Pipelines play an indispensable role in modern urban construction, including urban water supply and drainage systems, and oil and gas transmission systems. With increasing use over time, there will be a variety of damage to the pipeline, and thus leakage. Real-time monitoring of pipelines is very important. The current pipeline leak detection methods mainly include the negative pressure wave detection method [1], the pressure gradient monitoring method [2], the mass balance monitoring method [3], the acoustic wave detection method [4, 5], etc. These methods have problems such as generating false alarm signals, poor positioning accuracy, and not being applicable to small-scale leaks. The distributed fiber optic detection method is a new emerging pipeline leak detection method that has the advantages of strong resistance to electromagnetic interference, good insulation, corrosion and high voltage resistance, and being able to achieve long-distance detection [6–8]. Pattern recognition of the collected fiber optic vibration signals is required for the fast and accurate location of buried pipeline leakage points.

The current identification of fiber optic vibration signals can be divided into traditional, machine learning-based, and deep learning-based methods. For the traditional method, Xu et al. [9] developed an image-based signal processing method. First, the differential operation of all detection points in the whole detection range is completed and the differential data is divided according to the threshold. The picture is formed by extracting the event region according to a mathematical morphological method. Finally, the existing line segments are extracted using Hough transform and the slope is analyzed to determine the event type. As for machine learning-based methods, Zhong et al. [10] proposed a pipeline vibration classification method based on the support vector machine (SVM) algorithm. By extracting features such as the mean, variance, mean square error and signal power of a signal in time and frequency domains, a classifier based on the SVM algorithm is built to determine the vibration and identify the vibration patterns. As for the deep learning-based approach, Wu et al. [11] used multi-scale wavelet packet decomposition as pre-processing and then used the processed signal to train a 1D convolutional neural networks-based (CNN-based) model for vibration sensing recognition. After transforming the time domain signal in each segment to the frequency domain by fast Fourier transform, Wen et al. [12] rotated the amplitude and frequency by 90 degrees and mapped the spectral amplitude to grayscale (0 to 255) values. The higher the amplitude, the darker the corresponding region. The time-frequency representation of the signal is called a spectrogram. The spectrogram is classified by a 2D-CNN (multi-classification SVM instead of softmax). Guo et al. [13] used a leakage spectrogram to characterize the leakage signal and established a time-frequency convolutional neural network model to identify the leakage signal.

In this paper, a new method is proposed for the classification of fiber optic vibration signals. The vibration signal is converted into a 64 × 64 grayscale map, and the idea of image classification is used to realize the classification and recognition of fiber optic vibration signals. The grayscale map is divided into a training set and a test set, and the training set is input into the residual network (ResNet) 50 network for training to obtain a network model, and the test set is input into the model to achieve classification recognition. The method proposed in this experiment is applied to data from actual experiments, and the recognition accuracy obtained reaches 94.59%, which can accurately and efficiently realize the classification and recognition of fiber optic vibration signals.

2.1. Principle of Distributed Fiber Optic Vibration Sensing System

Light scattering in optical fiber produces various types of scattered light. The polarization, intensity, phase, and wavelength characteristics of light are modulated by physical quantities such as temperature, pressure, and vibration of the external environment, so the sensing of these physical quantities can be achieved by detecting the optical characteristics of scattered light.

The distributed fiber optic vibration sensing system is shown in Fig. 1. The laser modulation module is mainly completed in three parts: Pulse modulator, erbium-doped fiber amplifier (EDFA) module, and optical circulator. The laser is used as the system’s light source and emits a continuous beam at the transmitting end to provide an ultra-narrow bandwidth optical signal for the system to work properly. The laser is modulated into an optical pulse signal after passing through the pulse modulator, and the erbium-doped fiber amplifier then amplifies the optical pulse signal before passing through the circulator into the sensing and conduction mode, which is composed of a single-mode fiber and completes all the functions of this module. When the sensing fiber senses a perturbation event, it generates a backscattered Rayleigh signal that is transmitted through the annulus to the signal reception and processing module. The signal reception and processing module is shown in the figure in the data acquisition board section, and generally consists of a combination of a photodiode, an analog-to-digital converter, and a computer. The signal scattered towards Rayleigh after passing through distributed optical fibers is first received by a photodiode, and then amplified by an amplifier. The amplified signal is converted from an optical signal to an electrical signal through an analog-to-digital converter, and finally the converted electrical signal is stored on a computer.

Figure 1. Distributed fiber optic vibration sensing system.

2.2. Principle of Fiber Optic Vibration Signal Classification Recognition

For fiber optic vibration signal classification, the strategy of this research is to first convert the dataset into a 64 × 64 grayscale map, and then input the grayscale map into the ResNet50 network for training. The fiber optic vibration signal classification process is shown in Fig. 2. A grayscale diagram corresponding to the original data is shown in Fig. 3. Because the frequency of the acquisition card is 4,096 Hz, it is considered to be converted into a grayscale image of 64 × 64 size. In addition, each pixel of the gray image corresponds to a value between 0 and 255, so it is necessary to normalize the original data and multiply it by 255. Value 0–255 represents the difference between light and dark in the grayscale map. The larger the value of the pixel, the brighter the pixel.

Figure 2. Process of fiber optic vibration signal classification recognition.

Figure 3. Data pre-processing. (a) Original data, (b) 64 × 64 grayscal map.

The model structure of the ResNet50 network is shown in Fig. 4, and the CONV (convolutional layer) block and ID block are extracted separately in the figure to make the whole model structure clearer. The input module of the ResNet50 network consists of CONV and max pool (maximum pooling layer), with ReLU (activation function) and Batch norm (batch normalization layer) to enhance the network fitting ability. The four residual modules contain two residual structures, CONV block and ID block. The dimensionality of the CONV block output is different from that of the input and can be used to change the dimensionality of the network; the dimensionality of the ID block output is the same as that of the input and deepens the network by cascading. The output module downsample the features through AVG pool (average pooling layer), and then changes the feature vector to batch size × 2,048 through flatten layer, and the extracted feature vector is used to achieve classification by the K-nearest-neighbors (KNN) algorithm, and the corresponding category is output.

Figure 4. Structure of ResNet50.

3.1. Data Set

In this paper, the damage fiber optic vibration signal dataset of a buried pipeline obtained from a China State Shipbuilding Corporation (CSSC) simulation of an actual situation was used. The data were collected with and without external noise interference, with three different fiber placement methods (as shown in Fig. 5, where gray cylinders indicate pipes and red lines are fibers): Laying wrapped around the outer wall of the pipe, laying through the pipe, and a combination of the two layouts. For each laying method, the fiber vibration signal is obtained by tapping the cavity, tapping the breakage of the pipe, and tapping the normal position of the pipe.

Figure 5. Three different methods of laying fibers.

3.2. Research on the Effect of Laying Methods

The data obtained for the three different fiber laying methods are trained without external noise interference to research the effect of laying methods. The recognition accuracies are shown in Table 1 and the confusion matrix of classification results is shown in Fig. 6.

Figure 6. Confusion matrix of classification results for three fiber laying methods without external noise interference.

TABLE 1. Accuracy of classification results for three fiber laying methods without external noise interference.

Methods	Wrap Around Laying	Laying Through the Pipe	Mixed Laying
Accuracy (%)	94.81	87.86	88.94

The data obtained for the three different fiber laying methods are trained with external noise interference to research the effect of laying methods. The recognition accuracies are shown in Table 2 and the confusion matrix of classification results is shown in Fig. 7.

Figure 7. Confusion matrix of classification results for three fiber laying methods with external noise interference.

TABLE 2. Accuracy of classification results for three fiber laying methods with external noise interference.

Methods	Wrap Around Laying	Laying Through the Pipe	Mixed Laying
Accuracy (%)	92.73	89.74	79.36

The dataset with external noise interference and the dataset without external noise interference are combined and trained for the three different fiber laying methods to research the effect of laying methods. The recognition accuracy is shown in Table 3, and the confusion matrix of the classification results is shown in Fig. 8.

Figure 8. Confusion matrix of classification results for three fiber laying methods in the mixed dataset.

TABLE 3. Accuracy of classification results for three fiber laying methods in the mixed dataset.

Methods	Wrap Around Laying	Laying Through the Pipe	Mixed Laying
Accuracy (%)	993.66	86.85	81.42

Analyzing the three groups of experiments, it becomes evident that the wrap around laying technique exhibits the highest recognition accuracy regardless of external noise interference. On average, the recognition accuracy of the wrap around laying method is improved by 5.58% compared to laying through the pipe and by 10.50% compared to mixed laying.

The effectiveness of wrap around laying can be attributed to the more comprehensive transmission of vibration signals to the fiber. The wrap around laying method ensures that the vibration signal is more fully conveyed, enhancing recognition accuracy. Conversely, the results obtained from the mixed layup could be much better. This suggests that there are discernible disparities in certain characteristic components, such as energy distribution, between the fiber vibration signals obtained from laying through the pipe and those acquired through wrap around laying.

3.3. Research on the Effect of Training Strategy

In addition to the fiber vibration signals obtained in the training process by tapping the cavity, tapping the leakage of the pipe and tapping the normal position of the pipe, an additional noise class is added to investigate the effect of the training strategy on recognition accuracy. The additional noise class is composed of a randomly selected amount of data from the three classes in the presence of external noise interference.

In the strategy of wrap around laying, the cases of without external noise interference, with external noise interference and mixed dataset are trained separately. The hybrid dataset consists of the dataset without external noise and the dataset with external noise. The recognition accuracy is shown in Table 4, and the confusion matrix of the classification results is shown in Fig. 9.

Figure 9. Confusion matrix of classification results for wrap around laying.

TABLE 4. Accuracy of classification results for wrap around laying.

Noise	Without External Noise	With External Noise	Mixed Dataset
No Noise Class Added (%)	94.81	92.73	93.66
Added Noise Class (%)	96.43	93.72	93.62

In Fig. 9 and Table 4, it can be seen that the recognition accuracy of the model with added noise class but without the interference of external noise was the highest at 96.43%, due to the interference of the energy feature component of external noise interference on the recognition accuracy, when both wrap around laying were used. The recognition accuracy of the model with noise class added increased by 0.86% on average compared with that of the model without noise class added. This is mainly due to the higher recognition accuracy of the noise class itself, as can be seen in Fig. 9, where the model’s recognition accuracy for noise is close to 100%, improving the model’s overall recognition accuracy. It also indicates that the discrimination between noise and the three types of percussive vibration is good, and the daily noise is generally not recognized as the percussive vibration state, which verifies that the method proposed in this paper has good robustness and is little affected by external noise.

In this paper, three collected distributed fiber optic vibration signals were converted into 64 × 64 grayscale images, and ResNet50 was used to classify grayscale maps corresponding to the three distributed fiber optic vibration signals. Considering the laying method of the pipe, it was found that the recognition accuracy of the fiber optic vibration signal obtained by the wrap around laying method is the highest, and the average recognition accuracy can reach 93.73%. After adjusting the training strategy, it was found that adding noise classes to the data could improve the recognition accuracy of the model, and the average recognition accuracy could reach 94.59% after adding noise classes. This is an improvement of 0.86% compared to the model without noise classes.

The authors acknowledge the support given by the National Natural Science Foundation of China (Grant No. 62275153, 61875125, 62005165) and the development fund for Shanghai talents (Grant No. 2021005). In addition, the authors want to express appreciation for the editor and anonymous reviewers for their valuable comments and suggestions for this paper.

National Natural Science Foundation of China (Grant No. 62275153, 61875125, 62005165); Development fund for Shanghai talents (Grant No. 2021005).

The authors declare no conflict of interest.

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.