Data-Driven Blockage Detection in Sewer Collection Systems

Introduction The efficient operation of sewer collection systems is crucial for maintaining public health and environmental quality. Blockages in these systems can lead to severe consequences, including sanitary sewer overflows (SSOs), property damage, and increased maintenance costs. These incidents can disrupt communities, pose health risks, and strain municipal resources. Therefore, early detection and accurate identification of blockages are essential for preventing these adverse effects and ensuring the smooth operation of sewer infrastructure. Traditional methods of detecting blockages often rely on reactive approaches, such as responding to complaints or visible overflows. However, these methods are not sufficient for proactive maintenance and prevention. With advances in technology, new approaches integrating machine learning and data analytics have emerged, offering significant improvements in the accuracy and timeliness of blockage detection. Machine learning, in particular, has shown great promise in the field of blockage detection. By analyzing large datasets from various sources, including sensor data, maintenance records, and weather patterns, machine learning algorithms can identify patterns and anomalies that indicate potential blockages. This predictive capability allows for early intervention, reducing the likelihood of SSOs, flooding and minimizing the impact on communities. Typical blockage detection approaches (using anomaly detection methods or forecasting methods) are point-based detection approaches. However, a blockage forms and evolves over time. Furthermore, traditional approaches don't take into account factors such as blockages evolving during a rainfall-event and the effect that can have on the return back to baseline levels. This study aims to present a data-driven approach to detecting blockages in real-time through a time-series data annotation method called matrix profiling and state-based monitoring approach to triggering alerts. Compared to neural network based forecasting approaches or autoencoders, matrix profiling is easier and faster to train and requires less historical data. A state-based triggering approach allows us to monitor the evolution of a blockage, particularly during wet-weather, and trigger alerts based on a probabilistic approach. Methods Time-series sensor data for depth and rainfall were collected in New Bedford, MA. The city's sewers consists of 254 miles of pipes, 72 CSO regulators, and 27 CSO outfalls. Data was gathered from 94 locations for depth and 3 rain gauges for rainfall (Figure 1). Maintenance data, including SSOs and blockages, was also used to validate model predictions. Matrix profiling using time-series motifs was used to detect blockages in real-time while a state-based triggering method was used to detect changes due to rainfall and monitor evolution of a blockage to trigger alerts. This will allow the model to shift from a sleep state to a monitoring state (change in levels indicating a potential blockage) to a trigger state (blockage detected) depending on the evolution of the data in real-time (Figure 2). Results & Discussion The blockage detection model uses a distance-based matrix profiling approach to detect blockages. Distances are calculated against 'motifs' which are blockage patterns generated through a synthetic blockage generator which allows us to generate blockage data with varying characteristics. Blockages are, by definition, relatively rare occurrences in a sewer utility. The motifs allow us to generate synthetic data that resembles blockages but also allows us to add actual blockage data as motifs. Figure 2 shows an example detection of blockage using the matric profiling approach. Even accounting for rainfall-based changes in depth, the model is able to detect a blockage within 40 minutes of the blockage developing. The full presentation will outline details of the models along with more detailed results along with details in the state-based event detection and classification methods. We will also show results for validation against actual detected blockages and performance, in terms of time to detection and accuracy of detection. The complete work is to be completed by February 2025.