430 Figure 2 – Hierarchical multilayer control system. In a closed grinding circuit, Hodouin (2011) classifies process variables into four distinct categories: manipulated variables (u), such as ore feed rate and pump speed; controlled variables (y), including P80 and sump level; disturbances (d), such as variations in ore grindability; and other output/internal state variables (x), like mill power or circulating load. Artificial Intelligence attempts to perform tasks that traditionally require human intelligence and has evolved from systems based on predefined rules, known as knowledge bases, toward systems that acquire their own knowledge by identifying patterns from data, as is the case with Machine Learning (Goodfellow et al., 2016). Within this approach lies Deep Learning (DL), which incorporates Artificial Neural Networks (ANNs) as a technique inspired by the functioning of the human brain. A specific class of these techniques is Recurrent Neural Networks (RNNs), which are dynamic systems designed specifically to handle sequential data (Geron, 2019). Classical RNNs suffered from the problem of gradients either vanishing or exploding exponentially over time; consequently, when processing long sequences, information was lost at each time step. Long Short-Term Memory (LSTM) networks utilize short- and long-term memory cells and were proposed by Hochreiter and Schmidhuber (1997). Reinforcement Learning (RL) is defined as a computational approach to learning from interaction, focusing on goal-directed learning from trial and error (Sutton & Barto, 2018). In this framework, an agent faces a sequential decision-making problem where the goal is to find a policy, a function mapping states (St) to actions (at), to maximize the numerical reward (rt) over time. This objective is achieved through policy evaluation and optimization methods, such as value-based or policy-based learning, which rely on the agent discovering which actions yield the most reward by trying them Machine learning methods designed for time-series analysis, particularly Long Short-Term Memory (LSTM) neural networks, offer the capability to model nonlinear relationships and temporal dependencies that are difficult to represent using first-principles approaches alone. This work proposes a data-driven framework that shifts circuit management from reactive to predictive control by deploying Long Short-Term Memory (LSTM) networks as short-horizon, multi-step P80 forecasting models trained on historical multivariate sensor data. The LSTM captures the dynamic, nonlinear behavior of the grinding process and serves as a surrogate environment for offline reinforcement learning (RL), enabling an agent to identify optimal adjustments of controllable process variables without interrupting plant operations. Together, the LSTM–RL framework provides a foundation for adaptive, resilient grinding circuit control. 2. STATE OF THE ART Grinding is widely recognized as a core operation in mineral processing plants, and the P80 has long served as the critical product size indicator directly linking circuit performance to downstream metallurgical outcomes. Hodouin (2011) established the foundational taxonomy of process variables (manipulated, disturbance, internal state, and controlled) that underpins the variable selection and control formulation adopted in this work.
RkJQdWJsaXNoZXIy MTM0Mzk2