Scholarlink

<p>Deep learning architectures, especially deep recurrent neural networks (DRNNs), are widely recognized for their high accuracy in fault detection and diagnosis (FDD) applications. However, their high computational and memory complexity presents significant challenges, particularly when deploying them on edge devices with limited resources. In answer to these challenges, model compression techniques offer an attractive solution. This research introduces a novel model compression approach for fault detection and diagnosis in modular multilevel converters (MMCs). The data is initially represented as interval-valued to effectively handle uncertainties such as measurement errors, noise, and variable variability. The long-short-term memory (LSTM) model is then developed and tested for fault diagnosis across various measurement uncertainties. To further enhance the model’s efficiency, a model compression technique based on projection via principal component analysis (PCA) is applied, significantly decreasing the number of learnable parameters and memory requirements. Importantly, the compressed LSTM model achieves a reduction of around 50% in the learnable parameters and memory usage, while maintaining high accuracy. Furthermore, the model’s energy consumption is reduced by 70.34%, from 897.187 Joules (J) for the original model to 266.015 J after compression, highlighting the model’s suitability for deployment in resource-constrained environments without sacrificing performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PCA-based compression of deep LSTMs: a robust framework for uncertain industrial systems

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Zahra Yahyaoui,
  • Mansour Hajji,
  • Yassine Bouazzi

摘要

Deep learning architectures, especially deep recurrent neural networks (DRNNs), are widely recognized for their high accuracy in fault detection and diagnosis (FDD) applications. However, their high computational and memory complexity presents significant challenges, particularly when deploying them on edge devices with limited resources. In answer to these challenges, model compression techniques offer an attractive solution. This research introduces a novel model compression approach for fault detection and diagnosis in modular multilevel converters (MMCs). The data is initially represented as interval-valued to effectively handle uncertainties such as measurement errors, noise, and variable variability. The long-short-term memory (LSTM) model is then developed and tested for fault diagnosis across various measurement uncertainties. To further enhance the model’s efficiency, a model compression technique based on projection via principal component analysis (PCA) is applied, significantly decreasing the number of learnable parameters and memory requirements. Importantly, the compressed LSTM model achieves a reduction of around 50% in the learnable parameters and memory usage, while maintaining high accuracy. Furthermore, the model’s energy consumption is reduced by 70.34%, from 897.187 Joules (J) for the original model to 266.015 J after compression, highlighting the model’s suitability for deployment in resource-constrained environments without sacrificing performance.