An adaptive prognostics and health management system for early fault detection and life prediction in digital circuits

This technology uses advanced signal analysis and machine learning to monitor digital circuits, detect early signs of faults, predict their severity, and estimate remaining useful life, enabling proactive maintenance before failures occur—even with limited real-world data.

Background

Digital circuits are the foundational components across a vast array of modern industrial systems, including those in critical sectors like manufacturing, telecommunications, transportation, and healthcare. Their ubiquitous presence means that any operational failure can trigger severe repercussions, such as extensive system downtime, substantial financial losses, and significant safety hazards. Therefore, ensuring the sustained reliability and predictive maintenance of these circuits is essential for maintaining vital services and economic stability.

However, current approaches to managing digital circuit health suffer from notable deficiencies. Traditional testing methodologies, including Automatic Test Pattern Generation (ATPG), Built-In Self-Test (BIST), and scan-chain diagnosis, are inherently reactive, detecting faults only after they have already manifested as logical errors. This limitation means they lack the crucial capability to predict potential failures before they impact circuit functionality. Furthermore, developing accurate prognostic models is complicated by significant discrepancies between idealized simulation data and the complex, variable behavior observed in real-world operational environments.

Technology

This technology implements an adaptive Prognostics and Health Management (PHM) system for digital circuits by analyzing the subtle transition behaviors of output voltage signals, specifically their rising and falling edges. It extracts detailed temporal, spectral, and combined temporal-spectral characteristics from these transitions. To ensure accuracy despite discrepancies between simulated and real-world environments, the system incorporates domain adaptation techniques, such as Transfer Fisher Discriminant Analysis (TFDA), to align feature distributions. These adapted features then feed into machine learning models, including Support Vector Machines (SVMs) for classifying fault types and Adaptive Gaussian Process Regression (AGPR) for predicting fault severity. Ultimately, these predictions are integrated with degradation models to estimate the Remaining Useful Life (RUL) of circuit components.

This system is uniquely differentiated by its ability to provide early fault detection and prognosis before logical failures manifest, a significant advancement over traditional methods that only detect problems post-failure. It innovatively leverages the often-ignored transient behavior of signals to extract rich fault-related information. Furthermore, its sophisticated domain adaptation technology bridges the crucial gap between simulation data and real-world circuit performance, enabling accurate predictions with minimal experimental data. This allows for proactive maintenance and targeted intervention at the component level, offering a more robust and predictive approach to digital circuit reliability.

Benefits

  • Enables early detection of digital circuit faults before logical failures occur, allowing preventive maintenance.
  • Utilizes adaptive domain adaptation techniques to accurately align simulation and real-world data, improving fault diagnosis reliability.
  • Extracts rich temporal, spectral, and temporal-spectral features from voltage signal transitions for comprehensive fault characterization.
  • Provides component-level fault localization and severity prediction, facilitating targeted interventions.
  • Estimates Remaining Useful Life (RUL) of circuit components, supporting proactive lifecycle management.
  • Operates non-intrusively using only output voltage signals without requiring special test patterns.
  • Requires minimal real-world fault data due to effective transfer learning and domain adaptation methods.
  • Applicable to a wide range of digital circuits and extensible to hybrid digital-analog systems.

Applications

  • Predictive maintenance for industrial electronics
  • Early fault detection in automotive ECUs
  • Health monitoring for aerospace avionics
  • Prognostics in mission-critical medical devices
  • Condition monitoring in data center hardware

Additional info

US patent application pending