Modern Reliability and Maintenance Engineering for Modern Industry

Transcription

1 Modern Reliability and Maintenance Engineering for Modern Industry Piero Baraldi, Francesco Di Maio, Enrico Zio Politecnico di Milano, Department of Energy, Italy ARAMIS Srl, Italy

2 LOW SMART KID POSITIONING HIGH ARAMIS & LASAR : the identity card Advanced Reliability Availability & Maintenance for Industries and Services History Main Intervention Areas Research & Innovation consulting approach Large amount of deployed algorithms (more than 20yrs R&D) Approach University Mission Industry Knowledge Information Data Simulation, Modeling, Analysis, Research for Treasuring Knowledge, Information and Data RESEARCH INNOVATION MARKET DELIVERY

3 The Team LASAR Members (3): Piero Baraldi (Associate Professor, PhD) Francesco Di Maio (Assistant Professor, PhD) Enrico Zio (Full Professor, PhD) LASAR Post-Docs (3): Michele Compare (PhD) Rodrigo Mena (PhD) Sameer Al Dahidi (PhD) LASAR PhD students (7): Marco Rigamonti Francesco Cannarile Wei Wang Zhe Yang Shiyu Chen Alessandro Mancuso Edoardo Tosoni ARAMIS Members (7): Enrico Zio (President) Michele Compare (Partner, CEO) Piero Baraldi (Partner) Francesco Di Maio (Partner) Giovanni Sansavini (Partner) Francesco Cannarile (Research consultant) Rodrigo Mena (Research consultant) + Network of senior experts on specific technical areas + MSc students (10-12/y) + Visiting (4-6/y)

4 Industry

5 Problem statement Failures Prevented by Design for Reliability Maintenance Normal Degraded Failure Time

6 Modern Reliability Engineering (SMART) Reliability Engineering

7 Modern (smart) Reliability Engineering The Big KID

8 Big Knowledge(ID) Real world problem formulation Mathematical problem Real world solution interpretation Mathematical solution

9 Big (K)Information(D)

10 Big (KI)Data

11 Can the Big KID become SMART for Reliability Engineering?

12 Problem statement Failures Prevented by Design for Reliability Maintenance Normal Degraded Failure Time

13 SMART Reliability Engineering Big KID opportunities Reliability analysis for Design for Reliability: From (binary) failure modeling to degradation-to-failure modeling

14 Failure modelling (binary) Failure ON OFF As Good As New Failed X(t) 100% 0% t System unavailability U(t) = Pr[X (t) <100%] U(t) = Pr[X (t) =0%]

15 Degradation-to-failure modeling Multi-state: Failure Failure Mode 1 Degradation state 1 Degradation state n 1 ON OFF Failure Mode M Degradation state 1 Degradation state n M X(t) 100% 75% 50% 25% 0% t Demand of system performance System unavailability D(t) U(D,t) = Pr[X (t) < D(t)]

16 Degradation-to-failure modeling Multi-state: Failure Failure Mode 1 Degradation state 1 Degradation state n 1 ON OFF Failure Mode M Degradation state 1 Degradation state n M How to represent and model the item behavior? Integrating physics-of-failure knowledge in reliability models Multi-State Physic-Based Models

17 Reliability SMART Reliability Engineering Challenges Alloy 82/182 dissimilar metal weld of piping in a PWR primary coolant system Reliability? KID (Knowledge, Information, Data) Model Year Highly reliable Sufficient failure data Physics knowledge Expert judgment Field data Statistical models of time to failure Stochastic process models Physics-based models Multi-state models

18 SMART Reliability Engineering Multi-State Physic-Based Models Alloy 82/182 dissimilar metal weld of piping in a PWR primary coolant system Physical laws Multi-state physics model of crack development in Alloy 82/182 dissimilar metal weld

19 SMART Reliability Engineering Opportunities 1) Random shocks Degradation process Random shock process 2) Dependences in degradation processes Valve Internal leak Pump Initial state Pump Failure state λ 32 λ 21 λ

20 SMART Reliability Engineering Opportunities 20 3) Maintenance Degradation process a b Preventive maintenance (a) Corrective maintenance (b) 4) Uncertainty Valve Internal leak Uncertain parameters in degradation models Pump Initial state Pump Failure state λ 32 λ 21 λ

21 SMART Reliability Engineering Challenges 21 Valve Internal leak Pump Initial state Pump Failure state λ 32 λ 21 λ Piecewise-deterministic Markov process (PDMP) Finite-volume integration schemes Monte Carlo (MC) simulation

22 Problem statement Failures Prevented by Design for Reliability Maintenance Normal Degraded Failure Time

23 Maintenance Engineering Objectives Optimization of the maintenance of production assets: Maximize reliability (R) Maximize production availability (A) Minimize personnel, material, inventory costs Maximize efficiency/effectiveness of maintenance interventions (M) Trade-off internal/external maintenance efforts (M) Fulfill safety/regulatory constraints (S) Fulfill budgetary constraints (C) In other words RAMS(+C) for BETTER PERFORMANCES WITH LOWER COSTS

24 How to achieve the objectives Integrated maintenance process, supported and informed by knowledge of components/systems/process behaviors through high-quality data, real-time information effective models and methods to process the information effective organizational processes to implement the solutions KID + Intelligence

25 SMART Maintenance Engineering Big KID opportunities Maintenance: Integrating physics knowledge and data: Prognostics and Health Management (PHM)

26 Maintenance Intervention Approaches Modelling is in support to proper maintenance planning and Maintenance Intervention PHM Unplanned Planned Corrective Replacement or repair of failed units Scheduled Perform inspections, and possibly repairs, following a predefined schedule Conditionbased 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

27 Maintenance Maintenance Corrective Maintenance Planned Periodic Maintenance Condition Based Maintenance (CBM) Predictive Maintenance (PrM) Prognostics and Health Management (PHM) PHM is fostered by advancements in: Sensor Algorithm Computation power

28 PHM for what? PHM in support to CBM and PrM Fault Normal Conditions Vibration Detection Abnormal Conditions Equipment Temperature t t Fault Diagnostics Anomaly of Type 1 Anomaly of Type 2 Anomaly of Type 3 Decision Maker No Maintenance Maintenance Maintenance Decision Sensors measurements Fault Prognostics Remaining Useful Life (RUL)

29 PHM: why? (Industry) Increase maintainability, availability, safety, operating performance and productivity Reduce downtime, number and severity of failure and life-time cost

30 PHM: why? (Business) Improve cash flow, profit stream and utilization of assets Guarantee long term business Increase market share

31 PHM: how? (Fault detection) Feedwater Pump Measured Signals Signal Reconstructions x 1 x 1 x 2 x r t t MODEL OF EQUIPMENT BEHAVIOR IN NORMAL CONDITION x 2 x r t t t MEASUREMENTS = RECONSTRUCTIONS t NORMAL OPERATION

32 PHM: how? (Fault detection) Feedwater Pump Measured Signals Signal Reconstructions x 1 x 1 x 2 x r t t MODEL OF EQUIPMENT BEHAVIOR IN NORMAL CONDITION x 2 x r t t t t MEASUREMENTS RECONSTRUCTIONS ANOMALOUS OPERATION P. Baraldi, F. Di Maio, L. Pappaglione, E. Zio, R. Seraoui, Condition Monitoring of Electrical Power Plant Components During Operational Transients, Proceedings of the Institution of Mechanical Engineers, Part O, Journal of Risk and Reliability, 226(6) , Baraldi, P., Canesi, R., Zio, E., Seraoui, R., Chevalier, R. Genetic algorithm-based wrapper approach for grouping condition monitoring signals of nuclear power plant components (2011) Integrated Computer-Aided Engineering, 18 (3), pp

33 PHM: how? (Fault diagnostics) Signal measurements representative of the fault classes: «c 1,c 2, c n, class» BWR feedwater system (Swedish Forsmark-3) Measured Signals Fault Types (classes) CLASSIFIER c 1 c 2 c n Baraldi, P., Razavi-Far, R., Zio, E., Classifier-ensemble incremental-learning procedure for nuclear transient identification at different operational conditions, (2011) Reliability Engineering and System Safety, 96 (4), pp F. Di Maio, J. Hu, P. Tse, K. Tsui, E. Zio, M. Pecht, Ensemble-approaches for clustering health status of oil sand pumps, Expert Systems with Applications, Volume 39, pp , doi: /j.eswa

34 PHM: how? (Fault prognostics) Kalman Filter Particle filter Model Based Monte Carlo Simulation Autoregressive (AR) models Similarity-based methods Data- Driven Hidden Semi-Markov Models Artificial Neural Networks Neuro-fuzzy systems Physics-based model of the degradation process Measurement equation Current degradation trajectory A threshold of failure External/operational conditions Degrading component Degradation trajectories of similar components Life durations of a set of similar components Similar components

35 PHM: how? (Fault prognostics) Rotating machinery (e.g. pump) t Health Index FAILURE THRESHOLD t Prognostic model Health index prediction t p t f t t p t RUL ˆ Baraldi, P., Cadini, F., Mangili, F., Zio, E. Model-based and data-driven prognostics under different available information (2013) Probabilistic Engineering Mechanics, 32, pp E. Zio, F. Di Maio, A Data-Driven Fuzzy Approach for Predicting the Remaining Useful Life in Dynamic Failure Scenarios of a Nuclear Power Plant, Reliability Engineering and System Safety, RESS, /j.ress , F. Di Maio, K.L. Tsui, E. Zio, Combining Relevance Vector Machines and Exponential Regression for Bearing RUL estimation, Mechanical Systems and Signal Processing, Mechanical Systems and Signal Processing, 31, , 2012.

36 PHM: performance? Accuracy

37 PHM: performance? (detection) Accuracy Fault Detection: Low rate of False Alarms Low rate of Missing Alarms x 1 x 2 x r Detection Model Normal Condition False Alarm Rates Missing Alarm Rates 0.54% 0.98%

38 PHM: performance? (diagnostics) Accuracy Fault diagnostics: Low Misclassification rate Signals C 1 Diagnostic Model C 2 C 3 o = true = diagnostic model Misclassification rate = 2.58%

39 RUL(t) PHM: performance? (prognostics) Accuracy Prognostics Predicted KF RUL Single Model True RUL Prognostic model time [Days]

40 4 0 FAULT DETECTION

41 Application: NPP feedwater pumps Reactor Coolant Pumps of a PWR NPP x4 [ P. Baraldi, R. Canesi, E. Zio, R. Seraoui, R. Chevalier, "Generic algorithm-based wrapper approach for grouping condition monitoring signal of nuclear power plant components. Integrated Computer-Aided Engineering, Vol. 18 (3), pp , 2011

42 PHM: how? (Fault detection) Feedwater Pump Measured Signals Signal Reconstructions x 1 x 1 x 2 x r t t MODEL OF EQUIPMENT BEHAVIOR IN NORMAL CONDITION x 2 x r t t t t MEASUREMENTS RECONSTRUCTIONS ANOMALOUS OPERATION

43 The Auto-Associative Kernel Regression (AAKR) Empirical modeling refers to any kind of (computer) approximated modeling based on empirical observations rather than on mathematically describable relationships of the system modeled 1. Collect data measurements representative of the system behavior 2. Develop the empirical model using as training set the collected data measurements test measurements measurements test measurements reconstruction

44 Available information Historical measurements of 94 stationary signals during 1 year of operation EDF experts have considered 48 of the 94 measured signals as the most important for the component condition monitoring

45 Signal Grouping AUTO-ASSOCIATIVE MODEL OF PLANT BEHAVIOR IN NORMAL CONDITION MEASURED SIGNALS COMPARISON MODEL 1 MODEL 2 MODEL m

46 Challenge: Optimal Grouping Requirement: each signal should belong to only 1 group MEASURED SIGNALS COMPARISON HOW TO GROUP SIGNALS? MODEL 1 MODEL 2 MODEL m

47 Approaches to signal grouping HOW TO SPLIT THE SIGNALS INTO SUBGROUPS? A-PRIORI CRITERIA (signals divided on the basis of ) physical homogeneity location in the power plant correlation others

48 Results BEST A-PRIORI GROUPING: correlation signals WHITE: high correlation (0.7 1] 33 GREY: medium correlation ( ] BLACK: low correlation [0 0.4] 11 signal belonging to group 1 signal belonging to group 2 signal belonging to group 3 signal belonging to group 4 signal belonging to group 5 4 signals CRITICALITY: how to group the signals which have a low degree of correlation with all the others?

49 Approaches to Signal Grouping HOW TO SPLIT THE SIGNALS INTO SUBGROUPS? A-PRIORI CRITERIA (signals divided on the basis of ) location in the power plant correlation physical homogeneity others WRAPPER APPROACH n SIGNALS SEARCH ENGINE SEARCH ALGORITHM CANDIDATE GROUPS AAKR PERFORMANCE EVALUATION OPTIMAL GROUPING

50 The Wrapper Approach: Genetic Algorithms SEARCH ENGINE Genetic Algorithms CHROMOSOME n genes = n signals GENE signal 1 signal 2 signal 3 signal i signal n GROUP LABEL (integer number) PERFORMANCE EVALUATION fitness = Accuracy

51 Results OBTAINED GROUPINGS correlation GA MAIN DIFFERENCE: signals which have a low degree of correlation are divided in groups and/or mixed with signals with a high degree of correlation

52 Results: Wrapper Approach AUTO-ASSOCIATIVE MODEL OF PLANT BEHAVIOR IN NORMAL CONDITIONS ŝ 1 t COMPARISON MEASURED SIGNALS s 1 s 1 ŝ 1 (SEAL OUTCOMING FLOW) t DECISION t ABNORMAL CONDITION: seal deterioration NORMAL CONDITION ABNORMAL CONDITION

53 FAULT CLASSIFICATION

54 Signal Value Application: NPP turbine AVAILABLE DATA: Number of transients: N=115 Transient time length: 4500t Number of vibration signals: K= Time

55 Objective: Unsupervised Clustering for Fault Diagnosis Feature 2 The problem examples are unlabeled Scope Find hidden structure in data Feature 1

56 Objective: Unsupervised Clustering for Fault Diagnosis Feature 2 The problem examples are unlabeled there is no error or reward to evaluate a potential solution v 1 1 v 2 2 unsupervised clustering 3 v 3 Feature 1

57 Methodology Feature 2 Feature 2 Feature 2 STEP 1: feature extraction Signal 1 Signal 2 time STEP 2: unsupervised clustering Fault class 1 v 1 Feature 1 v 2 Fault class 2 Fault class 3 v 3 Feature 1 Feature 1

58 5 8 Feature Extraction: Fuzzy Similarity Analysis (1-D) 1- Transients pointwise difference computation: x(t) i-th transient T 2 i j i, j x t x t i 1,2,..., N j 1,2,..., N t i j-th transient Time (t) j-th transient i-th transient 2- Transients pointwise similarity computation: 3- Similarity Matrix W definition: N columns N rows 1 1,2... 1, N 2,1 1 2, ,2... N 1, N N,1... N, N 1 1 μ (i,j) is the membership value of the distance δ (i,j) to the condition of approximately zero Similarity between transients 2 and 3

59 Feature Extraction: Spectral Analysis 1- Similarity matrix W Fully connected graph G(v i,e ij ) 1 1,2... 1,9 2,1 1 2, ,2... 8,9 9,1... 9,8 1 v i =transient i eij i, j v j =transient j

60 Spectral Analysis: the scope Identify the most appropriate features for - partitioning the graph G(v i,e ij ) Similarities

61 Spectral Analysis: the scope Identify the most appropriate features for - partitioning the graph G(v i,e ij ) - shuffling matrix W to highlight blocks of similarity Clusters

62 6 2 Spectral Analysis: The procedure for feature selection 2- Compute the degree matrix D: i, j 3- Compute the Normalized Laplacian matrix L sym : 1/ 2 1/ 2 1/ 2 1/ 2 Lsym D LD I D W D d ii N j 1 4- Compute the first C eigenvalues of L sym 1, 2,..., C ( are very small, but is relatively large) C 1 5- Compute the first C eigenvectors u of L sym Patterns to be clustered u, i 1,..., N i u 1 u C N u 2 u 1 u 2 u C New Features

63 Values of the eigenvalues of Lrw Eigenvalues Plot 1 76 eigenvalues obtained by applying spectral analysis method to the 76*76 similarity matrix Eigenvalue number 3 Clusters

64 Cluster Number Result discussion The obtained clusters are representative of different operational conditions /maintenance actions done on the component (i.e., chronological order of the transients): Outliers Missings C1:Assigned C1:NotAssigned C2:Assigned C3:Assigned C3:NotAssigned Transients in Chornological Order

65 6 5 Results: Spectral clustering

66 FAULT PROGNOSTICS

67 Diathermic oil temperature Case study: LBE-XADS The model of the LBE-XADS has been embedded within an MC-driven fault injection engine to sample component failures. High-temperature failure mode Upper threshold safe Low-temperature failure mode Lower threshold 4 control/actuator faults 64 accident scenarios

68 Data-driven Fuzzy Similarity Approach Similarity-based approach for prognostics of the available Remaining Useful Life Library of reference trajectory patterns Data from failure dynamic scenarios of the system Fuzzy-similarity comparison New developing accidental scenario prediction Remaining Useful Life (RUL)

69 Methodology: Fuzzy Similarity Analysis 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: 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 μ (i,j) is the membership value of the distance δ (i,j) to the condition of approximately zero

70 Results

71 Modern Reliability and Maintenance Engineering for Modern Industry

72 Tools for a SMART Reliability and Maintenance Engineering for Modern Industry FTA ETA FMECA Hazop Bayesian Belief Networks Process and Stochastic Flowgraphs Neural Networks Monte Carlo Simulation Petri Nets Graph Theory Fuzzy Logic Systems Optimization Algorithms Clustering Algorithms Complex Network Theory

73