Application of CNN-LSTM in infrastructure performance grade assessment

Abstract

Bridges are crucial infrastructure of transportation. However, due to various adverse factors such as mechanical loading, followed by environmental corrosion and material aging, the structural performance of bridges almost inevitably degrade over a prolonged service life span. Accurate assessment of the refined performance grade, thereby, is of great significance not only for ensuring the safe utilization of bridges but also for formulating science-based maintenance strategies and achieving economy in life-cycle management. Conventional bridge performance assessment methods mainly depend on manual field inspection and simple statistical analysis. These traditional approaches are highly subjective and inefficient. Further, these methods are inevitably unable to identify both multi-component correlations and temporal degradation behavior, and therefore lack reliability of high refinement results such as seven-tier grading. This paper constructs a hybrid Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) model for infrastructure performance grade assessment. The model employs a one-dimensional CNN to extract local correlation features from multi-structural component scores of bridges, and leverages LSTM to capture temporal degradation trends from years of periodic bridge inspection data, enabling seven-level grade assessment of bridge infrastructure performance. The publicly available bridge condition dataset from the New York State Department of Transportation (NYSDOT) is used for empirical research. Five mainstream algorithms, including XGBoost, single CNN, single LSTM, Gated Recurrent Unit (GRU), and CNN-GRU, are set as control models. Four indicators, namely Accuracy, Macro-Precision, Macro-Recall and Macro-F1 Score, are adopted to comprehensively evaluate model performance. The experimental results show that the proposed CNN-LSTM model achieves an assessment accuracy of 84.6%, with Macro-Precision, Macro-Recall and Macro-F1 Score reaching 83.9%, 83.2% and 83.5% respectively. All performance indicators outperform those of all control models. Tests with random data perturbations show that the model only causes small changes in evaluation metrics and has strong robustness. The results of this study verify that the CNN-LSTM hybrid model is well suited to capture the temporal monitoring features of bridge structures. It can reliably complete seven-tier refined performance grading and provide solid technical support for the intelligent assessment and maintenance decision-making of bridge infrastructure.