1. Introduction
Nowadays, the Internet of Things (IoT) is becoming an extremely important communication model and has triggered numerous global research interests [
1]. IoT is an influential communication paradigm that links glove-wide entire network components and interconnects with each other [
1]. IoT is a state-of-the-art automation and analytics platform that uses incorporate technologies such as networking, sensing, big data, and artificial intelligence to realize an intelligent service sensing and monitoring in a wide-ranging system [
1,
2]. The sensor network, comprising several sensors, is used as a crucial component of the IoT framework to attain multi-point measurements for a given physical environment [
1,
2,
3]. From a various number of sensor network categories, Fiber Bragg Grating (FBG) sensors are widely used in complex environments because of their strong anti-electromagnetic interference, slight size, corrosion resistance, high-temperature tolerance, high sensitivity, and strong multiplexing capabilities [
3,
4,
5,
6,
7,
8]. Additionally, they displays massive bandwidths, have few broadcast victims, allow vast topographical coverage, and can be operated without electrical powering of local batteries to decrease the risk of sparking in flammable surroundings [
1,
3,
4]. Furthermore, the FBG sensor is an exceptional sensor element, which is appropriate to measure temperature, strain, pressure, slope, displacement, acceleration, load, and other paramters in IoT ecosystems as well as in various industrial applications for both motionless and active modes of operation [
3,
9,
10]. To realize and measure the above-listed parameters, several FBG sensors need to be multiplexed in one cable [
3,
4,
5,
6].
In the previous scheme, wavelength division multiplexing (WDM) [
11,
12] and time-division multiplexing (TDM) [
13,
14] techniques have been proposed to increase the number of FBG sensors that are multiplexed in a single fiber cable. However, in the WDM technique, each FBG has a limited operating range, which is not able to overlap. Thus, the number of FBG sensors multiplexed in WDM is limited through the operating spectra of the respective FBG sensor and the entire bandwidth light source [
11,
12]. Every FBG spectral signal in the WDM of the FBG sensor network is needed to have an independent operational region, which means that the adjacent spectrum does not overlap. In addition, the TDM method [
13,
14] is prone to delays, and it is very difficult to guarantee service to the desired receivers. This makes the number of FBG sensors in the sensor system very restricted.
Recently, to resolve the problems of WDM and TDM, researchers have proposed intensity wavelength division multiplexing (IWDM) techniques. IWDM is used to increase the number of multiplexed FBG sensors, and it has a low-complexity advantage [
3,
15,
16,
17,
18]. Several network topologies, such as tree, ring, mesh, star, and bus topologies, have been proposed in the FBG sensor network area [
3,
12,
13,
17] to implement different multiplexing techniques. However, each network topology has its drawbacks. Nevertheless, due to the simplicity of installation, cable cost reduction, and ease of fault troubleshooting, bus topology is the most widely available and preferred architecture for distributed sensor network applications [
4,
12,
13,
19] compared with the other topologies. Furthermore, because of its easily re-configurable structure, bus topology is a preferential structure over the other network topologies for incorporating many numbers of FBG sensors. IWDM-based FBG sensor network structures can be designed using one or a hybrid of one or several network topologies.
The IWDM approach allows the adjacent FBG reflection spectrum to overlap and crossover to each other without their identification being lost [
3,
15,
16,
17,
18]. This realized that IWDM is used to improve the multiplexing capacity and capability of the FBG sensor system and is capable of increasing the interrogated number of sensors twice more than conventional WDM [
15,
16,
17,
18]. However, in an IWDM-based sensor system, an unmeasurable gap is created when the bandwidth of one FBG is slightly greater than the neighboring FBG, which induces crosstalk between FBGs and increasing errors in Bragg wavelength measurements [
15,
16,
17]. Therefore, accurate determination of the FBG Bragg wavelength from the overlapping spectra is the primary task of the FBG sensor network. Recently, to boost the accuracy of Bragg wavelength measurements, several evolutionary algorithms such as genetic algorithms (GAs) [
20], differential evolution (DE) algorithms [
21,
22], particle swarm (PSO) algorithms [
23,
24], and distributed estimation algorithms (EDA) [
25] have been proposed. However, as the number of sensors increases, evolutionary algorithms typically need longer processing time to achieve better accuracy and have higher detection errors, which affect real-time monitoring of the Bragg wavelength measurement.
In successive works, to improve the overlap spectra detection speed and accuracy, several machine learning approaches, such as the least-squares support vector regression machine (LS-SVR) [
26] and the extreme learning machine (ELM) [
11,
17], have been proposed. However, these proposed algorithms have a lack of adaptability and a poor learning capacity for a greater number of FBG sensor systems, and during configuration changes, these algorithms also have reusability problems. In addition, in traditional machine learning techniques, features are manually extracted and consistency is limited, while deep learning ensures that data features and spectral features are automatically extracted via the multi-layer network structure [
3,
14,
18,
27,
28,
29,
30]. More recently, to enhance the accuracy of Bragg wavelength detection, several deep learning techniques such as long short-term memory (LSTM) [
27], convolutional neural networks (CNNs) [
28], back-propagation neural networks (BPNNs) [
29], a hybrid of discrete-wavelet transform (DWT) and LSTM [
18] techniques, and MLP [
14,
30] techniques have been proposed. Meanwhile, the speed of training time, accuracy of Bragg wavelength detection, and the reusability capability of the model are still problems. Therefore, the model’s reusability, accuracy, and computation time speed are still the limitations of these models as the number of FBG sensors increases.
The gated recurring unit (GRU) deep learning algorithm Bragg wavelength measurement technique fills the gap of those models’ learning efficiency, because GRU has a higher capacity to adapt data fluctuations and provide efficient generalization [
3,
31]. The GRU structure is relatively simple and its learning speed is fast, and it is currently one of the most effective methods studied for the spectral characteristics of overlapping FBGs [
3]. It is characterized by a high capacity of learning, and understanding of complex data, and an easy adaptability to dynamic and temporal changes in data [
3,
31].
In this paper, we proposed GRU-based Bragg wavelength detection techniques to solve IWDM’s crosstalk problem, to improve the Bragg wavelength detection accuracy, the computational time speed, and the multiplexing capability of the FBG sensor system. In general, the contributions of this paper are described as follows:
To design a reliable multi-point sensing bus structure FBG sensor network.
To increase the number of FBG sensors using IWDM techniques.
To demonstrate the promising GRU model for accurate Bragg wavelength detection of each FBG sensor from the overlapping spectra.
To confirm the preferentiality of the proposed GRU model in Bragg wavelength detection over other traditional and recently proposed machine learning models.
To verify the performance of the proposed FBG sensor network in terms of computational time, Bragg wavelength detection accuracy, and multiplexing capability as the number of FBG sensors in the network is increased.
The remaining portions of this paper are organized as follows: the operational principle and experimental setup is stated in
Section 2; the methodology part is discussed in
Section 3; and the results and discussion are presented in
Section 4. Finally, in
Section 5, the conclusion of this paper is presented.