Urban power communication networks (UPCNs), due to their complex topology and service diversity, are vulnerable to cascading failures. Existing single-metric methods lack accurate damage assessment and global optimization. This paper proposes a Black Widow-enhanced Slime Mould Algorithm (BSMA) for key area identification. A spatial model incorporating communication sites, links, and tower nodes is built, and a multi-source regional importance metric combining centrality, grid hierarchy, and service weights is designed. BSMA improves search performance via black widow mechanisms and uses the importance metric to locate vulnerable areas. Experiments on a UPCN in Jilin Province show BSMA outperforms two advanced methods, achieving 13.3% higher efficiency, 7.05% better connectivity, and 18.4 % stronger tenacity, validating its effectiveness in multi-scale key area identification.

Distributed temperature data collected by BOTDR systems exhibit significant spatiotemporal patterns, where temperature events often share similar magnitudes but differ in structural characteristics, making accurate classification challenging. To address this, we propose MFDS-GFNet, a Multi-Feature Dual-Stream Gate Fusion Network for multi-class temperature event recognition. The method begins with a multi-dimensional feature extraction framework that derives temporal gradients, spatial gradients, and local statistical features from raw data to capture event-specific patterns. A dual-stream architecture is then employed, where raw data and physical features are processed separately using residual networks to enhance deep feature learning and training stability. Finally, a gate fusion module adaptively integrates the two feature streams, automatically weighting their contributions for optimal classification. Experimental results show that MFDS-GFNet achieves 99.74% accuracy across five representative temperature events. Compared to five baseline models, it outperforms the Transformer by 1.07% and GoogLeNet by 1.53%. The model also supports real-time deployment with an inference latency of 0.40 ms per sample.

Multilocation time series classification, one of the most fundamental Φ-OTDR vibration signal applications, has not only gained substantial research attention but has also been widely applied in security monitoring. The current prevailing classifiers are almost all built upon convolutional operations. As a consequence, these methods only perform well in modeling local relationships of signal content but are limited to capturing long-range global interactions. Besides, previous research typically emphasized temporal feature exploration. Yet, the fusion of both spatial and temporal perspectives is pivotal for unearthing richer details, aiding models in enhanced classification decisions. In this paper, we propose a novel Transformer-based method, called ST-T, to capture long-term spatiotemporal features simultaneously. Specifically, we first propose to take time series subsequences which stem from different location variables, as the inputs. Second, we employ a two-dimensional variable-position encoding layer to keep the spatiotemporal information of each patch. Third, we introduce a novel cross-shape mask that enables self-attention mechanism to focus on the pertinent information that requires attention. We evaluate our method on a publicly available Φ-OTDR vibration dataset. ST-T achieves the best accuracy rank compared with several competitive methods. Further ablation studies validate the effectiveness of our model.

The distributed optical fiber sensing system based on φ-OTDR has a wide range of applications in the field of long-distance and large-scale monitoring due to its corrosion resistance, and strong anti-electromagnetic interference ability. Aiming at the problem of low accuracy and high false alarm rate of vibration sensing event classification in φ-OTDR system, a high-accuracy classification method based on Vision Transformer is proposed to identify and classify different types of vibration events, including background noise, digging, knocking, watering, shaking, and walking. Firstly, the two-dimensional space-time signal samples are divided into non-overlapping patches, these patches are flattened into a one-dimensional feature vector, and learnable classification tokens and position coding are embedded to form a feature vector sequence. Then, the global feature extraction and coding of the input feature vector sequence are performed through the Multi-Head Self-Attention mechanism of the Transformer, the backbone network. Finally, the output of the encoder is directly input into the multi-layer perceptron to output the classification results. The proposed method is compared with other three classification approaches based on deep learning, including 2DCNN, 1DCNN-BiLSTM, and an improved 1DCNN-BiLSTM known as 2DCNN-BiLSTM. The experimental results show that the average classification accuracy of the proposed method can reach 99.3%, the false alarm rate ranges from 0.2% to 1.8%, and all the precision, NAR and F1-score indicators are optimal.

CNN is commonly used for DAS signal events recognition, such as 1D-CNNs, 2D-CNN and 1DCNNs-BiLSTM. However, these methods do not make full use of spatiotemporal information to identify DAS signal events, and even have the problem of identifying events confusion. This paper proposes a new DAS signal events recognition method. This method combines 1D-CNNs with gMLP, so that each part can give full play to its own advantages and fully exploit the spatiotemporal information of DAS signal. Firstly, the signal time features on each spatial sampling node are extracted by multiple sets of parallel 1D-CNN, which are recorded as 1D-CNNs. Secondly, the spatial gating unit of gMLP is used to realize the spatial information interaction between the feature vectors of different sampling nodes, so as to mine the spatial relationship between the signals on each sampling node. Finally, the temporal features extracted by 1D-CNNs and the spatial features of gMLP mining are superimposed to form spatiotemporal information, and the fully connected layer is used to realize DAS signal events recognition. The public data set is used for experimental verification. The recognition accuracy of 1DCNNs-gMLP method for training 3 epochs can reach 98.5%, and the highest one is 99.71%. Our accuracy is 7.1%, 4.7% higher than that of 2D-CNN and 1D-CNNs respectively. Compared with the 1DCNNs-BiLSTM that extracts and utilizes detailed time information and overall spatial relationship, the accuracy is improved by 1.4%. The 1DCNNs-BiLSTM has the problem of identification confusion for background noise, digging and watering events. 1DCNNs-gMLP reduces the confusion of these three types of events, and background noise events confusion is decreased by 0.8%; digging events decreased by 1.99%; watering events decreased by 6.16%.

Intelligent recognition of events along the fiber has become an increasingly important issue for many safety monitoring to achieve fault diagnosis and anomaly warnings. The fiber signal acquisition environment is complex, there is a lot of nonstationary noise in the signal. The traditional denoising methods are usually based on the assumption of a constant noise level, which makes it difficult to adapt to different noise situations. Besides, to address the issue of feature extraction and recognition of vibration signals, many methods have been proposed, e.g., convolution neural networks and short-time Fourier transform. However, temporal order information is ignored, which leads to poor classification of certain events that are only clearly distinguishable in the temporal domain. Therefore, this paper designs an Adaptive Shrinkage Denoising and Sequence-state Learning Vibration Detector model (ASVD) that adaptively learns soft thresholds to eliminate noise-related information and capture the temporal sequence state evolution. Specifically, we embed the adaptive soft-threshold learning module into the deep framework to enhance the discriminative feature learning ability from noised signals. Then, we incorporate the sequential state encoder into the framework to capture temporal sequence state evolution. We conducted experiments on an open dataset and achieved an accuracy of 96.5%, a false positive rate of 0.7%, and a false negative rate of 3.53% averaged across the six events.

Since the incremental community detection method is affected by the initial network community and the incremental detection process, it is prone to generate the problem of error accumulation. To solve the above problem, a dynamic community detection algorithm is proposed. First, considering the closeness between nodes to detect the community structure of the first snapshot network. After that, active nodes that reflect network changes are identified in adjacent snapshots and allocated to communities. Then, during the incremental detection process, considering the influence of adding edges on the accuracy of the community, the edge-added node is defined. In this way, the allocated community structure can be split into multiple local communities and singleton communities constructed by edge-added nodes and other nodes. By merging and optimizing the split communities, the final community structure can be detected. Finally, experiments show that the normalized mutual information and the modularity are improved by an average of 5.23% and 5.00% on synthetic dynamic networks. The modularity is improved by an average of 2.39% on real dynamic networks.