Reliability Analysis and Risk Assessment in the Oil and Gas Industry: an outlook. Prof. Enrico Zio

Transcription

1 Reliability Analysis and Risk Assessment in the Oil and Gas Industry: an outlook

2 CONTENTS: 1. RELIABILITY ANALYSIS Maintenance => Prognostics and Health Management (PHM) Operation => Optimization 2. QUANTITATIVE RISK ASSESSMENT Modeling => Uncertainty

3 RELIABILITY ANALYSIS

4 RELIABILITY, AVAILABILITY, MAINTAINABILITY (RAM): Reliability of a component or system at time t is the probability that it survives up to time t. Availability of a component or system at time t is the probability that it is operating at time t. Maintainability of a component or system within a given period of time is the probability that it will be restored to specified conditions within the given period of time by a maintenance action

5 RAM analysis in O&G Project Lifecycle New Expl. initiative Discovery well Production Start Up Opportunity and Project Exploration System (OPES) Opportunity and Project Development System (OPDS) Opportunity and Production Operation System (OPOS) EXPLORATION & APPRAISAL DEVELOPMENT Evaluation Concept Selection Concept Definition Gate 1 Gate 2 Gate 3 Execution Hand over Production & Close Out PRODUCTION Handover Handover to Package Operations First Period Production Running Production & Improvement Gp 1 Gp 2 Running Production & Preparation to Decommissioning OPDS OPOS

6 WHY RAM analysis in Oil&Gas Support the Decision Process for Plant configuration and selection Give suggestion for configuration improvement and design review Give suggestion to select the best maintenance strategies in a global system vision Support operation and maintenance activities during the Plant lifecycle Validation of the estimated production targets Increase the Plant availability Preserve the value of assets and increase asset reliability Standardize work flows and procedures Give suggestions for spare parts stocking minimizing risks Management of RCM activities Traceability of results Support Company in cost, logistic aspects and manning level evaluations Support for on condition maintenance implementation Support the Plant Integrity management

7 HOW RAM Analysis Process description Reliability Data RAM Methodology Failure modes Failure rates and ART Maintainability Analysis (MA) Data Model Preparation Failure Mode & Effects Analysis (FMEA) Maintenance Workload Evaluation Maintenance Workload CMT PMT Criticality Analysis % Availability Analysis in terms of system Production Item Criticality Index Loss of Production RCM Approach Maintenance PLAN Recommendation Process (Design), QRA, O&M REVIEW Critical Points Plant Availability Configuration Improvements RAM Activity Phase II Overview RAM Output Optional Activity

8 MAINTENANCE

9 Maintenance Intervention Approaches Maintenance Intervention PHM Unplanned Planned Corrective Replacement or repair of failed units Scheduled Perform inspections, and possibly repairs, following a predefined schedule Condition- based Monitor the health of the system and then decide on repair actions based on the degradation level assessed Predictive Predict the Remaining Useful Life (RUL) of the system and then decide on repair actions based on the predicted RUL

10 PROGNOSTICS AND HEALTH MANAGEMENT (PHM)

11 Condition Monitoring and Prognostics and Health Management (PHM) Condition Monitoring Detection Diagnostics Prognostics Maintenance

12 Condition monitoring and PHM: what is it? PHM = DDP = Detect = Diagnose + + Predict Normal operation f 1 f 2 Remaining Useful Life

13 Ideal Maintenance T m =T f dt Component s life fully exploited Unavailability and further failures due to maintenance actions are avoided T m = maintenance time T f = failure time

14 Uncertainty and Maintenance The failure mechanisms have uncertainty associated with their occurrence in time: Inherent physical randomness of the degradation and failure processes Model used to assess the performance of the system imprecise reproduction of reality Early prediction Late prediction Failure When to perform maintenance: non-trivial decision

15 Diagnostics: fault detection and classification Prognostics: lifetime estimation f 1 Diagnostic system Normal operation f 2 FAULT DETECTION early recognition Forcing functions Measured signals f 1 Prognostic system Lifetime estimation Prognosis P FAULT CLASSIFICATION correct assignment f 2

16 FAILURE PROGNOSIS

17 Failure Prognosis Objective: residual useful life (RUL) estimation Mean Time To Failure approach: MTTF = Static, a priori (made at t=0) estimation. Prognostic approach: TTF = Dynamic estimation that evolves with time from the moment a component is firstly used until it has failed. TTF updated as more information becomes available. Residual Life Actual remaining time MTTF Time To Failure (prognostic) Current Time

18 Methods Methodology box Component Data & Information Component Health Assessment Fault Detection Diagnostics Prognostics

19 Approaches Different data & information Different approaches DATA ATA-DRIVEN MODEL-BASED Statistical distribution of failure times Degradation models o Markov models o Filtering approaches (e.g. Kalman and Particle Filtering) o Conventional numerical algorithms o o Linear regression Time series analysis Machine learning and data mining algorithms o o Artificial neural networks o Fuzzy logic systems Support vector machines o

20 DATA-DRIVEN DRIVEN EXAMPLE: FUZZY SIMILARITY

21 Fuzzy Similarity Comparison 1- Trajectory pointwise difference computation: 3- RUL i (t) estimation: n-long test trajectory pattern (Fig.1) n-long, j-th interval of the i-th treference trajectory pattern (Fig. 1) 2- Trajectory pointwise similarity computation: µ (i,j) is the membership value of the distance δ (i,j) to the condition of approximately zero 4- RUL estimation: Figure 1 Weighted sum of the RUL i, i=1,2,,n RUL =w i RUL i with i=1,2,,n

22 Results:

23 MODEL-BASED EXAMPLE: PARTICLE FILTERING

24 Model-Based Dynamic Degradation State Estimation HYPOTHESES: System model: x = hidden degradation state vector ω = i.i.d. random process noise vector f = non-linear dynamics function vector k = time step index Measurement equation: υ = i.i.d. random measurement noise vector h = non-linear measurement function vector x = f ( x, ω ) z k k k 1 k 1 Hidden Markov process = h ( x, υ ) k k k k Available Recursive estimate of the posterior pdf ( k 0: k ) p x z

25 Model-based filtering techniques Kalman Filter Exact only for linear systems and additive Gaussian noises Extended-Kalman or Gaussian-sum Analytical approximation Grid-based filters Numerical approximation (burdensome) PARTICLE FILTERING Numerical approximation which, in the limit, tends to the exact posterior pdf p ( x z ) k 0: k

26 Particle filtering: the logical sequential procedure Time step k i z k not yet collected i z k available Monte Carlo prediction of N state trajectories (= i particles) xk i = 1, K, N by importance sampling Observation Likelihood i p ( z x ) k k by importance sampling BAYES RULE Posterior (updated) distribution of the current system states p ( xk z0: k ) and failure probabilities CBM

27 Results Number of particles: 5000 Five measurements at time steps: k 1 = 100, k 2 = 200, k 3 = 300, k 4 = 400, k 5 = 400 ω k, υ k = Gaussian noises d * = 80 Crack depth evolution Expected cost per unit time 527 k 1 = k 2 = k 3 = k 4 = k 5 =500

28 PHM IN OIL&GAS: Scale Deposition (Data-Driven) Driven)

29 Scale Deposition in Oil Well Equipment Equipment: : tubulars and valves for offshore drilling Degradation mechanism: : scale deposition Degradation state: : thickness of scale deposition (x) not measurable during operation! Available information: : 32 laboratory tests (z, z,x) MEASURED SIGNALS (z) Orientation Location Roughness Initial Weight Test Temperature Test Pressure Brine Concentration Test Duration DEGRADATION STATE (x) Scale Thickness

30 Objective 4 Available: Pairs of data «measured signals-degradation state» (z, z,x) Objective: Measured signals (z) Empirical Model Degradation state x + uncertainty(x) Maintenance Planning

31 Modelling Approach: Ensemble of Neural Networks Measured signals (z) Neural Network 1 Neural Network 2 Degradation state (x) Aggregation Degradation state (x) + uncertainty P(x) x Neural Network B Different input signals Different training sets Diversity between neural networks outputs Mean, median, etc. Accuracy Uncertainty estimation (Carney et al., 1999)

32 Sensitivity Analysis Which signals have an influence on the scale thickness? Basic Idea Output Influence Output No Influence Large Separation Small Variance Small values Large values Input Signal Small values Large values Input Signal Classification Tree 32 data Brine 28 4 Pressure Orientation most influencing signals: Brine Pressure Orientation

33 Empirical Modelling: Ensemble of Neural Networks Input Signals 3 most influencing measured signals 5 most influencing measured signals Neural Network 1 Neural Network 100 Neural Network 101 Neural Network 200 Median Scale Thickness (x) All measured signals Neural Network 201 Neural Network 300

34 Model performance

35 PHM IN OIL&GAS: Example Choke Valve Erosion (Data-driven+Model-based)

36 Choke valve erosion: Health Assessment Component: choke valve Degradation mechanism: erosion Health indicator: valve flow coefficient (C v ) Not directly measurable Estimation WELL TESTS provide measures of C v C v measure

37 Choke valve erosion: Physics-based estimation Pressure drop ( P) Opening (θ) Oil flow rate (ṁ o ) Water flow rate (ṁ w ) Gas flow rate (ṁ g ) C v = Physics based model from fluid-dynamicsdynamics 27.3 F p m& P m& g m& ρ g J 2 m& + ρ w w m& o + ρ o C v estimate C v estimate - C v measure ERROR = Health indicator Number of well test

38 1 Choke valve erosion: Data-driven estimation C v measure used to train empirical models Health indicator Pressure drop ( P) Opening (θ) Oil flow rate (ṁ o ) Water flow rate (ṁ w ) Gas flow rate (ṁ g g) ) Group 1. Group 4 Parameters subsets Number of well test Model 1. Model 4 Kernel Regression Models Ensemble approach: increase robustness and accuracy Estimate 1. Estimate 4 Local Fusion C v estimate C v estimate C v estimate ERROR = ERROR = 4.441

39 Choke valve erosion: Data-driven estimation ENSEMBLE APPROACH: aggregation of physics-based and data-driven driven predictions Pressure drop ( P) Opening (θ) Oil flow rate (ṁ o ) Water flow rate (ṁ w ) Gas flow rate (ṁ g ) Physics based model Data-driven ensemble C v estimate ERROR = C v estimate ERROR = C v estimate ERROR = Local Fusion Hybrid Ensemble

40 OPTIMIZATION OF OPERATION (by EVOLUTIONARY METHODS)

41 Genetic Algorithms: the life cycle 2) Crossover Genes * random * bit coding 3) Evaluate chromosome performance * estimate fitness 1) Select parents * random * fitness Selected Parents Child fitness known 4) Promotion to the future * constant population size * Fittest favoured, weakest mistreated Population 5) Go to next iteration if optimum not found satisfactorily next Population 4

42 Experimental Example of Genetic Evolution

43 Parents selection

44

45

46 Differential Evolution

47 Manipulation for DE: Mutation Mutation: for each vector x i (target vector) a noisy vector v i is created choosing three mutually different vectors (r 1, r 2, r 3 ) from the population. The basis vector is perturbed by the weighted difference between the other two (step length). v i =x r1 +F*(x r2 -x r3 ) The weight F takes a value in (0,1] Using the weighted difference, the exploration becomes self-organized: the step length adjusts itself following the evolution of the vector population in the successive generations

48 Manipulation for DE: Crossover Crossover : a trial vector u i is created mixing the target vector x i and the noisy vector v i. A modified Bernoulli trial is performed on every element of the vector: u ij =v ij if U(0,1)<CR u ij =x ij otherwise Crossover Rate CR є[0,1] controls the inheritance of target vector into trial vector Avoid drastic change in the population Prevent premature stagnation Maintain some good properties from current population

49 Manipulation for DE: Selection Selection: the trial vector u i is compared with the target vector x i. The fittest one takes place in the population for the next generation. This is the only selection procedure used in DE The next generation will be better or at least equal to the previous one.

50 DE improvements with respect to GA self-organized: the vector differences enhance the flexibility in any domain dimension few parameters: NP, F and CR more reliable: the crossover procedure is particularly robust (CR<0.5) faster the exploration by weighted difference is efficient

51 DE application oil&gas real case Real case Oil&Gas industry Problem statement: optimal system variables setting for maximum production (SO) from an integrated hydrocarbon production asset. Integrated asset = two production environments (simulated by different programs): 1. The gathering system (GAP) 2. The process plant (HYSYS) Optimization particularly challenging: Several variables (14) Several constraints (15) Separate optimization of the two systems does not achieve satisfactory optimum

52 DE application to oil&gas real case 14 Variables: 8 inlet choke P 2 separators P P slug-catcher/1 separator Stabilizer head P Stabilizer T reboiler Stabilizer middle T 15 constraints: 8 FBHP Oil, Gas and Water entering the plant volume flow to the treating section CO 2 /H 2 S ratio Wobbe index Oil TVP

53 DE application to real case The integrated optimization by DE is compared with 2 classic separated optimizations: 1. network optimization with gas constraint (into the plant) and successive process optimization 2. network optimization with gas constraint + minfbhp and successive process optimization Oil standard conditions bbl/day GAP opt. Gas rate GAP opt gas + FBHP Integrated opt. Some constraints are not satisfied Cputime ~ 1 h Satisfaction of constraints but low productivity Cputime ~ 1 h Satisfaction of constraints and higher productivity Cputime ~ 4.5 h

54 QUANTITATIVE RISK ASSESSMENT

55 WHY QRA in Oil&Gas ACCIDENTS in oil and gas Onshore/Offshore E&P activities LOSS OF LIVES IMPACTS ON THE ENVIRONMENT LOSS OF ASSETS - ECONOMICAL LOSS LOSS OF REPUTATION Risks need to be managed to rationally address the health and safety aspects of the activities undertaken by the Company.

56 QUANTITATIVE RISK ASSESSMENT What undesired conditions may occur? Accident Scenario, A 2. What damage do they cause? Consequence, C 3. What is the likelihood of occurrence? Likelihood, l (U) RISK = (A, C, U)

57 QUANTITATIVE RISK ASSESSMENT 57 Alternative 1 Alternative 2 - Design configuration 1 - Redundancy allocation 1 - Evacuation plan 1 VS - Design configuration 2 - Redundancy allocation 2 - Evacuation plan 2 A RISK 1 RISK 2 C: How many fatalities C 1? U: What is the likelihood of having C 1 fatalities or more? C: How many fatalities C 2? U: What is the likelihood of having C 2 fatalities or more?

58 QRA in the O&G Project Life Cycle New Expl. initiative Discovery well Production Start Up Opportunity and Project Exploration System (OPES) Opportunity and Project Development System (OPDS) Opportunity and Production Operation System (OPOS) EXPLORATION & APPRAISAL DEVELOPMENT Evaluation Concept Selection Concept Definition Gate 1 Gate 2 Gate 3 Execution Hand over Production & Close Out PRODUCTION Handover Package Handover to Operations First Period Production Running Production & Improvement Gp 1 Gp 2 Running Production & Preparation to Decommissioning OPDS OPOS

59 QRA in the Development Phases Project Development Phases Evaluation G1 Concept Selection G2 Concept Definition G3 Execution Hand over Production Design Phases Pre-feasibility & Feasibility Studies Basic Design FEED Detail Design Revamping Regulations Execution Implementation Risk Analyses Approach HAZID Studies Blowout Study Preliminary Fire and Explosion (F&E) Analyses Preliminary Dispersion Analyses HAZID Studies Preliminary F&E Analysis Preliminary QRA HAZID Studies HAZOP Studies F&E Analysis Smoke & Dispersion Analysis TR Impairment QRA Escape, Evacuation and Rescue Analysis (EERA) SIMOPS Risk Evaluation Cost-Benefit Analysis (CBA) HAZOP Studies F&E Analysis Smoke & Dispersion Analysis QRA EERA SIMOPS Risk Evaluation Safety Integrity Level (SIL) Assessment HAZID & HAZOP for Revamping F&E Analysis for Implementation F&E Analysis for Regulation Execution Smoke & Dispersion Analysis for Regulation Execution QRA for Regulation Execution EERA for Regulation Execution

60 MODELING FOR QRA

61 (aleatory and epistemic) Uncertainty in QRA ALEATORY Initiator Event (IE) 1 p 1 Event 1: Shut-down valve 1 p 2 p 2 Event 2: Emergency and evacuation procedure {S 1, p S1, c 1 } {S 2, p S2, c 2 } p 1 1 p 2 {S 3, p S3, c 3 } EPISTEMIC p 2 {S 4, p S4, c 4 } Aleatory: variability, randomness (in occurrence of the events in the scenarios) Epistemic: lack of knowledge/information (on the values of the parameters of the probability and consequence models)

62 QRA Modelling 62 valve 1 X t 1 t INFORMAT TION AVAILABLE valve 2 valve N λ is between 10-3 and 10-2 [h - 1 ] λ is quite small X t 2 t N X t t SYSTEM RISK MODEL (UNCERTAIN) RISK MEASURES (A,C,U,M,K),K) REPRESENTATION OF UNCERTAINTY M UNCERTAINTY PROPAGATION

63 PRA Modelling 63 INFORMATI ION AVAILABLE valve 1 valve 2 valve N f T (t, λ) probability density function E r t λ is UNIFORM between 10-3 and 10-2 [h - 1 ] λ is less than 10-2 [h - 1 ] with probability 0.9 SYSTEM RISK MODEL f z (Z) (PROBABILISTIC) RISK MEASURES (A,C,U,P,K),K) PROBABILISTIC REPRESENTATION OF UNCERTAINTY (M=P) UNCERTAINTY PROPAGATION

64 UNCERTAINTY IN QUANTITATIVE RISK ASSESSMENT

65 Uncertainty representation UNCERTAINTY Types: Randomness (aleatory, e.g.: maximal flow, Q) Imprecision (epistemic, e.g.: valve coefficient, Cv) Representation: Probability distributions Probability distributions (Sufficient statistical information, i.e., data) Possibility distributions (Scarce information) Example: f Q (q) 7 8 x f Cv Cv 0.06 (C v ) 0.05 π Cv Cv (C v ) q C v C v Gumbel pdf Gaussian pdf The value of C v is between 5 and 60, with a preference for 30

66 Faithful representation of information and knowledge

67 Things I know: Information-based bounds cdf Do not add knowledge that is not included in the available information

68 Things I know: (expert) knowledge-based bounds cdf Do add expert knowledge when reliable

69 Things I know: (expert) knowledge-based bounds cdf Do add expert knowledge when reliable

70 CONCLUDING REMARKS

71 (Outlook) Points to retain-general 1. RELIABILITY ANALYSIS Maintenance: Methods of PHM, depending on information available (data-driven, driven, model- based) Operation: Methods for combinatorial optimization (evolutionary algorithms like GA, DE,...) 2. QUANTITATIVE RISK ASSESSMENT Methods for adequate representation of uncertainty, depending on information available (probability, possibility,...)

72 (Outlook) Points to retain- together 1. RESEARCH Projects collaboration 2. EDUCATION Master students WELCOME Continuing education course (Polimi-4 days): Advanced techniques for RAMS (Monte Carlo simulation, Neural Networks, Genetic Algorithms etc.) Specialized Oil&GAS courses (1-week) (ENIcorporateUniv) RAMS QRA MAINTENANCE ENGINEERING

73 (Outlook) Points to retain- together 1. RESEARCH+EDUCATION PhD students WELCOME (Polimi-deadline April)