Condition Monitoring ArchitectureTo Reduce Total Cost of Ownership

Eric Bechhoefer, NRG SystemsBrogan Morton, NRG Systems

Barriers to Sales of PHM Systems

• CBM/PHM System are Proven to Work– Low Penetration into Commercial Markets– Example: 3% of Wind Turbines

• Why? - Business Case is Hard to Make– Safety not the primary concern, cost avoidance is– Hard to Quantify Benefit

• Change Architecture to Improve Value– Lower “Costs” and Better Information

Current System Architecture• System Hardware

– 6 to 8 PZT Accelerometers• 5% Accuracy, .5 to 10,000 Hz

– Tachometer– Signal Conditioning

• 6 to 12 channels• Sample Rate: 60 to 80 KSPS

• Support/Monitoring Services– Human in the loop to turn data into a diagnosis– $1,000 to $1,500 per year per turbine

• IT Infrastructure– Data hosting on local server– Data also shipped to centralized analysis center

System Layout: Wind Turbines

From a System Perspective…

• How to Lower Total Ownership Cost– Hardware Considerations

• Costs driven by accelerometer

– Software/Support Considerations• Costs driven by knowledge creation (data to diagnosis)

– IT Infrastructure Considerations• Cost driven by local data storage and associated

maintenance

Accelerometers – MEMS vs. PZT

MEMS Advantages• Cost

– $6 to $30 vs. $100’s

• Bandwidth– 0 to 32,000 Hz vs. 0.5 to

10,000 Hz

• Accuracy– Typically 1% vs. 5% or 10%

Error

• Self Test– Can Enable BIT vs. No BIT

MEMS Disadvantages• Needs to be Packaged

– No Trivial Task

• Noisier– PDS is 2 to 40x higher

System Issues• 4 Wire?

– Power/Signal

• Local Conversion?– ADC, then Microcontroller

and RT

Sensor System Considerations

• Low cost target – move to MEMS– Analog vs. Digital Sensing

• If Digital– Local ADC,

• EMI is Reduced

– Microcontroller, RAM, Receiver/Transmitter• If Multi-Drop: RS-485

– If Microcontroller: Local Processing?• Many Smaller, Cheap Processors vs. One Larger Processor

• Low cost packaging– Alternative to Stainless Steel or Titanium– Transfer Function – Has to Be Stiff/Light

MEMS: A Sensor Solutions

• Noise Was Not An Issue– After Signal Processing,

Noise was Negligible

• Conductive Plastic Package– 40% Mass of Stainless– Similar Stiffness– 12% of Cost of Stainless– 6.5KHz Resonance, Flat

Response to 17 KHz

1000 mv/g vs. 70 mv/gMEMS Accel, 0.25 HzWithin 2% of Low G Accel

Embedded PHM• Micro with FPU Support

– 32MB RAM– 24 Bit ADC– Sample @ 300-100,000 kbps– R/T > 500 KB/S

• Local Vibe Processing– Time Synchronous Average

(TSA)– FFT/IFFT– Hilbert Transform

• Total Cost: Similar to PZT Accel

Software & Support Considerations

• Algorithmic– Digital signal processing of the vibration signals for

fault detection• Knowledge Creation

– Goal: Actionable information requiring little interpretation

Main Shaft

MainBearing

3-stage Gearbox

Generator

• 17 Bearings• 9 Gears• 8 Shaft

Low Speed Shaft

Int. Speed Shaft

High Speed Shaft

Car

rier

Pla

te

Typical Drivetrain Configuration

Algorithmic

• Process vibration signal into indications of faults– Data reduction without loss of information

• No Spectrums/Order Analysis– Configurable Analysis for Shafts, Gears and

Bearings,– Several Condition Indicators for Each Component

• Use Time Synchronous Average (TSA)

Why This Approach

• Large Variation in Wind Speeds Cause Large Changes in Rotor Speed

• 3/Rev Torque/Speed Ripple From Tower Shadow/Wind Shear

• Gearbox has many gear meshes; isolate gears of interest

• Due to Changes in Rotor Speed, Order Analysis or the PSD Cause Smearing of Frequency Content

• Example Main Rotor Shaft

Example of Spectrum Vs. TSA

1st, 2nd, and 3rd Harmonics of Ring Gear Frequency

The TSA

• Use Tachometer as Phase Reference on Shaft• Reduces Non-Synchronous Noise 1/sqrt(revolutions)• For Each Revolution (From Tach)• Resample length m = 2^Ceiling(log2(number of points in Rev))

Gear Fault Indicators

• No Single CI Works With All Fault Modes– Surface Disturbance, Scuffing, Deformation,

Surface Fatigue, Cracks, Tooth Breakage, Eccentricity

• Use a Number of Analysis to Cover All Fault Modes– Residual Analysis, Energy Operator, Narrow Band

Analysis, Amplitude Modulation Analysis, Frequency Modulation Analysis.

Gear Analysis

Knowledge Creation

• Recall Goal: Create actionable information requiring little interpretation– Convey what to fix and when

• Single Health Indicator for Each Component– Fusion of different condition

indicators– Common scale for every

component (0-1)

Health as a Function of Distributions

• HI Paradigm: Map the CIs into an HI– HI Ranges from 0 to 1, Where the Probability of

the HI exceeding 0.5 is the PFA– HI in Warning when between 0.75 and 1– HI is Alarm when Greater than 1.0– Continued Operations with HI > 1 could Cause

Collateral Damage

Controlling Correlation Between CIs

• All CIs have PDFs• Any Operation on the CI to

form an HI is a Function of Distributions– Max of n CI (an Order Statistics)– Sum of n CI– Norm of n CI

• Function of Distribution– PFA Correct if Distribution are

IID– Need to Whiten

ij CI 1 CI 2 CI 3 CI 4 CI 5 CI 6

CI 1 1 0.84 0.79 0.66 -0.47 0.74

CI 2 1 0.46 0.27 -0.59 0.36

CI 3 1 0.96 -0.03 0.97

CI 4 1 0.11 0.98

CI 5 1 0.05

CI 6 1

CI to HI Mapping

• Six CIs used in HI Calculation– Residual RMS– Energy Operator RMS– FM0– Narrowband Kurtosis– AM Kurtosis– FM RMS

• Statistics Generated from 4 test articles: 100 samples prior to fault propagation

IT Infrastructure

For Owner/Operator• No Seat License of the CMS

Database• No Local Servers to Host Data• Management of Software

Maintenance. • Allows Pooling of Dataset of

Similar Type/Model Turbines without Risk of Exposing Proprietary Information

For CMS Developer• Simplifies Software Maintenance

Cost– Only one Platform to Develop

and Test to, – Only one Platform to Deploy

Software Updates/Patches to,• Reduces the Cost of Certification

– Configuration Management is Greatly Simplified

• Scalability

Alternate to Local Server: Cloud Computing

Conclusion

• Significant value can be created by redesigning system architecture– Vibration sensing

• Non-traditional sensor, new packaging and design methods

– Advanced signal processing techniques• Increased sensitivity to faults under dynamic conditions

– Knowledge Creation• Automated fusion of fault modes• Actionable information with diagnostic support

– IT Infrastructure• Economy of scale using cloud services