All five projects share the same skeleton, so nothing is improvised on day 1:
- Mining layer (what for): stage of the value chain, user, KPI and constraints (safety, continuity, unions, shifts).
- Engineering layer (how it is solved): measurable problem, clean data, simple baseline, model or method, validation and quantified impact.
- Software layer (how it is built): ingestion and cleaning (pandas), logic or model (scikit-learn, rules, optimisation), consumable output (Streamlit or Power BI) and an explanation of how it integrates. In 3 days you build a credible MVP and an implementation plan.
Where each part comes from. The general topics come from public sources about Hackamine and BHP. The numbers in the exercises (40 trucks, 48 h, 3 points, 60 activities, 5 days, the shared crane) are made up for practice: practice parameter. Industry literature confirms that each problem exists.
A. Predictive maintenance for haul trucks
Mining layer
Critical components: engine, transmission, wheel motors and tyres. User: maintenance planner. KPIs: physical availability, MTBF and spare parts in the warehouse. It ties in with the 2021 challenge and with the autonomous haulage fleet.
Engineering layer
- Pareto of failures by hours and by cost.
- Define “major failure” from the work orders and label the previous 48 h (practice parameter).
- Features: alarm frequency, temperature and pressure trends, hours since the last maintenance.
- Rule-based baseline, then gradient boosting.
- Time-based validation (train on the past, test on the future).
- Threshold chosen by weighing the cost of a false negative against that of a false positive.
Software layer
Join alarms and work orders by equipment and time, store the dataset as parquet, persist the model with joblib and show a dashboard with the risk ranking per truck and the three features that explain each alert.
Evidence
- According to Guía Minera’s coverage of the third Hackamine (2021), one of the challenges was predicting component stock-outs from alerts.
- A C3 AI case study describes a model that predicted engine and cooling system failures in a fleet of 79 haul trucks up to 14 days in advance, whereas earlier systems gave little or no time to react to alerts.
- A Queen’s University thesis, “Predicting Turbocharger Failures in Mining Haul Trucks”, proposes predicting turbocharger failures with machine learning.
B. Flotation recovery
Mining layer
Variables: feed grade, P80, pH, reagents, cell level and air, plus mineralogy (chalcopyrite, clays). The classic trade-off is recovery versus concentrate grade. User: metallurgist and control room operator. It is Spence’s current pain point.
Engineering layer
- Align variables with the lag imposed by residence time.
- Segment by regime or ore type with clustering.
- Explainable model (trees plus SHAP).
- “Operating windows” built from the best shifts.
- Impact quantified with the fine copper formula.
Software layer
Lagged features, a “soft sensor” for expected recovery and a panel that suggests ranges to the operator. Always a recommendation, never automatic control.
Evidence
- In its Q3 FY26 results (MCh), BHP attributed the drop in recoveries to the complexity and variability of the ore.
- A study on a copper porphyry in northern Chile assesses whether the textural characteristics of the ore (extracted from hyperspectral images of the same drill holes used in the flotation tests) improve the prediction of copper recovery.
C. Major SAG mill maintenance
Mining layer
The SAG mill is the bottleneck: every hour of downtime is lost tonnage. The maintenance includes the liner change, energy lockout and isolation, permits, contractors and a shared crane (practice parameter).
Engineering layer
- WBS (work breakdown structure).
- Durations with three estimates (optimistic, most likely, pessimistic).
- Dependencies and critical path.
- Resource levelling and S-curve.
- Extra: Monte Carlo to estimate the probability of finishing in 5 days (practice parameter).
Software layer
Activities in a CSV, critical path with networkx, Gantt in Plotly and a dashboard of actual versus planned progress. A well-built MS Project or Excel file also works.
Evidence
- The Hackamine 2026 technical rules ask for a Gantt chart.
- A CIM paper (“Toward the 24 hour large SAG Mill Reline”, John Russell, 2008) argues that the liner change dictates the critical path of every maintenance shutdown and that, on an annual basis, it is the single largest influence on plant availability. It mentions the liner change systems of the SAG mills at Escondida Laguna Seca.
- At Vale’s Clarabelle mill, the SAG historically needed about 4 planned liner changes a year, roughly 4% of its annual availability. Later work by Vale describes a change of strategy to reduce that number.
D. Continuity: dispatch and LTE network outage
Mining layer
With an autonomous haulage fleet, if the network goes down the trucks stop: it is a safety and a production issue at the same time. A degraded operation protocol is needed.
Engineering layer
- FMEA and root cause analysis (5 whys, Ishikawa).
- Single points of failure: links, power, LTE core and servers.
- Redundancy, RTO and RPO.
- IT/OT separation following the Purdue model and cybersecurity with IEC 62443.
Software layer
Health monitor per antenna and per service (latency, packet loss, heartbeats), anomaly detection, alerts and a topology map. It can be demonstrated with synthetic logs and a runbook.
Evidence
- According to Guía Minera’s coverage of the third Hackamine (2021), in 2021 there were challenges on the fixed and LTE networks and on simultaneous system failure.
- At Thiess, if communication drops, the autonomous equipment stops.
- Komatsu describes the wireless network as what controls the operation of its autonomous trucks and also what stops it.
- At Fortescue’s Christmas Creek mine (February 2019), an autonomous truck collided at low speed with a parked one after a loss of Wi-Fi coverage. No one was injured.
E. Fatigue on the night shift
Mining layer
7x7 or 14x14 shifts, high potential incidents and critical controls. It is a sensitive topic: it is approached from safety and work design, never from monitoring the worker.
Engineering layer
- Incident rate normalised by man-hours, by clock hour and by night of the roster cycle.
- Simple statistical test (Poisson or chi-square).
- Controls following the hierarchy of controls: shift redesign and breaks before PPE.
- Measurable pilot.
Software layer
Risk heatmap of hour by day of the cycle, a break planning tool and privacy by design: aggregated data, no individual tracking.
Evidence
- In the Escondida mediation (Guía Minera), the union warned that the 14x14 roster and multi-skilling could increase fatigue.
- A study published by NIOSH identifies fatigue as a main factor in haul truck collisions, especially in the early morning hours.
- Important correction: an ARRB study of haul truck operators on 12-hour shifts found that, within a single night, time on task does not seem to affect fatigue (performance was even worse in the first eight hours). What matters is the circadian rhythm, the successive nights of the cycle and the variability between individuals. That is why case E is framed as “incidents by clock hour and night of the cycle”, not as “hours worked”.
Source quality. The C3 AI, Unikie or Oxmaint cases are vendor material: they help show that the problem exists, but their savings figures are marketing. For the pitch, prefer BHP reports, papers (AusIMM, CIM, NIOSH) and what the mentors say.