Advancing Indoor Positioning Systems with Machine Learning

Indoor Positioning Systems (IPS) have become essential tools in various industries, such as healthcare, sales, manufacturing, logistics, and construction. These systems help track items, manage proximity, and take inertial measurements, which can improve efficiency, accuracy, and safety in operations. However, existing IPS solutions often require external infrastructures, raising concerns about privacy and functionality, especially in sensitive factory environments. Additionally, deploying extra infrastructure can be costly and impractical in certain locations, like tunnels or mines.

Recent advances in Machine Learning (ML) offer a potential solution by relying solely on data collected from onboard sensors in Internet of Things (IoT) devices. Despite this potential, it is still uncertain which ML models work best with the limited resources typically found in these devices. This work focuses on creating an IPS that uses motion and Ambient Sensors to track moving objects in factories concerned about privacy.

#Problem Overview

The focus of this research is indoor positioning for tracking objects that follow a specific route in a factory. Standard approaches usually measure precise x-y coordinates, but this study prioritizes determining the relative position along a known path. Many traditional methods require external infrastructure to function effectively, such as radio frequency-based solutions that depend on consistent signal strength. However, for applications that only need to know where items are along a path (like in an assembly line), having an exact x-y coordinate may not matter as much.

Additional challenges arise with deploying infrastructure that could lead to privacy issues or accumulated errors. For instance, some approaches require prior information that can become outdated over time. In many cases, assumptions about motion or environmental conditions may not hold true, making the results less reliable.

This research aims to determine how to learn indoor positions from data collected by sensors, making it suitable for low-power devices like microcontrollers. By combining data from Motion Sensors (like accelerometers and gyroscopes) and ambient sensors (such as pressure and temperature sensors), the goal is to create a system that can provide accurate positioning estimates along a predetermined path.

#Application Scenario

To better illustrate the problem, consider a factory assembly line. In this context, workers often move products along specific routes designed to optimize the workflow. To keep track of these items in real-time, a robust localization solution is needed that fits within strict privacy guidelines.

The proposed solution involves evaluating lightweight ML models that can run on devices with limited memory and processing power. This includes testing algorithms such as decision trees, random forests, multilayer perceptrons, and convolutional neural networks. More complex models, like Long Short-Term Memory networks, may also be explored for their effectiveness in this application.

This work is unique because it tackles indoor positioning as a Multivariate Time Series classification problem using only motion and ambient sensors. The study will also analyze how well different ML models work under hardware constraints, showing their potential for real-world applications.

#Analyzing Existing Solutions

Current IPS solutions can be broken down into several categories based on technology, technique, and algorithms. Technologies range from satellite positioning to various sensor-based solutions. Unfortunately, many existing methods depend heavily on infrastructure, which can compromise privacy. Although there are wireless systems that protect privacy, most still require additional infrastructure, which can be impractical for the specific needs in factories.

The techniques in use include dead reckoning, vision-based methods, triangulation, and fingerprinting. Many of these methods necessitate external systems for operation. Dead reckoning, in particular, relies on initial positioning information and can suffer from cumulative errors over time. Vision-based systems can also raise significant privacy concerns.

Algorithmically, methods range from traditional statistical approaches, like least squares and maximum likelihood, to modern ML methods. Although ML has been increasingly applied to positioning systems, it has mainly been used in scenarios where the dynamics are more predictable.

#Defining the Problem

The task at hand is to localize assets within a building or factory accurately. Since assets may move along predetermined paths at varying speeds, the goal is to divide these paths into smaller segments. By doing this, the system can determine the most likely segment where the asset is located based on sensor data collected over time.

The research establishes that an asset will follow a specific path. However, the speed of movement and the time to traverse the path can differ with each instance. The focus here is not to measure exact coordinates but to identify which segment of the path the asset is currently on.

Data collected will consist of multivariate time series from motion and ambient sensors. These data will then serve as inputs for a classification task on edge devices, enabling real-time processing.

#Machine Learning Models

To tackle the indoor positioning challenge, various ML models will be implemented. Traditional techniques such as decision trees and random forests will be used due to their simpler structures, which make them suitable for low-performing devices. Additionally, more advanced approaches, including convolutions and LSTM layers, will be applied to seek enhanced pattern recognition among the gathered data.

1. Decision Trees (DT): This method works by splitting data into subsets based on certain criteria. The simplest configuration of DTs is fairly easy to interpret and can be suited for low-powered devices.

2. Random Forests (RF): This is an ensemble method based on multiple decision trees. It typically provides better accuracy by averaging the results from many trees but requires more memory.

3. Multilayer Perceptrons (MLP): A basic neural network model that includes multiple layers. It can be trained to recognize complex patterns but can be more demanding on hardware.

4. Convolutional Neural Networks (CNN): Initially designed for image processing, CNNs can also work on time-series data. By treating the data as a 2D grid, CNNs can identify local patterns and features.

5. Long Short-Term Memory (LSTM): This type of network is capable of learning long-term dependencies in sequential data, making it beneficial for time-series analysis.

Each model will be evaluated based on its accuracy, memory requirements, and speed in making predictions.

#Collecting and Preparing the Dataset

To verify the proposed method, a dataset called Motion-Ambient will be created. This dataset will comprise recordings from various sensors used in a practical scenario that mimics real factory conditions.

The data-logging setup will involve a movable device collecting data from various sensors, including inertial measurement units (IMUs), pressure sensors, temperature sensors, and others across three distinct paths. These paths will include both indoor and outdoor segments, ensuring a variety of conditions.

#Preprocessing the Data

Since sensors often collect data at different rates, it is necessary to preprocess the data to ensure consistency. Initially, any missing recordings will be filled using techniques that utilize nearby values in time. After that, a rolling mean filter will help smooth the data, followed by normalization to standardize the values.

#Selecting Features

From the raw data, important features will be chosen based on their relevance to the indoor positioning problem. The final feature set will be used for training the ML models, ensuring that the chosen attributes contribute effectively to making accurate predictions.

#Evaluating the Models: Results and Analysis

Once the models are trained using the prepared dataset, various metrics will be employed to evaluate their performance.

1. Accuracy: This represents the proportion of correct predictions made by the model. A higher accuracy indicates better performance in identifying the correct segment of the path.

2. Loc-score: This is a specialized metric that provides a more forgiving evaluation of classification accuracy by considering predictions that are close enough to the correct answer.

3. Memory Footprint: This indicates how much memory the model requires. Smaller models are more suitable for deployment on low-power devices.

4. Inference Latency: This measures the time it takes for a model to make a prediction. Lower latency is crucial for real-time applications.

5. Throughput: This metric shows how many predictions can be made in a given time frame, offering insight into the system’s efficiency.

#Model Performance

The results from the evaluations will be compared across all the models. It is expected that models like CNN-1D and MLP will show higher accuracy and acceptably low memory footprints and inference latencies, making them suitable for the intended application.

In summary, the goal is to present a balanced view of each model's strengths and weaknesses concerning the specific indoor positioning context. The findings will guide future implementations and potential improvements.

#Discussion and Future Directions

Data drift is a challenge common in ML applications, where changes in data over time can affect model performance. This research acknowledges that periodic retraining may be necessary to keep models up to date. The proposed indoor positioning method primarily focuses on following predetermined paths, but future work may involve introducing various scenarios, such as unexpected stops or obstacles.

There is also a need to generalize the models further, allowing them to handle various motions that occur in indoor settings, which can enhance their applicability across multiple environments.

Furthermore, aspects like seasonal effects on ambient conditions could be investigated to see how they influence model performance. Until now, the dataset used lacks representation for such variations.

The research aims to implement the developed system in real factory conditions, validating its effectiveness and practicality through real-world use.

#Conclusion

This study aims to develop a machine learning-based indoor positioning system that leverages motion and ambient sensors while avoiding reliance on external infrastructures. Through comprehensive model evaluation, the effectiveness of various machine learning approaches will be assessed in the context of factory environments.

The unique dataset generated as part of this work will contribute to future research in indoor positioning solutions. By focusing on practicality, privacy, and the needs of low-power devices, this project aspires to offer innovative solutions for industry applications, ensuring efficient tracking of assets while maintaining user privacy.