Since pipelines are the primary route for delivering natural gas, maintaining pipeline safety during industrial operations is essential [1].

Currently, many methods have been developed for the detection of gas pipeline leaks. The mainstream detection methods include the mass/volume balance method [2], the flow monitoring method [3], the statistical decision method [4], the transient model method [5], and the acoustic detection method [6]. Compared with other signals, acoustic signals can achieve non-contact detection. Distributed feedback fiber laser vibration sensors (DFB-FL) have the characteristics of narrow line width, small size, single frequency single mode, and anti-electromagnetic interference, and are suitable for sensing physical quantities related to acoustic vibration [7]. Therefore, this article uses DFB-FL for gas pipeline leak monitoring [8]. Efficient event recognition of acoustic signals obtained by DFB-FL has become an important research topic [9,10].

Researchers attempt to find leaks in the field of pipeline leakage detection by using machine learning techniques. At the moment, support vector machines (SVM) [11], random forests algorithms (RF) [12], and K-Nearest Neighbor (KNN) [13] are the most widely utilized machine learning techniques. Qu et al. [14] proposed a pipeline leakage detection warning system based on SVM, using SVM to identify the characteristic signals of pipeline leaks. Quy et al. [15] proposed an acoustic emission signal feature classification model based on the KNN algorithm for detecting leaks in natural gas pipelines. However, machine learning methods often rely on hand-crafted features [16], such as time-domain and frequency-domain features [17,18]. Because these methods only extract superficial features from the data and require human intervention, they have an impact on the classification accuracy of the signal.

Deep Learning techniques have been widely applied in the field of vibration event recognition because they may eliminate the need for human-powered feature extraction for vibration signal characteristics in traditional methods [19]. It has been demonstrated that Convolutional Neural Network (CNN) is crucial for fault diagnosis and leakage detection [20]. Lu et al. [21] proposed an intrusion pattern recognition scheme based on the Gramian Angular Field (GAF) and CNN. The method converts the vibration signal into a GAF graph and then combines it with CNN to complete the classification of intrusion events. Rahimi et al. [22] proposed a predictive scheme for tank leaks based on the combination of the Fast Fourier Transform (FFT) and CNN, which uses FFT to convert the signal into an image and then uses CNN to predict tank leaks. In order to categorize signals, these methods primarily turn them into images. However, these extra processes may result in the loss of feature data in addition to complicating the detection procedure. As a result, researchers have started to focus on studying how to use one-dimensional Convolutional Neural Networks (1DCNN) to finish the signal categorization task. Yang and Kang et al. [23,24] construct a 1DCNN to extract features from leakage signals to improve the accuracy of pipeline leakage detection. These methods primarily use 1DCNN for feature extraction, followed by other classifiers for data classification. The testing time will rise as a result. Multi-scale networks can increase the identification rate of the network by extracting information from several signal frequency bands. It is widely used in many fields. Li et al. [25] employed multi-scale networks to predict the remaining useable life (RUL). Liu et al. [26] designed a multi-scale convolutional neural network (MSCNet) to improve the model's feature representation ability and use it in the medical industry. Yu et al. [27] suggested a network security communication intrusion detection system using a multi-scale convolutional neural network. In order to extract multi-scale features and fuse features between channels to gain more detailed information, this article will employ a multi-scale network.

Based on a DFB-FL vibration sensor, the neural network model (mCNN-LFLBs) is suggested that combines local feature extraction blocks (LFLBs) with multi-scale convolutional neural networks (mCNN). The multi-scale structure is created by paralleling three 1DCNNs with different convolutional kernels, which enables the model to adaptively learn different scale features of the signal. The LFLBs further extract local features, and LFLBs can not only improve the local representation ability of the model and enhance its expressive power but also alleviate the problem of overfitting and reduce the computational burden. Compared with traditional methods, the original one-dimensional time-domain signal is taken as input, and relevant features can be extracted from the time-domain signal without human intervention. In addition, the leakage hole diameter of the pipeline can be distinguished while the leakage event is detected. Compared with the traditional methods, the model not only has a faster convergence speed but also has an average recognition rate of 98.6% and has more obvious advantages in the judgment accuracy of the leakage aperture size.