The industry is moving in full speed towards ‘Industry 4.0’ , where automated processes become even more efficient by using advanced technologies such as the Internet of Things and Advanced Analytics. For some applications, this even applies to expensive and complex consumer products, such as High Power Electric Vehicle Chargers.
A large and innovative producer of Electric Vehicle chargers was unhappy with the amount of issues their dedicated Customer Service team was dealing with. This also resulted in an unnecessary high Mean Time To Repair. The time between failure and repair was too high due to a time consuming chain from end-user noticing the fault to the Customer Service team handling the issue. This resulted in two negative aspects. Firstly, customers become unhappy as they cannot fully rely on the charger. Secondly, an unreasonably high work load for the Customer Service team.
Luckily, the existing Customer Service infrastructure contained a vast amount of labelled fault data. This means that we knew exactly which data belonged to a healthy charger and which data belonged to a specific fault. We first had a look at which faults happened frequently and were causing the largest downtime. This provided us with the insight of the faults we needed to predict first. With the data and the goals sorted we were able to quickly develop our analytical models. Given the data, we developed custom Classification algorithms. These models analyse all the sensor data and look for patterns which are unique to one of the classes (healthy / faulty). For future predictions, the model looks for these unique fault patterns to predict whether a part is likely to fail. This is when two challenges arose. Firstly, we only had failure data for some of the chargers, while we needed to develop models that worked well for all. Secondly, these chargers are subject to frequent changes in software, hardware or environment, causing their behavior to deviate from the learned patterns. To overcome these challenges, we combined two state of the art optimization techniques which incorporate deviations in the data distributions and continuously monitors the data to adapt the unique fault patterns.
The resulting models are able to detect approximately 20% of all failures up to 2 months in advance! After going through the proper design cycle, we integrated the models into the existing Customer Service infrastructure. The models are running on their AWS Lambda, having direct access to all sensor data from the AWS S3 data storage. The model automatically generates tickets into the Customer Services JIRA platform once it predicts a failure. This allows both the Customer Service Team as well as the Customer to schedule maintenance right before failure, reducing Mean Time To Repair and maximizing charger availability!