Interference prediction between lora and WIFI at 2.4 GHZ using machine learning

Abstract

Use of the 2.4 GHz frequency band for LoRa is beneficial for a range of applications due to absence of duty cycle and regulatory spectrum restrictions. However, operation in the 2.4 GHz Industrial, Scientific, and Medical (ISM) band exposes LoRa to interference from wideband technologies such as WiFi. The interference from the latter technology is very severe and impacts the reception of the LoRa packet. This thesis addresses that problem by developing a machine-learning-based interference prediction model between LoRa \,2.4 GHz and WiFi. The problem is formulated as a binary supervised-learning task in which packet reception success is predicted from physical signal parameters. The final feature vector consists of a received signal strength indicator (RSSI) difference term, LoRa RSSI, and the spreading factor. Three classifiers (Logistic Regression, Random Forest, and XGBoost) were trained and compared on a dataset constructed from real-world cross-technology interference experiments. Among them, XGBoost was selected as the final model, achieving 0.9681 test accuracy, 0.9808 F1-score, and 0.9867 receiver operating characteristic area under the curve (ROC-AUC) on the held-out experimental test set. To validate that it generalizes beyond training data, the selected predictor was evaluated on simulation-generated scenarios produced by a custom ns-3 LoRa \,2.4 GHz module, as part of a parallel thesis work by another student. The results reveal three distinct deployment regimes: short- and medium-range scenarios which are predominantly interference-limited, the 2000 m case behaves as a crossover point, and long-range 5000 m scenarios become sensitivity-limited, with higher spreading factors providing the best predicted packet reception ratio. The main contribution of this thesis is a validated interference-prediction component that can support future interference-aware resource-allocation mechanisms in LoRa \,2.4 GHz networks.