Retrofitting as the key to digitising existing machines
The forging industry largely manufactures safety-relevant products that are characterised by high strength. The high forces required for this are provided by presses and hammers, mostly of older generations. Many of these existing machines do not have consistent data interfaces, uniform documentation or integrated systems for production and quality data.
These machines often cannot be integrated into a digital environment, or only to a very limited extent. This complicates or prevents communication with Industry 4.0 structures, further automation and systematic data processing. Downtime, scrap and energy losses are detected, but without appropriate measurement systems, the causes can hardly be quantified, and measures can rarely be derived economically. At the same time, there is often a lack of processed operating data that could be used to specifically tap into manufacturing potential such as unit cost reduction, fault detection, statements on wear behaviour or the optimisation of cycle times.
This is why the DiReProFit research project was introduced. Through a digital retrofit, i.e. retrofitting with sensors and modern measurement technology, the process transparency and connectivity of existing, older machines could be improved without replacing core machine functionalities or interfering with the control system. For small and medium-sized enterprises in particular, this represents an economically viable alternative to the cost-intensive purchase of new Industry 4.0-compatible forming machines.
Avoiding bad investments in retrofitting
In the DiReProFit research project, the IPH and the LFM address a gap in common practice: companies need a preliminary assessment before they procure sensors, set up data infrastructure or commission external service providers in order to avoid potential bad investments. This is because technical retrofitting with sensors is now state of the art; the real challenge lies in selecting suitable sensor concepts, economically viable measurement data acquisition and, above all, the correct interpretation of the data obtained.
Small and medium-sized enterprises in particular have to plan retrofits with limited personnel, a tight budget and high production capacity utilisation. A wrong move ties up investments and blocks production time. Retrofit measures without clear objectives therefore harbour a considerable risk: if the sensor concept is unsuitable for the intended purpose, the desired measurement variables cannot be reliably recorded and investments remain without demonstrable benefits.
Regulation and liability pose a further obstacle. Many companies avoid interfering with existing control systems because a so-called “significant change” within the meaning of the Machinery Directive can trigger a new conformity assessment. In practice, this is known as CE marking: it documents that a machine meets basic safety requirements. If complete technical documentation is missing for old systems, this step becomes costly or fails altogether.
The DiReProFit research project addressed this conflict with an approach that reduced the planning effort and made it possible to assess the technical benefits of potential retrofit projects in advance.
Plannability, data quality and investment-proof decisions
The IPH and LFM created three building blocks that users can utilise for their own assessment.
First, the project team developed an overview of suitable sensor systems for digitalisation measures. In forging companies, robust, contamination-resistant measuring principles compete with requirements for cycle times, temperature ranges and installation space. The project results classify sensor types, measuring principles and typical integration paths in such a way that process managers and decision-makers can arrive at a reliable pre-selection more quickly.
To this end, the IPH and LFM designed exemplary retrofit concepts in a second project milestone and implemented them on a forging line during production: piece counting according to process steps, hot measurement for early quality assessment, and temperature and energy analyses for thermal process control. The implementation showed why pure piece counting via ejector signals from the control system can fail in series production: component-dependent multi-stroke sequences and set-up processes distort the interpretation. The LFM therefore combined control signals with external, redundant sensor technology. This created a database for the retrofit that can be integrated into the process context. However, maintenance must still be taken into account on the application side, for example for optical sensors that are exposed to scale, oil mist and other particles.
Hot part measurement provided another key benefit for users. LFM used the optical measurement technology of one of the companies involved in the project to detect geometric deviations in blanks while they were still hot. This sensor concept enables companies to identify tool wear or die damage at an earlier stage and plan tool changes in good time during ongoing production runs. This reduces quality costs, scrap and rework.
As a third milestone, the IPH developed a software demonstrator as a proof-of-concept that translates retrofit knowledge into application-oriented decision-making. The demonstrator is dialogue-based and provides text-based recommendations. At its core is a retrieval-augmented generation model (RAG model). This method combines a generative large language model with a pre-curated collection of documents: the system searches for suitable text passages in existing guidelines, literature, sensor overviews and practical reports and formulates an answer from them. This reduces the “hallucinations” typical of inference-based language models, i.e. AI outputs without a reliable basis.
Upon request, the demonstrator structures the planning of a digital retrofit for users, recommends sensor and integration paths, identifies typical risks, including the costs of CE marking, and remains manufacturer-independent. The demonstrator has been published on GitHub, together with guidelines and open source code, so that companies can understand and further develop the approach. In practice, the computing time depends on the performance of the hardware. While the Docker-based demonstrator can run on various operating systems, the institutes recommend installing it on a server with powerful hardware for distributed access.
Large language models with RAG systems as industrial knowledge tools
Companies that consistently use and further develop the approaches implemented in the project for a database-driven language model can use this to gain insights from their own stored documents and as an integral part of technical knowledge work. The language model then not only answers general questions, but also specifically accesses the company’s own knowledge base. This must be curated and, ideally, uniformly formatted. This is precisely where – as in the project – the advantage of the RAG approach lies, as it combines a generative large language model with answers based on verified documents, guidelines, sensor overviews, retrofit reports and other practical knowledge. This results in reliable, context-related answers without the language model itself having to be retrained for each new piece of information. Since answers to users’ questions are already available in the data material, they only need to be compiled. This also significantly reduces the number of answers that the language model has to generate itself from purely statistical word sequences.
In the long term, this approach will develop into a digital knowledge assistant for engineering, production, maintenance and investment planning. Employees tasked with planning retrofits use this system to ask questions about suitable sensor principles, integration methods, typical risks, documentation requirements or economic conflicts of interest. In addition to freely accessible technical information, the company stores its own documents, operating standards, empirical knowledge and project-specific documentation in the database. This expandability is key: user companies can supplement the general database with internal content, so that the answers become more case-specific and comprehensive as the knowledge base grows. The basic system architecture can also be used for other contexts with suitable data material.
Local AI preserves data sovereignty, is generalisable and makes knowledge usable
For companies, the strategic value of the local AI approach lies primarily in three points. Firstly, data sovereignty is maintained because the solution can be operated on a server or a workstation within the company, meaning that sensitive information does not have to leave the company. Secondly, generalisability increases: even companies without a long history of machine data or without fully digitised legacy systems can use the system to obtain reliable initial assessments and recommendations for action, which are valuable at the start of a project. Thirdly, the effort required to make new knowledge usable is reduced because new documents can be added to the database instead of repeatedly retraining a language model with high computing requirements.
The system implemented in the project is therefore not an autonomous decision-making AI system, but a reliable, document-based assistance system for industrial decisions. It supports specialists in preparing retrofits, selecting sensors and interfaces, evaluating technical and regulatory constraints, and converting empirical knowledge into reusable answers.
With DiReProFit, the IPH and the LFM have laid the foundation for this: a locally operable, manufacturer-independent and gradually expandable RAG system that becomes more useful with each additional knowledge source and can grow into a central tool for industrial knowledge work in the long term.
The final report of the research project is available in the research database of the Industrieverband Massivumformung e.V. (Industrial Association for Solid Forming) and on the IPH website.
