Date of Award

2024

Document Type

Campus Access Master's Thesis

Degree Name

Master of Science in Geological Engineering (MS)

Administrative Home Department

Department of Geological and Mining Engineering and Sciences

Advisor 1

Snehamoy Chatterjee

Committee Member 1

Xiang Li

Committee Member 2

Ashish Kumar

Abstract

The mining industry has traditionally relied on geostatistical models for estimating and managing mineral resources. In modern mining operations, real-time updates to these models are critical, as sensor data plays a vital role in characterizing deposit properties. However, noise often affects sensor measurements, compromising the accuracy and reliability of resource models. This study presents a novel reinforcement learning (RL) framework designed to dynamically denoise sensor readings and improve resource model updates.

The methodology begins by reviewing current approaches to integrating sensor data into geostatistical models, highlighting their vulnerability to noise. The proposed framework leverages a Deep Deterministic Policy Gradient (DDPG) algorithm with a recurrent neural network (RNN) agent. This self-learning system sequentially processes noisy sensor data, using state representations that incorporate noise characteristics, spatial locations of sensor data, and uncertainty metrics derived from Monte Carlo (MC) Dropout. The RNN-based actor network generates actions to iteratively denoise the sensor value, while the critic network evaluates the denoising performance to optimize the process.

The reward function is designed to prioritize the agent's proximity to the clean sensor value, minimize uncertainty, and promote exploration in high-variance scenarios. Training involves experience replay and interaction across multiple episodes, enabling the agent to learn robust denoising strategies. Experimental validation using a public dataset demonstrates improvements in signal quality and model reliability.

By integrating this RL-based denoising algorithm into geostatistical resource model updating frameworks enables more robust tracking of mineral quality variations and enhances decision-making during mining operations. This approach offers a scalable solution for real-time data processing, with potential applications in other domains requiring dynamic sensor data refinement.

Available for download on Wednesday, December 03, 2025

Share

COinS