In deep learning-based fault diagnosis of the wind turbine gearbox, a commonly faced challenge is the domain shift caused by differing operational conditions. Traditional domain adaptation methods aim to learn transferable features from the source domain and apply them to the target data. However, such methods still require access to target domain data during the training process, which limits their applicability in real-time fault diagnosis. To address this issue, we introduce an adversarial single-domain generalization network (ASDGN). It relies solely on data from a single length of data acquisition in wind turbine fault diagnosis. This novel approach introduces a more flexible and efficient solution to the field of real-time fault diagnosis for wind turbines.

Globally, natural wetlands have suffered severe ecological degradation (vegetation, soil, and biotic community) due to multiple factors. Understanding the spatiotemporal dynamics and driving forces of natural wetlands is the key to natural wetlands’ protection and regional restoration. In this study, we first investigated the spatiotemporal evolutionary trends and shifting characteristics of natural wetlands in the Northeast Plain of China from 1990 to 2020. A dataset of driving-force evaluation indicators was constructed with nine indirect (elevation, temperature, road network, etc.) and four direct influencing factors (dryland, paddy field, woodland, grassland). Finally, we built the driving force analysis model of natural wetlands changes to quantitatively refine the contribution of different driving factors for natural wetlands’ dynamic change by introducing the sparrow search algorithm (SSA) and extreme gradient boosting algorithm (XGBoost). The results showed that the total area of natural wetlands in the Northeast Plain of China increased by 32% from 1990 to 2020, mainly showing a first decline and then an increasing trend. Combined with the results of transfer intensity, we found that the substantial turn-out phenomenon of natural wetlands occurred in 2000–2005 and was mainly concentrated in the central and eastern parts of the Northeast Plain, while the substantial turn-in phenomenon of 2005–2010 was mainly located in the northeast of the study area. Compared with a traditional regression model, the SSA-XGBoost model not only weakened the multicollinearity of each driver but also significantly improved the generalization ability and interpretability of the model. The coefficient of determination (R²) of the SSA-XGBoost model exceeded 0.6 in both the natural wetland decline and rise cycles, which could effectively quantify the contribution of each driving factor. From the results of the model calculations, agricultural activities consisting of dryland and paddy fields during the entire cycle of natural wetland change were the main driving factors, with relative contributions of 18.59% and 15.40%, respectively. Both meteorological (temperature, precipitation) and topographic factors (elevation, slope) had a driving role in the spatiotemporal variation of natural wetlands. The gross domestic product (GDP) had the lowest contribution to natural wetlands’ variation. This study provides a new method of quantitative analysis based on machine learning theory for determining the causes of natural wetland changes; it can be applied to large spatial scale areas, which is essential for a rapid monitoring of natural wetlands’ resources and an accurate decision-making on the ecological environment’s security.

The hydraulic oil of marine equipment contains a large number of abrasive contaminants that reflect the operating condition of the equipment. In order to realize the detection of particulate contaminants, this research first proposes a shape-based classification method for oil abrasive particles, designs an oil abrasive particle collection system, and constructs a new dataset. After that, the research introduces deep learning target detection technology in computer vision, and uses GhostNet to lighten the network structure, the CBAM (Convolutional Block Attention Module) attention mechanism to improve the generalization ability of the model, and the ASPP module to enhance the model sensory wildness, respectively. A lightweight target detection model, WDD, is created for the identification of abrasive particles. In this study, the WDD model is tested against other network models, and the mAP value of WDD reaches 91.2%, which is 4.8% higher than that of YOLOv5s; in addition, the detection speed of the WDD model reaches 55 FPS. Finally, this study uses real ship lubricating oils for validation, and the WDD model still maintains a high level of accuracy. Therefore, the WDD model effectively balances the accuracy and detection speed of marine oil abrasive particle detection, which is superior to other oil abrasive particle detection techniques.