Here you will find a list of the most common questions about our product.
Classic systems are given rules to recognise defects. A ML model, on the other hand, learns from examples whether a part is OK or NOK. In this learning process, the model derives implicit rules that do not require explicit programming (e.g. “NOK, if scratch on top left”). Machine Learning allows Maddox AI to automate even very complex use cases, such as surface recognition.
In contrast to classic quality control systems, the Maddox AI system is very robust in the face of high error variability and changing external influencing factors. Typical problems of classic systems, such as high pseudo rejection rates, can be prevented.
Yes, Maddox AI’s software is designed in such a way that everybody can train AI models and put them into use after an initial thirty-minute training session.
Yes, Maddox AI bears the entire investment risk. Only when we have been able to show you that our model achieves the predefined performance KPIs (e.g. inspection accuracy and speed) you enter a paid licensing model.
This depends on the specific use case. Before we start with a Maddox AI project, we show you transparently how much the use of Maddox AI would cost. Based on this, you can calculate your business case.
A small series is undoubtedly more challenging for an AI system. However, we have already successfully implemented several systems in manufactory-like productions. Let’s talk about your potential use cases. There are some interesting possibilities that AI offers to create added value for you even at lower production volumes.
Maddox AI learns the textures and the characteristics of defects. Once the AI has learned what constitutes a scratch, it can recognise scratches in all variations on the entire component.
In the vast majority of cases, high production speed is not a problem. Our models run on our industrial PC at your site, which means that the AI model analyses images within milliseconds. In our experience, it is the part handling rather than the AI evaluation that becomes the possible automation bottleneck.
Yes, that is usually no problem. We adapt the recording hardware used (camera, lighting, mounting, etc.) completely to your use case. Since our system is manufacturer-independent, we can always provide the optimal camera setup for your individual inspection case.
Yes, this is one of the things that makes machine learning (ML) so valuable for quality control. An ML model learns, based on examples, whether a part is OK or NOK. Through this approach, the model derives implicit OK/NOK rules, that do not require explicit programming (e.g. “NOK if scratch on top left”). A cleanly trained model knows the properties of an OK part and therefore reacts very robustly to defect variance or a not quite constant image capturing conditions (e.g. different positions of the product in the image).
This is absolutely possible. Maddox AI can communicate with all common robots and also ERP systems via an interface. By default, we can output our data e.g. via OPCUA, MQTT, REST or TCP-IP. If you have further/different communication requirements, we would be happy to discuss these in more detail in a short call.
Partial automation is possible at any time. In addition to a high-performance ML-based inspection system, Maddox AI always comes with an industrial PC and an industrial touchscreen monitor. This setup shows you live, directly at the inspection point, whether a part has been assessed as OK/NOK and supports the skilled worker, who only has to carry out the subsequent work steps (sorting, packing, etc.).
This depends very much on how quickly we get enough data (pictures of OK/NOK parts). Since Maddox AI is an AI system, data is very important as a basis. We therefore always ask our customers to collect bad parts in advance. We can usually record these parts directly on the installation day and ideally already have enough data to train the model. In this case, the implementation only takes a few days. If we need to collect more data, the system can also run for a few weeks before the implementation is completed.
The important thing is not necessarily the amount of data, but its diversity. This means that different appearances of error types are relevant for a good and representative sample of defecets. We do not need many OK images, as the diversity in these images is usually smaller than for NOK images. Accordingly, our focus is on images with defects. We can give you the following rough rule of thumb: In 90% of the cases, 50 images per task (e.g. 50 NOK images of a defect class) are sufficient. In 10% of the cases, another 50 images (i.e. 100 in total) are needed to reliably adapt the Maddox AI system to your use case.
That is absolutely possible. Depending on the installed setup and cycle time requirements, the camera can be connected directly to our industrial PC or images can be shared via shared drives or file transfer protocols.
Yes, this is not a problem. Maddox AI makes the results of the AI available via several protocols that are common in the industry, such as OPC-UA, REST API, MQTT, Modbus TCP or simple digital I/O.
We normally store the image data and the associated metadata (AI classification result = OK/NOK, timestamp + product type if applicable). So far, our system is used by large medium-sized companies, but also listed companies. Of course, the data is not stored in a silo, but you as a customer can access it completely at any time and transfer it e.g. to your internal MES system or a data lake.
Yes, Maddox AI places great emphasis on this. The Maddox AI system provides the ability to recognise people in the images to blur or black them out before uploading them to the cloud. No one has access to the people’s identities – not even we can recover these images.
Yes, Maddox AI is a classic SaaS (software as service) solution. Accordingly, you will receive software updates regularly and free of charge.
No, the software updates, maintenance and other after-sales services are included in the licence fee.
This is not a problem, we are happy to re-train the Maddox AI system for new error classes free of charge.
We are ISO 27001 certified.
