Abstract
DC-DC converters are critical components in industrial systems, where their failure can lead to abnormal operation or catastrophic accidents. Traditional fault analysis often relies on complex mathematical modeling or invasive signal processing. This paper proposes a non-invasive fault analysis method based solely on the circuit's output voltage response. By monitoring the output signal, the system can simultaneously locate faulty components and predict their Remaining Useful Performance (RUP). To address the common issue of particle degeneracy in standard prognostic algorithms, a novel conditional particle filter (CPF) is introduced. This filter utilizes historical degradation data to maintain particle diversity and improve prediction accuracy. Case studies on two converter topologies demonstrate that the proposed method effectively identifies fault conditions and achieves high RUP prediction precision.
Introduction
DC-DC converters are essential for power management in industrial electronics but are highly susceptible to environmental stressors such as heat, vibration, and aging. When internal components like capacitors or inductors drift from their nominal values, system stability is compromised.
Current diagnostic approaches generally fall into two categories: model-based methods, which are computationally intensive and struggle with real-world non-linearities, and signal-processing methods, which often require internal node access. Probing internal nodes is frequently impractical for sealed or high-density industrial units. Furthermore, standard prognostic tools like Particle Filters (PF) often suffer from "particle degeneracy," where the statistical diversity of future state guesses collapses, leading to inaccurate failure forecasts. This research introduces a "black-box" diagnostic approach that identifies internal component degradation using only the output voltage wire.
Methodology
The proposed system employs a two-stage strategy for health monitoring: output-based fault analysis and RUP prediction via an improved particle filter.
Output-Based Fault Analysis
Instead of monitoring internal nodes, the system focuses on the Output Average Voltage ($V_{oa}$). The researchers pre-construct "fault-indicating curves" by simulating how specific sensitive components (e.g., filter inductors or feedback resistors) affect the output voltage as they deviate from nominal values. By matching real-time $V_{oa}$ readings against these stored signatures, the system can pinpoint the specific component undergoing degradation.
Exemplar-Based Conditional Particle Filter (CPF)
To predict RUP, the authors developed a Conditional Particle Filter that incorporates historical degradation data as "exemplars." Unlike standard PFs, which can lose diversity during the resampling step, the CPF uses historical failure trajectories to guide current predictions. This ensures that the statistical "particles" remain grounded in realistic failure patterns, significantly mitigating the degeneracy problem.
Experimental Setup and Datasets
The methodology was validated using a combination of software simulation and hardware testbeds.
- Simulation: Saber software was used to generate SPICE models for two converter types, allowing for the creation of comprehensive fault-indicating curves.
- Hardware Testbeds: Two physical circuits were constructed: a simple converter based on the MC34063a chip and a more complex PWM-controlled converter using the TL494 chip.
- Data Acquisition: National Instruments boards and LabVIEW were utilized to record real-time voltage data as component degradation was induced.
Key Insights and Results
The core finding of this research is that internal component drifts create unique, identifiable "fingerprints" in the output voltage. By focusing on sensitive components identified through theoretical screening (such as L, C, and feedback resistors), the system avoids monitoring components with negligible impact on output health.
- Diagnostic Accuracy: The system successfully categorized faults (e.g., "Inductor L decrease" or "Resistor R1 increase") across both simple and complex topologies.
- Prognostic Precision: The CPF demonstrated remarkable accuracy in predicting the End of Performance (EOP). For instance, when monitoring the timing capacitor of the MC34063a, the maximum prediction error was limited to only 4 cycles.
- Real-Time Feasibility: Because the method relies on output sampling rather than complex internal modeling during operation, it is computationally efficient enough for real-time industrial deployment.
Conclusion
This research provides a robust framework for the real-time health monitoring of DC-DC converters without the need for invasive probing. While the current model focuses on single-component faults with a 20% deviation threshold, future work will explore simultaneous multi-component failures and the impact of extreme environmental noise on diagnostic sensitivity. The integration of historical data into the particle filtering process represents a significant step forward in improving the reliability of power electronic prognostics.
Summary updated