A robust deep learning-based damage identification approach for SHM considering missing data

Abstract: Data-driven method for Structural Health Monitoring (SHM), that mine the hidden structural performance from the correlations among monitored time series data, has received widely concerns recently. However, missing data significantly impacts the conduction of this method. Missing data is a frequently encountered issue in time series data in SHM and many other real-world applications, that harms to the standardized data mining and downstream tasks, such as condition assessment. Imputation approaches based on spatiotemporal relations among monitoring data are developed to handle this issue, however, no additional information is added during imputation. This paper thus develops a robust method for damage identification that considers the missing data occasions, based on long-short term memory (LSTM) model and dropout mechanism in the autoencoder (AE) framework. Inputs channels are randomly dropped to simulate the missing data in training, and reconstruction errors are used as the loss function and the damage indicator. Quasi-static response (cable tension) of a cable-stayed bridge released in 1st IPC-SHM is employed to verify this proposed method, and results show that the missing data imputation and damage identification can be implemented together in a unified way.

• The paper proposes a novel deep learning framework integrating LSTM autoencoders and dropout to simulate missing data conditions.
• It demonstrates superior damage detection in SHM compared to conventional DNN models using reconstruction errors on real bridge data.
• The approach enhances real-time monitoring reliability by effectively addressing missing sensor data without reliance on imputation techniques.

Deep Learning-Based Damage Identification in Structural Health Monitoring Considering Missing Data

The paper "A Robust Deep Learning-Based Damage Identification Approach for SHM Considering Missing Data" proposes a novel framework to address the challenges posed by missing data in Structural Health Monitoring (SHM). The approach leverages a combination of Long-Short Term Memory (LSTM) architecture within an autoencoder (AE) framework to effectively identify structural damages in the presence of incomplete datasets.

Background and Motivation

In the domain of SHM, which relies heavily on time-series data collected from various sensors, the occurrence of missing data is a common obstacle. Such gaps can significantly hinder the accuracy and reliability of damage assessments and condition evaluations. Conventional methods typically address missing data through imputation techniques that fill gaps without integrating additional information into the damage identification process. This study introduces an innovative solution by integrating the dropout mechanism in the data-driven damage identification model to seamlessly deal with missing data scenarios.

Methodology

The core of the proposed method is structured around an LSTM-based autoencoder. LSTM networks are adept at capturing temporal dependencies, making them well-suited for processing sequential data typical in SHM applications. The model incorporates dropout layers to simulate the presence of missing data during training, enhancing its robustness. By randomly dropping out input channels and utilizing reconstruction errors as indicators, the model performs simultaneous data reconstruction and damage identification.

Key components of the method include:

1. LSTM Autoencoder Framework: Utilizes LSTM cells in an encoder-decoder setup, capturing spatiotemporal dependencies inherent in SHM data.
2. Dropout Mechanism: Simulates various missing data conditions during training, ensuring that the model learns to recognize and compensate for incomplete data inputs.
3. Reconstruction Error as Damage Indicator: The discrepancy between reconstructed and original data is utilized to signal potential structural damage, leveraging the model’s learned representation as a baseline.

Results and Analysis

Validation of the model was conducted using cable tension data from a real-world cable-stayed bridge, collected under the International Project Competition for SHM. The model was evaluated under various missing data scenarios, including discrete, continuous, and total channel loss. Experimental results demonstrated that the LSTM-based autoencoder significantly outperforms conventional DNN models in scenarios with high degrees of data loss, effectively reconstructing missing data and maintaining reliable damage detection capabilities.

The LSTM model exhibited an ability to memorize historical data patterns, allowing it to distill complex temporal correlations that purely spatial models, such as standard DNNs, could not capture. This results in more accurate reflections of structural health states even when significant portions of data are missing.

Implications and Future Directions

The implications of this study are notable for both theoretical advancement and practical application in SHM. By integrating dropout directly into the damage identification process rather than treating missing data as a preprocessing hurdle, the authors demonstrate a novel method for enhancing the robustness and efficiency of SHM systems. This holistic approach not only reduces the dependency on high-quality, complete data but also strengthens the reliability of real-time monitoring solutions.

Future research directions could involve extending this framework to different types of structural systems and incorporating additional data modalities to further increase robustness. Moreover, the exploration of adaptive models that fine-tune their parameters based on evolving conditions and datasets holds potential for enhancing the practical utility of deep learning in SHM applications.

In summary, the study provides a comprehensive and robust strategy to tackle the pervasive issue of missing data in SHM, paving the way for more resilient monitoring systems in complex engineering applications.