Threat event recognition is one of the widely researched topics in distributed fiber optic sensing. Deep learning is an important means for pattern recognition. The main challenges that limit its recognition accuracy can be categorized into two aspects: lack of generalization and existence of false recognition for some signals with low vibration intensity and obscure features. On one hand, this is due to ambiguous features, such that the target signal is often obscured in noise, and on the other hand, such signals are easily mislabeled in the process of constructing data sets. The classification accuracy of neural networks can be improved in three ways. The first approach is to preprocess the data from the front end of the network by applying various methods, such as band-pass filtering, wavelet denoising, and Hilbert transform. However, these methods have relatively limited positive effects and will contribute to the loss of detailed information to some extent. The second approach is to increase the extraction of features from training samples, such as inputting multiple features of the signal into the network simultaneously to improve the recognition accuracy through feature fusion. The third approach is to increase the means of feature extraction through various methodologies, such as increasing the number of convolutional layers, introducing recurrent neural networks (RNNs), and supplementing deep belief networks. The design of specific schemes should consider the data characteristics of the sensing system. In this study, the sampling rate and sampling points of the Mach-Zehnder interferometer (MZI) are 50 times greater than those of the phase-sensitive optical time-domain reflectometer. Thus, the third approach will substantially increase the computations and response time. In this paper, we attempted to implement the enhancement of vibration features from the perspective of signal sources for improving the recognition accuracy of weak vibration events under the traditional network framework.

The proposed recognition program employs the conventional waveform, frequency spectrum, and time-frequency spectrum feature extraction method; further, RNNs are not introduced in this method to avoid excessive computations. We optimized the conventional recognition strategy from the perspective of sample sources. We used empirical mode decomposition (EMD) to decompose the signal into multiple intrinsic mode functions (IMFs). The target signal exhibits clearer vibration features on certain IMFs. Specifically, the vibration part shows sharp pulses separated from the fundamental noise that exhibit higher contrast than the response pulses at the same location of the original waveform (Fig. 4). Furthermore, the frequency spectrum of the IMF shows distinct morphological features with suitable same-class consistency and inter-class differentiation (Fig. 5). These features are independent of the signal strength, which provides a feasible premise for the effective identification of weak vibration events. Based on the statistics of a large number of samples, IMF₂ and IMF₃ are selected to extract their waveform features and time–frequency spectrum features and forIMF₂-IMF₅ to extract their frequency features. The neural network consists of convolutional operators and fully connected networks. The waveform and frequency features are extracted using one-dimensional convolutional operators, while the time-spectrum features are extracted using two-dimensional convolutional operators (Fig. 6). In addition, based on the same feature extraction framework, four control groups were designed with the original signal as the sample source (Table 1).

By comparing val_accuracy, val_precision, val_recall, and other parameters (Fig. 8, Fig. 9, Table 2, Table 3, Table 4, and Table 5), the proposed model was found to be the best in terms of recognition accuracy, stability, and generalization. Although all the models exhibited high accuracy when using the training set, the metrics in the test and validation sets showed a significant decrease compared with the training set. Only the validation accuracy of Model 1 (our strategy) exceeded 90%, which remained relatively stable. The classification accuracy of the model for specific events can be characterized using precision, recall, and F1_score; all three parameters are greater than 90% in Model 1. The average response time of the proposed model is less than 0.07 s, indicating good feasibility and development space.

In this paper, we have proposed a classifier Model 1 based on EMD and convolution operators using the Mach–Zehnder interferometer (MZI) as the sensing system. Instead of the original signal, IMF₂ and IMF₃ were selected to extract contour features and time–frequency domain features, and IMF₂-IMF₅ were selected to extract the frequency spectral features. Four control groups were developed based on the same feature extraction framework with the original signal as input. The test_accuracy and val_accuracy of Model 1 are 97.02% and 94.88%, respectively, while the val_accuracy of the remaining control groups is less than 90%. The average val_precision and val_recall of Model 1 for the five vibration events are 95.51% and 94.42%, respectively. In particular, for two weak vibrations, Event 4 and Event 6, the recognition accuracy of Model 1 exceeded those of other control groups, thereby fully demonstrating the optimization of this scheme in terms of generalization and vibration recognition. This paper improves the conventional deep learning network from the perspective of sample sources; however, it does not change the conventional feature extraction, and further extension of the feature extraction dimension is required in the subsequent development stage to address the shortcomings of the existing classifier. In addition, the selection of IMF components in this study is based on the target signal features and the statistical results of large batch samples, and the selection in practical applications need not be limited to this paper.