Roller bearings are one of the most safety-critical components in many machines. Predicting the vibration-based remaining useful life (RUL) of roller bearings allows operators to make informed maintenance decisions and to guarantee reliability and safety. The health indices (HIs) for degradation assessment are constructed by extracting feature information from the collected data, which significantly influences the prognosis result. Conventional HI construction methods rely heavily on expert knowledge and also have limited capacity for learning health information from the raw data from roller bearings. Furthermore, outlier regions often occur in HIs developed by those methods, and these can easily result in false alarms. To address these problems, a novel HI construction method based on a deep multilayer perceptron (MLP) convolution neural network (DMLPCNN) model, which also considers outlier regions, is proposed in this paper. In the proposed model, a 1-D MLP convolution (Mlpconv) block, consisting of a convolution layer and a micro network, is applied to learn features directly from vibrational data. The learned features are then mapped into an HI using a global average pooling layer and a logistic regression layer. Finally, an outlier region correction method, based on sliding thresholds, is proposed to detect and remove outliers in the HI. The outlier region correction method is able to enhance the interpretability of the constructed HI. The effectiveness of the proposed method is verified using whole-life datasets of 17 bearings. The experimental results demonstrate that the proposed method outperforms conventional methods.

Traditional data-driven intelligent fault diagnosis methods for rotating component commonly assume that sufficient labeled data is available. However, the rotary machine works in a normal state most of the time in practical engineering, resulting in fault diagnosis scenarios with extremely limited and imbalanced data. This phenomenon deteriorates the performance of intelligent fault diagnosis methods. In this article, a data augmentation method based on an improved variational autoencoding generative adversarial network is proposed to address this problem. First, the self-attention mechanism and Wasserstein distance with gradient penalty are introduced to improve data generation capability and training stability of the generative model. Then, the generative models are used to generate synthetic samples with high similarity and diversity based on extremely limited frequency spectrum samples. Subsequently, data filters are established to automatically refine the generated samples. Finally, the refined samples are used to enrich imbalanced dataset and improve the fault diagnosis accuracy. The effectiveness of the proposed method is validated by vibration datasets of two rotating components. Experimental results show that the proposed method can generate high-quality samples and has the potential to enhance the fault diagnosis performance in scenarios with limited data.

As an important component of the transmission system, planetary gearboxes are widely used in equipment such as aircraft, wind turbines etc. The changing operating condition results in gear failure easily. The fault diagnosis method using vibration signals can be used to detect gear faults and then avoid accidents. However, there are small faulty samples in the real industrial scenarios, which limits the effectiveness of faults diagnosis methods based on traditional deep learning algorithms. In this paper, an ensemble of novel simplified graph wavelet neural network is proposed for fault diagnosis of planetary gearboxes. In the proposed method, the design of graph wavelet neural networks is simplified to make it more concise and appropriate for fault diagnosis. Different wavelet bases are used to build simplified graph neural networks for extracting diverse features of one sample. Also, a learnable weighting ensemble strategy is proposed to fuse the extracted features. In order to verify the effectiveness of the proposed method, vibration signals of different types of prefabricated faulty gears were acquired in the laboratory and the proposed method was applied for fault diagnosis. The experimental results show that the proposed method has a higher accuracy and a robust comparing with traditional convolutional neural networks and some other graph neural networks such as ChebyNet, GCN and GWNN.

Fault diagnosis of railway assets has drawn the interest of both the scholarly and engineering communities. Federated learning (FL) enables training models across distributed assets to preserve data privacy and reduce high data transfer costs, which has been applied in fault diagnosis. However, the imperfect data transmission problem due to communication errors easily results in low accuracy of FL-based fault diagnosis in the railway system. To solve the problem, a sequential and asynchronous federated learning framework is proposed for fault diagnosis of railway point machines (RPMs) in this work. First, a dual-branch network is proposed as the global model in asynchronous FL for reducing parameters, while maintaining high accuracy. Second, a time cycle mechanism based on sequential Kalman filtering is proposed for reducing the negative impact of data communication errors. Finally, experimental results demonstrates that the proposed method enhances the applicability of online RPM fault diagnosis training in real deployment scenarios.

Aerial image detection of transmission tower and power line in transmission lines is the key technology for unmanned aerial vehicles (UAVs) intelligent inspection, path planning, and obstacle avoidance. Different from previous transmission lines detection methods, this work uses an instance segmentation algorithm to detect transmission towers and power lines. However, the aerial survey image’s large size and high resolution result in a low signal-to-noise ratio, while the high length to diameter ratio of power lines present great challenges. To solve these issues, this paper conducts instance segmentation of aerial survey images based on the improved Solov2 network. Specifically, the neck of the model has been replaced by PaFPN to enhance the feature fusion ability and reduce feature loss. The designed MaskIou branch processes feature relationships before and after the Mask branch of Solov2 and calculates the segmentation loss of each type of mask. Transfer Learning is adopted to pre-train the network and learn data from a single type of lightning conductor so that the network can better identify linear features with high length to diameter ratio that is difficult to recognize. Comparison tests with multiple instance segmentation networks show that the proposed improved network can achieve higher precision in the segmentation.