AI Model Wiki

    Skip to main content
    15 min
    Lesson 6 of 6

    Common Pitfalls & How to Avoid Them

    From the course AI for Mining & Resources

    Learn from failed mining AI projects to prevent the most costly mistakes.

    Pitfall 1: The Data Silo Problem

    What happens: Your maintenance data lives in SAP, dispatch in Wenco, geology in Micromine, and process in OSIsoft PI. Nobody has ever joined these datasets. An AI vendor promises amazing results — then spends 80% of the engagement extracting and cleaning data.

    Prevention:

    • Before signing any AI vendor contract, run your own data audit (Week 1 of the implementation plan)
    • Budget 40% of the AI project for data engineering — not the AI itself
    • Establish a data lake (AWS S3, Azure Data Lake) as a prerequisite
    • Assign a dedicated data engineer to the project, not a "part-time" one

    Pitfall 2: Model Drift in Harsh Environments

    What happens: Your predictive maintenance model is trained in the dry season. Six months later, wet season conditions change vibration signatures, and the model starts missing failures it previously caught.

    Prevention:

    class DriftMonitor:
        def check_feature_drift(self, current_window, training_baseline):
            for feature in self.monitored_features:
                ks_stat, p_value = ks_test(
                    training_baseline[feature],
                    current_window[feature]
                )
    
                if p_value < 0.05:  # Distribution has shifted
                    self.alert_mlops_team(
                        feature=feature,
                        drift_magnitude=ks_stat,
                        recommendation='retrain_scheduled'
                    )
    

    Schedule model retraining quarterly at minimum. Track prediction confidence scores daily — a drop in average confidence is the earliest drift signal.

    Pitfall 3: Operator Resistance (The "Job Threat" Perception)

    What happens: Truck dispatchers have 15 years of experience. They see the AI dispatch system as a threat to their jobs and subtly undermine it by overriding recommendations, claiming the system "doesn't understand the mine."

    Prevention:

    • Involve operators in the design process — their tribal knowledge is your most valuable training signal
    • Frame AI as the operator's assistant, not their replacement: "The system flags it; you decide"
    • Make the AI's reasoning visible: show why it made a recommendation, not just what
    • Celebrate operator-AI collaboration wins publicly — the dispatcher who used AI insights to prevent a crusher blockage is a hero

    Pitfall 4: Single-Point-of-Failure Connectivity

    What happens: Your AI dispatch system requires a live connection to the cloud. A fiber cut or network outage at a remote mine site means no dispatch system at all — reverting to radio and paper.

    Prevention:

    • Always design for graceful degradation: AI recommends when connected, reverts to rule-based fallback when disconnected
    • Deploy edge computing on-site (Nvidia Jetson, AWS Snowball Edge) for core models
    • Maintain a cached "last known good" dispatch plan that operators can execute manually
    • Test disconnected mode quarterly in drills — like a fire drill for your tech stack

    Pitfall 5: Ignoring the Regulatory Landscape

    What happens: Your autonomous haulage pilot succeeds, then legal discovers that operating autonomous vehicles in your jurisdiction requires regulatory approval you haven't sought. 18-month delay while legal catches up.

    Prevention:

    • Engage your national mining regulatory body (MSHA in US, DMIRS in WA, MoM in South Africa) in week 1
    • Document AI decision-making for liability purposes — if a tailings dam AI failed to raise an alert, who is responsible?
    • Implement audit logs for every AI recommendation and human override
    • Join industry bodies (International Council on Mining & Metals, MICROMINE user groups) for regulatory intelligence

    Key Takeaways

    • •Budget 40% of AI project cost for data engineering — not the model itself
    • •Monitor feature drift seasonally; harsh environments shift data distributions more than most industries
    • •Operator buy-in requires involving them in design and framing AI as their tool, not their replacement
    • •Always design for disconnected mode — remote mine sites cannot depend on continuous cloud connectivity
    0 of 6 lessons marked complete

    Progress is kept in this browser only.