tools for sensor performance...
TRANSCRIPT
Tools for Sensor Tools for Sensor Performance Assessment Performance Assessment
Research Assistants/Staff: Faculty:L. L. JiangJiang D. D. DjurdjanovicDjurdjanovic, , J. NiJ. Ni
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
Case StudyCase Study
Future WorkFuture Work
• To detect sensor progressive degradation and predict sensor failures and reconstruct the degrading sensor readings
• Direct hardware redundancy method: use two or more sensors to measure the same key variable and assess sensor performance by comparing the consistency between different sensors, simple to implement but potentially costly
• Functional redundancy method: use multiple sensors to measure related variables and validate sensor readings by employing sensor fusion techniques, more complex but cheaper approach
• Zero-lag correlation based direct hardware redundancy method
Truncated Features
M – Step Window
Sensor 1: x1 (0) x1 (1) x1 (2) …… x1 (M-1) x1 (M) x1 (M+1) ……
Sensor 2: x2 (0) x2 (1) x2 (2) …… x2 (M-1) x2 (M) x2 (M+1) ……
Sensor 3: x3 (0) x3 (1) x3 (2) …… x3 (M-1) x3 (M) x3 (M+1) ……
Extracted Features from Collected Data
……
Exponential Weighted Moving Average (EWMA) Control Chart
Extract Key Elements: C21 (t) C31 (t) C32 (t)from the Symmetric Correlation Matrix W (t)
Zero-Lag Correlation
11 12 13
21 22 23
31 32 33
(1) (1) (1)(1) (1) (1) (1)
(1) (1) (1)
C C CW C C C
C C C
⎡ ⎤⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦
W (3)
Moving the WindowOne Step Ahead
W (2) ……
• Time-series modeling based functional redundancy method
General Model
0 01
2 0
( ) ( ) ( ) ( ) ( )( )( )( ) ( ) ( ) ( ) ( )
out out in
in out in
G s G s A s A s B sS sG sS s G s B s B s A s
= = =
Transfer Function
By fitting Auto-Regressive Moving Average Vector (ARMAV) time series model to the data within an extending window, the dynamics of the plant and input/output sensor can be identified by finding the zeros and poles of the transfer function.
Sensor with First Order Dynamics
System Function
H(s) = K / (τs + 1)
Time [s]
With Degradation
Without Degradation
Inte
nsity
[V2 ]
Output Signal with Noise
U (t)
t [ms]T0 T
A0
Input Signal
Product Height Related
Product Width Related
θ
Step Input Response Curve
Nor
mal
ized
Out
puts
Time [ms]
Tangent Line at t = 0 msτ = 1 / tanθ
Period Index
Tim
e C
onst
ant [
ms]
Key Parameter for Sensor Degradation Detection
Increasing Trend
Feature Extraction
Zero-lag Correlation Based Hardware Redundancy Method
With Degradation
Without Degradation
Tim
e C
onst
ant C
orre
latio
n [m
s2 ]
Period Index
UCL MEAN LCL
EWMA Curve
Degradation Detected Period Index = 345
Tim
e C
onst
ant C
orre
latio
n [m
s2 ]
Period Index
• Evaluate both methods with real industrial data instead of simulation data
• Develop algorithm for sensor data reconstruction
Methods
Detection of Intermittent Connections in Detection of Intermittent Connections in An Industrial NetworkAn Industrial Network
Research Assistants/Staff: Faculty:YY. . LeiLei J. Ni, D. J. Ni, D. DjurdjanovicDjurdjanovic
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture Work
SponsorsSponsors
• Provide network performance and diagnostic information at the device level and network level.
• Develop novel network health monitoring tools for plant floor network systems.
• Minimize network diagnosing time due to intermittent connections on the network.
• Classification Result of Error Source
• Intermittent Connection is difficult to identify.• No tool available for Intermittent Connection detection
• Network health monitoring framework
• Error Source Detection
DNETInterface
DAQInterface
Feature Extraction &
Error Source Detection
Network
Bus-off Hitting Time&
Error-Passive Hitting Time
Intermittent LocationInference
Network Traffic Stochastic Pattern
&Error Estimation
DNETInterface
DAQInterface
Feature Extraction &
Error Packet Pattern Recognition
Network
Node waveform Feature Database
• Developed Error Source Detection System
Error Packet
Packet
2.65 2.7 2.75 2.8 2.85 2.91.7
1.8
1.9
2
2.1
2.2
2.3
2.4
2.5
2.6
Mean Voltage of Common Mode Signal (V)
Mea
n V
olta
ge o
f log
ic 0
(V)
Node 3Node 10Error Packet
• Identification of Error Location • Intermittent connection scenarios
BB AA
B AEB C
Normal Packet
BB AA
A
BE
Normal PacketError Packet
AE
A B
Receiving Mode Transmitting Mode
• Determination of the error location• Assume the intermittent connection locate at the
drop cable of a node• If we observe two patterns happens
coincidentally– Node A send error packet after normal packets– Error packets followed by packet sent by node A
• Then Node A has intermittent connection problem.
• Conduct plant validations
• Construct intermittent connection model and estimate nodes bus-off time
• is the rate of transmission of non-corrupted messages for station k
• is the rate of corrupted messages•
0kλ
1kλ
0 0k k kλ λ λ= + Error
Active
ErrorPassive
BusOff
Special command
TEC > 255
REC <=127
& TEC <=127REC >127 or TEC> 127 & TEC <=255
ErrorActive
ErrorPassive
BusOff
Special command
TEC > 255
REC <=127
& TEC <=127REC >127 or TEC> 127 & TEC <=255
Init
• Network Node Health Modeling
• NSF I/UCRC for Intelligent Maintenance Systems
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachApproach
AccomplishmentsAccomplishments
Future WorkFuture Work
SponsorsSponsors
• Use only the load data (amount of current) from the spindle and 5 servo motor drives to track mill bit health
• Explore correlation between servo motor current and cutting force in an oblique cutting
• Employ Time Frequency Analysis and other recently-developed methods to extract features from load data.
Effect of filtering on data.
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
• Filter raw load data using low-pass filter
6500 6550 6600 6650 6700 67500
1
2
3
4
5
6
7
8
9
10
6500 6550 6600 6650 6700 6750
0
1
2
3
4
5
6
7
8
9
10
Load
(V)
Time (s)
Filte
red
Load
(V)
Time (s)
• Partition filtered data to separate milling passes based on varying levels in servo-motor current
• Extract features from filtered data using TFA
Spindle Motor Current Spindle Motor Current
Time Frequency
Distribution6500 6550 6600 6650 6700 6750
0
1
2
3
4
5
6
7
8
9
10
Filte
red
Load
(V)
Time (s)
Spindle Motor Current
Time (s)
Freq
uenc
y (H
z)• Track 5-axis mill health under dynamically-varying,
oblique loading conditions
• Improve tracking and prediction of process health for various machines through signal processing
• Propose a general method for tracking the health of a mill bit without adding sensors to the system
• Decrease the overall cost of milling compressor blades by using a condition-based method to change mill bits
Extraction of features using Time Frequency Analysis
Cutting Process Monitoring in Cutting Process Monitoring in Turbine Blade MachiningTurbine Blade Machining
Research Assistants/Staff: Faculty:Adam Brzezinski Adam Brzezinski D. D. DjurdjanovicDjurdjanovic, J. Ni, J. Ni
• This research is supported by General Electric Aviation. My contact at GE is Roger Lindle.
• Finish characterization of mill health
• Install sensors and begin collecting data on grinder
• Integrate and process data from multiple sensors with various frequency ranges for the grinder project
Typical Confidence Value Plot
Amount of Warning Time from Raw Data
Time (s)
Load
(V)
Service WindowMachine
Shutdown
% H
ealth
of M
ill B
it
% of Program Completion
Decay as mill bit dulls
• Found method for properly partitioning data
• Determined predictive model for health of mill bit
• Produced CV degradation trends for load data
• Provided warning time (based on raw data) before unexpected shutdown happens
X-axis data
Spindle data
Warning of shutdown based on level of X-axis signal
Decay of mill bit health based on drift of process features
Immune System Engineering for Immune System Engineering for Automotive SystemsAutomotive Systems
Research Assistants/Staff: Faculty:ShiminShimin DuanDuan J. NiJ. Ni
D. DjurdjanovicD. Djurdjanovic
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives Immune System for the Automotive Immune System for the Automotive Engine SystemEngine System
Proposed Research Task and Time TableProposed Research Task and Time Table
Project Leadership and ManagementProject Leadership and Management
• Develop novel approach to realize “immune systems”functionalities in automated systems.
• Robustly detect abnormal behavior, isolate source, and compensate for negative effects to achieve desired performance in spite of the presence of an anomaly.
An anomaly effects automated system like a virus effects human
• PI: Dr. Dragan Djurdjanovic, University of Michigan • Co-PI: Prof. Jun Ni, University of Michigan
Kenneth Marko, ETAS Inc
Problem StatementProblem Statement• The goal of the proposed research will be to incorporate
the natural immune system functionalities and uniqueness into automotive engine system. There are three primary parts for accomplishing this target:
• Anomaly Detection Agents (ADA-s)-Identify the intrusion depicted in the degraded system-Isolate the intrusion source by reconfiguring, reconnecting and decomposing detection agents.
• Diagnostic Agents (DA-s)-Recognize and describe the anomaly source character if input/ output patterns have been observed in the past.-In case input/output patterns have not been seen in the past, create a new DA for recognizing the new anomaly next time.
• Control Agents (CA-s)-Utilize fault characterization from DAs to postulate control laws for restoring the performance of the anomalous subsystem as much as possible.
Proposed immune system operation for an automotive system
Multiplication of ADA-s, DA-s and CA-s in the electronic throttle mechanism
Subtask 3.1: Design of CA-s for recovering performance
Task 3: Methods for Control System Reconfiguration
Subtask 3.2: Implementation on an engine subsystem
Subtask 2.3: Demonstration on an engine subsystem
Subtask 2.2: Development of identification methods
Subtask 2.1: Methods for realization of DA-s
Task 2: Methods for Diagnosis of Anomalous Condition
Subtask 1.3: Practical implementation of ADA-s
Subtask 1.2: Isolation of the Anomaly Source
Subtask 1.1: Multi-regime anomaly detection
Task 1: Methods for Anomaly Detection and Isolation
31-3625-3019-2413-187-120-6Tasks
Month
The research activity for over a three-year period