2025-11-11T11:16:09.378565

Lightweight Ac Arc Fault Diagnosis via Fourier Transform Inspired Multi-frequency Neural Network

Wang, Jia, Ding et al.
Lightweight online detection of series arc faults is critically needed in residential and industrial power systems to prevent electrical fires. Existing diagnostic methods struggle to achieve both rapid response and robust accuracy under resource-constrained conditions. To overcome the challenge, this work suggests leveraging a multi-frequency neural network named MFNN, embedding prior physical knowledge into the network. Inspired by arcing current curve and the Fourier decomposition analysis, we create an adaptive activation function with super-expressiveness, termed EAS, and a novel network architecture with branch networks to help MFNN extract features with multiple frequencies. In our experiments, eight advanced arc fault diagnosis models across an experimental dataset with multiple sampling times and multi-level noise are used to demonstrate the superiority of MFNN. The corresponding experiments show: 1) The MFNN outperforms other models in arc fault location, befitting from signal decomposition of branch networks. 2) The noise immunity of MFNN is much better than that of other models, achieving 14.51% over LCNN and 16.3% over BLS in test accuracy when SNR=-9. 3) EAS and the network architecture contribute to the excellent performance of MFNN.
academic

Lightweight AC Arc Fault Diagnosis via Fourier Transform Inspired Multi-frequency Neural Network

Basic Information

  • Paper ID: 2510.26093
  • Title: Lightweight AC Arc Fault Diagnosis via Fourier Transform Inspired Multi-frequency Neural Network
  • Authors: Qianchao Wang, Chuanzhen Jia, Yuxuan Ding, Zhe Li, Yaping Du
  • Classification: eess.SP (Electrical Engineering and Systems Science - Signal Processing)
  • Publication Date: October 30, 2025
  • Paper Link: https://arxiv.org/abs/2510.26093

Abstract

Lightweight online detection of series arc faults is critical for preventing electrical fires in residential and industrial power systems. Existing diagnostic methods struggle to simultaneously achieve rapid response and robust accuracy under resource-constrained conditions. To overcome this challenge, this study proposes a Multi-Frequency Neural Network (MFNN) that embeds prior physical knowledge into the network architecture. Inspired by arc current curves and Fourier decomposition analysis, an adaptive activation function EAS with superior expressiveness is created, along with a novel network architecture with branch networks to assist MFNN in extracting multi-frequency features. Experimental results demonstrate: 1) MFNN outperforms other models in arc fault localization, benefiting from signal decomposition in branch networks; 2) MFNN exhibits superior noise immunity compared to other models, with test accuracy 14.51% higher than LCNN and 16.3% higher than BLS at SNR=-9; 3) Both EAS and network architecture contribute to MFNN's superior performance.

Research Background and Motivation

Problem Definition

With increasing electrification of buildings and transportation systems, arc faults have become a critical issue in power systems. Arc faults can ignite surrounding combustible materials, posing significant risks to personal safety and property. Since arc faults typically exhibit high impedance in their early stages, conventional overcurrent and residual current protection devices cannot effectively detect such faults.

Limitations of Existing Methods

Existing arc fault diagnosis methods can be categorized into two types:

  1. Feature Extraction-Based Methods: Employ empirical wavelet transforms, wavelet transforms, and Fourier transforms to analyze time-domain and frequency-domain signals. However, these methods encounter difficulties when processing loads such as dimmers and cleaners, as non-arc signals span wide frequency ranges similar to arc signals.
  2. End-to-End Deep Learning Methods: While capable of automatically discovering potentially necessary features, these methods are computationally intensive, lack physical guidance, and their credibility is questionable.

Research Motivation

The core question of this research is: Can physical guidance in deep learning ensure the accuracy of arc fault diagnosis under limited computational resources? Through analysis of arc electrical signals, signals can be identified as combinations of multiple signals with different periods, where higher-order harmonics play a crucial role in achieving accurate arc fault detection.

Core Contributions

  1. Proposed MFNN Model: A novel unified neural network equipped with physical guidance for handling arc fault diagnosis under limited computational resources.
  2. Innovative Activation Function EAS: Introduced an adaptive activation function with superior expressiveness, featuring periodic positive components with trainable parameters and analytical negative components with asymptotic functions, ensuring network expressiveness under limited resources.
  3. Novel Network Architecture: Based on the core principles of discrete Fourier series analysis, proposed a new neural network architecture with embedded multiple branch networks for signal decomposition and adaptive capture of critical features.
  4. Comprehensive Experimental Validation: Conducted experiments using standard arc fault testing platforms, tested eight advanced deep learning-based arc fault diagnosis models, and performed ablation studies and hardware implementation tests.

Methodology Details

Task Definition

This study focuses on lightweight online detection of AC arc faults under limited computational resources. The input consists of current and voltage signals, and the output is fault classification results (normal or fault).

Theoretical Foundation

Discrete Fourier Series and Physical Guidance

Given a sampled input vector x=[x0,x1,...,xN−1]x = [x_0, x_1, ..., x_{N-1}], the discrete Fourier series transforms a finite sequence of equally-spaced function samples into a complex sequence representing frequency components:

Xk=∑i=0N−1xie−jki2π/NX_k = \sum_{i=0}^{N-1} x_i e^{-jki2\pi/N}

The time-domain input signal can be represented through frequency-domain signals:

xi=∑k=0N−1Xke−jki2π/Nx_i = \sum_{k=0}^{N-1} X_k e^{-jki2\pi/N}

This demonstrates that the input signal can be identified as a linear combination of periodic signals with different frequencies, embodying the decomposability and additivity of the input signal.

Super-Expressive Activation Functions

Super-expressiveness is based on the approximation capability of neural networks. Neural networks employing activation functions with periodic and analytical components can approximate any continuous function with arbitrarily small error, even with a fixed number of neurons.

Model Architecture

Enhanced Adaptive Sine (EAS) Activation Function

Inspired by original arc current curves and decomposition components of input signals, the EAS activation function is proposed:

\sin(\omega x + \phi) & \text{for } x \geq 0 \\ \frac{x}{1+|x|} + \sin(\phi) & \text{for } x < 0 \end{cases}$$ where the right component is a periodic "sine" function with trainable parameters $\omega \in [0,\infty)$ and $\phi \in [-\pi, \pi]$, and the left component is a real analytic asymptotic function. #### Multi-Frequency Neural Network Architecture MFNN comprises multiple parallel branch networks and a trunk network: $$Branch_i(X) = BranchNet_i(X, \theta_i)$$ $$AggregateF = Concat(\sum_{i=1}^{p} Branch_i(X))$$ $$Pr(Y|X, \theta) = Softmax(TrunkNet(AggregateF))$$ #### Branch Network Design Branch networks are designed to separately capture local features of arc current, particularly features with specific frequencies: $$Latent_1 = f_{AvePool1}(EAS(f_{Conv1}(x)))$$ $$Latent_2 = f_{Conv2}(Latent_1) + f_{ShortCut}(x)$$ $$Branch_i(x) = f_{AvePool2}(ReLU(Latent_2))$$ #### Trunk Network The trunk network analyzes features from branch networks: $$Latent_{trunk1} = f_{AvePool}(EAS(f_{Conv1}(x)))$$ $$Latent_{trunk2} = ReLU(f_{Linear1}(Flatten(Latent_{trunk1})))$$ $$y = f_{Linear2}(Latent_{trunk2})$$ ### Technical Innovations 1. **Physical Knowledge Embedding**: Directly maps the physical principles of Fourier decomposition to neural network architecture. 2. **Adaptive Frequency Extraction**: The EAS activation function adaptively extracts critical features from arc signals. 3. **Lightweight Design**: Through physical guidance, MFNN reduces parameter requirements to only 0.26M parameters, significantly fewer than other models. ## Experimental Setup ### Experimental Platform The experimental platform includes: - Voltage Source: 220V, 50Hz (CKAT10 Development Kit) - Oscilloscope: Rigol MSO5000 (250 mA/S) - Current Probe: CP8050A (50A/50MHz) - Voltage Probe: Rigol (250 MHz) - Arc Generator: Compliant with GB 31143 and UL 1699 standards ### Dataset Experiments were conducted on multiple load types, including: - Tungsten Filament Lamp (800W) - Electric Kettle (1800W) - Air Compressor (2200W) - Vacuum Cleaner (1500W) - Switching Power Supply (3.95A) - Dehumidifier (178W) - Microwave Oven (1300W) - Sterilizer Cabinet (800W) The dataset contains 8 tuples and 16 categories (normal and fault currents). Each fault underwent 10 experiments, with each experiment containing 1,000,000 samples and two features (current and voltage). ### Evaluation Metrics Test accuracy is used as the primary evaluation metric, with all test accuracies representing averages of five experiments. ### Comparison Methods Compared with 8 advanced arc fault diagnosis models: - AutoEncoder (AE) - Lightweight CNN (LCNN) - ArcNet - ArcNN - Transformer - CNN-LSTM - Broad Learning System (BLS) ## Experimental Results ### Computational Complexity Comparison | Model | Parameters | FLOPs | Peak Memory | |-------|-----------|--------|------------| | MFNN | 0.26M | 29.24M | 1.41MB | | LCNN | 1.04M | 1098.66M | 3.25MB | | ArcNet | 4.24M | 5052.29M | 8.16MB | | BLS | 0.21M | 6.68M | 0.32MB | MFNN achieves superior performance while maintaining a lightweight footprint. ### Performance at Different Sampling Times On noisy datasets (SNR=5), MFNN maintains 100% accuracy at sampling times ≥2.5×10⁻²ms, while other models achieve only approximately 99%. ### Noise Robustness Experiments Test accuracy under different SNR conditions: | Model | SNR=-9 | SNR=-5 | SNR=-1 | SNR=1 | SNR=3 | |-------|--------|--------|--------|-------|-------| | MFNN | 96.54% | 99.14% | 99.89% | 100% | 100% | | LCNN | 82.03% | 94.43% | 99.47% | 99.85% | 99.88% | | BLS | 80.24% | 84.52% | 93.69% | 97.12% | 98.06% | MFNN achieves 14.51% higher accuracy than LCNN and 16.3% higher than BLS at SNR=-9. ### Occlusion Experiments Through occlusion analysis, MFNN identifies four features, with two being primary features, providing more useful information for generalization. ### Ablation Studies - **Necessity of Branch Networks**: Removing branch networks significantly decreases test accuracy, validating the contribution of the Fourier transform-inspired network architecture. - **Necessity of EAS**: Replacing EAS with ReLU slightly decreases test accuracy (0.1%-0.4%), but maintains good performance. ### Hardware Implementation MFNN was implemented on Raspberry Pi 4B with an average runtime of 4.2ms, meeting IEC62606 standards. Hardware experiments showed almost no misclassifications among 896 tests. ## Related Work ### Classification of Arc Fault Diagnosis Methods 1. **Feature Extraction-Based Methods**: Employ wavelet transforms, Fourier transforms, and other techniques to analyze time-frequency domain features 2. **End-to-End Deep Learning Methods**: Directly use raw electrical signals as input and automatically extract features ### Advantages of This Work Compared to existing work, the main advantages are: 1. Systematically embeds physical prior knowledge into neural network architecture 2. Achieves superior noise robustness under lightweight constraints 3. Provides interpretable feature extraction mechanisms ## Conclusions and Discussion ### Main Conclusions 1. MFNN successfully embeds physical guidance into neural networks, achieving superior arc fault diagnosis performance under limited computational resources 2. Both the EAS activation function and branch network architecture contribute significantly to model performance 3. MFNN demonstrates excellent performance in noisy environments with strong practical value ### Limitations 1. **Prior Knowledge Dependency**: MFNN's performance largely depends on input prior knowledge, limiting its applicability 2. **Domain Specificity**: The proposed components are based on prior knowledge of arc currents and require input datasets with similar physical characteristics for applicability ### Future Directions 1. Extension of prior knowledge and domain adaptation 2. Adaptability studies in other industrial diagnostic applications 3. Enhancement of robustness in more complex noise environments ## In-Depth Evaluation ### Strengths 1. **Strong Innovation**: Cleverly maps the physical principles of Fourier transforms to neural network design 2. **High Practicality**: Achieves superior performance under lightweight constraints, suitable for practical deployment 3. **Comprehensive Experiments**: Conducts thorough comparative experiments, ablation studies, and hardware validation 4. **Solid Theoretical Foundation**: Based on super-expressiveness theory and signal processing theory ### Weaknesses 1. **Limited Applicability**: The method is highly dependent on specific physical prior knowledge 2. **Insufficient Theoretical Analysis**: Lacks in-depth theoretical analysis of why this architecture performs better in noisy environments 3. **Dataset Scale**: The experimental dataset is relatively small, and generalization capability requires further verification ### Impact 1. **Academic Contribution**: Provides new insights for physics-guided neural network design 2. **Practical Value**: Has direct application value in power system safety 3. **Reproducibility**: Provides detailed experimental setup and hardware implementation for ease of reproduction ### Applicable Scenarios 1. Arc fault detection in power systems 2. Other industrial diagnostic applications with periodic signal characteristics 3. Real-time fault detection in resource-constrained environments ## References The paper cites 27 relevant references covering multiple domains including arc fault detection, deep learning, signal processing, and neural network theory, providing a solid theoretical foundation for the research.