1 Introduction

Location awareness has long attracted much attention due to the increasing widespread of smartphones, smart vehicles, and Internet of Things (IoT) connected devices. In addition, the need of location information in various applications such as driverless cars, smart homes, passive keyless entry systems, etc. lead researchers to explore different aspects of localisation systems.

Localisation accuracy and broad availability of global navigation satellite systems (GNSS) such as global positioning system (GPS) and Galileo in smartphones and other devices are in most cases sufficient for outdoor applications. However, since their signals are too weak or degraded by multipath effects in indoor environments, other approaches are required for indoor applications. These approaches can be part of existing communication systems (e.g., WiFi access points) or be a dedicated technology like ultra-wide band (UWB). According to studies in the literature, UWB technology is one of the best candidates for indoor localisation. This is due to high time resolution, low power consumption, and low complex hardware of UWB technology. Therefore, manufacturers have recently started to integrate UWB technology into mobile devices for location-based services (LBS) [1].

With the help of high bandwidth communications, UWB radio technology (IEEE 802.15.4z [2]) can achieve high time resolution. As a result, localisation with high accuracy can be achieved. However, multipath propagation (MPP) and non-line-of-sight (NLOS) conditions in UWB-based localisation technology, which frequently occur in indoor environments, are the challenges that impede the achievement of high accuracy [3]. Another challenge that compromises the accuracy of localisation systems is an intrusion attack. Since those systems are often used in unattended environments, security threats are crucial concerns to consider [4].

To solve the challenges, different classical signal processing methods and AI-based approaches can be used. Many recent publications demonstrate the effectiveness of AI algorithms in extracting knowledge, learning important features, and improving localisation performance. The advantages of AI over classical approaches are as follows: AI approaches are significantly more effective than traditional approaches in complex nonlinear problems. Due to the ease of adjusting AI-based approaches in case new data sets are added, AI algorithms can offer scalable solutions. Furthermore, unlike statistical methods, AI algorithms can easily be expanded to offer stable performance in a variety of environmental circumstances. In comparison to traditional approaches, they can incrementally adapt to changing environmental scenarios, thanks to their online learning abilities.

In this chapter we will have an overview of the challenges in UWB-based localisation systems and of the AI algorithms that can address these challenges. We have investigated 26 existing work in the literature [5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30] accordingly.

2 Method Overview

MPP, NLOS, and attacks distort the accuracy of UWB-based localisation systems. All considered AI approaches in this chapter aim to improve accuracy with or without considering also energy efficiency or security concerns. Therefore, we focus on maintaining or improving the accuracy of UWB-based localisation systems while having different additional objectives.

To improve accuracy only, these objectives vary significantly, considering if the localisation method is range based or range free. For range based approaches, where the accuracy of the inter-node distance estimation is of most importance, these objectives are (i) feature selection for extracting the most relevant features, (ii) NLOS detection for eliminating erroneous ranging information, and (iii) NLOS mitigation for error compensation. There is ample existing research in the literature which provides solutions to select most relevant features, detect, and mitigate ranging errors using AI algorithms. Clustering and AI-based dimensionality reduction (DR) techniques are frequently used for feature selection. For the detection and mitigation of NLOS errors classification, regression, clustering, etc. techniques are extensively used.

In range free approaches fingerprinting determines the localisation accuracy. The processing of high-dimensional data is a common concern. The utilized AI algorithms in this domain mainly reduce the dimensions of the data and thus reduce computational complexity. Therefore, (iv) dimensionality reduction is the objective in range free UWB-based localisation systems. In most cases AI-based DR and clustering techniques are proposed here.

Higher accuracy often comes with higher complexity. Therefore, energy efficiency has to be considered in the case that power is limited. The objective for the aim of accuracy along with energy efficiency is (v) complexity reduction.

For the aim of accuracy with energy efficiency concern, most related works apply complexity reduction methods. There has been some research on using AI algorithms to enhance accuracy along with complexity reduction.

If accuracy is compromised through attacks, the integrity of UWB measurements deteriorates. Hence, the secure and reliable operation of the system is compromised. There has been limited research focused on maintaining the system security together with accuracy, while focusing on the objectives (vi) trustworthiness and (vii) intrusion mitigation. Security concerns in UWB-based localisation have gained some attention recently. The aforementioned three aims and seven objectives are depicted in Fig. 1.

Fig. 1
A flow chart of A l enhanced U W B localization systems from top to bottom. Only accuracy consists of range-based and range-free. Accuracy and energy deficiency consist of complexity reduction and accuracy and security consist of trust worthiness and intrusion mitigation, and others.

Classification of the objectives of AI-enhanced UWB-based localisation systems

3 AI for Solving UWB-Based Localisation Challenges

3.1 Localisation Challenges

As stated, localisation methods that are considered in the references presented in this chapter can be categorized into range based and range free approaches. Therefore, their challenges are described in detail in the following two subsections.

3.1.1 Range Based Approaches

With range based approaches, each anchor node has an estimate of how far it is from the target node. The estimates can be achieved through different measurements of the received signals. These measurements can be received signal strength indicator (RSSI) [31], time of arrival (TOA) [32], two-way ranging (TWR) and variants [33], time difference of arrival (TDOA) [34], and hybrid ones [35,36,37]. The range estimates along with the known position of the anchor nodes are used to compute the position of the target node. It is worth noting that in a two-dimensional space, at least three anchor nodes are required to compute the position of the target node. In an ideal case, the position of the target node is the intersection of imaginary circles with radius of the true distance, as shown in Fig. 2a. However, in real environments, due to several errors, these circles rarely intersect at a single point, which leads to an error in the computed position of the target node. For a single position estimate, as the distance estimates deviate from the true distance, it yields errors. The sources of this distance deviation either depend on the system, such as measurement noise or dilution of precision due to anchor node placement, or on the environment such as MPP and NLOS. While the system dependent errors cannot be influenced by algorithmic processing, MPP and NLOS can be observed by channel state measurements and can hence be integrated into the algorithmic design.

Fig. 2
A schematic of positioning techniques. It represents range-based and range-free. Range-based depicts 5 anchors with curved trajectories meeting at a point from all five anchors. Range free depicts straight-line trajectories meeting the target from all five anchors.

Schematic diagram of positioning techniques for ideal cases

3.1.2 Range Free Approaches

Range free approaches mainly comprise RF fingerprinting and AOA. In RF fingerprinting data from a target node is collected in an offline phase at access points (APs) for multiple known positions, the so-called reference points (RPs). This data is used to construct a radio map. By comparing measured data at APs with the radio map, it is possible to estimate the target node position in the online phase in real-time.

The accuracy of fingerprinting approaches depends, among others, on the number of RPs. Creating a radio map for a large area is thus a tremendous effort if a reasonable accuracy is required. Furthermore, if the position of even one AP is changed, the offline database needs to be recreated. Additionally, the high dimension of signals in most cases is another concern of fingerprinting approaches, which increases the computational complexity.

If AOA information is available at the anchor nodes, triangulation can be used to compute the position of the target node. It requires at least two anchor nodes to compute the position of the target node in a two-dimensional space. In an ideal case, the position of the target node is the intersection of imaginary lines, representing the angle estimates, as shown in Fig. 2b. However, in real environments, the lines do not cross at a single point resulting in a localisation error. Angle estimates that deviate from the true angle at a single position produce errors. Similar to the range based approaches, the error source can be either system-related or environmental, like MPP and NLOS. MPP and NLOS can be observed by channel state measurement and thus integrated into the algorithmic design, whereas system dependent errors cannot be affected by algorithmic processing.

3.2 AI Algorithms in UWB-Based Localisation Systems

Location accuracy, energy efficiency, and security are the most challenging problems in LBS that can be solved by AI algorithms. The ability of AI to learn useful information from input data with known or unknown statistics is its most significant advantage. As shown in Fig. 3, we briefly categorize the AI algorithms which have been used by researchers for solving the aforementioned challenges in UWB-based localisation systems. In the following, a brief background on the utilized AI algorithms is given.

Fig. 3
A flow chart depicts A l enhanced U W B-based localization from top to bottom. It consists of supervised learning including classification and regression, unsupervised learning with D R and clustering, and reinforcement learning consists of decision learning.

Classification of AI-based methods used for UWB-localisation in the literature

3.2.1 Supervised Learning

Supervised learning is a subclass of AI-methods that refers to algorithms that build a predictive model using data points with known outputs. As a general rule, supervised approaches are applicable to problems where labels are available. Supervised learning approaches can be separated into classification and regression.

Classification algorithms categorize new observations into categorical classes on the basis of training data. Regression algorithms model the relationship between a certain number of features and continuous target variables on the basis of training data.

Several approaches for carrying out the aforementioned tasks have been proposed in the literature which are briefly discussed in the following.

Naive Bayes (NB) is a probabilistic learning approach based on Bayes’ theory used for classification. This algorithm assumes that all variables in the data set are naive, which means that they are not related to each other. The implementation of NB is fast and straightforward. However, its drawback is the need for independent features. Typically, features in real-life applications are, however, correlated, which adversely affects the performance of this classifier.

K-Nearest Neighbors (KNN) is a supervised learning algorithm that is both simple and widely used in the field of AI. Assuming that \(K = 10\), the first 10 available data points in the data set with the smallest distance from a new data point are chosen to select the label of the new data point based on majority voting.

Modified K-Nearest Neighbor (MKNN) is a type of weighted KNN that selects the label of the new data point based on a modified majority voting in which the distances between the data point and its neighbors weights their votes.

Perceptron learning methods are based on neural networks and comprise a series of algorithms that try to recognize underlying relationships in data through processes that mimic the operation of a human brain. Networks including Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) are available in the literature. For the sake of brevity, the network architectures are not covered in this chapter. These methods can be used either for classification or regression tasks.

Logic-based approaches can also be used for performing classification or regression tasks. A fuzzy inference model uses uncertainty to estimate target variables. The Decision Tree (DT) is a map of possible outcomes of a series of related choices or options that allows an individual to weigh possible actions in terms of costs, probabilities, and benefits. A decision tree typically starts with an initial node, after which possible outcomes are branched from it, and each of those outcomes leads to other nodes, which in turn create branches of other possibilities. This branching structure finally turns into a tree-like diagram. An alternative to DT is Random forest (RF), which combines a number of decision trees, to predict the label based on the majority of votes in each tree. More trees in the forest lead to higher accuracy and avoid the problem of overfitting.

Kernel-based approaches map input data into a higher dimensional feature space using a kernel function. Then, the underlying relationships of data will be recognized in the feature space. Published kernel-based approaches for UWB-based localisation include Support Vector Machine (SVM), Relevant Vector Machine (RVM), and Support Vector Data Description (SVDD). These methods can be used either for classification or regression tasks.

3.2.2 Unsupervised Learning

In contrast to supervised learning, unsupervised learning refers to AI algorithms for the identification of patterns in data sets that contain neither classified nor labelled data points. Without the need for human intervention, these algorithms help discover patterns or data groupings. DR and Clustering are two typical fields of unsupervised learning approaches which are used in AI-enhanced UWB-based localisation systems.

DR is a process which transforms data from a high-dimensional space into a low-dimensional space. The aim is to learn relationships between features and represent the data using so-called latent features that relate to the original features of the data. The aim is usually to reduce the complexity of a model and avoid overfitting. Principal Component Analysis (PCA) is a well-known unsupervised learning technique for reducing data dimensionality. It is a linear projection from a higher-dimensional input space to a lower-dimensional output space. The catch is to minimize the amount of information or variance lost when variables are removed.

Clustering is the process of dividing unlabelled data points into groups (or clusters) such that data points in a group have a greater degree of similarity than data points in other groups. Expectation-Maximization (EM) is a model-based clustering technique that has been used for UWB-based localisation. It is an approach for performing maximum likelihood estimation in the presence of latent variables. This is accomplished by first estimating the values of the latent variables, then optimizing the model, and then repeating these two steps until convergence is achieved.

3.2.3 Reinforcement Learning

As the name implies, reinforcement learning (RL) involves the process of learning through trial and error. RL draws its inspiration from the learning behavior of humans, which use past experiences to react to new circumstances. RL decisions are made based on received rewards and penalties. The algorithm is rewarded for correct decision and penalized for an incorrect decision. Q-learning is a model-free reinforcement learning algorithm that determines how useful a given action at each state is in gaining some future reward. It can manage problems regarding stochastic transitions and rewards without requiring adaptations.

4 Overview of Related Work

In this section, we briefly describe the related work in the field of UWB-based localisation that utilize AI algorithms to enhance accuracy, energy efficiency, and security. We categorize related work into three aims (according to Fig. 1): accuracy only (Table 1), accuracy along with energy efficiency (Table 2), and accuracy together with security (Table 3). The objective(s), the AI algorithm(s), the description, and the remarks are all mentioned for each related work in the tables.

The majority of work in Table 1 concentrate on NLOS detection and mitigation. In the most cases, supervised and unsupervised learning approaches have been used for NLOS detection and NLOS mitigation, respectively.

Table 1 AI-enhanced UWB-based localisation work in literature whose aim is accuracy only

The majority of work in Table 2 concentrates on feature selection and dimensionality reduction. In the most cases, unsupervised learning approaches have been used for these purposes.

Table 2 AI-enhanced UWB-based localisation work in literature whose aim is accuracy and energy efficiency

Considering security aspects in AI-enhanced UWB-based localisation systems is less investigated in the literature. As shown in Table 3, there are only two publications in which supervised learning approaches have been used for security concerns.

Table 3 AI-enhanced UWB-based localisation work whose aim is accuracy and security

5 Application Example

In the following application example, which has been developed in the course of the InSecTT project, we focus on LOS/NLOS detection [8] and trustworthiness [29] objectives. We describe how KNN algorithms can effectively address these challenges.

5.1 KNN for LOS/NLOS Detection

As stated in [8], NLOS conditions in UWB-based localisation systems lead to errors and must be detected. This is a binary classification problem with LOS and NLOS classes. From a data perspective, [8] uses channel impulse response (CIR) estimates to calculate the received timestamps as input data (see Fig. 4).

Fig. 4
Three illustrations of the channel impulse response and a timestamp. The tag and anchor nodes from 0 to 4 of the first diagram are connected with slanting arrows. The sensor curves T X a, T X e, and T X b denote fluctuations.

The channel impulse response is estimated at each node (left side) from which the received timestamp is derived (right side) (from [8])

LOS and NLOS conditions are distinguished by comparing the received timestamps. The timestamps for all LOS cases are similar and the same is true for NLOS. Thus, KNN is well suited for this classification task. However, several key points have to be considered. Firstly, the used distance metric affects performance. Choosing the K value as hyperparameter is another concern. If one chooses a too small value, it will lead to unstable decision boundaries. However, a large value can be computationally very expensive. As a rule of thumb, the optimal K value is usually the square root of n, where n is the number of samples.

To formulate the KNN algorithm for this LOS/NLOS detection problem, assume that we have training data D given by

$$\begin{aligned} D=\{(x_1,y_1),(x_2,y_2), \dots , (x_n,y_n)\}\,, \end{aligned}$$
(1)

where \(x_i\) and \(y_i\) are the sample and the label of the ith input data, respectively. It is obvious that the samples are the vector of timestamps and the labels are either LOS or NLOS.

For a test point \(x \in \mathbb {R}\), a set \(S_{x}\subseteq D\) is defined as a set of K neighbors. Using the function dist that computes the distance between two points in \(\mathbb {R}\), a set \(S_{x}\) of size K can be defined as:

$$\begin{aligned} dist(x,x')\ge \max _{x''\in S_{x}}{dist(x,x'')}\,, \forall x' \in D\backslash S_{x}\,. \end{aligned}$$
(2)

Peterseil et al. [8] uses the KNN algorithm to detect weak NLOS conditions in the double sided two way ranging (DS-TWR) scheme and assigns a correct label to the measurement data. In [8], a packet-wise approach is proposed instead of a cyclic approach [38] which outperforms the cyclic approach by up to 15 cm improvement of measurement error in a collected data set by considering only those measurement data, which are classified as taken from LOS conditions.

5.2 KNN for Error Mitigation and Trustworthiness

NLOS mitigation is another objective in UWB-based localisation systems that can be addressed with KNN-based algorithms. The goal is to estimate the measurement error due to the NLOS condition and use it to correct the raw measurement. As we have seen, a NLOS condition can be detected by the KNN algorithm and thus it can be compensated for in the positioning computation. This translates in a regression problem. [29] uses first path power and mean excess delay derived from the received signal and the CIR, respectively.

The KNN algorithm assumes that the measurement error of training and testing samples is similar if they are close in feature space. The predicted measurement error

$$\begin{aligned} \hat{e} = \frac{1}{K}\sum _{i=1}^K e_i\, \end{aligned}$$
(3)

is therefore calculated as the average error of the K nearest neighbors in the feature space.

Fig. 5
Two graphs of first path power versus mean excess delay and n versus e k in meters. a. A circle is placed at (20, negative 95). b. It denotes a high at (0.5, 15). The values are approximate.

a Feature space; for an example feature measurement the circle depicts the KNN-neighborhood (for the sake of clarity, errors higher than 1 m are truncated to 1 m) and b Histogram of label values from neighborhood (from [29])

In practice it can happen that certain areas of the feature space contradict each other, leading to error-prone estimations. In [29], a modified KNN algorithm (mKNN) was proposed by additionally evaluating the standard deviation of the neighborhood

$$\begin{aligned} \hat{\sigma } = \sqrt{\frac{1}{K}\sum _{i=1}^K (e_i-\hat{e})^2}\,. \end{aligned}$$
(4)

In Fig. 5 one can distinguish between areas (or neighborhoods) with mostly low measurement error represented by green dots, areas with mostly high measurement error represented by red dots and areas with “mixed” error in which red and green dots appear with similar density. In neighborhoods with mostly green or red dots the standard deviation of the measurements is low because the condition is clearly either LOS or NLOS. In “mixed” neighborhoods, however, the LOS/NLOS condition is unclear an thus the standard deviation is high. Thus, by mapping \(\sigma \) to a range between 0 and 1, it can be interpreted as a trustworthiness measure, indicating whether the estimation of the measurement error, i.e. the correction term, is likely to be accurate or not.

Peterseil et al. [29] shows that the modified KNN algorithm reduces the ranging root mean square error from 36 cm if the whole data set is used, to 19 cm those 70% of the collected data with highest trustworthiness score are used.

6 Conclusion

With the growing demand for accurate and reliable location information in wireless networks, the use of UWB technology becomes increasingly important. UWB-based localisation systems face a variety of challenges that can be addressed by AI algorithms. All AI algorithms covered in this chapter aim to improve accuracy, with or without consideration of energy efficiency or security aspects. The larger part of the related work uses AI algorithms to improve the localisation accuracy. Also research on energy efficiency together with appropriate accuracy has been presented in several publications and was shown in an application example developed in the scope of the InSecTT project. However, AI algorithms to improve the security of localisation systems have only been covered in a few works and require more research.