Search

Wheat rust is one of the important factors leading to wheat yield decline. The traditional method of artificial identification of wheat rust remains inefficient. With the development of unmanned technology, unmanned aerial vehicles (UAVs) with advanced deep vision models can be used to monitor wheat growth, hence enabling timely disease detection and deployment of corresponding treatment measures. However, due to the limitation of embedded system hardware, it is a challenge to deploy a high‐accuracy and lightweight wheat rust detection model in an embedded system. In this paper, a training method based on transfer learning and sharpness‐aware minimization (SAM) is proposed to improve the accuracy of lightweight models. Specifically, the initial model is pretrained on an ImageNet dataset, and then the classifier of the model is fine‐tuned on the wheat rust training set; to prevent the risk of overfitting, the SAM method is applied to adjust the global parameters of the model. The experimental results on the Yellow‐Rust‐19 dataset show that the training method can effectively improve the accuracy of the wheat rust detection model, which is 3.56% higher than the existing methods. In addition, the parameter scale and computing resource requirements of four lightweight models are compared. The results indicate that MobileNetV3‐Small can achieve satisfactory accuracy with low requirements for storage and computing resource, and a detection frame rate of 101.36 frames per second in the Raspberry Pi, which is suitable for the unmanned detection of wheat rust. [ABSTRACT FROM AUTHOR]

Deep learning (DL)-based methods for remaining useful life (RUL) prediction have received increasing research attention due to excellent feature extraction abilities. Most DL methods rely on abundant labeled samples for supervised training. However, because of the adoption of the over-maintenance strategy of equipment, the monitored data for the degradation of equipment usually consists of few labeled samples and a large amount of unlabeled samples, which limits the performance of DL methods. To take advantage of the value of unlabeled samples, this paper proposed a contrastive learning framework for RUL prediction. First, an unlabeled sample augmentation is developed firstly to extend the sample set. Then, an unlabeled sample learning (USL) architecture is proposed to learn the information of degradation from unlabeled samples to promote general DL models' performance on RUL prediction. Based on the proposed framework, USL-convolutional neural network and USL-long short-term memory network are used to validate its performance based on datasets of turbofan engine and bearing. Results show that the performance of RUL prediction based on the proposed framework can be enhanced by unlabeled samples and verify the good scalability and generalization ability of the proposed framework. • Proposed framework boosts RUL prediction with unlabeled samples despite few labeled. • Proposed framework is designed for general deep learning methods like CNN and RNN. • Two case analyses validated the effectiveness and generalization ability. [ABSTRACT FROM AUTHOR]

The issue of uncertainty in Remaining Useful Life (RUL) prediction based on the Deep Learning (DL) framework has gained increased attention in recent years. The probabilistic distribution is usually the effective way to capture the associated uncertainty in RUL prediction. However, most existing research relies on a single distribution that assumes a singular underlying pattern in the data. This type of approach restricts the ability to describe dynamic industrial processes and lacks robustness in adapting to the complexity of time-varying degradation. This paper proposes an approach to improve uncertainty quantification in RUL prediction based on the Multi-Distribution Fusion (MDF) structure. Initially, multiple possible RUL prediction results are produced. Subsequently, the MDF is used to integrate the former results with different weights and outputs the final RUL prediction distribution. The proposed method excels in uncertainty capturing in complex scenarios and provides a deeper understanding of the underlying dynamics of the monitoring data. The application of MDF resulted in more enhanced and robust uncertainty concerning RUL predictions. To validate the effectiveness of the proposed method, two neural networks, the Long Short-Term Memory (LSTM) network and Convolutional Neural Network (CNN), are individually combined with MDF and applied to the C-MAPSS datasets. • The concept of multi-distribution is introduced to quantify the aleatoric uncertainty. • The proposed structure enhances the performance of RUL prediction with uncertainty quantification. • The proposed framework is designed for general deep learning methods, such as CNN and RNN. [ABSTRACT FROM AUTHOR]