Get in Touch

Tag: Machine Learning

  • Predictive AI Model For Crop Nutrient Needs

    Predictive AI Model For Crop Nutrient Needs

    Predictive AI Model For Crop Nutrient Needs

    Predictive AI Model For Crop Nutrient Needs

    Overview

    We worked with Wesfarmers Chemicals, Energy & Fertilisers to predict crop nutrient uptake needs so fertilisation can be scheduled promptly. Our predictive regression models achieved over 85 percent accuracy, meeting the 80 percent deployment threshold and aligning with lab result accuracy ranges of 80 to 90 percent.

    Problem

    Lab analyses of plant and soil samples did not always arrive in time to act optimally, given changing weather and varying crop needs.

    Solution

    • Predictive models: Developed regression models that estimate nutrient uptake needs with more than 85 percent accuracy.
    • Deployment-ready: Designed to meet a minimum 80 percent accuracy requirement, comparable to lab result accuracy of 80 to 90 percent, so estimates are available as soon as samples are taken.

    Result

    • Model accuracy: Over 85 percent for nutrient uptake estimates.
    • Benchmark: Within the 80 to 90 percent lab result accuracy range and above the 80 percent deployment threshold.
    • Operational impact: Earlier estimates enable timely fertilisation aligned to conditions, reducing risk of over or under fertilisation.

    < back to Case Studies

    Book Discovery Call
  • Proactive Centrifuge Maintenance

    Proactive Centrifuge Maintenance

    Proactive Centrifuge Maintenance

    Proactive Centrifuge Maintenance

    Overview

    We worked with Wesfarmers Chemicals, Energy & Fertilisers to predict upcoming centrifuge vibration levels using a machine learning model trained on historical sensor data, enabling timely washes and reducing wear.

    Problem

    Material build-up on centrifuge walls increases vibration and accelerates equipment wear. Scheduled periodic washes could not always occur in time to prevent prolonged high vibration, and were wasteful when no spike was imminent.

    Solution

    • Predictive model: Trained on historical data from several centrifuges to forecast near-term vibration levels.
    • Actionable timing: Designed to trigger targeted washes and maintenance when risk is rising, and avoid unnecessary cleans when it is not.

    Result

    • Early warning: Predicted when centrifuges would consistently exceed safe vibration levels up to 1 hour in advance, with over 90% accuracy.
    • Operational gains: Significant reduction in production downtime and water usage from unnecessary washes.

    < back to Case Studies

    Book Discovery Call
  • Automating UNSPSC Expense Classification

    Automating UNSPSC Expense Classification

    Automating UNSPSC Expense Classification

    Automating UNSPSC Expense Classification

    Overview

    We worked with the Department of Finance to automate mapping expenses to United Nations Standard Products and Services Code (UNSPSC) with a bespoke machine learning model, delivering high-accuracy classifications and freeing teams from manual coding.

    Problem

    Analysts spent significant time reading expense descriptions and assigning precise UNSPSC codes. Existing manual processes slowed throughput and introduced inconsistency.

    Solution

    • Custom ML classifier: Automated categorisation of financial transactions to UNSPSC codes.
    • Technique blend: Combined text processing, a custom learning mechanism, and semantic matching to improve coverage for future cases.
    • Accuracy focus: Tailored to the department’s data to maximise precision.

    Result

    • Over 99% classification accuracy, a 15% uplift versus out-of-the-box machine learning techniques.
    • Productivity and cost savings from automating a previously manual task.

    < back to Case Studies

    Book Discovery Call
  • Credit Card Fraud Detection

    Credit Card Fraud Detection

    Credit Card Fraud Detection

    Credit Card Fraud Detection

    Overview

    We worked with a university finance department to identify fraudulent corporate credit card transactions by building a bespoke, unsupervised pattern-discovery system that learns normal behaviour and flags meaningful deviations.

    Problem

    Periodic consultant reports and rule-based alerts were in place, yet a fraudster managed to bypass these controls for an extended period of time. The team needed a more accurate and timely way to detect suspicious activity.

    Solution

    • Unsupervised pattern discovery: Instead of pre-coded rules, the system learns the norm from cardholder data and surfaces deviations.
    • Risk-based ranking: Cardholders are ranked according to the number and utility of deviating patterns identified in their transactions, helping analysts focus on the highest-risk profiles first.

    Result

    • The prosecuted fraudster ranked second, with all suspicious patterns correctly identified, which made the behaviour easy to spot.
    • When run periodically, the tool provides peace of mind and early alerts to deviating patterns, enabling detection before significant losses occur.

    < back to Case Studies

    Book Discovery Call
  • SAG Mill Throughput Optimisation

    SAG Mill Throughput Optimisation

    SAG Mill Throughput Optimisation

    SAG Mill Throughput Optimisation

    Overview

    We helped BHP to improve Semi-autogenous grinding (SAG) mill throughput by deploying a machine learning model that ingests real-time, site-specific data across drilling, blasting, crushing, and SAG operations. Its explainable machine learning outputs show which factors drive each prediction, so operators can trust and act on recommended set-points.

    Problem

    Physics-based and commercial simulation tools were not accurate enough for real-time control. Just as importantly, they lacked explainability — operators couldn’t see which inputs influenced predictions, making it hard to adjust set-points with confidence.

    Solution

    • Integrated ML model: Uses live plant data from drilling, blasting, crushing, and SAG mill operations to estimate throughput and recommend set-points (e.g., crusher gap settings, SAG mill speed, percentage of solids) for the current operating context and constraints.
    • Explainable outputs: The model’s inbuilt explanatory power highlights the features influencing each prediction, helping operators understand why recommendations are made and increasing adoption in live plant operations.

    Result

    • Prediction uplift: 42% better than the existing physics-based model and 27% better than commercial simulation software on average.
    • Throughput gain: Applying the optimal set-points increased throughput by 12% on average.
    • Operator confidence: Clear explanations supported decision-making in live operations.

    < back to Case Studies

    Book Discovery Call