The theme of this year’s programming hackathon is IIoT: Asset Health Monitoring, Predictive Analytics, and Maintenance Optimization of industrial mobile assets. The goal is to create valuable applications that leverage haul trucks data from one of the Barrick Gold's sites.
In this blogpost you will see our approach to minimize fuel consumption (which represents a major expense for Barrick).
We decide to create a KPI called EEOI which would take into account:
- The trip distance
- The fuel consumption
- The load carried by the haul truck.
We decided to leverage Azure Machine Learning to create a model representing the EEOI. Once the model is trained and deployed in Azure, we use the model to compute an expected EEOI based on the real-time data in the PI system. We can then compare the expected EEOI versus the measured EEOI for a trip.
Here is our solution’s architecture:
We also leverage the PI System by using these tools:
- Event Frame
- AF Analysis
- Custom Data-Reference to query the Azure Machine Learning Web Service
The machine learning work in two steps:
- Train the model
- Use the model
In the training process we identified our model features (variables we think we can use to predict the EEOI). For the exercise, we focused on:
- The Truck Id
- The payload
- The average speed over the course.
To generate the data needed to train the model, we used AF Analysis to create Event Frames containing statistics for each truck voyage over a period. To achieve that we used historical data to compute the EEOI. Here is an event frame we used:
From this data, we created a model using Azure Machine Learning Studio. This purpose of the model is to mathematically represent a complex behavior which is difficult (near impossible) to be implement into a deterministic algorithm. For the purpose of the project we used a linear regression model.
Once the model is trained, we used a custom data-reference in AF to compute an expected EEOI by querying the Azure Machine Learning web service for each truck’s voyage. Now that we have an expected EEOI (based on the model) we can compare with the measured EEOI for each voyage and identify which speed target based on the payload and/or the truck. We used an AngularJS front-end to visualize the data.
We found that Machine Learning can be really powerful but there are challenges:
- Accurate sensors and data are essential
- Fuel rate wasn’t accurate due to the type of sensor used. Another approach (fuel level sensor) would give better result;
- Data contextualization must be well performed for a better result (feature selection);
- Segmentation of the road to enhance the model to take road configuration into account would also give better result.
New opportunities can be implemented based on the work done during the Hackathon. We identified two of them:
- Assist the operator in its driving behavior
- Optimize driving behavior on some route segments based on the route conditions
Integrating the PI System infrastructure with Azure Machine Learning is opening a new way to analyse and predict insdutrial process events. It's now the time for historical data to predict the future...