The development process for a neural network requires problem understanding, proof-of-concept testing, pre-production training, pre-production testing and finally, production acceptance testing.
- To understand the problem, a part description and list of defect causes and appearances are developed.
- To perform proof-of-concept testing a few examples of ‘good’ parts and parts with each defect are obtained. These examples should incorporate the range of ‘good’ and ‘defective’ parts.
- Proof-of-concept testing involves acquiring images of the example parts. These images must enable a person to correctly categorize the parts.
- Part delivery mechanisms (conveyors, flippers, pneumatic actuators, etc.) are developed and built.
- A larger number of example parts are provided. Ideally, these are obtained from different batches to capture the full range of process variations.
- This larger set of parts is used to obtain training images and train the neural network.
- Pre-production testing is performed with new parts. Network re-training may be required.
- Production acceptance testing is performed on freshly produced parts.