A low-cost methodology based on artificial intelligence for contamination detection in microalgae production systems

This work presents the development of a neural network for the classification of microalgae genera that allows the detection of contamination in the analysed sample by focusing on the output of the softmax layer. The data necessary to perform the classification are obtained by means of a spectral sweep in the 300-750 nm range, to which a baseline correction is applied in order to obtain a cleaner spectrum without unwanted fluctuations, facilitating the identification and analysis of the relevant spectral components. The developed model has been trained based on the data obtained from pure samples of 4 different genera of microalgae, Spirulina, Chlorella, Synechococcus and Scenedesmus, obtaining as a result a network that allows classification correctly and with high accuracy rate the genera of the pure samples introduced. The network was initially validated with new samples obtained in the laboratory after training obtaining a macro F1 score of 98.64 %. Afterwards, the resulting model was tested in three different photobioreactors, two raceway reactors, and one tubular reactor, to demonstrate the capabilities of the proposed methodology for contamination detection purposes. This makes it possible to monitor cultures without the need to invest in new acquisition devices or the time of qualified operators for technical monitoring. Moreover, this solution highly contributes to the maintenance of the reactor operation detecting possible contamination in advance.