Intelligent fault diagnosis algorithms based on machine learning and deep learning techniques have been widely used in industrial applications and have obtained much attention as well as achievements. In real industrial applications, working loads of machines are always changing. Hence, directly applying the traditional algorithms will cause significant degradation of performance with changing conditions. In this paper, a novel domain adaptation method, named generative transfer learning (GTL), is proposed to tackle this problem. First, raw datasets were transformed to time–frequency domain based on short-time Fourier transformation. A domain discriminator was then built to distinguish whether the data came from the source or the target domain. A target domain classification model was finally acquired by the feature extractor and the classifier. Experiments were carried out for the fault diagnosis of a wind turbine gearbox. The t-distributed stochastic neighbor embedding technique was used to visualize the output features for checking the effectiveness of the proposed algorithm in feature extraction. The results showed that the proposed GTL could improve classification rates under various working loads. Compared with other domain adaptation algorithms, the proposed method exhibited not only higher accuracy but faster convergence speed as well.
Introduction: The deployment of robots for automated weeding holds significant promise in promoting sustainable agriculture and reducing labor requirements, with vision based detection being crucial for accurate weed identification. However, weed detection through computer vision presents various challenges, such as morphological similarities between weeds and crops, large-scale variations, occlusions, and the small size of the target objects.Methods: To overcome these challenges, this paper proposes a novel object detection model, PD-YOLO, based on multi-scale feature fusion. Building on the YOLOv8n framework, the model introduces a Parallel Focusing Feature Pyramid (PF-FPN), which incorporates two key components: the Feature Filtering and Aggregation Module (FFAM) and the Hierarchical Adaptive Recalibration Fusion Module (HARFM). These modules facilitate efficient feature fusion both laterally and radially across the network. Furthermore, the inclusion of a dynamic detection head (Dyhead) significantly enhances the model’s capacity to detect and locate weeds in complex environments.Results and discussion: Experimental results on two public weed datasets demonstrate the superior performance of PD-YOLO over state-the-art models, with a modest increase in computational cost. PD-YOLO improves the mean average precision (mAP) by 1.7% and 1.8% on the CottonWeedDet12 dataset at thresholds of 0.5 and 0.5-0.95, respectively. This research not only presents an efficient and accurate weed detection method but also offers new insights and technological advances for automated weed detection in agriculture.
Problems: Plant Disease diagnosis based on deep learning mechanisms has been extensively studied and applied. However, the complex and dynamic agricultural growth environment results in significant variations in the distribution of state samples, and the lack of sufficient real disease databases weakens the information carried by the samples, posing challenges for accurately training models.Aim: This paper aims to test the feasibility and effectiveness of Denoising Diffusion Probabilistic Models (DDPM), Swin Transformer model, and Transfer Learning in diagnosing citrus diseases with a small sample.Methods: Two training methods are proposed: The Method 1 employs the DDPM to generate synthetic images for data augmentation. The Swin Transformer model is then used for pre-training on the synthetic dataset produced by DDPM, followed by fine-tuning on the original citrus leaf images for disease classification through transfer learning. The Method 2 utilizes the pre-trained Swin Transformer model on the ImageNet dataset and fine-tunes it on the augmented dataset composed of the original and DDPM synthetic images.Results and conclusion: The test results indicate that Method 1 achieved a validation accuracy of 96.3%, while Method 2 achieved a validation accuracy of 99.8%. Both methods effectively addressed the issue of model overfitting when dealing with a small dataset. Additionally, when compared with VGG16, EfficientNet, ShuffleNet, MobileNetV2, and DenseNet121 in citrus disease classification, the experimental results demonstrate the superiority of the proposed methods over existing approaches to a certain extent.
Fault diagnosis is of great significance to improve the production efficiency and accuracy of industrial robots. Compared with the traditional gradient descent algorithm, the extreme learning machine (ELM) has the advantage of fast computing speed, but the input weights and the hidden node biases that are obtained at random affects the accuracy and generalization performance of ELM. However, the level-based learning swarm optimizer algorithm (LLSO) can quickly and effectively find the global optimal solution of large-scale problems, and can be used to solve the optimal combination of large-scale input weights and hidden biases in ELM. This paper proposes an extreme learning machine with a level-based learning swarm optimizer (LLSO-ELM) for fault diagnosis of industrial robot RV reducer. The model is tested by combining the attitude data of reducer gear under different fault modes. Compared with ELM, the experimental results show that this method has good stability and generalization performance.
7 October 2024 in IEEE Transactions on Instrumentation and Measurement
This study confronts a demanding fault diagnosis task where no sensor signals from the target faults can be utilized for model training. Considering semantic descriptions of the target (or unseen) faults may be known in advance, zero-shot learning (ZSL) is employed to solve the zero-sample fault diagnosis task. Most of existing ZSL methods designed for computer vision tasks rely on pre-extracted features generated by a powerful feature extractor. However, collecting a large amount of fault data in advance for training a powerful and universal fault feature extractor is impractical. To overcome this problem, we propose a masked autoencoder via end-to-end zero-shot learning (MAE_EZSL) approach consisting of four steps, which are self-supervised masked autoencoder pretraining, shared latent space learning, zero shot classifier training, and unseen faults detection. The essence of our approach lies in the comprehensive utilization of masked autoencoders to deeply explore the features of seen faults. Subsequently, we align the distributions learned from sensor signals and fault semantic information to construct essential features associated with unseen faults. Experiments were meticulously conducted to assess the performance of the MAE_EZSL approach using datasets obtained from a benchmark bearing and a specialized test-rig. The obtained results demonstrate that the proposed MAE_EZSL method exhibits competitive performance when compared to state-of-the-art zero-shot learning algorithms.