The structural security of civil energy equipment is significant for the steady operation of power supply system, and porcelain bushing type terminal is a typical energy equipment that needs long‐term monitoring. As a nondestructive structural health monitoring method, ultrasonic guided wave (UGW) technology is extremely suitable for state detection of energy equipment. However, most current UGW methods still need to manually select the guided wave features, which rely heavily on the guidance of expert experience. This article presents a deep‐learning method to directly utilize original‐guided wave signals to quantitatively detect the liquid‐level state. Firstly, the original signals were fed into convolutional autoencoder (CAE) to catch the low‐dimension representation and realize the automatic feature extraction. Then, the low‐dimension representations were orderly input into the long short‐term memory (LSTM) recurrent neural network for liquid‐level regression. In feature extraction step, CAE method can effectively extract the useful features and remove the interference and signal distortion. In regression step, both the accuracy and the robustness of proposed method are better than backpropagation network and convolutional neural network. Experimental results show that proposed CAE‐LSTM method can accurately inspect the liquid level by original signals and implement maintenance monitoring.

中文翻译：

---

基于深度学习的导波检测在瓷套管式终端中的液位状态

民用能源设备的结构安全性对供电系统的稳定运行具有重要意义，瓷套管式终端是需要长期监控的典型能源设备。作为一种无损结构健康监测方法，超声波导波（UGW）技术非常适合于能源设备的状态检测。但是，大多数当前的UGW方法仍然需要手动选择导波特征，这在很大程度上取决于专家经验的指导。本文提出了一种深度学习方法，可以直接利用原始导波信号定量检测液位状态。首先，将原始信号输入到卷积自动编码器（CAE）中，以捕获低维表示并实现自动特征提取。然后，低维表示被有序地输入到长短期记忆（LSTM）递归神经网络中以进行液位回归。在特征提取步骤中，CAE方法可以有效地提取有用的特征，消除干扰和信号失真。在回归步骤中，所提方法的准确性和鲁棒性均优于反向传播网络和卷积神经网络。实验结果表明，提出的CAE‐LSTM方法可以通过原始信号准确地检查液位并进行维护监控。所提方法的准确性和鲁棒性均优于反向传播网络和卷积神经网络。实验结果表明，提出的CAE‐LSTM方法可以通过原始信号准确地检查液位并进行维护监控。所提方法的准确性和鲁棒性均优于反向传播网络和卷积神经网络。实验结果表明，提出的CAE‐LSTM方法可以通过原始信号准确地检查液位并进行维护监控。