摘要：

针对深海极端环境下水下温压传感器故障诊断面临的低信噪比、时空耦合特征提取不充分及故障样本不平衡等关键挑战，提出一种生成式数据增强CNN-LSTM网络(Generative Data-augmented CNN-LSTM Network,GD-CLNet)。该方法首先通过“斯皮尔曼等级相关系数-随机森林”二级筛选策略提取温度(T)、压强(P)等7项核心判别特征，优化模型输入维度；其次，利用深度卷积生成对抗网络(DCGAN)对稀疏故障样本进行增广，有效缓解数据类别不平衡问题；在此基础上，GD-CLNet协同发挥CNN局部空间特征提取与LSTM长时序依赖建模的优势，实现传感器健康状态的精准识别。实验结果表明，GD-CLNet在温度数据集上准确率达0.920、F₁值为0.892，在压强数据集上准确率为0.889、F₁值为0.912，性能显著优于传统机器学习与单一深度学习模型，为复杂水下环境装备的状态监测与预测性维护提供了高精度、高可靠性的技术解决方案。

关键词: 水下温压传感器, 故障诊断, 卷积长短期记忆网络(CNN-LSTM), 深度卷积生成对抗网络(DCGAN), 深度学习

Abstract:

To address the key challenges in fault diagnosis of underwater temperature-pressure sensors in deep-sea extreme environments—including low signal-to-noise ratio (SNR), inadequate extraction of spatiotemporally coupled features, and imbalanced fault sample distribution—this study proposes a Generative Data-augmented CNN-LSTM Network (GD-CLNet). First, a two-stage feature selection strategy combining Spearman’s rank correlation coefficient and Random Forest (RF) is employed to extract seven core discriminative features, such as temperature (T) and pressure (P), thereby optimizing the model’s input dimensionality. Second, the Deep Convolutional Generative Adversarial Network (DCGAN) is utilized to augment sparse fault samples, effectively mitigating the problem of imbalanced data categories. On this basis, theGD-CLNet synergistically leverages the advantages of CNN in extracting local spatial features and LSTM in modeling long-term temporal dependencies to achieve precise identification of sensor health states. Experimental results demonstrate that theGD-CLNet achieves an accuracy of 0.920 and an F₁-score of 0.892 on the temperature dataset, and an accuracy of 0.889 and an F₁-score of 0.912 on the pressure dataset. Its performance is significantly superior to traditional machine learning and single deep learning models, providing a high-precision and reliable technical solution for the condition monitoring and predictive maintenance of equipment in complex underwater environments.

Key words: underwater temperature-pressure sensor, fault diagnosis, CNN-LSTM, deep convolutional generative adversarial network(DCGAN), deep learning