins at −180° and the stable region is bounded below rather than above. A 52% loss of motor gain destabilises the loop entirely — the opposite of the usual failure direction. The comparison also reverses in steady state. Without integral action, state feedback tracked 20% low under a 20% plant gain error, where the PID integrator held to 1.2%. Neither design dominates: pole placement buys nominal performance, ZN buys steady-state robustness within a narrow stability envelope.

This project was about designing and building an excellent sound amplifier the kind that makes music come out of a speaker loud and clear. Specifically, we focused on a Class AB design, which is a great compromise between being energy-efficient and producing high-quality sound.