en

Predict maintenance requirements for car rental

View this sample project to learn how to prepare data and build a model that predicts when rental cars will break down to maximize customer satisfaction

A more advanced project using real world data for Predictive Maintenance can be discovered here.

Replacing Cars Before They Break Down

I am data scientist at a large car-rental company. To make sure our clients don’t rent out cars that could break down, we replace the cars that we believe are likely to have technical problems regularly. But we don’t want to replace them too often either, to limit cost. Part of our business is optimising that balance.

I recently noticed that 3 months ago more of our cars had technical issues than usual. Last August, an unusual proportion of our cars had a problem with the transmission.

We had collected data about the cars that where affected by this problem of course, and how these cars had been used in the months before they broke.

As you can imagine, our goal was to identify the reasons for this problem and find out how to prevent it from happening again in the next months.

Business Goals

  1. Understand why cars are having technical problems at a higher rate
  2. Identify the cars with the highest probability of having the same problem in the future

How Did We Do This ?

We gathered data to solve those two goals: – measures taken during the various failures (august) – data about how cars had been used in the weeks before their incidents (june & july) – information from the maintenance department (june & july)

We followed 4 steps to go from that raw data to achieving our business goals, and preventing the next failures:

Step 1: Explore and prepare the data: We made several transformations to the raw data so we could get valuable information from it. Typically, we transformed that data to create a new dataset aggregated at the car level, with as many relevant features as possible about each vehicle.

Step 2: Create the model: DSS trained a model to predict the feature we wanted to understand (failure or not failure), using the historical data we computed in the previous step.

Step 3: Apply the model to the new data: We deployed the model to have the probability that each car would have the same failure in the next few months

Step 4: Make a decision: Knowing the probability of failure, we identified two groups of cars : the cars we had to check in the 2 next days because they were most likely to have issues, and the cars that we could wait a few extra days to check.