Data Imputation of Soil Pressure on Shield Tunnel Lining Based on Random Forest Model

Min Wang, Xiao-Wei Ye, Xin-Hong Ying, Jin-Dian Jia, Yang Ding, Di Zhang, Feng SunSensors 2024 · DOI

TL;DR

This paper tackles the critical issue of missing data in shield tunnel monitoring, proposing a Random Forest (RF) model for accurate soil pressure imputation. By demonstrating RF's superior performance over traditional methods like mean and median imputation, the study offers a robust solution for maintaining the integrity of tunnel safety assessments.

What Problem Does This Paper Address?

Shield tunneling projects generate vast amounts of data from sensors monitoring forces and displacements. However, this data is prone to missing values due to various factors like sensor damage, connection issues, or power outages. These data gaps can compromise the accuracy of structural safety assessments, potentially leading to engineering hazards. Existing methods for handling missing data, such as simple deletion or basic imputation techniques (mean, median), often fail to account for complex relationships within the data or lead to significant information loss, especially when missingness is high. This paper aims to address this by developing and validating a more sophisticated data imputation technique specifically for shield tunnel soil pressure data.

Methods

The core of this research is the application and optimization of the Random Forest (RF) algorithm for data imputation. RF is an ensemble learning method that builds multiple decision trees and aggregates their outputs.

Here's a breakdown of their approach:

  • Random Forest Principle: RF uses a "bagging" approach, where multiple decision trees are trained on bootstrapped samples of the data. For regression tasks (like predicting soil pressure), the final prediction is typically the average of the predictions from individual trees. A key advantage is its ability to handle high-dimensional data and its inherent resistance to overfitting.
  • Parameter Optimization: The authors systematically tuned RF's hyperparameters to find the optimal configuration for their specific task. This involved:
  • Number of Decision Trees (ntree): Tested values of 50, 100, and 200.
  • Maximum Depth of Trees: Explored depths of 2 and 4.
  • Number of Features per Split (max features): Evaluated using 2, 4, and 8 features.
  • Cross-Validation: 5-fold cross-validation was employed to select the best parameter combinations by minimizing the Mean Squared Error (MSE).
  • Imputation Strategy: The RF model treats the feature with missing values as the "label" and the other features as predictors. It iteratively imputes missing values, starting with features that have the fewest missing values. During imputation, other missing features are temporarily set to 0.
  • Comparison Methods: The performance of the proposed RF imputation method was compared against two baseline methods:
  • Mean Imputation: Replacing missing values with the mean of the available data for that feature.
  • Median Imputation: Replacing missing values with the median of the available data for that feature.
  • Evaluation Metric: Mean Squared Error (MSE) was used to quantify the accuracy of the imputation. A lower MSE indicates a more accurate imputation.

Datasets

The study utilized a dataset of soil pressure time series data collected from fiber-optic grating pressure sensors installed on eight sections of a shield tunnel lining. The data represents 1 hour of monitoring with a 1-second sampling interval, resulting in approximately 3600 data points per monitoring point.

To evaluate the imputation methods, the authors artificially created incomplete datasets by introducing missing data at three different missing ratios: 20%, 40%, and 60%. These synthetic datasets were generated from a complete dataset, ensuring a controlled environment for comparison.

Key Insights and Ideas

  • RF's Robustness to Missingness: The study highlights that the Random Forest model's imputation error remains relatively stable (around 0.000250) even as the missing data ratio increases, demonstrating its robustness.
  • Beyond Simple Averages: The stark performance difference between RF and mean/median imputation underscores the importance of considering inter-feature relationships. Simple averaging methods fail to capture the complex dynamics of soil pressure, whereas RF leverages these relationships for more accurate predictions.
  • Parameter Sensitivity: While RF generally performs well, the study shows that parameter tuning (especially max depth) is crucial for optimizing performance and minimizing MSE. The number of features per split had a less significant impact compared to tree depth.
  • Practical Implications for Tunnel Safety: By providing a reliable method for filling data gaps, this research directly contributes to more accurate and continuous monitoring of tunnel structures, thereby enhancing safety assessments and potentially preventing failures.

Key Results

  • Superior Imputation Accuracy: The Random Forest model consistently achieved the lowest Mean Squared Error (MSE) across all tested missing ratios (20%, 40%, and 60%) compared to mean and median imputation.
  • Significant Error Reduction: At a 60% missing rate, the RF method achieved an MSE of approximately 0.000249, which was about 70% lower than the errors of the mean and median imputation methods.
  • Optimal Parameters: The study identified an optimal parameter combination for the RF model: 200 decision trees, a maximum depth of 4, and considering a maximum of 4 features at each node split.
  • Visual Agreement: Plots comparing the imputed values with original data showed a high degree of agreement for the RF method, visually confirming its effectiveness in filling missing data points.

Limitations and Future Work

  • Artificial Missing Data: The study used artificially generated missing data. While this allows for controlled evaluation, real-world missing data can exhibit more complex patterns (e.g., MNAR - Missing Not At Random), which might affect the model's performance.
  • Specific Data Type: The research focused specifically on soil pressure data from shield tunnels. Further validation on other types of sensor data or in different tunneling environments might be beneficial.
  • Computational Cost: While RF is generally efficient, the parameter optimization process can be computationally intensive. Future work could explore more efficient hyperparameter tuning strategies.
  • Real-World Data Application: The paper demonstrates the method's effectiveness on artificially created missing data. Applying and validating this method on actual datasets with naturally occurring missing data would be a valuable next step.

Summary updated

BibTeX
@article{wang2024data,
  author = {Min Wang and Xiao-Wei Ye and Xin-Hong Ying and Jin-Dian Jia and Yang Ding and Di Zhang and Feng Sun},
  title = {{Data Imputation of Soil Pressure on Shield Tunnel Lining Based on Random Forest Model}},
  journal = {Sensors},
  year = {2024},
  doi = {10.3390/s24051560},
}
Figure