Source-linked AI summary
A Hierarchical Deep Convolutional Neural Network and Gated Recurrent Unit Framework for Structural Damage Detection
Jianxi Yang, Likai Zhang, Cen Chen, Yangfan Li, Ren Li, Guiping Wang, Shixin Jiang, Zeng Zeng
TL;DR
Structural damage detection methods have treated spatial or temporal relationships separately, leaving their joint modeling as a research gap. HCG combines CNN and GRU components to learn both, and the paper reports improved performance across two structural-health-monitoring datasets.
Problem
Existing structural damage detection methods had not jointly modeled spatial relations among sensors and temporal sequence relations.
Source: §1 · p. 3Method
HCG uses a low-level CNN for spatial and short-term temporal features and a high-level GRU for long-term temporal dependencies.
Source: §1 · p. 4, §4.1 · p. 9Results
Across the IASC-ASCE benchmark and three-span bridge dataset, HCG outperforms other methods; it is 1.5 times faster than RNN-based approaches.
Source: §1 · p. 4, §6 · pp. 24–25Takeaways & Limitations
The reported evaluations support HCG’s use for structural damage detection on the two evaluated datasets.
Source: §6 · pp. 24–25The formulation assumes all sensors are synchronized.
Source: §3.2 · p. 7
Abstract
from arXiv · showhide
Structural damage detection has become an interdisciplinary area of interest for various engineering fields, while the available damage detection methods are being in the process of adapting machine learning concepts. Most machine learning based methods heavily depend on extracted ``hand-crafted" features that are manually selected in advance by domain experts and then, fixed. Recently, deep learning has demonstrated remarkable performance on traditional challenging tasks, such as image classification, object detection, etc., due to the powerful feature learning capabilities. This breakthrough has inspired researchers to explore deep learning techniques for structural damage detection problems. However, existing methods have considered either spatial relation (e.g., using convolutional neural network (CNN)) or temporal relation (e.g., using long short term memory network (LSTM)) only. In this work, we propose a novel Hierarchical CNN and Gated recurrent unit (GRU) framework to model both spatial and temporal relations, termed as HCG, for structural damage detection. Specifically, CNN is utilized to model the spatial relations and the short-term temporal dependencies among sensors, while the output features of CNN are fed into the GRU to learn the long-term temporal dependencies jointly. Extensive experiments on IASC-ASCE structural health monitoring benchmark and scale model of three-span continuous rigid frame bridge structure datasets have shown that our proposed HCG outperforms other existing methods for structural damage detection significantly.
1. Introduction
Structural damage detection needs to account for both relationships among sensors and temporal dependencies, which existing approaches had not jointly modeled. HCG combines CNN and GRU components and reports advantages over existing methods.
1. Introduction
- Prior structural damage detection approaches had not jointly considered spatial relations among sensors and temporal sequence relations.The introduction also notes challenges of slow RNN-based approaches and high memory use in CNN-based approaches.Source: §1 · pp. 3–4
- HCG combines CNNs for spatial and short-term temporal patterns with GRUs for long-term temporal dependencies.The framework is described as hierarchical, with a low-level CNN and a high-level recurrent model.Source: §1 · p. 4
- HCG significantly improves performance over existing deep-learning models on the IASC-ASCE benchmark.The introduction also names the three-span bridge dataset among the evaluation datasets.Source: §1 · p. 4
- HCG is reported to be 1.5 times faster than RNN-based approaches.Source: §1 · p. 4
- Compared with CNN-based approaches, HCG requires roughly 10% data memory usage and allows latent feature extraction.Source: §1 · p. 4
2. Related Work
Structural damage detection research includes model-driven methods and data-driven methods, including shallow and deep learning approaches. The related work identifies constraints in model-driven approaches and a gap in deep-learning methods’ treatment of spatial and temporal information.
2.1. Model-driven Method
- Model-driven methods use mathematical models and physical theorems, but complex model construction, real-excitation simulation, online updating, and quantitative interpretation present challenges.These constraints are described as limiting flexible application and requiring manual intervention.Source: §2.1 · p. 5
2.2. Shallow Models of Data-driven Methods
- Shallow data-driven methods include SVM, k-nearest neighbors, random forests, semi-supervised classification, PCA, and self-organizing maps.Examples include multi-sensor data fusion for damage detection and classification methods using structural time-series responses.Source: §2.2 · pp. 5–6
2.3. Deep Models of Data-driven Methods
- Deep-learning work includes MLP and CNN approaches for structural damage identification, including feature extraction, classification, and anomaly detection.Examples include methods evaluated on the IASC-ASCE benchmark and approaches that convert sensor time series into images.Source: §2.3 · p. 6
- The reviewed deep neural network methods had not considered acceleration time-series features and spatial correlations together.Source: §2.3 · pp. 6–7
3. Problem Definition and Analysis
The paper frames structural damage detection as diagnosing a structure’s state from historical sensor signals. It models synchronized sensor time series as a classification input and motivates a hierarchical CNN–GRU approach to learn spatial and temporal dependencies.
3.1. Problem Description
- Structural damage detection diagnoses the state of a structure’s materials and parts over its lifetime using recorded sensor history.The passage notes that this history can support prognosis, including damage evolution and residual life.Source: §3.1 · p. 7
3.2. Problem Modeling
- The model concatenates synchronized time series from all sensors into an input matrix and maps it to an estimated damage category.The categories include healthy and several damage cases; the task is formulated as classification.Source: §3.2 · pp. 7–8
3.3. Problem Analysis
- Bridge sensors can have spatially dependent patterns, while readings over time are also sequentially dependent, so both dimensions matter for detection.Source: §3.3 · p. 8
- The proposed hierarchy assigns spatial and short-term temporal feature extraction to a low-level CNN and long-term dependencies to a high-level GRU.Source: §3.3 · pp. 8–9
4. Hierarchical CNN and GRU framework (HCG)
HCG uses a convolutional component to capture interactions among sensors and short-term temporal patterns, then a recurrent component to model long-term dependencies and predict structural damage states.
4.1. Architecture of HCG
- HCG’s low-level convolutional component captures sensor interactions and short-term temporal dependencies, while its high-level recurrent component models long-term dependencies.The framework takes multiple-sensor time series as input and uses a softmax layer to predict the structure’s damage state.Source: §4.1 · p. 9, Figure 2 · p. 9
4.2. Convolutional Component
- Convolution kernels span all sensors but are short in time, learning spatial correlations and short-term temporal patterns whose latent-vector outputs preserve the input sequence length.Zero-padding keeps input and output lengths the same.Source: §4.2 · p. 10, Figure 3 · p. 10
4.3. Recurrent Component
- The recurrent component feeds convolutional outputs into GRUs to learn long-term temporal dependencies and uses the top-layer GRU’s final hidden state to predict damage category.A fully connected network and softmax layer generate the final output.Source: §4.3 · pp. 11–12, Figure 4 · p. 11
4.4. Loss Function
- Training uses mean squared error (L2-loss) as the objective, comparing predicted classes with ground-truth damage states.The loss is defined over the training samples.Source: §4.4 · p. 13
5. Performance Evaluation and Discussions
The evaluation uses two structural-health-monitoring datasets: a scale model of a three-span continuous rigid-frame bridge and the Phase I IASC-ASCE benchmark. Their sensor layouts, excitation, damage states, and signals define the experimental data.
5.1.1. TCRF Bridge Dataset
- The TCRF Bridge dataset uses a 20:1 scale model with 18 acceleration sensors and trolley-induced dynamic loading.Twelve sensors measure vertically at the beam bottom and six horizontally at the web; concentrated force creates floor cracks representing damage states.Source: §5.1.1 · pp. 13–14
- TCRF damage is represented by middle-floor main-span cracks, with four damage states whose degree depends on concentrated-force strength.Acceleration is sampled at 8 kHz as a car passes over the bridge; responses differ among damage states and are used by HCG to learn spatial and temporal features.Source: §5.1.1 · pp. 14–15
5.1.2. IASC-ASCE Benchmark Dataset
- The Phase I IASC-ASCE benchmark is a simulated structure widely used for structural-damage classification and provides a common comparison platform.The study uses the Phase I simulated benchmark; the benchmark program also includes Phase II and experimental benchmarks.Source: §5.1.2 · pp. 15–16
- The IASC-ASCE setup has 16 sensors, roof-shaker excitation, and 250 Hz acceleration recordings of length 20,000 from column 4 on each floor.The structure has 120 degrees of freedom and symmetric mass distribution; collected sensory data are fed to HCG.Source: §5.1.2 · p. 16
5.2. Experimental Setup
The experiments use held-out data splits, four classification metrics, and deep-learning baselines configured for a broadly fair comparison in trainable parameter count.
5.2.1. Experiment Settings
- The datasets are split into 60% training, 20% validation, and 20% test data.Models are built with Keras and trained on a server with four Tesla P100 GPUs and eight E5-2620V4 CPUs.Source: §5.2.1 · pp. 16–17
5.2.2. Evaluation metrics
- Accuracy, precision, recall, and F1 are used to evaluate HCG and the compared baselines.Accuracy is defined using true positives, true negatives, false positives, and false negatives.Source: §5.2.2 · p. 17
5.2.3. Compared Baselines
- For fair comparison, all deep-learning models were adjusted to have very similar numbers of trainable parameters and trained using Adam with learning rate 0.001 and batch size 64.HCG uses a two-layer convolutional component with 64 filters and a two-layer GRU with 128-dimensional hidden states in each layer.Source: §5.2.3 · p. 18
- The GRU and LSTM baselines each stack three recurrent layers, each with 64-dimensional hidden states.The LSTM baseline adapts the method in.Source: §5.2.3 · p. 18
5.3. Results of the TCRF Bridge Dataset
On the TCRF Bridge dataset, HCG outperforms DNN, CNN, LSTM, and GRU across all reported evaluation metrics. It also converges faster than the competing deep-learning models and achieves 94.04% final accuracy.
5.3. Results of the TCRF Bridge Dataset
- HCG converges more quickly than the other deep-learning competitors during training.The training curves are shown for loss and accuracy.Source: §5.3 · p. 19, Figure 9 · p. 19
- HCG outperforms DNN, CNN, LSTM, and GRU on all evaluation metrics for the TCRF Bridge dataset.The results average 10 runs; the authors report that HCG performs better than single CNN and GRU models and attribute this to considering spatial and temporal dependencies together.Source: §5.3 · pp. 18–19, Table 3 · p. 19
- 94.04% is HCG's final accuracy on the TCRF Bridge test set.The reported confusion matrices compare CNN, GRU, and HCG; the authors note that conditions 3 and 4 are harder to distinguish because their data characteristics are similar.Source: §5.3 · pp. 19–20, Figure 10 · p. 20
5.4. Performance Analysis of Hyper-parameters in the TCRF Bridge Dataset
The hyper-parameter analysis compares network depth and neuron count across DNN, CNN, LSTM, RNN, and HCG on the TCRF Bridge dataset. HCG retains higher accuracy across the tested hyper-parameter sets.
5.4. Performance Analysis of Hyper-parameters in the TCRF Bridge Dataset
- The network-structure experiment compares 2-, 3-, 4-, and 5-layer configurations with the same neuron count across models.The comparison includes DNN, CNN, LSTM, RNN, and HCG on the TCRF Bridge dataset.Source: §5.4 · p. 20, Table 4 · p. 21
- The neuron-count experiment uses the 4-layer configuration because it achieved higher accuracy than the other layer configurations.Different neuron counts are compared across the models; the table notation [40, 70, 32, 32] denotes the counts in the four layers.Source: §5.4 · p. 21, Table 5 · p. 21
- HCG has higher accuracy than the neural-network models across the tested hyper-parameter sets.The authors relate this performance to learning sensor interactions and short-term temporal dependencies at a low level, with a high-level component handling long-term temporal dependencies.Source: §5.4 · p. 21
5.5. Results of the IASC-ASCE Benchmark Dataset
On the IASC-ASCE benchmark, HCG shows advantages over the compared baselines, and it retains higher accuracy across the tested hyperparameter sets.
5.5. Results of the IASC-ASCE Benchmark Dataset
- The IASC-ASCE training process is illustrated with loss and accuracy curves for the deep-learning baselines and HCG.Figure 11 contains the accuracy and loss curves.Source: §5.5 · pp. 21–22, Figure 11 · p. 22
- HCG shows advantages over the compared baselines on the IASC-ASCE Benchmark Dataset.Table 6 summarizes the experimental results for this dataset.Source: §5.5 · p. 22, Table 6 · p. 22
5.6. Performance Analysis of Hyper-parameters in the IASC-ASCE Benchmark Dataset
- HCG has higher accuracy than neural network models across the tested hyperparameter sets.The comparisons examine network depth and neuron count; the neuron-count experiments use the four-layer set because it achieved higher accuracy than the others.Source: §5.6 · p. 23, Table 7 · p. 23, Table 8 · p. 23
5.7. Running Time and Model Sizes
HCG combines faster epoch training than GRU-based models with the smallest model size among the compared CNN, LSTM, and GRU methods.
5.7. Running Time and Model Sizes
- HCG-adopted models are roughly 1.5 times faster than GRU-based models for one epoch of training.Adding HCG’s hierarchical structure only slightly affects computational speed; Table 9 compares epoch time and model sizes.Source: §5.7 · pp. 23–24, Table 9 · p. 24
- HCG has the smallest model sizes compared with CNN, LSTM, and GRU.The reported GPU memory consumption includes model parameters and input data; parameter memory is the same across the models in these experiments.Source: §5.7 · p. 24, Table 9 · p. 24
6. Conclusions
The paper concludes that HCG outperforms other methods for structural damage detection. Its hierarchical CNN–GRU design captures spatial and short- and long-term temporal features, while future work may consider other machine-learning techniques to improve computational efficiency.
6. Conclusions
- HCG uses a low-level convolutional component for spatial and short-term temporal features and a high-level recurrent component for long-term temporal dependencies.The framework is evaluated on the ASCE benchmark and a three-span continuous rigid frame bridge structure dataset.Source: §6 · p. 24
- Experiments demonstrate that HCG outperforms other methods for structural damage detection.The experiments cover the ASCE benchmark and a three-span continuous rigid frame bridge structure dataset.Source: §6 · pp. 24–25
- Future research may consider stochastic configuration networks and graph convolutional neural networks to further improve computational efficiency.Source: §6 · p. 25