Detection of Voltage Anomalies in Spacecraft Storage Batteries Based on a Deep Belief Network

Xunjia Li, Tao Zhang, Yajie Liu2019 · DOI

TL;DR

This paper introduces a novel Deep Belief Network (DBN) based approach for detecting voltage anomalies in spacecraft storage batteries. By learning complex inter-parameter correlations from telemetry data, the DBN model can predict battery voltage and identify degradation before critical failures occur, offering a more reliable and timely solution than traditional methods.

What Problem Does This Paper Address?

Spacecraft power systems are critical for mission success, and storage batteries are essential components. As spacecraft age and batteries undergo numerous charge/discharge cycles, their performance degrades, leading to abnormal correlations between various operational parameters. Detecting these anomalies early is crucial for maintaining spacecraft health, preventing mission-ending failures, and ensuring operational reliability. Existing anomaly detection methods often struggle with the high dimensionality, complex interdependencies, and noise inherent in spacecraft telemetry data, necessitating a more robust and sophisticated approach.

Methods

The core of this research is the application of a Deep Belief Network (DBN) for anomaly detection. The DBN is a type of deep neural network composed of stacked Restricted Boltzmann Machines (RBMs) and a final neural network layer. The authors propose a two-stage training process:

  • Pre-training: An unsupervised greedy learning approach using RBMs to initialize the network weights and thresholds, capturing the inherent characteristic information within the data.
  • Fine-tuning: A supervised learning phase using a back-propagation neural network layer to optimize the DBN's parameters for accurate voltage prediction.

The DBN model learns the complex correlations between multiple telemetry parameters (charge/discharge currents, times, temperatures) and the battery voltage. Anomalies are detected by comparing the DBN's predicted voltage with the actual measured voltage. A significant deviation indicates degradation. The authors also employed the Levenberg-Marquardt algorithm for parameter optimization, noting its efficiency for large datasets.

Datasets

The study utilized telemetry data from an in-orbit satellite (referred to as satellite X) spanning three years (May-July 2014, May-July 2016, and May-June 2016). This data included parameters such as charge currents, input current, temperatures, spread spectrum systems, battery voltages, and charge/discharge times. The data was collected at a sampling frequency of once every two seconds. The 2014 data was used for training and testing the models, while 2015 and 2016 data were used for evaluating the anomaly detection performance.

Key Insights and Ideas

  • DBN's Superiority in Capturing Complex Correlations: The DBN's deep architecture allows it to effectively learn and represent the intricate, non-linear relationships between various battery parameters and voltage, which are often missed by shallower models.
  • Predictive Anomaly Detection: By predicting future voltage based on current parameters, the DBN can identify deviations before they become critical, enabling proactive maintenance and fault mitigation.
  • Early Warning Capability: The proposed method demonstrated the ability to detect anomalies in the battery's state-of-health (SoH) approximately one year before the actual storage battery fault occurred, highlighting its significant predictive power.
  • Efficiency Gains: The DBN model, particularly when optimized with the Levenberg-Marquardt algorithm, achieved faster training times and better convergence compared to traditional back-propagation neural networks.

Key Results

  • The DBN model with the Levenberg-Marquardt optimization algorithm converged in 639 iterations with an operating time of 533.91 seconds.
  • This DBN-LM model achieved a Mean Squared Error (MSE) of 7.25 × 10⁻⁴ and a Mean Absolute Error (MAE) of 0.0105 on the test data, outperforming other tested models.
  • The proposed DBN-based algorithm successfully detected anomalies in the satellite's storage battery SoH, identifying degradation levels that were not apparent to ground personnel at the time.
  • Compared to traditional multivariate anomaly detection methods like PCA and LLE, the DBN approach showed significantly better performance in detecting voltage degradation anomalies, especially with large volumes of non-linear data.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the DBN model is effective, the accuracy of anomaly detection is dependent on the quality and completeness of the telemetry data. Parameters like battery capacity and internal resistance, commonly used in ground-based diagnostics, are not always available in spacecraft telemetry. The study also notes that the conventional BP algorithm's slow convergence and tendency to fall into local minima are limitations addressed by the DBN.
  • Future Work: The authors suggest applying the proposed algorithm to detect anomalies in storage batteries of other in-orbit spacecrafts, with the aim of further refining and improving the method. They also plan to explore modifications and enhancements to the algorithm.

Summary updated

BibTeX
@misc{li2019detection,
  author = {Xunjia Li and Tao Zhang and Yajie Liu},
  title = {{Detection of Voltage Anomalies in Spacecraft Storage Batteries Based on a Deep Belief Network}},
  year = {2019},
  doi = {10.3390/s19214702},
}
Figure