What is KIPOS?
The acronym KIPOS stands for the joint project of a four-member consortium consisting of antares Informations-Systeme GmbH, Hahn-Schickard Gesellschaft e.V. and Weisser & Grießhaber GmbH. The project name stands for "Artificial Intelligence for Process Optimization in Injection Moulding".
The research project was funded by the Central Innovation Program for SMEs (ZIM) and lasted a total of 2.5 years.
The end result of the project was to be a new type of software solution that optimally supports the operator of the injection molding machines. With the help of integrated sensors and machine interfaces for inline measurement, process models were to be created to help predict process and component quality and optimize these with setting parameter recommendations. Modern artificial intelligence methods were used to analyze the data.
The participants

antares Informations-Systeme GmbH

Hahn-Schickard Society for Applied Research e.V.

Weisser & Grießhaber GmbH

IMS Gear SE & Co. KGaA
Project process
-
Start
Project start
The overall project is divided into 3 sub-projects, with each party taking on one sub-project depending on its own competencies.
As an expert in the field of software technologies, antares is developing an interaction platform for the worker in the injection molding process. The tool is intended to provide the operator with AI-supported parameter recommendations based on real-time process data and thus optimally improve the process.
Weißer & Grießhaber is responsible for the implementation and validation of the AI process model based on training data from industrial production.
The AI-based process model is being developed by Hahn-Schickard. It will be used for process and component optimization in the plastic injection moulding process. -
7/20
Design of a software architecture model
Based on current technology, a hybrid architecture was chosen that includes both local and cloud components at project partners. This chosen overall architecture provides the ideal basis for the research project, as it offers cloud-based scalability and enables the connection of local IoT devices such as cameras and sensors.
In order to collect, process and use data from various sources, interfaces to all databases and text sources as well as templates for data visualization are integrated. -
12/20
Development of a software solution for static design of experiments (DoE)
The web application allows you to create an experimental design with various settings, including input and output parameters as well as the definition of the name and the model to be used (full/partial factorial designs, designs according to the Taguchi method). Values for the previously defined factors and specific settings can then be entered.
The subsequent view of the resulting plan shows each test run in which the specified parameter values must be set on the machine and the output parameters listed at the end must be measured and entered. All created plans are displayed in the plan overview, including useful information such as parameters or model. The current status is indicated by different colors, and it is possible to delete or subsequently edit plans. -
2/21
Implementation of a model for the detection of surface anomalies on plastic components
In order to detect defects, the focus is mainly on the use of a Convolutional Neural Network (CNN). For this purpose, the images obtained from a previous production process are divided into five different classes, with one class describing good components and the remaining four different anomalies. Using this method, over 93% of workpieces are already correctly categorized.
Subsequent hyperparameter tuning and classification using an algorithm ensure a further increase in accuracy. Thanks to pre-processing optimization, the computing effort can also be reduced to such an extent that the processing time for an image is reduced to 1/100 of a second.
-
2/21
Development of the assistance system
Even before the development phase of the assistance system, the decision was made in favor of an architecture model that would ensure problem-free scalability and maintainability. To this end, a microservice was developed that links all data from the project partners' various data sources in a central database
Further components of the assistance system summarized:
- Generating and maintaining the various models for predicting component quality
- Front-end application for viewing all data in real time
- Individual assignment of roles and authorizations to users
- Backend (service) for (real-time) communication between all endpoints
-
11/21
Development of a process model together with Hahn-Schickard
Different types of models are used to analyze the initial data that will be used to create the process models. Based on the results, a selection is made to use a limited number of models for further experiments. The selected model types include: 1. linear regression, 2. random forest, 3. support vector machine and 4. neural network.
The first adjustments already lead to improved results for all four models, so that even at this stage none of the generated models provide unrealistic values when predicting quality features. -
4/22
Software solution for the automatic creation of a process report
The individual process reports have a modular structure and can be called up at any level of abstraction depending on the data displayed. The generated reports are based on both current and historicized data.
For example, general information on the respective machines can be called up at the highest level of abstraction and reports can be created to view individual orders or analyze errors in the process. -
9/22
Installation and adaptation of signal processing and process models in the field at Weißer + Grießhaber and Hahn-Schickard
Thanks to the container architecture used and the prior data collection and preparation, the installation of the software solution in the industrial environment is largely problem-free. The adjustments only affect the local conditions, such as the number of integrated machines.
By making a few modifications, the model quality could be increased considerably, in particular by reducing the parameters and quality characteristics used for the prediction.
-
3/23
Project completion
Although the processes involved in injection molding are extremely complex and each machine has individual characteristics, we have developed a solid basic architecture that will serve as the basis for future injection molding projects.
The DoE tool is already available as a ready-to-use standard solution.
The models we developed for process parameters and for detecting surface anomalies are another success of the project. With very high accuracy in the detection of defective components and precise defect type categorization, excellent results were achieved. Provided certain conditions are met, existing models can easily be supplemented with additional data and retrained.
Our findings
Through our involvement in KIPOS over several years, we have gained valuable insights into the theoretical and practical aspects of the plastic injection molding industry. These insights have broadened our understanding of process optimization, so we believe that the methods and algorithms can be transferred and expanded to other industries, especially to process-like workflows in our customer companies.
We therefore recognize the opportunity to expand our areas of expertise and solutions and offer highly covered standard solutions on the market.
This is how it continues
Utilization of the findings: Like Hahn-Schickard, antares aims to use the research results to develop smart assistance systems that effectively support the machine operator in reducing production costs, improving the quality of the end products and increasing production speed.
We will combine the knowledge and effects gained from the project under a new area of expertise that addresses software-supported process optimization and the identification of production deficits. We will focus primarily on industry-related applications and requirements. The market launch of the solutions in phase 1 will largely take place in our existing customer environment, around 60% of which is active in series production.
Market launch concepts with partner companies are also a priority, in addition to our own measures. We will try to gain both consulting partners and suppliers of machine technology for this and thus drive market penetration forward much faster. The first steps towards market launch have already been taken.