Perimeter security technologies, such as electronic fences and tension networks, are currently outperformed by the phase-sensitive optical time domain reflectometer (Φ-OTDR). Φ-OTDR, known for its antielectromagnetic interference, high concealment, and large monitoring range, provides efficient large-scale monitoring at a reduced cost. Moreover, it not only can locate intrusion events but also identify event types when combined with signal recognition methods. These unique attributes make it valuable for perimeter security applications. Existing identification methods for perimeter intrusion signals are predominantly reliant on machine learning and deep learning techniques. However, machine learning methods require a high level of expert knowledge and their classification efficacy heavily depends on the chosen combination of features and classifiers. Furthermore, the currently available deep learning methods suffer from inadequate learning ability for time-series signals and require complex calculations. To address these challenges, we propose a deep learning recognition model that incorporates a multiattention mechanism. This model was designed to enhance the extraction of critical signal features and improve network learning capabilities. We used the DAS system to gather signals from climbing, knocking, trampling, and no intrusion events, to validate the effectiveness of our proposed method. We also contrasted the recognition rate and efficiency of various deep learning models and assessed the differential impacts of machine learning and deep learning for large sample multiclassification issues.

We first extracted the time-domain waveform of the vibration signal using a signal demodulation technique and then employed a mobile difference method to locate the intrusion event. Following this, we introduced a multiattention temporal convolutional network (MATCN) recognition model, which provides the collected vibration signals directly for identification. This model utilized the channel attention mechanism to optimize the residual module, thereby enabling the selective learning of crucial information from different feature channels. Moreover, we employed the leaky rectified linear unit (Leaky ReLU) to mitigate the issue of neuron death during convolution and to enhance the model's robustness. Furthermore, we incorporated a temporal attention mechanism to help the network identify critical information-laden time slices. We determined the depth of the MATCN based on number of stacked layers in the residual module, informed by the changes in the validation sample's loss function value during training. We conducted ablation experiments to validate the proposed strategy's effectiveness. We also compared MATCN with other typical networks for timing signal recognition tasks, including long short-term memory networks (LSTM), convolutional layers incorporated into long short-term memory networks (CNN-LSTM), and temporal convolutional networks (TCN). An early stop mechanism was added during the network training process to prevent model overfitting. We compared the iteration speed, training epoch, and recognition rates of the different deep learning models. Lastly, we contrasted the recognition effects of MATCN and machine learning methods using two feature sets: zero crossing rate, kurtosis, energy entropy, and approximate entropy; zero crossing rate, kurtosis, skewness, and permutation entropy. These features were combined with common classifiers such as random forest (RF), K-nearest neighbor (KNN), and support vector machine (SVM) for recognition. We compared the recognition effectiveness of different feature group-classifier combinations.

We devised performance comparison experiments for different deep learning models, employing LSTM, CNN-LSTM, TCN, and MATCN to process the same training and validation samples. The network training effectiveness is evaluated by comparing the iteration time, number of epochs, training time, and recognition rate of validation samples throughout the training process for each network (Fig. 13, Table 2). Network performance was assessed based on the recognition rate of each event and the testing time for nontraining samples (Table 3). The results demonstrate that although the iteration speed of MATCN is marginally slower than that of TCN, MATCN require less training time to converge, resulting in the highest overall training efficiency. Moreover, the recognition rate of MATCN for nontraining samples reaches 98.50%, and the recognition time is a mere 0.53s, thus outperforming LSTM and CNN-LSTM. Machine learning methods were also employed to identify the same training and nontraining samples, revealing that the recognition efficacy of machine learning relies heavily on feature extraction and classifier selection. The highest recognition rate achieved by machine learning is 88.67%, falling short of MATCN and even LSTM, thereby underlining the advantages of deep learning for large sample multiclassification problems (Fig. 14).

To address the issue of high expert reliance in machine learning and inadequate learning ability in deep learning for critical time-series signal features in optical fiber perimeter security pattern recognition, we propose a MATCN-based optical fiber perimeter signal recognition model. This model considers the temporal sequence of vibration signals and combines channel and temporal attention mechanisms to extract critical information from various angles. It enhances network learning capability and employs Leaky ReLU to mitigate neuron death during the convolution process, thereby boosting the model's robustness. The recognition results for the four signals indicate that the recognition rate of MATCN for nontraining samples attains 98.50%, thus surpassing LSTM and CNN-LSTM. Furthermore, MATCN outperforms machine learning in handling large sample multiclassification problems. The proposed model can selectively learn critical information across different channels and time slices, facilitating precise and efficient identification for perimeter intrusion signals.