190 likes | 468 Views
Predictive Maintenance. 張智星 (Roger Jang) jang@mirlab.org http://mirlab.org/jang 台灣大學 資訊系 MIR 實驗室. Problem Definition. Given a set of training data Input: A set of wafers with their sensor readings Output: Maintenance stage (1-3) Goal Construct a model to predict maintenance stage.
E N D
Predictive Maintenance 張智星 (Roger Jang) jang@mirlab.org http://mirlab.org/jang 台灣大學 資訊系 MIR實驗室
Problem Definition • Given a set of training data • Input: A set of wafers with their sensor readings • Output: Maintenance stage (1-3) • Goal • Construct a model to predict maintenance stage
Important Statistics • Specs. of the problem • No. of recipe steps = 10 • No. of sensors = 57 • Statistics for a step: • mean, min, max, std, slope (mean is used here) • No. of wafers = 47 • By machines at different maintenance stages • No. of maintenance stages = 3 • Specs. of classification • No. of instances = 47 • No. of features = 10*57*5 = 2850 • No. of classes = 3
Average of 57 Sensor Readings • 47 curves for each subplot, 10 points for each curve
LDA Projection to 2D Plane for Each Sensor • LDA for dimensionality reduction (10 2)
Best Performance For Each Sensor • Best LDA projection for a single sensor
Comparisons • Result of prediction Groundtruth: Predicted:
Application Case 2 • Characteristics • Different set of recipes • More wafers
Important Statistics • 5 Recipe steps • Stabilize, strike, nitridation, dechuck, purge • No. of sensors = 50 • Used statistics = mean • No. of wafers = 1522 • No. of maintenance stages = 3
Best Performance For Each Sensor • After best LDA projection from 5 inputs
Best Perf. For All 435 Sensor Pairs • After best LDA projection from 5*2 inputs
Best Perf. For Top-100 Sensor Pairs • After best LDA projection from 5*2 inputs
Confusion Matrix and Error Positions • Confusion matrix • Error positions
Future Work • To explore further: • Which recipe is the most influential? • Within the most influential recipe, which sensors are the most influential? • How to further explore feature extraction/selection? • Machine-dependent modeling?
Application Case 3 • Characteristics • 9 maintenance stages
Best Performance For Each Sensor • After best LDA projection from 5 inputs
Best Perf. For All 435 Sensor Pairs • After best LDA projection from 5*2 inputs