A Simple Introduction to Ultra-Wideband Indoor Positioning Via Artificial Intelligence: Multi-Layer Perceptron Decomposition Architecture7 min readJust now--Indoor positioning (IP) for mobile Internet of Things (IoT) is one of the significant problems that must be solved to enable various applications such as navigation, proximity marketing, asset tracking, collision avoidance, and social distancing. While significant advances have been made in this area over the past decade, IP systems still suffer from low positioning accuracy due to non-line-of-sight (NLoS) scenarios. Ultra-wideband (UWB) technology has emerged as one of the key methods in achieving high positioning accuracy; however, its performance can significantly degrade due to multipath components in the wireless environment. Therefore, novel signal and information processing techniques are required to achieve robust positioning accuracy in indoor environments with multiple obstacles.In this article, I have introduced my conference paper, which presents UWB Indoor Positioning via novel Multi-Layer Perceptron (MLP) Decomposition Architecture. This article's technical content has been published as a conference paper (Cakan, E., ahin, A., Nakip, M., & Rodoplu, V. (2021, June). Multi-layer perceptron decomposition architecture for mobile IoT indoor positioning. In 2021 IEEE 7th World Forum on Internet of Things (WF-IoT) (pp. 253257). IEEE.)Artificial Intelligence (AI) techniques hold much promise for processing information in IP systems. In this article, I have introduced a novel AI-based architecture for mobile IoT indoor positioning, which is called Multi-Layer Perceptron (MLP) Decomposition. In our architecture, in the first stage, a bank of MLPs processes the position and distance information from each anchor. In the second stage, the outputs of the MLP bank are fed into a main MLP block. While the decomposition of MLPs into multiple stages has been used in the Machine Learning (ML) literature, to the best of the authors knowledge, this is the first work that uses a design based on MLP decomposition for indoor positioning.We demonstrated the performance of our architecture on actual data collected in an indoor environment with multiple obstacles. We proved that our architecture outperforms the benchmark processing techniques optimized on the same data: MLP, Linear Regression, Ridge Regression, Support Vector Regression, and the Least Squares Method for indoor positioning. These results show that our architecture can significantly advance the positioning accuracy of existing indoor positioning systems and enable mobile IoT indoor applications that require high positioning performance.Demonstration area for data collecitonThe Quest for Precise Indoor LocationAccurate indoor positioning is fundamental for numerous emerging applications. Imagine navigating a complex hospital building with turn-by-turn directions on your phone, efficiently locating a specific tool in a vast warehouse, or receiving location-based promotions in a shopping mall. These scenarios, and many others, rely on knowing the precise location of devices and individuals within indoor spaces.While various technologies have been explored, achieving high accuracy consistently remains a hurdle. Traditional methods often struggle with Non-Line of Sight (NLoS) conditions, where obstacles block the direct signal path, leading to inaccurate distance measurements. Ultra-Wideband (UWB) technology has emerged as a leading contender for high-precision IP due to its wide bandwidth. But even UWB systems can be affected by multipath signals bouncing off surfaces.Illustration of LoS and NLoS signal (left), Signal Strength x Time of Multipath (right) (Source: Ubisense)Introducing MLP Decomposition: A Two-Stage AI ApproachTo address these challenges, we have developed a novel processing architecture based on the principles of Machine Learning. The core idea is to employ a Multi-Layer Perceptron (MLP) Decomposition strategy, breaking down the problem into two distinct stages of processing.The architecture is designed for an indoor positioning system with multiple anchors (fixed reference points with known coordinates) and a mobile tag (the device whose position needs to be determined). The system relies on measuring the distance between the tag and each anchor.Stage 1-Individual Anchor Processing. In the first stage, for each anchor in the system, a dedicated MLP block processes the distance measurement to that specific anchor along with the known x and y coordinates of that anchor. This individual processing allows the system to learn the specific characteristics and potential biases associated with each anchors signal propagation. The output of each of these individual MLP blocks is a set of learned features.Stage 2-Data Fusion with a Main MLP. The second stage involves a central Main MLP Block that takes the features generated by each of the individual anchor MLP blocks as its input. This main block then fuses this information to produce the final estimated x and y coordinates of the mobile tag.This two-stage decomposition allows for a more nuanced and effective way of learning the complex relationships between the raw measurements and the actual location compared to using a single, monolithic MLP. Importantly, the entire architecture is designed to be end-to-end trainable, meaning all the individual MLP blocks and the main MLP block are trained together to optimize the overall positioning accuracy.How Does the MLP Decomposition Architecture Work?In the figure below, the MLP Decomposition architecture is displayed.First Stage: Each MLP Block fuses the di distance between T (the tagged mobile device) and Ai (the i-th anchor) with the xAi and yAi coordinates of Ai (which are inputs to this MLP block) to produce a set of features {{fi,q}q{1,,Q}}. Here, fi,q is the q-th output of MLP Block i.Second Stage: The Main MLP Block fuses these sets of features (received as the outputs of the MLP blocks in the first stage) to produce (xT, yT), which is the estimate of the current position of tag T.MLP Decomposition Architecture for Indoor PositioningEach MLP Block i consists of L hidden layers and an output layer (the L+1st layer). The Main MLP Block consists of Lmain hidden layers and an output layer with two neurons (corresponding to xT and yT). The entire architecture is end-to-end trainable, meaning that each MLP block in the first stage and the Main MLP block in the second stage are not trained separately.Experimental Setup and ResultsWe demonstrated the empirical results of our architecture based on a setup of a UWB system. We compared the results with alternative techniques that process the same set of positioning data. Our results indicate that our architecture outperforms all the alternative techniques studied.In our experiment, the deployment area for the anchors and the tag was a furnished living room of 3 m 2.1 m. It is emphasized that this environment has a rich multipath profile. In the experiment, n = 4 anchors were used and located near the four corners of the room. The room was divided into a 15 cm 15 cm two-dimensional grid, and the tag visited these grid points. At each visit point, 5 distance measurements were taken from all anchors at 4-second intervals.Application AreaThe table below presents the performance comparison of all techniques using 10-fold cross-validation. Performance was measured in terms of three metrics: (1) Mean positioning error; (2) Standard deviation (STD) of positioning errors; (3) Coefficient of determination R, which takes values in the range.MLP-based techniques significantly outperform Linear Regression, Ridge Regression, SVR, and the Least Squares Method. Most importantly, the MLP Decomposition architecture outperformed all other examined techniques, including MLP, in terms of mean positioning error. Compared to MLP, MLP Decomposition reduced the mean positioning error from 8.98 cm to 7.68 cm. This represents a significant improvement of 14.5% over MLP. Furthermore, the standard deviation of the positioning error for the MLP Decomposition architecture is comparable to that of MLP and is among the two lowest values across all techniques. Finally, the R coefficient of the MLP Decomposition architecture achieved the highest value among all techniques.Performance Results via 10-Fold Cross ValidaitonThe figure below shows the Cumulative Distribution Function (CDF) of the positioning error under the MLP Decomposition model. The results show that 48.03% of the positioning errors are within 5 cm, 84.94% are within 10 cm, and 98.03% are within 30 cm.The CDF of positioning error for MLP Decomposition modelConclusionIn this study, we developed a novel Artificial Intelligence-based processing architecture called the Multi-Layer Perceptron (MLP) Decomposition architecture for the indoor positioning of mobile IoT devices. Our architecture takes the positions of the anchors and the measured distance between each anchor and the mobile IoT device as inputs and produces the estimated position of the tag.We measured the performance of our architecture in a demonstration using ultra-wideband (UWB) as the underlying technology for distance measurement to each anchor. We found that our MLP Decomposition architecture outperforms all the following techniques in mean positioning error: MLP, Linear Regression, Ridge Regression, Support Vector Regression, and the Least Squares Method for indoor positioning. In particular, our cross-validated results showed that our architecture reduces the positioning error by 14.5% compared to MLP, which exhibited the second-best performance. This improvement is significant, considering the current competitive landscape of indoor positioning techniques.Although we have only presented our results for UWB in this paper, the processing architecture we have developed for indoor positioning is general and can be applied to any positioning technology that provides a measurement of the distance from the mobile IoT device to each anchor. Our MLP Decomposition architecture shows promise for ultra-high-precision next-generation indoor positioning of mobile IoT devices to enable applications such as navigation, proximity marketing, asset tracking, collision avoidance, and social distancing.