Ruslan Abdulin

In the following journal paper, supported by NSF government agency and the ECE department at California State University of Northridge, under the supervision of Professor Rasoul Narimani, I have designed a synthetic dataset generation algorithm based on NYISO load profiles and implemented multiple supervised ML models for online detection and localization of cyberattacks in power grids. The proposed method integrates ARMAConv spectral filter for local dependencies extraction and connectivity embeddings with an Encoder-only Transformer for global dependencies extraction and sequence processing. The evaluation on IEEE 14- and 300-bus test cases reveal that our model significantly outperforms the benchmarks as the number of buses in a test case increases. Technologies used: Python and libraries, such as PyTorch and Pandapower.

• See the paper

The onboard computer is responsible for UART communication with any flight controller that supports the MAVLink protocol in order to promptly pick up a dynamic payload upon landing. The computer utilizes the combination of a state-machine approach and an RTOS to ensure the pick-up procedure takes place with correct timing, allowing potential repetition of the mission without rebooting. The procedure includes sending a variety of infrared commands, lowering the gate until it touches the ground, recognizing that the payload is inside, raising the gate with the payload inside, and signaling to the drone that it is time to take off.

The computer is placed on a custom soldered circuit. The circuit(PCB prototype) has gone through the bring-up and hardware-in-the-loop testing. Additionally, I debugged the code using JTAG connection via OpenOCD with GDB. The circuitry enables reliable connections between peripherals and the computer, limits pick-up mechanism extraction by utilizing a DC-encoder, as well as debounces GPIO pins using an RC-based low-pass filter.

• Code

The dynamic payload is responsible for autonomously finding a path to the pick-up mechanism located on the bottom of the drone. The payload utilizes contrast-based line detector sensor, DC motors, camera that actively looks for a reflective tape wrapped around the pick-up mechanism and a flashlight for consistent lighting and therefore increased reliability of the computer vision algorithm.

There is a PCB specifically designed for integration of the peripherals with the MCU(ESP32S3)

• Code

At Guardinova LLC, I was responsible for integrating the ESP32S3 MCU with a OV3660 camera to stream its output via WiFi (connected as a station) to an HTTP server whenever the microcontroller is close (proximity is calculated based on RSSI using NimBLE) to a BLE Beacon.

• The code is not available for public disclosure

By following the official NRF24L01+ documentation, I've implemented a fully-configurable driver for TX and RF radio communication with STM32F407 Discovery board. The setup requires the HAL library and SPI connections between the board and the radio module. The driver was debugged using an external SWD debugger via OpenOCD with GDB.

The driver features dynamic payload sizing, adressing, CRC, auto-acknowledgement, and auto-retransmission for a custom amount of times in case no-ACK is received.

• Code

In accordance with the official documentation for the STM32F407-Disc1 board, I have implemented and well documented a bare-metal GPIO and SPI drivers. The drivers include a set of macros fully aligning with the memory map as well as bit positions and values for each register.

The drivers provide both polling and ISR-based APIs with callback functions.

• Code

The configurable driver for STM32 boards enables both polling and ISR-based data transfers to a Liquid Crystal Display via an I2C module which reduces the total number of pins.

• Code

At CSUN Advanced System Labs, I implemented an object detection algorithm that computes object vertex coordinates relative to the camera’s origin. The vertices are derived using an approximation based on combined convex hulls. The convex hulls are obtained from Luxonis OAK-D depth camera that is connected to a Linux-based NVIDIA Jetson board for processing. The vertices are passed to a pathfinding algorithm, which determines next camera positions to enable real-time obstacle avoidance.

Additionally, I was responsible for camera calibration using a checkerboard pattern.

• Code