OORJA
Grinding-circuit optimization combining graph-based reinforcement learning, constrained control and a virtual ball-charge estimator.
2025
SIH 2025 · Optimized Ore Reduction using Joint Agents.
Grinding as a coupled control problem
OORJA links ore characteristics, mill operating state and product requirements in a single grinding-circuit architecture. The central problem is specific energy per tonne: lowering instantaneous power is useful only when throughput, particle size and grinding-media consumption remain within their operating targets. Harder feed, changing ball charge and downstream slurry conditions make isolated setpoint rules insufficient.
The design combines a Digital NeuroGraph, advance ore-hardness forecasting and a virtual ball-charge estimator. A strategic graph reinforcement learning layer selects a target operating state. Model predictive control translates that target into bounded recommendations for the operator, with observed outcomes feeding the next estimation cycle.
Technology and role
Scroll horizontally to compare.
| Component | Responsibility |
|---|---|
| Kafka · Flink | Ingest SCADA, acoustic, vibration and mine-plan events; normalize the streaming inputs. |
| InfluxDB · Neo4j | Retain high-frequency telemetry separately from the graph of operating states and transitions. |
| LSTM · physics/ML estimator | Forecast ore hardness and estimate ball charge from wear dynamics, power/load relationships and learned residuals. |
| Graph RL · MPC | Choose an operating strategy, then enforce load, power, temperature and rate-of-change constraints. |
| FastAPI · React · TypeScript · Three.js | Serve model outputs and live operator views with scenario comparisons and setpoint explanations. |
| Hyperledger Fabric · Docker · Kubernetes · NGINX | Record recommendation/override history and define the service deployment boundary. |
Architecture walkthrough
- Acquisition: SCADA mill power, feed rate, speed, water and sump density join acoustic/vibration signals and mine-plan ore context. Stream processing places heterogeneous measurements into a shared operational state.
- State representation: graph nodes describe ore hardness, load, energy and throughput conditions. Edges express operational transitions and their energy costs, allowing the strategy layer to reason about coupled effects.
- Estimation and forecasting: an LSTM supplies incoming-feed context. Wear kinetics, Morrell power-load correlation and a neural residual correction supply complementary evidence for the otherwise difficult-to-observe ball-charge state.
- Recommendation: graph RL proposes a target; MPC checks whether a feasible path exists under the circuit constraints. The operator view compares alternatives, explains the proposed changes and allows overrides.
- Feedback: predicted versus observed behavior updates the state and model parameters. Ball-addition logs and physical charge observations provide calibration anchors for the estimator.
Core mechanisms
- Multi-objective optimization balances specific energy, peak-demand penalties and grinding-media loss while retaining throughput and product-size/P80 requirements.
- Ore-hardness forecasting moves the decision point ahead of the resulting power spike, giving control recommendations feed context.
- The wear model tracks ball-diameter change against ore abrasivity and operating time. Power/load correlation supplies an independent estimate; learned residuals account for departures from the physical model.
- MPC constrains load fraction, power ceilings, temperature and vibration bounds, plus the rate at which setpoints can change.
- An infeasible control path retains the current operating state and requests a revised strategic target.
- Outcome-based parameter correction addresses changing ore and equipment behavior without treating an old calibration as permanent.
Engineering decisions
Scroll horizontally to compare.
| Decision | Reason |
|---|---|
| Graph plus time-series storage | Coupled state relationships and dense chronological telemetry have different query needs. |
| Physics plus learned residuals | Wear dynamics provide structure when physical charge measurements are sparse; residual correction captures plant-specific behavior. |
| Separate strategy from constrained control | An attractive target still needs a feasible transition within operating limits. |
| Operator review and decision history | What-if comparisons and overrides keep recommendations inspectable and attributable. |
System deliverables
The SIH 2025 system design specifies the ingestion layer, Digital NeuroGraph, ball-charge estimator, forecasting/control split and operator workspace. Its technical contribution is the connection between a latent process-state estimate and a constraint-aware recommendation path, rather than a dashboard that merely displays power readings.