monitorowanie i diagnostyka w systemach sterowania - przykłady... · lecture notes in computer...

Monitorowanie i Diagnostyka w Systemach Sterowania

Wydział Elektrotechniki i Automatyki

Katedra Inżynierii Systemów Sterowania

Dr inż. Michał Grochowski

Politechnika GdańskaWydział Elektrotechniki i Automatyki

Katedra Inżynierii i Systemów Sterowania

Przykłady aplikacji diagnostycznychna podstawie:

Swędrowski L., Duzinkiewicz K., Grochowski M., Rutkowski T. Use of neural networks in diagnostics of rolling-element bearingof the induction motor. V Międzynarodowy Kongres Diagnostyki Technicznej, 2012 Kraków

Nowicki i Grochowski. Kernel PCA In Application to Leakage Detection In Drinking Water Distribution System. Lecture Notes in Computer Science. ICCCI 2011, Part I, LNCS 6922, pp.497-506. Prezentacja na ICCCI 2011 - Gdynia;

Cyra M. i Kamowski D. Praca dyplomowa inżynierska: Detekcja uszkodzeń obiektów przemysłowych przy użyciu sztucznych sieci neuronowych., Gdańsk 2011.

Borowa A., Mazur K., Grochowski M., Brdyś M.A., Jezior K.MultiRegional PCA for leakage detection and localisation in DWDS – an approach. 8th Conference on DIAGNOSTICS OF PROCESSES AND SYSTEMS - DPS’2007 - Słubice, Poland

Opracował: dr inż. Michał [email protected]

[email protected]


na studiach II stopnia specjalności: Systemy Sterowania i Podejmowania Decyzji



Wprowadzenie

Przykłady:

• Wykrywanie uszkodzeń łożysk tocznych w silniku;

• Wykrywanie uszkodzeń w silniku obcowzbudnym

(wprowadzenie do laboratorium);

• Wykrywanie wycieków w sieci dystrybucji wody pitnej

(wprowadzenie do laboratorium);

MiDwSS W11: Przykłady aplikacji diagnostycznych

V Międzynarodowy Kongres Diagnostyki Technicznej3 – 5 września, 2012 Kraków

Gdansk University of TechnologyFaculty of Electrical and Control Engineering

Use of neural networks in diagnostics of rolling-element bearing of the induction motor

Authors:

Leon Swędrowski

Kazimierz Duzinkiewicz

Michał Grochowski

Tomasz Rutkowski


Types of induction motor faults

Introduction

bearings faults 41%

stator faults 37%

rotor faults 10%

others faults 12%


V Międzynarodowy Kongres Diagnostyki Technicznej, 3 – 5 września, 2012 Kraków


Introduction

Examples of the real bearing faults




The methods used in order to diagnose the damage in induction motors are based on various physical phenomena: vibrations, temperature, acoustic and current based;

These techniques require installing expensive sensors, especially for smaller induction motors;

Techniques which are based only on the analysis of the supplying currents are not expensive and are particularly useful when it is impossible to install diagnostic devices directly on the induction motor.

Introduction




Relation between bearing fauts and stator currents:

Theoretically, while damaged elements of bearing contacts, it results in periodic intensive mechanical vibration in the machine;

These mechanical vibrations result in air-gap eccentricity, hence oscillations in air-gap length induce variations in the flux density;

This leads to corresponding changes of the stator currents -additional harmonics of the stator currents are produced.

Through detailed analysis (e.g. spectral) of the changes in these currents one can find the induction motor fault and even distinguish its kind

Introduction




Typically, the current signals diagnostics are based on one of the two models:

model based on air gap eccentricity,

model based on torque variations.

As an alternative to model-based techniques, data-driven soft computing techniques related to the artificial intelligence techniques (neural network, fuzzy logic, genetic algorithms and various hybrid systems) might be used.

Introduction




The currents Clark vector components i and i are functions of thestator (supply) currents ia, ib, ic and are stationary according to the stator:

Clark transform approach




Under an ideal condition (sinusoidal supply currents and healthy motor), three-phase currents lead to the Clark vector with the following components:

I - maximum value of the supply phase current,

f - supply frequency,

t - a time.





The Clark vector curve i = f(i) for „ideal condition”


-3 -2 -1 0 1 2 3-3

-2

-1

0

1

2

3

i = f(i

)

i [A]

i [

A]




Examples of typical stator (supply) currents trajectories ia(t), ib(t), ic(t) for „laboratory condition”


0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04-2

0

2

ia(t)

i a [

A]

t [sec]

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04-2

0

2

ib(t)

i b [

A]

t [sec]

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04-2

0

2

ic(t)

i c [

A]

t [sec]




Examples of the Clark vector trajectories i(t) and i(t) for „laboratorycondition”


0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

-2

-1

0

1

2

i(t)

i [

A]

t [sec]

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

-2

-1

0

1

2

i(t)

i [

A]

t [sec]




Examples of the Clark vector curve i = f(i) for healthy bearing and „laboratory condition”


-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

Clark vector curve i = f(i

)

i [A]

i [

A]




• The Clark vector curve has different from a circle shape, as a result of not exactly sinusoidal supply currents or faulty motor,

• In the case of faulty motor it is a result of the additional harmonics presence in the stator current, generated by the bearing fault,

• The curve i = f(i) is a very simple pattern, that allows for detection of abnormal bearing condition by monitoring the deviations of Clark vectors patterns from the reference pattern.





Examples of Clark vector curve i = f(i) : outer ring defect

Clark transform approach - Bearing faults

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5


)

i [A]

i [

A]

• healthy bearing (blue line)• outer ring defect (red line)



-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5


)

i [A]

i [

A]


Examples of Clark vector curve i = f(i) : inner ring defect


• healthy bearing (blue line)• inner ring defect (red line)



-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5


)

i [A]

i [

A]


Examples of Clark vector curve i = f(i) : ball ring defect


• healthy bearing (blue line)• ball defect (red line)




Neural Network Approach

Proposed fault diagnosis system

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5


)

i [A]

i [

A] Artificial Neural

Network

Healthy

Faulty

Inner ring fault

Outer ring fault

Ballfault



Clark vectors


Neural diagnostic system structure


induction motor

input data(Clark vector components)

output data(fault indicator <-1, 1>)

neural model 0 healthy engine

neural model 1 outer ring fault

neural model 2

neural model 3

inner ring fault

ball fault




Neural diagnostic system structure


-1

-1

1

-1

Diagnosis

induction motor

input data(Clark vector components)

output data(fault indicator <-1, 1>)

neural model 0 healthy engine

neural model 1 outer ring fault

neural model 2

neural model 3

inner ring fault

ball fault




Description of the experimental setup

Laboratory experiments

Induction motor

Laser system for precise alignment of the shafts

Magnetic coupling

Separate base of motor mounted on the absorbers of vibrations

Electromagnetic load




Description of the experimental setup

The induction motor utilised during the test is STg80X-4C:Pn = 1,1 kW, Un = 380 V, nn = 1400 RPM, In = 2,9 A

The motor was supplied directly from three phase supply network, with the voltage of 400 V and frequency 50 Hz

The bearings installed in the motor were the 6204 type, the bearings were artificially deteriorated

Three kinds of bearing faults were tested: outer ring fault inner ring fault and the ball spot





Laboratory experiments - artificially introduced bearing faults

Outer ring fault

Inner ring fault

Ball fault




Artificially introduced bearing faults

Three kinds of bearing faults were tested:

outer ring fault: longitudinal scratch (lengths: 3mm, 6mm; depths: 0.5mm, 0.7mm, 1 mm; width: 1mm),

cross scratch (lengths: 3mm; depths: 0.5mm, 0.7mm, 1 mm; width: 1mm),

spot (diameters : 1mm, 1.5mm, 2mm; depths: 0.5 mm, 0.7mm, 1mm).

inner ring fault: cross scratch (lengths: 3mm; depths: 0.5mm, 0.7mm, 1mm; width: 1mm),

spot (diameters : 1mm, 1.5mm, 2mm; depths: 0.5mm, 0.7mm, 1mm).

the ball spot: spot (diameters : 1mm, 1.5mm, 2mm; depths: 0.5mm, 0.7mm, 1 mm),

flat (depths: 0.5 mm, 0.7mm, 1mm).





Measuring system

Separate currents signal processing channels (consecutively from the right: shunt, transformer, signal conditioning circuit and analog-digital converter),

Electronic modules installed inside the separate magnetic screen,

Signal processing channels are located inside the separate magnetic screen.




LOAD

Clark Transform

Dataselection

Datanormalization

Measuring system

Data processing

Bank of neural network models

Fault Detection

Fault Classification

Induction motor

USB

Diagnosis system

PC

AC

Power

Supply


Block diagram of the neural network based diagnosis system




Measurement data

During the researches the data were gathered with healthy bearings in order to have the reference values of the stator currents,

Afterwards, the experiments were repeated with artificially deteriorated bearings,

The stator (supply) currents were sampled with a 65 kHz sampling rate and transmitted by USB to PC where were processed in Matlab/Simulinkenvironment (3x1600 data points/samples for one period),

Each of the experiments lasted for 8 seconds (approximately 1 million data points - 400 currents signal periods).





Data preprocessing

Step 1: Clark transformation of an induction motor stator (supply) currents

input vector of size 2x1300 = 2600 data points for one period (65kHz), data matrix 2600x400 for each experiment;

Step 2: measurements data down-sampling to fp = 16 kHz, inputs vector of size 2x320 = 640 data points for one period (16kHz), data matrix 640x400 for each experiment;

Step 3: selection of representative currents periods, e.g. data matrix 640x50 for each experiment;

Step 4: normalizing into [0 - 1] range;

Step 5: dividing the data into: training 70%, validating 15%, testing 15%.





Neural network details

four neural networks;

multilayer feed forward networks;

two nonlinear layers;

hyperbolic tangent sigmoid transfer functions;

size of the inputs 640;

15 neurons in the hidden layer;

1 neuron in the output layer;

learning method: Scaled conjugate gradient;

simulation environment: Matlab, Neural Network Toolbox.





Laboratory experiments – Neural network details

Training

performance

Confusion

matrix

Input-output

regression

Error

histogram




Simulation results

Summary

Kind of dataFault detection Fault classification

Efficiency [%]

Training data * 100 100

Testing data 77 56

Summary 92 84



* Training data: training 70%, validating 15%, testing 15%.


The number of experiments with 25 artificially deteriorated bearings and 15 healthy bearings have been carried out;

The research confirmed the effectiveness of presented approach to fault detection and diagnosis of induction motor using data driven methods based solely on measurements of motor currents at a constant engine speed;

accuracy of 77% in fault detection and 56% in case of fault classification;

The most common mistakes of faults classification process during the tests carried out, consisted in incorrect distinguishing between the faults of inner and outer ring;

Studies have shown the need for a much larger number of measurement data from various types and degree of damage as well as more advanced pre-processing e.g. by wavelet transformation in order to make easier performing a more in-depth analysis of the type of the faults.

Summary





Diagnozowanie uszkodzeń przy użyciu sieci neuronowych – wprowadzenie do

laboratorium




PROCESU - wejścia

F - uszkodzenia

Y - wyjścia

Generacja residuówModel obiektu

R - Residua

S - Sygnały diagnostyczne

Ocena wartości residuów

Lokalizacja uszkodzeń

F - Uszkodzenie

KlasyfikatorR → S

RelacjaS →F

Schemat diagnozowania z wykorzystaniem modeli

procesu

Diagnozowanie uszkodzeń przy użyciu sieci neuronowych

źródło: Korbicz i inni, 2002




źródło: Korbicz i inni, 2002


Schemat diagnozowania z wykorzystaniem modeli procesu




Schemat diagnozowania z wykorzystaniem modeli procesu

źródło: Cyra Kamowski, 2011





Przydatne komendy

• ntstool• gensim• netc = closeloop(net)• closedLoopPerformance = perform(netc,tc,yc)• [Xs,Xi,Ai,Ts,EWs,shift] = preparets(net,Xnf,Tnf,Tf,EW)• [y,wasMatrix] = tonndata(x,columnSamples,cellTime)• setsiminit(sysName,netName,net,xi,ai,Q)• con2seq• seq2con• gsubtract(a,b)• gmultiply(a,b)





Przydatne komendy

[x,t] = simplenarx_dataset;net = narxnet(1:2,1:2,10);view(net)[xs,xi,ai,ts] = preparets(net,x,{},t);net = train(net,xs,ts,xi,ai);y = net(xs,xi,ai);netc = closeloop(net);view(net)closedLoopPerformance = perform(netc,tc,yc)[xs,xi,ai] = preparets(netc,x,{},t);y = netc(xs,xi,ai);[sysName,netName] = gensim(net,'InputMode','Workspace',...

'OutputMode','WorkSpace','SolverMode','Discrete');setsiminit(sysName,netName,net,xi,ai,1);x1 = nndata2sim(x,1,1);





Authors:

Borowa A., Mazur K., Grochowski M., Brdyś M.A., Jezior K.

MultiRegional PCA for leakage detection andlocalisation in DWDS – an approach

8th Conference on DIAGNOSTICS OF PROCESSES AND SYSTEMS - DPS’07

10-12 September - Słubice, Poland 2007

Leakage Detection in Water Distribution Systems

Methods:• acoustic-based

• flow- and pressure- based

• thermal analysis

Data-based models• Repetetive nature of the

changes

Leakage symptom• Relationship is

no longer preserved źródło: Nowicki A, Grochowski M.Kernel PCA in Application to LeakageDetection in Drinking Water Distribution System. ICCCI 2011 3rd

International Conference on Computational Collective Intelligence –Technologies and Applications

Water Supply Network Model

Characteristics:

• Occupies large areas

• Local impact of leakages

• Measuring instruments concentrated in chosen nodes

Local models

Monitoring nodes(local models)

źródło: Nowicki A, Grochowski M.Kernel PCA in Application to LeakageDetection in Drinking Water Distribution System. ICCCI 2011 3rd

International Conference on Computational Collective Intelligence –Technologies and Applications

Chojnice DWDS

pump stationsreservoir

tank

nodes

pipes

• 2 water reservoirs• 1 water tank• 177 nodes• 284 pipes• 3 pump stations• number of potential measurements:467

• measuring time step 5 min

Drinking Water Distribution Systems

Reservoir

Pumps

Pipes

Tank

Nodes or

Demand nodes

Simple DWDS

MR-PCA for leakage detection –simple example

Regional PCA models

pipe flows

nodal heads

nodal demands

Simple DWDS

Max T2 = 2,65

Max SPE = 0,61

Max T2 = 2,7

Max SPE = 0,6

Max T2 = 9

Max SPE = 83,8

Max T2 = 5,5

Max SPE = 16

Max T2 = 0,7

Max SPE = 8,5

MR-PCA for leakage detection –simple example

Simple DWDS

Reduction

PCA models designing

• Data selection

• Data acquisition

• Training process:– selection of important PCs

(Captured Percent of Variance),

– T2 and SPE thresholds

Degree of PCA models dimensions reduction

Useful heuristics

The water main screening effect

T2 – Leakage at node 033

SPE - Leakage at node 033

Monitoring at node 97

leakage nodesmonitoring nodeswater mains

training dataleakage case

Useful heuristics


T2 - Leakage at node 043





Useful heuristics







Useful heuristics



SPE - Leakage at Node 043




Useful heuristics


Range of leakages detection possibilities – T2

Monitoring at node 097 Monitoring at node 103

Useful heuristics

Influence of the water tank on leakage detection possibilities

T2 – Leakage at node 023

SPE - Leakage at node 023leakage nodemonitoring nodes




Useful heuristics


SPE - Leakage at node 023leakage nodemonitoring nodes

Influence of the water tank leakages detection possibilities




Selection of subnetworksRange of leakage detection possibility in the network


Useful heuristicsWater diameters influence on

leakages detection possibilities

Network diameters in thenetwork

Diamater


Selection of subnetworksNetwork decomposition onto subnetworks

▶ DWDS structure:

◦ water mains,

◦ tanks,

◦ other pipes diameters

▶ measuring devices

III

II

IV

I

Selection of subnetworksMonitoring at node 051 -

range of leakages detection ability in the network

Leakages start at 11 AM and end at 5 PM

Sensor fault case

Failure of the flow sensor in pipe 152

123

Elements of PCA model designed for monitoring node123:

▶ water flow in pipe 152,



▶ pressure at node 123.

Sensor fault case


Monitoring at node 152 Monitoring at node 044


Failure of the flow sensor in pipe 152

Leakage localizationCase I – leakage in area III




III

II

IV

I

leakage node

Leakage localizationCase I – leakage in area III



leakage node

III


Leakage localizationCase II – leakage in area II

leakage node

II

Leakage localizationCase III – leakage in area I

leakage node

I




Kernel PCA in Application to LeakageDetection in Drinking Water Distribution

System – wybrane slajdy

• Adam Nowicki, Michał Grochowski

• Gdansk University of Technology

ICCCI 20113rd International Conference on

Computational Collective Intelligence – Technologies andApplications

Gdynia, 22nd September 2011

Lokalny model sieci wodociągowej

Symulacja wycieku w pobliżu węzła

monitorowanego

- ciśnienie:

P92

- przepływy

Q144 , Q145 , Q160 , Q161 , Q254



MiDwSS W7: Metoda kernel PCA (kPCA)

Liniowe vs nieliniowe PCA

Przykład:

dzień 1 dzień 2 dzień 3 dzień 4 dzień 5 dzień 6 wyciek






Przykład:

• Przypadek uproszczony:2 pomiary

• Przypadek rzeczywisty6 pomiarów






Przykład:

PCA

VQPCA

KPCA





•Modele statystyczne: PCA VQPCA – Vector Quantization PCAkPCA – kernel PCA





PCA

VQPCA

KPCA

•Modele statystyczne: PCA VQPCA – Vector Quantization PCAkPCA – kernel PCA

Wyniki eksperymentów kPCA

Węzeł pomiarowy

Wyciek z tego węzła wykryty

Wyciek z tego węzła częściowo wykryty

Wyciek z tego węzłaniewykryty

WynikiWielkość wycieku:QW = 0.32 m3/h

Wyniki

Dobry kandydat na węzeł monitorujący

Słaby kandydat na węzeł monitorujący

Wyniki

Węzeł 87

Węzeł 30

Węzeł 80

Wielkość wycieku:Q = 1.92 m3/h

Lokalizacja wycieków przy pomocy Kernel PCA

Wyniki

PCA VQPCA KPCA

PCA - 10 VQPCA - 19 KPCA - 4

Ilość fałszywych alarmów

Dziękuję za uwagę



monitorowanie i diagnostyka w systemach sterowania - przykłady... · lecture notes in computer...

Documents