应用CNN-BiLSTM-SEAttention模型预测电磁 超声测厚间隙

打开文本图片集

中图分类号：TB9;TB552;TP183 文献标志码：A 文章编号：1674-5124（2025）09-0158-09

Abstract： Reasonable measurement gap is an important factor in ensuring the stability and accuracy of electromagnetic ultrasonic non-contact thickness measurement.However， there is stillalack of electromagnetic ultrasonic measurement methods capable of simultaneously monitoring the gap during the thickness measurement process.To more accurately monitor the gap state during ultrasonic non-contact thickness measurement， a hybrid prediction model （CNN-BiLSTM-SE Attntion） is proposed by combining the characteristicsof electromagneticultrasonic signals.The model integrates a convolutional neural network （CNN），a bidirectional long short-term memory network （BiLSTM），and a squeeze-and-excitation attention mechanism （SE Attention）. A dataset was established and preprocessed through gap calibration experiments. The network model was constructed to extract local features of ultrasonic echo signals using CNN，capture long-term dependencies bidirectionally via BiLSTM， and incorporate the SE attention mechanism to automatically assgn weights to hidden layers，thereby enhancing key ultrasonic feature information. Based on the gapcalibration data，the model was trained and optimized. When tested on datasets with different signal-tonoise ratios， the model achieved an average accuracy of 95.74% with prediction errors within 0.10mm Compared to the CNN，BiLSTM， CNN-BiLSTM models，and the fitted function method， the proposed model demonstrates better prediction accuracy and noise resistance. It can effectively predict the gap state during electromagnetic ultrasonic thickness measurement.

Keywords： electromagnetic ultrasonic thickening; gap prediction; convolutional neural network; bi-directional longand short-term memory network; SE attention mechanism

0 引言

电磁超声测厚具有测量速度快、精度高、非接触等特点，在大型薄壁件的壁厚在机扫描测量中具有优势[1]。（剩余8958字）