Neural network solder paste inspection system

• US 5,751,910 A
• Filed: 05/22/1995
• Issued: 05/12/1998
• Est. Priority Date: 05/22/1995
• Status: Expired due to Fees

• Current Assignee: Eastman Kodak Company
• Original Assignee: Eastman Kodak Company
• Inventors: Loewenthal, Kenneth H., Bryant, Steven M.
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/447,282
• Time in Patent Office: 1,086 Days
• Field of Search: 395/22, 395/23, 395/3, 395/61, 395/900, 382/145, 382/147, 382/157, 382/159
• US Class Current: 706/2
• CPC Class Codes: G06N 3/043 based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
• CPC Class Codes: Y10S 706/912 Manufacturing or machine, e.g. agricultural machinery, machine tool

Abstract

A solder paste brick inspection and physical quality scoring system 10 employs a neural network 70 trained with a fuzzified output vector. An image of solder paste bricks 64 on a printed circuit board 12 is acquired by a CCD camera 30. Values of a predetermined set of brick metrics are extracted from the image by a computer 28 and used as a crisp input vector to trained neural network 70. A defuzzifier 76 converts a fuzzy output vector from neural network 70 into a crisp quality score output which can be used for monitoring and process control.

Claims

1. A training a neural network, comprising the steps of:

providing a first training data base of crisp input and crisp output pairs;

fuzzifying crisp outputs of said first training data base to create a second training data base of crisp input and fuzzified output pairs; and

training the neural network with said crisp input and fuzzified output pairs to solution conversions, wherein;

said crisp outputs are fuzzified in accordance with a predetermined nearness relationship between different crisp outputs, the neural network is employed for classification of objects, the crisp output of each crisp input and crisp output pair is representative of an image of a particular object, and the crisp output of each such crisp input and crisp output pair represents a desired classification of the particular object, and the objects comprise solder paste bricks on at least one printed circuit board, the crisp input of each crisp input and crisp output pair represents a machine vision image of a particular solder paste brick, and the crisp output of each such crisp input and crisp output pair represents an operator generated physical quality score for the particular solder paste brick.

2. The method of claim 1 further comprising the step of extracting values for a predetermined set of metrics from the machine vision image of a particular solder paste brick and using said extracted values as the crisp input.

3. A method of determining a quality score of a test solder paste brick with a neural network trained in accordance with the method of claim 2, comprising the steps of:

acquiring a machine vision image of the test solder paste brick;

extracting values for the predetermined set of metrics for the test solder paste brick from the acquired machine vision image of the test solder paste brick;

applying the extracted values for the test solder paste brick as an input to the trained neural network to produce a fuzzy network output vector; and

defuzzifying the fuzzy network output vector to provide a quality score for the test solder paste brick.

4. A method for training a neural network to automatically provide a quality score of the physical quality of a test solder paste brick, comprising the steps of:

acquiring an image of each brick of a training set of solder paste bricks;

extracting values for a predetermined set of metrics from the acquired image for each brick of the training set;

defining a desired quality score for each brick of the training set;

fuzzifying the desired quality score for each brick of the training set into a fuzzified quality score vector for each brick in accordance with a predetermined nearness relationship between different quality scores; and

applying the extracted values as network inputs and the fuzzified quality score vector as a fuzzy network output vector for each brick of the training set to the neural network in training the neural network to solution convergence.

5. A method of determining a quality score of a test solder paste brick with a neural network trained in accordance with the method of claim 4, comprising the steps of:

acquiring an image of the test solder paste brick;

extracting values for the predetermined set of metrics for the test solder paste brick from the acquired image of the test solder paste brick;

applying the extracted values for the test solder paste brick as inputs to the trained neural network to produce a fuzzy network output vector; and

defuzzifying the fuzzy network output vector to provide a quality score for the test solder paste brick.

6. A method for automatically determining a quality score for a test solder paste brick, comprising the steps of:

acquiring an image of the test solder paste brick;

extracting values for a predetermined set of metrics from the acquired image;

applying the extracted values as inputs to a neural network trained using a fuzzified network output vector to produce a fuzzy network output vector for the test solder paste brick; and

defuzzifying the fuzzy network output vector to provide a quality score for the test solder paste brick.

7. The method of claim 6 wherein the image of the test solder paste brick is acquired using off axis illumination, dark field reflection and cross polarization.

8. The method of claim 7 applied to multiple test solder paste bricks, and further comprising the step of graphically displaying quality scores for the multiple test solder paste bricks to provide a quality trend indication.

9. The method of claim 8 applied to real time inspection of the physical quality of solder paste bricks for fine-pitch surface mount components in a printed circuit board assembly line, and further comprising the step of employing the quality trend indication as real time feedback for solder paste brick quality control.

10. The method of claim 6 wherein the predetermined set of metrics includes brick width, brick length, brick area, longitudinal symmetry, mean intensity, brick edge raggedness and brick corner roundness.

11. Apparatus for inspecting fine-pitch surface mount solder paste bricks on printed circuit boards and automatically providing a quality score of the physical quality of the bricks, comprising:

a camera for acquiring an image of a solder paste brick being inspected;

a computer for extracting values of a predetermined set of brick metrics from the acquired image; and

a neural network hosted by the computer and trained with a fuzzy network output vector, said neural network having the extracted values applied as inputs to the neural network and outputting a fuzzy quality score for the brick being inspected.

12. The apparatus of claim 11 further comprising a defuzzifier that defuzzifies the fuzzy quality score outputted by the trained neural network to provide a crisp quality score for the brick being inspected.

13. The apparatus of claim 12 further comprising an off-axis source of polarized light for illuminating the brick being inspected, and a cross polarizer at an input of said camera.

14. The apparatus of claim 13 further comprising means for controllably positioning the camera relative to a printed circuit board such that the camera acquires an image of a particular solder paste brick desired to be inspected.

15. The apparatus of claim 14 wherein the camera comprises a CCD camera, the source of light comprises a halogen light source, and the means for controllably positioning comprises an X-Y linear motor translation stage.