On-line Alert Systems for Production Plants

1
On-line Alert Systems for Production Plants

• A Conflict Based Approach

2
Alert Systems

• Based on sensor readings, raises a flag in case
  of an abnormal situation
• Researchers construct a system using Bayesian
  Networks to detect abnormal situations and
  generate alerts

3
Typical Production Plant
S1
S2
Sn
C1
C2
Cn
4
Common Issues

• Engineers knowledge of plant is not sufficient
  for providing a causal structure
• Process is too complex to specify the possible
  faults and how to detect them based on sensor
  readings
• Difficult to determine the delay from event to
  effect
• Faults are so rare that statistics cannot be used
  to learn neither the structure nor the parameters
  of a model of the faults
• Difference between a true value and its sensor
  reading true values should appear as hidden
  variables

5
Causal Approach

• Collect as much causal structure as possible and
  combine with a data driven learning method
• Limitations
• Learning algorithms cannot cope with domains with
  a massive set of hidden variables
• It is not obvious how such a model could be used
  for classifying abnormal behavior

6
Proposed Method

• Learn a Bayesian Network representing normal
  operation only
• Does not require information about possible
  faults or modeling of abnormal behavior
• Faults are detected by measuring conflict between
  model and sensor readings

7
Proposed Method

• Consists of two steps
• Learning a model of the sensors for normal
  operation
• Using the learned model to monitor the system,
  initiate alerts and perform on-line diagnostics

8
Learning a Model

• Can be done in many different ways
• In this paper the researchers analyzed the
  database of sensor readings during normal
  operation where the variables are the sensors and
  hidden variables are the components of the system

9
(No Transcript)
10
Initiation of Alerts

• Sensor readings are received in a constant flow
• Readings are chopped up into time steps of 1
  second
• Therefore every second we have evidence for every
  variable in the model

11
Conflict Measure of Evidence

• Let evidence be e e1,,en and the conflict
  measure of the evidence is

12
Detecting Conflict

• In general during normal operating we expect
• conf(e) 0
• An indication of an abnormal situation is
  detected when
• conf(e) gt 0

13
Problem with Conflict Measure

• A negative conflict value does not necessarily
  imply that we have a normal situation
• If sensors are strongly correlated during normal
  operation, the conflict level will be very
  negative
• A few conflicting sensors therefore will not
  cause the entire conflict to be positive
• To detect watch for jumps in the conflict level

14
Tracing Source of Alert

• Greedy Conflict Resolution
• Recursively remove the sensor reading that
  reduces the conflict the most
• Stop when conflict is below a predefined
  threshold
• Can be performed very fast using fast
  retraction, lazy propagation or arithmetic
  circuits

15
Conflict Resolution
16
Coal Power Plant Network
17
Sample Data set

• Used sample data from normal operation to learn
  model for coal mill
• Two data sets covered actual errors/abnormal
  situations
• The fall-pipe leading coal into the power plant
  becomes clogged
• A temperature sensor becomes faulty

18
Clogged Fall-Pipe Data
19
Clogged Fall-Pipe Conflict Resolution

• Indicated that the sensor measuring the
  water-percentage in the coal can explain all the
  conflicts
• There is no sensor for Fall-Pipe clog detection
  so this is the closest sensor that could explain
  the conflict
• Result was consistent with the analysis of the
  plant Engineers

20
Faulty-sensor Data
21
Faulty Sensor Conflict Resolution

• Indicated that six significant sensors could
  explain the conflict
• Engineers indicated that four of the six were
  actually significant, the other two were
  anomalies due to the model

22
Oil Production Facility

• Simulated cases for normal system operation
• Two other data sets covered errors/abnormal
  situations
• Simulated faults in the pumping system
• Simulated faults in the cooling system

23
Pump and Cooling Data
24
Pump Data
25
Cooling Data
26
Change Point Detection

• Note that in the previous plots that the conflict
  measures where all negative indicating no faults
• However, remember the problem if the sensors are
  strongly correlated (slide 15)
• In order to perform conflict resolution the
  conflict threshold should be based on values
  observed during normal operation
• In order to detect changes in system operation
  need to track jumps in the conflict measure

27
Change Point Detection

• Assume conflict values for normal operation are
  independent samples from normal distribution with
  fixed mean and variance
• Model the lth conflict value as random variable
  with normal distribution f where mean ?l and
  variance ?l2 are estimated using the last m
  observations

28
Change Point Detection

• To detect change point calculate the logarithmic
  loss for the last n observations and raise alert
  in case the value is above a predefined
  threshold

29
Change Point Detection

• This approach is sensitive to fluctuations and
  results in false positives also has difficulty
  detecting drifts in conflict measure
• To alleviate this compare model with another
  model f where mean and variance estimated by
  shifting s observations back

30
Change Point Detection

• Compare models (score difference)

31
Change Point Detection

• For normal operation score should be within the
  interval -? ?
• ? will determine the ratio of false positives and
  false negatives
• n determines the response time of the system
• m determines the relevant history
• s the number of observations to shift back

32
Clogged Fall-Pipe Data(nm5, s20)
33
Clogged Fall-Pipe Data(nm10, s40)