Scholarlink

<p>To address the issues of insufficient utilization of temporal information and limited feature extraction capability leading to low diagnostic accuracy in existing analog circuit fault diagnosis models, a fault diagnosis model combining an Improved Bidirectional Temporal Convolutional Network (IBiTCN) and a Bidirectional Long Short-Term Memory Network (BiLSTM) is proposed. Firstly, the residual connections of the Bidirectional Temporal Convolutional Network (BiTCN) are improved through adopting only unidirectional residual connections to guarantee gradient stability and diminish model complexity, Meanwhile, The traditional Rectified Linear Unit (ReLU) activation function is superseded by the Gaussian Error Linear Unit (GELU) using pre-activation manner to optimize the residual blocks structure and enhance the model’s nonlinear expressive capability for initial local feature extraction. Secondly, Residual Self-Attention mechanism (RSA) is incorporated after bidirectional feature fusion to highlight critical fault features. Finally, the advantage of BiLSTM in learning long sequences is utilized to construct global features, and further fault classification is achieved through the Softmax layer. The study validates the proposed model using three circuits as test subjects: the Sallen-Key bandpass filter circuit, a two-stage four opamp biquad low-pass filter circuit, and a leapfrog filter circuit. Across 20 tests, the average diagnostic accuracy reached 99.72%, 96.13%, and 91.17%, respectively. The experimental results demonstrate that the proposed model effectively improves diagnostic accuracy while maintaining strong stability.</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

Fault Diagnosis of Analog Circuits Using an Improved BiTCN Combined with BiLSTM

  • 期刊论文

  • 发表日期：

    2025年8月23日

查看全文
  • Xianjun Du,
  • Shenglong Jia,
  • Yuxiang Hu,
  • Yaqi Wang

摘要

To address the issues of insufficient utilization of temporal information and limited feature extraction capability leading to low diagnostic accuracy in existing analog circuit fault diagnosis models, a fault diagnosis model combining an Improved Bidirectional Temporal Convolutional Network (IBiTCN) and a Bidirectional Long Short-Term Memory Network (BiLSTM) is proposed. Firstly, the residual connections of the Bidirectional Temporal Convolutional Network (BiTCN) are improved through adopting only unidirectional residual connections to guarantee gradient stability and diminish model complexity, Meanwhile, The traditional Rectified Linear Unit (ReLU) activation function is superseded by the Gaussian Error Linear Unit (GELU) using pre-activation manner to optimize the residual blocks structure and enhance the model’s nonlinear expressive capability for initial local feature extraction. Secondly, Residual Self-Attention mechanism (RSA) is incorporated after bidirectional feature fusion to highlight critical fault features. Finally, the advantage of BiLSTM in learning long sequences is utilized to construct global features, and further fault classification is achieved through the Softmax layer. The study validates the proposed model using three circuits as test subjects: the Sallen-Key bandpass filter circuit, a two-stage four opamp biquad low-pass filter circuit, and a leapfrog filter circuit. Across 20 tests, the average diagnostic accuracy reached 99.72%, 96.13%, and 91.17%, respectively. The experimental results demonstrate that the proposed model effectively improves diagnostic accuracy while maintaining strong stability.