In December 2016, the wastewater treatment plant of Baarle-Nassau, Netherlands, failed. The failure was caused by the illegal disposal of high volumes of acidic waste into the sewer network. Repairs cost between 80,000 and 100,000 EUR. A continuous monitoring system of a utility network such as this one would help to determine the causes of such pollution and could mitigate or reduce the impact of these kinds of events in the future. We have designed and tested a data fusion system that transforms the time-series of sensor measurements into an array of source-localized discharge events. The data fusion system performs this transformation as follows. First, the time-series of sensor measurements are resampled and converted to sensor observations in a unified discrete time domain. Second, sensor observations are mapped to pollutant detections that indicate the amount of specific pollutants according to a priori knowledge. Third, pollutant detections are used for inferring the propagation of the discharged pollutant downstream of the sewage network to account for missing sensor observations. Fourth, pollutant detections and inferred sensor observations are clustered to form tracks. Finally, tracks are processed and propagated upstream to form the final list of probable events. A set of experiments was performed using a modified variant of the EPANET Example Network 2. Results of our experiments show that the proposed system can narrow down the source of pollution to seven or fewer nodes, depending on the number of sensors, while processing approximately 100 sensor observations per second. Having considered the results, such a system could provide meaningful information about pollution events in utility networks.

Illegal discharges of pollutants into sewage networks are a growing problem in large European cities. Such events often require restarting wastewater treatment plants, which cost up to a hundred thousand Euros. A system for localization and quantification of pollutants in utility networks could discourage such behavior and indicate a culprit if it happens. We propose an enhanced algorithm for multisensor data fusion for the detection, localization, and quantification of pollutants in wastewater networks. The algorithm processes data from multiple heterogeneous sensors in real-time, producing current estimates of network state and alarms if one or many sensors detect pollutants. Our algorithm models the network as a directed acyclic graph, uses adaptive peak detection, estimates the amount of specific compounds, and tracks the pollutant using a Kalman filter. We performed numerical experiments for several real and artificial sewage networks, and measured the quality of discharge event reconstruction. We report the correctness and performance of our system. We also propose a method to assess the importance of specific sensor locations. The experiments show that the algorithm’s success rate is equal to sensor coverage of the network. Moreover, the median distance between nodes pointed out by the fusion algorithm and nodes where the discharge was introduced equals zero when more than half of the network nodes contain sensors. The system can process around 5000 measurements per second, using 1 MiB of memory per 4600 measurements plus a constant of 97 MiB, and it can process 20 tracks per second, using 1.3 MiB of memory per 100 tracks.

Deep neural networks (DNN) have outstanding performance in various applications. Despite numerous efforts of the research community, out-of-distribution (OOD) samples remain a significant limitation of DNN classifiers. The ability to identify previously unseen inputs as novel is crucial in safety-critical applications such as self-driving cars, unmanned aerial vehicles, and robots. Existing approaches to detect OOD samples treat a DNN as a black box and evaluate the confidence score of the output predictions. Unfortunately, this method frequently fails, because DNNs are not trained to reduce their confidence for OOD inputs. In this work, we introduce a novel method for OOD detection. Our method is motivated by theoretical analysis of neuron activation patterns (NAP) in ReLU-based architectures. The proposed method does not introduce a high computational overhead due to the binary representation of the activation patterns extracted from convolutional layers. The extensive empirical evaluation proves its high performance on various DNN architectures and seven image datasets.