Novel integrated matching algorithm using a deep learning algorithm for Wi-Fi fingerprint-positioning technique in the indoors-IoT era

Methodology

The first step of the proposed approach is collecting an extensive dataset of WPA RSSI in the offline phase. However, as explained in the next section, a preprocessing technique should be applied to construct a comprehensive dataset in a harsh environment, including RSSI multipath issues. Further, a Mean and Standard deviation based Augmentation TEchnique (MSATE) was proposed to enhance the quantity of RSSI samples for each class or RP. One of the methods used during data analysis to increase the quantity of RSSI data samples is known as data augmentation. Data augmentation involving the creation of freshly-produced synthetic RSSI data based on previously gathered datasets, or the addition of slightly modified copies of exisiting RSSI data, was used for data analysis. This serves as a regularizer and aids in reducing overfitting while training a DL model. However, in addition to the benefits, augmentation has certain limitations. The fundamental disadvantage of data augmentation is data bias, which means that the enhanced data distribution may deviate significantly from the original. Because of this data bias, conventional data augmentation strategies perform sub-optimally for time series data (Wen et al., 2021; Wong et al., 2016). Therefore, to determine the best data augmentation approach, additional research is required to generate new or synthetic data records with advanced applicability. The suggested approach, MSATE, produces a new N number of RSSI values for each WAP at a specified position (p) based on the mean and standard deviation computation for the pre-collected RSSI values at point (p). There will therefore be N new RSSI records at (p). The augmented records are expressed in Eq. (1): (1)${\displaystyle re{c}_{j,pAug}=\left(RS{S}_{1,pAug},RS{S}_{2,pAug},RS{S}_{3,pAug},\dots ..RS{S}_{M,pAug}\right)}$

where rec_j,p Aug is the jth augmented record at point p (for this study j is between 1 ≤ j ≤ 15), M is the number of the utilized WAPs (for this study M is equal to 14 WAPs), and RSS_M,p Aug is the augmented RSS values for each WAPs at point p, as it is expressed in Eq. (2): (2)${\displaystyle RS{S}_{i,Aug}=RS{S}_{m,i,p}\pm RS{S}_{std,i,\ast }+Rand}$

where the i value is between (1 ≤ i ≤ 14) since the number of the utilized WAPs is 14, RSS_m,i,p is the mean of ith WAPs RSS values at point p, RSS_std,i,∗ is the mean of the standard deviations (RSS_std,i,p) of all points for each WAPs, as presented in Eq. (3), and Rand is the randomized value between (−2 ≤ Rand ≤ +2). (3)${\displaystyle RS{S}_{std,i,\ast }=\frac{\sum _{p=1}^{no\text{\_}RPs}RS{S}_{std,i,p}}{no\text{\_}RPs}}$

where no_RPs is the number of RPs of the setting area, and RSS_std,i,p is the standard deviations of the RSS values of ith WAP at point p. The proposed augmentation procedure is depicted in the flowchart below, Fig. 2.

In the second step, the proposed solution applies the normalization process over the expanded dataset. Normalization is a method that is frequently used as a preprocessing mechanism in the preparation of machine learning data. The purpose of normalization is to maintain the RSSI values’ ranges while converting the values of a dataset’s numeric RSSI columns to a comparable scale. Not every dataset has to be normalized for machine learning and it is only necessary when features have separate ranges. Here, the normalization was done to reduce fluctuation problems, particularly those that arise when using a variety of platforms and devices to improve the performance of the LSTM (Maghdid et al., 2022). The RSSI values were then transformed to a comparable scale that was used to assist in LSTM training. The z-score method was used as the applied normalization method, which generates a data set with a mean of 0 and a standard deviation (std) of 1. This scaling approach works well when the data follows a Gaussian distribution (normal distribution) (Tabbakha et al., 2021). A snapshot of the normalization formula is expressed in Eq. (4). (4)${\displaystyle re{c}_{jNorm}=\left(RS{S}_{1Norm},RS{S}_{2Norm},RS{S}_{3Norm},\dots ..RS{S}_{MNorm}\right)}$

where rec_j Norm is the jth record in the dataset, M is the number of the utilized WAPs (for this study M is equal to 14 WAPs), and RSS_M Norm is the normalized RSS values for each WAPs at any records, as it is expressed in Eq. (5): (5)${\displaystyle RS{S}_{M,Norm}=\frac{RS{S}_M-RS{S}_{m,i}}{RS{S}_{std,i}}}$

where the i value is between (1 ≤ i ≤ 14) since the number of the utilized WAPs is 14, RSS_m,i is the mean of ith WAPs RSS values, and RSS_std,i is the standard deviation of the RSS values of ith WAP.

The third step uses the LSTM as a matching approach. It is used for position estimation after the RSSI data preparation. As a rule of thumb, the LSTM network is an extension of RNN utilized in deep learning and it can effectively train enormous structures of data. However, using LSTM without applying data normalization and augmentation will not provide accurate results, specifically for positioning estimation; this is due to the large number of RPs available. To this end, the LSTM technique was used after making some data preparations, including RSSI data augmentation (MSATE) and normalization (Norm). The data preparation addressed the challenge of too many RPs/classes that were available throughout the setting area. However, the aim of using the LSTM was to minimize the time and memory complexity of the WkNN. Figure 3 depicts the design of the LSTM and shows that the network will be fed the sequence from the input layer, which is a series of normalized RSSI values for each RP or fingerprint.

Afterward, the LSTM layer was connected to the fully-connected layer. The fully-connected layer was in charge of connecting its nodes to all other nodes in the LSTM layer. This layer is responsible for deep data feature analysis, with hidden layers automatically performing feature extraction. Each node in the fully connected layer can be used as an activation function across the RSSI vector on a linear combination of all the features learned by the previous layers. A set of weights were used as the parameters for the linear combination expression. These weights were applied to each layer (Abdul & Al-Talabani, 2022). The last fully-connected layer divided the node outputs into as many nodes as there were classes in the classification task, which, in this case, was 2,214 classes. As a result, in the last fully-connected layer, the number of classes equaled the output size parameter. Because the region included several classes/RPs, the Softmax layer was utilized to compute each class’s probability using the classification layer’s network inputs.

LSTM background

The LSTM network is an extension of RNN utilized in deep learning. In the network structure, the LSTM comprises a more complex repeating module. The primary distinction between regular RNN and LSTM is that the LSTM can detect long-term dependencies. It employs an architecture that can overcome the gradient vanishing problem. The LSTM is a deep learning algorithm that classifies and regresses time-series data like sounds and RSSI values by taking feature changes into account at each time step. The LSTM layers establish long-range connections between the sequence data’s RSSI values. The structure of a memory block LSTM is depicted in Fig. 4 to clarify the LSTM layer. Equations (6) and (7) calculate the cell and hidden states, respectively (Abdul, Al-Talabani & Ramadan, 2020). (6)${\displaystyle C_t=R\left(t\right)\odot C_{t-1}+I_t\odot \widetilde{C_t}}$ (7)${\displaystyle H_t=O_t\odot tanh\left(C_t\right)}$

where ⊙ denotes the vectors’ element-by-element multiplication and tanh() is the hyperbolic tangent function, which is the state activation function. The input gate and forget (remember) gate are denoted by I_t and $R\left(t\right)$. The cell candidate gate is presented by C_t, the candidate for cell state is indicated by $\widetilde{C_t}$, which is determined through Eq. (8), and the output gate is O_t in this study. (8)${\displaystyle \widetilde{C_t}=tanh\left(W_C\left[H_{t-1},X_t\right]+B_c\right)}$

where $\widetilde{C_t}$ indicates the weight for the input gate, X_t is the input at timestamp t and B_c

is a bias for the input gate.

Environment setting

The OMNeT++ simulator simulated the third floor of the Faculty of Engineering building from Koya University. This floor was chosen and constructed as an experimental setting in the proposed solution. The simulated area was approximately 56 ×77.6 square meters, and was comprised of four corridors, sixteen study halls, four computer laboratories, nine staff offices, three restrooms, and a large 13m ×19 m meeting hall. A total of 14 WAPs were deployed in the survey area, as depicted in Fig. 5.

Data collection

The entire region of the third floor was included for data collection. The area was divided into grids of 1 m ×1 m, dividing the region into 2,214 points, as illustrated in Fig. 6. At each sample point, RSSI data was recorded for one second. A total of 22,140 records were collected.

A set of parameters were configured for the overall WAPs and a smartphone was simulated in the network to deploy the simulation. Table 1 shows the primary network parameters used to construct the experimental environment in the OMNeT++ simulator, with a simulation period of 2,214 s.

Data preprocessing

Due to the restricted coverage range of the Wi-Fi, neighboring interfering devices, the existence of obstacles, and other factors, not all WAPs were detected among all of the RPs. The missing WAP RSSI values were required to have an RSSI value before data processing could take place. As a result, the missing RSSI values were fed with pre-known RSSI data from the related WAPs and at the same RPs. Additionally, the RSSI value dataset ranged from -17 dBm to -100 dBm and records with less than -100 dBm, or greater than -17 dBm, were ignored. After this process, ten RSSI records were kept at each RP. Consequently, there were 2,214 RPs and 22,140 preprocessed RSSI record RPs in the entire region.

Setting up the parameters of the LSTM and WkNN

The ML approach, LSTM, was used to train the model based on RSSI data obtained from the OMNeT++ environment and augmentation process. The trained approach was configured and tuned to provide improved positioning results.

The selection of hyperparameters was given special consideration here. The LSTM model was updated until the optimal hyperparameters were chosen and an optimized LSTM model was obtained. The Bayesian optimizer was then applied to determine the optimal hyperparameters and training choices for the LSTM model. The Bayesian optimizer is a practical approach to sweep hyperparameters in such experiments. In this research, the MatLab toolbox’s Experiment Manager was used to look for a combination of hyperparameters to improve the accuracy of the proposed solution after giving a range of values for each hyperparameter. Table 2 shows the set of parameters used to train LSTM.

The RSSI dataset was divided into two portions during the experimental setup to evaluate the proposed solution, including the LSTM and WkNN. The training used 80 percent of the whole fingerprint data, including the original and augmented data. Twenty percent of the remaining data was utilized during the testing phase. As indicated earlier, each RP had 25 RSSI recordings after preprocessing and augmentation. During data separation (80% for training, 20% for testing), every fifth record out of every five records was chosen for testing purposes, and the remaining four records were used for training.

The proposed Norm_MSATE_LSTMT solution was compared with the WkNN technique to evaluate its performance. When estimating the position (x, y) of a sample RSSI vector, we used the WkNN at a city block distance (Prabhakar & Rajaguru, 2016). An exhaustive search algorithm was used as the nearest neighbor search method. In this work, the value of k was chosen empirically and was initialized at 3. The weight values (w1, w2, w3) were added for the three selected RP coordinates {(x1, y1),(x2, y2),(x3, y3)}. The weight values were updated for each iteration when applying the WkNN algorithm, as expressed in Eq. (9). (9)${\displaystyle \frac{S=\sum _{i=1}^k\frac{1}{d_i}}{w_i=\frac{\frac{1}{d_i}}{S}}}$

where d_i is the distance between target RSSI sample vector at (x, y) coordinates and ith RSSI vector with (xi, yi) coordinates from the three selected RPs. The target coordinate (x, y) was then calculated by weighting the coordinate positions using Eq. (10). (10)${\displaystyle \left\{\begin{array}{c}x={\sum }_{i=1}^3{w}_i{x}_i\hfill \\ y={\sum }_{i=1}^3{w}_i{y}_i\hfill \end{array}.\right.}$