Data Mining Techniques Applied in Advanced Manufacturing PRESENT BY WEI SUN

Background  Complex production process  Large volume of production data  Explore the underlying dependencies of production data  Improve production rate  Reduce defective products

Application Gap  Capacity what the data looks like  Capability how the data can be utilized  Knowledge how to perform knowledge discover and management

Application Gap

Solution: PDP-Miner  A data analytics platform customized for process optimization  Data mining techniques  Big data infrastructures

Concrete Case: PDP Manufacturing  Production line: 6000 meters  Production process: 75 assembling routines  Production time: average production time requires 76 hours  Production equipment: 279 major production equipment  Controlling parameters: more than 10,000 controlling parameters

Workflow

Key Questions  What are the key parameters whose values can significantly differentiate qualified products from defective products  How the parameter value changes affect the production rate  What are the effective parameter recipes to ensure high yield rate

Achievement 1  DATA ANALYTICS PLATFORM WITH 3 FUNCTIONS  1. Cross-language data mining algorithms integration  2. Real-time monitoring of system resource consumption  3. Balancing the node workload in clusters

Load Balance in Dynamic Environment

 The system can balance the workloads in a dynamically changed cluster  The entire system can be linearly extended with resources of different computing power.

Achievement 2  Regression modeling to describe the relationship between product quality and various parameters  Association based methods to identify feature combinations that can significantly improve the quality of product

Regression Analysis  Linear regression models  1. Ridge Regression  2. Lasso Regression

Regression Analysis Outcomes  1. The various of the humidity of the air has positive correlation with the yield rate  The pressure of air has positive correlation with the yield rate. The less the pressure changes, the higher the yield rate would be

Regression Analysis Outcomes

 Temperature and humidity have significant correlations with the product quality  When the surrounding temperature is under 27 centigrade, the number of defective products increases dramatically

Regression Analysis Outcomes

Association Based Classification  CARs: Class association rules  {r: F y}, F: a subset of the entire feature value set y: class label

Association Based Classification  For each CAR Support: s Confidence: c  Indicate how many records contain F and the ratio of records containing F that are labeled as y  A rule based classifier is built by selecting a subset of the CARs that cause the least error

Association Based Classification  Early detection strategy  If CARs refer to the features in the early manufacturing process, it could identify semi-finished defective products quickly, which prevent further resource waste.

System architecture: PDP-Miner  Data Analytics Platform  Data Analysis Modules

SYSTEM ARCHITECRTURE

DATA ANALYTICS PLATFORM  Easy operation for task configuration  Flexible supports for various programs  Effective resource management

DATA ANALYSIS MODULES  1. Data Exploration  2. Data Analysis  3. Result Management

DATA ANALYSIS MODULES: 1. Data Exploration  Comparison Analysis  1. Quickly identify parameters whose values are statistically different between two datasets  2. Be able to extract the top-k most significant parameters  Data Cube  1. Explore high dimensional data  2. Be able to conduct multi-level inspection of data

DATA ANALYSIS MODULES: 2. Data Analysis  Important Parameter Selection  Regression Analysis  Discriminative Analysis

DATA ANALYSIS MODULES: 3. Result Management  The important parameter list  Parameter value combinations  The regression model

Discriminative Analysis  Association based classification  Low-Support discriminative pattern mining

Low Support Discriminative pattern mining  Production process generates high dimension data, and it’s time-consuming when utilizes standard association rule based methods  SMP: Low support pattern mining pattern  Integrate SMP algorithm into PDP-Miner

Discover key parameters

Discover Key Parameters  Step 1: Separate data set into 2 categories based on product quality: GOOD(Qualified products) and SCRAP(Defective products)

Discover Key Parameters  Step 2: Generate hundreds of frequent parameter value combinations for each given dataset

Discover Key Parameters  Step 3: Extract the frequent combinations in SCRAP that are not frequent in GOOD and obtain the value combinations that result in defective products

Important Parameter Discovered Big red crosses indicate that the values present densely on SCRAP products Such a parameter value combination should be avoided in the production practice

Deployment Practice  1. Be able to control product quality and reduce cost by decreasing defective products  2. Be able to quick diagnosis on parameter values  3. Be able to verify and validate analytic results  4. PDP yield rate increased from 91% to 95%

Thank You ANY QUESTIONS?