Open Access
Add to BibSonomy
Abstract
A distributed optic fiber perimeter security system is proved to be an effective strategy for the security monitoring of some vital targets, such as power plants, power substations and telecommunication base stations. However, this method can hardly distinguish different categories of the intrusion behavior and is easily mis-triggered by different kinds of environmental interference. To distinguish different intrusion patterns and different interference events effectively, a vibration pattern recognition algorithm is proposed and demonstrated based on the merged Sagnac interferometer structure. The method consists of two parts: the pre-processing algorithm and the multi-layer perceptron neural networks (MLP-NNs). The pre-processing algorithm is applied to retrieve and extract the vibration signal from the captured source signal, and the MLP-NN is used to realize pattern recognition from each type of input. Typically, a high-dimensional vector group which contains hundreds of orders of vibration signal’s power frequency is obtained to cover as many signalized features as possible. Moreover, results of the experiment deployed on a 10 kilometer long perimeter fence in the transformer substation show that the proposed classification-based model achieves 97.6% classification accuracy in the test. Through multiple comparison tests, the proposed model gives a solid performance in the subsequent integrated evaluation to classify each intrusion pattern.
© 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
At present, distributed fiber-optic perimeter security system has been widely applied in vital targets monitoring such as security monitoring in airport, crude-oil transportation, power plant and border security [1,2]. Due to its anti-electromagnetic interference, high resolution, low power cost and distributed character, distributed fiber-optic perimeter security system has taken the place of the traditional perimeter security monitoring solutions [3].
Fiber-optic interferometer is the method for detection of phase sensitive events [4,5]. Based on photoelastic effect, the vibration events are captured by the distributed sensors reflected on the phase-shift of the coherent light. And thanks to highly sensitive performance, the monitoring length of fiber-optic sensor probably reach hundreds of kilometers [6].
Because of the long distance of the monitoring scale, the quality of the signal is likely interfered by the environmental noise around the sensor deployment through the monitoring object, such as optical communication links and oil pipelines. Therefore, the signal feature extraction is particularly vital to the intrusion behavior monitoring. Some optical structures and optimization algorithms have been proposed to optimize the interference caused by the environmental noise. Double-beam interferometric structure has been reported due to its stable sensitivities of dynamic measurement [7,8]. In those structures, signal features are obtained by a balanced-detection configuration, in which the signals from the two detectors are subtracted to have the phase difference induced by the disturbance signal. The possible configurations like unidirectional Mach-Zehnder interferometer (MZI), line-based Signac interferometer (SI) and combination of MZI and SI are deployed in engineering projects [9]. The issue is that fiber polarization is likely affected by measurement conditions, which limits the performance of the sensors, and it is hard to form sensing structure with a single freely extending fiber. Thus, the combination structure like merged MI-SI interferometers is presented [10]. The coupler using in Sagnac loop provides passive biasing of the interferometer which improve the SNR, and meanwhile the merged structure can be deployed in single fiber-core mode.
Based on the analysis of large numbers of engineering testing data, the feature of response between environmental interference and the real intrusion events are almost the same, such as the fiber sensor vibration caused by the passing vehicles, the waving movements generated by blowing-winds and striking actions by rains directly act on the fiber perimeter fence.
In long-distance security monitoring projects, in order to reduce the difficulty of manual inspection and classify the vibration behaviors to evaluate the intrusion risk, it is necessary to distinguish the real intrusion behaviors and the interference events. Therefore, composite monitoring method like linkage camera is applied to capture the trigger activities. But the monitoring range of the linkage camera is limited, which can only be deployed at some key nodes. Besides, the method requires manual confirmation in the final stage of the operation which increases the workload and risk of false positive alarming. Later, some researchers have proposed a number of intrusion pattern recognition methods which mainly focus on time-domain character analysis such as the analysis of short time energy feature and the fluctuation rate. This type of methods is able to distinguish some notable different modes, however, in other general cases, when the intrusion behavior is more similar to the interference in time domain, they performance with less accuracy and lower efficiency. Therefore, this type of methods can hardly be applied in practical monitoring projects.
To accurately classify the different patterns of the vibration behaviors and distinguish the real intrusion events, the vibration pattern recognition algorithm based on the merged Sagnac interferometer (SI) fiber-optic structure is proposed and demonstrated in this paper. From theoretical analysis and multiple experiments, we have found that compared to the time domain approaches mentioned above, the classification model based on the merged Sagnac interferometer (SI) fiber-optic structure is able to extract the power vectors from the vibration signal in frequency domain. Through the extracted power vectors, the high-dimensional vector group can be formed to cover feature information as much as possible and to achieve the accurate classification performance. Then, we investigate and test the model trained with multi-layer perceptron neural network (MLP-NN) along a 10 kilometers long perimeter fence in the transformer substation. The result has shown the high accuracy and efficiency in classification of the intrusion behaviors in this work. Since it can not only decrease the system complexity and the hardware cost like the linkage camera application, but also reduce false alarming ratio and the manual workload, thus making it more promising in security monitoring projects.
2. Related work
2.1 Merged MZI-SI interferometry structure
Compared with the conventional Sagnac interferometer, the structure used in this paper is an improved wide-bind Sagnac interferometer. The configuration of the structure, based on a 3×3 coupler, is shown in Fig. 1. A super-luminescent diode (SLD) is applied to produce light with a short coherence length. The light is coupled into a single-mode fiber and is then split into two light beams by the 3×3 single-mode fiber coupler (Coupler1). One light beam passes through the time-delay coil with a physical length of C experiencing a delay to a 2×2 single-mode fiber coupler (Coupler2), while the other light beam propagates through the direct leg to the Coupler2. The Faraday rotator mirror (FRM) reflects these lights back to Coupler2 subsequently. The interference occurs between lights when travelling around optical paths with equal length due to the short coherence length light source. Therefore, only lights through path 1 (1-5-8-9-10-11-10-9-7-6-4) and path 2 (1-4-6-7-9-10-11-10-9-8-5) can interfere at Coupler1. Modulated interference light which carries the location information of disturbance point (10) is captured and transformed by Photodetector_1 and Photodetector_2. Processed by data processing method, the position of the disturbance signal can be measured.
2.2 Artificial neural network (ANN)
Accurate and reliable behavior classification is an important part of composite perimeter security monitoring system. It is a challenge task since the features of response between environmental interference and real intrusion events are almost the same. Some kinds of environmental interference such as the fiber sensor vibration caused by the passing vehicles and the waving movements caused by blowing-wind contain similar character compare to the real intrusion events [11]. In the recent literature, several approaches to distinguish the intrusion behaviors and the interference have been proposed and evaluated [12,13].
In general, the existing intrusion behaviors classification approaches and strategies can be summarized into two major classes: the fast distinguish between intrusion and interference, the specific and accurate categorizing of intrusion behaviors. The fast distinguishing between intrusion events and interference is commonly applied in real-time security warning system [14]. In this application scenario, an accurate and fast intrusion alarming is required to help to form the decisions. Based on the analysis of large numbers of engineering testing data, the feature of response between environmental interference and the real intrusion events are almost the same, such as the fiber sensor vibration caused by the passing vehicles, the waving movements generated by blowing-winds and striking actions by rains directly act on the fiber perimeter fence.
Compared with the strategy of multiple layer neural networks, Support Vector Machine(SVM) is a fast classifier in binary classification problems. The basic idea of SVM is to find the optimal hyper plane in database of binary classification to build the classification model and classify the data based on OHP. SVM is a small sample learning method with a solid theoretical foundation. Different from the existing statistical methods, this method avoids the traditional process from induction to deduction, realizing efficient “transduction inference” from training samples to forecast samples, and greatly simplifies the usual classification and regression issues. In [15], short term energy and short term threshold rate are extracted and form feature vector based on source vibration signal of security system. In subsequent reports, more features have been extracted to realize compound classification in distributed sensor deployment system. Utilizing threshold rate, the number of alarming data frame and the dynamic range of the signal, Mahmoud. S et al. [16] proposes a fast classification model between intrusion and the weather influence. With the only best approximation, Radial Basis Function Neural Network(RBF-NN) is another promising method for intrusion behavior classification. Based on Empirical Mode Decomposition(EMD), Li et al. [17] presents the classifier with the extraction of kurtosis features, realizing multiple intrusion categorizing. However, the higher-order features are obtained through a calculation process of multiple iterations, which limits the real-time performance. MLP-NN has been proven to be an effective method for accurately behavior categorizing, due to their strong capability of capturing the complex and nonlinear dynamics. A comparative study of major approaches for intrusion behavior categorizing in perimeter security system indicates that MLP-NN outperforms the other methods in terms of predictive accuracy [18]. Considering that convolutional operation has been utilized to extract features which can reduce the scale of the network, the approach could provide real-time classification information with high accuracy.
3. Experiment and discussion
3.1 System architecture
In this work, our goal is to develop an enhanced intrusion behavior classification application in distributed perimeter security system. The overall architecture of proposed method in perimeter security monitoring system is shown in Fig. 2. The merged SI interferometer forms up the distributed sensor structure which can be deployed flexibly according to the monitoring demands. The composite evaluation terminal acquires source monitoring data from the photodetectors with 1G/s bandwidth which is fit for real-time monitoring system. To extract the features of the active vibration signal information accurately and entirely, the source data is processed with the vibration signal retrieval algorithm before classification.
The features of the monitoring signal are key to the accuracy of the classification model, therefore, the pre-processing retrieval algorithm is of great importance. The proposed system is based on the distributed interferometric structure which is phase-sensitive. And it is because of the phase sensitivity that the merged SI structure features with a long distance monitoring range. Thus, we apply the method to retrieve the phase shift of the coherence light caused by the vibration.
In order to get the detailed category information of the vibration behaviors to make an accurate judgement in perimeter security structure, a Multi-Layer Neural Network classification model is developed due to its merits of classify accuracy, efficiency and applicability for real-time monitoring implementation. Based on the monitoring information provided by the signal retrieval algorithm and the vibration behavior classifier, the composite intrusion evaluation terminal is capable of constructing an accuracy and reliable monitoring platform. The platform is highly integrated which provides real time intrusion alarm signals, intrusion location information, and intrusion behavior analysis including behavior categorizing, intrusion severity information and the live camera information of the nearest camera node. In this architecture, as highlighted above, introducing the active vibration classifier into the distributed monitoring system to classify the intrusion behavior is the major contribution of this paper.
3.2 Experiment setup
The experimental setup of the merged interferometric structure is shown in Fig. 3. A super-luminescent diode (SLD) source with a central wavelength spectrum of 1550 nm, 35nm bandwidth (minimum) is utilized as the light source. Between the 3×3 coupler and the 2×2 coupler, there is a 10km-long time-delay fiber (TDF) to ensure that the structure gives a high response to a common disturbance. To simplify the following location algorithm, the splitting ratio of the 3×3 coupler is set at 1:1:1; the initial phase difference of the 3×3 coupler outputs is outputs is set at ${120^ \circ }$.
Fig. 5. Comparison between source signal and retrieval result. (a) Source signal of detectors. (b) Retrieval result.
Additionally, all the fibers in the system are single-mode fiber. In the experimental setup, all the devices except the sensing fiber are deployed at the terminal side, which are equipped with anti-vibration measures to prevent the environmental interference. Connected with the terminal, the distributed sensing period which is shown as period L in Fig. 3 form a complete the overall interferometric structure.
A 10 kilometers long sensing fiber is distributed in the test sensing area, as shown in Fig. 4, which covers a same length period of perimeter fence. We set the test point at the distance of 0km, 2km, 4km, 6km, 8km respectively to cover the whole sensing period.
At each test point, the typical behaviors are performed to create vibration signals, which include pounding, striking, tapping, climbing events, the vibration caused by passing vehicles and the plant slapping caused by wind. Among the behaviors aforementioned, the pounding, striking, tapping, climbing behaviors are intrusion events, the vibration caused by passing vehicles and the plant slapping caused by wind are interference.
3.3 Pre-process
The vibration incident at each test point along the sensing fiber will be detected and captured by the evaluation terminal. The data-acquisition platform is at 500k sample rate. As it is mentioned that the merged interferometric structure is phase sensitive, to obtain the vibration signal from raw data, we apply the phase retrieval algorithm to retrieve the vibration signal and the retrieval result is shown in Fig. 5.
The detected signals at PD1, PD2 are given by:
where A, B are determined by the light power and the transimpedance of the system. $\alpha$ and $\beta$ are the nonreciprocal phase bias induced by the symmetrical 3×3 single-mode fiber coupler which enhances the resolution of the interfering signals. ${C_1}$ and ${C_2}$ are direct light current of system. $\Delta \phi$ is the phase shift generated by the disturbance. Calculus retrieval method is used to simplify equation that contains the $\Delta \phi$. The ultimate purpose of simplifying process is to retrieve the $\Delta \phi$, Subsequently, locate the disturbance. Through removing direct light current and transformation,By analyzing the integration, the phase shift caused by vibration is measured.
It is demonstrated from the Fig. 6 that the raw data contains many “invalid” information in every retrieval result. For the behavior classification model, the period which contains the vibration signal plays an important role in the raw data. Some research works divide the raw data into the behavior data frame according to the fluctuation rate counting, which may cause “behavior interception deviation”. Since the deviation data frame cannot objectively describe the behavior character, the application of these data frame may probably affect the classification accuracy of trained model.
Fig. 6. Retrieval results of different intrusion behaviors. (a) Pounding. (b) Striking. (c) Climbing. (d) Tapping. (e) Vibration caused by passing vehicles. (f) Plant slapping.
Fig. 7. Comparison of interception method between take-off point location method and fluctuation rate counting. (a) Take-off point location method. (b) Fluctuation rate counting. (c) Fluctuation rate counting.
Table 1. Data Sets Captured in Each Test Points
Therefore, in this work, the retrieval results of the raw data are tracked and intercepted to assure that each data frame contains the test vibration behavior. The schematic and comparison diagram is shown in Fig. 7. More specifically, the trigger threshold has been set to trace the take-off point in retrieval results, and the sliding window with fix length is utilized to partition the retrieval signal data series into a number of segment pairs to form up the data frame. We intercept 1000 points of each retrieval series to form the frame, and each data frame contains a respectively complete vibration behavior. Through this intercept operation, we can collect the valid vibration signals as much as possible to form up the data space. For each mode of behavior, we acquire about 400-700 sets at each test point, which is summarized in Table 1. In total, we obtain 2900, 3300, 2300, 2800, 2600, 1900 sets for pounding, striking, tapping, climbing events, the vibration caused by passing vehicles and the plant slapping caused by wind respectively. 80% of the total sample sets are randomly selected for training the intrusion classification model while the rest 20% of the total sample sets are randomly selected for testing.
4. Methodology
4.1 Power spectrum of the vibration signal
A reliable and accurate classification of intrusion behavior is essential for efficient perimeter security monitoring. In distributed fiber-optic perimeter security system, the detected signals are generally categorized into 3 situations: normal situation, intrusion events and interference which can be concluded from the retrieval results aforementioned. A number of studies have evaluated various time domain approaches for classifying the vibration behaviors in perimeter security system. In time domain analysis, short time energy feature and the fluctuation rate are usually utilized as the features to classify the vibration. When the system operates at normal state without obvious vibration acts such as the intrusion and the environmental interference on the fiber sensor, amplitude and fluctuation remain low and quiet. On the contrast, when intrusion events occur like the climbing or striking behavior act the fiber fence, both the amplitude and the fluctuation rate remarkably arise. Meanwhile, in another situation, when the distributed sensor structure is interfered by some slight environmental striking, amplitude remains unchanged but the frequency of signal changes notably. Thus, with the features of short time energy and the fluctuation rate, the model is able to distinguish some notable different modes. However, in other general cases, when the intrusion behavior is more similar to the interference in time domain, it performances with less accuracy and lower efficiency. In the comparison test, we choose 4 types of common intrusion behaviors and 2 types of interference as our classified behaviors, including (a) Pounding (b) Striking (c) Climbing (d) Tapping (e) Vibration caused by passing vehicles (f) Plant slapping. The first model is trained and tested when only including four types of intrusion behaviors, while the other model includes 4 types of intrusion behaviors and the other 2 types of interference. We can summarize from the classification result, which is shown in Table 2, that the classification model depending on time-domain character is effective when the signals are notably distinguished, however it performs with less accuracy when adding the interference.
Different from the aforementioned model using the features in time domain as the inputs, we aim at developing an intrusion behavior classification model with power frequency series extracted from the retrieval signals. As in frequency domain, the power spectrum is able to present change of power with frequency, therefore the power spectrum of the vibration signal gives us another way to analyze the vibration signals. Besides, in frequency domain, the calculation which is complex in time domain become more concise. In this work, since the sensor network structure is phase sensitive, the vibration is eventually reflected through the phase shift of the detected signals, and the differential of the phase information is the exact frequency, therefore, it is more concise and intuitive to analyze the signal in frequency domain.
In the study, power spectrum is applied. The energy matrix of characteristic frequency points, which contains more message of the signal, is applied as the input of the classification model. We obtain the spectrum sequence through N FFT, and N is chosen according to the frequency resolution of the sequence. It is demonstrated from the spectrum sequence that the frequency characters are mainly distributed in lower band, thus, we select the frequency band from 0Hz to 2000Hz to form up the frequency sequence, which is shown in Fig. 8. In each sequence, we pick one point every 4 points and serialize the chosen 400 frequency points as the input of the MLP-NN model.
Fig. 8. Frequency Spectrum of Each Behaviors. (a) Pounding. (b) Striking. (c) Climbing. (d) Tapping. (e) Vibration caused by passing vehicles. (f) Plant slapping.
4.2 Learning based intrusion behavior classification model
Due to its classification accuracy and generalization ability, Multi-Layer Neural Network (MLP-NN) is selected as our approach for behavior classifier in composite perimeter security monitoring. In order to avoid overfitting and reduce the scale of the parameters to be trained, a convolutional layer and a maxpool layer are utilized to extract the vibration signal features. Here, since we obtain the frequency series from 0-2000 Hz of each vibration signal data frame as the input of the network, which is a one-dimensional vector, we apply 1D-convolution to extract the features. Through a convolutional layer (filters=10, kernel_size=10×10, stride=1), and maxpool layer (pool_size=10×10, stride=1), we obtained a 532×1 vector, and then put it into the fully-connected (FC) layers.
The MLP-NN based vibration behavior classifier has a feed-forward neural network framework with the fully-connected layers, and each node of the layers contains a nonlinear activation function as shown in Fig. 9. Besides, the nonlinear activation needs to be pre-defined for each node to activate neurons in the fully-connected layer. Here, we choose the improvement of Relu function, that is Elu function, as the activation function for the first fully-connected layer. Because of the non-linear character, the Elu function is fit for non-linear mapping in network, and compared with sigmoid and tanh function, it is more concise to find the derivative of the loss function, making it shorter time to train the network. The function is formulated as:
In this formula, ${\varphi _j}$ is the activation function of node j; x is the input vector for node j; ${w_{kj}}$ is the output weights and ${b_{kj}}$ is the constant bias; $\sum j$ indicates the overall shape of the activation function. Eventually, at the output softmax layer, the output of each node is computed as a linear combination of the output nodes:
4.3 Evaluating the performance of intrusion behavior classification model
The evaluation of the intrusion behavior classification model based on the MLP network is conducted by using the test data collected from the experimental architecture, which is distributed on a 10 kilometers long perimeter fence to simulate a real situation. The program is written in MATLAB and python and it is evaluated on a computer with i7 CPU @ 2.80GHz and 32 GB memory.
In the training process, the batch size is set to 32 feeding into the network to improve the training efficiency to accelerate the update rate of the parameters. Here, we choose Adam optimizer to optimize the training process. The summary of the testing results of intrusion behavior classification based on MLP-NN is demonstrated in Table 3 in terms of the classification accuracy and computational cost for both training and testing. To evaluate the performance of the classification model, some comparison MLP-NN based models with different structures are set, and the performance is also presented in Table 3.
Table 2. Classification Accuracy of Model Based on Fluctuation Rate
Table 3. Classification Accuracy of MLP-NN-based Model
For all collected behaviors, the applied MLP-NN classification model has a solid performance, the total classification accuracy of testing set is up to 97.6%, matching up the accuracy of training set, showing the solid good performance with low bias and low variance. The classification speed represents for a given series of resource monitoring signal, how long it takes the trained model to return the classification results, it is noted that, as shown in Table 3, the time cost for applied classification of MLP-NN model is short within 0.01 s. As the comparison, the MLP-NN based model without convolutional layer is also tested, the comparison result demonstrates that convolutional operation brings sparse feature to the model, and make the applied classification model more efficient. Different neuron numbers of fully connected layer and different optimizers are also tested in the experiment. Through the multiple experiments, we obtained the applied MLP-NN based classification model with appropriate parameters.
Figure 10 shows the groundtruth and the classification results of normal situation, intrusion events and interference respectively. As shown in Fig. 10, MLP-NN is able to provide reliable results with satisfactory prediction accuracy for each scenario. It is noted that the model is able to distinguish the intrusion events, interference and normal situation clearly and accurately, the classification accuracy reaches 99.6%, 95% and 98.6% respectively. For all the intrusion events, the false negative rate of the proposed model is very low, for only 56 events have not been recognized, which ensures the reliability of the monitoring system. Meanwhile, the false positive rate of the proposed model is very low, for there are few interferences have been recognized as the intrusion events, which cut down the cost of manual inspection. Table 4 shows the statistic classification accuracy of each intrusion behavior and the accuracy of each behavior surpasses 94%. Besides, in case of some uncommon intrusion behaviors, 3 extra behaviors: fierce hitting, cutting and pile driver driving nearby are supplemented to retrain another MLP-NN based classification model, and the result shows that the classification accuracy of some uncommon intrusion behaviors also surpass 95%, which proves the generalization of the classification method. Compared with some classification methods with time domain features, the main approach of the proposed model is giving an accurate classification result for each behavior, which provides an effective proof for the monitoring judgement.
Fig. 11. Evaluation results of proposed model in perimeter security monitoring system. (1) Striking. (2) Climbing. (3) Tapping. (4) Vibration caused by passing vehicles. (5) Pounding. (6) Plant slapping.
Table 4. Classification Result of Each Intrusion Behaviors
For the last step, we have tested the proposed model in real monitoring system. From 12 p.m. of July 8th to 12 p.m. of July 18th 2020, we applied the intrusion behavior classifier in perimeter security monitoring system to evaluate the proposed method. The monitoring system was responsible for perimeter security monitoring of Dongtun transformer substation in Changsha, Hunan. With the assistance provided by the related personnel of management, we could obtain the real-time source monitoring signal from the original interferometric perimeter security system which was the part of the whole monitoring system. The source signal was evaluated by the proposed model, meanwhile, the manual inspection was implemented in main monitoring room through the distributed camera nodes as the groudtruth of the test, and the evaluation results are shown in Fig. 11. During the testing period, the proposed method successfully detected 21 events of total 23 events (17 sabotage events and 6 interference events), which was confirmed by the professional inspector through the camera nodes. Among the 21 detected events, 19 events were correctly categorized. The result has proved that the method can effectively distinguish the sabotage behaviors in real time security monitoring system. Meanwhile, some interferences were incorrectly categorized by the proposed model, which left us a topic to continually reduce the false positive rate of the classification model in realistic applications in the future.
5. Conclusion
This research proposes an intrusion behavior classification-based perimeter monitoring system for real time intrusion monitoring which proves to be an effective method to classify the categories of intrusion behavior. Combined with different types of monitoring method, such as the camera monitoring and vibration pre-warning system, the integrated platform provides an effective monitoring method for some vital target. The testing results indicate that the proposed MLP-NN model can realize a high accuracy and cost less computation time for real time monitoring implementation. Based on the preprocess method, or more precisely, the phase retrieval algorithm, the proposed MLP-NN model proves to be an accurate intrusion behavior classifier, even if it is under the condition of the environmental interferences. The classification results show that the total classification accuracy of testing set is up to 97.6%, matching up with the accuracy of training set. At last, we apply the proposed model in perimeter security monitoring system, and it also gives a high performance in each intrusion behavior category.
Funding
Science and Technology Commission of Shanghai Municipality (17DZ2280600).
Acknowledgments
The authors would like to thank Bo Jia and Qian Xiao for their support of this work. This paper is based upon the work supported by Science and Technology Commission of Shanghai Municipality under 17DZ2280600.
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. A. Rogers, “Distributed optical-fiber sensing,” Meas. Sci. Technol. 10(8), R75–R99 (1999). [CrossRef]
2. C. J. Oton, L. Tozzetti, F. Pasquale, and F. D. Pasquale, “High-Speed FBG Interrogation with Electro-Optically Tunable Sagnac Loops,” J. Lightwave Technol. 38(16), 4513–4519 (2020). [CrossRef]
3. B. Lee, “Review of the present status of optical fiber sensors,” Opt. Fiber Technol. 9(2), 57–79 (2003). [CrossRef]
4. Y. B. Ying, C. L. Zhao, and L. Y. Hou, “Demodulation method of Fabry-Perot sensor by cascading a traditional Mach-Zehnder interferometer,” Opt. Laser Technol. 118, 126–131 (2019). [CrossRef]
5. Y. P. Cao, H. M. Zhang, and B. Li, “Simultaneous measurement of temperature and refractive index based on microfiber Bragg Grating in Sagnac loop,” Opt. Fiber Technol. 47, 147–151 (2019). [CrossRef]
6. P. R. Hoffman and M. G. Kuzyk, “Position Determination of an Acoustic Burst Along.a Sagnac Interferometer,” J. Lightwave Technol. 22(2), 494–498 (2004). [CrossRef]
7. X. Fang, “Fiber-optic distributed sensing by a two-loop Sagnac interferometer,” Opt. Lett. 21(6), 444–446 (1996). [CrossRef]
8. Q. Sun, D. Liu, J. Wang, and H. Liu, “Distributed fiber-optic vibration sensor using a ring Mach-Zehnder interferometer,” Opt. Commun. 281(6), 1538–1544 (2008). [CrossRef]
9. A. A. Chtcherbakov, P. L. Swart, and S. J. Spammer, “Mach–Zehnder and Modified Sagnac-Distributed Fiber-Optic Impact Sensor,” Appl. Opt. 37(16), 3432–3437 (1998). [CrossRef]
10. Q. H. Song, Q. Xiao, H. Y. Wu, and B. Jia, “Improved localization algorithm for distributed fiber-optic sensor based on merged Michelson-Sagnac interferometer,” Opt. Express 28(5), 7207–7220 (2020). [CrossRef]
11. S. R. Xie, Q. L. Zou, L. W. Wang, M. Zhang, Y. H. Li, and Y. B. Liao, “Positioning error prediction theory for dual Mach-Zehnder interferometric vibration sensor,” J. Lightwave Technol. 29(3), 362–368 (2011). [CrossRef]
12. T. G. Liu, Z. Yu, J. F. Jiang, K. Liu, X. Z. Zhang, Z. Y. Ding, S. Wang, H. F. Hu, Q. Han, H. X. Zhang, and Z. H. Li, “Advanced of some critical technologies in discrete and distributed optical fiber sensing research,” Wuli Xuebao 66(7), 070705 (2017). [CrossRef]
13. J. Y. Yao, S. Zhang, W. Zhen, Z. Qi, and M. Liu, “Investigation of fiber-optic Michelson interferometer system,” Study Opt. Commun. 2, 45–48 (2007).
14. K. Liu, T. J. Chai, T. G. Liu, J. F. Jiang, and C. He, “Muti-area optical perimeter security system with quick invasion judgement algorithm,” J. Opt. Laser 26(2), 288–294 (2015). [CrossRef]
15. N. Lu and X. J. Lu, “Signal recognition algorithm of fiber-optic security system based on time-domain features,” Transducer and Microsystem Technologies 36(4), 150–152 (2017).
16. S. S. Mahmoud and J. Katsifolis, “Elimination of rain-induced nuisance alarms in distributed fiber-optic perimeter intrusion detection systems,” Proceedings of SPIE Fiber-Optic Sensors and Application International Society for Optics and photonics2009:1–11
17. Z. C. Li, K. Liu, J. F. Jiang, P. F. Ma, and T. G. Liu, “A high-accuracy event discrimination method in optical fiber perimeter security system,” Indared Laser Eng. 47(9), 922002 (2018). [CrossRef]
18. Z. X. Zhou, S. H. Duan, J. Tian, X. H. Wang, and B. L. Xu, “Research on optical fiber distributed vibration sensor and pattern recognition technique of vibration signal,” Optical Instruments 35(6), 11–15 (2013).
