[Paper Review] Automated Estimation of Construction Equipment Emission Using Inertial Sensors and Machine Learning Models

Farid Shahnavaz|arXiv (Cornell University)|Feb 26, 2022

Vehicle emissions and performance81 references24 citations

TL;DR

This study proposes an IoT and machine learning framework that uses inertial sensor data (accelerometer and gyroscope) from construction equipment to predict real-time emissions of CO, NOX, and CO2 without relying on engine or fuel data. The Random Forest model achieved R² values of 0.94 for CO and CO2, and 0.91 for NOX, demonstrating high accuracy in emission estimation using only motion-based inputs.

ABSTRACT

The construction industry is one of the main producers of greenhouse gasses (GHG). With the looming consequences of climate change, sustainability measures including quantifying the amount of air pollution during a construction project have become an important project objective in the construction industry. A major contributor to air pollution during construction projects is the use of heavy equipment. Therefore, efficient operation and management can substantially reduce a project’s carbon footprint and other environmental harms. Using unintrusive and indirect methods to predict on-road vehicle emissions has been a widely researched topic. Nevertheless, the same is not true in the case of construction equipment. This paper describes the development and deployment of a framework that uses machine learning (ML) methods to predict the level of emissions from heavy construction equipment. Data is collected via an Internet of Things (IoT) approach with accelerometer and gyroscope sensors as data collection nodes. The developed framework was validated using an excavator performing real-world construction work. A portable emission measurement system (PEMS) was used along with the inertial sensors to record the amount of CO, NOX, CO2, SO2, and CH4 pollution emitted by the equipment. Different ML algorithms were developed and compared to identify the best model to predict emission levels from inertial sensors data. The results show that Random Forest with the coefficient of determination (R2) of 0.94, 0.91, and 0.94, and normalized root-mean-square error (NRMSE) of 4.25, 6.42, and 5.17 for CO, NOX, and CO2, respectively, was the best algorithm among different models evaluated in this study.

Motivation & Objective

• To develop a non-intrusive, real-time method for estimating construction equipment emissions using only inertial sensor data.
• To eliminate reliance on expensive PEMS or engine parameter monitoring by leveraging equipment motion as a proxy for emission levels.
• To validate the framework using real-world excavator operations with simultaneous PEMS and sensor data collection.
• To compare multiple machine learning models to identify the optimal algorithm for emission prediction from motion data.
• To establish a foundation for scalable, low-cost emission monitoring in construction projects.

Proposed method

• Deployed an IoT system with synchronized accelerometer and gyroscope sensors mounted on an excavator’s cabin and arm to capture dynamic motion data.
• Collected simultaneous ground-truth emission data using a Portable Emission Measurement System (PEMS) measuring CO, NOX, CO2, SO2, and CH4 in ppm.
• Preprocessed sensor data by extracting time-domain features such as mean, standard deviation, kurtosis, and frequency of fluctuations.
• Trained and compared four supervised machine learning models—Random Forest, XGBoost, Gradient Boosting, and Multivariate Linear Regression—on the feature-engineered data.
• Evaluated model performance using the coefficient of determination (R²), with cross-validation to ensure robustness.
• Selected the best-performing model based on R² scores across CO, NOX, and CO2 predictions.

Experimental results

Research questions

• RQ1 Can inertial sensor data alone accurately predict real-time emissions from construction equipment without engine or fuel data?
• RQ2 Which machine learning algorithm performs best in predicting CO, NOX, and CO2 emissions from motion-based features?
• RQ3 How does the predictive performance of ML models trained on inertial data compare to existing emission estimation methods for on-road vehicles?
• RQ4 Is there a strong correlation between equipment movement patterns and emission rates, enabling indirect estimation?
• RQ5 Can a low-cost, non-intrusive sensor system replace traditional PEMS or engine monitoring for emission tracking?

Key findings

• The Random Forest model achieved the highest R² values of 0.94 for CO, 0.91 for NOX, and 0.94 for CO2, outperforming other models.
• The study demonstrated that equipment motion patterns captured by inertial sensors are strongly correlated with actual emission levels.
• The framework enables real-time, non-intrusive emission estimation using only two smartphone-mounted sensors, eliminating the need for PEMS or engine data.
• The results show that machine learning can effectively learn complex emission patterns from kinematic data alone.
• The method is scalable and cost-effective, offering a practical solution for monitoring emissions across diverse construction projects.
• The study establishes a foundation for future work using deep learning and multi-machine generalization.