3 April 2024 in IEEE Transactions on Plasma Science
Partial discharge (PD) is widely recognized as the early sign of fatal insulation failure of high-voltage (HV) power equipment and is considered a major factor in accelerating insulation degradation. However, the detection of weak PD signals in noisy environments and the identification of PD modes based on poor-quality data are rather challenging. In this work, using compact and implantable optical detection tools, optical emission spectroscopy (OES) and phase-resolved (PR) ultraviolet (UV) are combined to monitor PD. Based on the obtained dataset, a Transformer-based deep learning model is built and its recognition performance is optimized to realize the fast recognition and classification of spark, corona, and surface flashover, with a precision of 100% and an accuracy of 99%. This capability will help to alarm potential HV equipment failure and assess the risk in advance.
14 March 2024 in IEEE Transactions on Instrumentation and Measurement
The real-time prelaunch condition monitoring of the flight control systems for rocket requires a fast response speed, which results in missing signals of partial sampling points if it is unmet. The topological inconsistency of sampling points caused by missing signals increases the difficulty in extracting hidden fault features and reduces the accuracy of condition monitoring. This study proposes a missing-signal tolerant condition monitoring method via multiscale features and domain adaptation. The appropriate size of the sliding window is selected by the fuzzy entropy index to improve the real-time response speed of the proposed method. The spatial features of sampling points in different topologies are aggregated by a graph neural network. A multiscale feature extractor with a channel attention mechanism is designed to extract multiscale temporal features. The domain generalization performance of the monitoring model for limited information is improved through unsupervised domain-adversarial training of neural networks to avoid domain shifts caused by missing signals. A test dataset with missing signals is obtained from the virtual signals generated by a ground test simulator. The experimental results of this dataset verify the feasibility and effectiveness of the proposed method.
9 August 2024 in IEEE Access
In Unmanned Aerial Vehicle (UAV) systems, packet loss during sensor data transmission causes data missing, which reduces fault features in sensor signals and causes the accuracy of state monitoring to decrease. This study proposes a state monitoring algorithm combining a convolutional neural network (CNN) with a random forest (RF) for data missing scenarios. CNN algorithm is designed to extract the distributed fault information from the available signals and acquire the state features of the system. Random forest algorithm processes the state features and judges the system state. The integrating strategy utilizes the automatic feature extraction capability of CNN and the superior discrimination capability of an RF classifier to improve the state monitoring accuracy. The experimental results show that the accuracy of state monitoring in data missing condition reaches 92.74%. The comparative experiments verify the validity of the proposed algorithm.
9 September 2025 in IEEE Sensors Journal
The shortage of labeled historical data, particularly the reduction of partial sensor signals in flight logs, has diminished the accuracy of UAV fault detection methods during long-term flights. The limited prior spatial information derived from scarce and incomplete historical data causes overfitting in detection models, particularly when addressing large-scale and heterogeneous online data. This paper proposes a self-supervised prototypical network with a graph sample augmentation method to leverage a small amount of available training samples and enhance the generalization performance of the fault detectors. Missing sensor signals are reconstructed by exploiting the remaining sensor signals in the historical data to create complete monitoring samples. Subsequently, a subset of sensors is randomly removed from these complete samples, and additional samples are resampled from them to augment the training dataset. The augmented training samples are grouped and aggregated intomultiple prototypes based on their categories. Online data are sequentially matched to the prototypes corresponding to various fault types and identified based on their similarity. For unlabeled unknown faults, a meta-trained detector is designed to quickly learn and classify anomalies by utilizing prior knowledge from related meta-tasks. The experimental results, based on datasets from three UAVs, demonstrate the effectiveness of the proposed method.