A Lightweight, Transferable, and Self-Adaptive Framework for Intelligent DC Arc-Fault Detection in Photovoltaic Systems

This paper proposes LD-framework, a lightweight, transferable, and self-adaptive learning-driven system that achieves near-perfect DC arc-fault detection in photovoltaic systems by effectively overcoming spectral interference, hardware heterogeneity, and long-term operating condition drift through compact spectral learning, cross-hardware alignment, and cloud-edge collaborative adaptation.

The Gist

Imagine your home solar panel system as a busy highway. Most of the time, the cars (electricity) flow smoothly. But sometimes, a loose connection or a frayed wire causes a spark—a DC arc fault. This is like a tiny, invisible fire that can start a real blaze if not stopped immediately.

To stop these fires, we use a safety guard called an AFCI (Arc-Fault Circuit Interrupter). Think of the AFCI as a very sharp-eyed security guard standing at the gate. Its job is to spot the "spark" (the arc) and instantly shut down the power.

The Problem:
In the real world, this job is incredibly hard.

1. The Noise: The solar panels and batteries are constantly talking to each other, switching on and off, and dealing with clouds or changing loads. This creates a lot of "static noise" that sounds suspiciously like a fire. Old-school guards get confused by this noise and either shut down the power for no reason (a "nuisance trip") or, worse, miss a real fire.
2. Different Hardware: Every solar inverter (the brain of the system) is built differently. A guard trained on one brand of inverter might get confused when they see a different brand.
3. Aging: Over years, wires get old, weather changes, and the "voice" of the system changes. A guard trained in 2024 might not recognize the sounds of 2028.

The Solution: The "LD-Framework"
The paper introduces a new, smart security system called the LD-Framework. Instead of a static guard with a rulebook, imagine a super-smart, adaptable AI team that learns and evolves. It has three special members:

1. LD-Spec: The "Frequency Detective" (The On-Device Guard)

• What it does: This is the guard standing right at your house.
• The Analogy: Imagine listening to a crowded room. A human might hear "noise," but a frequency detective uses a special pair of glasses to see the colors of the sound.
• How it works: Arc faults have a unique "color" (a specific broadband energy pattern) that is different from the normal "static" of the inverter. LD-Spec is a tiny, lightweight AI that looks at these colors. It's so smart and efficient that it can run on a small chip inside your inverter without slowing anything down. It ignores the noise and only screams "Fire!" when it sees the specific color of an arc.

2. LD-Align: The "Universal Translator" (The Cross-Hardware Expert)

• What it does: This member helps the system work across different brands of solar equipment.
• The Analogy: Imagine you trained a guard to recognize a fire in a brick house. Now you move them to a wooden house. The smoke looks different! LD-Align is like a translator that teaches the guard, "Hey, even though the smoke looks different here, it's still a fire."
• How it works: It takes the knowledge learned on one type of inverter and "aligns" it so it works perfectly on a different type. The best part? It doesn't need to relearn everything from scratch. It just needs a tiny bit of new data (like 0.5% to 1% of the usual amount) to adjust its understanding.

3. LD-Adapt: The "Cloud-Connected Learner" (The Long-Term Evolution)

• What it does: This member ensures the system stays smart over many years, even as the environment changes.
• The Analogy: Imagine a guard who gets a daily report from a central headquarters. If a guard in a specific neighborhood starts seeing a new type of "fake fire" (like a weird weather pattern), they send a sample to HQ. HQ analyzes it, updates the guard's rulebook, and sends it back.
• How it works:
  • Detection: If the local guard is confused by a new situation, it sends a "mystery sample" to the cloud.
  • Verification: Experts check if it's a real fire or a false alarm.
  • Evolution: If it's a new pattern, the cloud updates the AI model and sends it back to all devices (like a software update on your phone).
  • Safety: They use a "Canary Deployment" strategy—like testing a new recipe on one chef before serving it to the whole restaurant—to make sure the update doesn't break anything.

The Results: Why This Matters

The researchers tested this system with over 53,000 real-world examples. Here is what happened:

• Perfect Accuracy: It caught 99.99% of real fires.
• Zero False Alarms: It never shut down the power for no reason, even during tricky moments like starting up the system or switching loads.
• Fast Recovery: When the system encountered a totally new environment (like a real rooftop vs. a lab), it could "teach itself" to get back to 95% accuracy just by tweaking its settings, without needing a massive overhaul.

In Summary:
This paper presents a self-driving safety system for solar power. Instead of a rigid, dumb switch that gets confused by noise or new hardware, this system is a lightweight, learning AI that:

1. Sees the fire clearly through the noise.
2. Adapts to different hardware brands instantly.
3. Learns from the cloud to stay smart forever.

It turns a static safety device into a living, breathing guardian that gets better the longer it works, keeping our homes safe from solar fires without annoying false alarms.

Technical Summary

1. Problem Statement

The paper addresses the critical safety challenge of DC series arc faults in residential Photovoltaic (PV) and Battery Energy Storage Systems (BESS). Unlike AC parallel arcs, DC series arcs are sustained by continuous current flow and exhibit stochastic, broadband spectral signatures.

Key Challenges Identified:

• Intra-System Variability: Normal operations (MPPT perturbations, grid transitions, load switching) generate broadband spectral disturbances that overlap with arc signatures, causing nuisance trips or missed detections in traditional threshold-based detectors.
• Cross-Hardware Heterogeneity: Different inverter platforms vary in switching frequency, semiconductor technology, and topology. These differences create distinct spectral distributions, making models trained on one device fail on another without costly retraining.
• Long-Term Temporal Drift: Over time, component aging, environmental changes, and installation-specific factors shift baseline spectra. Static models degrade in performance, leading to elevated false-alarm rates or reduced sensitivity.
• Embedded Constraints: Practical deployment requires millisecond-level latency and operation on microcontrollers (MCUs) with strict memory and computational budgets, ruling out heavy deep learning models.

2. Methodology: The LD-Framework

The authors propose a unified, learning-driven framework called LD-framework, consisting of three integrated components designed to operate in a closed-loop cloud-device ecosystem.

A. LD-Spec: Lightweight On-Device Detection

• Architecture: A compact, frequency-domain Convolutional Neural Network (CNN) optimized for MCUs.
• Feature Engineering: Instead of time-domain analysis, the system uses a frequency-centric pipeline:
  1. Current signals are segmented into 1024-sample frames.
  2. Hanning windowing and Discrete Fourier Transform (DFT) are applied.
  3. DC components are removed, and magnitudes undergo logarithmic scaling.
  4. Adjacent frequency bins are aggregated to reduce sensitivity to narrow-band switching harmonics.
• Model Design: The CNN operates exclusively along the frequency axis using lightweight convolutional blocks (Conv-BN-ReLU-Dropout-MaxPool). It avoids residual connections to save memory.
• Decision Logic: A Dual-Threshold Mechanism is employed. A frame-level classifier outputs a probability, but an arc event is only declared if a counter of consecutive positive frames exceeds a predefined threshold. This suppresses transient false alarms.

B. LD-Align: Cross-Hardware Transfer

• Goal: To enable a model trained on a source inverter to generalize to a target inverter with minimal target data.
• Mechanism: A mixed-domain fine-tuning approach that preserves the source model's decision boundary while adapting to the target.
• Loss Function: The total loss combines:
  • $L_{tgt}$: Supervision on the small target dataset.
  • $L_{src}$: Replay of source data to prevent catastrophic forgetting.
  • $L_{L2}$: Regularization to keep weights close to the pre-trained source.
• Strategy: It uses layer-wise learning rate scheduling (lower rates for the backbone, higher for the head) and early stopping to ensure robust transfer with only 0.5%–1% labeled target data.

C. LD-Adapt: Self-Adaptive Cloud-Device Evolution

• Goal: To handle long-term drift and unseen operating regimes in the field.
• Workflow:
  1. Novelty Detection: Devices transmit potential false alarms or novel regimes to the cloud. Experts verify these to establish ground truth.
  2. Two-Stage Evolution:
     • Stage 1 (Parameter Tuning): Adjusts hyperparameters (learning rates, loss weights) and fine-tunes model parameters using new field data. This is fast and preserves the model structure.
     • Stage 2 (Micro-Architecture): If Stage 1 saturates, limited structural changes (kernel size, channel width) are allowed, constrained to <5% additional FLOPs to maintain MCU compatibility.
  3. Canary Deployment: Updated models are tested on a small subset of devices before fleet-wide rollout to ensure safety.

3. Key Contributions

1. Deployment-Oriented Spectral Backbone (LD-Spec): A frequency-domain CNN that captures arc-discriminative features while strictly adhering to embedded latency and memory constraints.
2. Cross-Hardware Transfer (LD-Align): A method that preserves arc-relevant spectral structures across heterogeneous inverters, requiring minimal target supervision (0.5%–1%) to achieve high performance.
3. Lifecycle-Aware Adaptation (LD-Adapt): A cloud-device collaborative mechanism that detects novel regimes and performs controlled model evolution, preventing performance degradation over time.
4. Industrial-Scale Validation: Comprehensive testing across multiple inverter frequencies, UL 1699B scenarios, and real-world field deployments.

4. Experimental Results

The framework was validated using over 53,000 labeled samples across laboratory and field conditions.

• Detection Performance:
  • Achieved 0.9999 Accuracy and 0.9996 F1-score.
  • 0% False-Trip Rate across diverse nuisance-prone conditions (start-up, grid transitions, load switching, harmonic disturbances).
  • 100% Arc Fault Detection rate.
• Cross-Hardware Transfer:
  • Reliable adaptation achieved using only 0.5%–1% of labeled target data.
  • A source model trained with only 30% supervision was sufficient for effective transfer.
• Field Adaptation:
  • In a real-world scenario where a baseline model dropped to 21% precision due to unseen conditions, LD-Adapt recovered precision to 95% using only Stage 1 parameter updates.
  • No Stage 2 structural evolution was required for the tested drifts, validating the efficiency of the hierarchical approach.
• Scaling Analysis: Performance gains diminished significantly beyond 100k samples, suggesting a practical data saturation point for embedded deployment.

5. Significance

This paper presents a paradigm shift in AFCI technology, moving from static, laboratory-trained classifiers to dynamic, maintainable, and deployment-oriented protection systems.

• Scalability: The ability to transfer models across different hardware with minimal data solves the "data scarcity" problem inherent in collecting dangerous arc-fault data for every new device.
• Reliability: The self-adaptive mechanism ensures that safety systems do not degrade over time due to environmental drift or component aging, a critical requirement for long-term PV-BESS safety.
• Feasibility: By proving that high-accuracy detection is possible on resource-constrained MCUs without sacrificing robustness, the framework provides a practical foundation for next-generation smart safety standards in distributed energy systems.