Client

Our client is one of the leading car manufacturers from Europe that produces and sells more than half a million cars per year, and generates annual revenue of over €100 billion (2019). 

Business Challenge

As a part of the production process, molten aluminium is injected into a die cavity, mounted in a machine, in which it solidifies quickly and forms an engine block. Engineers visually check these engine blocks to verify if they are casted well. This project aimed at developing and productionizing a custom algorithm that will decide whether an engine block will be produced correctly during the casting process in order to avoid further processing and scraps, which would lead to financial savings for the manufacturers.

Our solution

Machine Learning

During the first set of consultations with our client, we identified two thermal cameras that emit thermal images of each engine block as a data source for the algorithm’s training and inference. However, using only the thermal source hindered the algorithm’s ability to correctly classify casted engine blocks, which we proved by experimenting with Computer Vision model only in the first iteration of development. We advised and identified multiple sensors that emit measurements such as pistons speed, intensification pressures etc., and used them as an additional data source that will enhance our Computer Vision model.

Our final solution used a two-branch Deep Learning model:

• First branch processed tabular sensor data
• Second branch processed thermal images data - four thermal images per engine block

Data Engineering

We developed two main data pipelines for training process: 

• One for transferring the data from our client’s factory server to their remote, powerful server capable of running complex Deep Learning model
• Second for transforming and normalising the dataset, and feeding it to the model

For the inference phase, data traverses to the model through the same data pipelines, our algorithm produces predictions and sends them back to the factory’s server. 

Additionally, a cron job runs every 5 days, retrains the algorithm, and discards the outdated data. 

Service Provided

Consultations, Design, Machine Learning, Data Engineering, Testing, Maintenance.

Project Scope

5 months project delivered by a cross-functional team of a Machine Learning Engineer, Data Engineer, product owner and a scrum master. 

Results 

Our solution provided an accuracy comparable to an expert engineer and provided significant time savings in manufacturing process, and financial savings.

The algorithm and data pipelines have been tested in production and were observed to be reliable, and able to handle the load of the client’s needs.