Pro1plan Consultants

Introduction

In large-scale infrastructure projects, accurate forecasting of diaphragm wall installation duration is critical for sequencing, resource planning, and risk mitigation. Leveraging Azure Machine Learning Designer, I developed a regression model to predict installation time using geotechnical and structural features. This article outlines the step-by-step methodology, from data preparation to deployment.

📊 Step 1: Data Preparation

The dataset comprised 1,095 records, each representing a diaphragm wall panel installation event. Key features included:

• Soil classification (e.g., SM, CL, ML)
• Panel thickness
• Substructure length (SubLength)
• Tunnel section identifiers

🧼 Cleaning & Normalization

• Numerical features such as panel thickness and SubLength were normalized using MinMaxScaler and MaxAbsScaler to ensure consistent model behavior
• Categorical features like soil type and tunnel section were encoded using LabelEncoder and OneHotEncoder
• Panel ID was excluded from encoding, as it served only as a nominal identifier without predictive value

This preprocessing ensured that all features were scaled appropriately and semantically meaningful for model training.

🔀 Step 2: Train-Test Split with Stratification

To ensure fair evaluation:

• Data was split into 90% training and 10% testing
• Stratification was applied based on soil type, preserving class balance across both sets

This approach ensured the model could generalize across varied geotechnical conditions.

🌲 Step 3: Model Selection – Boosted Decision Tree Regression

Given the moderate dataset size, I selected Boosted Decision Tree Regression for its:

• Efficiency on small-to-medium datasets
• Ability to model non-linear relationships
• Robustness to outliers and missing values
• Interpretability via feature importance metrics

This model type is well-suited for structured construction data with mixed feature types.

⚙️ Step 4: Hyperparameter Tuning

Using Azure ML’s Tune Model Hyperparameters component:

• Sweeping mode was set to Entire Grid for exhaustive search
• Performance metric was Accuracy, interpreted as prediction closeness within acceptable tolerances
• Label column was defined as Duration

This ensured the model was optimized across all relevant parameter combinations.

✅ Step 5: Model Evaluation

The trained model was validated using the 10% holdout set. Key metrics:

While the R² was modest, the MAE of 1.83 days is practical for early-stage planning and sequencing.

🧠 Step 6: Stacking Ensemble (Optional Extension)

To improve performance, I am exploring stacking:

• Trained multiple base models (Boosted Tree, Linear Regression, Decision Forest)
• Extracted predictions using Convert to CSV
• Combined outputs and trained a meta-model (e.g., LightGBM) on the stacked predictions

This ensemble approach improved generalization and reduced overfitting, especially across soil types.

Training Pipeline for predicting the duration of Dwall construction

🚀 Step 7: Deployment & Validation

Using Designer’s Real-Time Inference Pipeline:

• Added Web Service Input, Enter Data Manually (replacing the Dwall dataset) components

Inference Pipeline to enable the deployment  

• Deployed to an Azure Container Instance (Some errors in getting the endpoints as shown in the status of endpoints below)

• Validated predictions using manual input and REST API calls.

This enabled real-time duration forecasting for new panel configurations, supporting proactive planning and risk control.

📌 Conclusion

Azure ML Designer provided a robust, visual framework for building, tuning, and deploying a predictive model tailored to construction workflows. The methodology balances technical rigor with operational practicality—making it ideal for engineering teams seeking reproducible, scalable ML solutions.