inaugural lecture Kang

PowerPoint Presentation

Data, intelligent systems and control

1Prof Kang LiEnergy, Power and Intelligent ControlSchool of Electronics, Electrical Engineering and Computer ScienceQueen's University Belfast

10/04/[email protected]

ContentResearch AreasData driven Modelling, Classification, Monitoring, and ControlInternational outreaches and industrial applicationsFuture work

2

2

3

Operation settingsPower, Voltage, FrequencyInputsPlantOutputs

Data

MonitoringSystem modellingClassificationControlOptimizationFault diagnosis1. Data Driving approaches

3

1. Data Driving approaches

A brief view of control history5System An interconnection of elements and devices for a desired purpose.Control System An interconnection of components forming a system configuration that will provide a desired response.

Process The device, plant, or system under control. The input and output relationship represents the cause-and-effect relationship of the process.

A brief view of control history6

Multivariable Control System

Open-Loop Control Systems utilize a controller or control actuator to obtain the desired response.Closed-Loop Control Systems utilizes feedback to compare the actual output to the desired output response.

A brief view of control history718th Century James Watts centrifugal governor for the speed control of a steam engine.1920s Minorsky worked on automatic controllers for steering ships.1930s Nyquist developed a method for analyzing the stability of controlled systems1940s Frequency response methods made it possible to design linear closed-loop control systems1950s Root-locus method due to Evans was fully developed1960s State space methods, optimal control, adaptive control and1980s Learning controls are begun to investigated and developed.Present and on-going research fields. Recent application of modern control theory includes such non-engineering systems such as biological, biomedical, economic and socio-economic systems


Watts Flyball Governor (18th century)Greece (BC) Float regulator mechanism Holland (16th Century) Temperature regulator


Example of modern control systemThe Utah/MIT Dextrous Robotic Hand

Put an Video here about formation flightPut an Video here about Honda ASIMO robot10A brief view of control history

A brief view of machine learning11Data Mining, web information extracting.Analysis of astronomical dataHuman Speech RecognitionHandwriting recognitionFraudulent Use of Credit CardsDrive Autonomous VehiclesPredict Stock RatesIntelligent Elevator ControlWorld champion BackgammonRobot SoccerDNA ClassificationWhere machine learning can be used.

A brief view of machine learning12IBM: From Deep Blue to wastonIBM's WatsonSupercomputer Destroys Humans inJeopardy

A brief view of machine learning13A video on google self-driving car

A brief view of machine learning14Decision tree learningArtificial neural networksNaive BayesBayesian Net structuresInstance-based learningReinforcement learningGenetic algorithmsSupport vector machinesExplanation Based LearningInductive logic programmingTechniques used in machine learning

History of data-driven approach15




19

Huge Data or Small DataInput Selection, Curse of dimensionalityModel order selectionOnline-estimation, adaptive learningModel interpretationParametric or non-parametric modellingComputational complexity, sparse modelUnder-fitting or over-fittingData Driving approaches

19

20Data Driving approachesInput selection and curse of dimensionality(Suppose each inputs has 10 operating points)

20

Solutions:Optimal Experimental DesignPrincipal component analysisInput selectionObjective: Reduce input dimension and data redundancy Data Driving approachesInput selection and curse of dimensionality

22Data Driving approachesHuge data or small data

101010011010001001001010110100010111101010101010101101101011010101101011110101010101001010100101010101010110110101101010100010101011010101001001010101010101000101010010100010110101010001010101010101010100101001011101100101010101010101010101001010101001010101010101010101010101010101010010101010101001010100101010010101010100101001010010101010100101010101010101010010101001010101001010Large data:Increased the training/learning timeHeavy computational complexityBetter estimation performanceSmall data:Not sufficient for model trainingDifficult to validateSmall computational complexityCross validation

22

23Data Driving approachesImbalanced data

Solution: Generate artificial data to balance two classes

23

24Data Driving approachesNon-parametric modellingModel Structure + Equations

Priori information on system structure is unknown

Model a system directly with its responses

24

Data Driving approachesComputational complexity, sparse model

Subset selectionForward selectionBackward eliminationStepwise selectionTwo-Stage Selection [2]Fast recursive algorithm [1][1] K. Li, J. X. Peng, and E. W. Bai, A two-stage algorithm for identification of nonlinear dynamic systems, Automatica, vol. 42, no. 7, pp. 11891197, 2006.[2] K. Li, J. X. Peng, and G. W. Irwin, A fast nonlinear model identification method, IEEE Transactions on Automatic Control, vol. 50, no. 8, pp. 12111216, 2005.

Data Driving approachesUnder-fitting or over-fitting

To monitor the performance of the processTo study historical data representing good past process behaviorProjection methods of Multivariate Data Analysis (PCA, PCR, PLS)To plot multivariate score plots to monitor the process behaviorData Driving approachesMultivariate statistical process control(MSPC)

Low dimensional projectionData Driving approachesMultivariate statistical process control(MSPC)

Original variable set

Mapping function

Nonlinear scores

Demapping function

Predicted variable set

NPCA using 5-layer ANNData Driving approachesMultivariate statistical process control(MSPC)

Input DataReal systemActual outputData driven model (nonlinear)ControllerData Driving approachesData-driven control

Why model Nox

Environmental impact Legislation

NOx reduction techniques

Primary or combustion modification-based technologies including real-time advanced control systemsSecondary or flue-gas treatment-based technologies.

Types of models ANN model Identification modelGrey-box modelCFD modelBlack-box methodGrey-boxmethodWhite-boxmethod

Data Driving approachesApplications using data driving approaches - NOx emission modelling

32Data Driving approachesApplications using data driving approaches - NOx emission modelling

32

Data Driving approachesApplications using data driving approaches - NOx emission modelling

Flame Velocity profile

Temperature profileData Driving approachesApplications using data driving approaches - NOx emission modelling

NOx formation process Data Driving approachesApplications using data driving approaches - NOx emission modelling

36Data Driving approachesApplications using data driving approaches Clinical chemical analysis

Objective: Predict Misuse of Growth PromotingHormones in Cattletotal protein,potassium, phosphate,creatinine, cholesterol, albumin, alkaline phosphatase (ALP),aspartate transaminase (AST), alanine transaminase (ALT), Gamma glutamyltransferase (GGT), lactate dehydrodenase (LDH), sodium,chloride, carbon dioxide (CO2), urea, glucose, calcium,total bilirubin, urate,corrected calciumMetabolic Markers

36

(a) The support vector machine (SVM) trained from heifer data and its decision values on the 140 training heifers across 14 different days. (b) The support vector machine trained from steer data and its decision values on the 140 training steers across 14 different days.Data Driving approachesApplications using data driving approaches Clinical chemical analysis

Sensitivity = TP/(TP+FN)Specificity = TN/(TN+FP)HeiferSteerSensitivity94.67%97.33%Specificity87.69%93.85%

Data Driving approachesApplications using data driving approaches Clinical chemical analysis

Melt pressureMelt temperatureFeed rateBarrel temperatureScrew speedViscositycontrol

Melt temperature Melt pressure Melt viscosityto Optimize

Energy usage Product qualityData Driving approachesApplications using data driving approaches Polymer extrusion monitoring

Data Driving approachesApplications using data driving approaches Polymer extrusion monitoring

Video on polymer extrusion

Soft sensor with a feedback structure.Data Driving approachesApplications using data driving approaches Polymer extrusion monitoring

MIMO PID control Linear-quadratic regulator (LQR) Model predictive control (MPC) Fuzzy control Neuro-fuzzy controlModel-basedModel-freeData Driving approachesApplications using data driving approaches Polymer extrusion monitoring

43Data Driving approachesApplications using data driving approaches Economic dispatch

Differential evolution Economic dispatch is the short-term determination of the optimal output of a number of electricity generation facilities, to meet the system load, at the lowest possible cost.Emergency reserve is required on the purpose of maintaining the reliability in case of the loss of a main unit or an increase in load.

43

Extrusion blow moldingInjection blow moldingStretch blow molding

Data Driving approachesApplications using data driving approaches ISBM

Data Driving approachesApplications using data driving approaches ISBM

Table 1: The comparison of forward and backward selectionAdvantageDisadvantageForwardFast/less computing Constrained minimizationBackwardSlow/much computing Unconstrained minimization

Forward selection method (constrained minimisation)yX1

X1 1e = y X1 1yX1

X1 = y X1 1-X2 2X2X2 2

e 1 MethodologyFRA and Two-Stage selection

12kn

j

Selected termsStage 1: Forward model selectionStage 2: Backward model refinement - Loop 1 .. - Loop 2 .. - Loop 3 .. Candidate terms pool Two-stage selectionRemains efficient and effective from FRAEliminates optimization constraint in FRAReduces the training error without increasing model sizeMethodologyFRA and Two-Stage selection

A famous tipping problem comprising of three rules (based on tipping practices in the U.S.):

Fuzzy logic expresses the system in terms of a set of human understandable rules.MethodologyFuzzy Neural Model

Table 1.1: Fuzzy components in a rule-based systemMethodologyFuzzy Neural Model

Takagi-Sugeno-Kang fuzzy system (a multi-model approach):

Number of rules

Example for single input single output system -Linguistic termsFuzzy sets Local models

MethodologyFuzzy Neural Model

FIGURE: Fuzzy inference process.MethodologyFuzzy Neural Model

51

Fuzzy Neural Networks (FNNs) =Fuzzy Inference System +Neural Networks

Rule premises Rule consequentsRule formulationMethodologyFuzzy Neural Model

MethodologyEng-genes

The PCA algorithm relies on a coordinate transformation of data matrix to produce a reduced set of principal components

is the score matrix, is the loading matrix

P is dominant eigenvectors of the covariance matrix:

correlated variablesuncorrelated variablesx

T*PT+E

MethodologyPCA

PCA allows the construction of two statistics, a Hotellings T2 statistic and a Q statistic

MethodologyPCA

The ith solutionThe updating velocity for the ith particleInertia weightThe best position found by the ith particleThe global best solutionAcceleration coefficientsRandom numbersParticle swarm optimizationMethodologyMetaheuristicalgorithms (PSO, DE, GA,etc)

Differential evolution

1. Mutation2. Crossover3. Selection

MethodologyMetaheuristicalgorithms (PSO, DE, GA,etc)

3. International outreaches and industrial applicationsCommercial MSPC monitoring software tools emerging from longstanding collaboration with DuPont (now Invista) and Annex6SUPMAX distributed networked monitoring and control systems with Shanghai Automation Instrumentation Co., Ltd (SAIC)Wired/wireless hybrid networked monitoring and control system embedded with the referenced networked control and MSPC techniques has been deployed in several hot-rolling production lines in Baosteel Group

59International outreaches and industrial applications

UK-China Science Bridge ProjectOn Sustainable Energy and Built Environment

59

3. International outreaches and industrial applicationsScience bridge Project - achievements

Academic exchanges


To embed in-house electronic software and hardware capability including processes and systems

61


Integrating energy efficiency monitoring, control and optimization for plastics industry

62

Energy monitoring and control,

ProcessControllerData-driven model (Optimization)Energy efficiencyFuture work

64Future work

Integrating EVs and renewable power to smart gridIntegrating EVs into electricity load in smart grids - smart charging technology, shifting EV charging to off-peak period. EV being used as distributed large energy storage sources for the grid.Vehicle-to-grid (V2G), Vehicle-to-building (V2B) and other technologies.

64

Questions ?

65Kang LiEPIC Research [email protected]

text

text

. . .

. . .

. . .

. . .

. . .

X

X'

tNLz

function H1

function H2

x1

xk

x'1

x'k

tNL1

Chart130204

No.Talent cultivation

Sheet1No.Full time PhD30Exchanged students20Research fellows4

Chart130412

No.

Sheet1No.Invited lectures30Research training4Honours12

Chart17549178

No.Research outputs

Sheet1No.Projects7Books5Journal papers49Conference papers178

inaugural lecture Kang

Engineering

Transcript of inaugural lecture Kang