August, 2021

Digital Twin: Prediction of product quality data

Q-DAS’s ( main objective in the InterQ project is the development of a digital twin of a real production line of its project partners.

The purpose of this digital twin is the prediction of product quality. Predictive quality can be used in various ways, e.g. to reduce the measurement effort, for maintenance scheduling or even for closed loop solutions.

The heart of the digital twin is an AI engine based on supervised machine learningSupervised learning requires labelled training data of a certain quality.

While often plenty of data is being produced in production processes, one of the main difficulties is the missing link between production and quality data.

Therefore, during the initial step of the project, in an analysis of the production domain and its data and quality data management the following question need to addressed:

  • Can process, sensor and quality data be linked?
  • Are changes possible to the data management to enable data association,
    e.g. by using product ID’s and time stamps?

Figure 1: A very rough sketch of the data collection idea.

The following picture shows an example of process data and product data which both share the same Product ID and Feature ID. In this example the mapping of the desired product quality data and the related process machining data is possible.

    Figure 2: Example of Product and Process data which both share the same Product ID and Feature ID

    When a good data collection process is implemented and training data is available, we enter the model training phase.

    The model training module consists of a data loader (e.g. loading specific data from the database), preprocessing (data cleansing, encoding) and an auto-machine-learning engine (executing hyperparameter search) A successfully trained model is persisted in the AI-model database. 

    Figure 3: Model Training

    If the prediction quality is validated, we enter the prediction phase. In the prediction phase, a prediction engine is deployed to the production line. It loads the appropriate model from the AI-model-database.

    While the production process is going on it generates quality predictions from process and sensor data and intermittently evaluates the model validity. When predictive and measured quality deviate, a retraining may become necessary.

    It is also possible to feed back data to the training data set from interactive assessment of the predicted quality. The following picture is a visualization of the concept.

    Figure 4: Model prediction