NUS-ISS Learning Day 2018- Application of analytics in manufacturing sector
- 1. Application of Analytics in Manufacturing Sector Mr. Prashant Jain Mr. Zixin Wang NUS-ISS, 13 July 2018 #ISSLearningDay2018 HP Inc. Singapore
- 2. AGENDA • Business Understanding • Think Approach • Process Understanding • Data Understanding and Exploration • Modelling Results and Conclusion #ISSLearningDay2018
- 3. Business Understanding Solving a problem is useless if it carries no business value. #ISSLearningDay2018
- 4. Let’s talk business @ HP INDIGO • Electro-Ink Manufacturing Facility for CMYK • The Indigo Electro-Ink are used in HP Indigo Digital Presses. • HP has proprietary, patented technology and a business model which sells both presses, consumables and services.
- 5. Let’s talk business @ HP INDIGO 1. •Typical batch process is several hours long •The batch processes are highly repeatable 2. •Several raw materials are added in a tank •Lot of process operations are carried out 3. •Finally, batch quality is checked at the end. •Batch is unloaded and packaged. • Electro-Ink Manufacturing Facility for CMYK • The Indigo Electro-Ink are used in HP Indigo Digital Presses. • HP has proprietary, patented technology and a business model to sells both presses, consumables and services.
- 6. Characteristics of Batch Manufacturing GOALS Produce maximum yield Best product quality Minimum amount of waste CHALLENGES Highly complex process Hidden interlinked attributes Huge amount of data Easily miss atypical process variation relationships WHY PREDICTIVE ANALYTICS ? Summarize all SPC charts End-of-Batch Quality Prediction Fault Prediction
- 7. Think Approach Come up with logical approach #ISSLearningDay2018
- 9. Comparing Approaches at hand UNIVARIATE OR SPC •Doesn't take into account interactions among inputs •Interactions between process variables and output quality characteristics MULTIVARIATE SPC (MSPC) •Compiles many acceptable batches •Accounts for process relationships •Allows a much more accurate analysis Understanding Golden Batch Understanding Golden Profile VS
- 10. Process Understanding Domain knowledge is essential to any data analytics project #ISSLearningDay2018
- 11. Understanding Ink production Process Agitator Cooling water in Cooling water out Material-1 Material-2 Material-3 GRINDING TANK RECIRCULATION • The Ink grinding process inside Attritor is several hours long with many stages before transferring. • Pain Point: If batch-end Ink quality is poor, we won’t know until the last hour. • Objective: To come up with a predictive model to predict the quality of batch half way through the process so that suitable action can be taken when we still have time to improve the batch. Parameter Type Data points per batch Process (every min logging) 800-1000 Quality (end of batch) 1-data point
- 12. 12 7 hrs midway 5 stages passed Stage-1 Stage-2 Stage-3 Stage-4 Stage-5 Stage-6 Stage-7 Stage-8 Stage-9 Parameter-1
- 13. Data Understanding and Exploration Data collection, Integration, Preparation and Transformations #ISSLearningDay2018
- 14. Data collection and SUCCESS Criteria Started with Magenta Color for Modelling: • Process data (sample shown in the table) is collected for 4 months. • Ink Quality has various parameters and is very subjective in itself. • Process Team chose ‘DE’ to be the dependent parameter for modelling. • Quality depends on properties of Raw Materials also – 3rd data set. Success Criteria: • Predictive modeling for DE(Quality parameter) with at least 90% Model accuracy.
- 15. Dependent variable Distribution • 4 months of data is collected • Dependent Variable histogram looks Normal. • Some outlier batches removed. • Shapiro test: NHT accepted, data is normally distributed. 15
- 16. Correlation analysis Correlation close to +1 –> Positively correlated Correlation close to -1 –> Negatively correlated • Carried out correlation analysis and discovered that process parameters have strong correlation with corresponding setpoints. Insights: Certain parameters have high correlation with other parameters and techniques that tend to combine highly correlated parameters into single parameter can be used – Dimension reduction
- 17. Predictive Modelling workflow Input parameters: • 4 months of Magenta batches • 900 data points per batch • 16 variables per batch Output parameters data: • 14 quality related variables Raw Material properties data: • RM properties data • Total 46 variables Batch Data Validation Batch Wise Unfolding Data Integration Dimension Reduction Modelling Evaluation Prediction
- 18. Batch Data validation – Golden profile Batch Categories Green: • Perfect batch • Process attributes are within limits • Quality parameters are fine Yellow: • Imperfect batch • Some process attributes have variations • Quality parameters are fine Red: • Failed batch • Process attributes are out of limits • Quality parameters are out of limits 18 0 20 40 60 80 100 120 140 160 180 Batch Categories of 4 months of data Red Green Yellow
- 19. PROCESS DATA – 3D FORMAT 19 Batch Data Not of uniform length & 3D Format Raw Materials data (46 variables ) Output paramete rs (1- variable) Input Paramete rs (16 variables) Process data - 3D format can’t be modelled directly.
- 20. Batch alignment - Dynamic Time Warp • Sometimes operator and event-initiated processing halts and restarts are part of the batch process design, such as adding a special ingredient. • However, the batch data used in model development must be somehow aligned in order to facilitate data analysis. • To achieve uniform batch length. the data at a certain time in the batch could be simply chopped off or compressed. • Better results may be achieved by applying a newer technique known as dynamic time warping (DTW). • DTW aligns batch data with the reference trajectories by minimizing total distance between them. • The batch with median time duration can be used as an initial DTW reference batch. 20
- 21. 21 • Process data is logged at every minute. • 3D data format can’t be modelled directly. It needs to be unfolded into 2D format. • Used Batch-wise unfolding technique to change to 2D format. • Resulting data is a flat matrix with over 16000 columns. • For POC, need not take data at every minute. • Sampling at every 10min of first 500 data points (approx. 8.3 hours) to get 830 columns. BATCH-WISE UNFOLDING
- 22. Data integration Input parameters Approx. 830 variables Primary key = Batch ID Output parameters 1 variable – Dependent (y) Primary key = Batch ID Raw Material Data 46 variables Primary key = Batch ID Batch-ID Primary Key (150 batches) Batch-wise Unfolded Input parameters (sampling at every 10min – 830 variables) Raw Material Data 46 variables DE (Selected Output Parameter) Final 2D Integrated Dataset Still 800+ columns, need to compress
- 23. Dimension reduction with principal component analysis • Reduced 815 variables to just 100 Principal components. • Selected on the basis of screen plot eigen value criteria which explain 99.6% of the total variance. • Modeling with PCs – Neural Nets, Decision Trees, Random Forest, Gradient Boost Machines, MLR & etc. 3 dimensions >>> 2 dimensions
- 24. Modelling Results and Conclusion Model selection and discussion #ISSLearningDay2018
- 25. Multiple Linear regression – Best models • Divide dataset into 70:30 for training and testing. • Built model on Training data and checked % error on Test data. 25 Model No. Variables Adj.R2 MAPE on test M1 All PC (100) 0.75 9.07% M2 35 significant PC from M1 summary 0.74 9.76% M3 35 PC + Removed 2 outliers using Cook’s plot 0.77 10.1% M4 33 significant PC from M3 summary 0.76 10.4% M5 26 significant PC 0.71 10.8% Best Models
- 26. y Prediction interval at 95% confidence 26 Base Model Prediction is Average y 9 points out of prediction interval. Base Model Prediction accuracy is 80%. Does not accounts for parameter variations. Regression Model 3 points out of prediction interval Model Prediction accuracy is 92.8%. Advantage: Can predict variations in data better than base model. Base Model
- 27. CONCLUSION • This approach, when applied to analyse the impact of processing conditions on final-product quality, can provide operators with continuous and accurate predictions of end-of-batch quality parameters. • A successful application of batch analytics may result in minimizing batch variations and improving batch quality, validated for a batch process.