Scholarlink

<p>Traditional deep learning methods for motor fault diagnosis have primarily focused on signal-based classification. Intelligent operation and maintenance encompasses both fault diagnosis and maintenance decision-making. However, transitioning from vibration fault signal diagnosis to generating maintenance decision texts presents challenges such as data heterogeneity between numerical signal data and textual decision data. Relying solely on existing large models is insufficient to achieve a closed-loop process from intelligent diagnosis to maintenance decision-making. Thus, this paper proposes a motor fault diagnosis method based on a Dual-Channel Convolutional Neural Network with Probabilistic Feature Fusion and Lightweight Transformer (DCCNN-LTransformer), as well as a decision-making method that utilizes a fault knowledge graph. In the feature extraction stage, a novel probabilistic feature fusion mechanism is introduced to extract features from vibration signals. A dual-channel 1D convolutional neural network framework is employed to capture feature signals and obtain positional encodings. For feature classification within the Transformer, an Enhanced Dynamic Window-hybrid Attention (EDWHA) mechanism is proposed to improve the local analysis capability for fault signals. Finally, after the fault diagnosis algorithm classifies the faults, the BERT-BiLSTM-CRF algorithm is employed to extract knowledge related to current fault diagnosis and maintenance decision-making, thereby constructing a structured fault knowledge graph database. By integrating a general-purpose LLM with the fault knowledge graph, the system can generate maintenance decisions for various motor faults, ultimately forming a closed-loop maintenance decision-making process. To validate the algorithm’s adaptability under various conditions, three distinct fault datasets were utilized. Experimental results demonstrate that the proposed algorithm outperforms CNN, Transformer, and CNN-Transformer models, confirming its effectiveness. A fault knowledge graph was constructed, and a motor fault diagnosis and maintenance decision-making platform was established, thereby verifying the reliability of this study.</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

Research on motor diagnosis and maintenance decision-making technology based on DCCNN-L transformer and fault knowledge graph

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月27日

查看全文
  • Wu Shufang,
  • Wang Xinwen,
  • Liu Zeyu,
  • Shi Gaoxian

摘要

Traditional deep learning methods for motor fault diagnosis have primarily focused on signal-based classification. Intelligent operation and maintenance encompasses both fault diagnosis and maintenance decision-making. However, transitioning from vibration fault signal diagnosis to generating maintenance decision texts presents challenges such as data heterogeneity between numerical signal data and textual decision data. Relying solely on existing large models is insufficient to achieve a closed-loop process from intelligent diagnosis to maintenance decision-making. Thus, this paper proposes a motor fault diagnosis method based on a Dual-Channel Convolutional Neural Network with Probabilistic Feature Fusion and Lightweight Transformer (DCCNN-LTransformer), as well as a decision-making method that utilizes a fault knowledge graph. In the feature extraction stage, a novel probabilistic feature fusion mechanism is introduced to extract features from vibration signals. A dual-channel 1D convolutional neural network framework is employed to capture feature signals and obtain positional encodings. For feature classification within the Transformer, an Enhanced Dynamic Window-hybrid Attention (EDWHA) mechanism is proposed to improve the local analysis capability for fault signals. Finally, after the fault diagnosis algorithm classifies the faults, the BERT-BiLSTM-CRF algorithm is employed to extract knowledge related to current fault diagnosis and maintenance decision-making, thereby constructing a structured fault knowledge graph database. By integrating a general-purpose LLM with the fault knowledge graph, the system can generate maintenance decisions for various motor faults, ultimately forming a closed-loop maintenance decision-making process. To validate the algorithm’s adaptability under various conditions, three distinct fault datasets were utilized. Experimental results demonstrate that the proposed algorithm outperforms CNN, Transformer, and CNN-Transformer models, confirming its effectiveness. A fault knowledge graph was constructed, and a motor fault diagnosis and maintenance decision-making platform was established, thereby verifying the reliability of this study.