Our team worked with a supplier in the nuclear energy industry to help them solve a problem with maintenance supply chain issues.
In this instance most of the spare parts required for the power plant were no longer manufactured. Thus the supply available was fixed, and although new parts could be custom built where required it was becoming prohibitively expensive.
The customer had teams searching databases manually for parts they needed, the problem however was that the parts were often described differently in each database. Making it difficult to compare them.
They wanted a system that was able to learn from the data to compare parts and make recommendations to the user about which parts may be suitable.
For this problem we developed a similarity engine powered by deep siamese networks. Each part was described in a limited number of characters within the different databases. To add additional challenge to this problem, various engineering terminology was used interchangeably and parts could be described with either in metric or english units.
Our team worked with the customer to develop pre-processing methods to counter some of these problems. Our architecture for the solution was based on an active learning system that presented suggested part matches for the user to sentence as good or not. Each pair of parts sentenced formed more labelled training pairs used to improve the model. Given that the existing workflow was based on a combination of SQL query and excel, we were able to integrate into this workflow and provide an enhanced user experience.
Please contact us to help you solve any challenging NLP problems that you have.