Field Deployment of Natural Gas Pipeline Pre-Warning System With CEEMDAN Denoising Method

This work utilizes the CEEMDAN algorithm to analyze the interference of Rayleigh back-scattering signals in standard communication optical fibers. The technology has several advantages, such as anti-electromagnetic interference, improved electrical insulation, corrosion resistance, higher sensitivity, and the capability for long-distance monitoring. In this study, in-situ monitoring data from a 53.2 km natural gas pipeline in a terrain area in Southwest China were analyzed. The results demonstrate that, using the CEEMDAN algorithm for a blind test conducted over fourteen days, a 100% recognition accuracy for mechanical tamping and a Nuisance Alarm Rate (NAR) of less than 1% were achieved.


I. INTRODUCTION
N ATURAL gas pipelines are one of the most cost-effective means of energy transportation [1].However, natural gas leaks [2] can pose a serious threat to the state, society and the environment [3].Therefore, a potential leak forecasting system for pipeline operations is necessary to detect and limit the consequences of leaks at an early stage [4].Fiber optic sensors are another widely used technology for natural gas pipeline monitoring.Gong et al. [5].developed an array of fiber optic pressure sensors that measure instantaneous pressure and wave direction information in pipelines, improving the ability to detect and locate pipeline defects.However, this technique has low detection accuracy in long-distance pipelines due to wave energy attenuation.Ren et al. [6] invented a fiber Bragg grating (FBG) annular strain sensor that can measure pipe circumferential strains caused by internal pressure changes and detect leaks.However, these optical methods are point-based detection techniques and cannot locate potential leaks.
A promising pipeline monitoring technique that has been widely investigated by researchers is the phase-sensitive optical time-domain reflectometer (Φ-OTDR), which uses the de-ployed optical fiber cable parallel to the buried pipeline as a distributed micro-vibration sensor [7].The sensing principle is based on the coherent Rayleigh backscattering effect in standard telecommunication fibers [8].To improve the recognition rate and reduce the false alarm rate, Adeel et al. [9] proposed an adjacent Spearman correlation assisted algorithm that extracts the impact region of the perturbation without involving idle data region and reduces the number of data traces in the pattern recognition process.Wu et al. [10] proposed an intensity and phase stacked convolutional neural network (CNN) that utilizes both the intensity and phase information from a distributed acoustic sensing (DAS) with coherent detection, achieving a classification accuracy of 88.2% with 1 km sensing length.Wang et al. [11] proposed a combination recognition method by using empirical mode decomposition (EMD) and extreme gradient boosting (XGBoost) to identify interference events and reduce the interference alarm rate.The recognition accuracies for five threat events were all above 90%.Shuqing et al. [12] proposed a pipeline leak detection method based on Hilbert-Huang Transform (HHT) and approximate entropy, which achieved higher detection accuracy than the traditional methods.However, these recognition results were only satisfactory for discerning some significantly different event signals in laboratory environment, not for various unknown complicated interferences in practical applications.To address this issue, Javier Tejedor et al. [13], [14] presented a pipeline integrity threat detection and identification system that employs DAS + PRS technology and evaluated it on realistic field data, opening the path to future improvements towards fully-functional pipeline threat detection systems operating in real conditions.However, most of the analyses were based on time domain data, and real-time frequency analysis was rarely mentioned.
In this paper, we propose a Natural Gas Pipeline Prewarning System (NGPPS) based on the coherent Rayleigh back-scattering principle in standard telecommunication fiber and use real-time FFT analysis in frequency domain to monitor potential threats along the natural gas pipeline.We analyze and demonstrate the in-situ monitoring data and results of a 53.2 km natural gas pipeline in a terrain area in Southwest China.We achieve 100% recognition accuracy for illegal mechanical tamping and a Nuisance Alarm Rate (NAR) of less than 1% using Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) noise reduction method for blind test over fourteen days.

II. SENSING PRINCIPLE
The NGPPS discussed in this paper consists of two parts: a Distributed Optical Fiber Micro-Vibration System (DMVS), which acquires and locates environmental vibration data along the sensing optical fiber; and a pattern recognition software, which uses the collected data for threat event identification.The basic principle of DMVS is to detect the phase change induced by the coherent Rayleigh scattering from two sensing points along the optical fiber, as shown in Fig. 1.The disturbance can be monitored at the receiving end by observing the interferogram intensity of the Rayleigh back-scattered signal.The detailed description of the sensing principle can be found in [15], [16].
The DMVS used in this paper is a self-developed vibration sensing equipment, whose schematic configuration is shown in Fig. 1.The equipment outputs a continuous light wave from a Narrow Line-width Laser (NLL) with a pulse width varying from 50 ns to 200 ns, corresponding to 5 m to 20 m spatial resolution.The light is modulated at a repetition of 1 kHz by a commercially available Acoustic-Optic Modulator (AOM) and amplified by an Erbium-Doped Fiber Amplifier (EDFA), then injected into the sensing optical fiber through an optical circulator.The Rayleigh back-scattered light along the optical fiber under test (FUT) is amplified and transformed into electrical signals by a Photo Detector (PD) with a bandwidth of 150 MHz.The pulsed electrical signals are collected by a 14-bit data acquisition card (PCIe-9802DC of JiANI Technology) with a maximum sampling rate of 250 MSa/s.The card has an on-board FPGA function that performs FFT and RMS algorithms.The processed results are then sent to an industrial computer for automatic feature extraction and threat event recognition.

A. Background and NGPPS set-up
The test site is located in Southwest Sichuan Province, where a 53.2 km pipeline operated by PetroChina transports purified gas with an operating pressure of 4.5-5.5 MPa.This pipeline is located in an active seismic zone in southern Sichuan, which has a subtropical humid monsoon climate with sufficient rainfall.As a result, earthquakes of different magnitudes often occur locally.Therefore, the pipeline buried in the slope region is often exposed to the risk of being torn by strong stress when the mountain collapses due to earthquakes or rainstorms.Moreover, due to the scarcity of land resources in southwest mountainous areas, the risk of illegal pipeline occupation and third-party damage is frequent, but it is difficult to detect small construction activities near the pipeline in time.
To address these challenges, we developed and used a Distributed Optical Fiber Micro-Vibration System (DMVS) to measure the Rayleigh back-scattered signal change of incident light in optical fiber due to external vibration when a threatening event occurs near the pipeline, thus obtaining the vibration signal near the pipeline.Through a unique identification algorithm, the vibration signal can be qualitatively analyzed, and the pipeline threat events can be found, tracked, and located in time.Fig. 2 shows the schematic set-up of the Natural Gas Pipeline Prewarning System (NGPPS).The right part of Fig. 2(a) shows the monitoring pipeline GIS road map of 53.2 km in total.Optical fiber cables with a length of ∼55 km are buried a few centimeters away and parallel to the pipeline, which is deployed across highways, hills, farmland, villages, and valleys, etc.The whole system is installed in the center control room at the pump station 1 and mainly consists of UPS, DMVS, industrial computer, and a server.DMVS is mainly responsible for real-time monitoring of the data information acquired by the optical fiber cable and extracting the features for pattern recognition.In the case of mechanical tamping or other events, the program alarm interface will indicate the exact alarm location, the frequency of event occurrence, and the event type, etc.
Fig. 3(a) shows the relationship between the intensity of Rayleigh back-scattered light and the length of the optical fiber cable along the natural gas pipeline.It can be seen that without perturbation, the output amplitude of Rayleigh back-scattered signal remains unchanged for the consecutive 55 km.However, once a continuous vibration is applied at the 8.25 km along the optical fiber length, the accumulated RMS energy can be seen from Fig. 3(b).It is clear that at the vibration point, the amplitude of RMS signal is much larger compared with other no vibration points along the sensing fiber.

B. Feature Extraction Based on Complete Ensemble Empirical Mode Decomposition With Adaptive Noise (CEEMDAN)
Empirical Mode Decomposition (EMD) is an adaptive and efficient method applied to decompose vibration and acoustic signals in time domain [17].It aims to extract a series of Intrinsic Mode Functions (IMFs) from the target signal by sifting stage by stage.However, mode mixing often occurs during the decomposition of threat events, which affects the accuracy and reliability of the results.Therefore, Ensemble Empirical Mode Decomposition (EEMD) [18] is proposed to denoise the original vibration signals and address the problem of EMD.EEMD sifts an ensemble of white noise-added signals and treats the mean as the final true result by adding the white noise with sufficient number of trials.EEMD method is suitable for the analysis of nonlinear and non-stationary signals as it is a noise-assisted data analysis method.However, it introduces a new problem as the noise of EEMD signal may produce different number of modes.
In terms of the application proposed in this paper, the vibration signal along the monitoring pipeline extracted from DMVS usually contains nonlinear noise.The mode aliasing issue often arises by the EMD/EEMD algorithm when the original vibration signal is superposed with white Gaussian noise.Therefore, a new denoising method called Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) [19] represents a substantial improvement over the original EEMD.It provides an exact reconstruction of the original signal and better spectral separation of modes with low computational cost.Fig. 4 shows the general flow chart of CEEMDAN process.The benefit of the decomposition using the CEEMDAN is that the added white noise series cancel each other in the final mean of the corresponding IMFs, thus significantly reducing the chance of mode mixing.The general process of CEEMDAN algorithm can be expressed with the following mathematical equations.Suppose the original vibration signal acquired by DMVS is x[n], w[n] is the white noise and n is the sampling number.x[n] is first decomposed by EMD I realizations to obtain their first mode IMF1[n], and compute: Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.The first residue can be expressed as: IMF * 1 is obtained in the same way as in EEMD process.After that, the first EMD mode is computerized over an ensemble of r 1 [n] plus different realizations of a given noise obtaining IMF2 by averaging.The next residue is defined as: The process continues with the rest of the modes until the stopping criterion is reached.As can be seen from Fig. 4, for k = 3, …K calculated the k th residue: The decomposition is completed The final residue meets the following expression: Where K is the total number of modes.In this case, the final vibration signal acquired by DMVS can be given: Eq. ( 6) is the reconstruction of the original data.The results achieved by the CEEMDAN depend on the choice of the ensemble number (N) and the amplitude of added noise (AM) and their relationship can be expressed by the following equation: where ε is the final standard deviation of error calculated as the difference between the original signal and the sum of the IMFs resulting from the EEMD.
In the pipeline pre-warning application environment, various threat events can occur along the pipeline.Fig. 5, 6, and 7 show the decomposition results of three major threat events:  mechanical tamping, manual digging, and truck passing by CEEMDAN respectively.To extract the features of different threat events along the pipeline, we use a radar map to depict the intensity distribution of eigenvalues of each threat event.Each event is a time series divided into a length of 1024 sampling points.The eigenvalues listed in Table I of six different kinds of threat events are processed by algorithm to eliminate the noise and reconstruct the time series signal.It can be clearly seen from Fig. 8 that the eigenvalues of most of the threat events are different.We first investigate the two major threat events along  the pipeline: mechanical tamping and manual digging.From Fig. 8, it is obvious that mechanical tamping has the largest amplitude in terms of energy, since tamping-induced ground vibration can be easily acquired by DMVS.In the frequency domain, it also has a larger mean frequency peak compared with other five events.From Fig. 8, the calculated kurtosis number is larger than other events.
Therefore, these eigenvalues are sufficient for us to determine the mechanical tamping event, and give an alarm once such event happens along the pipeline.On the other hand, noise event produced by environmental fluctuation has the smallest value for each eigenvalue, thus it is easy to determine whether a potential threat occurs or not.It is more difficult to identify the manual digging event, since some of the eigenvalues of manual digging are quite close to truck passing signal.In this case, we employ the valley factor to avoid the interference of truck passing signal against manual digging, as shown in Fig. 8.Although the vibration duration of truck passing and manual digging in time domain are quite close to each other, as seen from Figs. 6  and 7, the intensity of truck passing signal is much larger than manual digging.The valley factor of truck passing is 29, while this number for manual digging is only 11.
Figs. 9 and 10 show the eigenvalues radar map of five major threat events along the monitoring pipeline: excavator, mechanical tamping, truck passing, water flowing, and manual digging.The eigenvalues of each threat event within the processed time frame (1024 points) are calculated five times to ensure the repeatability of each event.It can be seen that for a specific threat event, the calculated eigenvalues from Table I are relatively stable after repeating five times, and only a small fluctuation is observed for each time.The results indicate that the selected eigenvalues can effectively represent the threat events.

C. Signal Recording and Labeling
As mentioned in the previous section, we used an active natural gas pipeline operated by PetroChina for field experiment in a real scenario.We monitored the activities near the pipeline by sensing the optical fiber cable deployed parallel to the pipeline with an average depth of 1.5 meters and a distance of ∼1 meter along the whole pipeline.To monitor the potential threat events and discriminate them from other environmental signals, we recorded different kinds of activities by DMVS during fourteen consecutive days at five measurement zones with different environmental conditions listed in Table II.The location difference of the threat events enabled us to test the system behaviors under different environmental conditions.The most hazardous threat events are third-party interventions for pipeline monitoring purposes.These events mainly refer to mechanical tamping, excavator digging, and manual digging.Fig. 11 shows typical nuisance alarm events such as truck passing from the expressway and water flowing.These events usually have an impact on the correct recognition of the system performance.Therefore, it is necessary to analyze the signal characteristics of these events and reduce the NAR of the whole system.The bandwidth of the acquired vibration signals from DMVS covers frequencies up to 512 Hz, determined by the repetition rate of the NLL.However, we carried out experiments by analyzing frequencies up to 200 Hz, since frequencies of most of the monitored events are below 200 Hz and do not convey meaningful information.We set the low limit of the spectral range to 1 Hz, since the window size in the FFT is 1024.The relevant parameters related to the energy-in-bands computation in the feature extraction are: the acoustic frame size (which is set to 1 second in the system), the acoustic frame shift (set to 10 milliseconds), and the number of FFT points (set to 1024) for time frame analysis.We chose these values based on their best performance with a few trial experiments.We conducted classification on a frame-by-frame basis.Therefore, we calculated the feature vectors mentioned in Section III-B for each 1024 points within every 60-second length recorded file.We used all the vectors in the pattern classification stage, either as training or testing data.
IV. DISCUSSION Fig. 12 shows the real NGPPS installed at the control station of the monitoring pipeline.We also investigated the radial monitoring range of the sensing cable, since it is of great significance to predict the pipeline intrusion event in advance.Fig. 13(a) depicts a schematic cross-section view of the radial sensing range of NGPPS, while (b) shows the radial monitoring range of mechanical tamping vertical to the 55 km optical fiber cable using the traditional EMD and the CEEMDAN algorithm.CEEMDAN denoising extends the radial sensing range by more than ten meters compared with EMD method within the first 40-45 km sensing cable.However, as the sensing distance continues to increase, the radial sensing range rapidly decreases due to the effect of optical power attenuation.Fig. 14 shows the specific algorithm flow of feature extraction using CEEMDAN denoising algorithm and recognition of event types by BP.The confusion matrix obtained from Fig. 15 is based on the training data mentioned above.During the blind test, we selected seven major events that occur occasionally 1 The buried natural gas pipeline buried closed to a river. 2 GIS map of the buried natural gas pipeline with different geographic characteristics.3 The buried natural gas pipeline across a highway.4 The natural gas pipeline buried crossing a village.along the pipeline to analyze the event recognition rate and NAR among each threat event.By using CEEMDAN denoising algorithm, we achieved above 91% recognition rates for all the seven events due to the improvement of the SNR.Regarding the major threat events such as excavator and mechanical tamping, we identified all of them correctly during the blind test period.Manual digging, mechanical tamping, and excavation are three  common events of third-party intervention.Before using CEEM-DAN, 6% of excavator event and 2.6% of truck passing noise were recognized as manual digging, while after CEEMDAN denoising, only 2.2% of truck passing noise were recognized as manual digging.In terms of excavator and mechanical tamping, the NAR was reduced from 4.4% to 0% by using CEEMDAN denoising.

V. CONCLUSION
The NGPPS proposed in this paper uses the interference effect of Rayleigh backscattered light wave in communication optical fiber to detect the phase change of light wave caused by external vibration.We use CEEMDAN denoising algorithm to improve the SNR of the potential threat events and increase the radial monitoring range up to 35 meters with a sensing distance of 55 km.Compared with EMD, our method achieves 100% recognition accuracy for manual digging and mechanical tamping during the blind test.At the same time, we achieve a low NAR of less than 1% for both manual digging and mechanical tamping events.

Fig. 2 .
Fig. 2. (a) Field set-up of NSPPS with DMVS installed in one of the pump station with pipeline route map shown in the right.(b) Signal processing procedures of NSPPS to monitor potential threat events, including data acquisition, pattern recognition, field blind test and alarm platform.

Fig. 4 .
Fig. 4. The schematic flowchart of CEEMDAN algorithm to denoise the vibration signals generated by different threat events along the pipeline.

Fig. 8 .
Fig. 8. Contrast radar map for seven eigenvalues distribution of different threat event.

Fig. 9 .
Fig. 9. Radar map shows the six different events.Each map shows the intensity distribution of eigenvalues which are calculated for five times to demonstrate their repeatability.

Fig. 10 .
Fig. 10.Radar map shows the six different events.Each map shows the intensity distribution of eigenvalues which are calculated for five times to demonstrate their repeatability.

Fig. 11 .
Fig. 11.1 The buried natural gas pipeline buried closed to a river. 2 GIS map of the buried natural gas pipeline with different geographic characteristics.3The buried natural gas pipeline across a highway.4The natural gas pipeline buried crossing a village.

Fig. 12 .
Fig. 12. Picture of NSPPS installed at the pipeline control station.

Fig. 13 .
Fig. 13.(a) Cross-section diagram of radial monitoring range of optical fiber cable by NSPPS (b) Radial monitoring range of mechanical tamping vertical to the pipeline direction using EEMD and CEEMDAN algorithm respectively.

TABLE I EIGENVALUES
USED TO RECOGNIZE DIFFERENT THREAT EVENTS