Advanced Software Technologies Advanced Pattern Recognition for Anomaly Detection Chance Kleineke/Michael Santucci Engineering Consultants Group Inc. August 16, 2017 Tampa Convention Center Tampa, Florida
Power Plant Monitoring Scope Typical Power Plant has ~ 2000 I/O per Unit Temperature, Pressure, Flow, Vibration, etc. Units typically have 20-40 Critical Assets Pumps, Fans, Turbines, HeatXchangers, etc One Plant Operator may monitor 2 or 3 units Hard Limit Alarm Thresholds for each sensor Plant Reacts to Alarms Too Late! 2
Alarm Thresholds Sensor Alarm limits must be set outside of normal operating range and cover all ambient conditions Hi Alarm Limit 3
Expected Value - Classical Approach Developing the Expected Value Classical Approach: Use Design information Owner Manual Consider Influencing Factors - Ambient Temperature, Running speed, etc Use first principals equations to calculate expected pressure; Boyle s Law, MLR Pressure = Function (Press_Design, Ambient Temp, Tire Speed) Expected Pressure = Constant + A1*Temp + A2* Speed + A3*Other (Mult Lin Regress) Short comings... Other Factors, Tread, wear, Passengers, road condition, Bias/Radial etc. What happens if you loose one of your inputs? One sensor Calibration issue? 4
Statistical Approach New Statistical Approach Considers What are the pressures of the other tires? Use other correlated sensors to determine where the subject sensor should be Expected values generated from history (PI Archive) Includes all higher order effects Similarity Based Modeling: Relies on the correlation between variables, not the variables themselves Uses history which incorporates all the flaws in the data and higher order effects Robust - Can run with missing inputs Precise can detect small disturbance in process ie. Slow Leak Expected Pressure = F(History : Press_Tire1, Press_Tire2, Press_Tire3, Press_Tire4) Other Factors including, wear, Passengers, road condition, etc. are already included! 5
Advanced Pattern Recognition APR is an empirical modeling technique deploying algorithms used to detect process anomalies and performance degradation in real-time. Non parametric models are superior to other techniques in their ease of deployment, simplicity and computational overhead. 6
How Does IT Work? APR software uses historical tag values to create models for assets based on past performance Powerful algorithms detect subtle changes in equipment behavior days, weeks and even months before conventional monitoring techniques Get History from All Operating Conditions Create Model and Calculate Tolerances Select Correlated Sensors Remove Redundant Data and Outliers Alarm on any Statistically Significant Deviation 7
Equipment and OEM Agnostic Rotating Equipment Turbines Pumps Fans Pulverizers Generators Motors Compressors Etc. Nonrotating Equipment Heat Exchangers Cooling Towers Condensers Transformers Precipitators Blowers Reactor Vessels Etc. 8
Model Sensors Expected Value with Limits Actual Real-time value from data historian Expected Predicted value from Predict-It Estimator Deviation Difference between Actual and Expected 9
Equipment Failure Life Cycle PredictIt Alarm Conventional Monitoring System Alarm Develop options: 1. Change operating conditions 2. Re-sequence with other maintenance 3. Better planned outage People Making the right decisions when it matters. Photos courtesy of Reliabilityweb.com 10
Case Study Generator Winding Failure Generator Cooled by water flowing through slots in generator. This case illustrates abnormal trending of a winding temperatures was detectable over a year before the generator failed. Three Years Planned Outage 1 year before Failure Anomaly Detected 11
Model Building 12
Modeling Technique Nearest Neighbors Training Data Difference Data Sum LF RF LR RR LF RF LR RR Sum( ) 25 24 25 27-5 -5-5 -5 20 26 25 26 28-4 -4-4 -4 16 27 26 26 28-3 -3-4 -4 14 28 27 27 29-2 -2-3 -3 10 29 28 28 30-1 -1-2 -2 6 30 29 29 31 0 0-1 -1 2 30 30 30 32 0 1 0 0 1 31 30 31 33 1 1 1 1 4 32 31 31 34 2 2 1 2 7 33 32 32 35 3 3 2 3 11 34 33 33 36 4 4 3 4 15 Snapshot or Test Data LF RF LR RR 30 29 30 32 Minimum Difference 13
APR Advantage A proven statistical technology solution that can be used centrally to provide an effective failure early warning system to the business across a diverse energy generation portfolio of assets Intuitive Graphical Representation of Model Results Easy to set up and use ensuring a fast road to value realization Fast and efficient model execution speed Supports Real Time Causal Network Diagnostics 14
Demo? 15
Model Building - Select Process Points 16
Model Building - Select Training Data 17
Model Building - Data Remove Outliers 18
Model Building - Run Model 19
Diagnostics Bayesian Networks Inputs: Real-time Residual Deviation Alarms Manual Tech Exam (Oil analysis) Case Failure Database Output Probability of Fault Causation Mapping Self Learning 20
Asset Network Fault cases/rules Observations/ Symptoms 21
Fault Learning and Inferrence Expert Decision Support - Diagnosis - What-If New Faults Database of Fault Signatures Asset Variable Behavior (Alarms) Asset Knowledge Causal Network Real-Time Diagnostics 22
Thank You! Mike Santucci, ECG Inc. santuccim@ecg-inc.com Phone: 330-807-7661 Chance Kleineke, ECG Inc. kleinekec@ecg-inc.com Office: 330-869-9949 Stop by Booth 404 23