Fig 1 - uploaded by Bijan Alizadeh
Content may be subject to copyright.
Source publication
In recent years, using distributed fiber optic sensors based on Brillouin scattering, for monitoring pipelines, tunnels and other constructional structures have gained huge popularity. However, these sensors have a low Signal to Noise Ratio, which usually increases their measurement error. To alleviate this issue, ensemble averaging is used which i...
Contexts in source publication
Context 1
... recent years, the ANN has found a widespread of applications in different fields such as pattern recognition, signal classification, and signal processing. The ANN architecture is inspired by the human brain functionality and it is basically a network of interconnected processing elements called neurons. Fig. 1 shows the architecture of a common ANN called multilayer perception which consists of one input layer, one output layer and n hidden layers. The functionality of a neuron can be described as ...
Context 2
... In j is the input, W i, j is the weights of the neurons and B i is the bias. The proposed architecture for the first hidden layer is shown in Fig. 11. As shown in Fig. 11, every output of the input layer is multiplied by the weights of the first hidden layer which are stored in a memory. Then the results of these multiplications are added to each other, and finally, the bias is added to them. Implementing the first hidden layer requires 21 memory with the depth of 41. This is ...
Context 3
... In j is the input, W i, j is the weights of the neurons and B i is the bias. The proposed architecture for the first hidden layer is shown in Fig. 11. As shown in Fig. 11, every output of the input layer is multiplied by the weights of the first hidden layer which are stored in a memory. Then the results of these multiplications are added to each other, and finally, the bias is added to them. Implementing the first hidden layer requires 21 memory with the depth of 41. This is because each neuron ...
Context 4
... are in this range. On the other hand, the sigmoid function has a symmetric structure which means by storing the values on the positive side, the information on the negative side is also accessible. So, since only the values between 0 and 3 are needed and the quantization step size is 2 −10 , a ROM with (3 − 0/2 −10 ) = 3072 word should be used. Fig. 12 shows the proposed architecture for the sigmoid function. First, the input enters the Pipeline_Reg, and then it is added by 3072 which is 3 × 2 10 . This is because the input address cannot be a negative address so the −3-0 part of the sigmoid function is moved to 0-3. After that, the input is compared to 0 and 6144 and will be limited ...
Context 5
... Output Layer: The proposed architecture for the output layer has been shown in Fig. 13. This layer takes 41 input samples, add them by 1 × 2 10 , divide them by gain, and then add the biases to them. It should be noted that there are no dividers in the implemented architecture, instead (1/Gain) has been calculated, and then the dividers are replaced by ...
Context 6
... with the state-of-the-art methods, 105 000 noisy spectrums have been created. After that, the noisy spectrums are given to the ANN in [9], an estimator based on cross correlation which was introduced in [5] and the proposed ANN in this paper with and without moving average, and then the average of absolute error of all the methods is calculated. Fig. 15 illustrates the accuracy comparison of the proposed ANN with other state-of-the-art methods. It should be noted that, in this paper, the BFS is detected by finding the position of the maximum amplitude of the ANN output and no estimation method such as fitting algorithm is used. Hence, it is expectable that an estimation method like ...
Similar publications
Estimation of the accurate value of Brillouin frequency shift (BFS) in a Brillouin optical correlation domain analysis (BOCDA) is challenging due to the contributions to the Brillouin gain spectrum (BGS) from the locations other than corresponding to the correlation between the pump and the probe. In this scenario, we demonstrate optimal post-proce...
Citations
... Cracking is a critical issue in structures as it can lead to their failures, and, consequently, put the safety of humans in danger [1][2][3]. Precise measurement of cracks is thus important for maintaining the durability of structures and ensuring protect environment for humans. To achieve this purpose, * Author to whom any correspondence should be addressed. ...
Optical fiber sensors based on the Brillouin optical time domain analysis (BOTDA) have good accuracy of crack opening displacement (COD) measurements. In this paper, we propose a method for COD quantification based on the area under the Brillouin frequency peaks induced by a crack. The study adopted a three-dimensional (3D) finite element model (FEM) to simulate the strain distribution within a segment of an optical fiber. The simulation results revealed that an increase in COD was associated with an increase in the Brillouin frequency peak area. The peak strain increased by 93 με when the COD increased from 30 μm to 110 μm. The numerical findings were proved experimentally by employing a BOTDA interrogator for distributed sensing of strains. Two cracks in a 15-m-long steel beam were detected with the smallest error of 13%. The COD was predicted from the areas under the crack-introduced strain peaks under varying loads of 97, 196, 294 and 392 N. The effect of different spatial resolutions (10, 20 and 50 cm) and intervals (1, 2.5 and 5 cm) on sensing performance was discussed. Compared to previous research, the 3D FEM not only accurately predicted the changes in distributed optical fibers with cracks but also simplified traditional theoretical analysis. For the first time, a method has been introduced to predict cracks by comparing the area under the Brillouin peaks. This approach not only enhanced linearity but also reduced errors. The proposed method can be easily implemented in engineering practice for multi-point crack sensing in civil infrastructure.
... A typical machine learning-empowered sensing system, however, involves data acquisition and data processing as two separate procedures [33], [34], [35], [36], [37], [38], [39], [40], [41]. First of all, it is necessary to observe and capture the raw BGS data of interest, which last for a finite time. ...
... In some other works regarding time-domain sensing techniques, machine learning has already been deployed to the system based on the hardware platform, and fast BFS extraction has been achieved [34], [35]. The signal processing is, however, still offline, and the measurement is not truly real time. ...
... In the real-world BOCDR, the shapes of BGSs usually no longer possess the similarity to the simulated signals, so the NN models trained with the ideal BGS datasets cannot provide correct and accurate BFS prediction for real BGSs. Different from most works regarding machine learning-based optical fiber sensors [27], [28], [31], [32], [34], [35], here we endeavor to train the NN models with real-world data, which requires collecting the training datasets from the experiment and labeling the BFSs for the datasets manually. ...
In this paper, a fiber-optic sensor based on Brillouin optical correlation-domain reflectometry (BOCDR) with a high repetition rate and real-time signal processing function is demonstrated with the assistance of swallow neural networks (NNs) constituted of no more than three hidden layers. In the proposed scheme, the signal processing time for every single Brillouin spectrum is compressed to less than the acquisition period of the spectrum by designing the structure of the NNs, on the basis of the knowledge of the relationship between the implementation time and the preliminary calculation count involved in the NNs. In the experiments with both simulated data and the real-world system, the performances of NNs with different sizes are studied from the perspectives of timing and accuracy. By training NN models with the data acquired from the experiments, real-time dynamic strain measurement is realized with a repetition rate of up to 20 kHz and a dynamic range of 4000 με. Different from other works regarding machine learning-empowered measurement acceleration in distributed optical fiber sensors with an offline signal processing phase, the method proposed in this paper enables consecutive monitoring of the parameters under test.
... More importantly, it is reported that the accuracy in BFS is said to be significantly higher than that of the curve fitting method at a low signal-to-noise ratio (SNR) [17]. Therefore, the method is used widely in BFS estimation [20][21][22][23]. ...
The cross-correlation method has a low computational burden and is less sensitive to noise, but the method may have a long spectra measurement time, especially for Brillouin spectra with an asymmetric frequency sweep span. To improve the real-time performance of spectra measurement, the influence of the frequency sweep span, sweep span deviation, and frequency step on the error in the estimated Brillouin frequency shift (BFS) is systematically investigated. Based on the results, the optimal sweep span and its influencing factors are investigated. The results reveal that when the frequency sweep span is not wide enough and there is a sweep span deviation, the BFS error will not decrease with a decreasing frequency step. The error decreases rapidly with an increasing frequency sweep span, and then it tends to a stable value. The linewidth and sweep span deviation have an important effect on the optimal sweep span. An estimation formula for the optimal sweep span is presented, and an improved cross-correlation method is proposed based on it. The proposed method is compared with existing classic cross-correlation methods. The results reveal that the proposed method can ensure high BFS accuracy and decrease measurement time.
... They are also adapted by researchers when the data dimension, image type, and content complexity are strictly restricted. The application scenarios are extensive for these networks, such as real-time high-level computer vision tasks [2][3][4][5], complex-valued image processing tasks [6,7], practical industries that take advantage of machine intelligence [8,9], and platforms that realize real-time visual processing [10]. However, the shallow structures and few layers with trainable parameters of these networks constrain their ability to simulate a non-linear function precisely, which bring big challenges to the researchers in a broad field. ...
Abstract Lightweight or mobile neural networks used for real‐time computer vision tasks contain fewer parameters than normal networks, which lead to a constrained performance. Herein, a novel activation function named as Tanh Exponential Activation Function (TanhExp) is proposed which can improve the performance for these networks on image classification task significantly. The definition of TanhExp is f(x) = x tanh(ex). The simplicity, efficiency, and robustness of TanhExp on various datasets and network models is demonstrated and TanhExp outperforms its counterparts in both convergence speed and accuracy. Its behaviour also remains stable even with noise added and dataset altered. It is shown that without increasing the size of the network, the capacity of lightweight neural networks can be enhanced by TanhExp with only a few training epochs and no extra parameters added.
... In order to further improve the real-time performance of the BOTDA system, parallel and pipeline mechanisms of FPGA can be used to reduce the time consumption of data processing [19][20][21]. The hardware architecture of the proposed method implemented on an FPGA is shown in Figure 3, consisting of three parts: (1) data caching module, (2) moving NCC module, and (3) result output module. ...
To address the issues of high time consumption of frame synchronization involved in a scanning-free Brillouin optical time-domain analysis (SF-BOTDA) system, a fast frame synchronization algorithm based on incremental updating was proposed. In comparison to the standard frame synchronization algorithm, the proposed one significantly reduced the processing time required for the BOTDA system frame synchronization by about 98%. In addition, to further accelerate the real-time performance of frame synchronization, a field programmable gate array (FPGA) hardware implementation architecture based on parallel processing and pipelining mechanisms was also proposed. Compared with the software implementation, it further raised the processing speed by 13.41 times. The proposed approach could lay a foundation for the BOTDA system in the field with the associated high real-time requirements.
To address the challenge of deploying Convolutional Neural Networks (CNNs) on edge devices with limited resources, this paper presents an effective 4-bit quantization scheme for CNN and proposes a DSP-free multiplier solution for deploying quantized neural networks on Field-Programmable Gate Array (FPGA) devices. Specifically, we first introduce a Threshold-Aware Quantization (TAQ) method with a mixed rounding strategy to compress the scale of the model while maintaining the accuracy of the original full-precision model. Experimental results demonstrate that the proposed quantization method retains a high classification accuracy for 4-bit quantized CNN models. Additionally, we propose a Compact Lookup-based Multiplier (CLM) design that replaces numerical multiplication with a lookup table of precomputed 4-bit multiplication results, leveraging LUT6 resources instead of scarce DSP blocks to improve the scalability of FPGA to implement multiplication-intensive CNN algorithms. The proposed 4-bit CLM only consumes 13 LUT6 resources, surpassing existing LUT-based multipliers in terms of resource consumption. The proposed CNN quantization and CLM multiplier scheme effectively save FPGA resource consumption for FPGA implementation on image classification tasks, providing strong support for deep learning algorithms in unmanned systems, industrial inspection, and other relevant vision and measurement scenarios running on DSP-constrained edge devices.
Transcendental functions cannot be expressed algebraically, which brings a big challenge to efficient and accurate approximation. Lookup table (LUT) and piecewise fitting are common traditional methods. However, they either trade approximate accuracy for computation and storage or require unaffordable resources when the expected accuracy is high. In this paper, we have developed a high-precision approximation method, which is error-controllable and resource-efficient. The method originally divides a transcendental function into two parts based on their slope. The steep slope part is approximated by the method of LUT with the interpolation, while the gentle slope part is approximated by the range-addressable lookup table (RALUT) algorithm. The boundary of two parts can be adjusted adaptively according to the expected accuracy. Moreover, we analyzed the error source of our method in detail, and proposed an optimal selection method for table resolution and data bit-width. The proposed algorithm is verified on an actual FPGA board and the results show that the proposed method’s error can achieve arbitrarily low. Compared to other methods, the proposed algorithm has a more stable increase in resource consumption as the required accuracy grows, which consumes fewer hardware resources especially at the middle accuracy, with at least 30% LUT slices saved.
