AI tested in service – availability higher than 99%
Artificial intelligence (AI) is on everyone’s lips, as it opens up a wide range of possibilities, including improving products.
“Predictive maintenance”, also abbreviated PMX, is another buzzword. Predictable maintenance maximizes availability. The prerequisite for this is that you have historical data on the devices.
The conditions for PMX are good at SBRS. Our award-winning backend system Chargeview collects the data of the charging processes and statuses over a year.
Together with the Niederrhein University of Applied Sciences, a study was carried out in which all data was systematically combed through with an AI to identify patterns for possible predictable errors.
To train the network, the data sets of all charging stations in the period from January to September 22 and in the period from April to December 23 were considered. This data is used to look for the occurrence of a “non-repairable” failure of the charging station, which requires maintenance and repair by a service technician. The so-called “ChargingState” then assumes a certain error value.
The most important finding:
The total number of failures that have occurred at all charging stations so far is low. This means that the prerequisite mentioned at the beginning of the study in form of a sufficiently large amount of training data is not given. However, the availability of the largest possible amount of training data is the decisive factor for the successful application of AI approaches.
On the one hand, this is good, or rather, excellent as a figurehead for our technology. On the other hand, the final jump from perhaps 99% to almost 100% will probably not be possible with the help of AI.
Another difficulty for AI is the structural diversity of the stations, which have divergent feature values as a result of technical changes and further developments, or do not have certain features at all. Training networks with different data sets as input values of the network is difficult and requires either a focus on certain types of charging stations or a significantly larger amount of data for training.
Interesting first findings. We’ll stay tuned!
