Session: 05-15-03: General Topics in Biomedical and Biotechnology - III

Paper Number: 98982

98982 - Implementing and Evaluating a Cost-Effective Optical Biosensor With Integrated Artificial Intelligence and Machine Learning 

For this project, the concentration and color of a sample of water were determined using machine learning and regression principles paired with an optical biosensor that has the ability to detect the amount of red, blue, and green pigment in a sample. The setup of the optical biosensor consisted of a small, 3D-printed cavity. A three-color LED, a photoresistor sensor, and a microcontroller are integrated into the cavity that collected the necessary data to train the neural network by momentarily flashing red, blue, and green light into the sample cuvette thereby allowing the color pigments to be obtained as data. The concentration of each is then displayed on the interface of the circuit made to perform this function. Data on the amount of red, blue, and green values, which indicate concentration of the pigment color, can then be recorded. Training data consisted of six different classes of colors: purple, blue, green, red, orange, and transparent (no color) each with concentration intervals of sixteen percent, excluding samples with no color. Extensive training data was taken by measuring the color and concentration of each sample thirty times. Data is then transferred to a computer using a micro-USB cord where it can be easily deciphered and prepared for training. Using the data, the color of the sample and its concentration are successfully detected using a two-stage AI/ML architecture. First, a multi-class classification neural network is trained to detect the color of the sample. Many trials are recorded and evaluated with varying number of layers and neuron density per layer. After finding a satisfactory fit to the training data and thus successfully predicting the color of the sample, an appropriately trained regression model is selected to determine the concentration of the sample from the red-blue-green values. The most effective model is found to be linear regression. Keras and Tensorflow are used to train and evaluate said networks. Various combinations of different parameters such as activation function, layer number, and neuron density of the neural networks are combined and evaluated to determine the most effective combination. The machine learning models were then uploaded to an ESP32 microcontroller to perform evaluation of color and concentration directly from the physical setup. Performance is then evaluated after uploading the models. Ideas explored in this project raise the possibility of machines more efficiently predicting other traits and concentrations of a fluid using machine learning and artificial intelligence with an optical biosensor.  

Presenting Author: Ethan Regal Gannon University

Presenting Author Biography: Fourth year undergraduate pursuing a Bachelor of Science in Mechanical Engineering. I have worked as a research assistant for over a year in the field of machine learning. I have also began work in thermal science research recently. Other work includes working part time as a tutor for various STEM courses and as a teaching assistant. I represent my university as the president of our ASME chapter and as a member of the Pi Tau Sigma mechanical engineering honors society. Current interests include but are not limited to; deep learning, CFD, and heat transfer.

Authors:

Ethan Regal Gannon University
Pezhman Hassanpour Gannon University