Deep Learning & deep analysis: The interplay between neural networks and complex system simulations Presentation by Peter Mas November 3 2017 ESI conference Realize innovation.
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 2
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 3
Siemens - Simcenter TM : Simulation and Test Solutions Enabling Digital Twin for Closed Loop Performance Engineering Data analytics using data & models: content of this presentation Systems-Driven Product Development Design System Test Simcenter TM Portfolio Predictive Engineering Analytics Exploration Digital twin Analytics Utilization CAE Reporting Teamcenter - Digital Continuity for Multi-Domain Traceability, Change and Configuration Page 4
System simulation What is a Mechatronic System & how to model it? Control Mechanic Hydraulic LMS Imagine.Lab Amesim system model Control Electric Hydraulic Mechanic Electric A group of multi-physic components which influence each other and are controlled active or passive Thermodynamic Page 5
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 6
Machine Learning In One Slide Supervised learning Standard Neural network Recurrent Neural network Convolutional Neural network Basic idea For time series inputs For n-dimensional data Page 7
Combining data with simulation models Examples in this presentation VIRTUAL SENSING how to get the actual data that is needed (without increased cost) PREDICTIVE MONITORING How to forecast a problem using the huge amount of data system system Virtual sensing Predictive Monitoring System control Predicted by model Trained by model Page 8
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 9
Virtual sensing toolchain Customer s objectives Goal Develop a generic tool to create Virtual Sensors from a Digital Twin, ready for deployment on ECU What is an Extended Kalman Filter (EKF)? Kalman Filter: predictor-corrector algorithm to estimate non-measured variables from a model and a set of measured variables Extended: Kalman Filter applied to non-linear models (but required data can be huge) Applications of EKF Variable or parameter estimation Sensors removal (cost savings) Redundancy / operating range extension Data fusion (merging estimations) Page 10
Virtual sensing toolchain Improved workflow with Machine Learning Simplified workflow Model reduction done automatically using Principal Component Analysis (PCA) and whitening LPV model replaced by a Recurrent Neural Network state-space model Benefits Improved overall accuracy (loss of accuracy 2) Total memory dropped from 100 Mb to less than 100 kb Fast training from simulation results (~5 min) Automatic model reduction thanks to PCA (here 20 states + 5 inputs reduced to 5 inputs for NN model) New workflow with Neural Network State-Space Model States & control input data f, h data and Jacobian A, B, C, D data PCA (with whitening) Training/Fitting Neural network model simulation (NEDC, WLTC etc.) Change settings (number of neurons, hidden layers etc.) Performance Evaluation EKF synthesis Page 11
Virtual Sensing Tool chain Combustion engine EGR ratio estimation Page 12
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 13
Condition monitoring & machine learning The ITEA Reflexion Project 11.4 M, three year European cluster research project (ITEA) Consortium: Siemens, Oce, Philips, Barco, Axini, Synerscope, TNO-ESI, Yazzoom. Project focus: Research applications of machine learning to: Improve data analysis of inservice products to drive innovation in the design phase. Further valorize existing design models by leveraging them as a source of machine learning training data. Page 14 (Work in Siemens performed by Cameron Soby)
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Overview Anomaly Detection: A project collaborator posed a challenge to find a behaviour change in an anonymized machine data solved using deep auto-encoding CNNs. Encode Decode Prediction Fault detection: A combination of simulation-generated data, signal preprocessing, and CNNs are used to detect bearing faults. Simulation, Preprocessing CNN Prediction Root cause analysis: Analysis of time-series sensor data to determine the location, severity and orientation of shaft imbalances causing vibration using a simple RNN Signal Prediction Page 15
Simulation driven machine learning Shaft dynamics analysis Anomaly detection Page 16 Application example: condition monitoring Shaft Dynamics - Motivation Shafts are the most common and most efficient mode of power transportation for rotating machinery. In the oil and gas and power generation sectors, large shafts connecting electrical motors or gas turbines to pumps, compressors transmit power in the megawatt range at speed approaching 10k RPM. The value of the machines and the losses incurred by their downtime necessitate monitoring methods that do not require experiments. Goal: Train machine learning method using simulations to determine location and magnitude of imbalance. Large turbomachinery is used to create pressure and drive flow in gas pipelines. The shafts driving these machines must be carefully balanced. (Siemens, 2017)
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Shaft Dynamics - Motivation Large Drive (>10 MW) Generator Measurement Planes Imbalance Planes Application case: a large electric motor drives a generator, which are connected by a 10m, 20 ton shaft. Shaft displacement is measured at four points, and imbalance corrections are possible at four other locations. Page 17
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Shaft Dynamics Imbalance Prediction Page 18
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Bearing Fault Diagnosis Failure cause prob. distribution for induction motors. (Bonnett and Young, 2008) Failure cause prob. distribution for wind turbines. (Fischer et. al, 2012) In the course of normal operation in bearings and gears develop faults from fatigue loading; Is it possible to detect faults before they cause a system failure using simulations? Page 19
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring High Frequency Bearing Model LMS Amesim bearing model an outer race defect as a force impulse. Three DOF bearing model resulting in an inhomogeneous system of 3 ODE s. Sassi et al. (2007) Acceleration time signal of outer race the envelope function can be modeled as one or pair of exponential decays depending on the system parameters. Comparison to a 60 DOF model (Nakhaeinejad 2010) shows that the simple model captures the key behavior well. Page 20
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Signal Synthesis From high-frequency resolution simulations, longer time signals are synthesized: Gaussian broadening Half-normal delay Noise injection with P ω 1/ω (pink noise) Deviations from no-slip assumption are introduced: Gaussian broadening: play between the rollers and the carrier Half-normal distribution: slippage from roller-defect interaction. Page 21
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Bearing Signal Analysis Modeling and Experiment Here, we will apply domain knowledge where possible to target the characteristic bearing fault response. Synthesis Create time-series envelope signal using simulation information Augment with noise Signal Processing Synchronously average at expected defect periods Produces several averages per signal. Machine Learning Extract features or directly use sync. avg. to train supervised algorithms. Page 22
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Machine Learning Methods - Feature-based Goal: characterize signal with statistical measures capturing key differentiators. Skewness Kurtosis Wavelet Energy f 1 f 2 f 3 f N Logistic Regression Random Forest MLP Neural Network Features used: skewness, kurtosis (ASA and spectral), crest factor, margin factor, wavelet decomposition energies. Algorithms used: logistic regression, random forests, k-nearest neighbors, SVM-RBF, and neural networks. Page 23
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Machine Learning Methods - CNN Convolutional Neural Network (CNN) CNNs learn filters that exploit the spatial morphology of the data Applying this method to an angle-synchronous average is more logical than the power spectrum because of the spatial sensitivity of the power spectrum. A small, portable network is sufficient to reach high accuracy because of the significant preprocessing. filter 15x10 filters 50% dropout Flatten 50 dense units 50% dropout Nominal Faulty 128 samples Page 24
Shaft dynamics analysis Simulation driven machine learning Anomaly detection Application example: condition monitoring Experimental Data Simulation-driven machine learning methodology. The best statistical classifier (random forest) performs the best when in-service data for the same machine is available; otherwise, a CNN or NNDTW is preferable. Page 25
Shaft dynamics analysis Simulation driven machine learning Predicted Fault State Inner Race Outer Race Anomaly detection Application example: condition monitoring Wind Turbine Inner Race Fault 1 0.8 0.6 0.4 0.2 0 50 45 40 35 30 25 Days Until Failure 20 15 10 5 0 1 0.8 0.6 0.4 0.2 0 50 45 40 35 30 25 Days Until Failure 20 RF CNN NNDTW 15 10 C. Sobie, C. Freitas, and M. Nicolai (2017) Simulationdriven machine learning: Bearing fault classification. Accepted to Mechanical Systems and Signal Processing. 5 0 Page 26
. Agenda Introduction: Complex system simulation Introduction: Deep learning Application example: Virtual sensing Application example: Condition monitoring Conclusions Page 27
Deep Learning & deep analysis Conclusions Increased amount of streaming test data is available to describe the state of a system. Analysing it becomes more and more complex as a job. System simulation is a rapidly growing physical modeling capability allowing to model complex mechatronic systems Deep learning is a rapidly growing capability that allows to learn complex relationships between available information In this presentation it has been shown how both system simulation and deep learning can be used to make maximum use of the available test data: In virtual sensing, physics based models allow to predict sensor information virtually, supported by deep learning technology to make it practical In predictive modeling, neural networks can be thought to monitor the system state of health supported by physics based models Page 28
Thank you for your attention Questions? Peter Mas Simcenter Engineering & consulting Siemens Test & Simulation segment E-mail: Peter.mas@siemens.com Realize innovation. Page 29