Softwareentwicklung
Pro Juventute: MINT Skills bei Jugendlichen in der Berufswahl fördern
Mehr erfahren
Homepage zuehlke.com
Öffentlicher Sektor & Verwaltungen
Der Earth Control Measure Evaluator des PUB: Entwicklung eines Prototyps mit maschinellem Lernen
-
Bei der Wasserbehörde von Singapur gehen jedes Jahr über 1500 Zeichnungen ein, deren manuelle Prüfung Zeit und ein geübtes Auge braucht.
-
Die Herausforderung? Wie lässt sich mit maschinellem Lernen der Zeitaufwand bei der manuellen Prüfung reduzieren, ohne Fehler zu übersehen?
-
Hier erfahren Sie, wie in der Cloud ein vollständiger Prototyp entwickelt und für die Erkennung bestehender Zeichnungen und darin vorkommender häufiger Fehler trainiert wurde, um den Gesamtprozess zu optimieren.
Situated near the equator, Singapore has a typically tropical climate with abundant rainfall. It rains an average of 167 days a year, with an annual total rainfall of around 2,500mm. During heavy rainfall, silt can be washed from exposed earth surfaces in construction sites into waterways. Good Earth Control Measures (ECM) implemented in construction sites prevent silty runoff from being washed into waterways, keeping them clean and beautiful.
PUB, Singapore’s national water agency, receives more than 1,500 drawing submissions per year seeking approval for construction works that require Earth Control Measures (ECM) to be implemented. This process requires time as well as an experienced eye to manually review the drawings and ensure that the ECMs are properly designed.
Improve process efficiency with machine learning
Together with PUB, Zühlke found a way to augment the existing workflow using machine learning. Machine learning algorithms can analyze the input plans drawing and verify their design against ECM rules for compliance. Such an augmentation could reduce the overall time necessary to spend on verifying these plans by handling the more standardized cases and only requesting feedback from PUB officers for cases with higher complexity.
Scale implementation with speed and accuracy
Over a span of four months, the team built a corresponding end-to-end prototype incorporating:
- Multiple models of computer vision (object detection and instance segmentation) trained on past ECM submissions to recognise key design elements
- Optical character recognition to extract relevant text
- Rule engine to check compliance based on the recognitions and a prototypical user interface to interact with PUB officers.
Designed for sustainable growth
Thorough analysis of the algorithms’ performance has identified the working scenarios and their shortcomings. The prototype is deployed on Amazon Web Services (AWS), allowing to further probe the models with new data. In combination, these findings allow to adjust the business case and steer the submission format towards desired standardisation levels.
More about the Automated Earth Control Measure Submission Evaluator innovation challenge.