3 October 2023 in IEEE Transactions on Industrial Informatics
Anomaly detection is a crucial matter to ensure the spacecraft stability. During the spacecraft operation, sensors and controllers generate a large volume of multidimensional time series telemetry data with long periodicity, and one key point to detect the anomaly inside the spacecraft timely and precisely is to extract essential features from the sheer amount of telemetry data. However, great challenges exist owing to the complex coupling relationships and the temporal characteristics inside the telemetry data. To address this issue, we propose a novel approach called maximum information coefficient attention graph network (MAG). The basic frame is a graph neural network, which utilizes embedding vectors to describe the intrinsic properties of each dimension, correlation analysis to investigate long-term dependencies, an attention mechanism to determine short-term interactions among dimensions, and long short term memory (LSTM) to extract temporal features. The fusion of these modules through a graph neural network results in the construction of the MAG model, allowing for a comprehensive analysis of complex variable relationships and temporal characteristics leading to successful detection of various types of anomalies. Since telemetry data has heterogeneous characteristics, we adapt the loss function and design an unsupervised anomaly scoring method suitable for MAG. To verify the effectiveness of the proposed algorithm, we conducted experiments using two publicly and two new available spacecraft telemetry datasets, and the results demonstrate that our algorithm is more efficient and accurate in detecting spacecraft data anomalies than several other state-of-the-art methods.
16 September 2024 in IEEE Transactions on Instrumentation and Measurement
Due to the increasing complexity of industrial systems and the rise in automation levels, their safety and reliability have assumed greater importance. In operation, actuators and sensors generate a large amount of multivariate time series data, wherein anomalous patterns could indicate malfunctions. However, existing methods for detecting anomalies in multivariate time series often struggle with accuracy, particularly in complex industrial environments, and frequently lack interpretability of detected anomalies. Specifically, many current approaches fail to effectively capture temporal dependencies and parameter coupling relationships, leading to suboptimal detection performance. To address these issues, we propose a Triple Attention Graph structure (TAG), which integrates a triple attention mechanism and a normalized flow-based explainable graph encoder. The graph encoder applies Attention LSTM to capture temporal features, focusing particularly on detecting temporal anomalies. Subsequently, through edge attention, we emphasize parameter coupling relationships and highlight parameters that trigger anomalies via a global attention mechanism. Finally, we integrate the features extracted by the graph encoder with conditional probability and normalized flow to assess the density of the time series. This model facilitates anomaly detection through anomaly density estimation and provides explanations for anomalies through attention coefficients. To validate TAG’s effectiveness, we tested it on two publicly available datasets. The results demonstrate superior detection accuracy and recall rates compared to existing methods. Furthermore, we present a detailed analysis of our approach’s interpretability in terms of temporal characteristics, associations, and global perspectives through a case study.
5 August 2025 in Knowledge-Based Systems
29 November 2024 in Measurement Science and Technology
Anomaly detection in spacecraft plays a pivotal role in ensuring the safety and reliability of aerospace missions. As contemporary spacecraft systems become intricate with the integration of numerous subsystems and components, it is crucial to have effective methods to detect anomalies to maintain the normal operation of the spacecraft and prevent potential failures. However, due to the complexity of the space environment and the high level of integration of systems, conventional anomaly detection methods are often inadequate in dealing with multi-source, high-dimensional, and nonlinear data within spacecraft. To address these problems, this paper proposes a spacecraft end-to-end anomaly detection framework based on Mutual Information Neural Estimation (MINE) and long short-term memory networks-variational autoencoder (LSTM-VAE). This framework employs MINE to capture complex dependencies between features and utilizes LSTM-VAE for temporal modeling and anomaly detection, while an attention mechanism is also incorporated to enhance model interpretability. Specifically, MINE is used to learn nonlinear, high-order dependencies among variables in the aerospace system. LSTM-VAE manages to blend the temporal modeling capability of recurrent neural networks and the advantaged latent representation learning of variational autoencoders, providing deeply integrated modeling of the telemetric data dynamics and latent structure. The introduction of the attention mechanism elevates the model’s interpretability, making it possible to trace anomaly detection results back to specific time points and related features. The research conducted comparative experiments on two novel aerospace telemetry datasets. The results indicate that the proposed framework significantly outperforms existing methods in terms of anomaly detection accuracy, precision, recall rate, and F1 score. Additionally, through case study analysis, we demonstrate this framework’s application in actual spacecraft anomaly detection tasks, thereby illustrating its potential in aiding anomaly source tracing in spacecraft.
20 November 2025 in IEEE Transactions on Instrumentation and Measurement
The health condition and remaining useful life prediction of lithium-ion batteries are of vital importance to the safety of energy supply systems. Nonetheless, existing methods for health status and remaining useful life prediction fail to effectively reveal the correlation among features during the charging and discharging process. Moreover, they neglect the temporal relationship of features themselves. Besides, there are obvious limitations in current temporal network methods for predicting the degradation process of different instances, as they are not adequate in capturing long-term dependency relationships. To address these issues, this study introduces a two-stage MINE-Graph-Mamba (MGM) method. In the first instance, the MINE is employed to extract the correlation among features, to serve as edge information of the graph. These relationships are then processed through a graph attention mechanism, while merging new correlation information, and feeding this combination into the Mamba to extract temporal relations and to generate feature representations. The first stage models the multidimensional time-series features of each charge-discharge cycle, producing a unified charge-discharge feature representation. In the second stage, these representations are input into the MGM model to obtain multi-level feature evolution representations that describe the entire degradation process, enabling accurate life prediction. Experimental validation using real battery data demonstrates that the proposed method significantly improves health status and RUL predictions, outperforming existing data-driven approaches.
25 March 2025 in IEEE Transactions on Instrumentation and Measurement
As industrial systems become increasingly complex and automated, fault diagnosis based on operational sequential data has attracted significant attention. However, in real-world production, the intricate spatiotemporal relationships within industrial systems often result in diverse and unseen fault types. Moreover, economic and safety constraints frequently limit the availability of sufficient fault samples. Therefore, it is crucial to develop methods that effectively extract the coupled dimensional and temporal features from multidimensional industrial timeseries data and diagnose unseen faults using knowledge of common seen faults. To address these challenges, we propose an innovative fault diagnosis method that integrates zero-shot learning with graph structures to extend diagnostic capabilities from seen to unseen faults. To analyze complex variable relationships and temporal characteristics, Long Short-Term Memory (LSTM) is employed to extract temporal features across dimensions, which serve as graph nodes. Simultaneously, Mutual Information Neural Estimation (MINE) is used to learn inter-dimensional relationships, forming the graph edges. These relationships are transmitted and integrated through the graph structure. A public attribute matrix is established to bridge seen and unseen faults, which is then projected onto an embedding vector space to infer unseen fault types based on embedding similarity. This method, termed LSTM-MINE Graph network (LMG)-zero-shot learning (ZSL), relies solely on training with seen fault samples while effectively detecting unseen faults. Experiments conducted on a simulation platform and two real industrial test platforms demonstrate that LMG-ZSL outperforms other advanced methods, achieving higher accuracy and lower standard deviation in identifying unseen faults and fault attribute prediction. Stability tests further confirm its superior performance and robustness.