Periodic Reporting for period 1 - ProbSenS (Probabilistic neuromorphic architecture for real-time Sensor fusion applied to Smart, water quality monitoring systems)
Período documentado: 2017-09-01 hasta 2019-08-31
The ProbSenS project aimed to develop a novel low-power event-driven probabilistic Very Large-Scale Integration (VLSI) architecture for real-time, adaptive and robust multisensor integration. Current commercial instrumentation is only able to analyze a small fraction of the markers targeted in everyday applications, and is generally bulky, monoparametric, and requires periodical calibration in front of environmental changes and sensor non-idealities. Most analyses are still performed in laboratories, resulting in increased costs, hindered logistics, and delayed detection. By using principles of how biological systems rapidly combine multisensory information and generate meaningful features in dynamic and uncontrolled real-world conditions, the neuromorphic networks designed in the project are connected to a set of Si-based microsensors to fuse multivariate data in situ for a true EU societal challenge: the real-time monitoring of water pollutants. The resulting system is targeted to be faster, smaller, energy efficient, and highly resilient to noise, nonlinearities, matrix effects, and drifts associated with the sensors.
During the first phase of the project, we established the architecture to fuse the readout from electrochemical sensors fabricated in mature solid-state technologies in the clean room of the Institute of Microelectronics of Barcelona. System specs were defined in collaboration with all partners, and multiparametric probes were installed in a water treatment plant of the industrial partner Agbar, together with their respective sensors and readout electronics. Measurements from the probes and quality analyses were collected all along the project, and incorporated into a comprehensive database. Correlations and causal relations were identified between measurements.
Once sensors, electronics and database were in place, we investigated two deep neural network (DNN) architectures, and determined conversion and learning rules to run the algorithms in the spiking domain on a programmable neuromorphic substrate. First, a convolutional neural network (CNN) architecture was realized, where temporal structure is mapped in sequentially activated cells and each sensor modality constitute an input branch. The neural network was trained offline with regular real-valued ReLU activations on the data recorded at the water treatment plant, and could be converted to the spiking domain by using the open source SNN toolbox. On-chip learning was also studied through use of local rules with surrogate gradients, where learning is modulated by local estimations of the error. The CNN was validated for predicting two measurements of interest for the final user. Besides this first approach, we also explored implementations based on recurrent neural networks (RNNs). In these networks, the dynamics of selected analyses are reproduced through projections of neural spatiotemporal representations generated by excitatory/inhibitory (E/I) activity.
In the last stage of the project we investigated a compressive spiking readout scheme to encode electrochemical sensor measurements. The chosen architecture is targeted to attenuate non-idealities such as noise, offset, and drifts. We also defined the on-site multisensory setup to validate the performance of the two DNN architectures developed in the project. Results were demonstrated in international conferences and public events (ISCAS 2019, Science is Wonderful! Exhibition 2019), and presented in periodic follow-up meetings with the industrial user and an H2020 meeting organized by the REA. An invention disclosure on this work has been submitted. The patent application was being prepared at the moment of composition of this summary.
Once sensors, electronics and database were in place, we investigated two deep neural network (DNN) architectures, and determined conversion and learning rules to run the algorithms in the spiking domain on a programmable neuromorphic substrate. First, a convolutional neural network (CNN) architecture was realized, where temporal structure is mapped in sequentially activated cells and each sensor modality constitute an input branch. The neural network was trained offline with regular real-valued ReLU activations on the data recorded at the water treatment plant, and could be converted to the spiking domain by using the open source SNN toolbox. On-chip learning was also studied through use of local rules with surrogate gradients, where learning is modulated by local estimations of the error. The CNN was validated for predicting two measurements of interest for the final user. Besides this first approach, we also explored implementations based on recurrent neural networks (RNNs). In these networks, the dynamics of selected analyses are reproduced through projections of neural spatiotemporal representations generated by excitatory/inhibitory (E/I) activity.
In the last stage of the project we investigated a compressive spiking readout scheme to encode electrochemical sensor measurements. The chosen architecture is targeted to attenuate non-idealities such as noise, offset, and drifts. We also defined the on-site multisensory setup to validate the performance of the two DNN architectures developed in the project. Results were demonstrated in international conferences and public events (ISCAS 2019, Science is Wonderful! Exhibition 2019), and presented in periodic follow-up meetings with the industrial user and an H2020 meeting organized by the REA. An invention disclosure on this work has been submitted. The patent application was being prepared at the moment of composition of this summary.
ProbSenS provided the opportunity to validate the resulting prototypes as portable instrumentation for the real-time monitoring of water supplies working hand in hand with a leading water management company, encouraging technology transfer to a core EU 2020 sector. Water quality assurance remains a global health concern, needing new multisensor solutions able to: (i) categorise running samples with high confidence under its wide, tangled and dynamic pollutant matrix; (ii) offer long-term stability in front of continuously changing environmental conditions (e.g. temperature, interference effects, sensor drifts); and (iii) deliver real-time results before hazardous situations like waste discharges and microbial contamination spread in fast-flowing water and reach human consumption. ProbSenS fits this challenge by: (i) adopting neural sensor fusion algorithms to reliably model multivariate spatiotemporal dependencies from raw input sensory data; (ii) accelerating computation in low-power neuromorphic circuits to provide in-field auto-calibration; (iii) developing asynchronous readout schemes to minimize monitoring delays; and (iv) applying the last two approaches to miniaturize the system and boost energy autonomy in view of future system integrations to access remote and narrow spots along the water distribution network.
Results from the project show promise in predicting quality analyses conducted on drinking water. A system that uses the architecture developed in the project could be applied to other domains needing immediate analysis of fluids on the spot. Examples are the continuous manufacturing of chemicals along the production chain and the non-invasive assessment of athletic performance using wearable technologies.
Results from the project show promise in predicting quality analyses conducted on drinking water. A system that uses the architecture developed in the project could be applied to other domains needing immediate analysis of fluids on the spot. Examples are the continuous manufacturing of chemicals along the production chain and the non-invasive assessment of athletic performance using wearable technologies.