The Predictions and Detections of Time Series in Wireless Communications Using Machine Learning Approaches

Abstract

Wireless channel prediction has been the subject of significant research interest in recent years; however, it remains an open topic. Both conventional mathematical and statistical models have proven incapable of forecasting time-varying channel variables within coherent time windows. This is due to the fact that the mechanisms of these models are designed to fit a range of scenarios for specific categories, targeting minimal mean square errors over the entire dataset, rather than time-consequent prediction within coherent windows. Thus, these models are not well-suited for predicting time-consequent patterns within coherent windows. Conversely, artificial neural network (ANN) based approaches have been proposed and investigated for wireless channel prediction. In principle, ANN can learn and identify patterns that are hidden in data. However, extracting and predicting wireless channel variables using ANN presents significant challenges for researchers. Additionally, the commonly utilized long-short term memory (LSTM) approach has not yielded satisfactory results due to the presence of interferences, outliers, and noise that cannot be avoided. The impact of noise on the performance of machine learning models in predicting future values in time series data can be substantial. In the context of long-term predictions, noise can be particularly problematic as small errors in the prediction can accumulate over time, leading to significant deviations between predicted and actual values. Noise is characterized by random variations or errors in the data that can obscure the patterns that machine learning models are designed to recognize and learn. Additionally, the noise can lead to overfitting of the model, whereby it fits not only the underlying patterns in the data but also the noise. This can result in poor generalization performance, where the model performs well on the training data but poorly on new, unseen data. In response to these concerns, two novel machine learning models have been designed. One model targets long-term time series predictions, while the other is designed for time series pattern detections. For the prediction model, we reconfigure the conventional LSTM cell by introducing new Infinite impulse response (IIR) gates into the LSTM cells, which are specifically designed to remove noise and interferences. These gates are self-adaptive through backpropagation during the training phase to optimize the model. To validate the proposed IIR-LSTM prediction model, experimental Global Position Systems (GPS) distance error datasets were utilized to assess the effectiveness of the model. In addition, the proposed model was evaluated alongside several commonly used time series forecast models, and the results demonstrate its ability to perform well on diverse datasets, with the best performance of long-term predictions achieved. For the pattern detection model, we proposed a novel model that utilizes the position information of the softmax layer to estimate the confidence of prediction and determine the patterns, if there were. To increase detection tolerance to noise, the positions of the most likely class and its adjacent classes were utilized. A dedicated standard deviation layer was introduced to calculate the position variance of the Softmax outputs while assigning a set of weights to them to optimize the model prediction performance under various noise conditions. The model was tested by using both a simulated dataset and real wireless sensor data. The results demonstrate that the presented method achieves a higher detection accuracy, particularly under high noise conditions, when compared to the one dimensional Convolutional Neural Network (1D-CNN) Auto-encoder