From diver-based to data-driven maintenance of the locks at Port of Antwerp-Bruges

A maritime powerhouse with a commitment to data

Port of Antwerp-Bruges is a crucial economic hub, supporting approximately 164,000 jobs and contributing nearly 21 billion euros in added value. Effective management of docks, bridges, locks, quay walls, roads, tugboats, service buildings, and land is essential to maintain smooth operations. This involves constant monitoring and maintenance to ensure uninterrupted maritime traffic on the River Scheldt.

Since 2015, the Port of Antwerp-Bruges has been dedicated to collecting vast amounts of real-time data from various assets, both above and below the water surface. The Advanced Port Information & Control Assistant (APICA) plays a crucial role in this data collection. "APICA gives us a live image of our port operations, invaluable for operational decision-making. Our goal was to extend this capability to our maintenance practices,” says Matthias Lootens, Asset Management Implementation Manager.

“Inspecting the condition of lock gates has always been a challenge, relying on specialized divers, while preventive maintenance can lead to unnecessary downtime and premature component replacements. So we were wondering whether we could leverage the data we amassed already in order to optimize our maintenance schedules.”

Bringing AI into the equation

Recognizing the potential of data to revolutionize their maintenance strategies, the port partnered up with Yitch, a specialist in data-driven solutions. Peter Mees, Manager Data & Intelligence at Yitch, describes their approach: "Our expertise lies in translating raw data into actionable predictions, organizing it with the ISA-95 standard, and utilizing it to train machine learning models. The port had already invested in comprehensive data collection. Our job was to make sense of this data to predict maintenance needs.”

The most promising data for this purpose was the power consumption of the drives controlling the lock gate motors. “When there’s a lot of sludge on the tracks over which the carriages of the lock gates move, this translates into significantly higher and less homogeneous power consumption, especially during the final closing movement of the lock gate,” explains Mees. By analyzing these power consumption patterns, Yitch aimed to predict maintenance needs, thus reducing the necessity of diver inspections as well as the cost of unneeded preventive maintenance.

Three objectives, three solutions

The initial use case of utilizing current flows to predict maintenance needs at a lock gate expanded to using three different AI machine learning principles. This allowed the prediction of maintenance needs not just for one, but across numerous lock gates simultaneously, while also identifying other anomalies at an early stage without the model being specifically trained for them.

1. Predicting the maintenance need of a single lock door, using Supervised Machinelearning (Classification)
   
   Supervised machine learning was employed to analyze and predict maintenance needs based on electrical consumption data from a single lock door. This involved collecting detailed current profiles, resulting in 300 labeled datasets indicating normal (OK) or abnormal (NOK) operations. "We meticulously labeled these profiles, identifying key features such as slope, standard deviation, minimum, maximum, and average current values," says Lootens.
   
   A supervised machine learning algorithm was then trained with this labeled dataset to develop a classification model. This model could predict whether a new current profile indicated normal or abnormal behavior. The accuracy of this model was rigorously evaluated to ensure reliable predictions.
    
2. Predicting the maintenance need of numerous lock doors, using Unsupervised Learning (Clustering)
   
   To extend predictive maintenance capabilities across multiple lock doors, unsupervised machine learning techniques, particularly clustering, were utilized. Unlike supervised learning, clustering does not rely on pre-labeled data. Instead, it enabled the identification of common patterns and deviations across all lock doors without requiring individual labeling. By analyzing these clusters, it became possible to detect maintenance needs and operational anomalies throughout the entire lock system.
    
3. Detecting / triggering other anomalies for which the model was not trained (Golden Batch)
   
   The port also needed to detect issues beyond sludge buildup, such as a lock door’s undercarriage derailing. Initial models, trained solely on historical data, could not recognize such anomalies. The solution was to create a Golden Batch: a reference dataset representing optimal operational conditions. Data were re-clustered into four clusters, each labeled as OK or NOK. For each cluster, the average and standard deviation of the current profiles were calculated for every time index.
   
   This was visualized with a line representing the average and a deadband set at three times the standard deviation, encompassing 99.7% of the data. "New current profiles were compared against the Golden Batch to detect significant deviations, resulting in a score that highlighted potential maintenance needs," explains Mees. "Large, sustained deviations indicated severe issues, such as an undercarriage derailing. This approach allowed the creation of triggers and alarms for previously unseen anomalies, enhancing the system’s robustness and reliability."

Conclusion

By integrating these machine learning techniques - Supervised ML using classification for individual lock door analysis, Unsupervised ML through clustering for system-wide maintenance prediction, and the Golden Batch for previously unseen anomaly detection - Yitch developed a sophisticated and comprehensive predictive maintenance system for Port of Antwerp-Bruges. "Implementing AI-driven maintenance has been a game-changer for us," says Lootens. "We no longer react to problems; we anticipate and prevent them. This has made our operations more efficient and reliable."