A Dual-Modality Deep Learning Framework with Received Signal Strength Indicator–Fine Timing Measurement Fusion for Accurate Wi-Fi Indoor Localization

The rapid development of indoor environments that include intelligent buildings, warehouses, and healthcare is increasing the need to have accurate and reliable indoor positioning systems [1, 2]. Many applications that are context-dependent, such as indoor navigation, asset tracking, and patient monitoring simply cannot work without precise spatial data [1]. Unlike in outdoor environments, where the Global Positioning System (GPS) technology normally exhibits a satisfactory level of accuracy, indoor environments pose greater challenges, as signals are attenuated, multipath-propagated, and may be affected by walls, furniture, and human movement [3-6]. These challenges are further compounded by the common usage of Internet of Things (IoT) technologies [4]. Complex indoor environments consist of an increasing number of dense networks of interconnected devices, which create large-scale heterogeneous wireless data streams [7]. As a result, indoor localization systems are expected to be not only accurate, but also scalable and robust in noisy and dynamically changing signal conditions [7]. Standard positioning approaches developed for idealized environments remain insufficient to address these specifications [7]. Classical methods such as triangulation and trilateration determine target location by measuring distance or angle from anchor nodes. While these approaches are effective in line-of-sight conditions, their performance degrades substantially in complex indoor settings. The reliability of these techniques is therefore limited by environmental constraints and measurement uncertainty, making them unsuitable for many real-world indoor deployments [7, 8]. Conversely, Wi-Fi-based localization has become a practical choice because it is cost-effective and reliant on existing wireless local area network (WLAN) infrastructure [9]. Among Wi-Fi-based approaches, fingerprinting has gained considerable attention because it does not require explicit distance estimation. Instead, it exploits distinctive patterns in wireless signal measurements—most commonly Received Signal Strength Indicator (RSSI) values—collected from multiple access points [9-11]. This localization framework typically involves two phases [10-12]. In the offline phase, RSSI readings are recorded at predetermined locations to construct a spatial radio map [10-12]. During the online phase, real-time RSSI measurements are matched against this radio map to determine the user’s exact location [10-12]. Although it is practical, RSSI-based fingerprinting has a number of shortcomings. RSSI is very sensitive to changes in the environment, and in crowded indoor environments, it tends to be ambiguous [11]. New developments in wireless standards have offered Fine Timing Measurement (FTM), which provides better ranging information, using signal propagation time [13]. Although FTM provides better distance estimation, it is highly susceptible to noise and environmental variability when it is used alone [14]. These observations highlight the necessity of localization models that are capable of incorporating heterogeneous wireless measurements and at the same time are resilient to indoor noise and signal distortion. Instead of using the properties of one modality or just geometric considerations, the combination of complementary signal properties through learning-based approaches is a promising direction. Following this trend, the current study combines two complementary and inexpensive Wi-Fi measurements in a single learning model, which can be used in a commodity device with no extra infrastructure to enable an accurate indoor localization system.

The principal contributions presented in this paper are summarized as follows:

• Proposing a robust hybrid deep learning model combining RSSI and FTM signals from ESP32C3 devices to significantly enhance indoor positioning accuracy.
• Introducing an integrated lightweight architecture, which consists of an independent per-modality normalization followed by a hybrid dense–convolutional design, where the two heterogeneous signals are first balanced in scale and then fused by layers that learn their global and local dependencies, providing a simpler alternative to fusion methods based on separate correlation analysis.
• Developing a hierarchical neural network that leverages the fused RSSI and FTM data to capture the interdependency between global and local features through dense and convolutional layers to precisely localize in 2D indoor environments.

The remainder of this paper is organized as follows: A comprehensive review of existing work in the field of indoor localization and learning-based fingerprinting is provided in Section 2. The dataset description process and the construction of the dual-modality dataset are presented in Section 3. Section 4 presents the results and discussion. Section 5 presents the conclusion and future directions of the proposed model.

This section presents the dataset used in this paper, the studied problem formulation, and the proposed model for robust indoor localization based on RSSI and FTM readings.

3.1 Dataset description

The dataset utilized in this paper is the ESP32C3 Wi-Fi FTM RSSI dataset [20], which was acquired from ESP32C3 devices (Wi-Fi enabled). These devices—in addition to measuring RSSI—incorporate the capability to extract FTM readings. The dataset environment consists of eight fixed devices acting as access points (APs) to collect synchronized RSSI and FTM readings from a mobile device moved by an autonomous robot Turtlebot3, to ensure high-precision ground truth while simulating realistic user trajectories. Readings are samples associated with a reference position (x, y) that serve as multimodal position labels to support AI techniques for localization tasks. The smartphone was mounted at a height of 75 cm to mimic the typical elevation of a device in a user's pocket. The dataset features were collected in a setting of a multi-floor building, covering diverse indoor layouts to collect the dataset entries. Table 1 illustrates basic statistics of the collected features.

Table 1. Features statistics for all floors

The values in Figure 1 show a non-homogeneous distribution of values, which indicates difficulty in combining those values to assemble a larger dataset for training.

Although the values of RSSI and FTM have common values areas, those common readings do not reflect the same spatial locations. Consequently, the same APs readings at an instance of time do not indicate the same location at each floor. The modality represented in this dataset includes RSSI and FTM. RSSI readings resemble the strength of the signal received by the user’s equipment. The readings are recorded in decibel-milliwatts (dBm) and with ranges indicated in Table 1. On the other hand, FTM readings represent the round-trip timed distance features in meters. The benefit of FTM observations is that they act as assistant features to enhance localization due to their resilience against noise and other physical phenomena like multipath loss. In this dataset each row of readings (RSSI and FTM vectors) indicates a spatial 2D location (x, y).

(a) Received signal strength indicator (RSSI)

(b) Fine timing measurement (FTM)

Figure 1. Received signal strength indicator (RSSI) and fine timing measurement (FTM) values distribution

3.2 Problem formulation

The presented indoor localization task can be classified as a supervised regression problem. The goal is to estimate the localization of a user in a heterogeneous wireless environment. The dataset encoded two streams of readings from each access point, namely RSSI (r) and FTM (f) signals, where $r \in \mathbb{R}^{N_r}$ and $f \in \mathbb{R}^{N_f}$ represent RSSI and FTM streams respectively, while $N_r$ and $N_f$ indicate the number of access points in the current setting. The objective of the model is to approximate the nonlinear location mapping system.

$\mathcal{F}:(r, f) \rightarrow l$, where $l=(x, y) \in \mathbb{R}^2$          (1)

The system is designed to optimize location accuracy by lowering the mean squared error (MSE) between the estimated and the ground truth locations provided in the dataset.

3.3 Data preprocessing

As seen in Table 1, the wireless signals under consideration, RSSI and FTM, have extremely different numerical ranges. This means that if these heterogeneous data streams are ingested into the network architecture directly without some sort of calibration, it creates huge scaling mismatches. An independent normalization approach is essential to create a common numerical domain and to ensure a fair representation of features. If this preprocessing is not performed, the optimization process may end up with a significant convergence bias, which is that the model is biased towards the spectrum of features that has large absolute values. Accordingly, each feature set was normalized independently using min–max scaling, which rescales the values of each modality into a common [0, 1] range. Min–max scaling was used instead of z-score standardization as it maintains the boundedness of the signals and both modalities would be scaled to the same range, which would result in a balanced contribution to the fusion and prevent the larger-scale modality from dominating. This separation is for balancing the training phase and also for ensuring a fair contribution from both signal sources. Also, manual feature engineering was not performed in this study, so that the model can learn spatial representations directly from the raw received signals.

3.4 Proposed hybrid localization model

The proposed model is a parallel-input neural network architecture that processes the RSSI and FTM signal streams simultaneously, making sure that the characteristics of each signal stream are preserved as shown in Figure 2. Each heterogeneous input stream is independently normalized using min–max scaling to avoid the dominance of the larger scale and to ensure a fair integration of features. After this key calibration, the normalized streams are then combined together to form a coherent and homogeneous feature map. This fused representation is then fed to a dense neural encoder with a particular focus on learning global interdependencies between the signals. The dense encoder is first applied because there is no natural topological order between access points in the concatenated RSSI–FTM vector and a convolution applied directly to raw fused values would not have any meaning. The dense layers thus embed the fused features into a learned representation that captures the global cross-modal dependencies and then unfold it into a structured sequence that the 1D convolution extracts local correlations from. The encoder uses a Leaky ReLU activation function (Eq. (2)) that can effectively prevent dead neurons.

$f(x)=\max (a x, x)$          (2)

where, a is a small non-negative hyperparameter to set the direction of the function’s slope for negative input x. Two other techniques applied here: first, L2 regularization to improve generalization, and dropout to reduce overfitting. The resultant stream is fed into a 1D-CNN to learn localized dependencies across features. This convolution helps in assessing the exploitation of structured correlations among the fused features stream. The output regression layer acts as a spatial localization estimator for the input 2D location. The layout of the proposed model is illustrated in Figure 2, and Table 2 shows the parameters used in this model.

Table 2. Structure and parameters of RSSI–FTM localization model

3.5 Model training

For model training, the Adam optimizer was selected to update weights, while the MSE was employed as the loss function to guide the learning process and decrease the difference between true and estimated locations in Euclidean space as shown in Eq. (3).

$M S E=\frac{1}{n} \sum_1^n\left(L_i-\widehat{L}_i\right)^2$          (3)

where, n is the batch size, L represents the true 2D location, and $\widehat{L}$ is the 2D location estimated by the model.

For evaluating the model’s predictive capability, the data was partitioned into training and unseen samples. Table 3 illustrates the hyper-parameters used to train the model. These training settings were selected to obtain a balance between stability and generalization. The Adam optimizer was chosen due to its adaptive learning behaviour that converged reliably on regression tasks with the common default learning rate of 0.001. Batch size is 32, which is a compromise between gradient stability and memory costs. The dropout rate is set to 0.2 and the L2 weight is 1 × 10⁻⁴, which is a light regularization for the size of the dataset. Lastly, the maximum number of epochs is 750 and early stopping is used, which means that the training will automatically stop when the validation loss stops improving.

Table 3. Training parameters for the proposed model