Helping a Logistics Company save millions on future maintenance costs
In August this year, someone in my network who works as a Data Scientist for a Logistics company in India reached out for advise on a Data Science project. What got me interested and involved in the project was that this person who reached out was extremely clear on the problem he was looking to build the initial solution around vs saying – I want to develop a predictive maintenance model since it is the next shiny thing.
This company has a fleet of hundreds of Diesel engine trucks and leadership was trying to be proactive about BS VI regulations, specifically regulations around Diesel engines, that will be coming out in 2020.
To be proactive, they had started fitting trucks with Diesel particulate filters (DPFs). A DPF is a filter that captures and stores exhaust soot (some refer to them as soot traps) in order to reduce emissions from diesel cars. While DPFs are certainly great for the environment, they can be a headache for the average fleet to manage. Why? Because not only must DPFs need to be kept clean, but there are also numerous factors that affect how quickly soot builds up in them.
The new trucks that were being fitted with DPFs came equipped with multitude of sensors as well. The analytics team was thinking about developing a fault detection and predictive maintenance model to minimize additional cost from maintaining the DPFs.
Developing a Python Model for fault detection and preventive maintenance
In less than three months, I was able to help a small team of three pretty “Green” new grads create a production scale model. Credit goes solely to those three bright individuals who were extremely sharp to pick up suggestions and run with it. Credit also goes to two OEM engineers from the truck manufacturers who not only helped with sensor installation but also helped the team in India setup a device management infrastructure to effectively collect, process and visualize sensor data.
Below, I have discussed two variations of the model.
Example 1: Fault Detection
The team developed a Predictive Maintenance and fault detection model that collects and analyzes sensors data.
DPF failures are very common on trucks because, like any other filter, they’re prone to clogging over time. However, DPF failure on trucks is critical -A vehicle not being able to pass exhaust fully through its system, is akin to a vehicle not being able to breathe. And if this happens, this may lead to malfunctioning of other systems and equipments on the truck, and the engine will start to deliberately constrain its power.
To capture this behavior, we created a model that analyzes data from multiple sensors on the truck to understand the vital operating factors that impact fuel usage — including acceleration, speed and temperature.
Once the model was deployed and went live on the assets, it quickly identified that some trucks have been burning up to 20% more fuel than normal over the past four months. To put that cost into perspective: The cost of the excess fuel consumed for all of the trips taken by that one truck during that time amounts to approximately INR 18,000; that cost more than doubles the cost of the original repair of INR 7,000, had the issue been known about and addressed at the onset.
By alertingfleet managers to the problem and equipping their maintenance team with this actionable insight, technicians were able to prioritize the truck for repair, fix the DPF before other problems resulted, and get the truck out of the shop and back onto the road with optimal fuel economy.
Example 2: Predictive Maintenance
Sensors on the DPF setup detect and collect key data points to measure critical parameters. Based on the data, you then determine which signals can be used to tune your model. In the specific case of DPF, you can tune your DPF model using two unique signals:
- Differential pressure: Sensors at both the inlet and outlet of the DPF measure the differential pressure of the aftertreatment system. A clogged DPF causes the pressure to be much higher on the front end — where exhaust is having difficulty passing through the filter — than on the back end.
- Temperature: When soot builds up inside a clogged DPF, forced regeneration needs to happen. This triggered event forces up the temperature of the aftertreatment system in an attempt to burn off some of the soot. It requires the driver to pull over and initiate a self-cleaning process for the truck, which can take up to an hour to complete. Forced regeneration becomes less effective the more clogged the DPF is.
By correlating these signals and detecting anomalous patterns early on — or, in other words, when the engine heats up and tries to clear itself, but an insufficient amount of exhaust is able to clear the filter — the model can accurately predict future DPF failures. This can give fleet owners and operators weeks of lead time before those failures occur.
Armed with advance warning, fleets can proactively schedule their trucks to come into the shop for planned DPF maintenance. They effectively minimize the risk of those vehicles breaking down on the side of the road while upholding their commitment to driver safety and on-time shipments.
Next steps that the team is working on now is extending the model to a holistic fleet management scenario. I will not be involved in this journey going forward from I am pretty sure that the team has all the ingredients to be successful.
Based on my own advisory role and inputs from Data Scientists at a top five Logistics company in India.