Patent DE 10 2024 100 703

For efficient production processes, it is crucial to detect faults in machines and systems and their components, such as bearings, at an early stage, minimise downtimes and avoid costly repairs. Key concepts in the field of industrial maintenance are therefore condition monitoring and predictive maintenance. Wireless sensors ensure continuous monitoring and condition detection in real time. A new method for data-reduced communication between wireless sensor nodes and monitoring systems makes a major contribution to reducing the energy consumption of wireless sensors, enabling them to be used in numerous locations over many years, even those that are difficult to access.

When monitoring industrial systems, data from wireless sensors is analysed to detect potential problems or deviations from normal operating conditions at an early stage and to predict maintenance requirements. However, continuous real-time monitoring via wireless sensors can have limitations: If the sensors are located in many places or in places that are difficult to access and are expected to work reliably for years, their energy consumption is critical in terms of battery life. The sensors, which are interesting for monitoring, also generate a large amount of data. To date, this can only be transferred to the central monitoring device with a high energy consumption. In addition, defects in machines occur very rarely or not at all during normal operation and certain error states, such as defective bearings and machine parts, cannot be easily triggered.

The idea behind the invention is an algorithm that collects data and learns the normal machine behaviour patterns directly on the sensor, largely unsupervised. The regular behaviour patterns of a system and certain novelty criteria, which stand for deviating behaviour, are learned from the recorded data on the wireless sensor. Data is only sent to a central monitoring system if the calculated novelty value deviates from the normal criteria for "healthy behaviour". This completely eliminates the need to transmit raw data to a central monitoring system. This on-device learning and subsequent retraining make the algorithm generalisable for various industrial scenarios and robust against possible data drift. The method for recognising the novelty value of data combines mathematical processes in a specific sequence for ongoing condition monitoring and predictive maintenance over many years.

On-device data reduction using singular value decomposition and correlation: In the initialisation phase, the proper operation (good condition) of a bearing, for example, is recorded and a threshold value for detecting irregular conditions is determined by means of automated feature extraction using singular value decomposition. This is further improved in the observation phase by learning certain anomalies and states. By using only the first k dominant singular values and the corresponding parts of the singular value decomposition matrices, a reduced data set is generated on the sensor node for transmission to the monitoring system. In the application phase, a feature vector of newly recorded vibration values is generated. Canonical correlation analysis is used to analyse how different the learned features of the good state are compared to the newly recorded feature and thus qualitatively record changes in state.

Advantages of the invention

• Data processing on the sensor node
• Data reduction in radio transmission
• No need to record fault data and error states in advance
• Reduction in the energy requirements of wireless sensors for increased service life