Predictive maintenance

Official figures on production robot downtime are difficult to come by – and the figures that are available vary greatly, making it hard to get a realistic picture of the downtime costs incurred. However, we do know for a fact that downtime costs in the automotive industry add up to tens of millions per company every year.

Rigid maintenance models are the order of the day

Robot manufacturers address this problem with more or less rigid maintenance models. Corrective and preventive models are currently common (whereby selective corrective maintenance leads to faster wear). In practice, robots are often only serviced when they require repair, when the operating hours require it, or when preventive maintenance is carried out without taking the machine’s condition into account.

Robot manufacturers have already reacted to this problem and have established condition monitoring systems in many places. In this case, a prediction for future maintenance requirements is made based on the actual condition of a machine. However, the forecasts remain relatively vague. At the end of the day, condition monitoring is only a refined form of preventive maintenance. The path to predictive maintenance is not far.

Data collection: “The more the better” – but how to transfer it?

The main problem in assessing the future state of a machine is likely to be the availability of data. This is because a thorough observation, e.g. of vibration data, often generates a large amount of data that can hardly be transferred via the network infrastructure. On the other hand, many machines are still monitored with a few sensors, some of which are not very powerful, that only capture a partial spectrum of the data. However, unlike in many other areas of life, the principle of “the more the better” applies to data collection – so the more data that can be used for monitoring, the more reliable statements about the future machine condition can be.

This difficulty is usually countered today with the help of so-called edge solutions. Here, an algorithm or even an AI tries to filter out the relevant data and transfer only this to the control system, where the actual evaluation then takes place – and still requires a high degree of (costly) computing power.

Higher precision at lower costs

So if you want to minimize the risk of robot failure while also reducing costs, the obvious solution is to evaluate the sensor data at the point of origin. This has been possible for a few years now thanks to the increasing performance of semiconductors, on which an AI “embedded” runs using highly developed special processes. These embedded AI sensors only transmit the evaluation result, thus reducing the transmission volume to a minimum. At the same time, however, the ability to process even very large amounts of data is increasing, enabling a significantly deeper and more precise evaluation. This not only allows the current degree of wear to be detected (condition monitoring), but also enables precise predictions to be made about the service life of a component or even the entire machine (predictive maintenance). The advantage of AI over algorithms is that it can recognize complex and otherwise unpredictable events as anomalies and trigger appropriate actions.

However, embedded AI not only has the advantage of deeper data evaluation, but also significant cost advantages due to its low resource requirements. This means significantly more performance for less money. And accurate predictions of an impending failure make servicing more flexible, faster and more cost-effective. An efficient service strategy replaces maintenance intervals. This creates a win-win situation for manufacturers and customers.

Instead of a modular system: customer- and use case-specific development

Sensors equipped with AI are developed as “embedded AI system components” on a customer-specific basis. The focus is always on the specific use case for which the AI and the sensor board are developed and can later be mass-produced. Such system components are designed to establish connectivity with the bus used (Can, Lin, etc.). In addition to the development of the AI and electronics, the appropriate installation space is also sought where the sensors can be optimally placed. Compared to conventional AI modular systems, such systems are clearly at an advantage.

With the help of such components, drives, joints, gearboxes, bearings or even hydraulic drives can be monitored using vibration, ultrasound or lasers. Lasers are also suitable when the sensor can only be attached outside the robot. Complex use cases may require the use of multiple sensors (sensor fusion).

Better grip

But embedded AI offers even more possibilities in robotics: collaboration can be improved with the help of gesture or voice control; person recognition solutions ensure more security without affecting data protection issues. And actuators and tools can also be improved using embedded AI. For example, grippers can be equipped with pressure sensors that can provide real-time feedback on whether an object has been gripped correctly.

AITAD GmbH in Offenburg focuses on the emerging field of embedded AI and is the only cross-industry provider of this technology worldwide. Here, an interdisciplinary team develops AI models and electronic components that ensure greater reliability and security. The core competencies are predictive maintenance, user interaction and functional innovations.