Quality prognostics system and method for manufacturing processes

• US 7,493,185 B2
• Filed: 06/02/2005
• Issued: 02/17/2009
• Est. Priority Date: 06/03/2004
• Status: Active Grant

• Current Assignee: National Cheng KUNG University (Government of The Republic of China)
• Original Assignee: National Cheng KUNG University (Government of The Republic of China)
• Inventors: Huang, Guo-Wei, Hung, Min-Hsiung, Su, Yu-Chuan, Cheng, Fan-Tien
• Primary Examiner(s): Shechtman; Sean P
• Application Number:
  US11/142,316
• Publication Number:
  US 20050288812A1
• Time in Patent Office: 1,356 Days
• Field of Search: 700/28, 700/30, 700/31, 700/33, 700 48- 52, 700103-105, 700/108, 700/109, 700/121, 702 81- 84, 706/1, 706/2, 706/15, 706/62
• US Class Current: 700/108
• CPC Class Codes: G05B 19/41875 characterised by quality surveillance of production
• CPC Class Codes: G05B 2219/32194 Quality prediction
• CPC Class Codes: G05B 2219/35346 VMMC: virtual machining measuring cell simulate machining process with modeled errors, error prediction
• CPC Class Codes: Y02P 90/02 Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
• CPC Class Codes: Y02P 90/80 Management or planning

Abstract

A quality prognostics system and a quality prognostics method for predicting the product quality during manufacturing processes are disclosed, wherein the current production tool parameters sensed during the manufacturing process and several previous quality data collected from the measurement tool are utilized to predict the future product quality, and a conjecture modeling step and prediction modeling step are performed respectively. The conjecture modeling step itself also can be applied for the purpose of virtual metrology. Further, a self-searching step and a self-adjusting step are performed for searching the best combination of various parameters/functions used by the conjecture algorithm or prediction algorithm; and meeting the requirements of new equipment parameters and conjecture/prediction accuracy.

Claims

1. A quality prognostics system for manufacturing processes, comprising:

conjecture modeling means for using a set of current input data of a production tool to conjecture a conjecture value for a product lot currently manufactured in said production tool, wherein said conjecture modeling means is built in accordance with a conjecture method, and said conjecture method is selected from the group consisting of a first neural network technique, a fuzzy logic technique and a stepwise regression technique; and

prediction modeling means for using said conjecture value of said product lot together with at least two actual measurement values of at least two previous product lots to predict a prediction value for a next product lot, wherein said at least two previous product lots are respectively produced in at least two production cycles, and said at least two actual measurement values are one-to-one corresponding to said at least two previous product lots respectively, and said prediction modeling means is built in accordance with a prediction method, and said prediction method is selected from the group consisting of a weighted moving average technique and a second neural network technique.

2. The quality prognostics system of claim 1, further comprising:

raw-data pre-processing means for converting a set of raw data inputted from said production tool to a set of desired input data of a specific format.

3. The quality prognostics system of claim 1, further comprising:

self-searching means for finding an optimal combination of parameters/functions required by said conjecture method or said prediction method when said conjecture method or said prediction method is just selected and related initial values are just set up, thereby increasing the prediction/conjecture accuracy.

4. The quality prognostics system of claim 1, further comprising:

self-adjusting means for tuning system parameters to meet the prediction/conjecture accuracy requirement of said production tool when a prediction accuracy or a conjecture accuracy is lowered than a predetermined lower bound.

5. The quality prognostics system of claim 1, further comprising:

a selection-and-setup interface for assisting users to select said conjecture method and said prediction method properly, and setting up related initial values.

6. The quality prognostics system of claim 1, wherein said conjecture modeling means has the function of virtual metrology.

7. The quality prognostics system of claim 1, further comprising:

an evaluation index used to evaluate the conjecture/prediction accuracy for said manufacturing processes.

8. The quality prognostics system of claim 1, wherein said conjecture modeling means is built in accordance with said first neural network technique, and said prediction modeling means is built in accordance with said second neural network technique which is the same as said first neural network technique, thereby forming said conjecture modeling means and said prediction modeling means as a combined conjecture-prediction modeling means with outputting said conjecture value for said product lot currently manufactured in said production tool, and said prediction value for said next product lot.

9. A quality prognostics method for manufacturing processes, comprising:

providing a conjecture modeling step for using a set of current input data of a production tool to conjecture a current conjecture value for a product lot currently manufactured in said production tool, wherein said conjecture modeling step is based on a conjecture method, and said conjecture method is selected from the group consisting of a first neural network technique, a fuzzy logic technique and a stepwise regression technique; and

providing a prediction modeling step for using said current conjecture value of said product lot together with at least two actual measurement values of at least two previous product lots to predict a prediction value for a next product lot, wherein said at least two previous product lots is respectively produced in at least two production cycles, and said at least one actual measurement value is one-to-one corresponding to said at least one previous product lot, and said prediction modeling step is based on a prediction method, and said prediction method is selected from the group consisting of a weighted moving average technique and a second neural network technique.

10. The quality prognostics method of claim 9, further comprising:

performing a raw-data pre-processing step for converting a set of raw data inputted from said production tool to a set of desired input data of a specific format.

11. The quality prognostics method of claim 9, further comprising:

performing a self-searching step for finding an optimal combination of parameters/functions required by said conjecture method or said prediction method when said conjecture method or said prediction method is just selected and related initial values are just set up, thereby increasing the prediction/conjecture accuracy.

12. The quality prognostics method of claim 11, wherein said conjecture method adopts said first neural network technique for building neural networks, said conjecture modeling step further comprising:

collecting sensor data of said production tool, and selecting a plurality of input parameters for said production tool;

assigning a possible number of hidden layers, possible numbers of nodes and possible transfer functions;

selecting a plurality of training data sets and a plurality of test data sets;

normalizing said sensor data and said input parameters before entering a training step of said neural networks, and de-normalizing an output value derived from said first neural network technique after said training step is done;

setting up conditions for terminating said training step;

setting up initial weights for said neural networks;

running said neural networks for obtaining a conjecture output value;

computing an error value between said conjecture output value and said actual measurement datum;

checking if said conditions for terminating said training step have been met, wherein said conditions includes that the number of training cycles for said training step reaches a predetermined limit value, and that said error value is lower than a predetermined error value; and

said training step of said neural networks is terminated when either one of said conditions has been met;

when said conditions have not been met, returning said error value to said step of setting up said initial weights for re-tuning, and then running said neural networks;

checking if said self-searching step is finished and generating a checking result;

when said checking result is no, selecting a new possible number of hidden layers, new possible numbers of nodes and new possible transfer functions; and

when said checking result is yes, said initial weights, said possible number of hidden layers, said possible numbers of nodes and said possible transfer functions are selected as an optimal combination, and completing the establishment of said conjecture modeling.

13. The quality prognostics method of claim 9, wherein said conjecture modeling step can be used for performing virtual metrology.

14. The quality prognostics method of claim 9, further comprising:

providing an evaluation index used to evaluate the conjecture/prediction accuracy for said manufacturing processes.

15. The quality prognostics method of claim 9, wherein said prediction method adopts said weighted moving average method, said weighted moving average technique using said actual measurement values (y_i−

1, y<sub>i−

2</sub>) of two previous product lots obtained from a measurement tool together with said current conjecture value (ŷ

_i) to predict said prediction value ({tilde over (y)}_i+1) of said next product lot, wherein the formula of said weighted moving average method is {tilde over (y)}_i+1=w_iŷ

_i+w_i−

y<sub>i−

1</sub>+w<sub>i−

2</sub>y<sub>i−

2</sub>wherein w_i, w_i−

, w<sub>i−

2</sub> are assigned weights.

16. The quality prognostics method of claim 9, further comprising:

performing a self-adjusting step for tuning system parameters to meet the prediction/conjecture accuracy requirement when a prediction accuracy or a conjecture accuracy is lowered than a predetermined accuracy bound.

17. The quality prognostics method of claim 16, wherein said self-adjusting step comprises:

inputting sensor data and a plurality of process parameters from said production tool;

performing said conjecture modeling step or said prediction modeling step for obtaining said current conjecture value or said prediction value for said next product lot;

evaluating said conjecture accuracy or said prediction accuracy for said current conjecture value or said prediction value;

determining if said conjecture accuracy or said prediction accuracy is lower than said predetermined lower bound, and generating a determination result; and

when said determination result is yes, adding the current set of production sensor data and measurement datum of which said conjecture accuracy or said prediction accuracy is lower than said predetermined lower bound to said training data sets, and re-training and rebuilding said conjecturing modeling step or said prediction modeling step, and then subsequently performing said conjecture modeling step or said prediction modeling step for said next product lot.

18. The quality prognostics method of claim 17, wherein when said determination result is no, performing said conjecture modeling step or said prediction modeling step for a new next product lot following said next product lot.

19. The quality prognostics method of claim 9, wherein said conjecture modeling step is based on said first neural network technique, and said prediction modeling step is based on said second neural network technique which is the same as said first neural network technique, thereby forming said conjecture modeling step and said prediction modeling step as a combined conjecture-prediction modeling step with outputting said conjecture value for said product lot currently manufactured in said production tool, and said prediction value for said next product lot.