Predictive Analytics in Manufacturing

Predictive Analytics in Manufacturing

• 1.
  Kittiphan Pomoung
• 2.
  About Me • MrKittiphan Pomoung • Education : Master degree in Recording Technology @ KMITL • Experiences : – 21 years working experience in Hard Disk Drive companies – 9 months in Data Mining Project Collaboration with IBM • Email Address : [email protected]
• 3.
  Topics • Industry Revolusion(Trend) . • Power of prediction: assess results in advance, identify key challenges and how to overcome them. • A taste of success: simple data modeling applied to a real case in manufacturing process with a satisfactory result.
• 4.
  Special Thanks • EakasitPacharawongsakda, PhD. • Aimamorn Suvichakorn, PhD. • Kosit Bunsri, M. Eng.
• 5.
  Industry 4.0 • Waterand Steam Power • First Power Loom- 1784 • Electric Energy • Assembly belt , 1870 • Electronics and information • Programmable Logic Control – PLC, 1969 • Cyber-Physical System • All tools will communicate and Data will be shared to each other . • Product and Machine talk together • Build Per Order, (flexible with RFID) 2nd Revolution 1st Revolution 3rd Revolution 4th Revolution
• 6.
  Prediction in Manufactory •Market and Demand Forecast • Machine Utilization • Preventive Maintenance • Quality Improvement
• 7.
  Challenges • High expectationin prediction accuracy • Unknown factors and variables – Oli Price – Market’s demand • Inadequate resources – Knowledgeable staffs – Tools • Limited data and understanding.
• 8.
  Part 2
• 9.
  Reliability Prediction • Reliabilitytest can take very long time (>1000 hrs), sometimes with temperature variation. – Tyre, chair, motor and HDD. • What if, we can predict the result earlier, “before the test starts”. – Traditional Method – Advance / Numerical Predictive Method
• 10.
  10 Components in anHDD • Can be more than 17 components • One component possibly comes from 2 suppliers • At least 34 variables in total! , Many data stored
• 11.
  Reliability Prediction :Background Electrical Test Assembly Components > 16 parts Reliability Test Done • Data : > 200 parameters (attributes and variables). 1mil data entries per week. • Duration: • Some components manufacturing process > 60-90 days • Reliability test 700-1200 hrs • Worst case total processing time is 4 months • What if, the predictive model can predict the result earlier Basic Hard disk Process
• 12.
  Basic Hard DiskDrive Reliability Test Process • Test under stress condition • 700-1200 hrs test time • Limited samples for training • 200-300 drives per batch • Only 1-2 failed units per batch • Some failures occurred at late test hours (worn-out). Reliability Prediction : Background
• 13.
  • Objective :To predict the Reliability test result for new material qualification in term of failure rate. • Benefit : Time saving ($$) and Quality Improvement. • Background : – New material qualification usually takes 3 months. – Failure could occur at the last minute of Reliability test, at the last test station. – If happens, to re-design and re-qualification again. • Challenges : – Limited failed samples to form the correlation – Reliability test is more stressed than usual electrical tests Reliability Prediction : Project
• 14.
  Data Preparation Classification Test Train Deployment Validation Predictive Model FeatureSelection • To improve efficiency and accuracy • 200 parameters down to ~ 20 attributes Classification : • Rule Base  Moderate • *Decision Tree  Good • Fusion (Naive Bayes + Decision Tree)  Best In Class Techniques : Limited failed drives • *Over sampling / Boosting • Under Sampling • Result ~ 65-70% accuracy when implemented. • The classification model is continuingly optimized by training with new samples. Reliability Prediction : Workflow
• 15.
  • Different productsrequire different modelling techniques. • Classification method could be constrained when implementing • Rule Base --> Moderate • Decision Tree  Good results, easy for implementation • Fusion (Naive Bayes + Decision Tree)  Best In Class • Future Works • Defined key parameters input variables (KPIV) • Establish KPIV/KPOV that correlating to component level . • Establish Predictive model at component level (prior to HDD assembly) Reliability Prediction : Lessons
• 16.
  Part 3
• 17.
  Process/Model (Transfer Function) Inputs Parameters Output Defect/Pareto Performance /Distribution Factors(Power/ Temp) PerformancePrediction : Objective • Objectives : – Product Boundary or Product Capability – Relationship / transfer Function – Know it earlier , as fast as it possible – What is the effect if input change (optimization)
• 18.
  Performance Prediction :Process-1 • Average and Stdev of Input Population • Buy off Distribution Type of output Population
• 19.
  Performance Prediction :Process-2 Output Input • Transfer function – Output= 15.328527 + 7.012858*Input- 1.5895329*(Input-3.2)^2 Trails -2 -1 CT +1 +2 Input (Avg) 2.7 3.0 3.2 3.4 3.6 Output (Avg) 34 36 37.7 39.3 40.7 • Calculate (simulate ) the Average of output distribution
• 20.
  Performance Prediction :Process-3 • Generate (pseudo) output distribution with Random Technique – Excel/JMP : Random Normal (Output’s Avg, sigma) – SS = 1000* • Iterations and Sample size are able to improve accuracy Trails -2 -1 CT +1 +2 Input (Avg) 2.7 3.0 3.2 3.4 3.6 Output (Avg) 34 36 37.7 39.3 40.7
• 21.
  2.48 2.72 2.963.2 3.44 3.68 3.922.48 2.72 2.96 3.2 3.44 3.68 3.92 Output Input Failure Rate (CDF Plot) Spec/Limit (LCL) Good • 3.4 is minimum requirement to meet product capability Performance Prediction : Result Output Input • Product Performance(Output) vs Incoming Performance (Input)
• 22.
  • Linear Regressionand Monte Carlo Techniques – Simple – Decent and Fair Result • Iterations and Sample size – More is Good, better accuracy – Sample size > 50k , 5 times per each input. – Averaged the result for each run . Performance Prediction : Lessons
• 23.
  End