Data-driven cyber-physical networked systems for autonomous cognitive control and adaptive learning in industrial & urban water environments

Water infrastructures in both urban and industrial environments face new challenges that prompt novel data processing and control systems capable of ensuring certain water parameters and cost-efficient solutions. For example, water distribution networks need to optimize pumping operations to minimize energy costs and are vulnerable to accidental contamination spills, which can compromise the lives of inhabitants. Aquaponics industrial plants, which combine aquaculture with hydroponics to symbiotically raise plants and fish, constitute another example, which promise the generation of sustainable food if water and climate parameters are intelligently controlled. Unfortunately, state-of-the-art tools in engineering are unable to meet these demands. INDURB introduces a cohesive and multidisciplinary approach that leverages recent advances in data science, optimization and machine learning to develop smart sensing and control mechanisms for water infrastructures. Key tasks of our system include a) reliable control of water parameters and optimization of variables in smart water networks; b) improved decision making and future planning for service operations; and c) interaction with the end-users. From a scientific perspective, we have developed algorithms for data preprocessing for time alignment of time series, learning dependency data models from the observed time-series, learning normal operational models and predicting and detecting events as well as the design of deep reinforcement learning based control methods. We have also implemented the overall system and network architecture for incremental online data acquisition, analysis, and related machine learning. Examples of impact in applications are: a) control of pumps in water distribution network in order to satisfy the user water demand and maintain a balance between the pressures, decreasing also the costs of maintenance, b) estimation of flow data values in the water network in locations where there are no sensors or there are failures, c) maximizing plant growth while ensuring the wellbeing of the fishes in Aquaponics plants. This work has been done in collaboration with partners NIVA, Feedback Aquaculture and Kristiansand Municipality.

- Design of overall system architecture for online data acquisition of Water Distribution Networks, including the necessary protocols for real-time analysis of data, with an online incremental data acquisition. In the case of Aquaponics, we have also developed a data pipeline involving data pre-processing, dimensionality reduction, feature selection, and prediction capabilities. - Design of online learning of causal (spatial and temporal) dependencies from multiple streaming data time-series, which are dynamic and non-stationary. This includes online kernel-based learning algorithms for both linear and non-linear vector autoregressive models, where in the latter case, the high dimensionality of the inference problem is tackled by using a Fourier-based random feature approximation. These data-driven model learning algorithms have been tested for data from Water Networks, allowing to perform both prediction of dynamics and data imputation. - For scenarios where the data samples from different sensors are not aligned in time, we have designed zero-delay reconstruction of correlated multivariate sensor time series under certain smoothness constraints, associated to the water network data. We learn about the space of signals of interest and optimise the acquisition process at hand to improve the accuracy of the reconstruction under zero-delay processing requirements. This allows resampling and aligning temporal samples simultaneously across multiple time series. - Design of an estimation algorithm to approximate the posterior distribution between states and observations for complex dynamic systems, such as the water quality variables in the Aquaponics systems, and considering also an stochastic risk measure based on the variance of the estimated values, which is also part of the optimisation criterion. - Design of a framework for data-driven adaptive control based on safe deep reinforcement to manage the variables in safety-critical environments, e.g. Aquaponics (maximise plant growth and quality while keeping a healthy fish), providing partial information about potential risks to enhance the training for the control agents. - Design of graph filtering techniques for different inference tasks, such as least squares estimation, denoising, consensus, and change-point or anomaly detection. First, the design of distributed robust graph filters against quantization noise and stochastic errors; second, the design of graph shift operators to maximise the efficiency of the associated inference tasks; third, the joint design of network protocols and graph filters for distributed implementation over a wireless sensor network. - Design of reinforcement learning algorithms for data-driven cost-efficient pump control in water distribution networks. In addition to minimising the energy costs, we have also considered balancing the pressures across the water network. This leads to minimising water losses due to leakage events and b) maintenance costs.

The main objective of INDURB is the design and implementation of a data-driven networked cyber-physical system (CPS) for autonomous control and continuous learning from data in both industrial and urban water environments, which will be represented by Aquaponics industrial plants and urban Water Distribution Networks (WDN). These networked systems are composed of a large number of wirelessly interconnected components with local data processing and control capabilities over large geographic areas or with high spatial densities. The malfunctioning of one component, either due to accidents or intentional attacks, may have devastating effects. Unfortunately, currently existing scientific and engineering methods do not consider smart sensing and data analytics components, and proper integration methods to offer timely detection, cognitive control and adaptive learning and therefore, are very conservative and highly sub-optimal. INDURB offers a new complete multidisciplinary approach, providing the following features: a) a highly reliable and low-cost water health protection with respect to both chemical and microbiological contamination, predicting and reacting through actuation (chemical dosage, smart valves and pumps), ensuring that the water parameters are within the correct limits adapting to the application demands; b) improved decision making and future planning for service operations; and c) a real-time bidirectional interaction between our CPS and the end-users (water utilities, food and water consumers), which will also provide data into the system, interplaying with cost-benefit trade-offs and allowing an automatic end-user complaint surveillance to address water quality problems and improve satisfaction. The results will be demonstrated for both scenarios, by means of a real WDN pilot and a real Aquaponics fish factory, in cooperation with partners NIVA (world wide leader in water technology), Municipality of Kristiansand (smart city objective) and FAqua.