This study proposes a framework for predicting the remaining useful lifetime (RUL) of concrete structures using acoustic emission (AE) data. This framework presents two primary contributions: state indicator (SI) construction based on the Mann-Whitney test (MWUT) and RUL prognosis using the damage accumulation indicator (DA) calculated from the SIs. The proposed indicators display an excellent description of the instantaneous and accumulated damage to the specimen, respectively. In the first stage, a deep neural network (DNN)-based constructor learns how to compute the SIs using raw AE data with MWUT-based SIs as training labels. The proposed SIs represent the instantaneous damage sustained in a time-step by highlighting the difference in the AE activity between the data of an unknown state from this time-step and a reference normal working condition. Through the aggregation of the SI values over time, a DA can be obtained, which represents the accumulation of damage in the time steps from the beginning to the time of inspection. Given a partial DA curve as the input, a prognosis model based on the gated recurrent unit (GRU) learns to predict the future DA values, from which RUL can be derived in the second stage. To evaluate the proposed method, multiple reinforced concrete beams (RCBs) were loaded under a four-point bending test scenario to collect AE data throughout their lifetimes. The results obtained from the validation show that the proposed SI can provide meaningful insight into the concrete structure’s deterioration in comparison to the other indicators constructed from statistical features. Furthermore, the GRU-based prognosis model also outperforms reference methods based on recurrent neural networks (RNNs) and long short-term memory (LSTM) with a significantly smaller prediction error.

中文翻译：

---

使用 Mann-Whitney U 测试状态指标和深度学习的混凝土结构剩余有用寿命预测模型


本研究提出了一个使用声发射（AE）数据预测混凝土结构剩余使用寿命（RUL）的框架。该框架提出了两个主要贡献：基于曼惠特尼测试 (MWUT) 的状态指标 (SI) 构建和使用从 SI 计算的损伤累积指标 (DA) 的 RUL 预测。所提出的指标分别很好地描述了样本的瞬时损伤和累积损伤。在第一阶段，基于深度神经网络 (DNN) 的构造函数学习如何使用原始 AE 数据和基于 MWUT 的 SI 作为训练标签来计算 SI。所提出的 SI 通过突出显示该时间步的未知状态数据与参考正常工作条件之间的 AE 活动差异来表示该时间步中遭受的瞬时损坏。通过对随时间变化的 SI 值进行聚合，可以获得 DA，它表示从开始检查到检查时的时间步长内损伤的累积。给定部分 DA 曲线作为输入，基于门控循环单元 (GRU) 的预后模型学习预测未来的 DA 值，从中可以在第二阶段导出 RUL。为了评估所提出的方法，在四点弯曲测试场景下加载多个钢筋混凝土梁 (RCB)，以收集其整个生命周期的 AE 数据。验证结果表明，与根据统计特征构建的其他指标相比，所提出的 SI 可以提供对混凝土结构恶化的有意义的了解。 此外，基于 GRU 的预后模型还优于基于循环神经网络 (RNN) 和长短期记忆 (LSTM) 的参考方法，且预测误差明显更小。