Phase sensitive optical time-domain reflectometer (Φ-OTDR) system has a wide range of application scenarios, and the types of perceived signals are complex and varied. Therefore, research on the recognition method of Φ-OTDR signals is crucial. To improve recognition accuracy and achieve shorter recognition time, a pattern recognition method based on Markov transition field (MTF) and MobileNetV2 for the Φ-OTDR signal is proposed.

We first decompose the two-dimensional Φ-OTDR spatiotemporal signal into a set of one-dimensional signals, and use downsampling to shorten the length of the original signal and reduce the amount of data. Next, based on the MTF principle, the preprocessed one-dimensional signal is encoded into a two-dimensional image. This image encoding method has good noise resistance characteristics and can amplify and capture the time-domain features of one-dimensional signals. The encoded image is input into four lightweight neural network models for signal pattern recognition. The experimental results indicate that MobileNetV2 has the best recognition performance for encoded images. Finally, transfer learning methods are used to train the network model, effectively accelerating the convergence of the model and improving the recognition accuracy.

This method achieves high recognition accuracy and fast recognition speed, with an average recognition accuracy of 96.0% for six signals and recognition time of 0.2047 s for a single signal. Among the six signal modes, the method proposed in this paper has high recognition accuracy for the four signal modes of digging, knocking, watering, and walking.

Comparing the proposed method with the latest research on recognition, as shown in Table 3, it demonstrates advantages in average recognition accuracy compared to traditional convolutional neural network (CNN) methods and particle swarm optimization-support vector machine (PSO-SVM) based methods. The method presented in this paper demonstrates better classification performance for the four types of Φ-OTDR signals: digging, knocking, watering, and walking. These four signals exhibit a certain degree of suddenness in their temporal and spatial variations, corresponding to the appearance of dark block structures in MTF images. The background noise signal changes slowly, has a certain periodicity and long-term trend, and has weak dynamic characteristics. In the MTF image corresponding to the shaking signal, block features and line features are mixed, with rich details and a certain degree of confusion compared to other signal pattern images. Therefore, the classification difficulty of these two signals is relatively high.

Comparing the recognition speed of the method proposed in this paper with those in other studies, as shown in Table 4, the preprocessing time for a single signal in this paper is 0.1707 s, the recognition time for a single encoded image is 0.0340 s, and the total recognition time is 0.2047 s. Not only did it demonstrate the advantage of processing speed in the signal preprocessing and feature extraction stages, but it also showed faster recognition speed compared to YOLO (you only look once) based methods due to the use of lightweight neural networks.

The experimental results show that encoding one-dimensional time series signals into two-dimensional images based on Markov transition field principle can better explore the changing characteristics of the signals. MTF images effectively preserve the dynamic transformation characteristics of the original signal by visualizing the conversion probability of signal amplitude, eliminating the complex feature extraction steps in traditional pattern recognition tasks. They can effectively amplify the features of the signal and improve recognition efficiency. The encoded Markov transition field image has complex details, and MobileNetV2, as a classic lightweight network model, exhibits significant advantages in recognizing this special encoded image. Simultaneously using transfer learning methods and preloading model parameters, compared with training directly without transfer learning, accelerates the convergence of the model and significantly improves the accuracy of network classification.