The Fast Detection and Identification Algorithm of Optical Fiber ... - MDPI

7 downloads 0 Views 7MB Size Report
Aug 23, 2018 - For intrusion signals detection, the Constant False. Alarm Rate (CFAR) algorithms effectively overcome the adverse effect caused by a large ...
algorithms Article

The Fast Detection and Identification Algorithm of Optical Fiber Intrusion Signals Zhiyong Sheng, Dandan Qu *, Yuan Zhang and Dan Yang School of Electrical and Information Engineering, North China University of Technology, Beijing 100144, China; [email protected] (Z.S.); [email protected] (Y.Z.); [email protected] (D.Y.) * Correspondence: [email protected] Received: 21 June 2018; Accepted: 20 August 2018; Published: 23 August 2018

 

Abstract: With the continuous development of optical fiber sensing technology, the Optical Fiber Pre-Warning System (OFPS) has been widely used in various fields. The OFPS identifies the type of intrusion based on the detected vibration signal to monitor the surrounding environment. Aiming at the real-time requirements of OFPS, this paper presents a fast algorithm to accelerate the detection and recognition processing of optical fiber intrusion signals. The algorithm is implemented in an embedded system that is composed of a digital signal processor (DSP). The processing flow is divided into two parts. First, the dislocation processing method is adopted for the sum processing of original signals, which effectively improves the real-time performance. The filtered signals are divided into two parts and are parallel processed by two DSP boards to save time. Then, the data is input into the identification module for feature extraction and classification. Experiments show that the algorithm can effectively detect and identify the optical fiber intrusion signals. At the same time, it accelerates the processing speed and meets the real-time requirements of OFPS for detection and identification. Keywords: OFPS; fast algorithm; real-time performance; feature extraction

1. Introduction Optical fiber sensors have many advantages such as anti-electromagnetic interference, passive, good electrical insulation, high sensitivity, and adaptability to a wide range of complex environment [1–5]. Therefore, they are widely used in security pre-warning for border lines, airports, military bases, power plants, and oil pipelines [6]. The principle of Φ-OTDR optical technology is adapted in Optical Fiber Pre-Warning System (OFPS), the light pulse is injected from one end of the fiber, and the light detector is used to detect the backward Rayleigh scattered light. When the fiber link has a disturbance, the light intensity at the corresponding position changes, and the same time is measured before and after the measurement. The difference in energy of the position detects the vibration position. When an intrusion occurs above the ground, the vibration is transmitted to the buried fiber and causes it to vibrate. The purpose of OFPS is to detect intrusion events with high resolution and to identify the detected intrusions with high accuracy. How to improve the real-time performance of OFPS during data processing of optical fiber intrusion signals is a problem. In recent years, with the continuous improvement of random signal processing techniques and pattern recognition methods, the research into optical fiber intrusion signals detection and recognition algorithms has achieved significant progress. For intrusion signals detection, the Constant False Alarm Rate (CFAR) algorithms effectively overcome the adverse effect caused by a large number of non-stationary random interference signals, and greatly improve the detection performance of OFPS [7–10]. However, in practical applications, it is found that there are a large number of false alarms in CFAR detection results. In the embedded processing, the interruption time is 1 s. When the data sampling time reaches 1 s, the timer interruption is entered, and the sampled data is placed in the Algorithms 2018, 11, 129; doi:10.3390/a11090129

www.mdpi.com/journal/algorithms

Algorithms 2018, 11, 129

2 of 20

buffer. After reading the data of the buffer, continue the next second sampling. The CFAR algorithms need to process a large amount of data in a unit of time. If the data is not processed completely within 1 s, the remaining data will be left to be processed in the next second. This will lead to disordered data and will seriously affect the real-time performance and accuracy of the system. Therefore, this paper proposes a single-channel detection method, and uses the dislocation calculation method to further optimize the performance of the algorithm. It has been observed that the false alarm rate of this single-channel detection method is lower than CFAR. For the identification of intrusion events, the traditional analysis method is to extract the signal features in the time domain [11], frequency domain [12] or time-frequency domain [13–17] and perform the recognition based on features. However, these methods are limited to general-purpose computers and the processing speed is slow. In practical applications, embedded implementations can not only run outdoors, but also have the characteristics of high real-time performance, simple operation, and portability. With the development of embedded technology and OFPS, the outdoor operations demand of OFPS is more and more strict, and the requirements for real-time processing are getting higher and higher. Therefore, this paper uses Fast Fourier transform (FFT) to extract the features of optical fiber intrusion signals and identify them, which can achieve the purpose of rapid identification and meet the real-time performance of the system. Based on this, this paper presents a new fast algorithm for the detection and identification of optical fiber intrusion signals. The algorithm is simple and easy to implement, and the computational complexity is much lower than the traditional methods. The dislocation summation of the detection part and FFT of the identification part both greatly improve the real-time performance of the system. Moreover, the parallel mode and Ping-Pang structure of the two DSPs in the embedded platform make the actual operation faster. We collected relevant data through field experiments for verification of algorithm performance. Through many experiments, it is shown that this method can effectively remove false alarms and increase the recognition rate. At the same time, the actual operating speed has increased by 31%. The relative amount (31%) mentioned here means that the system runtime is reduced by 31% after using the algorithm proposed in this paper relative to the direct use of the CFAR algorithm. This efficiency is improved by a single-channel detection algorithm, coupled with the parallel use of DSP in hardware implementation. This article is organized as follows. In Section 2, we introduce the fast algorithm for detection and identification. Section 3 presents the implementation details of the embedded processor algorithm and experimental analysis of the actual data. Finally, Section 4 gives the conclusion. 2. Design of Fast Detection and Identification Algorithm for Optical Fiber Intrusion Signals According to the real-time demand of OFPS platform and the characteristics of artificial signal, this paper presents a fast algorithm for the detection and recognition of the optical fiber intrusion signals. The number of operations of this algorithm is O(n), and the number of operations of CFAR algorithm is O(n2 ). The algorithm can be divided into two parts, single-channel detection and recognition. As shown in Figure 1, the collected original data is divided into two paths—I-channel and Q-channel, and the two channels are respectively subjected to summation, high-pass filtering, and then the filtered result is normalized. When the time reaches 1 s, the norm summation of the normalized data of I-channel and Q-channel are obtained. Then, the norm summation of the I-channel and Q-channel is compared, and the larger normalized data are used as the identification data for the next step. The results of the FFT and the number of alarm points are used for identification, and the specific process is described in detail below.

Algorithms 2018, 11, x FOR PEER REVIEW Algorithms Algorithms2018, 2018,11, 11,129 x FOR PEER REVIEW Algorithms 2018, 11, x FOR PEER REVIEW

3 of 19 33of of20 19 3 of 19

10

 n=2

10 50 2 2 2 10 50 n n n 2 2 n =11 2 n=2 50 n 2 2n 2 2n n n=2 n n n =11 n n =11

a

 a + b /  a + b > threshold 2 b > threshold 2 + b a / + b a/ + b a > +threshold 2 

64 50 2 2 2 64 50 n n n 2 2 n =11 2 n = 50 50 n 2 2n 2 2n n n = 50 n n n =11 n n =11

2

n

2

n

 a + b /  a + b > threshold1 b > threshold1 + b a / + b a/ + b a > +threshold a 1   64

n = 50

2

n

2

n

Figure 1. The whole flow of the algorithm. Figure1.1. The The whole whole flow flowof ofthe thealgorithm. algorithm. Figure Figure 1. The whole flow of the algorithm.

2.1. Single-Channel Detection Algorithm 2.1. Single-Channel Single-Channel Detection DetectionAlgorithm Algorithm 2.1. 2.1. Single-Channel Detection Algorithm As shown in Figure 2, the original signals are summed adopting the dislocation processing As shown shown in in Figure Figure 2,2, the the original original signals signals are are summed summed adopting adopting the the dislocation dislocation processing processing As method, andinthe data is at a timeare interval of 4 ms. The total of theprocessing optical cable is As shown Figure 2, accumulated the original signals summed adopting the length dislocation method, and the data is accumulated at a time interval of 4 ms. The total length of the optical cableisis method, the data is cable accumulated at a time interval 4 ms. The length of optical the optical cable 0.72 and km. and We divide into at 45aparts with equal Accordingly, the of the columns method, the data is the accumulated time interval of length. 4 of ms. The totaltotal length of number the cable is 0.72km. km. We Wedivide dividethe the cableinto into45 45 parts parts with with equal equal length. length. Accordingly, Accordingly,the thenumber numberof ofthe the columns columns is setWe 45 in our subsequent processing and the length of each column isthe 16 number m. All columns are divided 0.720.72 km. divide the cablecable into 45 parts with equal length. Accordingly, of the columns isset set45 45 in in our our subsequent subsequentprocessing processingand andthe thelength lengthof ofeach eachcolumn columnisis16 16m. m. All All columns columns are are divided divided into four and the sum value filtered when thecolumn cumulative reaches 4 ms. shown in is setis 45 in ourgroups subsequent processing andisthe length of each is 16 time m. All columns are As divided into four four groups groupsand andthe thesum sumvalue valueisisfiltered filteredwhen whenthe thecumulative cumulativetime timereaches reaches44ms. ms. As Asshown shownin in the 3, and the number of columns to be when filtered one millisecond is 1/4 of4the columns, intointo fourFigure groups the sum value is filtered theincumulative time reaches ms.total As shown in and the Figure 3, the number of columns to be filtered in one millisecond is 1/4 of the total columns, and the 3, number the number of columns tothe beQ-channel filtered in data one millisecond is 1/4 of total the total columns, the Figure same processing is at the same time. the Figure 3, the ofperformed columns toon be filtered in one millisecond is 1/4 of the columns, and and thesame sameprocessing processingisisperformed performedon onthe theQ-channel Q-channeldata dataatatthe thesame sametime. time. the the same processing is performed on the Q-channel data at the same time.

Figure 2. 2. Flow Flow of of single–channel single–channel detection. detection. Figure Figure 2. Flow of single–channel detection. Figure 2. Flow of single–channel detection.

Figure Figure 3. 3. Summation Summation of of every every 44 ms ms in in I-channel. I-channel. Figure 3. Summation of every 4 ms in I-channel. Figure 3. Summation of every 4 ms in I-channel.

High-pass filtering is performed on the summation results expressed as follows: High-pass filtering is performed on the summation results expressed as follows: High-pass filtering is performed on the summation results expressed as follows:

y ( n ) = d1 x ( n ) + d 2 x ( n − 1) − c2 y ( n − 1)

y ( n ) is the current filtering result, x ( n ) is the current summation result, x ( n -1) rep

Algorithms 129 20 of the l mmation result of2018, the11,previous time unit, y( n -1) represents the filtering4 ofresult

addition, d1 = 0.9313 , d 2 = −0.9313 , c2 = 0.8626 .

High-pass filtering is performed on the summation results expressed as follows:

en the filtered data is normalized as follow

y ( n ) = d1 · x ( n ) + d2 · x ( n − 1) − c2 · y ( n − 1)

(1)

y ( n)

where y(n) is the current filtering result, x (n) is the current summation result, x (n − 1) represents the summation result of the previous time unit, y(n − 1) represents the filtering result of the last time unit. 1024 In addition, d1 = 0.9313, d2 = −0.9313, c2 = 0.8626. Then the filtered data is normalized as follow n=2 y(n) ∼ z(n) = (2) 1024 ∑ (| x (n) − x (n − 1)|)

z (n) ≅

 ( x(n) − x(n − 1) )

en the comparison between the normalized data and the threshold is performed. n =2 e selection process of the identification data is shown in Figure 4. The norm summatio Then the comparison between the normalized data and the threshold is performed. ized data points of theprocess I-channel and Q-channel are calculated respectively. Then The selection of the identification data is shown in Figure 4. The norm summation of 128 normalized data points of the I-channel and Q-channel are calculated respectively. Then the two are compared, and the data with larger norm summation is selected as the identificati results are compared, and the data with larger norm summation is selected as the identification data.

Figure 4. Extraction of identification data.

Figure 4. Extraction of identification data. 2.2. Feature Extraction and Identification

process of feature extraction and identification is shown in Figure 5. The features of ture Extraction The andwhole Identification OFPS signals are not obvious in the time domain. If these signals are transformed into the frequency domain, the corresponding features are clearer. In addition, FFT can extract the spectrum feature of a signal, which is often used in spectrum analysis. When FFT is performed on N points, the FFT results of N points can be obtained. Because of the symmetry of the FFT results, we usually only use the first half of the results in the frequency domain. The time domain and frequency domain of the data are analyzed, and the following conclusions are obtained: Mechanical signals have a stronger energy between 50 and 64 Hz in the FFT results, while other signals do not have this feature; The walking signal has stronger energy between 2–10 Hz in the FFT result, while other signals do not have this feature; The FFT result of the picking signal does not have obvious features, but its time domain duty cycle is significantly smaller than other signals.

e whole process of feature extraction and identification is shown in Figure 5. The fea ignals are not obvious in the time domain. If these signals are transformed into the fre , the corresponding features are clearer. In addition, FFT can extract the spectrum fe l, which is often used in spectrum analysis. When FFT is performed on N points, of N points can be obtained. Because of the symmetry of the FFT results, we usually o t half of the results in the frequency domain. The time domain and frequency domai e analyzed, and the following conclusions are obtained: Mechanical signals have a s between 50 and 64 Hz in the FFT results, while other signals do not have this featu g signal has stronger energy between 2–10 Hz in the FFT result, while other signals is feature; The FFT result of the picking signal does not have obvious features, but duty cycle is significantly smaller than other signals.

the first half of the results in the frequency domain. The time domain and frequency domain of the data are analyzed, and the following conclusions are obtained: Mechanical signals have a stronger energy between 50 and 64 Hz in the FFT results, while other signals do not have this feature; The walking signal has stronger energy between 2–10 Hz in the FFT result, while other signals do not have this2018, feature; Algorithms 11, 129The FFT result of the picking signal does not have obvious features, but its5time of 20 domain duty cycle is significantly smaller than other signals.

Algorithms 2018, 11, x FOR PEER REVIEW

5 of 19

We use a lot of experimental data (or you can edit the number) to find the energy ratio, and fit the obtained result to the probability density function. The probability density function curve is shown in the figure. As can be seen from the figure, thresholds selection is most appropriate at the Figure 5. Feature extraction and identification. intersection of the curves. Figure 5. Feature extraction and identification. The process of feature extraction and identification is shown in Figure 6. The 128 points FFT is We use a lot of experimental dataNext, (or you edit the number) to find energy ratio, fit the applied for each column of the data. thecan moduli of points 50–64 andthe points 11–50 areand summed obtained result to the probability density function. The probability density function curve is shown in up, respectively. Then the former is divided by the latter and the quotient is compared with the the figure. Asthreshold can be seen from the figure,signal. thresholds appropriate at the intersection recognition of the mechanical If the selection quotient is most greater than the threshold as shown of the curves. in Equation (3), where threshold_1 = 0.9, it is judged as mechanical signal. The process of feature extraction and identification is shown in Figure 6. The 128 points FFT is 64 50 2 2 2 2 applied for each column of the data.anNext, 50–64 and points 11–50 are summed (3) up, + bn the / moduli an of + bpoints n > threshold _1 11 respectively. Then the former nis=50divided by then =latter and the quotient is compared with the recognition threshold the mechanical signal. thethreshold_1, quotient is greater thanisthe thresholdas aswalking shown insignal. Equation If theofquotient is smaller than Ifthe the signal determined The (3), where threshold_1 = 0.9, it is judged as mechanical signal. discrimination method is shown in Equation (4), where threshold_2 = 0.55



64 10p

∑

n=2 n=50



2

2

+ bbnn 2 // aann2 +

50 50

 ∑

=11 nn= 11

p2 anan+2 + bn 2bn>2 threshold _2 > threshold_1

(3) (4)

Figure 6. The selection processing of thresholds. Figure 6. The selection processing of thresholds.

If the quotient in Equation (4) is larger than threshold_2, it is judged as walking signal. If the If theis quotient is smaller than the threshold_1, signal is determined as walking quotient smaller than threshold_2, the signal will be the further identified if it is picking signals.signal. The discrimination method is shown in Equation (4), where threshold_2 = 0.55 The picking signal is identified based on the number of alarm points rather than the result of FFT. And if the number of the alarm points is not within the range between 2 and 15, it is ultimately 10 p 50 p 2 2 determined as false alarm. ∑ an + bn / ∑ an 2 + bn 2 > threshold_2 (4) n =2

n=11

3. The Implementation of Recognition Algorithm and the Experimental Analysis 3.1. The Implementation of Fast Identification Algorithm As shown in Figure 7, the optical fiber early warning system is equipped with an optical pulse transmitting board, an optical pulse receiving board, a signal processing board, a data switching board, a motherboard, and a display. Optical pulse-emitting plates are used to modulate, amplify, and emit light

Algorithms 2018, 11, 129

6 of 20

If the quotient in Equation (4) is larger than threshold_2, it is judged as walking signal. If the quotient is smaller than threshold_2, the signal will be further identified if it is picking signals. The picking signal is identified based on the number of alarm points rather than the result of FFT. And if the number of the alarm points is not within the range between 2 and 15, it is ultimately determined as false alarm. 3. The Implementation of Recognition Algorithm and the Experimental Analysis 3.1. The Implementation of Fast Identification Algorithm As shown in Figure 7, the optical fiber early warning system is equipped with an optical pulse transmitting board, an optical pulse receiving board, a signal processing board, a data switching board, a motherboard, and a display. Optical pulse-emitting plates are used to modulate, amplify, and emit light pulses. The optical pulse-receiving plate is for receiving backscattered light in the optical fiber and converting the optical signal into an electrical signal. The signal processing board is used to collect and process data, complete the detection and identification of the intrusion signal, and send the alarm result to the upper computer. The software architecture designed in this paper is implemented in the signal processing board. The motherboard runs the host computer software, together with the display to complete display of REVIEW the alarm results and the setting of the system parameters. Algorithms 2018,the 11, x FOR PEER 6 of 19

Figure Figure 7. 7. Optical Optical fiber fiber Pre-warning Pre-warning monitoring monitoring system. system.

As shown in Figure 8, the fiber-optic warning system signal processing board is equipped with As shown in Figure 8, the fiber-optic warning system signal processing board is equipped with two TS201 processors, one FPGA processor and one PowerPC processor, as well as multiple memory two TS201 processors, one FPGA processor and one PowerPC processor, as well as multiple memory and interface circuits. The signal processing board inputs the in-phase analog signal Sig_I and the and interface circuits. The signal processing board inputs the in-phase analog signal Sig_I and the quadrature analog signal Sig_Q, and the ADC module with the sampling frequency of 100 MHz is quadrature analog signal Sig_Q, and the ADC module with the sampling frequency of 100 MHz is converted into a digital signal; the FIFO queue in the FPGA module is used to read the data in the converted into a digital signal; the FIFO queue in the FPGA module is used to read the data in the ADC and cache, and the FPGA takes 1 ms as the the cycle simultaneously initiates an interrupt ADC and cache, and the FPGA takes 1 ms as the the cycle simultaneously initiates an interrupt request request to DSP1 and DSP2; DSP1 and DSP2 respond to the interrupt request and read the cache data to DSP1 and DSP2; DSP1 and DSP2 respond to the interrupt request and read the cache data from the from the FIFO queue of the FPGA. In addition, the peripheral interface of DSP1 and DSP2 is FIFO queue of the FPGA. In addition, the peripheral interface of DSP1 and DSP2 is implemented in the implemented in the FPGA to ensure data between the two DSPs and the PowerPC. Transparent FPGA to ensure data between the two DSPs and the PowerPC. Transparent transmission; PowerPC is transmission; PowerPC is responsible for communication with the host computer, and managing DSP responsible for communication with the host computer, and managing DSP software upgrades. software upgrades.

ADC and cache, and the FPGA takes 1 ms as the the cycle simultaneously initiates an interrupt request to DSP1 and DSP2; DSP1 and DSP2 respond to the interrupt request and read the cache data from the FIFO queue of the FPGA. In addition, the peripheral interface of DSP1 and DSP2 is implemented in the FPGA to ensure data between the two DSPs and the PowerPC. Transparent transmission; PowerPC Algorithms 2018, 11, 129 is responsible for communication with the host computer, and managing DSP 7 of 20 software upgrades.

Figure 8. Signal processing board module diagram. Figure 8. Signal processing board module diagram.

As shown in Figure 9, the entire system is consisted of DSP and host computer. The DSP module As shown in Figure 9, the entire system is consisted of DSP and host computer. The DSP module is used for optical fiber intrusion signals acquisition at the test site and subsequent signal processing. is used for optical fiber intrusion signals acquisition at the test site and subsequent signal processing. A mechanical drill is used to shovel the soil near the underground optical fiber, and the mechanical A mechanical drill is used to shovel the soil near the underground optical fiber, and the mechanical drilling signals are collected. Near the underground optical fiber, electric pick is used to excavate the drilling signals are collected. Near the underground optical fiber, electric pick is used to excavate the soil and generate mechanical picking signals. We generate the walking signals by walking around soil and generate mechanical picking signals. We generate the walking signals by walking around the the optical fiber. optical fiber. Algorithms 2018, 11, x FOR PEER REVIEW 7 of 19

Figure 9. 9. Frame Frame diagram diagram of of hardware hardware implementation. implementation. Figure

According to the characteristics and requirements of algorithm implementation on embedded According to the characteristics and requirements of algorithm implementation on embedded hardware, the computational complexity and operation time of each algorithm module are analyzed. hardware, the computational complexity and operation time of each algorithm module are analyzed. The algorithm flow is decomposed into different processors on the board. The processing flow after The algorithm flow is decomposed into different processors on the board. The processing flow filtering is distributed in two parallel-processing DSPs. The two DSPs work in parallel and the after filtering is distributed in two parallel-processing DSPs. The two DSPs work in parallel and intermediate results are input to the Synchronous Dynamic Random Access Memory (SDRAM) the intermediate results are input to the Synchronous Dynamic Random Access Memory (SDRAM) through the data bus. The external raw data is input to the DSP through the AD interface and LINK through the data bus. The external raw data is input to the DSP through the AD interface and LINK port. Finally, the identification results are sent to the host through the PCI interface. port. Finally, the identification results are sent to the host through the PCI interface. The processing flows in the two DSPs are shown in Figure 10. First, the raw data is received from The processing flows in the two DSPs are shown in Figure 10. First, the raw data is received the TTL interface. Then it is sent to DSP1 through the link port, and stored in the Ping-Pang SDRAM from the TTL interface. Then it is sent to DSP1 through the link port, and stored in the Ping-Pang memory. At the same time, DSP1 obtains the data of 1024 ms from the Ping-Pang memory for the SDRAM memory. At the same time, DSP1 obtains the data of 1024 ms from the Ping-Pang memory for subsequent high-pass filtering, and the results are stored in the on-chip RAM. After the detection and the subsequent high-pass filtering, and the results are stored in the on-chip RAM. After the detection identification processing of the first 512 ms filtered data is completed, the results are uploaded and identification processing of the first 512 ms filtered data is completed, the results are uploaded through the PCI bus. Then DSP1 waits for the next 1024 ms of data. At the same time, DSP2 obtains through the PCI bus. Then DSP1 waits for the next 1024 ms of data. At the same time, DSP2 obtains the the filtered data of the second 512 ms from the Ping-Pang structure for subsequent processing. DSP1 filtered data of the second 512 ms from the Ping-Pang structure for subsequent processing. DSP1 and and DSP2 work in parallel mode, and finally determine the output signal type. In this way, parallel DSP2 work in parallel mode, and finally determine the output signal type. In this way, parallel flow flow processing is realized through the Ping-Pang processing of DSP1 and DSP2. The processing time is shortened and the real-time performance is improved.

the TTL interface. Then it is sent to DSP1 through the link port, and stored in the Ping-Pang SDRAM memory. At the same time, DSP1 obtains the data of 1024 ms from the Ping-Pang memory for the subsequent high-pass filtering, and the results are stored in the on-chip RAM. After the detection and identification processing of the first 512 ms filtered data is completed, the results are uploaded through the2018, PCI11,bus. Algorithms 129 Then DSP1 waits for the next 1024 ms of data. At the same time, DSP2 obtains 8 of 20 the filtered data of the second 512 ms from the Ping-Pang structure for subsequent processing. DSP1 and DSP2 work in parallel mode, and finally determine the output signal type. In this way, parallel processing is realized through the Ping-Pang time is flow processing is realized through the Ping-Pangprocessing processingofofDSP1 DSP1and andDSP2. DSP2. The The processing processing time shortened and the real-time performance is improved. is shortened and the real-time performance is improved.

Figure Structure diagram of parallel processing. Figure 10.10. Structure diagram of parallel processing.

3.2.3.2. Experiment Analysis of Actual Data Experiment Analysis of Actual Data ToTo verify thethe algorithm, wewe designed experiments to to collect actual data in Shangweidian Village, verify algorithm, designed experiments collect actual data in Shangweidian Village, Mentougou District, Beijing. The tester takes four kinds of of actions of of mechanical drilling, mechanical Mentougou District, Beijing. The tester takes four kinds actions mechanical drilling, mechanical picking, walking dividedinto intothree threegroups groupsand andlasts lasts 2 picking, walkingand andpicking pickingrespectively. respectively. Each Each action is divided forfor 2 min. min. of each group of data is 122,880 rows and columns. order observe detection TheThe sizesize of each group of data is 122,880 rows and 45 45 columns. In In order to to observe thethe detection and and identification results, collection of data is performed in a specific geographical It is identification results, thethe collection of data is performed in a specific geographical area. area. It is evident evident from the results the signals are mainly concentrated in columns 41 The and abscissa 42. The abscissa from the results that thethat signals are mainly concentrated in columns 41 and 42. represents represents the locations and the between distance adjacent betweencolumns adjacentiscolumns is 16 m. Therepresents ordinate represents the locations and the distance 16 m. The ordinate time and the time and the unitThe is second. The number of false alarms the major factor for theFor detection. For the unit is second. number of false alarms is the majorisfactor for the detection. the identification, the mechanical drilling and mechanical picking signals are both classified as mechanical signals. 3.2.1. The Detection and Identification Results of Mechanical Signals In the figures of detection results, the black dots represent the optical fiber vibration signals. FFT is applied to the identification data extracted from the detection results, and the sum of the FFT results of different frequency bands is compared with the recognition threshold. The intrusion type of signals is judged as the identification result. In the figures of the identification results, black circles represent mechanical signals, red squares represent walking signals, and green triangles represent picking signals. The detection results of mechanical signals are as Figures 11 and 12: It can be observed from Figures 11 and 12 that the alarms of the mechanical signals are mainly concentrated in the 41st column and 42nd column. There are only a few false alarm points near the 10th column and 11th column. Overall, the detection results are relatively ideal. And in the subsequent identification processing, the alarm point will be further verified. The identification results of mechanical signals are shown in Figures 13 and 14: The statistical recognition result of mechanical drilling is shown in Table 1. The drill 1, drill 2 and drill 3 at the Table 1 represent three groups of mechanical drilling signals. The picking, mechanical and walking in the first row of Table 1 represent the number of the signals which are finally recognized as these three types. The statistics method for recognition rate of following three types of signals is the same as the mechanical drilling signals. The recognition result of mechanical drilling signals is shown in figures above. The ratio of the number of each kind of signal to the total number is calculated, respectively, and the recognition rate is obtained. The statistical recognition rate is shown in Table 2.

applied to the identification data extracted from the detection results, and the sum of the FFT results of different frequency bands is compared with the recognition threshold. The intrusion type of signals is judged as the identification result. In the figures of the identification results, black circles represent mechanical signals, red squares represent walking signals, and green triangles represent picking signals. Algorithms 2018, 11, 129 9 of 20 The detection results of mechanical signals are as Figures 11 and 12: 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(a) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(b) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(c) Figure Figure 11. The results drilling signals. (a) first Thegroup first of group of mechanical 11. detection The detection resultsofofmechanical mechanical drilling signals. (a) The mechanical signals; (b) The second groupof of mechanical mechanical drilling signals; (c) The mechanical drillingdrilling signals; (b) The second group drilling signals; (c)third Thegroup thirdofgroup of mechanical drilling signals. drilling signals.

Algorithms 2018, 129 PEER REVIEW Algorithms 2018, 11, 11, x FOR

10 of 20

9 of 19

120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(a) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(b) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(c) Figure 12.12. The detection mechanical picking signals. first group of mechanical Figure The detection results results ofof mechanical picking signals. (a) The (a) firstThe group of mechanical picking signals; (b)The Thesecond second group of of mechanical picking signals;signals; (c) The third group of mechanical picking signals; (b) group mechanical picking (c) The third group of mechanical picking signals. picking signals.

It can be observed from Figures 11 and 12 that the alarms of the mechanical signals are mainly concentrated in the 41st column and 42nd column. There are only a few false alarm points near the 10th column and 11th column. Overall, the detection results are relatively ideal. And in the subsequent identification processing, the alarm point will be further verified. The identification results of mechanical signals are shown in Figures 13 and 14:

Algorithms 2018, 129 PEER REVIEW Algorithms 2018, 11,11, x FOR

11 of 20

10 of 19

120 100

time(s)

80 60 40 20 0

0

mechanical walking picking 10

20 30 location(16m)

40

(a) 120 100

time(s)

80 60 40 20 0

0

mechanical walking picking 10

20 30 location(16m)

40

(b) 120 100

time(s)

80 60 40 20 0

0

mechanical walking 10

20 30 location(16m)

40

(c) Figure 13.13. The results of mechanical drilling (a) group The first group of mechanical Figure Theidentification identification results of mechanical drilling signals.signals. (a) The first of mechanical drilling signals; (b) The second group of mechanical drilling signals; (c) The third group of mechanical drilling signals; (b) The second group of mechanical drilling signals; (c) The third group of mechanical drilling signals. drilling signals. 120 100

time(s)

80 60 40 20

mechanical walking picking

(c) Figure 13. The identification results of mechanical drilling signals. (a) The first group of mechanical drilling signals; (b) The second group of mechanical drilling signals; (c) The third group of mechanical Algorithms 11, 129 12 of 20 drilling2018, signals. 120 100

time(s)

80 60 40 mechanical walking picking

20 0

0

10

Algorithms 2018, 11, x FOR PEER REVIEW

20 30 location(16m)

40

11 of 19

(a) 120 100

time(s)

80 60 40 20 0

0

mechanical walking picking 10

20 30 location(16m)

40

(b) 120 100

time(s)

80 60 40 20 0

0

mechanical walking picking 10

20 30 location(16m)

40

(c) Figure 14.14. The results of mechanical picking (a) group The first group of mechanical Figure Theidentification identification results of mechanical picking signals.signals. (a) The first of mechanical picking signals;(b) (b)The The second second group of mechanical picking signals;signals; (c) The third group of mechanical picking signals; group of mechanical picking (c) The third group of mechanical picking signals. picking signals. Table 1. The statistical identification results of Mechanical drilling signals.

The statistical recognition result of mechanical drilling is shown in Table 1. The drill 1, drill 2 and drill 3 at the Table 1 represent three groupsMechanical of mechanical drilling signals. The picking, Picking Walking mechanical and walking in drilling1 the first row of14Table 1 represent the number of the signals which are Mechanical 93 8 Mechanical drilling2 30 107 10 finally recognized as these three types. The statistics method for recognition rate of following three Mechanical drilling3 0 210 10 types of signals is the same as the mechanical drilling signals. The recognition result of mechanical drilling signals is shown in figures above. The ratio of the number of each kind of signal to the total number is calculated, respectively, and the recognition rate is obtained. The statistical recognition rate is shown in Table 2. As shown in Figure 15, the reason why the mechanical drilling signal is recognized as picking signal is that the feature extracted by FFT has strong energy from 50 Hz to 64 Hz, but there is an

20 0

0

walking picking 10

20 30 location(16m)

40

(c) Figure 14. Algorithms 2018, 11,The 129 identification results of mechanical picking signals. (a) The first group of mechanical 13 of 20 picking signals; (b) The second group of mechanical picking signals; (c) The third group of mechanical picking signals. Table 2. The identification rate of Mechanical signals.

The statistical recognition result of mechanical drilling is shown in Table 1. The drill 1, drill 2 Picking Walking and drill 3 at the Table 1 represent three groups of Mechanical mechanical drilling signals. The picking, Mechanical 80.9% mechanical and walking in drilling1 the first row of 12.1% Table 1 represent the number 7.0% of the signals which are drilling2 72.8% 6.8% finally recognizedMechanical as these three types. The 20.4% statistics method for recognition rate of following three Mechanical drilling3 0 95.5% 4.5% types of signals isMechanical the same picking1 as the mechanical drilling signals. The recognition 6.3% 92.1% 1.6% result of mechanical drilling signals is Mechanical shown in figures ratio of the number picking2above. The 10.4% 86.5% of each kind 3.1% of signal to the total Mechanical picking3and the recognition 4.5% 91.6% 3.9%statistical recognition number is calculated, respectively, rate is obtained. The rate is shown in Table 2. As shown shown in in Figure Figure 15, 15, the thereason reasonwhy whythe themechanical mechanical drilling drilling signal signal isis recognized recognized as as picking picking As signal is that the feature extracted by FFT has strong energy from 50 Hz to 64 Hz, but there is an an signal is that the feature extracted by FFT has strong energy from 50 Hz to 64 Hz, but there is obvious peak value at 12 Hz resulting in a small ratio of 0.7737 much below the threshold. As shown obvious peak value at 12 Hz resulting in a small ratio of 0.7737 much below the threshold. As shown in Figure Figure16, 16,the the reason reasonwhy whythe themechanical mechanicaldrilling drillingsignal signalisisrecognized recognizedas aswalking walkingsignal signalisisthat that the the in characteristics of mechanical drilling are not obvious where there is no remarkable high frequency characteristics of mechanical drilling are not obvious where there is no remarkable high frequency componentand andthe thelow lowfrequency frequencycomponent componentisislarge. large. component

X:12 Y:0.1358

(a)

(b)

Figure 15. Mechanical signal recognized as as picking picking signal. signal. (a) (a) Time Timedomain; domain;(b) (b)Frequency Frequencydomain. domain. Algorithms 2018, 11,Mechanical x FOR PEERsignal REVIEW 12 of 19 Figure 15. recognized

(a)

(b)

Figure 16. Mechanical drilling signal signal.(a) (a)Time Timedomain; domain; Frequency domain. Figure 16. Mechanical drilling signalrecognized recognizedas as walking walking signal. (b)(b) Frequency domain. Table 1. The statistical identification results of Mechanical drilling signals.

3.2.2. The Detection and Identification Results of Walking Signals Picking Mechanical The detection results of walking signals are as Figure 17: Mechanical drilling1 14 93 Mechanical drilling2 30 107 Mechanical drilling3 0 210

Walking 8 10 10

Table 2. The identification rate of Mechanical signals.

Mechanical drilling1 Mechanical drilling2 Mechanical drilling3 Mechanical picking1 Mechanical picking2

Picking 12.1% 20.4% 0 6.3% 10.4%

Mechanical 80.9% 72.8% 95.5% 92.1% 86.5%

Walking 7.0% 6.8% 4.5% 1.6% 3.1%

Mechanical picking3

4.5%

91.6%

3.9%

3.2.2. The Detection and Identification Results of Walking Signals Algorithms 2018, 11, 129 The detection results

of walking signals are as Figure 17:

14 of 20

120

time(s)

100 80 60 40 20 0

0

10

Algorithms 2018, 11, x FOR PEER REVIEW

20 30 40 location(16m)

50 13 of 19

(a) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(b) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(c) Figure 17.17. The detection walking signals. (a)first Thegroup firstofgroup ofsignals; walking (b) The Figure The detection results results ofof walking signals. (a) The walking (b) signals; The second group of walking signals; (c) The thrid group of walking signals. second group of walking signals; (c) The thrid group of walking signals. alarm pointsof of the the walking are are mainly concentrated in the 41st and 42nd and 42nd TheThe alarm points walkingsignals signals mainly concentrated incolumn the 41st column column, which is the location of the collected data. The false alarm point is relatively the least among column, which is the location of the collected data. The false alarm point is relatively the least among these four types of signals. these four types of signals. The identification results of walking signals are shown in Figure 18:

The identification results of walking signals are shown in Figure 18: 120 100

)

80

second group of walking signals; (c) The thrid group of walking signals.

The alarm points of the walking signals are mainly concentrated in the 41st column and 42nd column, which is the location of the collected data. The false alarm point is relatively the least among these four types of signals. Algorithms 2018, 11, 129 15 of 20 The identification results of walking signals are shown in Figure 18: 120 100

time(s)

80 60 40 20

walking picking

0

0 Algorithms 2018, 11, x FOR PEER REVIEW

10

20 30 location(16m)

40

14 of 19

(a) 120 100

time(s)

80 60 40 20 0

walking picking

0

10

20 30 location(16m)

40

(b) 120 100

time(s)

80 60 40 20 0

walking picking

0

10

20 30 location(16m)

40

(c) Figure 18. The identification results of walking signals. (a) The first group of walking signals; (b) The Figure 18. The identification results of walking signals. (a) The first group of walking signals; (b) The second group of walking signals; (c) third The third group of walking signals. second group of walking signals; (c) The group of walking signals.

As shown in Table 3, the recognition rate of the walking signals is very high, and there are no As shown in Table 3, the recognition rate of the walking signals is very high, and there are no walking signals recognized as mechanical signals. In a few cases, walking signal is recognized as walking signals recognized as mechanical signals. In a few cases, walking signal is recognized as picking signal. As shown in Figure 19, the walking signals’ features extracted by FFT have no obvious picking signal. As shown in Figure 19, the walking signals’ features extracted by FFT have no obvious peak value in the range of 2–10 Hz. The ratio calculated in this part is 0.3421, which is far below the peak value in the range of 2–10 Hz. The ratio calculated in this part is 0.3421, which is far below the threshold 0.55, so it will not be judged as walking signal. At the same time, the number of alarm threshold 0.55, so it will not be judged as walking signal. At the same time, the number of alarm points points which belongs to this second is 7, which is also in the range of 2–15, so it is judged as a picking which belongs to this second is 7, which is also in the range of 2–15, so it is judged as a picking signal. signal. Table 3. The identification rate of walking signals.

walking 1 walking 2 walking 3

(a)

Mechanical

Walking

Picking

5.6% 1.6% 2.7%

94.4% 98.4% 97.3%

0 0 0

(b)

walking signals recognized as mechanical signals. In a few cases, walking signal is recognized as picking signal. As shown in Figure 19, the walking signals’ features extracted by FFT have no obvious peak value in the range of 2–10 Hz. The ratio calculated in this part is 0.3421, which is far below the threshold 0.55, so it will not be judged as walking signal. At the same time, the number of alarm points which belongs to this second is 7, which is also in the range of 2–15, so it is judged as a picking Algorithms 2018, 11, 129 16 of 20 signal.

(a)

(b)

Figure 19. Walking signal recognized as picking signal. (a) Time domain; (b) Frequency domain; as picking signal. (a) Time domain; (b) Frequency domain; 15 of 19

Figure 19. Walking recognized Algorithms 2018, 11, x FORsignal PEER REVIEW

Table 3. The identification rate of walking signals.

time(s)

The Identification Results of Picking Signals 3.2.3. 3.2.3. The Identification Results of Picking Signals Mechanical Walking The detection results of picking signals are Figure 20: The detection results of picking signals are asasFigure 20: walking 1 5.6% 94.4% walking 120 2 1.6% 98.4% walking 100 3 2.7% 97.3%

Picking 0 0 0

80 60 40 20 0

0

10

20 30 40 location(16m)

50

(a) 120

time(s)

100 80 60 40 20 0

0

10

20 30 40 location(16m)

50

(b)

120

time(s)

100 80 60 40 20 0 0

10

20 30 40 location(16m) (c)

50

Figure 20. The detection results of picking signals. (a) The first group of picking signals; (b) The

Figure 20. The detection results of picking signals. (a) The first group of picking signals; (b) The second second group of picking signals; (c) The third group of picking signals. group of picking signals; (c) The third group of picking signals.

The noise signal is different from the previous two types of signals. The alarm points are not only concentrated in the 41st and 42th columns. The alarm points in the 10th, 11th, and 31st columns are also very obvious. However, it does not mean that the 10th column, 11th column, and 31th column are vibration signals. And in the subsequent identification processing, the alarm point will be further verified. The identification results of picking signals are shown in Figure 21:

Algorithms 2018, 11, 129

17 of 20

The noise signal is different from the previous two types of signals. The alarm points are not only concentrated in the 41st and 42th columns. The alarm points in the 10th, 11th, and 31st columns are also very obvious. However, it does not mean that the 10th column, 11th column, and 31th column are vibration signals. And in the subsequent identification processing, the alarm point will be further verified. Algorithms 2018, 11, x FOR results PEER REVIEW 16 of 19 The identification of picking signals are shown in Figure 21: 120 100

time(s)

80 60 40 20 0

0

mechanical picking 10

20 30 location(16m)

40

(a) 120 100

time(s)

80 60 40 20 0

picking 0

10

20 30 location(16m)

40

(b) 120 100

time(s)

80 60 40 20 0

0

mechanical picking 10

20 30 location(16m)

40

(c) Figure The identification results picking signals.(a)(a)The Thefirst first group picking signals; The Figure 21.21. The identification results ofof picking signals. group ofof picking signals; (b)(b) The second group of picking signals; (c) The third group of picking signals. second group of picking signals; (c) The third group of picking signals.

As shown in Table 4, the recognition rate of picking signals is almost completely right. The As shown in Table 4, the recognition rate of picking signals is almost completely right. recognition of one group of data is 100%, and the other group is 0.9618. A very few of picking signals The recognition of one group of data is 100%, and the other group is 0.9618. A very few of picking are recognized as mechanical signals. It is shown in Figure 22 that the low frequency component of signals are recognized as mechanical signals. It is shown in Figure 22 that the low frequency component picking signal is obvious and the ratio is 0.7223, which is obviously larger than the threshold value, of picking signal is obvious and the ratio is 0.7223, which is obviously larger than the threshold value, 0.55, so it is judged as a mechanical signal. 0.55, so it is judged as a mechanical signal.

Algorithms 2018, 11, 129

18 of 20

Table 4. The identification rate of picking signals.

Picking 1 Picking 2 Algorithms 2018, 11, x FOR PEER REVIEW Picking 3

Picking

Mechanical

Walking

96.2% 100% 95.6%

3.8% 0 4.4%

0 0 0

(a)

17 of 19

(b)

Figure 22. Picking signal recognized as walking signal. (a) Time domain; (b) Frequency domain; Figure 22. Picking signal recognized as walking signal. (a) Time domain; (b) Frequency domain; Table 4. The identification rate of picking signals.

3.2.4. The Performance Comparison of Fast Algorithm with Other Methods Picking Mechanical Walking As shown in Table 5, the false alarm rate of single-channel detection algorithm proposed in this Picking 1 96.2% False3.8% 0 article is 2.6% lower than Cell Average Constant Alarm Rate (CA-CFAR). Picking 2 100% 0 0 Pickingperformance 3 95.6% of single-channel 4.4% 0 algorithm with CA-CFAR. Table 5. Comparison of detection detection Method Realization Method Alarm with Rate Other Methods 3.2.4. The Performance Comparison of Fast False Algorithm detection 5.2% Embedded platform Assingle-channel shown in Table 5, thealgorithm false alarm rate of single-channel detection algorithm proposed in this CA-CFAR 7.8% Embedded platform article is 2.6% lower than Cell Average Constant False Alarm Rate (CA-CFAR).

The recognition performance each method is shown indetection Table 6.algorithm The recognition rate of the Table 5. Comparison of detectionof performance of single-channel with CA-CFAR. methods used in this paper is relatively high, and the average recognition rate is 93.51%. Method False Alarm Rate Realization Method single-channel detectionofalgorithm 5.2% of each cognitive Embedded platform Table 6. Comparison Recognition performance strategy. CA-CFAR 7.8% Embedded platform Recognition Rate Method performance of each method is shown in Table 6. The recognition rate of the The recognition Picking Mechanical Walking Average methods used in this paper is relatively high, and the average recognition rate is 93.51%. This strategy 97.3% 86.6% 96.7% 93.5% Literature [3] 93.5% 84.3% 91.8% 89.9% Table performance of each cognitive strategy. Literature [5]6. Comparison 86.4%of Recognition 80.0% 84.1% 83.5%

Recognition Rate Method The real-time performancePicking of each method is shown in Table 7 that is better than other strategies Mechanical Walking Average with the time consumption s. It can be 86.6% seen that the method this paper not only This strategy of 0.659 97.3% 96.7% adopted in93.5% guarantees the recognition but also is superior to other algorithm Literature [3] rate, 93.5% 84.3% 91.8% in time efficiency. 89.9% Literature [5] 86.4% 80.0% 84.1% 83.5% Table 7. Comparison of Recognition performance of each cognitive strategy.

The real-time performance of each method is shown in Table 7 that is better than other strategies Method Realization Method Time/s with the time consumption of 0.659 s. ItProcessing can be seen that the method adopted in this paper not only This strategy 0.6590 Embedded guarantees the recognition rate, but also is superior to other algorithm inplatform time efficiency. Literature [3] 0.9867 Embedded platform Literature [7] 0.8923 Embedded platform Table 7. Comparison of Recognition performance of each cognitive strategy.

Method This strategy Literature [3] Literature [7]

Processing Time/s 0.6590 0.9867 0.8923

Realization Method Embedded platform Embedded platform Embedded platform

Algorithms 2018, 11, 129

19 of 20

4. Conclusions In recent years, the research on detection and identification algorithms of optical fiber vibration signals has made great progress. Scholars have proposed many algorithms and strategies, but most of them are in the simulation stage. Based on this, this paper presents a fast detection and identification algorithm for optical fiber intrusion signals that can be implemented on embedded platforms. The algorithm has been implemented on a DSP-based hardware board and actually tested the algorithm in Mentougou District, Beijing. Several common intrusion vibration signals are detected and identified. The results show that the detection performance is good, and the average recognition rate is above 90%. The threshold training of this algorithm is pre-processed on the personal computer (PC) side, so it saves a lot of time when implemented on the embedded platform, and the application of the parallel method can further increase the running speed. The specific time consumption is 0.659 s, which proves the effectiveness of the algorithm proposed in this paper. Author Contributions: Conceptualization, Z.S. and Y.Z.; Methodology, Z.S.; Software, D.Y.; Validation, D.Q. and Y.Z.; Formal Analysis, D.Q.; Investigation, Y.Z.; Resources, D.Q.; Data Curation, D.Y.; Writing-Original Draft Preparation, D.Q.; Writing-Review & Editing, D.Q.; Visualization, Z.S.; Supervision, Y.Z.; Project Administration, Y.Z.; Funding Acquisition, Y.Z. Funding: This research was funded by [the National Natural Science Foundation of China] grant number [61571014 and 61601006]. Acknowledgments: The authors wish to express their gratitude to the anonymous reviewers and the associate editor for their rigorous comments during the review process. In addition, authors also would like to thank experimenters in our laboratory for their great contribution on data-collection work. They are Sun Chengbin, Tan Lei, Feng Tingliang. This work was supported by the National key R&D Program Project [grant No. 2017YFB1201104]. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.

Liu, K.; Chai, T.J.; Liu, T.G.; Jiang, J.F.; Chen, Q.N.; Pan, L. Multi-area optical perimeter security system with quick invasion judgement algorithm. J. Optoelectron. Laser 2015, 26, 288–294. (In Chinese) Jing, K.; Zhi-Hong, Z. Time prediction model for pipeline leakage based on grey relational analysis. Phys. Procedia 2010, 25, 2019–2024. [CrossRef] Liang, W.; Lu, L.; Zhang, L. Coupling relations and early-warning for ‘equipment chain’ in long-distance pipeling. Mech. Syst. Signal Process. 2013, 41, 335–347. [CrossRef] Liang, W.; Zhang, L.; Xu, Q.; Yan, C. Gas pipeline leakage detection based on acoustic technology. Eng. Fail. Anal. 2013, 31, 1–7. [CrossRef] Zhang, T.; Tan, Y.; Yang, H.; Zhao, J.; Zhang, X. Locating gas pipeline leakage based on stimulus-response method. Energy Procedia 2014, 61, 207–210. [CrossRef] Zhu, C.H.; Qu, Y.Z.; Wang, J.P. The vibration signal recognition of optical fiber perimeter based on time-frequency features. Opto-Electron. Eng. 2014, 41, 16–22. Qu, H.; Zheng, T.; Pang, L.; Li, X. A new detection and recognition method for optical fiber pre-warning system. Optik 2017, 137, 209–219. [CrossRef] Zhang, R.; Sheng, W.; Ma, X. Improved switching CFAR detector for non-homogeneous environments. Signal Process. 2013, 93, 35–48. [CrossRef] Weinberg, G.V. Management of interference in Pareto CFAR processes using adaptive test cell analysis. Signal Process. 2014, 104, 264–273. [CrossRef] Shi, B.; Hao, C.; Hou, C.; Ma, X.; Peng, C. Parametric Rao test for multichannel adaptive detection of range-spread target in partially homogeneous environments. Signal Process. 2015, 108, 421–429. [CrossRef] Tao, P.; Yan, F.; Liu, P.; Peng, W.; Li, Q.; Feng, T. Fiber-optic intrusion recognition system based on Mach-Zehnder interferometer. Chin. J. Quantum Electron. 2011, 28, 183–190. Wang, S. M-Z interferometer Optical Fiber Sensing System of distributed disturbance pattern recognition method. Infrared Laser Eng. 2014, 43, 613–618.

Algorithms 2018, 11, 129

13. 14. 15. 16. 17.

20 of 20

Qu, H.; Ren, X.; Li, G.; Li, Y.; Zhang, C. Study on the algorithm of vibration source identification based on the optical fiber vibration pre-warning system. Photonic Sens. 2015, 5, 180–188. [CrossRef] Qu, H.; Liu, S.; Wang, Y.; Li, G. Approach to identifying raindrop vibration signal detected by optical fiber. Sens. Transducers 2013, 160, 85–92. Qin, Z.; Chen, L.; Bao, X. Continuous wavelet transform for non-stationary vibration detection with phase-OTDR. Opt. Express 2012, 20, 20459–20465. [CrossRef] [PubMed] Hui, X.; Zheng, S.; Zhou, J.; Chi, H.; Jin, X.; Zhang, X. Hilbert-Huang Transform Time-Frequency Analysis in phi-OTDR Distributed Sensor. IEEE Photonics Technol. Lett. 2014, 26, 2403–2406. [CrossRef] Zhu, C.; Zhao, Y.; Wang, J.; Li, W.; Zhang, Q. Ensemble Recognition of Fiber Intrusion Behavior Based on Blending Features. Opto-Electron. Eng. 2016, 43, 6–12. © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).