A phase-sensitive optical time-domain reflectometer (φ-OTDR) system is a front monitoring and early warning technology that can acquire the location of disturbances in space and phase information of disturbances in time. With the advantages of high resolution, wide monitoring range, and strong anti-interference capability, this technology has been widely used in pipeline safety maintenance, intrusion warning, and large-equipment monitoring. However, due to the complex diversity of the application environment, the system suffers from low recognition accuracy and insufficient stability in actual use, particularly when similar signals are recognized in the system application. To solve these problems, this study proposes a similar-signal recognition method based on multiscale feature fusion. This method can effectively improve the recognition accuracy of similar signals while maintaining the recognition accuracy of the base signal.

The original signal is first decomposed into sub-signals in different frequency ranges using empirical mode decomposition (EMD) and wavelet packet decomposition (WPD). The original signal and individual sub-signals are then subjected to time-frequency feature extraction and approximate entropy feature extraction. The time-frequency features are used to evaluate the details of the time and frequency variations of the signal, the approximate entropy features are used to evaluate the complexity and regularity of the signal, and the multiscale signal decomposition and multi-feature extraction are used to amplify the feature differences between similar signals. Because the multiscale and multi-feature approach increases the dimensionality of the data, the proposed method utilizes principal component analysis (PCA) to combine high-dimensional features and reduce the dimensionality of system features, thereby improving system efficiency. Finally, the fused features are passed into a lightweight back-propagation (BP) neural network as input variables for signal data processing. Compared to other traditional neural networks, BP neural networks have the advantages of lightweight structures and high speed, enabling them to process signal data quickly.

Sub-signals decomposed by EMD and WPD have multiscale characteristics ranging from low to high frequencies. Each sub-signal contains a part of the signal domain within the main frequency-band range of the original data. Decomposition helps to amplify the feature gaps between different signals and facilitates subsequent multidimensional feature extraction (Fig. 10). Following feature extraction and fusion, the four signals show significant differences in the feature space. Thus, even with a simple classifier, signal classification and recognition can be achieved (Fig. 11).

A comparison among extracting multi-features from original signal [Fig. 12(a)], the CNN model [Fig. 12(b)], and the multi-scale feature fusion[Fig. 12(c)] reveals that the multi-scale feature fusion has higher recognition accuracy, where knocking and shaking-signal recognition accuracies reach 100% and trolleying and walking-signal recognition accuracies reach 98.5% and 98.0%, respectively. A comprehensive analysis reveals that the comprehensive recognition accuracy of the proposed method is increased by 8.4 and 9.0 percentage points over extracting multi-features from original signal and CNN model, respectively, and the similar-signal recognition accuracy is increased by 13.5 and 12.4 percentage points (Fig. 13), respectively. These results verify that the method has high recognition accuracy.

Experimental results show that the decomposition method using EMD combined with WPD can obtain sub-signals at different scales. The time-frequency domain and approximate entropy features can in turn be extracted from the original signal and sub-signal to enhance the differentiation of similar-signal features more effectively. The PCA algorithm can then reduce the dimensionality of high-dimensional data, thus effectively reducing the number of training features. A well-designed six-layer lightweight BP neural network model can also effectively identify different types of signals when identifying signal features with significant differentiation. Compared with the extraction of features directly from the original signal, the proposed method can improve the integrated and similar-signal recognition accuracies by 8.4 and 13.5 percentage points, respectively. Compared to those of the CNN method, the overall recognition accuracy is improved by 9.0 percentage points, and the similar-signal recognition accuracy is improved by 14.3 percentage points. This method effectively improves similar-signal recognition while maintaining the recognition accuracy of underlying signals, which is of great value for expanding the applications of φ-OTDR systems.