GAOBO HU

Project Overview:
This project aims to deepen the understanding of channel characteristics in wireless communications by analyzing the received power and amplitude distribution in indoor short-range wireless systems, as well as the directional characteristics of Multiple-Input Multiple-Output (MIMO) systems. The tasks include studying impulse response, Power Delay Profile (PDP), RMS delay spread, and estimating the transmission and arrival directions of signals in MIMO systems.

Project Content:
The frequency transfer function of an indoor short-range wireless system is measured using a Vector Network Analyzer (VNA) to analyze its temporal variation characteristics, impulse response, and amplitude distribution. The Power Delay Profile (PDP) and RMS delay spread are calculated to study multipath effects and signal dynamic properties. Additionally, measurement data from an 8×8 MIMO system is used to estimate the transmission and arrival directions of signals with traditional beamforming techniques and the Capon beamformer. The directional characteristics of signals are analyzed in conjunction with different modulation schemes, aiming to optimize the design and performance of wireless communication systems.

Project Outcomes:
The analysis of broadband characteristics and MIMO directional properties was successfully completed. The results demonstrate that increasing bandwidth improves delay resolution, and the signal amplitude distribution follows a Rayleigh distribution, indicating multipath propagation characteristics. Beamforming analysis based on the MIMO system effectively estimated the transmission and arrival directions of signals, providing theoretical support and experimental evidence for the design and optimization of wireless communication systems.

Project Overview: Project Content: Project Outcomes:

This experiment aims to analyze EEG signals using machine learning techniques and develop a system capable of predicting epileptic seizures. The system is designed to assist patients and caregivers in taking preventive measures before seizures occur. By utilizing CNN (Convolutional Neural Network) to process time-series data, the system identifies complex signal patterns to classify the risk of seizures.

• Data Preprocessing: The raw EEG data was cleaned to remove outliers and normalized; the dataset was divided into 80% training and 20% testing sets.
• Model Construction: A CNN architecture was designed, including convolutional layers for extracting local features, pooling layers for dimensionality reduction, and fully connected layers for classification.
• Model Training and Optimization: The model was trained using the cross-entropy loss function and Adam optimizer, combined with an early stopping mechanism to prevent overfitting. Grid search was employed to optimize hyperparameters.
• Model Evaluation: Model performance was assessed using metrics such as accuracy, precision, and recall. The training process was visualized by plotting loss and accuracy curves.

The experiment demonstrated that the model could effectively predict epileptic seizures, achieving high accuracy on the test set. The feature extraction capabilities of the convolutional layers, combined with the optimized model structure, made it highly effective in processing complex signals. This provides an efficient solution for seizure prediction.

Project Overview:
This experiment aims to deploy Fully Connected (FC) and Convolutional Neural Networks (CNN) on the Arduino Nano 33 BLE Sense for epilepsy prediction. It investigates model training, optimization, and performance on embedded devices.

Project Content:
The CHB-MIT dataset's EEG data was used, normalized, and divided into training, validation, and test sets. Two neural network models were designed: the Fully Connected (FC) model consists of two dense layers with L1 regularization, while the Convolutional Neural Network (CNN) is a lightweight 2D convolutional network comprising convolutional layers, pooling layers, and dense layers. After training, the models were converted to TensorFlow Lite format and further transformed into Arduino-compatible .h files using appropriate tools. Model optimization was performed by reducing the number of convolutional kernels and applying dynamic range quantization to minimize storage requirements, making the models suitable for operation on memory- and storage-constrained embedded devices.

Project Outcomes:
The deployment of FC and CNN models was successfully achieved. Results showed that the CNN model performed better in terms of memory usage and storage efficiency, while the FC model exhibited faster computation speed. The experiment validated the feasibility of running machine learning models on embedded devices and proposed efficient optimization methods.

Project Overview:
This project aims to familiarize with the iterative decoding methods of Low-Density Parity-Check (LDPC) codes and explore their applications in modern communication systems. The tasks include implementing a bit-flipping decoder to decode a small LDPC code and a larger code transmitted through a Binary Symmetric Channel (BSC). The decoder's Bit Error Rate (BER) performance is evaluated through Monte Carlo simulations.

Project Content:
First, a parity-check matrix is set up, and the received bit sequence is decoded using a bit-flipping decoder. The decoding success is determined by calculating composite metrics. Next, a larger LDPC code is used to decode text data transmitted through the BSC, and the decoded bit sequence is converted back into text. Finally, simulations under different channel error probabilities are conducted to evaluate the BER and Block Error Rate (BLER) performance of LDPC codes under various conditions. The results are compared with the performance of uncoded transmissions.

Project Outcomes:
By implementing the bit-flipping decoder and decoding LDPC codes of different lengths, the project demonstrates a significant improvement in BER and BLER within low to moderate channel error probability ranges, validating the effectiveness of LDPC codes in enhancing data transmission reliability. Although performance degrades at high error probabilities, LDPC coding still provides better protection compared to uncoded transmission.

Project Overview:
This project aims to achieve precise base station localization using Received Signal Strength Indicator (RSSI). The main tasks include estimating the path loss exponent and determining an unknown location based on RSSI values from multiple base stations.

Project Content:
First, data preprocessing is performed using a channel attenuation model, including steps such as outlier removal, median filtering, and Kalman filtering, to accurately estimate the path loss exponent and simulate environmental characteristics. Linear regression analysis is then applied to determine the signal attenuation index. By combining this with a weighted least squares method, the final location information is iteratively calculated based on signal strength from multiple base stations. Finally, using measurement data from various base stations, the distances between the receiver and the base stations are estimated through weighted least squares. Iterative calculations are conducted to obtain precise location coordinates.

Project Outcomes:
The project successfully developed and optimized the base station localization technology, enhancing localization accuracy through data preprocessing and linear regression methods. Despite limitations in the amount of data and the number of base stations, the method demonstrated its effectiveness in complex environments and provides theoretical and practical references for future research on localization technologies.