ADAPTIVE MACHINE LEARNING BASED WIRELESS CHANNEL INTELLIGENCE

Abstract

The performance of a wireless communication link depends heavily on the state of the radio channel between the transmitter and the receiver. Path loss, shadowing by buildings and vegetation, multipath fading, Doppler shift caused by movement, and interference from other transmitters all change the received signal from moment to moment. Modern systems such as Wi-Fi, LTE, and 5G adapt their modulation, coding, and power to the channel, but the conventional methods they use rely on fixed thresholds and simple models that do not always follow real conditions well. This paper presents an adaptive machine learning based wireless channel intelligence system that learns to classify channel conditions, predict future signal quality, and recommend suitable link parameters. The system uses two sources of data. The first is a simulated dataset generated in Python, in which signals modulated with BPSK, QPSK, 16-QAM, and 64-QAM are passed through additive white Gaussian noise, Rayleigh, and Rician fading channels with different Doppler spreads and signal-to-noise ratios. The second is a measured dataset collected with ESP32 boards and a low-cost software-defined radio receiver, recording received signal strength, signal-to-noise ratio, packet error rate, and link quality indicators at different distances, indoor and outdoor locations, and movement conditions. Both datasets are cleaned, synchronised, and converted into windows for analysis. Features that describe the channel are extracted from each window, including mean and variance of received power, level crossing rate, average fade duration, estimated Doppler spread, Rician K-factor estimate, error vector magnitude, and statistics of packet errors. Several machine learning models are trained and compared. Random Forest, Support Vector Machine, K-Nearest Neighbours, and XGBoost classifiers identify the channel type and condition, regression models including Random Forest and LSTM forecast the signal-to-noise ratio a short time ahead, and a classifier maps the predicted channel state to the modulation and coding scheme expected to give the highest throughput within a target error rate. Evaluation uses accuracy, precision, recall, and F1-score for classification, RMSE and MAE for signal quality prediction, and achieved throughput and packet error rate for link adaptation. The results show that the learned models identify channel conditions with high accuracy and that the LSTM predictor tracks fading trends well at pedestrian speeds. Link adaptation based on predicted rather than last-measured quality achieves higher average throughput than a fixed threshold table while keeping errors close to the target. An online update step allows the models to adjust as new measurements arrive. A dashboard displays live link measurements, the current channel classification, the predicted signal quality, the recommended modulation scheme, and the history of adaptation decisions. The work illustrates how machine learning can add adaptive intelligence to the physical and link layers of wireless systems. Future work includes channel state information from multiple antennas, reinforcement learning for adaptation, and implementation on programmable radio hardware.