Wind power penetration has increased rapidly in recent years. In winter, the wind turbine blade imbalance fault caused by ice accretion increase the maintenance costs of wind farms. It is necessary to detect the fault before blade breakage occurs. Preliminary analysis of time series simulation data shows that it is difficult to detect the imbalance faults by traditional mathematical methods, as there is little difference between normal and fault conditions. A deep learning method for wind turbine blade imbalance fault detection and classification is proposed in this paper. A long short-term memory (LSTM) neural network model is built to extract the characteristics of the fault signal. The attention mechanism is built into the LSTM to increase its performance. The simulation results show that the proposed approach can detect the imbalance fault with an accuracy of over 98%, which proves the effectiveness of the proposed approach on wind turbine blade imbalance fault detection.

With the steady development of technology, electric motors (EMs) have become one of the most important components in modern industry. To ensure stable industrial production, detecting and classifying the EM faults is crucial. A novel intelligent deep-learning-based multi-fault detection method for EMs under varying working conditions is proposed in this paper. This method involves two steps: first, a 2D convolution network without pooling layer is proposed to extract features from raw EM data. In addition, a long short-term memory (LSTM) network is applied to extract the fault features for comparison. Second, a capsule network (Caps-Net) based on a dynamic routing algorithm is used as a classifier to realize intelligent multi-fault detection and improve the generalization performance of the proposed model. The proposed method is applicable to raw physical signals of EMs, which improves the overall efficiency of the fault detection. Moreover, the proposed method has a strong generalization ability. The simulation results demonstrate that the proposed approach can achieve higher accuracy than various benchmark methods. Moreover, its accuracy is at least 3% and 10% higher than those of other state-of-the-art models under two working conditions, in which the load type and size of the EM are changed, respectively.

In the construction industry, it is common occurrence for head injuries caused by workers not wearing a helmet. However, the current models for detecting safety helmet either have insufficient detection accuracy or insufficient generalization ability. For this reason, an improved convolutional neural network model, called YOLOv7-WFD, is proposed for the detection of workers without helmets in this paper. Firstly, a new module called DBS in this paper is proposed to strengthen the ability of model to extract target features. This module consists of a Deformable Convolutional, a Batch Normalization layer and a SiLU activation function. Secondly, the Content-Aware ReAssembly of Features (CARAFE) module is introduced to perceive effective features, which improves the model’s ability to reconstruct details and structural information during image up-sampling. Thirdly, Wise-IoU, which is a loss function with dynamic focusing mechanism, is adopted as the loss function to calculate localization loss, which enhances the generalization capability of model and accuracy of detection. Wise-IoU also can evaluate the "outlier" of the anchor box quality, and attenuate the negative impact of low-quality samples in the dataset and enhance the generalization ability of the model. Finally, the experiment shows that the improved YOLOv7-WFD achieves a mAP of 92.6% and a FPS of 79.3 when tested on SHEL5K dataset.

High-resolution underwater imagery is essential for advancing marine exploration, offering critical data to support resource analysis and discovery. However, underwater images frequently face issues, such as color loss and diminished contrast. To overcome these challenges, we propose CCMCS-Net, a network for enhancing underwater images that operate across multiple color spaces and are structured in two stages: color correction and contrast augmentation. A color correction subnetwork and a multicolor-space stretching subnetwork are the two primary parts of the network. Initially, the static correction module (SCM) leverages green channel data to adjust the red and blue channels, correcting color distortion in degraded images. At the same time, the dynamic correction module (DCM) obtains weight mappings through end-to-end training. It combines SCM to achieve adaptive dynamic compensation of the image, which generalizes the deviation during correction to a certain extent. Next, the RGB, LAB, and HSI color distributions are individually adjusted by applying histogram stretching to the corrected image. The adjusted channels are then concatenated and fed into a dynamic fusion module (DFM) for feature fusion across color spaces. This further enhances image visibility and contrast. The DCM and DFM are trained on our modified U-shaped network architecture. The network can adaptively select and emphasize the most valuable features for feature aggregation so that the aggregated features better promote high-quality image reconstruction. Extensive experiments demonstrate that CCMCS-Net delivers outstanding results in both visual assessments and objective metrics for real and synthetic underwater images.

As a key part of data preprocessing, namely attribute reduction, is effectively applied in the rough set field. The purpose of attribute reduction is to prevent too many attributes from affecting classifier operations and reduce the dimensionality of data space. Presently, in order to further improve the simplification performance of attribute reduction, numerous researchers have proposed a variety of methods. However, given the current findings, the challenges are: to reasonably compress the search space of candidate attributes; to fulfill multi-perspective evaluation; and to actualize attribute reduction based on guidance. In view of this, forward greedy searching to $\kappa$-reduct based on granular ball is proposed, which has the following advantages: (1) forming symmetrical granular balls to actualize the grouping of the universe; (2) continuously merging small universes to provide guidance for subsequent calculations; and (3) combining supervised and unsupervised perspectives to enrich the viewpoint of attribute evaluation and better improve the capability of attribute reduction. Finally, based on three classifiers, 16 UCI datasets are used to compare our proposed method with six advanced algorithms about attribute reduction and an algorithm without applying any attribute reduction algorithms. The experimental results indicate that our method can not only ensure the result of reduction has considerable performance in the classification test, but also improve the stability of attribute reduction to a certain degree.

As one of the key topics in the development of neighborhood rough set, attribute reduction has attracted extensive attentions because of its practicability and interpretability for dimension reduction or feature selection. Although the random sampling strategy has been introduced in attribute reduction to avoid overfitting, uncontrollable sampling may still affect the efficiency of search reduct. By utilizing inherent characteristics of each label, Multi-label learning with Label specIfic FeaTures (Lift) algorithm can improve the performance of mathematical modeling. Therefore, here, it is attempted to use Lift algorithm to guide the sampling for reduce the uncontrollability of sampling. In this paper, an attribute reduction algorithm based on Lift and random sampling called ARLRS is proposed, which aims to improve the efficiency of searching reduct. Firstly, Lift algorithm is used to choose the samples from the dataset as the members of the first group, then the reduct of the first group is calculated. Secondly, random sampling strategy is used to divide the rest of samples into groups which have symmetry structure. Finally, the reducts are calculated group-by-group, which is guided by the maintenance of the reducts’ classification performance. Comparing with other 5 attribute reduction strategies based on rough set theory over 17 University of California Irvine (UCI) datasets, experimental results show that: (1) ARLRS algorithm can significantly reduce the time consumption of searching reduct; (2) the reduct derived from ARLRS algorithm can provide satisfying performance in classification tasks.