TL;DR
This paper introduces an Extreme Learning Machine (ELM) approach to predict the settlement of soft foundations beneath fishbone-shaped dikes in the Yangtze River estuary. The ELM model demonstrates high accuracy and efficiency, even with limited training data, offering a valuable tool for assessing dike stability.
What Problem Does This Paper Address?
Fishbone-shaped dikes are crucial for coastal protection and channel management, particularly in dynamic environments like the Yangtze River estuary. These dikes are often constructed on soft, compressible soils, making their foundation settlement a critical factor for stability and long-term effectiveness. Traditional methods for predicting settlement can be complex, require extensive data, or have limitations in capturing future trends. This paper addresses the need for a more efficient and accurate method to predict and assess the stability of these soft foundations, especially when dealing with limited field data.
Methods
The core of this research is the application of the Extreme Learning Machine (ELM), a type of single-hidden-layer feedforward neural network. ELM is chosen for its ability to achieve high regression and classification accuracy with a fast learning speed, even with small training datasets. The ELM model is trained using input parameters such as ballast days, monitoring interval time, and daily settlement from previous periods. The study explores three ELM models (M5, M10, M15) with varying numbers of hidden neurons (5, 10, and 15) to determine the optimal configuration. The sigmoid function is used as the activation function in the hidden layer. The models are trained on data from the first 150 days of monitoring and validated using data up to 165 days. Further predictions are made for up to 800 days to assess long-term stability.
Datasets
The study utilizes a dataset collected from a real-world in situ monitoring case of fishbone-shaped dikes at the Manyusha beach protection project in the Yangtze River estuary. Settlement data was collected at two monitoring points (P1 and P2), with the paper focusing on P1. Settlement sensors were placed at various depths (13.89m, 18.9m, 13.35m, and 25.38m), and 165 days of settlement data were recorded. The training data comprises the first 150 days of monitoring results, while data from 158 and 165 days are used for validation.
Key Insights and Ideas
- ELM's Efficacy on Limited Data: The study highlights ELM's strong performance in predicting soft foundation settlement even with a relatively small training sample size, a common challenge in geotechnical engineering.
- Hidden Neuron Count Matters: The number of hidden neurons in the ELM model significantly impacts prediction accuracy, with the M15 model (15 hidden neurons) outperforming M5 and M10.
- Long-Term Stability Prediction: The ELM model can effectively predict future settlement trends, allowing for an assessment of when the soft foundation is likely to stabilize.
- Comparison with Traditional Methods: The ELM model's predictions for final settlement are compared with traditional methods (three-point, Asaoka's, hyperbolic), showing a reasonable agreement and offering a more dynamic prediction capability.
Key Results
- The M15 ELM model achieved the highest prediction accuracy, with an average error of 0.02% and a maximum error of 0.1% for the first 150 days of prediction.
- The M10 model showed an average error of 0.22%, and the M5 model had the highest average error of 0.48%.
- Performance metrics like Correlation Coefficient (CC) and Index of Agreement (Ia) were very high for the M15 model (CC=0.9999, Ia=0.9999), indicating excellent agreement between predicted and measured values.
- The ELM model predicted that the soft foundation would reach a daily settlement below 0.01 mm/d within approximately 335 days after construction, indicating stabilization.
- The final predicted settlement by the M15 model was 17.08 mm, falling within the range of traditional methods but offering a more robust prediction.
Limitations and Future Work
- Monitoring Interval Influence: While the study found that the monitoring interval time (T△) had little influence on the prediction value between the M15-D5 and M15-D10 models, further investigation into its impact across a wider range of intervals could be beneficial.
- Generalizability: The study focuses on a specific site. Future work could involve applying and validating the ELM approach on other fishbone-shaped dike projects with varying geological conditions and environmental factors.
- Incorporating More Factors: While the current model uses key parameters, future research could explore incorporating additional influential factors like wave loading, current velocity, and material properties to further enhance prediction accuracy.
- Secondary Consolidation: The paper notes that traditional methods may underestimate settlement by not considering secondary consolidation. The ELM's ability to predict time-varying settlement could implicitly account for this, but explicit investigation into this aspect would be valuable.
Summary updated