Undetected corrosion damage can lead to incalculable economic and environmental costs during the operation of industrial pipelines. For cost reasons, the early detection of such damage (predictive maintenance) is usually performed by passively operated robots, which are routed through the pipeline system during its normal operation. Such robots, however, record huge amounts of data. In this project, methods of machine learning and signal processing, among others, were adapted to achieve the fully automatic analysis of this sensor data.
The inspection of pipeline systems nowadays takes place using inspection tools, which can be routed through pipelines during normal operation. During this process a lot of different sensory data is collected. The currently used manual evaluation is related to a very high personnel and temporal expenditure. To reduce these costs and to improve the accuracy of the evaluation - which leads to an enormous increase of competitiveness – parts of the processing chain can be automated with the use of modern Machine Learning methods.
The significant development goals contain a new procedure for automatic sensor calibration and the automatic segmentation of pipeline defects and damages, as well as any other anomalies. In order to reduce the incorrect segmentation of basically harmless signal characteristics, various pipeline characteristics and interfering objects (e.g. valves) need to be taken into account to enable a distinction to actual defects. The goal is to reduce the manual amount of work for up to 80% and the reduction of detection errors for up to 70% (1st type of error) or rather 50% (2nd type of error).