Data-Driven Digital Shadows for Process Manufacturing
Author: Iason Tzanetatos
26th June 2025
Circular TwAIn is a Horizon EU project which aims to research, develop, validate, and exploit a novel AI platform for circular manufacturing value chains. This AI platform can support the development of interoperable circular twins for end-to-end sustainability.
Within the scope of our activities in the Circular TwAIn project, CORE IC is responsible for the design, development and deployment of a data-driven Digital Shadow for an industrial process, such as that of a petrochemical plant (the Circular TwAIN end user, SOCAR).
In process manufacturing, particularly in complex and highly regulated industries like petrochemicals, the ability to monitor, analyse, and optimise operations in real time is essential for maintaining efficiency and competitiveness. The integration of advanced digital technologies is reshaping traditional operations, enhancing performance, and driving innovation. To be able to realise this model, we have utilised historical operational data from our partners at SOCAR, that they have kindly shared with the consortium.
Architectural Design of Process Digital Shadow
Working with our partners at TEKNOPAR and SOCAR, we defined their requirements from the technology and identified the following key objectives:
1. Perform anomaly detection on real-time sensorial data, that depict the current conditions of the plant
2. Predict what the sensor readings will be in a 10-minute horizon (i.e., predict short-term future state of the plant)
3. Perform anomaly detection on the forecasts of the plant.
We started the development of a Digital Shadow for a manufacturing process by identifying the involved assets of the production line. Since we are dealing with a process where the involved assets interact with one another, the outputs of some of the assets are considered as inputs to other assets down the line.
After the involved assets and their interactions had been mapped, we moved to the mapping of the sensorial inputs/outputs of each asset. Relevant information such as SCADA schemes were used to determine the sensors that are considered as inputs and outputs of each machine.
Data Augmentation Techniques – Asphalt Use case
To successfully identify any anomalous conditions for each sensorial signal, irrespective of the input/output characterization of the involved sensors, we monitored each signal individually.
We proceeded with training an Autoencoder-like Deep Learning model, to identify anomalous conditions on a per sensor level, meaning that the model examines the data point of each sensor individually.
Autoencoder Neural Network Architecture
As depicted in the figure above, an Autoencoder model comprises three main components:
1. The Encoder, where the input information is compressed
2. The Bottleneck layer, where a compressed low dimensional representation of the input is determined by the model
3. The Decoder, which reconstructs the input relying only on information retrieved by the Bottleneck layer
By training an Autoencoder model with high quality, normally characterised operational data, we essentially have a model that can identify significant deviations on the operations of the involved assets.
However, since this use-case is particularly complex, niche models such as Reservoir Computing deep learning models. This family of models is best suited for time series data with complex patterns, similarly to the operational data of a manufacturing process.
From: Quantum reservoir computing implementation on coherently coupled quantum oscillators
By adopting the approach of the Autoencoder model, we formulate the problem in the exact manner, only we swap models and use a Reservoir Computing model.
Forecasting
To proceed with forecasting 10-minutes ahead of the manufacturing process, our model needs to follow the same interactions of the assets as in the actual plant.
Production flow on Petrochemical line
The process comprises three main components, a Reactor, an Absorber and a Stripper. Each asset interacts back and forth with each other, and these interactions must be replicated in the digital realm as well.
A Digital Shadow has been developed for each physical asset. We utilised the same family of Deep Learning models to perform forecasting, by switching the objective of the models. As a final step, we connected each model by following the physical domain, as previously mentioned.