Substations are critical components of power systems, serving as essential hubs for power transmission and distribution. Intrusions caused by human activity pose significant threats to substations, potentially leading to severe safety issues and substantial economic losses. The security and protection of these facilities are paramount, as any intrusion or damage can result in widespread power disruptions and compromise grid stability. Traditional perimeter security methods such as video surveillance and infrared detection systems face several limitations: video surveillance systems often have blind spots and lack automatic alarm capabilities, while infrared detection systems require flat installation environments and suffer from poor reliability. To enhance performance, distributed acoustic sensing (DAS) technology based on phase-sensitive optical time-domain reflectometry (φ-OTDR) is employed for perimeter security in this paper. By deploying optical fiber in the designated area, vibration signals are collected, and intrusion events are classified using a feature extraction and classification model. These systems offer distinct advantages, including strong anti-electromagnetic interference capabilities, high sensitivity, compact size, and excellent stability in harsh environments. Moreover, DAS enables continuous, real-time monitoring along the entire fiber length, effectively eliminating blind spots. However, accurate identification and classification of different intrusion events using DAS signals remains challenging due to signal complexity and environmental interference. Conventional recognition systems primarily rely on extracting features from signals in either the time or frequency domain, followed by classification using methods such as support vector machines, neural networks, or other deep learning models. These approaches face challenges such as underutilization of signal information across both domains, redundant classification parameters, and computational complexity. To address these issues, we propose a novel intrusion event identification method that combines DAS with advanced signal processing and deep learning techniques. The method integrates the Gramian angular difference field (GADF) for signal encoding and a multi-scale convolutional neural network (MSCNN) enhanced with a cross attention fusion module (CAFM). This approach aims to improve the accuracy of substation perimeter security monitoring by overcoming the limitations of conventional single-scale neural networks and enhancing feature extraction capabilities.

A DAS system is implemented using a narrow-linewidth laser (1550.12 nm) with a pulse repetition rate of 1 kHz and pulse width of 40 ns. The sensing fiber (155 m) is deployed in an S-shaped configuration, with vibration events simulated at the 8-m end section. The system collects vibration signals for five different scenarios: no invasion, striking, climbing, trampling, and shoveling. Raw vibration signals are obtained through in-phase/quadrature (I/Q) demodulation, and the signals are segmented into 1-s frames (1000 data points). The data is then resampled to 224 points. The GADF technique converts time-domain signals into 224 pixel×224 pixel images, which are used as input to the CAFM-MSCNN. After feature extraction through the first convolution layer, the model performs convolution operations using kernels of sizes 3, 5 and 7. The multi-scale fusion framework is built using these different kernel sizes. Each convolutional channel contains several convolutional layers and max pooling layers to extract features and capture complementary information of different scales. The features extracted from the three channels are then fused into feature vectors. The spliced feature vectors are classified by a classifier layer to determine the vibration categories. The model is built based on the PyTorch framework, and the program is written in Python 3.9. The optimal parameters are determined through repeated experimentation: batch size is set to 32, the maximum number of epochs is set to 100, the learning rate is set to 0.001, and the Adam optimizer is used. The confusion matrix is adopted as the evaluation metric of the model.

The experimental results show that the classification accuracies for the five types of vibration events by the proposed model are 100%, 97.65%, 97.22%, 98.32%, and 99.25% (Fig. 9), respectively. Compared with convolutional neural networks (CNN), long short term memory (LSTM) networks, time convolutional neural networks (TCN), CNN-LSTM, and MSCNN, the classification accuracy of the proposed model improves by 3.27 percentage points, 15.16 percentage points, 8.06 percentage points, 7.12 percentage points, and 0.83 percentage points, respectively (Table 4).

We propose a novel approach for substation perimeter security using DAS combined with GADF-CAFM-MSCNN architecture. The method achieves high accuracy in identifying different types of intrusion events, with an average accuracy of 98.49%. The integration of multi-scale feature extraction and cross-attention mechanisms effectively addresses the limitations of traditional approaches. The system’s robustness under various noise conditions, along with its improved performance over existing methods, demonstrates its potential for practical substation security applications. The proposed method offers a new and effective technical solution for enhancing the safety and reliability of intrusion detection.