Supervised Machine Learning using Convolutional Neural Networks
Location: Coralville, Iowa 52241, USA
Industry: Manufacturing, Life Sciences, Medical
Business Functions:
- Engineering
- Quality Assurance
- Manufacturing
Problems
The client is a manufacturer of customized DNA Oligos. One step of the process in this manufacturing is to fill up different plates with DNA or RNA membranes. The issue was
- Manual analysis of the filled DNA or RNA plates. These different plates contain many wells which need to be filled with different membranes. The robots being used here often miss certain wells and the entire plate needs to be refilled.
- The optimization required automating the analysis part so we can identify missing wells using machine learning.
Solutions
Allied provided solution by:
- Using machine learning, specifically, Deep Neural Networks; developed a reliable inference model which improved the client’s engineering process. The existing data that we had on the plates was engineered and then used to train a logistic regression model.
- Working on feature engineering, regularization, and data entry to produce a reliable model to automatically classify the filled plates as either empty, non-empty or tilted.
- Creating custom labeling application for maintaining the assets with the trained model.
Benefits / Results:
- 99.98% accuracy in categorizing new images
- Improved client’s manufacturing process
Tech Stack:
ReactJS
A front-end application was developed to train the images. A single image contained several different pieces of information which needed to be identified and isolated. The react application helped to standardize the process.
.net Core
An API was developed which made inferences using the ONNX model, developed using Python and TensorFlow/Keras. Several front-end components were also developed inside Microsoft Blazor Server tool which were later refactored into react.
TensorFlow/Keras
EfficieNet-B0 algorithm was used to train the image data. Batching was implemented due to the large size of the input dataset. Several models were trained and evaluated during this process of training.
Get In Touch With Us
we believe in cutting edge solutions and are committed to your success