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.
