Process Data Analytics State of the art and applications in oil sands industry

Transcription

1 Industrial Research Chair Control of Oilsands Processes Process Data Analytics State of the art and applications in oil sands industry Biao Huang Department of Chemical and Materials Engineering University of Alberta Sept. 2nd, 2016 IFAC MMM Vienna - BH

2 Outline Data Analytics State of the art Oil Sands Industry Process Data Analytics in Applications Analytics Toolboxes in Progress Conclusion 2

3 Data Analytics 3

4 Data Analytics 4

5 Data Era 5

6 Data Science 6

7 History of Data Analytics Analytics 1.0 Traditional Analytics (mid-1950s ) Analytics 2.0 Big Data (early 2000s - today) Analytics 3.0 Data Economy (future) 7

8 Data Analytics vs Data Analysis 8

9 Increasing Trend of Data Analytics 9

10 Application of Data Analytics: Engineering Conclusion: Growth of natural resources industry stagnated; where we are Big data in natural resources industry has great potential. 10

11 Typical Algorithms in Data Analytics Supervised learning Regression: LASSO, Decision tree, PLS, MLR Classification: Logistic regression Hybrid: Gaussian Process, Neural Network, SVM/SVR Unsupervised learning Dimension Reduction: PCA Clustering: k-means Inference Maximum Likelihood, Expectation Maximization Bayesian Method, Variational Bayesian, Bayesian Network 11

12 Data Analytics Software Platform and Toolboxes 12

13 Process Data Analytics(Engineering) 13

14 High Dimensionality * 14

15 High Dimensionality of Data - Decoding PCA/PLS/ICA and Applications: Dimensionality Reduction Pre-processing for many data mining tasks (Noise Reduction) Analyze data and to find patterns 15

16 Robustness Dealing with Irregular Data Modeling the noise: Gaussian distribution vs others Noise Time Histogram fitted by a t distribution Histogram fitted by a Gaussian distribution 16

17 Multimodality Process Pattern Model A Operation mode A Operation mode C Operation mode B Model C Model B 17

18 Nonlinearity Local Solution JIT modeling = Locally weighted modeling = Relevance-In-Space modeling = Lazy modeling Output Input Contributed by Sanghong Kim from Kyoto University 18

19 Nonlinearity Gaussian Process,,,,, Analytical expression of likelihood exists 19

20 Time varying time delays Multirate sampling with time Varying time-delays: Dual rate: fast rate input while slow rate output Time delay is varying at every sample 20

21 Errors-In-Variables (EIV) Noise-corrupted measurements:, Additive noise:, Unknown noise-free input and output:, 21

22 Process Knowledge - Bayes Methods Queries Observations py ( xpx ) ( ) p( x y) kp( y x) p( x) py ( ) Posterior (estimation of unknown) Prior (knowledge) Likelihood (model fit) 22

23 Oil Sands 23

24 Canada s Oil Sands 141,000 square kilometres deposit 1.7 trillion barrels of bitumen 170 billion barrels recoverable second largest oil reserve 1.3 million barrels crude oil per day Source: 24

25 From Oil Sands to Sweat Crude Oil 1. Mining 2. Crush & Conveyance 3. Extraction 4. Froth Treatment 5. Upgrading Source: 25

26 Extraction Source: 26

27 Source: 27

28 Process Data Analytics in Oil Sands 28

29 Shovel/Truck Oil Sands Feeder/Crusher Slurry Preparation Soft Sensor Feed Surge Hydrotransport to Extraction Inferential Control Froth Treatment Solvent Recovery Solvent Primary Separation Vessel Bitumen Froth Floatation Low Energy Extraction Thickener Thickened Fine Optimization Tails Sand Tailings FDI Control Monitoring Bitumen Storage Upgrader 29

30 Rapid Technology Transfer Platform MATLAB PI Database PCN LCN OPC Servers DCS Application Platform Individual Soft Sensor Individual Soft Individual Sensor Application Individual Soft Sensor Individual Application Valves Transmitters Analyzers 30

31 Application Platform PI Database PCN LCN OPC Servers DCS MATLAB Individual Soft Sensor Application Individual Monitor Individual Application Individual Soft Sensor Individual Application Platform Only One Platform Multiple OPC Multiple Individual Applications Rapid Technology Transfer Efficient Implementation Convenient Maintenance Valves Transmitters Analyzers 31

32 Data Analytics in Image Processing 32

33 Introduction of PSV Three layers due to density difference Froth/Middling interface level is the most important control variable Image source: 33

34 PSV interface measurements Density profiler Interface measurements D/P Cell Camera Best performance, however, only reliable when the performance index (PI) is higher than a given constant. Objective: Improve the camera reading reliability when the performance index is low. Image Analysis Image source: 34

35 Experimental Design Oil Camera Water 35

36 Image Captured by Camera The image shown on the right is the original image observed by the camera. Oil Water 100%?? 100% Pure Oil Mixture Pure Water Objective: Segment the captured image as a binary image (+1 for oil/-1 for water/ 0 for interface) Note: the image size for all images is (600 pixels 600 pixels) 36

37 Theory of Data Based Image Analysis Images can be modeled using Markov random field (MRF). Each pixel is considered as a random variable (RV) Each random variable (pixel) has a corresponding observation (corrupted with noise) Noisy observations Aim: to recover clean pixels from noisy observations MRF is employed to perform image segmentation and classification. Random variables (clean pixels) 37

38 Principle of MRF Estimation Potential of observation Potential of neighbors Energy minimization: Total energy = potential of observation + potential of neighbors 38

39 Image Segmentation Image Segmentation 39

40 Pixel Values Profile 0 Averaged horizontal pixel values profile 100 Oil: Water:

41 Mixture Boundary Determination 0 Averaged horizontal pixel values profile Pure Oil Mixture Pure Water

42 Mixture Boundary Indication 0 Averaged horizontal pixel values profile Next step: Identify the interface based on the pixel value close to zero 42

43 Interface Estimation 0 Averaged horizontal pixel values profile D/P Cell estimation: cm Image Processing estimation: cm 0.2% 43

44 Image under Different Condition D/P Cell estimation: cm Image Processing estimation: cm 0.5% 44

45 Data Synthesizing - Field Applications 45

46 Process & Motivation Froth Slurry H Middling Tailings It is very important to correctly control the Froth/middling interface height to avoid unwanted consequences: Increasing the possibility of sanding Reducing bitumen recovery Increasing water content in Froth increase the processing load on downstream Causing environmental impact due to increased bitumen content in tailings 46

47 Process & Motivation Froth 100% Sight glass There is a camera that reads the interface level 0% Middling Accurate only if: Tailings The interface level within the sight glass There is no accumulated materials on the sight glass 47

48 Process & Motivation Froth % Profiler Sight glass There is a camera that reads the interface level 20 0% Middling 30 Accurate only if: Tailings The interface level within the sight glass There is no accumulated materials on the sight glass Therefore, a profiler has been installed to help in measuring the interface when the Camera readings are not available 48

49 Process & Motivation The profiler Two dip pipes assembly A narrow dip pipe emits low energy gamma Another dip pipe holds an array of gamma detectors Due to difference in density, each phase attenuates the signal by different amounts These signals are transmitted to DSC as density measurements 49

50 Process & Motivation Froth Middling % 0% Objective: Ensure the availability of interface level readings that is: Continuous Accurate Anywhere in the PSV. Tailings

51 Problem description & data visualisation Froth % Temperature Profiler readings Density Middling 20 0% Tailings There is no clear characteristic behavior of profiler data around the interface 51

52 Problem description & data visualisation Froth Camera readings Temperature Profiler readings Density zoom zoom Middling 52

53 Problem description & data visualisation Profiler readings Froth Camera readings Density Temperature Temperature Density zoom zoom Middling There is no clear characteristic behavior of profiler data around the interface. The majority of them move with the interface Data-based modeling 53

54 Problem description & data visualisation Profiler readings Froth Camera readings Density Temperature Temperature Density zoom zoom Middling 54

55 Method/Regression Profiler readings: Density & Temperature We choose data-based modeling technique where: Databased Model Camera readings model is built to predict interface from profiler (D & T) readings by learning from the camera as an accurate reference, by means of regression between: Profiler data X Camera readings Y 55

56 Method/Regression/PLS Profiler readings: Density & Temperature Camera readings We choose: Partial Last Squares regression (PLS) Solves the collinearity issues among the X variables PLS Avoids inverting covariance matrix compared to OLS RPLS suitable for online DCS Application Dimensionality reduction 56

57 Method/Regression/PLS Density [ g / cm 3 ] Camera readings Off-line model training/fitting 57

58 Method/Regression/PLS Density [ g / cm 3 ] Camera readings Off-line model training/fitting Model validation (Prediction) However, due to variations in process conditions, the off-line model becomes outdated 58

59 Method/Regression/Recursive PLS Choose a representative training set ( X 0, Y0 ) Calculate the covariance matrices offline T ( X X) ( T X y ) 0 0 Update it online whenever a new sample ( x t, yt ) becomes available T T T ( X X) ( X X) x x ( X T y) t t ( X T y) t 1 t 1 x T t t y t t 1 0 Forgetting factor The model will be updated with the most recent profiler & camera data and be used when the camera data are not available 59

60 Results IF at 60 [min] Density sensors Temperature sensors The RPLS prediction is able to track the reference when the camera readings are not available 60

61 Results IF at 60 [min] PLS helps in dimensionality reduction in X Density sensors Temperature sensors The RPLS algorithm allows to select the input variables that have the highest importance Reduce dimensionality 61

62 Results Density sensors Density selection sensors selection Temperature Temperature sensors selection sensors selection The RPLS algorithm allows to select the input variables that have the highest importance Reduce dimensionality 62

63 Results Corr= y ref Corr= minutes prediction 10 minutes prediction y ref Corr= minutes prediction 60 minutes prediction Corr= y ref y ref Closer we are to the current (update) time, better the interface prediction performance is 63

64 Analytics Toolboxes in Progress 64

65 Soft Sensor Analytics Data preprocess Cumbersome Resample, outlier detection, rearrange, normalize, detrend. Data modeling Complex OLS, LASSO, RR, PCA, PLS, nonlinear regression. Platform Costly MATLAB, Unscrambler. 65

66 Soft Sensor Analytics Excel SQL Matlab Data preprocess Resample Normalize Rearrange Cumbersome Outlier removal Detrend Impute missing JPEG PDF RGBA Resample, outlier detection, rearrange, normalize, Time Trenddetrend. Data visualization Histogram.pkl Data modeling OLS PCA Complex RR PLS LASSO Nonlinear Cross validation OLS, LASSO, Simulation RR, PCA, PLS, nonlinear regression. Adaptation Platform Python Costly MATLAB, Unscrambler. 66

67 Process Diagnosis Analytics - Toolbox Setting a control limit Advanced algorithms Detection Trouble shooting Abnormal Event Management Character ization Learning the category and characteristics of fault Diagnosis Finding the source of fault 67

68 Process Diagnosis Analytics - Toolbox Causal Analytics: Extracts causality relations among the variables from data Oscillation Diagnostics: Detects and characterizes oscillatory type of faults 68

69 Conclusion Data analytics is an emerging area of research and applications Great potential, demands and opportunities Applicable in every sector Opportunity for everyone 69

70 Acknowledgments Zheyuan Liu, Fadi Ibriham, Ruomu Tan, Anahita Sadeghian, Ming Ma, Elham Naghoosi NSERC Industrial Research Chair in Control of Oil Sands Processes AITF Industry Chair in Process Control University of Alberta 70