الفهرس 
AbstractOnly 14 pages are availabe for public view
 |  | from | 108 |  |  |
| | | from | 108 | | |
Abstract
Accurate localization systems using visible light communication (VLC) are proposed based on several machine learning (ML) algorithms and received signal strength (RSS). In this thesis, our main objective is to provide accurate and compatible localization systems with distinct environments. Therefore, several models are proposed. Specifically, underwater and indoors models. The general mainframe of the proposed thesis comprises two phases: dataset collection and ML training. Firstly, datasets are collected with the aid of MATLAB and Ze-max OpticStudio Monte Carlo ray tracing software. Each line of sight (LoS) and nonline of sight (NLoS) signals are considered in building 2D and 3D environments. The second phase is training the acquired datasets with
several ML algorithms using a data mining tool. In order to validate the robustness of the proposed systems, several evaluation metrics are applied. Specifically, we estimate the following parameters: classification accuracy (CA), area under the curve (AUC), training time, testing time, precision, F1-score, recall, logloss and specificity. As well as confusion matrices, receiver operating characteristic (ROC) curve. Moreover, root mean square error (RMSE) is used as additional metrics.
More in detail, in this thesis, 2D underwater localization systems are proposed with configuring 40000 receivers in 10 × 10 m^2 area for measuring the channel gain in the pure sea water. The RSS values are estimated with the aid of Ze-max software. An observable note here is that the NLOS signals are considered. The obtained data set is trained with a neural network (NN). Moreover, several trials are applied to determine the superb neuron number, activation function and training algorithm. The optimum trial achieves results as follows: 99.1% for AUC and 98.7% for CA, precision, F1-score, and recall. Moreover, the
gained results for logloss and specificity are 7.3% and 99.3% respectively. Aside from, a 2D indoor localization system is proposed with an area of 5 × 5〖 m〗^2. LOS channel model is constructed. 10000 RSS values are estimated and trained with distinct ML algorithms. The mentioned evaluation metrics are applied and achieve the following results: 99.5% for AUC, and 99.4% for CA, precision, F-score, and recall. The logloss is 4%. While, precision is 99.7%. Further,
we were able to achieve RMSE of 0.1 cm.
In related context, by following the same mechanism in building 3D indoor
localization system of an office with size of 6 × 6 × 4〖 m〗^3. The random forest algorithm, nominated ML one in indoor localization work, is utilized in training 43200 RSS values. Here, an accurate localization system is able to obtained AUC of 97.9%, CA, precision, F1-score and recall are 96.3%. Whilst, the logloss and precision perform 16% and 98.2% respectively. Further, RMSE is estimated to evaluate the robustness of our 3D localization system. Accordingly, we achieved RMSE of 0.024 cm.