This paper proposes an intrusion pattern recognition scheme based on Gramian Angular Field (GAF) and convolutional neural network (CNN) for the dual Mach-Zehnder Interference (DMZI) distributed fiber perimeter security system, which possesses the advantages of fast recognition speed and high recognition accuracy rate. Compared with traditional recognition algorithms with the inadequate ability to extract deep features, GAF algorithm is used to transform the intrusion signals of one-dimensional time series to two-dimensional images which can present more deep features and maintain the time domain dependence of the signals, meanwhile, each intrusion signal corresponds to a unique fingerprint. Then, this GAF algorithm combined with the appropriate CNN model is able to extract deep features and recognize the intrusion patterns accurately. Besides, GAF algorithm is insensitive to the fluctuation of power source in the optical path, so the robustness as well as the practicability of the system are improved effectively. The experiments demonstrated that the average recognition accuracy rate of three common natural intrusion events and three human intrusion events on fence (wind blowing, light rain, heavy rain, knocking, impacting and slapping) can reach to 97.67%, and the detection response time is only about 0.58 s, which can meet the requirements of real-time emergency monitoring.

Accurately detecting man-made intrusion from different events is of great significance for distributed optical fiber sensing perimeter security system. Most traditional algorithms lack the ability to reject various events of unknown class which are mainly from natural disturbance, and greatly decline the accuracy of intrusion recognition in field application. In order to solve this problem, we proposed a novel robust intrusion event recognition scheme based on convolutional prototype network (CPL), which realized end-to-end feature extraction and recognition based on the similarity of intrusion signals by integrating relevant variables of prototype learning into the training process of multiscale convolutional neural network (MSCNN) as trainable parameters, and had the ability to recognize and reject the unknown disturbance events. In field experiments, the average recognition accuracy of intrusion events as known class can reach 84.67%, with the rejection rate of disturbance events as unknown class is about 83.75%, which ensure the accuracy of intrusion events monitoring in complex field environments. And the recognition response time is about 17 ms, which also meets the need of real-time monitoring.

Visual information can be easily obtained nowadays with more access to various photographic facilities. Appropriate logical analysis and processing of massive visual information could apply to flexible environment pollution detection. In this paper, we present a visual-based non-contact sensing method to detect the low frequency vibration pollution in daily life. The edge points of vibrating objects with more vibration information are extracted using bilateral filtering and Sobel operator. The Sobel operator weights the influence of the pixel position, which make the edge feature extraction effect better and the edge feature obtained more obvious. After that, the continuous frames of the vibration objects with edge information are transformed into several one-dimensional pixel level signals as the input of multi-scale network, which also improves the generalization ability. Experiments on daily objects under low frequency vibration are conducted. The detection accuracy is above 99%, which proves the reliability of the proposed method for the detection of low frequency vibration circumstance pollution.

Low frequency vibration monitoring has significant implications on environmental safety and engineering practices. Vibration expressed by visual information should contain sufficient spatial information. RGB-D camera could record diverse spatial information of vibration in frame images. Deep learning can adaptively transform frame images into deep abstract features through nonlinear mapping, which is an effective method to improve the intelligence of vibration monitoring. In this paper, a multi-modal low frequency visual vibration monitoring system based on Kinect v2 and 3DCNN-ConvLSTM is proposed. Microsoft Kinect v2 collects RGB and depth video information of vibrating objects in unstable ambient light. The 3DCNN-ConvLSTM architecture can effectively learn the spatial-temporal characteristics of muti-frequency vibration. The short-term spatiotemporal feature of the collected vibration information is learned through 3D convolution networks and the long-term spatiotemporal feature is learned through convolutional LSTM. Multi-modal fusion of RGB and depth mode is used to further improve the monitoring accuracy to 93% in the low frequency vibration range of 0–10 Hz. The results show that the system can monitor low frequency vibration and meet the basic measurement requirements.