The recent development of robotics has opened up the possibility of an online internal cleaning system for power transformers. This paper presents an innovative approach to the internal cleaning of power transformers during online operation. The proposed solution is an "OilBot"—a robot equipped with a multi-layered shielding capsule, a sophisticated cleaning system, and various sensors. The shielding design, which includes mu-metal, aerogel, polyimide, and silver nano-sheets, is described in detail. The cleaning system is equipped with an ultrasonic cleaner, micro-jet accumulator, centrifugal separator, coarse and fine filters, chemical and absorption layers, a thermal treatment component, and a vacuum dehydrator, all optimized for performance. The engineering drawing and compact view of the OilBot is presented. The control system of the OilBot is driven by a PID controller and model predictive control, with cleaning points determined by a mix of pre-programmed and AI-generated waypoints. The controller’s accuracy in following these waypoints is evaluated. The performance of the OilBot is tested. The long-term economic analysis, benchmark, process optimization on OilBots and chemical kinetics are also provided. This methodology provides a unique solution for maintaining transformer cleanliness and enhancing performance without disrupting operation.

The widespread deployment of Internet of Things (IoT) devices in intelligent school environments provides abundant data for enhancing video recognition systems with advanced behavior recognition technologies, which is crucial for ensuring campus safety. Traditional models primarily rely on visual cues and often struggle to identify complex behaviors against dynamic backgrounds. This study addresses these challenges by introducing a novel Context-Embedded Contrastive Learning (CECL) approach, which leverages data collected from IoT devices to synergistically integrate textual and visual modalities through contrastive learning, thereby refining behavior recognition capabilities. In addition, the model incorporates a Context-Embedded Transformer Encoder that utilizes temporal and geographic context cues captured by IoT sensors, processed through the proposed Context-Aware Multi-Head Attention mechanism. This enhancement significantly improves the model’s ability to discern subtle behavioral nuances by harnessing the rich contextual data from IoT-based surveillance systems, which conventional methods frequently overlook. Through this integration, the model achieves a more sophisticated understanding and recognition of abnormal behaviors in campus settings, as evidenced by the CABRH8 dataset. The CECL model has demonstrated outstanding performance, achieving a Top-1 accuracy of 80.26% and a Top-5 accuracy of 99.38% on the challenging CABRH8 dataset, thereby outperforming existing models reliant solely on visual data. Its robustness and adaptability have also been validated across additional datasets, including CABR50, UCF-101, and HMDB-51, demonstrating its potential for widespread deployment in various IoT-enabled surveillance applications.

Vehicle Make and Model Recognition (VMMR) is a fine-grained classification task in Intelligent Transportation System (ITS). Recent works address VMMR through feature encoding schemes, part-based methods or attention modules. Despite their astounding results, these techniques concentrate on the high-level semantic features. This practice cripples the feature expressive ability of the networks as the granular traits of the vehicle distilled from the early convolution layers are not embedded into the final feature representations. In this work, by contrast, a Cross-Granularity (CG) module which is responsible for the integration of macroscopic and microscopic components is proposed. By incorporating the CG module into a Convolutional Neural Network (CNN), the resultant network i.e. CGNet reinforces the feature extraction ability by amalgamating the feature maps from different scales to render a balanced mix between local contextual information and global semantic details. To validate the proposed framework, experiments are conducted on four publicly available datasets. We report competitive performance on web-nature Comprehensive Cars, Stanford Cars, Car-FG3K and surveillance-nature Comprehensive Cars datasets with 98.3%, 95.4% 86.4% and 99.1% accuracies. Furthermore, we demonstrate the ability of the CGNet to pinpoint distinctive fine-grained details via the Gradient-Weighted Class Activation Mapping (Grad-CAM) technique and compare it against the baseline which learns on deep features alone. The generalization ability of the CG module on other CNNs is also examined and the results suggest a high compatibility between the two.

Environment perception is the premise for intelligent vehicles to drive safely and stably. Despite the rapid development of road detection technology based on visual images, it is still challenging to robustly identify road areas in visual images due to the influence of illumination changes and noise. In order to solve this problem, we introduce a new optimized lidar and camera sensor fusion method for road environment sensing of intelligent vehicles. In road boundary detection based on laser data, a median point filtering method of ordered pole cloud is proposed. A method of boundary search, boundary seed point growth and obstacle clustering is proposed to identify road boundary. In the lane line classification based on visual image, a lane line search classification method is proposed, which can effectively classify lane lines and extract single lane lines. On the basis of the optimization of sensors, several constraint conditions are proposed based on the fusion of the two data, and the location of missing lane lines is predicted by using the road information identified by lidar and image, and the lane lines are identified again. Finally, a large number of experiments are carried out on kitti-Road benchmark data set, and a test platform is built to verify the results of the identification method proposed in this paper in rainy day, cloudy day, night and other special scenarios. Experimental results show that this method is superior to existing methods.

Taiwan's agriculture is an important national economic industry. Ensuring food security and stabilizing the food supply are the government's primary goals. The Agriculture and Food Agency (AFA) of the Executive Yuan's Council of Agriculture has conducted agricultural and food surveys to address those issues. Synthetic aperture radar (SAR) images will not be affected by climatic factors, which makes them more suitable for the forecast of rice production. This research uses the spatial-temporal neural network convolutional long short-term memory network (ConvLSTM) to identify rice fields from SAR images. The results show that ConvLSTM can greatly reduce the proportion of model false positives to 51.16%, produced higher average precision of 95.70%, and F1-score of 0.9648. The ConvLSTM neural network has produced good results for rice field identification compared with state-of-the-art neural networks.

Driver distraction is the main factor of severe traffic accidents and has become an essential issue in the traffic safety field. Hence, driver inattention systems are crucial in ensuring the safety of road users. With the introduction of Vision Transformer for computer vision tasks, there is a lack of comprehensive evaluation of various models for distracted driver detection. Hence, we raise the question - does vision transformers outperform convolutional neural networks (CNNs) in the field of detecting driving distraction? In this work, we evaluate and perform in-depth evaluations of various state-of-the-art CNN and Vision Transformer models to detect the distracted driver. We believe this will aid future researchers in this field in benchmarking their novel models with state-of-the-art models. We select ResNet, VGGNet, DenseNet, and EfficientNet as the candidates for CNN, while ViT, Swin Transformer, DeiT, and CaiT for Vision Transformer. We perform our benchmark on the American University of Cairo Distracted Driving Dataset (AUC-DDD) which consists of ten distracted classes. It is observed that CNN should be considered first if the downstream task is specific and the available dataset is small. An in-depth discussion and analysis are included in this work.

Most of the impulse denoisers are either median filter‐based or fuzzy filter‐based, which can only perform well in low noise conditions. This study presents an efficient convolutional neural network (CNN) with particle swarm optimisation (PSO) model for high‐density impulse noise removal. The proposed high‐density impulse noise detection and removal model mainly consists of two parts: the impulse noise removal and impulse noisy pixel detection for restoration. The authors’ model initially leverages the powerful ability of deep CNN architecture to separate noise from the noisy image, then adopts PSO to pinpoint the most optimised threshold values for detecting impulse noisy pixels. An ensemble of these algorithms is an intelligent and adaptive solution, producing a clean output while preserving significant pixel information. Targeting to solve high‐density impulse noise problems, the authors have trained their model with a massive collection of natural images and 14 standard testing images are used for validation purposes. In order to validate the robustness of the proposed method, different levels of high‐density impulse noise are considered. Based on the final denoised images, their model has proven its reliability, in terms of both visual quality and quantitative evaluation, on greyscale and colour images.

Vehicle model recognition plays a crucial role in intelligent transportation systems. Most of the existing vehicle model recognition methods focus on locating a large global feature or extracting more than one local subordinate-level feature from a vehicle image. In this paper, we propose the principal component analysis network-based convolutional neural network (PCNN) and pinpoint only one discriminative local feature of a vehicle, which is the vehicle headlamp, for vehicle model recognition. The proposed model eliminates the need for locating and segmenting the headlamp precisely. In particular, PCNN ascertains the effectiveness of both principal component analysis and CNN in extracting hierarchical features from a vehicle headlamp image and also reducing the computational complexity of the traditional CNN system. To further enhance the training procedure while still keeping the discriminative property of the network, the fully connected layer is updated by backpropagation optimized with stochastic gradient descent. The proposed method is validated using a data set that comprises 13 300 training images and 2660 testing images, respectively. The model is robust against various distortions. Experiments show that PCNN outperforms state-of-the-art techniques with an average accuracy of 99.51% over 38 vehicle makes and models using the PLUS data set. In addition, the effectiveness of the proposed method is also validated using the public CompCars data set, achieving 89.83% accuracy over 357 vehicle models.

Recent works in image captioning have shown very promising raw performance. However, we realize that most of these encoder-decoder style networks with attention do not scale naturally to large vocabulary size, making them difficult to deploy on embedded systems with limited hardware resources. This is because the size of word and output embedding matrices grow proportionally with the size of vocabulary, adversely affecting the compactness of these networks. To address this limitation, this paper introduces a brand new idea in the domain of image captioning. That is, we tackle the problem of compactness of image captioning models which is hitherto unexplored. We showed that our proposed model, named COMIC for compact image captioning, achieves comparable results in five common evaluation metrics with state-of-the-art approaches on both MS-COCO and InstaPIC-1.1M datasets despite having an embedded vocabulary size that is 39×-99× smaller.

This study presents a model to effectively recognise image noise of different types and levels: impulse, Gaussian, Speckle and Poisson noise, and a mixture of multiple types of the noise. To classify image noise type, the convolutional neural network (CNN) method with backpropagation algorithm and stochastic gradient descent optimisation techniques are implemented. In order to reduce the training time and computational cost of the algorithm, the principal components analysis (PCA) filters generating strategy is deployed to obtain data adaptive filter banks. The authors validated their designed CNN with PCA for noise types recognition model with degraded images containing noise of single and combination of multiple types, with a total of 11,000 and 1650 datasets for training and testing purposes, respectively. The variety and complexity of data have never been addressed before in any other research work. The capability of their intelligent system in handling images degraded under this complicated environment has surpassed human‐eye performance in noise types recognition. The authors’ experiments have proven the reliability of the proposed noise types recognition model by having achieved an overall average accuracy of 99.3% while recognising eight classes of noise.