I’ve always been fascinated by the LLC resonant converter.
When looking at traditional step-down topologies like the Flyback or Buck converter, they rely on "hard-switching." If you try to push the switching frequency higher to shrink the size of the magnetic components, the switching losses skyrocket, leading to severe heat dissipation bottlenecks.
This is exactly where the LLC topology shines. By utilizing a resonant tank (Lr, Lm, Cr) and a push-pull bridge configuration, it achieves Zero Voltage Switching (ZVS). This drastically minimizes heat generation, allowing for ultra-high-frequency switching. Combined with its inherent galvanic isolation, it’s no wonder this topology is the gold standard for high-power EV chargers and AI data centers.
I thought to myself: "If I can master diagnosing and managing this specific topology, I’ll have a serious edge as an engineer." So, I fired up PSIM and started designing.
🔍 The Hypothesis: Where is the System's "Pulse"?
To build a Predictive Maintenance (PdM) system, I needed a diagnostic indicator. After reviewing various papers and running preliminary simulations, I deduced a critical pattern: The voltage across the resonant capacitor (VCr) reveals exactly what is failing inside the system.
I designed a scenario where an MCU monitors this voltage. Since standard MCUs typically operate at 3.3V, I couldn't feed the raw high voltage directly into it. I had to carefully step it down using a voltage divider to create a scaled, MCU-friendly signal (V_ADC).
📊 Simulating the 4 States of Health
I simulated four distinct scenarios and analyzed how the V_ADC waveform responded:
1. Normal State: The average DC bias is strictly 0V.
2. Cr Aging (Degradation): I dropped the capacitance from 50nF to 30nF to simulate aging. The average voltage remained 0V, but the peak-to-peak voltage noticeably increased.
3. Switch 1 & 4 Failure: I simulated a dead switch using a 30MΩ resistor. The V_ADC waveform lost its balance and shifted entirely to a positive (or negative) DC bias.
4. Switch 2 & 3 Failure: The waveform shifted to the exact opposite DC polarity.
🧠 Building the Machine Learning Logic
Seeing these distinct patterns, I realized that analyzing the DC bias and peak values of V_ADC was enough to classify the faults. I exported the PSIM data to Python and trained a Machine Learning model using a logical decision tree I structured:
When I fed random fault data from PSIM into this algorithm, it diagnosed the exact issue flawlessly!
Case1: CR = 35nF
Case2: sw1 fault & CR = 42nF
⚠️ The Engineer’s Dilemma: Sampling & Aliasing
While the AI worked perfectly in theory, I had to consider the physical limitations of real-world hardware. LLC converters operate at extremely high frequencies, meaning a massive influx of data. However, an MCU’s ADC sampling rate is inherently limited.
I realized a critical issue: if the sampling frequency happens to be an exact multiple of the converter's control frequency, we run into aliasing—potentially missing the crucial peak voltages entirely. I concluded that to make this viable in the real world, we must find the optimal data slicing method and utilize asynchronous sampling to guarantee the true peaks are always captured.
🌱 Future Scope & Reflections
This project was incredibly exciting. However, I know there is more to solve.
Handling Transients: How do I ensure the MCU doesn't mistake a temporary load transient for a permanent fault? My next step is to program a time-delay logic where the MCU only flags a fault if the abnormal data persists over consecutive cycles.
Cascading Faults: I also want to train the model on chain-reaction scenarios—for example, how Cr degradation eventually leads to a switch blowout, and vice versa. I need to keep studying to accurately model these cascading failures in PSIM.
Through this project, I firmly realized that Predictive Maintenance (PdM) is an area where AI gives hardware engineers "wings." With the explosive construction of AI data centers and semiconductor fabs, the value of PdM experts who truly understand both power electronics and data analytics is only going to soar.
Citation: Xiong, S.; Pei, Y.; Wang, W.;
Liu, W.; Zhang, P.; Liu, Y. An
Open-Circuit Fault Diagnosis Method
for LLC Converters. Energies 2024, 17,
Top comments (0)