FIP MEC Srl is an Italian company with a long tradition in developing system that are safety critical. These components are expensive to manufacture and always need to pass stringent tests that measure the. Each time a prototype fails at testing the company needs to bear the cost of material, labour, and disposal. We engaged with them in an investigation to demonstrate how a digital twin built on previous data would be able to drastically reduce development time and costs.

Challenges

FIP MEC Srl had a state of the art laboratory but lacked the infrastructure needed to clean up and organise the unstructured results produced during testing and the formal certification process. This caused a bottleneck in the development process, as engineers had to spend a lot of time manually analysing data and running simulations to understand the results of tests and to design new prototypes. Furthermore, in cases when the test data was lost or corrupted, the company had no way to recover it, leading to costly delays and the need to repeat tests.

Solution Design

Given the challenges faced by FIP MEC Srl, we designed a solution that would allow them to clean up and organise their data, and to build a digital twin that could be used to simulate the behaviour of their components under different conditions. The digital twin was built using machine learning algorithms that were trained on the historical data produced during testing. This allowed us to create a model that could predict the behaviour of the components under different conditions, and to identify potential issues at design time, before the components were manufactured and tested.

Implementation

Data Collection

The first step in the implementation process was to collect and clean up the data produced during testing. We worked closely with the engineers at FIP MEC Srl to understand the data and to identify the relevant features that could be used to train the machine learning algorithms.

The raw data was composed of a series of reports coming from different testing machines in the internal lab and from the external certification process. We developed a pipeline that was able to extract the relevant information from these reports and to organise it in a structured format that could be used for keeping the Engineering team informed and for training machine learning models.

Digital Twin

Once the data was collected and organised, we trained a machine learning model to create a digital twin of the components. Given that this model was going to be used in a safety critical context, we focused on building a model that was not only accurate but also interpretable, so that engineers could understand the predictions and the underlying reasons behind them. We built the model by combining different sub-models that were able to capture different aspects of the behaviour of the components, such as their mechanical properties, their response to different loads, and their failure modes. This allowed us to create a digital twin that was able to simulate the behaviour of the components under different conditions without losing interpretability.

Deployment

FIP MEC had an IT Team and on premise cluster to run most of the software for internal use. Hence, we designed the system in such a way that could be delivered both on prem and cloud.

Summary

This was a landmark project for QuasiScience: it represented our first engagement outside the UK and the first consulting project. As a byproduct of this work, our team developed core capabilities to take a product from inception to market in a very short time frame and with a clear focus on the client.

The software originally developed to answer FIP MEC’s business needs has seen many updates and iterations over time. We called it is now available to all businesses that look for a tailored and effective experiment tracking software and results and accelerating the development process.