Project “DEXAI”

The problem with rare errors

State-of-the-art IOT devices produce continuous data streams. When it comes to training AI models for predictive maintenance, however, the most important data − error data − is often missing. As devices produced by manufacturers like Rational AG are very reliable, critical failures occur extremely seldom. In addition, the error patterns are often very different, which makes it even more difficult to generalise conventional models.

Conventional deep learning approaches often fail here, as they require large quantities of example errors. In addition, service technicians often find themselves facing acceptance problems: if an AI solution recommends replacing an expensive component but cannot explain why, the recommendation is often ignored. What are needed, therefore, are AI solutions which can learn using small amounts of data and whose decisions are transparent for humans.

DEXAI uses smart data rather than big data

Our approach consistently focuses on two core mechanisms to address both the predictive maintenance data dilemma and the barriers to acceptance by service technicians:

Maximum data efficiency: Rather than uploading quantities of raw data to the cloud, we use intelligent data preparation and sampling strategies to make an optimum selection of training data. We employ synthetic data generation methods to train generalisable models using minimal training data and maximum generalisability.

Anomaly detection reimagined (XAI approach): Anomalies detected by an AI model must enable conclusions to be drawn about malfunctions and be understandable by service technicians. We’re developing methods of explaining anomalies in high-dimensional time series data and drawing conclusions regarding potential error sources in order to enable targeted maintenance.

Added value for industry and environment

The DEXAI technologies not only aim to improve maintenance, but they also make IoT applications more economical and sustainable:

Sustainability and economy: The tremendous reduction in the quantities of data required means that cloud storage and computing resources are significantly reduced.

Reduced operating costs through XAI-supported diagnosis: Technicians not only receive error messages, they are provided with complete diagnoses. This significantly increases the First-Time Fix Rate (FTFR) and prevents unnecessary callouts and downtime, thus considerably reducing costs.

Industry-independent application: Although our research is based on the example of food preparation systems, the framework developed for synthetic data and feature selection is industry-independent and can be transferred to production or logistics systems.