As a promising and non-contact detection technique, machine vision has been widely used in fault diagnosis of substation equipment. The rapid and accurate detection of substation equipment in infrared images is one of the key steps for automatic fault diagnosis. However, the complexity of image background, the low contrast of infrared images and the rotational targets in infrared images pose a great challenge to detection task. This study aims to improve the detection accuracy of the model while having real-time detection speed, and propose a lightweight power equipment detection network (PEDNet) based on YOLOv4-tiny. First, a novel global information aggregation module is constructed to guide the network to focus on the salient regions where the target equipment is located. Second, an improved spatial transformer network is introduced to reduce the impact of rotational targets on detection accuracy. Finally, a feature enhanced fusion network is designed through the use of a multi-scale feature cross-fusion structure. It can fully fuse the feature information of the salient region, the rotational targets and the strong semantic information. The experimental results show that the proposed PEDNet can reach 92.66% detection accuracy and 107.07 FPS real-time detection speed on testing datasets. Compared with YOLOv4-tiny, there is a small sacrifice in detection speed, but the detection accuracy is improved and significantly higher than the existing state of the art (SOTA) object detection models.

Field Programmable Gate Array (FPGA) has become an efficient accelerator for convolutional neural network (CNN) inference due to its high performance and flexibility. To further improve the performance of CNN inference on FPGAs, an Intellectual Property core (IP core) called Deep Learning Processor Unit (DPU) is released by Xilinx. Unlike previous FPGA-based hardware designs focusing on specific functions or CNNs, the DPU IP supports ample basic functions of deep learning, and the developers can take advantage of DPUs to accelerate CNN inference conveniently. In DPU-based CNN acceleration platform, an encapsulated scheduler plays a crucial role in task scheduling between heterogeneous ARM and multiple DPUs. However, the current scheduler is unsatisfactory because its low schedule efficiency. This paper thus presents a high performance task assignment framework built upon Xilinx hybrid CPU-FPGA MPSoC devices. We first evaluate the main causes of low schedule efficiency problem. Then, we explore the scheduler rules and improve shedule efficiency through purposeful observations and analysis. Finally, we integrate our optimizations, and propose an efficient task assignment framework to maximize performance on DPU-based CNN acceleration platform. Experimental results on Xilinx Zynq UltraScale+ MPSoC zcu104 show that our efficient task assignment framework significantly boosts schedule efficiency for small-scale CNNs (from 36% to 70%), medium-scale CNNs (from 65% to 95%), and large-scale CNNs (from 77% to 99%) compared with original schedule strategy.

Multi-Access Edge Computing (MEC) is a promising paradigm that providing cloud-like service for handling the high-complexity and latency-sensitive applications on user equipment (UE) via computation offloading. However, the execution reliability is rarely considered in current MEC studies, which is an important factor to guarantee the quality of service (QoS). For that, this paper considers an energy-saving offloading to satisfy the reliability and latency requirements of the application. Specifically, we formulate an optimization problem to minimize the UE’s energy consumption with reliability and latency constraints. To tackle this NP-hard problem, we first divide the entire application into multiple directed-acyclic-graph-(DAG)-based subtasks, where the subtask can be executed on the UE locally or MEC server remotely. Then, we decompose the overall reliability and latency requirements into multiple constraints for each subtask. Finally, we propose a fast heuristic algorithm to find a solution satisfying the constraints. Simulation results demonstrate the proposed algorithm obtains lower energy consumption compared with the local execution and random assignment and costs less runtime compared with the greedy algorithm.

As a critical component of power battery, the quality of Nickel-plated punched steel strip (NPSS) is closely related to the performance of battery. In practice, however, the real-time detection for defects of NPSS with limited computating resources is a challenging task. Current researches on Vision-based strip defect detectors emphasize the use of large neural network models to pursue high accuracy, while ignoring the training cost and hardware requirements. To accelerate the model update and deployment cycles, and reduce the application cost, a lightweight detection network for accurate and fast detection of surface defects in NPSS is developed in this paper. First, a new lightweight backbone network, LEDNet, is designed to extract features. Depth dynamic convolution is used as the basic structure of the network to reduce the computational complexity while adaptively extracting effective features. Thereafter, a refine-residual bidirectional aggregation network is built to enhance the bidirectional propagation and reuse high-level and low-level features by further optimizing and adjusting the extracted features. On this basis, feature aggregation capability is improved. Finally, the intersection over union k-means algorithm is used to adjust the anchor box size, which balances the significant difference in aspect ratio of surface defects of NPSS. Experimental results on the NPSS defect data-set show that our network detection accuracy can reach 87.17%, the single-sheet detection rate is 0.09s, and the computational complexity is reduced. Compared with the state-of-the-art detectors, it achieves a strong competitive accuracy and low computational effort, which fully meets the accuracy and real-time requirements in the production environment of NPSS.

Regular inspection of photovoltaic (PV) modules is a crucial measure to ensure the safety of solar power generation. To address the challenges posed by small, multi-scale infrared defects in PV modules, this paper proposes a method called FAA-YOLO (Filter and Adaptive assisted YOLO) for detecting small infrared defects in PV modules. Firstly, we design a MPED (Multi-scale Parameter-free Edge Detection) module to enhance shallow features such as texture and edges, thereby improving the network’s capability to extract features from small defect targets. Secondly, we introduce an improved feature fusion network named SlimNet, which mitigates the challenges of small infrared defects in PV modules and the tendency of such features to be overlooked by fusing information from different scales. Lastly, we propose an Adaptive Auxiliary Detection Layer (AAD-Layer). The multi-scale adaptive auxiliary detection layer enables feature reuse and enriches defect feature representation by cascading features of different scales across layers and combining deep and shallow features. Experimental results show that the proposed FAA-YOLO model achieves a detection accuracy of 87.5%. Compared to others methods, it also strikes a good balance between accuracy and computational complexity.