An Open-Circuit Fault Diagnosis Method for Three-Level Neutral Point Clamped Inverters Based on Multi-Scale Shuffled Convolutional Neural Network

Yan Yan, Jiaqi Wu, Yanfei Cao, Bo Liu, Chen Li, Tingna ShiSensors 2024 · DOI

TL;DR

This paper introduces a novel Multi-Scale Shuffled Convolutional Neural Network (MSSCNN) for diagnosing open-circuit faults in three-level Neutral Point Clamped (NPC) inverters. The MSSCNN achieves high accuracy (99.91%) and robustness against noise while maintaining a lower model complexity compared to existing deep learning models.

What Problem Does This Paper Address?

Three-level Neutral Point Clamped (NPC) inverters are widely used in power electronics due to their high voltage handling capabilities and low harmonic distortion. However, they contain a larger number of power switching devices compared to simpler inverters, increasing the probability of faults. Open-circuit faults, while not immediately catastrophic like short-circuit faults, can lead to significant current distortion, potentially damaging the inverter or connected loads if not promptly detected and diagnosed. Existing fault diagnosis methods often struggle with complex data preprocessing, high model complexity, or sensitivity to noise. This paper aims to address these challenges by developing a more efficient, accurate, and robust fault diagnosis method for open-circuit faults in NPC inverters.

Methods

The core of this work is the proposed Multi-Scale Shuffled Convolutional Neural Network (MSSCNN). The MSSCNN architecture is designed to efficiently extract fault features from the three-phase output current of the inverter. Key techniques employed include:

  • Depthwise Separable Convolution: This technique significantly reduces computational complexity and the number of parameters by separating the convolution operation into two steps: depthwise convolution (applying a single filter per input channel) and pointwise convolution (using 1x1 convolutions to combine channel information).
  • Channel Shuffle: This operation is used to improve information flow between channels, allowing for better feature learning and integration, inspired by architectures like ShuffleNet.
  • Multi-Scale Feature Extraction: The network utilizes convolution kernels of different sizes (e.g., 3x1 and 9x1) within its downsampling modules to capture fault characteristics at various scales.
  • Batch Normalization (BN) and ReLU Activation: Standard deep learning techniques are used for stabilizing training and introducing non-linearity.
  • Depthwise Separable Convolutional Neural Network (MSSCNN) Basic Module and Downsampling Module: These custom-designed modules form the building blocks of the MSSCNN, incorporating the above techniques for efficient feature extraction and dimensionality reduction.

The overall development process involves data acquisition, normalization, dataset production using a sliding window approach, model training, and finally, testing and validation.

Datasets

The study utilizes a dataset generated from experimental measurements of the three-phase output current of a three-level NPC inverter. The data was collected under various operating conditions, including different modulation indices (0.3, 0.6, 0.9) and fundamental frequencies (20 Hz, 30 Hz, 50 Hz). Faults were simulated by deactivating driving signals to individual power switching devices. The dataset comprises 13 conditions (12 single-switch open-circuit faults and one fault-free state). The collected data was split into a training set (5460 samples) and a test set (1170 samples), with each sample having a size of 3x250x1. No new benchmark datasets were introduced; rather, experimental data was collected and processed.

Key Insights and Ideas

  • Depthwise Separable Convolutions are Key to Efficiency: The paper strongly emphasizes how depthwise separable convolutions drastically reduce computational load and model parameters, making the MSSCNN lightweight without sacrificing accuracy.
  • Multi-Scale Features are Crucial for Robustness: Using kernels of different sizes allows the network to capture a broader range of fault-induced current distortions, leading to better noise resistance.
  • Channel Shuffle Enhances Information Flow: This technique, borrowed from efficient network designs, helps integrate features across channels more effectively, contributing to improved diagnostic performance.
  • Simplicity in Preprocessing: The MSSCNN approach avoids complex preprocessing steps like wavelet transforms or Clark transforms, simplifying the overall fault diagnosis pipeline.
  • Clear Separation of Fault States: The t-SNE visualization of the MSSCNN's output data shows a remarkable separation of the 13 different fault and healthy states, indicating excellent feature extraction capabilities.

Key Results

  • High Accuracy: The MSSCNN model achieved an impressive average accuracy of 99.91% across 10 experiments, outperforming other benchmark models.
  • Superior Noise Resistance: In various noise environments (6 dB, 8 dB, 10 dB SNR), the MSSCNN consistently demonstrated higher diagnostic accuracy compared to CNN, ResNet, ShuffleNet V2, and MobileNet V3. For instance, at 10 dB SNR, MSSCNN achieved 99.65% accuracy, while the next best (ResNet) reached 99.14%.
  • Lower Model Complexity: The MSSCNN model exhibits significantly lower complexity in terms of parameters (256,248 vs. >378k for others), FLOPs (2.26M vs. >4M for others), and memory usage. This makes it more suitable for real-time applications with limited computational resources.
  • Faster Convergence: The loss function comparison curves show that the MSSCNN converges faster and reaches a lower loss value than the other models.

Limitations and Future Work

  • Limitations:
    • The study focuses solely on open-circuit faults. The performance on short-circuit faults or combinations of faults is not discussed.
  • While noise resistance is demonstrated, the paper doesn't explore extreme noise levels or different types of noise (e.g., impulse noise).
  • The dataset is generated from a specific experimental setup; its generalizability to different inverter hardware or operating conditions might need further validation.
  • Future Work:
    • The authors suggest that the proposed model can effectively improve fault determination and localization capabilities by clustering the originally mixed data feature distributions, implying potential for more precise localization.
  • Extending the method to diagnose other types of faults (e.g., short-circuit faults) and combinations of faults would be a logical next step.
  • Investigating the application of this MSSCNN architecture to other power electronic systems or different types of faults could be explored.

Summary updated

Abstract

This study constructs a power switching device open-circuit fault diagnosis model for a three-level neutral point clamped inverter based on the multi-scale shuffled convolutional neural network (MSSCNN) and extracts and classifies the fault information contained in the output current of inverters. The model employs depthwise separable convolution and channel shuffle techniques to improve diagnostic accuracy and reduce model complexity. The experimental results show that the new model has lower model complexity, better noise resistance and higher average diagnostic accuracy compared with fault diagnosis models based on CNN, ResNet, ShuffleNet V2 and Mobilenet V3 networks.
BibTeX
@inproceedings{yan2024open,
  author = {Yan Yan and Jiaqi Wu and Yanfei Cao and Bo Liu and Chen Li and Tingna Shi},
  title = {{An Open-Circuit Fault Diagnosis Method for Three-Level Neutral Point Clamped Inverters Based on Multi-Scale Shuffled Convolutional Neural Network}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24061745},
}
Figure