Incipient fault detection in power distribution systems is crucial to improve the reliability of the grid. However, the nonstationary nature and the inadequacy of the training dataset due to the self-recovery of the incipient fault signal make the incipient fault detection in power distribution systems a great challenge. In this article, we focus on incipient fault detection in power distribution systems and address the above challenges. In particular, we propose an adaptive time–frequency memory (AD-TFM) cell by embedding the wavelet transform into the long short-term memory (LSTM), to extract features in time and frequency domains from the nonstationary incipient fault signals. We make scale parameters and translation parameters of the wavelet transform learnable to adapt to the dynamic input signals. Based on the stacked AD-TFM cells, we design a recurrent neural network (RNN) with the attention mechanism, named the AD-TFM-AT model, to detect incipient fault with multiresolution and multidimension analysis. In addition, we propose two data augmentation methods, namely, phase switching and temporal sliding, to effectively enlarge the training datasets. Experimental results on two open datasets show that our proposed AD-TFM-AT model and data augmentation methods achieve state-of-the-art (SOTA) performance of incipient fault detection in power distribution system. We also disclose one used dataset logged at State Grid Corporation of China to facilitate future research.

Cloud-edge collaboration and edge intelligence have greatly driven the growth of the Industrial Internet of Things (IIoT). However, the jittery network delay and limited computational resources of edge servers make it difficult to meet the stringent latency requirements in IIoT, and so far there is no good solution to solve this problem. To this end, we introduce scalable neurocomputing, which provides neural networks with different utilities and computation resource requirements, to be deployed on edge servers of cloud-edge IIoT systems. We then optimize such systems by formulating data scheduling and system computational resource allocation as an infinite horizon optimization problem, considering that the data collection from end devices is an infinite long-term process. To solve this hard problem, we design a rolling prediction-optimization framework that transforms the infinite horizon problem into a truncated finite horizon optimization that maximizes the average system utility while satisfying the stringent delay constraints. We have conducted extensive simulations and built a prototype system, which verify the feasibility and performance of our proposed scheme.

Accurate location of Single Phase-to-Ground Faults (SPGFs) in Resonantly Grounded Networks is crucial for timely fault clearing, requiring precise dual-time detection of fault initiation and duration. However, existing methods struggle to address the weak and nonlinear characteristics of fault currents. To tackle these challenges, we propose a paradigm that integrates neural networks with an enhanced Digital Signal Processing (DSP) method. This approach replaces traditional feature extraction with DSP-based methods and dynamically optimizes DSP parameters through network feedback. Specifically, we introduce a transient weak feature-enhanced attention mechanism and a learnable nonlinear waveform reconstruction mechanism, culminating in a parameter-driven DSP-enhanced Transformer named 2TDformer. Furthermore, we construct a dual-time SPGF detection dataset and publicly release it to facilitate model evaluation and support broader research. Experimental results demonstrate that our method achieves maximum errors of less than 1.4 ms for fault initiation detection and 2.8 ms for duration detection. Compared to existing fault detection methods and the state-of-the-art time series anomaly detection models, 2TDformer reduces detection errors by 30%, setting a new benchmark in performance.

Sensor devices continuously generate large volumes of time series data in the Internet of Things (IoT) environment. These voluminous streams require models that scale to massive data while discerning the intricate, multi-scale patterns embedded in diverse temporal sequences. Transformer models have been widely used for IoT time series analysis due to their strong feature representation and global modeling capability. However, existing architectures struggle to explicitly capture temporal structures and adapt to non-stationary data, limiting classification performance. To address these issues, we propose a novel attention mechanism based on the autocorrelation function, named A2T, which leverages lag characteristics to unify temporal modeling and feature extraction. We further introduce a Parameterized Wavelet Transform Module that learns scale and bandwidth end-to-end and uses an attention gate to fuse multi-resolution coefficients. Building on this, we design a Dual-Channel Time-Frequency Feature Extraction module to improve adaptability to distribution shifts. Integrating these components, we develop A2Tformer for IoT time series classification. Experimental results on the UCR dataset demonstrate that A2Tformer achieves an average accuracy of 84.49% and ranks first on 26 out of all datasets, outperforming state-of-the-art Transformer-based models.