Usually, it is subjective, laborious, and inefficient about the traditional water-based air-tightness test. In this article, a mechanism and method of leak detection for a pressure vessel is proposed. First, through collecting and analyzing sequential pressure data, the beginning time of the air-tightness test phase is determined. Second, pressure vessels are located using a deep convolutional neural network (DCNN)-based method and relocation strategy. Third, a method that employs a subregion voting strategy, background modeling, and sequential pressure data is proposed to determine whether, where, and how a pressure vessel leaks. An improved pressure vessel air-tightness test mechanism is developed to provide the basic support for leak detection framework and improve test efficiency. On the independent data set, the accuracy of the proposed leak location method is 98%, and the maximum error of the leakage calculation is 2.14 mL/min. Experiments show that the proposed leak detection framework and mechanism can greatly improve the efficiency and reliability of pressure vessel leak detection.
10 June 2024 in IEEE Transactions on Sustainable Energy
Condition monitoring of wind turbines (WTs) is essential for advancing wind energy. Existing data-driven methods heavily rely on deep learning and Big Data, leading to challenges in distinguishing true faults from false alarms, impacting operational decisions negatively. Thus, this paper proposes a spatio-temporal graph neural network framework that incorporates prior knowledge. Prior WT knowledge is utilized by establishing a spatially structured directed graph embedded in a graph attention network (GAT). The features in WTs' supervisory control and data acquisition system are indicated by the nodes in GAT. Then, the global and local attention embedding layers as well as long short-term memory layers are employed to combine spatio-temporal information from each node. Finally, the condition monitoring in WTs' graph and node-level are established, and a fault propagation chain at node-level is constructed for explaining condition monitoring results. To demonstrate the explainability, robustness and sensitivity of the proposed approach, a comparative analysis between a true fault case and a false alarm case are given, and anomaly detection results are also reported.
28 August 2023 in IEEE Transactions on Instrumentation and Measurement
Remaining useful life (RUL) prediction is of great significance to ensure the safety and reliability of equipment. Graph neural network-based methods show great potential to improve RUL prediction performance by extracting spatiotemporal features from sensor monitoring data. However, current methods construct sensor-based homogeneous graphs without considering equipment component structure data and prior knowledge, which cannot characterize the dependency between sensors and studied equipment accurately. To solve this problem, we propose a temporal knowledge graph (TKG) informer network for RUL prediction. A TKG of equipment health status integrates sensor data with structure data through prior knowledge, so as to characterize various spatiotemporal features accurately. The graph structure and node information (spatial-domain features) of the TKG at each moment is embedded in a low-dimensional temporal graph representation (TGR). An informer network extracts variable information (temporal-domain features) to generate TGR for RUL prediction. The proposed method was evaluated on public datasets and was found to achieve much higher performance than other state-of-the-art models. The TKG datasets are available at IEEE DataPort: https://dx.doi.org/10.21227/jgs2-kt12.
Health status assessment plays an important role in guaranteeing equipment operating safety and reducing maintenance cost. Traditional deep learning methods require large amounts of labeled samples for training, making them ineffective in most practical scenarios with few-shot labeled samples. To address this challenge, we propose a novel knowledge graph metric learning network for few-shot health status assessment. We first construct a health status knowledge graph to fuse sensor monitoring data, component data, and prior knowledge, providing a new approach to describe the spatiotemporal characteristics of equipment. Then, even with few-shot samples, our knowledge graph metric learning network effectively extracts spatiotemporal features from the health status knowledge graph. This network can embed overall structure and node features of health status knowledge graph into low-dimensional vectors, and can learn the similarity of few-shot samples in the vector space. By transforming the health status assessment into a few-shot classification problem in the graph representation space, our model achieves high performance with few-shot labeled samples. Experimental results on public datasets demonstrate that our method is significantly comparable to existing full-shot approaches.