Open Access
Add to BibSonomy
Abstract
Fiber Bragg grating (FBG) sensors have been widely applied in various applications, especially for structural health monitoring. Low cost, wide range, and low error are necessary for an excellent performance FBG sensor signal demodulation system. Yet the improvement of performance is commonly accompanied by costly and complex systems. A high-performance, low-cost wavelength interrogation method for FBG sensors was introduced in this paper. The information from the FBG sensor signal was extracted by the array waveguide grating (AWG) and fed into the proposed cascaded neural network. The proposed network was constructed by cascading a convolutional neural network and a residual backpropagation neural network. We demonstrate that our network yields a vastly significant performance improvement in AWG-based wavelength interrogation over that given by other machine learning models and validate it in experiments. The proposed network cost-effectively widens the wavelength interrogation range of the demodulation system and optimizes the wavelength interrogation error substantially, also making the system scalable.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Fiber Bragg Grating (FBG) sensors have been attracted considerable study in recent years due to they have advantages of small size, lightweight, anti-electromagnetic interference, and so on. Especially as non-electrical sensors have been extensively applied for structural health monitoring (SHM) [1–3]. FBG sensor follows the principle that their central wavelength shifts in response to externally applied perturbation (e.g., temperature, strain, etc.).
Reflecting the external disturbance by the shift of the central wavelength is the main demodulation approach of FBG. Many FBG wavelength interrogation techniques are based on Mach–Zehnder interferometers [4], Michelson interferometry [5], CCD spectroscopy [6,7], tunable Fabry–Perot filters [8,9] and ring cavity fiber-laser scanning [10] have been reported. However, their implementation requires expensive equipment and complex system support. The edge filtering method is the most common mean used to interrogate the wavelength from an FBG sensor. It converts the FBG signal into an optical power signal using the linear filtering characteristics of the filter. Demodulation by measuring the shift in optical power. The volume optical filter suppresses the effect of light source output power fluctuations during demodulation and ensures fast response time in the case of linear output [11]. Wavelength Division Multiplexer (WDM) is capable of interrogating FBG sensors with high-resolution while enabling an all-fiber system to reduce optical loss [12]. Asymmetric Fabry-Perot interferometric (FPI) cavities are used to interrogate the central wavelength shift of FBG sensors with good linearity and a wide dynamic range [13]. A multi-band demodulation system that consists of multiple Coarse WDM (CWDM) was applied to demodulate weak FBG signal [14], which effectively improves demodulation sensitivity and reduces the effect of background noise. Furthermore, many novel approaches based on arrayed waveguide grating (AWG), long-period grating (LPG), light sources, have been proposed with the development of linear filters. Compared to AWG filters, LPG is extremely sensitive to external disturbances and prone to bending in the grid area in actual use, which affects measurement stability and accuracy. The light source is not so flexible that its measurement range is limited. The AWG-based filtering method offers high flexibility in the measurement range and has the advantages of fast response, high resolution, reusability, and easy packaging and integration. Two adjacent channels of AWG can be utilized simultaneously to demodulate the FBG sensing signal [15], the simultaneous employ of more channels provides the system with a wider wavelength demodulation range [16]. Optimization of the AWG demodulation algorithm and temperature compensation of the AWG-based demodulation system can boost interrogation precision and reusability [17]. Weng et al. [18] applied 8 channels SOI-based AWG to an interrogation micro-system for a phonic integrated FBG, and encourages it to circumvent the resolution loss under the large dynamic range. Most traditional approaches compromise on cost and system complexity in the pursuit of better performance. Yet the wavelength interrogation range and accuracy limitations are still prominent shortcomings.
With the rapid development of machine learning, especially artificial neural networks (ANNs). The cross-fertilization of ANN and FBG wavelength demodulation technologies has attracted significant attention [19–25]. Chan et al. [19] proposed to utilized an adaptive linear network called ADALINE Network to improve the wavelength detection accuracy of FBG sensors. Barino et al. [20] utilize ANN to process the spectral information of LPG sensors extracted by FBG arrays, which increase the flexibility of spectrum demodulation and reduce the cost of filtering. Basu et al. [21] employed two LPG to modulate the reflectance spectra of multiplexed FBGs, and their output light intensity is received by the detector and passed into the ANN as input for interrogating the FBG sensor. Ren et al. [25] improve the accuracy of the demodulation system by using cascaded neural network to match multiple FBG filters (FBGFs) for FBG wavelength interrogation. However, no study of ANN applied to AWG demodulation exists to date. Simultaneous use of multiple channel demodulation to increase the range of wavelength interrogation, which requires the same number of photoelectric probes (PDs) as the number of channels. This significantly increases the complexity of the system. In addition, crosstalk between channels can negatively impact interrogation accuracy.
To solve the aforementioned problems, the cascaded neural network technique was applied to the AWG-based FBG wavelength demodulation system in this paper. The first net of the network was based on a convolutional neural network (CNN) for mediating the relationship between demodulate channels and intensity. A network combining ResNet [26] and Backpropagation neural network [27] called Residual-Backpropagation neural network was introduced in tandem with CNN, to demodulate signals from FBG sensors with high accuracy. Experiments have proved that the cascaded neural network effectively widens the wavelength interrogation dynamic range of the AWG-based FBG demodulation system, and provides outstanding wavelength interrogation performance.
2. Methods
2.1 Principle of FBG demodulation based on AWG
The framework of FBG wavelength demodulation system based on AWG is shown in Fig. 1(a). The light emitted by the ASE source reaches the sensing FBG (Sen-FBG) through the circulator, and the beam that meets the Bragg condition will be reflected by the FBG and enter the AWG. The PD array was employed to receive the transmittance intensity from each channel and transmit it to the subsequent signal processing module. Its spectrum is shown in Fig. 1(b), two adjacent channels of the AWG are combined as a filter to demodulate the signal from the Sen-FBG, $CH_{m}$ and $CH_{m+1}$ are the spectrum of the two adjacent channels of the AWG. $P_{m}$ and $P_{m+1}$ are the transmission intensities of the $m^{th}$ and $m+1^{th}$ channels, respectively. $\lambda _{m}$, $\lambda _{m+1}$ and $\lambda _{FBG}$ are the center wavelength of the AWG channel and the Sen-FBG respectively. The transmitted intensity of these two adjacent channels varies with the axial stress on the Sen-FBG. The algorithm for obtaining the central wavelength shift of Sen-FBG can be formulated as Eq. (1).
Fig. 1. (a)AWG-based FBG demodulation system, Amp: Amplifier, DAQ: Data Acquisition. (b)Spectra of AWG dual-channel demodulation of FBG: the shaded part where the FBG spectrum and the AWG channel intersect is the transmission part of the AWG.
2.2 AWG-based FBG wavelength interrogation system using a cascaded neural network
The AWG-based wavelength interrogation system we utilized is shown in Fig. 2. The light generated by the ASE source enters the Sen-FBG through the circulator. Two translation stages were utilized to fix the Sen-FBG. One of them is equipped with an electric translation stage, which was used to continuously stretch the FBG in a constant step length to apply axial strain. The reflected light from the Sen-FBG is equally divided into two beams by the splitter. One of these beams will be measured directly by the optical spectrum analyzer (OSA), and the other enters the AWG. The optical power meter (OPM) following the AWG was utilized to detect the transmittance intensity of AWG channels. Three adjacent channels were utilized in this method for building two demodulation filters (A_F_1 and A_F_2). The transmission intensity of the three channels measured by OPM is recorded as $P_{CH_{m-1}}$, $P_{CH_{m}}$ and $P_{CH_{m+1}}$, their corresponding center wavelengths are $\lambda _{m-1}$, $\lambda _{m}$ and $\lambda _{m+1}$. The center wavelength of the Sen-FBG obtained by OSA is recorded as $\lambda _{FBG}$.
The processing flow of the proposed method was shown in Fig. 3. The original data is received by the proposed cascade neural neural after pre-processed, which consist of two networks connected in sequence, and the output of the Net1 will be directly input to the Net2 as its input. Net1 constructed by convolutional neural network (CNN) acts as an adaptive channel selector, to form demodulation filters by selecting appropriate channels for Sen-FBG from input data. When $\lambda _{FBG}$ is in the wavelength range of A_F_1 and satify the inequality $\lambda _{m-1} < \lambda _{FBG} < \lambda _{m}$ , the A_F_1 is selected as the filter for demodulating the Sen-FBG. When $\lambda _{m} < \lambda _{FBG} < \lambda _{m+1}$, the wavelength of Sen-FBG is in the wavelength range of the A_F_2 and it is used as the demodulation filter, as shown in Fig. 4. The other will automatically be deactivated when one of A_F_1 or A_F_2 was selected. The proposed Net2 was called Residual-Backpropagation neural network. It was applied to achieve high accuracy interrogation of the Sen-FBG center wavelength shift under the condition that Net1 was work well. The effective combination of Net1 and Net2 provides a wider wavelength interrogation range as well as lower interrogation errors. In the following, we will describe details of data pre-processing and specific architecture of Net1 and Net2.
Fig. 4. Principle of demodulation filter selection. A_F_1 consist of $CH_{m-1}$ and $CH_{m}$, $CH_{m}$ and $CH_{m+1}$ form A_F_2. The red and blue shaded areas represent the transmitted intensity of FBG at A_F_1 and A_F_2, respectively.
2.2.1 Data pre-processed
The original data must be pre-processed before they are fed into the proposed cascade neural network. Providing sufficient prior information by manually labeling the raw data is a prerequisite to ensure the accuracy of subsequent. According to the experimental situation, each set of data recorded by the AWG-based demodulation system corresponds to the filter selection, etc. is marked by us. One-Hot Encoding was applied to preprocess the filter selection results, A_F_1 and A_F_2 are coded as 10, 01 respectively. The transmitted intensity needs to be normalized since the data distribution of them is not uniform, which can improve the convergence speed, accuracy, and reliability of the model. Min-Max Normalization is utilized as the strategy to linearly transform the original data so that the results fall in [0, 1], the conversion function can be defined as Eq. (2).
where $P'_{CH_{k}}$ represents the normalized result of the transmission intensity of the $k^{th}$ channel.2.2.2 Net1: convolutional neural network (CNN)
The selection of demodulation filter is essential in this work due to it determines whether the subsequent wavelength interrogation can be carried out smoothly. It can be considered as a nonlinear classification problem since the three adjacent channels of the AWG can form two filters.
A CNN was designed as Net1 to achieve the correct demodulation filter selection. The difference between CNN and ordinary neural networks is that the former employs convolutional algorithms in at least one layer instead of multiplication, which gives the network powerful capabilities of feature extraction. As illustrated in Fig. 5, the proposed Net1 utilizes 1-D convolutional operation in entirely. Each convolutional layer with nonlinear activation layers is followed by another one with a normalization layer before the nonlinear activation, then a pooling layer is attached. The same substructure as above was replicated three times in the network. Thereafter, the output of the previous structure is compressed by a flatten layer, and then two fully connected layers with different nonlinear activation are connected to obtain the final output of the network. A straightforward and efficient regularization approach called batch normalization (BN) [28] was utilized as a normalization layer in the convolutional layer to obtain better performance. Rectified linear unit (ReLU) [29] was employed as an activation function in each convolutional layer and penultimate fully connected layer, to increase the generalization and representation capability of the network. And softmax [30] was placed in the output layer of the network as the activation function.
2.2.3 Net2: residual-backpropagation neural network
After Net1 makes a selection of the demodulation filters, the transmitted intensity of its corresponding channel was immediately received by Net2. Net2 was served as an interrogator deriving the central wavelength from an FBG signal, by learning the correlation between the intensity and the actual central wavelength of the Sen-FBG.
ResNet employs residual learning to solve the performance degradation problem of the network, which is centered on "shortcut connection" that allows the network layer to learn new features based on the input features, thus bringing better performance to the network. Inspired by ResNet, "shortcut connection" was incorporated into our backpropagation neural network. The product of this combination: Residual-Backpropagation Neural Network, regarded as Net2. The architecture of the proposed Net2 is shown in Fig. 6(a), which consists of an input layer, several hidden layers with different numbers of neurons, and an output layer. The shortcut connection combines the input and output of each hidden layer as the input of the next hidden layer, in a permutation and combination manner. This means that the input of each hidden layer contains not only the processed information of the previous layer but also a series of underlying information with different processing degrees. Net2 was permitted to learn more information to achieve lower error wavelength interrogation via shortcut connection. The principle of shortcut connection is illustrated in Fig. 6(b). It directly connects the input layer and the information processed by the stacked layer, which preserves the underlying information so that the performance will not decrease at least. The strategy of Net2 consists of a forward computation process and a reverse computation process, as shown in Fig. 6(c). Its forward computation process simply follows the network sequence, with data processed layer by layer from the input layer through hidden layers and finally fed into the output layer. The network will shift to reverse computation that returns the error signal along with the original propagation order and minimize the error by correcting the weight of each neuron if the output layer fails to obtain the desired output after a forward computation.
Fig. 6. (a)Architecture of Net2: each neuron is calculated as $y=f(wx+b)$, where $w$ is the weight, $b$ denote the bias, and $f$ is the activation function. (b) Principle of skip connection: the information of input can be defined as $X$, thee stacked layer that contains of Layer1, Layer2, and Layer3, is $F(X)$, the output is $X+F(x)$. (c)The arithmetic strategy of Net2: the black line and red line represent the forward propagation process and the backpropagation process respectively.
3. Experiments
3.1 Experimental setup
The experimental platform is shown in Fig. 7. The platform is mainly composed of FBG sensor demodulation system, ASE light source, and OSA (YOKOGAWA AQ6370D), the demodulation system is utilized to obtain the measured value of Sen-FBG central wavelength, and OSA is employed to obtain the actual central wavelength, as shown in Fig. 7(a). The FBG sensor demodulation system consists of an dual-channel optical power meter, three-axis translation stage, an AWG (40 channels, 125 GHz), an electric translation stage, a power supply, a stepper motor driver card, a circulator, a splitter, a PC for signal processing, and a Sen-FBG, as shown in Fig. 7(b). The grating used as Sen-FBG is based on SMF-28e, with a central wavelength of 1558 nm, 90$\%$ reflectivity, and 15 dB side lobe suppression ratio (SLSR). These FBGs have FWHM 0.2163 nm, 0.2165 nm, 0.2159 nm, 0.2145 nm, and 0.2174 nm, respectively. The experimental ambient temperature was maintained at 26$^\circ$C and all device in the system was preheated for thirty minutes before utilization.
Fig. 7. Experimental platform setup diagram:(a) Overall structure of the proposed system; (b) Detailed structure of FBG sensor demodulation system.
For the experimental setup, two three-axis translation stages were fixed on an optical plate, one of which was mounted an electric translation stage driven by a driver card. The Sen-FBG is fixed to a fiber optic fixture on a consistent height translation stage. In the experimental preparation stage, we stretched the Sen-FBG by controlling the electric translation stage until it was tightened ($\lambda _{FBG} \approx 1558.62 nm$). This wavelength was set as the reference wavelength as the initial state of the Sen-FBG in subsequent experiments. The center wavelength interval of two adjacent channels of AWG utilized in the experiment is $\sim$0.8 nm, after actual measurement, the FWHM of each channel is $\sim$0.456 nm. To ensure that the center wavelength of the Sen-FBG falls within the coverage of the three channels of AWG, channels 30, 31, and 32 were selected and formed two filters as demodulation filters. For convenience, the selected channels are noted as $CH_{1}$, $CH_{2}$ and $CH_{3}$ respectively, A_F_1 and A_F_2 are the names of the demodulation filters they were composed of.
During the experimental phase, the spacing between the two translation stages was adjusted to 27.5 cm, 32.0 cm, 36.0 cm, and 38.0 cm, respectively, to ensure sufficient data and feature diversity. Each FBG sensor prepared for the experiment was stretched and relaxed at different translation stage distances, and the fixed FBG sensor was stretched and relaxed by using an Arduino development board to control the electric translation stage with a step resolution of 20$\mathrm{\mu}$m. With the continuous effect of the electric translation stage on the Sen-FBG, the transmitted light intensity of each channel obtained by the optical power meter is acquired and saved by the PC, and the actual FBG center wavelength value from the OSA is recorded.
3.2 Training of cascade neural network and the results
All models in this work were built on the Tensorflow platform and the scikit-learn framework, and all algorithms were run on the Intel Core i7-9750H CPU (2.6GHz) and Geforce RTX2080 Max-Q GPU (8GB).
3.2.1 Configuration, training and results of Net1
A series of settings need to be specified before training Net1. The number of convolutional kernels per convolutional layer in the network is 32, 64, 128, 256, 512, 1024, respectively. The network was trained using Adam optimization that uses a dynamic learning rate algorithm to self-adaptively adjust the network weights, which encourages better convergence of the network to the optimal solution. The initial learning rate (LR) was preset to 0.001. The loss function is cross-entropy. The early stopping was used as a training trick in the Net1 training process. It is based on the principle that training is stopped when the performance on the validation set starts to decline, to avoid the problem of overfitting caused by continued training.
The original data was divided into training set, test set and validation set with the ratio of 75$\%$, 15$\%$ and 10$\%$ respectively after the data pre-processing module. $P_{CH_{1}}$, $P_{CH_{2}}$, $P_{CH_{3}}$ and $\Delta _{FBG}$ serve as the inputs of Net1, Filter and $P_{F}$ were utilized as the output of Net1.
To demonstrate the superiority of the performance of Net1, a series of machine learning classification algorithms are used as benchmark models for comparative experiments using identically processed datasets. These machine learning algorithms include Decision Tree, Discriminant Analyse, Logistic Regression, Naive Bayes, Support Vector Machine (SVM), Nearest Neighbor, and Ensemble. Neural network classification algorithms include Backprop Network, Probabilistic Neural Network, Learning Vector Quantization (LVQ) Neural Network, Elman Neural Network, and Self Organizing Maps (SOM) Neural Network were also incorporated as comparison models. 10-fold Cross-Validation was considered as validation strategy to select the optimal model in the full training cycle. The dataset under this strategy is randomly divided into ten parts, nine of which are utilized for training and another one for testing, and this process can be repeated ten times, each time with a different test set.
The accuracy of the filter selection in the experimental results is shown in Table 1. Net1 achieves the best result of 97.70$\%$ in the selection of filters. The reason is that, as a multi-feature binary classification task, the correlation between features and output categories needs to be mastered, and the convolution operation in Net1 empowers the network with powerful feature extraction capabilities. The confusion matrix drawn according to the actual classification of each model is presented in Fig. 8. Among the selected samples, the number of misclassified samples for Net1 totaled 2.3$\%$ of the total, and the probability of misclassifying A_F_1 into A_F_2 and A_F_2 into A_F_1 was similar (about 1.2$\%$ and 1.1$\%$). Based on the aforementioned problems, the following reasons can be analyzed:(1) The similar trend in the transmission intensity during the phase when the selection of demodulation filter was transformed, the classification results of these samples were confounded by Net1; (2) Insignificant changes in transmission intensity and insufficient differences in other features are not enough to influence the final classification decision of the Net1.
Fig. 9. Absolute error distribution of wavelength interrogation for the four training algorithms in the test dataset: ROE is the range of error, the frequency of the vertical coordinate implies the number of occurrences of the error of the algorithm in the test set.
Table 1. Comparison of the proposed Net1 and comparison algorithms in terms of filter selection accuracy.
3.2.2 Configuration, training and results of Net2
Four training algorithms were applied to train Net2 to the network with optimal performance: gradient descent with momentum (GDM), one-step segmentation (OSS), Powell-Beale conjugate gradient (CGB), and scaled conjugate gradient (SCG). The preprocessed dataset was used to train Net2 using each of the training algorithms described above. A series of network parameters were specified before training. The initial LR was set to 0.001. Tanh [31] was chosen as the activation function of each neuron, and the number of neurons in each hidden layer is 256, 256, 512 respectively. Data were randomly shuffled during network training. Minimizing mean squared error (MSE) that as the loss function is the training objective of the network. The early stopping was utilized as a training strategy to prevent the network from overfitting, and training stops when the validation error does not decrease for 20 consecutive epochs.
Wavelength interrogation error was defined as the difference between the output value from the network and the measured value of the OSA. We employ the absolute value of the wavelength error to evaluate the models that have been trained using each of the four algorithms. The error distribution of the four algorithms on the test dataset is shown in Fig. 9. Their absolute interrogation error is clustered in the range of 0-15 pm. The percentage of interrogation errors below 7.5 pm for the networks trained by SCG, CGB, GDM, and OSS algorithms are 76.1$\%$, 78.8$\%$, 49.9$\%$ and 67.3$\%$, respectively. This indicates that SCG and CGB algorithms are more suitable than other algorithms for the training of Net2. We performed a statistical of their absolute errors, as shown in Table 2, to select the optimal training algorithm for Net2. The minimum error of the network trained by the SCG algorithms is 0.003 pm among the error of the test set, which is the best among these test algorithms. And its average error is 5.98$\%$, 27.36$\%$ and 16.8$\%$ lower than that of CGB, GDM, and OSS respectively. In summary, SCG is considered by us to be the most suitable algorithm for training the proposed Net2.
Table 2. Statistical evaluation of the absolute wavelength interrogation error of Net2 trained by each of the four algorithms.
Machine learning regression algorithms such as Regression Decision Tree (Tree), Linear Regression, Support Vector Regression (SVR), and Gaussian Process Regression (GPR), and neural network-based algorithms such as Elman Neural Network, Generalization Regression Neural Network, and Backpropagation Neural Network were selected to perform the comparison experiments. The 10-fold cross-validation was adopted as the validation strategy for these models. Root Mean Squared Error (RMSE), R-Squared, MSE, and Mean Absolute Error (MAE) were employed to comprehensively evaluate the performance of these models, they can be formulated as Eq. (3)$\sim$ Eq. (6).
The results of the evaluation of baseline algorithms and Net2 using four different metrics are presented in Table 3. Net2 trained using the SCG algorithm achieves optimality in all metrics. The neural network correlation algorithm outperforms far than Tree, Linear Regression, SVR, and GPR on this dataset, which means that the neural network algorithm has more advantages in feature formulation and regression fitting. And the performance of the proposed Net2 is still superior compared to the selected neural network algorithm. The above statistical results demonstrate that Net2 achieves advanced results on the test dataset. The performance improvement of the demodulation system generated by Net2 is much higher than baseline models. This means that it is capable of performing highly accurate wavelength interrogation. Net2 has the highest R-Squared of 0.99, indicating that the model has outstanding regression and fitting performance. This additionally demonstrates the excellent generalization and robustness of Net2, ensuring the stability and accuracy of wavelength interrogation.
Table 3. Comparison of proposed Net2 and machine learning algorithms on error performance evaluation metrics RMSE, R-Squared, MSE, and MAE.
4. Discussion
The proposed cascaded neural network has been shown to significantly improve the performance of AWG-based FBG sensor demodulation system. With the proposed algorithm, we achieve three-channel ranges of interrogation wavelengths (at least 2.0 nm) in the system using a dual-channel optical power meter (Equivalent to two PD-Amp-DAQ sequences) and the system has some scalability. The trained network can be used to extend its dynamic range as well as wavelength interrogation accuracy. The wavelength interrogation range of FBG sensor from 31.2 nm to 40.0 nm can be obtained based on the proposed system, which makes our method valuable in a realistic application.
The cost-effectiveness we mention is firstly reflected by the fact that our proposed method allows more channels of AWG to be used simultaneously to extend the interrogation range of the system, without the need for a comparable number of channels of PDs, which significantly reduces the cost of the system and the complexity of signal processing. Furthermore, although the proposed algorithm is more complex than baseline models, it can reduce the complexity and cost of optical systems, and the signal processing part of most SHM applications is delivered to the could server to complete [32–35], without considering the hardware and computational cost of signal processing when monitoring.
The core of the FBG sensor demodulation system is the chosen demodulation device. We refer to relevant research and conducted a qualitative analysis for four aspects of their performance, cost, complexity, and reusability because their application scope is not consistent, as shown in Table 4. It can be seen that the stability and reusability of demodulation system are mainly determined by the stability of the device, such as EDF-based devices whose performance is severely degraded by RIA (radiation-induced effects) [36], FPI whose performance is uncontrollable and unstable, blazed grating which is used only for specific wavelengths resulting in their limitations, and LPG is too sensitive to external influences leading to unstable performance, etc. These factors make their corresponding demodulation systems weak in terms of stability and reusability even if they can obtain good performance. The AWG’s low crosstalk, low loss, low-temperature dependence, flat spectral response, high reliability, and low cost make it sufficiently stable and reusable when used as a demultiplexer in demodulation systems [37]. The proposed algorithm is based on a stable system to improve its performance and reduce the cost increase caused by the performance improvement. Although the performance of our proposed system cannot be optimal, its cost-effectiveness, low cost, low complexity, acceptable performance, stability, and reusability are noteworthy.
Table 4. Qualitative analysis of related FBG sensor demodulation systems.
To further improve the performance and robustness of the proposed network, it is straightforward to consider incorporating more features into the training dataset. This encourages the network to yield more accurate results via learning complex feature associations, both in terms of channels selection and wavelength interrogation.
The uninterrupted stretching of Sen-FBG may cause irreversible damage to their grids. The performance of the proposed network will be impacted by the massive wavelength shifts it causes. An additional data cleaning layer can be placed in front of the network to reduce the impact of abnormal data on network training.
5. Conclusion
This paper presents a novel approach to improve the performance of AWG-based FBG wavelength interrogation using a cascaded neural network, which consists of CNN and Residual-Backpropagation neural network in series. This network interrogates the central wavelength of the Sen-FBG by receiving the transmitted intensity from the channels and automatically selecting the channels based on the wavelength shift of the FBG. Multiple channels of AWG were used to extend the interrogation range of wavelengths but does not increase the complexity of the system. The experimental results show that the network has 97.70$\%$ accuracy in the selection of the demodulation channel, and its absolute average wavelength interrogation error reaches $\sim$5.672 pm. The demodulation performance is better than these selected baseline models by error evaluation metric analysis. The study extends the application of the neural network approach to AWG-based FBG wavelength interrogation, which provides a cost-effective, flexible customizable solution for remote wide-range wavelength interrogation.
Funding
National Natural Science Foundation of China (61762033, 61865005, 62175054); Natural Science Foundation of Hainan Province (2019CXTD400, 620RC554, 617079); Major Science and Technology Program of Haikou City (2021-002); Open Project Program of Wuhan National Laboratory for Optoelectronics (2020WNLOKF001); National Key Technology Support Program (2015BAH55F04, 2015BAH55F01); Major Science and Technology Project of Hainan Province (ZDKJ2016015); Scientific Research Starting Foundation of Hainan University (KYQD(ZR)1882).
Acknowledgments
CHEN S. C. thanks YAO F. F. and REN S. F. for their experimental assistance, WANG G. J. for critical comments on the manuscript, and HUANG M. X. for technical support.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. K. Loupos and A. Amditis, “Structural health monitoring fiber optic sensors,” in Fiber optic sensors, (Springer, 2017), pp. 185–206.
2. K. Bremer, M. Wollweber, F. Weigand, M. Rahlves, M. Kuhne, R. Helbig, and B. Roth, “Fibre optic sensors for the structural health monitoring of building structures,” Procedia Technol. 26, 524–529 (2016). [CrossRef]
3. H. Murayama, D. Wada, and H. Igawa, “Structural health monitoring by using fiber-optic distributed strain sensors with high spatial resolution,” Photonic Sens. 3(4), 355–376 (2013). [CrossRef]
4. L. Li, L. Xia, Z. Xie, and D. Liu, “All-fiber mach-zehnder interferometers for sensing applications,” Opt. Express 20(10), 11109–11120 (2012). [CrossRef]
5. C. da Silveira, J. Carvalho, P. Jorge, J. Costa, M. Giraldi, J. Santos, E. C. Júnior, D. Figueiredo, and O. Fraz ao, “Interferometric optical fiber inclinometer with dynamic fbg based interrogation,” in International Conference on Applications of Optics and Photonics, vol. 8001 (International Society for Optics and Photonics, 2011), p. 800145.
6. A. Ezbiri, A. Munoz, S. Kanellopoulos, and V. Handerek, “High resolution fibre bragg grating sensor demodulation using a diffraction grating spectrometer and ccd detection,” in IEE Colloquium on Optical Techniques for Smart Structures and Structural Mmonitoring (Digest No. 1997/033), (IET, 1997), pp. 1–5.
7. B. Das and V. Chandra, “Fiber-mzi-based fbg sensor interrogation: comparative study with a ccd spectrometer,” Appl. Opt. 55(29), 8287–8292 (2016). [CrossRef]
8. J.-j. Bai, L.-x. Wang, C.-y. Ran, and N.-j. Chen, “Demodulating and calibrating for fiber bragg grating based on Fabry-Perot filter,” Transducer and Microsystem Technologies 25, 10–14 (2006).
9. W. Zhu, J. Wang, J. Jiang, X. Liu, and T. Liu, “A high-precision wavelength demodulation method based on optical fiber fabry-perot tunable filter,” IEEE Access 6, 45983–45989 (2018). [CrossRef]
10. H. Wei and S. Krishnaswamy, “Comparative assessment of erbium fiber ring lasers and reflective soa linear lasers for fiber bragg grating dynamic strain sensing,” Appl. Opt. 56(13), 3867–3874 (2017). [CrossRef]
11. S. M. Melle, K. Liu, and R. Measures, “A passive wavelength demodulation system for guided-wave bragg grating sensors,” IEEE Photonics Technol. Lett. 4(5), 516–518 (1992). [CrossRef]
12. M. Davis and A. Kersey, “All-fibre bragg grating strain-sensor demodulation technique using a wavelength division coupler,” Electron. Lett. 30(1), 75–77 (1994). [CrossRef]
13. H. Chen and Y. Liang, “Analysis of the tunable asymmetric fiber fp cavity for fiber strain sensor edge-filter demodulation,” Photonic Sens. 4(4), 338–343 (2014). [CrossRef]
14. Y. Xiong, N. Ren, M. Wu, H. Liang, J. Han, and W. Yang, “Sensitivity-enhanced fbg demodulation system with multi-sideband filtering method,” Opt. Commun. 382, 246–252 (2017). [CrossRef]
15. H. Su and X. G. Huang, “A novel fiber bragg grating interrogating sensor system based on awg demultiplexing,” Opt. Commun. 275(1), 196–200 (2007). [CrossRef]
16. V. R. Marrazzo, M. Riccio, L. Maresca, A. Irace, and G. Breglio, “Wide range awg-based fbg interrogation system with improved sensitivity,” in 2019 15th Conference on Ph. D Research in Microelectronics and Electronics (PRIME), (IEEE, 2019), pp. 149–152.
17. H. Li, Y. Li, E. Li, X. Dong, Y. Bai, Y. Liu, and W. Zhou, “Temperature-insensitive arrayed waveguide grating demodulation technique for fiber bragg grating sensors,” Opt. Laser Technol. 51, 77–81 (2013). [CrossRef]
18. S. Weng, P. Yuan, W. Zhuang, D. Zhang, F. Luo, and L. Zhu, “Soi-based multi-channel awg with fiber bragg grating sensing interrogation system,” in Photonics, vol. 8 (Multidisciplinary Digital Publishing Institute, 2021), p. 214.
19. C. Chan, C. Shi, W. Jin, and D. Wang, “Improving the wavelength detection accuracy of fbg sensors using an adaline network,” IEEE Photonics Technol. Lett. 15(8), 1126–1128 (2003). [CrossRef]
20. F. O. Barino and A. B. dos Santos, “Lpg interrogator based on fbg array and artificial neural network,” IEEE Sens. J. 20(23), 14187–14194 (2020). [CrossRef]
21. M. Basu and S. Ghorai, “Sequential interrogation of multiple fbg sensors using lpg modulation and an artificial neural network,” Meas. Sci. Technol. 26(4), 045104 (2015). [CrossRef]
22. D. Li, H. Wang, and Z. Li, “Accurate and fast wavelength demodulation for fbg reflected spectrum using multilayer perceptron (mlp) neural network,” in 2020 12th International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), (2020), pp. 265–268.
23. X. Geng, S. Lu, M. Jiang, Q. Sui, S. Lv, H. Xiao, Y. Jia, and L. Jia, “Research on fbg-based cfrp structural damage identification using bp neural network,” Photonic Sens. 8(2), 168–175 (2018). [CrossRef]
24. B. Li, Z.-W. Tan, P. P. Shum, C. Wang, Y. Zheng, and L. jie Wong, “Dilated convolutional neural networks for fiber bragg grating signal demodulation,” Opt. Express 29(5), 7110–7123 (2021). [CrossRef]
25. N. Ren, Y. Yu, X. Jiang, and Y. Li, “Improved multi-grating filtering demodulation method based on cascading neural networks for fiber bragg grating sensor,” J. Lightwave Technol. 37(9), 2147–2154 (2019). [CrossRef]
26. Z. Wu, C. Shen, and A. Van Den Hengel, “Wider or deeper: Revisiting the resnet model for visual recognition,” Pattern Recognit. 90, 119–133 (2019). [CrossRef]
27. A. T. Goh, “Back-propagation neural networks for modeling complex systems,” Artif. Intelligence Eng. 9(3), 143–151 (1995). [CrossRef]
28. S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International conference on machine learning, (PMLR, 2015), pp. 448–456.
29. A. F. Agarap, “Deep learning using rectified linear units (relu),” arXiv preprint arXiv:1803.08375 (2018).
30. E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,” arXiv preprint arXiv:1611.01144 (2016).
31. E. Fan, “Extended tanh-function method and its applications to nonlinear equations,” Phys. Lett. A 277(4-5), 212–218 (2000). [CrossRef]
32. H. V. Dang, M. Tatipamula, and H. X. Nguyen, “Cloud-based digital twinning for structural health monitoring using deep learning,” IEEE Trans. Indust. Inform. (2021), early access.
33. H.-F. Chang and T.-K. Lin, “Real-time structural health monitoring system using internet of things and cloud computing,” arXiv preprint arXiv:1901.00670 (2019).
34. Y. Cao, S. Miraba, S. Rafiei, A. Ghabussi, F. Bokaei, S. Baharom, P. Haramipour, and H. Assilzadeh, “Economic application of structural health monitoring and internet of things in efficiency of building information modeling,” Smart Struct. Syst. 26(5), 559–573 (2020). [CrossRef]
35. F. Zonzini, C. Aguzzi, L. Gigli, L. Sciullo, N. Testoni, L. De Marchi, M. Di Felice, T. S. Cinotti, C. Mennuti, and A. Marzani, “Structural health monitoring and prognostic of industrial plants and civil structures: A sensor to cloud architecture,” IEEE Instrum. Meas. Mag. 23(9), 21–27 (2020). [CrossRef]
36. T.-S. Peng, Y.-W. Huang, L. A. Wang, R.-Y. Liu, and F.-I. Chou, “Photo-annealing effects on gamma radiation induced attenuation in erbium doped fibers and the sources using 532-nm and 976-nm lasers,” IEEE Trans. Nucl. Sci. 57(4), 2327–2331 (2010). [CrossRef]
37. S. E. Elrofai, “Review paper of array waveguide grating (AWG),” International Research Journal of Engineering and Technology (IRJET) 2, 1020–1029 (2015).
38. N. Xiang, Z. Li, X. Gui, F. Wang, Y. Hou, and H. Wang, “Research on a high-precision calibration method for tunable lasers,” Meas. Sci. Technol. 29(3), 035201 (2018). [CrossRef]
39. U. Tiwari, K. Thyagarajan, M. R. Shenoy, and S. C. Jain, “Edf-based edge-filter interrogation scheme for fbg sensors,” IEEE Sens. J. 13(4), 1315–1319 (2013). [CrossRef]
40. H. Kim and M. Song, “Linear fbg interrogation with a wavelength-swept fiber laser and a volume phase grating spectrometer,” in 21st International Conference on Optical Fiber Sensors, vol. 7753 (International Society for Optics and Photonics, 2011), p. 77537Y.
41. K. Zhou, A. Simpson, X. Chen, L. Zhang, and I. Bennion, “Fiber bragg grating sensor interrogation system using a ccd side detection method with superimposed blazed gratings,” IEEE Photonics Technol. Lett. 16(6), 1549–1551 (2004). [CrossRef]
42. S. Kim, S. Kim, J. Kwon, and B. Lee, “Fiber bragg grating strain sensor demodulator using a chirped fiber grating,” IEEE Photonics Technol. Lett. 13(8), 839–841 (2001). [CrossRef]
43. Y. Ni, X. Qin, H. Yao, Y. Guo, and X. Chen, “A high efficient method for demodulation of fbg sensors,” Opt. Fiber Technol. 38, 70–74 (2017). [CrossRef]
44. B. Jiang, J. Zhao, C. Qin, Z. Huang, and F. Fan, “An optimized strain demodulation method based on dynamic double matched fiber bragg grating filtering,” Opt. Lasers Eng. 49(3), 415–418 (2011). [CrossRef]
45. Y. Li, B. Lu, L. Ren, H. Chen, Y. Qiu, B. Mao, P. Zhou, C. Zhao, and X. Dong, “A highly precise fbg sensor interrogation system with wavemeter calibration,” Opt. Fiber Technol. 48, 207–212 (2019). [CrossRef]
46. H. J. Patrick, A. D. Kersey, J. Pedrazzani, and A. M. Vengsarkar, “Fiber bragg grating sensor demodulation system using in-fiber long period grating filters,” in Distributed and Multiplexed Fiber Optic Sensors VI, vol. 2838 (International Society for Optics and Photonics, 1996), pp. 60–65.
