Image analysis system for testing in manufacturing
A vision analytics and validation (VAV) system for providing an improved inspection of robotic assembly, the VAV system comprising a trained neural network three-way classifier, to classify each component as good, bad, or do not know, and an operator station configured to enable an operator to review an output of the trained neural network, and to determine whether a board including one or more “bad” or a “do not know” classified components passes review and is classified as good, or fails review and is classified as bad. In one embodiment, a retraining trigger to utilize the output of the operator station to train the trained neural network, based on the determination received from the operator station.
1. A vision analytics and validation (VAV) system for providing an improved inspection of robotic assembly, the VAV system designed to quickly adapt to new defects, new lighting conditions, and new parts without requiring tuning, the VAV system comprising:
a processor implementing a trained neural network three-way classifier, to classify each component as good, bad, or do not know;
an operator station configured to enable a human operator to review an output of the trained neural network three-way classifier, and to determine whether a board including one or more “bad” or a “do not know” classified components pass review and is classified as good, or fails review and is classified as bad;
a clustering system to implement a clustering model wherein a cluster of data defining “good” components is build using statistical features of the “good” components, and wherein an object which is not within the cluster of data is classified as bad;
a synthetic training data generator to create synthetic training data based on the cluster of data, representing elements within the cluster of data that do not occur in training data.
2. The VAV system of claim 1 , further comprising:
an image combiner to combine a plurality of images from one or more camera-based inspection machines, the image combiner to create a unified image;
a grid and ROI logic to divide the unified image into regions of interest including an element and associated pads;
an OCR and element match to identify elements on the board based on optical recognition;
a masking logic to mask extraneous information within a viewing window, prior to passing data to the trained neural network three-way classifier.
3. The VAV system of claim 1 , further comprising:
a machine learning system to train the trained neural network three-way classifier, the machine learning system training the neural network three-way classifier with data from the synthetic training data generator, the human operator, and a final inspection, enabling the VAV system to rapidly adapt to new defects, new lighting conditions, and new parts.
4. The VAV system of claim 1 , wherein the synthetic training data generator uses CAD files to create synthetic images of “good” and “bad” components with tags.
5. The VAV system of claim 4 , wherein one or more of the synthetic training data comprises only isolated subparts of the components.
6. The VAV system of claim 5 , wherein the isolated subparts comprise solder balls isolated from other parts of a circuit element.
7. The VAV system of claim 1 , further comprising:
the synthetic training data generator use a focused subset of the training data to generate the synthetic training data, the focused subset comprising the training data with portions that not relevant to the inspection masked out.
8. The VAV system of claim 7 , further comprising:
a masking logic to utilize CAD data about the component to match elements on a combined image and mask the portions that are not relevant to the inspection.
9. A system for providing an improved inspection comprising:
a processor implementing a trained neural network three-way classifier, to classify each component as good, bad, or do not know;
an operator station to enable an operator to verify an output of the trained neural network three-way classifier and to verify a classification of at least a subset of components;
a clustering system to implement a clustering model wherein a cluster of data defining “good” components is build using statistical features of the “good” components, and wherein an object which is not within the cluster of data is classified as bad;
a synthetic training data generator to create synthetic training data based on the cluster of data, representing elements within the cluster of data that do not occur in training data.
10. The system of claim 9 , further comprising:
the operator station configured to enable a human operator to review the output of the trained neural network, and to verify whether a board including one or more “bad” or a “do not know” classified components passes review and is classified as good or fails review and is classified as bad.
11. The system of claim 9 , wherein the synthetic training data generator uses CAD files to create synthetic images of “good” and “bad” components with tags.
12. The system of claim 9 , wherein one or more of the synthetic training data comprises only isolated subparts of the components.
13. The system of claim 12 , wherein the isolated subparts comprise solder balls isolated from other parts of a circuit element.
14. The system of claim 9 , further comprising:
the synthetic training data generator use a focused subset of the training data to generate the synthetic training data, the focused subset comprising the training data with portions that not relevant to the inspection masked out.
15. The system of claim 14 , wherein the masking out utilizes optical character recognition to identify and mask out writing.
16. The system of claim 14 , further comprising:
a masking logic to utilize CAD data about a component to match elements on a combined image and mask the portions that are not relevant to the inspection.
17. A method for providing an improved inspection of robotic assembly comprising:
classifying each component as good, bad, or do not know, using a processor implementing a trained neural network three-way classifier;
providing an output the trained neural network three-way classifier to an operator station to verify a classification of at least a subset of components;
clustering data to build up a cluster of data defining “good” components using statistical features of the “good” components, and wherein an object which is not within the cluster of data is classified as bad;
creating synthetic training data based on the cluster of data, representing elements within the cluster of data that do not occur in training data.
18. The method of claim 17 , further comprising:
enabling a human operator to review the output of the trained neural network three-way classifier and to verify whether a board including one or more “bad” or “do not know” classified components passes review and is classified as good, or fails review and is classified as bad at the operator station.
19. The method of claim 17 , wherein the synthetic training data is generated using CAD files to create synthetic images of “good” and “bad” components with tags.
20. The method of claim 18 , wherein one or more of the synthetic training data comprises only isolated subparts of the components.