In this study, Earth's magnetic field signals have been investigated to determine mobile user's location. In theory, Earth's magnetic field does not change during the day at a certain point. But, the various noise effects that are exposed during the measurement causes deviations in the measured signal. In this study; Kalman Filter, LOESS, Savitzky-Golay filters are adapted with two different approaches to purge Earth's magnetic field values from noise. KNearest Neighbour and Random Forest models have been trained with filtered signals and the locations of the mobile user are determined. Relevant systems have been tested by using RFKONDB which is existed in literature. The purpose of this study is to measure how these filters should be adapted to an Earth's magnetic field based indoor localization systems. Digital sensors, which are integrated mobile devices, can use different measurement techniques. In a heterogeneous environment, noise reduction filters can show a different effect. Two different test scenarios and two different noise reduction models, with the 3 noise reduction techniques, are developed to find the best case.
Indoor localization is particularly difficult due to the high dynamics and obstacles of
the environment. Therefore, there is no generally accepted positioning system in
indoor areas. The methods used for localization in closed areas are considered as
deterministic and probabilistic methods [
At first glance, it seems to be the most logical way to use the Earth's
electromagnetic field values in the design of a generally accepted indoor positioning system with
a unique magnetic field value in every part of the Earth. However, the Earth's
magnetic field information cannot be accurately measured due to disturbances caused by
ferromagnetic objects in indoor locations [
At the same time, electromagnetic signals collected from an environment are
exposed to the hard-iron effect caused by substances such as nickel-cobalt in the
environment. According to this, it is aimed to eliminate the noise values in measured
signals to create an indoor localization system independent from infrastructure hardware
by using the magnetic field values of the Earth. For this purpose, the RFKONDB[
The organization of the paper is as follows: In Chapter 2, indoor localization studies carried out by using Earth’s magnetic field values are detailed. The signal maps, filtering methods, filtering approaches and test scenarios used in the study are described in detail in Chapter 3. Finally, test results obtained from test scenarios are presented in Chapter 4.
The use of earth's magnetic field signals in indoor localization systems provides a powerful approach, which can be designed completely without infrastructures and allows for continuous positioning. The design is difficult due to the electromagnetic noise in the environment affects the sensor measurements of electromagnetic data.
A fingerprint-based approach is proposed in the AMID [
[
[
In this section, methods are introduced detailed. 3.1
K-Nearest Neighbour technique is a machine learning algorithm that has an easy implementation. It is a classification model based on finding the similarity rates between samples. The distance between samples is calculated with a Euclidean distance formula. If it is considered x=x1, x2, …, xn and y=y1, y2, …, yn as two samples taken from indoor areas the Euclidean distance calculation is presented in Equation 1. (1)
Large Euclidean distance symbolizes that two samples are like each other, while a smaller distance symbolizes that two samples are less resembled. To classify sample data; The Euclidean distance value for all points in the data set is calculated. It is classified according to the majority vote of the nearest K neighbors. 3.2
Random Forest is a tree-based classification method. It creates multiple decision trees in the dataset to be used for classification. The forest is formed with more than one decision tree. In the obtained forest, the subclass that represents the best value in the dataset is selected and graded. The decision tree algorithm tries to solve the problem by using tree representation. Each internal node of the tree corresponds to an attribute, and each leaf corresponds to the node class label. It places the best attribute of the dataset on the tree root. The subset divides the training set. Each subset is created to contain data that has the same value for an attribute. This process is done for all branches of the tree.
The primary challenge in implementing a decision tree is to determine which
attributes should be considered as the root node and count of levels. The feature
selection approach is adapted for this. Gini Index, entropy and twoing are metrics to
determine how often a randomly selected item is detected incorrectly.
3.3
The Kalman filter is described as one of the most important discoveries of the 20th
century. Although it is named as a filter, it is used in linear systems for estimating the
next step. With its recursive structure (re-inputting the outputs into the filter) is the
only filter that minimizes the estimation error in the existing filters. The Kalman filter
has two equations for estimation and correction [
The measurement value of a signal is obtained in the previous case . The control signal and are the previous operational noises. A, B and H are general representations of matrices. These values can be treated as numerical numbers. The A matrix is the state transition model, the B matrix is the control model, the H matrix is the measurement model. The measurable value of a signal is presented in Equation 3. 3.4
Savitzky-Golay Filter (SGF) is presented in 1964 by Abraham Savitzky and
Marcel J. E. Golay [
LOESS is one of the most popular kernel type smoothing filters. The LOESS filter is
a nonparametric method for estimating regression surfaces and has great flexibility.
There are no global assumptions about the parametric form of the regression surfaces.
LOESS fits nonparametric models, supports the use of multidimensional data,
supports multiple dependent variables and performs iterative reweighting to provide
robust fitting when there are outliers in the data [
Assume that i=1,…, n and Yi represents the measurement value of the corresponding xi where f(x) represents an unknown function and represents a random error in the observations or variability from sources not included in the . The regression function f(x) can be locally approximated by the value of a function in some specified parametric class. Such a local approximation is obtained by fitting a regression surface to the data points within a chosen neighborhood of the point x [23].
LOESS is ideal for modeling complex processes as a very flexible filter. But it requires very large and densely sampled datasets to produce good models. Increasing the size of the dataset results in increases experimental costs. 3.6
In this study, the effect of noise reduction/cleaning filters, which can be used in the localization/positioning systems based on Earth’s magnetic field values in indoor areas, is investigated. In theory, the Earth’s magnetic field may change in time referred to as short-term and long-term deviations. Long-term deviations caused by solar flares, the axis of the earth, etc. which are and ignored in this study. Short-term deviations caused by the metal effects. We propose two different approaches: Device-Free and DeviceDepended to clear/reduce the noise from the measured signals. Deviations caused by mobility, by building materials such as iron, nickel, etc. in the indoor area are considered as noise. To reduce the noise, it is recommended to filter only measurements at the reference points, as independent from the measuring devices. According to our theory, the magnetic field signals at the reference points should not change during the day and the deviation in the signals could be within a certain range. To examine this theory, signal measurements are grouped according to the reference points. the signal measurements in each group (DB k=1,…, n ) are subjected to filtering independent from the measured device information shown in Fig 1. The signal map obtained after filtering (DBkfiltered) is used for the training of the positioning model.
The Device-Depended approach recognizes that the noise source in the signals originates from the digital sensors of the devices which are used for measurements. The digital sensors are developed with four different approaches. Hence, the signal filtering process should be designed according to the device. The signal map (Fig. 2.) is divided into groups k=1,…, n). Each subgroup data is grouped again according to the device information. The measurements of each device are filtered, and the noise is reduced. Then, it is used to train the positioning models.
70% of the signal maps are used in the training of KNN and RF positioning models. The remaining 30% of the dataset is used to test the models. Accuracy rates of the models are calculated according to the estimations on the test data. Accuracy is calculated by finding the percentage ratio of the correct estimates in the submitted test data size to the total test data size. In a well-trained model, high accuracy is expected. Two different test scenarios (Fig. 3) are proposed to test the models that are trained in this study.
In Scenario 1, user can measure magnetic field signals of the Earth and the measurements are sent to the filtering unit. This filtering unit may be located on a remote server or integrated into a mobile device. The related unit reduces noise according to the filtering approach. Noise-free signals are transmitted to the positioning module. The positioning module is a unit where the user's position is determined. The specified position information is transmitted to building supervisor, user or to another application. In this scenario, processing time for determining the user's location is represented as ∆t. (5) ∆t is depended on filtering processing time and classification time (Eq. 6). In Scenario 2, user measures the magnetic field signals with his/her mobile device. The measured data is transmitted to the positioning module without any filtering process. The positioning module detects the position of the user and delivers to the relevant units.
Delays in the network are ignored in both models. For the same positioning model, the absence of filtering time in Scenario 2 seems advantageous in terms of decreasing processing time. 4
In this study, RFKONDB is used which is obtained by the fingerprinting technique. Measured area of The RFKONDB is divided into 1.2 x 1.2 m2 squares and signals are captured from in the middle of related squares. Fingerprinting signal maps are divided into two parts as 30% for testing and 70% for training. Training data is used to train KNN and RF methods. Test data is used to test the accuracy of the trained models. While creating fingerprint signal maps, NaN values are not used in the training phase of the models. The number of neighbors in the KNN method is 1. In the RF method, the nodes of the tree are selected according to the Gini index.
The accuracy values obtained localization models by using the RFKON dataset with the Device-Free filtering approach are given in Table 1. Cross-validation accuracy rates obtained as 95% by using KNN and RF as classification methods.
In the Device-Free filtering approach, the accuracy rates of the models by using Kalman filter and SGL filter have decreased. The LOESS filter did not change the accuracy of the model but it is observed that Test Scenario 2 is more successful than the original situation. Test Scenario 2 reduces localization processing time because it contains unfiltered test data. Therefore, the LOESS filter is an advantageous filtering method.
Device-Depended approach advocates that the cause of the noise in the signals is due to the measurements of the digital sensors. Therefore, to observe the accuracy rates of the filters, 6 different devices are used in the RFKON dataset. The signal measurements of the devices are grouped and filtered according to the device id value. The KNN and RF positioning models are trained by using 70% of the signal map obtained after filtering. The test results are shown in Table 2. The device-depended approach increased the accuracy of models. The Kalman filter is the most successful noise removal filter in this approach. 98% accuracy is obtained in the KNN and RF method SGF showed similar performance results to the Kalman filter. A decrease in accuracy values is observed in Test Scenario 2 by using Kalman and SGF in preprocess phase. It is observed that the models trained after LOESS filter are more successful in Test Scenario 2. Especially the RF method is a very successful localization method with a 98% accuracy rate.
RF
Validation Results KNN 0.95
RF Kalman LOESS SGF
The highest accuracy in the proposed localization models trained with RFKONDB data set is obtained by using SGF and LOESS filters with the Device-Free approach. Fig. 4 shows the 3D graph of the original measurements in the RFKONDB signal map and the 3D graph after the SGF and LOESS filters are used. Same colored dots symbolize measurements taken from the same reference points. The SGF and LOESS filters are provided similar results. When both filters are examined, it is seen that some measurements between the color groups are considered as noise and cleaned. 5
In this study; the filtering methods which are developed to clean the noise from the Earth’s magnetic F]ield signal values and their usage with positioning models have been investigated. For this reason, Kalman filter, LOESS and SGF filters, which are frequently used for cleaning the signals, are preferred. Fingerprinting technique is chosen as the method of indoor positioning. In the fingerprinting technique, signals are measured with different devices from the same reference points and these measurements are used to create signal maps. Two different approaches have been studied to reduce/clear the noise in the signals. The first approach; the noisy measurements in the signals are based on the digital sensor which is used to collect signals and therefore the measurements of each device are filtered independently, and the fingerprint map is generated. The relevant approach is called Device-Depended. The second approach, called Device-Free, is intended to clean the measurement noise from the environment, depending on the time taken at the reference point. All the signals in the fingerprint map at the same reference point are passed through the respective filters together. The main purpose of this approach is to collect signal measurements at a reference point in a narrower range. KNN and RF algorithms are used in the localization phase. In order to realize the proposed positioning models, RFKONDB signal map is used. Two different test scenarios have been proposed to compare the accuracy of the developed systems. Obtained test results are analyzed and the propositions related to these analyses are presented.
The accuracy of the models tested with 10k cross-validation and the validity of the test data and the models are calculated. Two different scenarios are developed for validation test data. In the first scenario, a filtered validation data set is presented to the models that are trained with the filtered data set. The aim is to present the data to be used in the determination of the user position by eliminating the noise with the same filter. In this scenario, the processing time of the data spend on filtering is disadvantageous. The second scenario includes the transmission of measurements from the user whose location is to be determined without any pre-processing. Although it is advantageous in terms of time, its positioning success decreases due to noisy measurements caused by digital sensors.
In the Device-Free approach, by using the RFKONDB data set, the success of the models proposed by the Kalman Filter has decreased. The SGF filter reacted similarly to the Kalman filter. When using the LOESS filter, the accuracy is not affected. We recommend using the LOESS and SGF filters for indoor positioning methods in the Device-Free approach.
In the Device-Depended approach, it has been observed that all filters increase positioning accuracy. The highest accuracy value found at 99%. With this approach, we recommend that the Kalman filter and the KNN positioning model to be created according to Scenario 1 for the positioning model performed by filtering.
In both approaches, it has been observed that validation data (Scenario 1) should also be filtered. It can be provided to send the data to the model after filtering on the mobile user's device where the location would be determined. It is also possible to filter the measurement data of the same user after being sent to a central server. Due to the development of technology; due to the usage of fast and powerful hardware the filtering process time has ceased to be a major problem. If the positioning model based on Scenario 2 would be created, it is observed that the most suitable filter would be LOESS filter. The accuracy rate obtained from models developed by using this filter is higher than the unfiltered validation test results. Especially the usage of the RF method with the LOESS filter is recommended.