SLIDE 1
Industrial Sensory Data Analytics Introduction, Analysis Goals & - - PowerPoint PPT Presentation
Industrial Sensory Data Analytics Introduction, Analysis Goals & - - PowerPoint PPT Presentation
Industrial Sensory Data Analytics Introduction, Analysis Goals & Methods/Tools Industrial Sensory Data Analytics Introduction Pattern Recognition in Time Series Data The field of Natural Language Processing (NLP) has seen a number of
SLIDE 2
SLIDE 3
3
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Introduction
Industrial Sensory Data Analytics
Pattern Recognition in Time Series Data ▪ The field of Natural Language Processing (NLP) has seen a number of deep learning based advances. ▪ These examples include speech recognition, voice recognition and speaker separation, and are based on recognizing patterns in the time series data that characterizes sound.
Speaker Analog signal Analog to digital conversion Pattern recognition Application
SLIDE 4
4
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Introduction ▪ The field of Natural Language Processing (NLP) has seen a number of deep learning based advances. ▪ These examples include speech recognition, voice recognition and speaker separation, and are based on recognizing patterns in the time series data that characterizes sound.
Tool Analog signal Analog to digital conversion Pattern recognition Quality
Industrial Sensory Data Analytics
Pattern Recognition in Time Series Data
SLIDE 5
5
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Signal Forecasting Soft Sensors & Signal Construction Interpretability and Explainability Signal Classification & Anomaly Detection
Industrial Sensory Data Analytics
Analysis Goals
SLIDE 6
6
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
▪ Supervised Machine Learning for Classification and Anomaly Detection. ▪ Unsupervised Machine Learning and Dynamic Time Warping for Similarity Analysis. ▪ Deep Learning and Long-Short-Term Memory Networks for Time Series Forecasting. ▪ Web Based Development Tools for Interactive Visualization Dashboards.
Industrial Sensory Data Analytics
Analysis Methods and Tools
SLIDE 7
Selected Research Projects and Applications Industrial Applications
SLIDE 8
Selected Research Projects and Applications
Industrial Applications
Deep Drawing of Car Body Parts
SLIDE 9
9
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
▪ Manufacturing of car body parts with a deep drawing tool. ▪ ~80 different parts with variable size, shape and curvature.
in collaboration with
Deep Drawing of Car Body Parts
The Product
SLIDE 10
10
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Coarse cutting Deep drawing Die cutting Water jet cutting Quality Control Cleaning
Deep Drawing of Car Body Parts
The Manufacturing Process
SLIDE 11
11
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Extension of the tool by a sensor system for continuous data acquisition during manufacturing. Using the sensory data for process monitoring and failure prediction.
Deep Drawing of Car Body Parts
The Deep Drawing Tool
SLIDE 12
12
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
▪ Strain gauge sensors give information about the force exerted upon the metal sheet. ▪ Course of the sensor signal indicates process failures.
Deep Drawing of Car Body Parts
The Data
SLIDE 13
13
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
▪ Strain gauge sensors give information about the force exerted upon the metal sheet. ▪ Course of the sensor signal indicates process failures.
Deep Drawing of Car Body Parts
The Data
SLIDE 14
14
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Network Input Prediction Target
Regular Production Cracking Metal Sheet
▪ Forecast of the strain gauge signal based on a limited cutout. ▪ Extraction of relevant information to predict the course of the sensor signal.
Deep Drawing of Car Body Parts
The Learning Problem
SLIDE 15
15
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Strain gauge cutout Strain gauge forecast
30 30 30 2
Classifier
LSTM Input LSTM LSTM
128 128 128 128
bi-LSTM Input bi-LSTM bi-LSTM bi-LSTM Output
Regressor
Deep Drawing of Car Body Parts
The Learning Model
SLIDE 16
16
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Deep Drawing of Car Body Parts
Results of the Signal Forecast
SLIDE 17
17
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
From manual labor and visual inspection of the product to automated and algorithmic data driven quality control.
Deep Drawing of Car Body Parts
In-line Signal Forecasting & Failure Prediction
SLIDE 18
18
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
What are the important parts of the signal that lead to the prediction of a process failure?
Baseline (trained on the complete signal) Prediction case (trained on the cutout signal)
Deep Drawing of Car Body Parts
Interpretability of the Model’s Decision
SLIDE 19
Selected Research Projects and Applications
Industrial Applications
Soft Sensors for Prototype Vehicles
SLIDE 20
20
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Determination of operating strengths and design loads for kinematic components in production vehicles.
in collaboration with
Extension of the car chassis of prototype vehicles with sensors to acquire load data during operation. Development of soft sensors for series production vehicles.
Soft Sensors for Prototype Vehicles
The Product, Task & Goal
SLIDE 21
21
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Internal control units External sensors Preprocessing Soft sensor models
Soft Sensors for Prototype Vehicles
The Soft Sensor Training Process
SLIDE 22
22
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Saving of electronics/hardware in series production vehicles by replacing real sensors with AI driven soft sensor models.
Soft Sensors for Prototype Vehicles
Evaluating Soft Sensors in Test Drives
SLIDE 23
Selected Research Projects and Applications
Industrial Applications
Failure Classification for Railway Switches
SLIDE 24
24
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
in collaboration with
Remote monitoring of switching operations via current signals. Deep learning based classification of current signals to identify faulty operations 60k switches in German railway network maintained by DB Netz AG.
Failure Classification for Railway Switches
Problem setting and Analysis Goal
SLIDE 25
25
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Database Historical Signals of ~ 5k Switches Failure Classification for Railway Switches
Classification of Characteristic Failure Offsets
Pre-processing DTW Matching Offset Extraction Trimming
Input: Offset Signal Convolutions (128, 256, 128) BatchNorm, ReLU Global Average Pooling Input: Operation Time Output (Softmax)
Failure Classification
1D-Convolutional Neural Network
FC Layer
SLIDE 26
26
Industrial Sensory Data Analytics Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Failure Classification for Railway Switches
Overcoming Sparse Data by Semi-Supervised Training Reducing time to maintenance in case of operation failures!
Low accuracy of traditional ML model Deep neural network requires many manually labeled curves (i.e. failures) Data enrichment by using ML-Tool for decision support and labeling
Small data catalog Nearest neighbor classifier
Classifier validation tool Classification accuracy: 84% Classification accuracy: 97%
1D-convolutional neural network Validated training data
SLIDE 27
Richard Meyes, M.Sc. Tel: +49 (0)202 439 1046 meyes@uni-wuppertal.de Chair for Technologies and Management of Digital Transformation
- Univ. Prof. Dr. Ing. Tobias Meisen
https://www.tmdt.uni-wuppertal.de/ Campus Freudenberg Rainer-Gruenter-Str. 21 D-42119 Wuppertal Germany University of Wuppertal School of Electrical, Information and Media Engineering