Customized video analytics feature extraction for next generation of digital twins in Industry 5.0: lightweight microservice-based on-the-edge model powered by bootstrap learning and deep edge technology

Digital twin of an injection moulding process is a total solution predicting supply chain, production cycle, needed maintenance, workforce management, etc. Data gathering from the machine and the production line is the key to realize the virtual model in an as close as possible fashion to its real-world representation. The challenge is twofold:

1. Data should be already available to train AI models.
2. Limited data is available from installed sensor networks on injection moulding machine as well as production line: environmental humidity/temperature, geometry check of produced parts, black spots, etc.

The data gathering phase is typically used in current digital twin products to address the first issue. The second issue is also typically solved using already available sensor networks. The first issue hampers easy and cost-effective adoption of the digital twin concept. The second issue limits the real-world accuracy of the virtual model.

eTwin addresses both issues. The RGBT video of the machine and production line products offers rich visual features. This is highly favorable for any digital twin or smart predictive maintenance platform. Video features are extracted in real-time on-the-edge, for which no external GPU or cloud access is required. In addition to extracting video features, eTwin trains a lightweight edge model using minimum amount of data, namely a few videos. No preprocessed clean dataset is used beforehand to train huge models. The lightweight trained model improves itself through bootstrap learning over time.

In this experiment, eTwin will be showcased in Elvez, a Slovenian manufacturing SME. Elvez has expertise in the production and manufacturing of cable harnesses, cable sets/accessories and wiring, various power cords, and injection molding of plastic parts. Elvez is our manufacturing company partner and willing to carry out this experiment in their production line. We have also an expression of interest from Trygons3, a Greek manufacturing company. eTwin can be used stand-alone for predictive maintenance and quality control purposes using our lightweight bootstrapped trained model. The manufacturing company can investigate and delve into data and our interpretation using the dashboard provided on an on-premise server. The data can also be transferred using IIoT protocols to other digital twin or similar products already installed in place.