AI-Enhanced Wireless Localization Technologies: A Comprehensive Review

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Nov 29 2023

In a paper recently published in the journal Applied Sciences, the authors reviewed different artificial intelligence (AI)-assisted wireless localization technologies that can effectively address the limitations of existing wireless localization technologies. Specifically, they reviewed the use of AI algorithms to deal with the issues of localization performance deterioration factors.

Preventing deterioration of signal quality

Signal quality deterioration issues due to noise in ionospheric scintillation, atmospheric conditions, and wireless signals can be addressed by excluding all external factors and retaining only the original wireless signal. A method based on artificial neural network (ANN) uses all the link quality indicator (LQI), humidity, temperature, and received signal strength indicator (RSSI) values compared to the existing measurement model that utilizes only the received wireless signal RSSI value to estimate location.

Similarly, the ionosphere layer based on long short-term memory (LSTM) can be used for localization error reduction in long-distance over-the-horizon (OTH) situations. The approach uses the external factors in place of eliminating them. However, the method can only be employed if the ionosphere layer is stable, which is a major disadvantage.

Resolving spatiotemporal asynchronization

The issue of spatiotemporal asynchronization can occur in different situations. If the internal clocks of the device and anchor are different, the device receiving the signal remains unaware of the signal transmission start time, which leads to localization errors.

A method based on a Lagrange programming neural network (LPNN), a type of deep neural network (DNN), can address this problem. In this approach, the LPNN utilizes a time of arrival (ToA) measurement set as input training data, which includes the transmission start time uncertainty due to clock asynchronization.

The LPNN can simultaneously determine the terminal position, clock asynchronization, and the starting transmission time from the input. These data can be used to extract the refined device position. Spatiotemporal asynchronization can also be caused by the relative changes in the device positions. In such scenarios, proper features are extracted and then probabilistic data are obtained based on the features to adjust the location.

A convolutional neural network (CNN)-based approach can be employed where a CNN is used to learn the multiple-input and multiple-output (MIMO) channel spatial–temporal-frequency characteristics from channel state information (CSI) data. Initially, the CSI data within MIMO-orthogonal frequency division multiplexing (OFDM) Wi-Fi systems is collected, and a three-dimensional (3D) phase and amplitude matrix is constructed. 

Subsequently, the method utilizes this constructed matrix as the CNN input and learns phase and amplitude features as the data passes through the 3D pooling and convolutional layers. The output layer then generates probabilistic classification results for location estimation depending on the newly obtained MIMO data.

Non-line-of-sight (NLoS) event identification

Discriminating NLoS and line-of-sight (LoS) paths is a critical aspect of wireless localization. Recently, AI-based classification techniques have gained significant attention for discriminating NLoS and LoS. For instance, a recurrent neural network (RNN) model was proposed that uses CSI to discriminate NLoS and LoS channels. Although determining the existence of LoS is difficult through CSI measurements, the LoS existence can be inferred through variation patterns in CSI over time.

The proposed RNN model utilizes a CSI measurement sequence for discriminating NLoS and LoS channels, which enables localization system implementation in challenging NLoS environments. Similarly, a support vector machine (SVM)-based method was proposed that distinguishes NLoS and LoS using an SVM-based signal classification, and then CNN is applied to the distinguished NLoS signal to correct errors and recognize patterns. A hybrid weighting algorithm is eventually applied to the obtained signal data and the target point coordinates are obtained.

A CNN method was proposed that uses channel impulse response (CIR) for discriminating NLoS and LoS. This approach uses a fully convolutional network using CIR data as an input dataset in ultra-wideband (UWB) systems to extract features. Subsequently, the information required for classification is strengthened using the self-attention algorithm to improve the NLoS and LoS classification accuracy. However, the method can require additional information to supplement the limited UWB CIR information/data.

In another method using graph neural network (GNN), a data-driven approach based on GNN is utilized to address mixed NLoS/LoS environments and realize effective large-scale network localization. Moreover, an ANN-based approach, designated as a channel impulse response-based neural network (CIRNN), can be used to combine the features of channel parameters and CIR. In this approach, NLoS and LoS can be more accurately classified by inputting a small number of channel parameters and the CIR into the CIRNN.

Improving performance using miscellaneous methods

Several studies have been performed to enhance the localization technology performance using different AI algorithms. The device location can be estimated using simple classification tasks when a sufficient RSSI dataset can be used as training data. Different AI algorithms can be employed to improve the performance, including K-nearest neighbor (K-NN), random forest (RF), generative adversarial network (GAN), and variational autoencoder (VAE).