AI-Based Sensor Anomaly Detection for Collection System Sensor Networks

Introduction As utilities are installing more sensors to monitor their systems and applying digital twins with dynamic data to gain insights, sensor data quality becomes key to drive actionable and improved outcomes. The Louisville and Jefferson County Metropolitan Sewer District in Kentucky (Louisville MSD) is operating and maintaining a complex sewer network which consists of more than 3,200 miles of sanitary and combined sewers and 5 regional wastewater treatment plants. To monitor and operate this complex network, more than 300 sensors are installed in the collection system. The system is also equipped with a predictive real-time control (RTC) system which modulates gates and pumps during wet weather based on real time sensor information to optimally minimize combined sewer overflows (CSOs). The operational performances of the sewer system and the quality of the data available for engineering design are largely dependant on the accuracy and the reliability of the sensors. In a hostile environment such as sewers, sensors fail for many reasons (Campisano et al., 2013) and failure may lead to sewer overflows or sewer backup that could have been otherwise avoided. To improve data quality and availability, and to reduce sensor maintenance costs, Louisville MSD, as part of its predictive maintenance program, is seeking innovative solutions to quickly detect faulty sensors. An algorithm to detect faulty sensors in locations with redundant level sensors was implemented with success in the recent past (Pleau et al., 2023). This article presents a new tool being implemented, based on artificial intelligence (AI) to detect anomalies in all the sensors of a given facility, even if they are not redundant. Methodology The developed sensor anomaly detection methodology consists of two steps. In the first step, basic anomaly detection algorithms are used to flag extreme values, periods of time where a sensor has a constant value (if it applies to the sensor behavior), abrupt changes in the behavior (if it applies to the sensor), etc. The second step of the anomaly detection consists of using an AI model to estimate the expected values for each sensor across time based on the observed conditions measured through the surrounding sensors. From a set of rules, notably by using the error between the AI model expected value and the observed value, it is determined whether the sensor is in fault. AI models used are deep learning models built around the Transformer architecture (Vaswani et al. 2017). The models were trained using available historical data where some sensors had less than 1 year of historic, and others had more than 3 years, each with a time step of 5 minutes. The AI sensor anomaly detection tool is integrated through a web platform. A schedule task collects data from the SCADA system and its data historian and runs the anomaly detection algorithms. Dashboards summarize the information to easily identify which sensors are abnormal and charts allow for a review of the anomalies detected by operators to confirm the need for a maintenance on the sensor. Results and Benefits The trained AI model was able to reproduce the normal behavior of the sensors with good accuracy. Anomalies have been detected by the procedure on historical data never seen by the model in training and validation. Figures 2 and 3 compares measurements with AI model estimates for two level sensors of the Main Diversion Structure, one of MSD's network facilities. Figure 1 presents a schematic of the Main Diversion Structure with the sensor locations. Anomalies that can be detected where the behavior of the sensor deviate significantly from the behavior expected by the model. Based on the encouraging results, the sensor fault detection algorithm is being implemented during winter 2024. The post-implementation operational data will be shared at WEFTEC. Among other benefits, the model architecture is meant to be used on multiple control facilities. Hence, we can use the same model structure and training procedure on other system sites, saving a lot of times when another site is added to the system maintenance procedures. The developed model can also be trained on historical data where behavior can change across time. Most other deep learning architectures require consistent behavior through the training data set and the implementation. This gives the model the flexibility to evolve across time as sensors behaviors can change because of structural modifications. The demonstration study using a full year of measurements shows that the proposed sensor fault detection algorithm is effective and reliable. Instead of regular routine maintenance of all sensors, MSD now expects to use in part this data-driven algorithm as part of overall predictive sensor maintenance strategy, to detect drifting sensor before sensor failure occurs and generate a work order for maintenance only when needed. This innovative algorithm is also easy to implement. It does not require an estimation of the measured process variable and can be configured to detect in real time a large set of anomalies including incipient failures. The algorithm contributes to significantly reduce the amount of poor-quality data used in the decision-making process for real time operation. Its ability to detect incipient failures is also a key feature of the predictive maintenance program aimed at reducing operation and maintenance costs.