Method for prognostic maintenance in semiconductor manufacturing equipments

1. A method for prognostic maintenance in semiconductor manufacturing equipments, comprising the steps of:

collecting a plurality of raw data and preprocessing the plurality of collected raw data to filter out meaningless null detection values existing in the plurality of raw data and generate detection values of normal pattern;

performing classification through a statistic classification model on the plurality of preprocessed raw data to generate a plurality of health indices;

performing classification on the plurality of generated health indices by a prescribed classification method to generate a plurality of health information;

using a regression analysis method to process the plurality of health information to generate a plurality of health reports; and

performing repairs and maintenance actively by in-situ engineers based on the plurality of generated health reports.

2. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 1 , wherein the statistic classification model is a Neural Network Model (NN Model).

3. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 2 , wherein the plurality of raw data consist of a plurality of historic data and a plurality of newly added data, in which the plurality of historic data indicate the data outputted by the semiconductor equipments under healthy condition, and the plurality of newly added data represent the data outputted by the semiconductor equipment under unknown condition.

4. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 3 , wherein the steps taken by the statistic classification model comprise:

through a learning phase, generating a plurality of historic neural points from the plurality of historic data, and building a distribution configuration of the plurality of historic neural points;

through an on-line monitoring phase, generating a plurality of newly added neural points from the plurality of newly added data, and building a distribution configuration of the plurality of newly added neural points; and

by means of a comparison phase, analyzing the distribution configuration of the plurality of historic neural points and the distribution configuration of the plurality of newly added neural points to acquire a plurality of health indices.

5. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 4 , wherein, in the learning phase, the plurality of historic neural points are acquired from the plurality of historic data through classification by a neural network classifier; then the plurality of historic neural points are processed to build the plurality of historic neural points presenting an elliptic distribution configuration.

6. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 5 , wherein, in the on-line monitoring phase, the plurality of newly added neural points are acquired from the plurality of newly added data through processes by the neural network classifier; then the plurality of newly added neural points are processed to build the distribution configuration of the plurality of newly added neural points within the elliptic distribution configuration.

7. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 6 , wherein the comparison phase consists of:

after acquisition of the distribution configuration of the plurality of newly added neural points within the elliptic distribution configuration, comparing the distribution configuration of the plurality of historic neural points with the distribution configuration of the plurality of newly added neural points based on the elliptic distribution configuration, thus generating a plurality of comparison values; and

processing conjunctively the plurality of comparison values to generate the plurality of health indices.

8. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 1 , wherein the prescribed classification method is the Principal Component Analysis (PCA).

9. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 8 , wherein the classification performed by the prescribed classification method on the plurality of health indices consists of:

performing a linear conversion operation on the plurality of health indices for classification, in which the linear conversion operation converts the plurality of health indices in an original coordinate system into a new coordinate system, and the new coordinate system has a plurality of new coordinate axes, and the plurality of new coordinate axes are respectively the first new axis, the second new axis, . . . , and the Nth new axis;

finding projection values of the plurality of health indices projected onto the plurality of new axes, acquiring a plurality of first principal component values over the first new axis, a plurality of second principal component values over the second new axis, . . . , and a plurality of Nth principal component values over the Nth new axis;

processing the plurality of first principal component values, plurality of second principal component values, . . . , and plurality of Nth principal component values based on a plurality of confidence indices built by in-situ engineers to acquire a plurality of principal component characteristic values;

generating a plurality of health information according to the plurality of principal component characteristic values.

10. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 9 , wherein the confidence indices simplify the plurality of health indices through retaining lower-rank principal component values and ignoring higher-rank principal component values.

11. A method for prognostic maintenance in semiconductor manufacturing equipments, comprising the steps of:

collecting a plurality of raw data, the plurality of raw data being provided by a Fault Detection and Classification (FDC) system, and the plurality of raw data indicate variation values detected in real-time on each wafer by the FDC system during semiconductor processes, and preprocessing the plurality of collected raw data to filter out meaningless null detection values existing in the plurality of raw data and to generate detection values of normal pattern;

performing classification through a Neural Network (NN) model on the plurality of preprocessed raw data to generate a plurality of health indices;

performing classification on the plurality of generated health indices by the Principal Component Analysis (PCA) to generate a plurality of health information;

using the Partial Least Squares Discriminated Analysis (PLS-DA) to process the plurality of health information to generate a plurality of health reports; and

performing repairs and maintenance actively by in-situ engineers based on the plurality of generated health reports.

12. A method for prognostic maintenance in semiconductor manufacturing equipments, comprising the steps of:

collecting a plurality of raw data and preprocessing the plurality of collected raw data, the plurality of raw data consisting of a plurality of historic data and a plurality of newly added data, in which the plurality of historic data indicate the data outputted by the semiconductor equipments under healthy condition, the plurality of newly added data represent the data outputted by the semiconductor equipment under unknown condition;

performing classification through a Neural Network Model (NN Model), on the plurality of preprocessed raw data to generate a plurality of health indices;

performing classification on the plurality of generated health indices by a prescribed classification method to generate a plurality of health information;

using a regression analysis method to process the plurality of health information to generate a plurality of health reports; and performing repairs and maintenance actively by in-situ engineers based on the plurality of generated health reports.

13. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 12 , wherein the steps taken by the NN Model comprise:

through a learning phase, generating a plurality of historic neural points from the plurality of historic data, and building a distribution configuration of the plurality of historic neural points;

through an on-line monitoring phase, generating a plurality of newly added neural points from the plurality of newly added data, and building a distribution configuration of the plurality of newly added neural points; and

by means of a comparison phase, analyzing the distribution configuration of the plurality of historic neural points and the distribution configuration of the plurality of newly added neural points to acquire a plurality of health indices.

14. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 13 , wherein, in the learning phase, the plurality of historic neural points are acquired from the plurality of historic data through classification by a neural network classifier; then the plurality of historic neural points are processed to build the plurality of historic neural points presenting an elliptic distribution configuration.

15. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 14 , wherein, in the on-line monitoring phase, the plurality of newly added neural points are acquired from the plurality of newly added data through processes by the neural network classifier; then the plurality of newly added neural points are processed to build the distribution configuration of the plurality of newly added neural points within the elliptic distribution configuration.

16. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 15 , wherein the comparison phase consists of:

after acquisition of the distribution configuration of the plurality of newly added neural points within the elliptic distribution configuration, comparing the distribution configuration of the plurality of historic neural points with the distribution configuration of the plurality of newly added neural points based on the elliptic distribution configuration, thus generating a plurality of comparison values; and

processing conjunctively the plurality of comparison values to generate the plurality of health indices.

17. A method for prognostic maintenance in semiconductor manufacturing equipments, comprising the steps of:

collecting a plurality of raw data and preprocessing the plurality of collected raw data;

performing classification through a statistic classification model on the plurality of preprocessed raw data to generate a plurality of health indices;

performing classification on the plurality of generated health indices by Principal Component Analysis (PCA) to generate a plurality of health information, wherein the plurality of health indices consists of:

performing a linear conversion operation on the plurality of health indices for classification, in which the linear conversion operation converts the plurality of health indices in an original coordinate system into a new coordinate system, and the new coordinate system has a plurality of new coordinate axes, and the plurality of new coordinate axes are respectively the first new axis, the second new axis, . . . , and the Nth new axis;

finding projection values of the plurality of health indices projected onto the plurality of new axes, acquiring a plurality of first principal component values over the first new axis, a plurality of second principal component values over the second new axis, . . . , and a plurality of Nth principal component values over the Nth new axis;

processing the plurality of first principal component values, plurality of second principal component values, . . . , and plurality of Nth principal component values based on a plurality of confidence indices built by in-situ engineers to acquire a plurality of principal component characteristic values; and

generating a plurality of health information according to the plurality of principal component characteristic values;

using a regression analysis method to process the plurality of health information to generate a plurality of health reports; and

performing repairs and maintenance actively by in-situ engineers based on the plurality of generated health reports.

18. The method for prognostic maintenance in semiconductor manufacturing equipments according to claim 17 , wherein the confidence indices simplify the plurality of health indices through retaining lower-rank principal component values and ignoring higher-rank principal component values.