Truck Productivity Analytics Toolkit 

Interested in using this toolkit? Please contact the PI for further discussion and collaboration. 

The Truck Productivity Analytics Toolkit is a web-based research and engineering utility developed by Dr. Chengkai Fan based on his doctoral research completed at the University of Alberta. The toolkit was created to translate machine learning models for truck productivity prediction into an accessible and interactive platform. It is intended to support academic research, teaching, student training, model demonstration, and exploratory engineering analysis.

The toolkit integrates operational, equipment, and weather-related variables to predict truck productivity across hourly, daily, weekly, and monthly temporal resolutions. All models and examples are presented without disclosing company names, mine locations, equipment identifiers, raw operational records, or other confidential information.

Main Features

• Multi-Temporal Productivity Prediction
  Users can select hourly, daily, weekly, or monthly analysis. Each temporal resolution is connected to a separately trained and validated prediction model. This structure recognizes that truck productivity patterns and influential variables may change when operational data are aggregated over different time periods.

• Manual Input of Operating Conditions
  Users can enter operational and environmental variables directly through the webpage. The current inputs include haul distance, number of trucks, number of shovels, empty truck speed, ambient temperature, humidity, precipitation, and wind speed.

• Model-Specific Prediction
  The toolkit automatically loads the trained model and preprocessing procedure associated with the selected temporal resolution. The input variables are processed in the same order and format used during model development before the productivity prediction is generated.

• SHAP-Based Model Interpretation
  The toolkit uses SHAP analysis to explain how each input variable influences a prediction. The results identify variables that increase or decrease the predicted productivity and quantify the relative contribution of each factor.

• Local SHAP Analysis
  Users can generate an explanation for an individual operating scenario. This analysis helps answer why a specific productivity value was predicted and which operational or weather variables had the strongest influence on that result.

• Global SHAP Analysis
  When a representative dataset is provided, the toolkit can evaluate overall model behavior across multiple observations. Available outputs may include mean absolute SHAP importance, feature rankings, summary plots, beeswarm plots, and feature dependence analysis.

• Temporal Resolution Comparison
  The same input scenario can be evaluated using the hourly, daily, weekly, and monthly models. This comparison helps users examine how the selected temporal resolution affects the predicted productivity and the interpretation of influential variables.

• Input Range Alerts
  The toolkit compares user-entered values with the ranges represented in the original model development data. A warning is displayed when an input falls outside the corresponding range, indicating that the prediction may involve extrapolation and should be interpreted with additional caution.

• Interactive Web Interface
  The toolkit is implemented using Streamlit and can be accessed through a standard web browser. Users can enter data, run predictions, generate SHAP explanations, compare models, and review results without writing Python code.

• Downloadable Results
  Prediction outputs and SHAP results can be exported for further analysis, reporting, classroom exercises, or research documentation.

Model Development and Performance

The models integrated into the toolkit were developed through a structured research process that included data cleaning, temporal aggregation, feature analysis, preprocessing, algorithm comparison, hyperparameter selection, validation, and performance evaluation. Different algorithms were selected for different temporal resolutions because one modeling method did not provide the best performance across all datasets. The toolkit therefore preserves the best-performing model identified for each temporal scale rather than applying a single algorithm to all analyses. Model performance was assessed using established regression indicators, including the coefficient of determination, root mean squared error, mean absolute error, predicted-versus-observed comparisons, and error distributions. The doctoral research showed that temporal resolution, data preprocessing, and model selection can substantially affect truck productivity prediction.

Key Characteristics

The toolkit combines several capabilities within a single platform:

• Truck productivity prediction at four temporal resolutions

• Integration of operational and weather-related variables

• Model-specific preprocessing

• Local and global SHAP interpretation

• Cross-resolution model comparison

• Input applicability and extrapolation warnings

• Web-based interaction

• Privacy-conscious presentation of research outcomes

Unlike a basic prediction calculator, the toolkit connects prediction with interpretation. It not only estimates truck productivity but also helps users understand how and why the models produce their results.

Research and Educational Impact

The toolkit converts doctoral research outputs into a practical and reusable engineering application. It demonstrates how trained machine learning models can be moved beyond research scripts, journal figures, and performance tables into an interactive decision-support environment.

For research and teaching, the toolkit can help users explore:

• The relationship between operating conditions and truck productivity

• The effects of temporal data aggregation

• Differences among machine learning and statistical models

• The interpretation of model predictions using SHAP

• The risks associated with prediction outside the model development range

• The translation of data-driven research into engineering tools

The platform can support courses and student projects in mining engineering, mine transportation, mine operations, machine learning, data analytics, explainable artificial intelligence, and engineering decision support.

Potential Engineering Impact

The toolkit provides a framework for evaluating truck productivity scenarios and identifying important productivity drivers. It may support exploratory analysis of fleet configuration, haul distance, truck speed, loading resources, and weather conditions.

The current platform also provides a technical foundation for future extensions involving:

• Productivity optimization

• Fuel consumption estimation

• Truck-shovel allocation

• Weather-aware operational planning

• Prediction uncertainty

• Autonomous haulage analysis

• Digital decision-support systems

Privacy and Confidentiality

The public-facing toolkit does not identify any company, mine, site, employee, vehicle, or equipment unit. It does not provide access to confidential dispatch records, production schedules, geographic information, operating procedures, or raw industrial datasets.

Important Notice

This toolkit is intended for academic research, education, student training, model demonstration, and exploratory engineering analysis. It should not be treated as a universal productivity model or used as the sole basis for operational, financial, or safety-critical decisions.

The models were developed using data from a specific operational context. Their accuracy may decrease when they are applied to equipment systems, operating environments, weather conditions, or production practices that differ substantially from those represented during model development.

SHAP values explain the behavior of the trained models. They do not independently establish physical causality. Predictions outside the original model development range should also be interpreted with caution.

Interested in using this toolkit? Please contact the PI for further discussion and collaboration.