Improving Indoor Localization Based on Artificial Neural Network Technology

Wireless networks are ubiquitous nowadays and hence provide a promising approach for indoor localization. Many algorithms have been proposed for exploiting wireless signals for localization purposes. Among the methods, ANN-based methods have attracted particular attention due to their robustness in complex signal environments. However, their accuracy is still degraded by multi-path effects, signal fluctuations, and so on. Accordingly, this study commences by examining the effects of fluctuations in the received signal strength indicator (RSSI) measurement on the accuracy of an ANN-based localization algorithm. This study list some strategies and illustrate by simulation exper-iment. Based on the investigation results, a systematic methodology is proposed for improving the localization performance by increasing the number of APs. The feasibility of the proposed method is demonstrated by means of numerical simulations


Introduction
Location-based service (LBS) technology has entered the mainstream as the location accuracy of the global positioning system (GPS) has improved to within 10 m.However, GPS signals have only a limited ability to penetrate building exteriors, and hence GPS-based localization methods perform poorly in indoor environments.Nonetheless, indoor localization is important for many applications, ranging from healthcare, to asset management, augmented reality, retail management, sales promotions, and so forth.As a result, enhanced localization services are urgently required.
Wi-Fi, Bluetooth, Infrared/Laser, Radio Frequency Identification (RFID), ZigBee and Ultra-wideband (UWB) have all attracted significant attention as enabling technologies for indoor localization [1].Among these techniques, Wi-Fi is regarded as a particularly promising solution due to the proliferation of wireless networks in both outdoor and indoor environments nowadays and the Wi-Fi capability implemented as standard on most handheld and portable devices (e.g., smartphones, tablets, watches, and so on).In typical Wi-Fi localization services, the position of the user (mobile device) is determined by applying a mathematical model (typically, triangulation) to the received signal strength indication (RSSI) measurement obtained from multiple access points (APs).However, in more recent years, the feasibility for implementing indoor localization based on fingerprinting technology and artificial neural networks (ANNs) (or some other form of neural network) has attracted growing attention.The training procedure performed in ANNs improves the robustness of the localization performance toward environmental effects compared to simple triangulation-based schemes.It has been shown that ANNs enable a localization error of as little as 2 m in ideal scenarios.[2] As a result, they vastly outperform conventional GPS methods.However, their accuracy is still adversely affected by RSSI signal fluctuations.While several studies have shown that this problem can be mitigated by adopting the channel state information (CSI) to build the fingerprint rather than the RSSI [3], further improvements in the ANN localization accuracy are still required to implement finer LBS solutions.
Accordingly, this research commences by examining the effects of RSSI signal fluctuations on the performance of an ANN-based localization scheme given the simple case of a single AP and a straight-line transmission path.A systematic methodology is then proposed for improving the localization performance by adding additional APs to the sensing field.The feasibility of the proposed approach is demonstrated by means of numerical simulations.Notably, the results provide a useful reference for improving not only the performance of ANN-based localization methods, but also that of conventional triangulation schemes.

Related Work
Many radio propagation models are available for distance prediction in wireless communications.These models are generally based on the signal-strength decay, and comprise the free space propagation model, the tworay ground model, and the log-distance model [4] The models perform well in outdoor, uncluttered environments, and are thus widely deployed in traditional wireless sensor networks (WSNs).However, the authors in [2] showed that the RSSI signal fluctuations produced in general indoor environments by scattering and other physical effects severely degrade the accuracy of the distance estimates in WSNs.Unfortunately, it is most common problem in every wireless indoor location system research.Various studies have suggested exploiting the channel state information (CSI) rather than the RSSI to construct the fingerprint [5].Generally speaking, the CSI provides a more robust estimation ability than the RSSI [3].However, the CSI has the form of a high-dimensional vector, and hence some form of data preprocessing, e.g., data compression, is required before it can be used for estimation purposes.As a result, the complexities of the localization algorithm and associated hardware are inevitably increased.Additionally, the extremely environment sensitive of CSI is still a problem.Moreover, the minimum average localization error is still around 2.4 m even for simple straight-line observations [5].Consequently, further improvements in the localization preformation of indoor distance estimation schemes are still required.

Preliminaries
In the RSSI-based localization scheme proposed in [6], the robustness of the distance estimates to signal intensity fluctuations was improved by using a base station (BS) as a reference point to detect the RSSI of three other anchor APs.In particular, the BS received the RSSI values from any of the three anchors and these signals were used to compute a dynamic distance coefficient to support the localization process.The authors in [7] proposed an RSSI real-time correction method based on the particle swarm optimization -back propagation neural network (PSO-BPNN) RSSI-distance model presented in [8].In the proposed approach, a terminal was established to collect the RSSIs of the surrounding APs and the RSSI measurements were then adjusted intelligently in realtime using RSSI fluctuation data stored on a local server.However, BPNNs fail to properly converge when presented with nonlinear separable input data [9,10].Consequently, only a limited improvement in the distance estimation is obtained.

RSSI Observations
The authors in [2] examined the typical signal fluctuations observed in RSSI samples collected in indoor environments.The results showed that small variations in the RSSI signal occurred in samples collected at the same distance from the AP, while large variations occurred in samples collected at different distances.It was hence inferred that the RSSI strength does not decay uniformly with increasing distance.To further investigate this phenomenon, the present study commenced by investigating the RSSI decay at eight observation points arranged uniformly along a straight indoor corridor with a spacing of 5 m between them (see Figure 1(a)).To investigate RSSI decay in shorter distance, we also gathered data at eight observation shown in Figure1(b).For each observation point, 10 RSSI samples were collected at 1-s intervals over a 10-s period.
Figure 2 shows the average value of the RSSI obtained at each of the eight observation points in Figure 1(a).Figure 3 shows the variation of the 10 RSSI measurements collected at at each of the eight observation points in Figure 2. In general, the results presented in the two figures show that the RSSI varies almost randomly.For example, the RSSI values collected at the same distance from the AP fluctuate arbitrarily (see Fig. 3), while the RSSI value at a distance of 30 m from the AP is greater than that at 25 m (see Fig. 2), while that collected at 10 m is slightly higher than that collected at 5 m.In other words, the variation of the RSSI strength with the sensing distance differs markedly from that predicted in the ideal logarithmic decrement model [11][12].

RSSI Signal Fluctuations
Given the uncluttered environment considered in Figure 1, the RSSI variations shown in Figures 2 and 3 can be attributed most likely to the multi-path effect [5].In practice, the non-linear variation of the RSSI strength with an increasing sensing distance, and the fluctuations of the RSSI measurement at the same collection point, pose significant challenges in implementing RSSI-based localization schemes.For example, referring to Figure 2, an average RSSI value of -46.2 dB may map to a sensing distance of 30 m, but may also conceivably map to a distance of 15 m.Thus, even if a robust ANN scheme is employed for distance estimation purposes, the training model may still fail to converge if the RSSI distribution is not taken into account.Furthermore, for sensing distances of 25 ~ 35 m, the detected RSSI values vary widely (and non-linearly) in the range of -46.2 dB ~ -57.2 dB.Consequently, even a well-fitting model may fail to accurately predict the sensing distance from the AP.In general, the results confirm that a single input feature (i.e., a single RSSI measurement) is insufficient to perform straight line distance location estimation in even simple, uncluttered environments.

RSSI Decay Simulators
To facilitate the present investigations, a sample generator was constructed to simulate the RSSI decay behavior based on the actual RSSI measurements obtained in the sensing field.Figure 4 shows five sets of RSSI-distance data collected in the sensor field shown in Figure 1(b) on five different dates, respectively.In general, the results suggest that two different RSSI simulators are required to properly model the variations of the RSSI strength at different distances from the AP and at the same distance from the AP, respectively.For simplicity, in constructing the first model, the complexity of the environmental effects on the RSSI strength was modeled using a Gaussian random distribution with a mean and variance determined in accordance with the measurements obtained in the real environment.Figures 5~7 show 5

Artificial Neural Network
A multi-layer neural network [13,14,15,16] was constructed to predict the sensing distance from the AP based on the measured RSSI value.The network contained one hidden layers; each with 20 nodes and a tangent hyperbolic activation function.The model was trained using part of the 720 samples shown in Figure 9 based on a nonlinear polynomial fitting function, and was designed to output the sensing distance given an input value of the RSSI.The details of the neural network structure are shown in Table 1   The supervised learning process for the ANN comprised three steps, namely (1) data preprocessing, (2) training and (3) testing.In the data preprocessing step, the RSSI data produced by the second sample generator described above was split into a training set and a testing set in the ratio of 80% : 20%.The samples in the testing set were manually labeled with the sensing distance and encoded into 8 classes accordingly using the one-hot encoding method.The training process was conducted for a maximum of 1000 epochs with a termination criterion of an accuracy of 0.75 (75%).It was found that the training process failed to converge to the required accuracy level even after 1000 epochs.Nonetheless, the testing process was performed anyway under the consideration that the testing outcome might yield useful insights into possible improvement strategies.Table 2 shows the confusion matrix for the testing results.It is seen that the test data associated with the fifth class (15 m) are incorrectly predicted as belonging to the third class (10 m), while the testing data associated with the sixth class (17.5 m) are similarly incorrectly predicted as belonging to the fourth class (12.5 m).

Table 2. Confusion matrix for testing data
In general, the training and testing results presented above indicate that the Multi-Layer Neural Network fails to properly converge when using a single RSSI input to predict the sensing distance if the RSSI measurements are subject to fluctuations.Notably, this situation occurs not only in the simple straight-line distance estimation problem considered here, but also in the two-dimensional estimation problem involving three APs.This kind of situation is common in indoor location system that predicting location flipping between several points losing the accuracy.[17] 2, the Multi-Level Neural Network fails to properly estimate the sensing distance from the AP if the RSSI signals contain fluctuations and environmental factors cause the RSSI signal strength to decay non-linearly over the sensing distance.It can be further inferred that, under such conditions, a single feature space (i.e., a single RSSI input) is insufficient to estimate the straight line distance from the AP.Accordingly, this study proposes the use of an enhanced localization scheme based on two APs (AP 1 and AP 2 ) separated by a distance of 20 m and having the same RSSI distributions. Figure 11 shows the simulated RSSI distributions of the two APs over a sensing distance of 20 m in figure 10 .Note that for simplicity, the two distributions are simply mirror images of one another.Figure 12 shows the detailed RSSI fluctuations at each sampling point in the two cases.Finally, Figure 13 plots the RSSI measurements of AP 2 (on the Yaxis) against those of AP 1 (on the X-axis).The plotted line of each point shows the propinquity of the two RSSI data distributions as the observation point moves between AP 1 and AP 2 .In other words, Figure 13provides an insight into the ability of the combined feature space to identify the sensing distance.In this particular case, the absence of any crossing-points in the plotted line implies that the feature space is sufficient to perform accurate sensing distance estimation.To verify this inference, we take 800 RSSI sample sets (100 sample sets in each 8 observation points) as showed in Figure 12.The two-dimensional RSSI data wereseparated into training/testing sets with 80 : 20 ratio ( 640 training sets and 160 testing sets) and training sets were input to an ANNwith the same structure as that described above.The network was found to converge with an accuracy approaching 99.58% after 200 training epochs.Having trained the model, 160 RSSI samples were used for testing purposes.Table 3 shows the resulting confusion matrix.12 The results presented in Table 3 confirm the efficacy of adding a second AP to the sensing field to improve the localization accuracy.However, it is still necessary to confirm whether the addition of this AP improves the localization accuracy for all possible distributions of the RSSI-distance profile.Accordingly, the RSSI generators were used to produce 6 random sets of RSSI data for AP 1 and AP 2 , as shown in Figure 14(a)~(f).In each depict, the circled regions show the RSSI discontinuity points which cause confusion when implementing the ANN using a single input feature space.Figure 15(a)~(f) shows the corresponding AP 2 vs. AP 1 RSSI signal plots.The results show that the addition of the second AP resolves the confusion problem in three out of the six cases.In other words, while the deployment of an additional AP is beneficial in improving the average accuracy of the localization process, it cannot guarantee a precise localization outcome in every case.For illustration purposes, consider the RSSI plot shown in Figure 14(e).Inputting the two-dimensional feature training data into the ANN, the localization accuracy is found to be just 0.88 (88%).Furthermore, the confusion matrix shows that the Class 5 outcome (corresponding to a sensing distance of 15 m) and Class 2 outcome (corresponding to a sensing distance of 7.5 m) are confused.This result is reasonable since the RSSI data at 15 m (AP 1 : -46 dB, AP 2 : -46 dB) are identical to those at 7.5 m (AP 1 : -46 dB, AP 2 : -46 dB ).Since the distance between the two confusion points is just 7.5 m, the 0.12 error rate (i.e., 1.0-0.88=0.12)prevents the ANN from discriminating between them.It is possible that this problem can be overcome simply by adding a third AP.However, this inevitably introduces a cost-accuracy tradeoff.Consequently, an alternative method for improving the localization reliability is required, as described in the following section.

Ap2 Placement Adjustment
In practice, the distance between AP 1 and AP 2 is flexible.Hence, the potential exists to solve the confusion problem described above by adjusting the distance between the two APs.Accordingly, the distance between AP 1 and AP 2 was increased from 20 m to 25 m.The corresponding RSSI distributions are shown in Figures 16  and 17.It is seen in Figure 17 that a continuous RSSI profile with no crossing points is obtained.In other words, the region of confusion is eliminated.The trained model converged to an accuracy of 0.95 (95%) after 200 epochs.The confusion matrix for the test results is shown in Table 5 and confirms the improvement in the estimation results.In other words, the results confirm that the localization performance can be improved without any increase in the hardware cost simply by adjusting the distance between the APs.

Conclusion
This paper has shown that when using a single AP, the localization performance of RSSI-based methods is seriously degraded due to a non-linear decay of the RSSI signal strength with an increasing sensing distance and fluctuations of the RSSI signal at the same observation point.A similar problem may occurs even when using 3 APs to define a two-dimensional coordinate location.Notably, the deployment of additional APs in an attempt to improve the localization performance also introduces an inevitable tradeoff between the localization performance and the hardware cost.Accordingly, this study has proposed a low-cost solution for improving the localization performance in indoor environments using a minimal number (two) of APs.The detailed steps in the proposed approach can be summarized as follows: 1.
Collect historicalreal-world RSSI data in the target sensing field and analyze the RSSI distribution (mean, variance and standard deviation).

2.
Plot the RSSI data of two APs against one another in the X-Y coordinate space and identify any possible regions of confusion (as indicated by crossing-points in the AP 2 vs AP 1 RSSI profile).

3.
Check if the confusion problem can be resolved by adjusting the distance between the two APs. 4.
As a last resort, deploy an additional AP within the sensing field.The feasibility of the proposed methodology has been demonstrated by means of numerical simulations.This study has used a Multi-Layer Neural Network for estimation purposes.However, the results regarding the feasibility of a two-dimensional feature space for fingerprinting purposes are also applicable to a k-NN localization algorithm.

Figure 2 .Figure 3 .
Figure 2. Average RSSI decay over distance of 35 m from AP in Figure 1(a) , 50 and 100 sets of simulated RSSI data obtained at distances of 5 ~ 35 m from the single APin Fig 1 (a).

Figure 4 .Figure 5 .
Figure 4. Experimental RSSI decay over distance of 20 m from AP in Figure 1(b)

Figure 6 .
Figure 6.Fifty sets of simulated RSSI decay data over distance of 35 m from AP in Figure 5

Figure 7 .
Figure 7. 100 sets of simulated RSSI decay data over distance of 35 m from AP in Figure 5 The second RSSI generator was used to simulate the fluctuations in the RSSIstrength at the same distance from the AP.Based on an inspection of the standard deviations of the RSSI signals measured experimentally at each of the observation locations shown in Fig. 1 (b), the RSSI generator was constructed using a Gaussian distribution random output with a variance value of 0.5 dB. Figure 8 shows a typical RSSIdataset produced by the first sample generator.Figure 9 shows the 720 samples (i.e., 8 sampling locations with 90 Gaussian random points per location)produced by the second generator.From a close inspection, the samples generated at distances of 10 m and 15 m, respectively, have similar distributions, while those at 12.5 m and 17.5 m, respectively, also have similar distributions.In other words, despite prior training, the ANN may struggle to differentiate between samples obtained at 10 m and 15 m, respectively, or 12.5 m and 17.5 m, respectively.

Figure 8 .Figure 9 .
Figure 8. Simulated RSSI decay over distance of 20 m from AP

Chi
Han Chen and Rung Shiang Cheng EAI Endorsed Transactions on Internet of Things 10 2018 | Volume 4 | Issue 16 | e5

Improving
Indoor Localization Based on Artificial Neural Network Technology EAI Endorsed Transactions on Internet of Things 10 2018 | Volume 4 | Issue 16 | e5

Figure 14 (
Figure 14(a)~(f).Six sets of RSSI data and corresponding confusion regions

Chi
Han Chen and Rung Shiang Cheng EAI Endorsed Transactions on Internet of Things 10 2018 | Volume 4 | Issue 16 | e5

Improving
Indoor Localization Based on Artificial Neural Network Technology EAI Endorsed Transactions on Internet of Things 10 2018 | Volume 4 | Issue 16 | e5 9 Improving Indoor Localization Based on Artificial Neural Network Technology below.EAI Endorsed Transactions on Internet of Things 10 2018 | Volume 4 | Issue 16 | e5

Table 1 .
Structure of adopted Multi-Layer Neural Network

Table 3 .
Confusion matrix for testing results obtained using RSSI data in Figure

Table 5 .
Confusion matrixfor testing results obtained using RSSI plots in Figure16