Intelligent Decision Support Systems
Deep Learning for Smart Production
Curriculum Development
of Master’s Degree Program in
Industrial Engineering for Thailand Sustainable Smart Industry
Smart manufacturing
• Smart manufacturing refers to using
advanced data analytics to
complement physical science for
improving system performance and
decision making.
• Manufacturing machines are fully
connected through wireless
networks, monitored by sensors,
and controlled by advanced
computational intelligence
Smart manufacturing
• Key supporting technologies to
advance modern manufacturing are
Internet of Things (IoTs), Cloud
Computing, and Cyber-Physical
System (CPS)
• Data at different stages of product’s
life is collected and processed
Smart manufacturing
• Data driven intelligence models the complex multivariate nonlinear
relationships among data, with no in-depth understanding of system physical
behaviors required
Smart manufacturing
• The potential benefit and successful application examples of typical machining
learning techniques including Bayesian Networks, Instance-based Learning,
Artificial Neural Network, and ensemble methods
• Deep Learning provides an advanced analytics tool for smart manufacturing in
the Big Data Era
Deep Learning VS Traditional Machine Learning
a) Traditional machine learning
b) Deep learning
Deep Learning VS Traditional Machine Learning
Feature learning Model construction Model training
Traditional Explicit engineered Use extracted features to Each module is
machine features extracted with construct data-driven trained step-by-step
learning expert domain model, usually with
knowledge shallow structures
Deep Features are learned by An end-to-end high Parameters are
learning transforming data into hierarchical model trained jointly
abstract structure with nonlinear
representations combination of mult-layers
Deep learning enabled advanced
analytics for smart manufacturing
Deep learning models
Model Principle Advantages Disadvantages
Convolutio Abstracted features are Reduced parameter High computational
nal Neural learned by stacked number, invariance of complexity for high
Network convolutional and shift, scale and distortion hierarchical model
(CNN) sampling layers training
Restricted Hidden layer describes Robust to ambiguous Time-consuming for
Boltzmann variable dependencies input and training label is joint parameter
Machine and connections not required in pre- optimization
(RBM) between input or training stage
output layers as
representative features
Deep learning models
Model Principle Advantages Disadvantages
Auto Unsupervised feature Irrelevance in the input is Error propagation
Encoder learning and data eliminated, and layer-by-layer and
(AE) dimensionality meaningful information is sparse
reduction are achieved preserved representations are
through encoding not guaranteed
Recurrent Temporal pattern Short-term information is Difficult to train the
Neural stored in the recurrent retained, and temporal model and save the
Network neurons connection correlations are captured long-term
(RNN) and distributed hidden in sequence data dependence
states for time-series
data
Applications to Smart Manufacturing
Concept
Sustainment
development
Design &
Performance
Operation Design
Materials,
Operation &
Processing
Maintenance
&Manufacturing
Production Test &
Evaluation
Descriptive analytics for production quality
inspection
• Machine vision and image processing techniques is used to detect surface
defect
• Deep learning has been investigated to learn-level generic features and
applied to a wide range of textures or difficult-to-detect defects cases
• Applications:
• Automated defect identification - CNN
• Feature extraction from the pixel representation of steel defect images - CNN
• Automatically inspect dirties, scratches, burrs, and wears on surface parts - CNN
Diagnostic analytics for fault assessment
• To monitor machinery conditions, identify the incipient defects, diagnose the
root cause of failures, and incorporate the information into manufacturing
production and control
• Applications:
• Image analysis – CNN
• Fault diagnosis of aircraft engine, chemical process, reciprocating compressor, rolling
element bearing, high speed train, and wind turbine – DBN
• Signal denoising, fused feature extraction, and rolling bearing fault diagnosis - AE
Predictive Analytics for Defect Prognosis
• Historical data is important for prediction temporal behavior of manufacturing
processes, such as maintenance, tools and machine lifetime.
• Applications:
• Mechanical system or component defect propagation and useful life, Machine health
monitoring tasks – RNN
• Useful life of an aircraft turbofan engine – LSTM
• Material removal rate and chemical mechanical polishing process parameters, RUL
prediction of a ceramic bearing, job schedule and balance the computational load - DBN
Case Study
• A Quality Control Application on a Smart Factory Prototype Using Deep
Learning Methods
Features:
• A camera placed over the assembly line in a smart factor model to detect object and
classification (Visual Quality Control)
• Deep learning methods classify “OK” or “NOT OK” product
• PLC controls the production line
Case Study
• A Quality Control Application on a Smart Factory Prototype Using Deep
Learning Methods
Smart Factory training prototype
assembly line
• Festo’s “Pick and Place”, “Muscle Press”
and “Separating” modules
• CPU 1214C model of Siemens S7-1200
PLC series
• Festo’s pneumatic and electropneumatic
valves, pistons, vacuum elements,
proximity sensors, DC motors
Case Study
• A Quality Control Application on a Smart Factory Prototype Using Deep
Learning Methods
Programming HMI
• TIA Portal V13 software
• The position of the product
• The quality control result
• The number of “OK” and “NOT OK” products
• Real time result
Case Study
• A Quality Control Application on a Smart Factory Prototype Using Deep
Learning Methods
Object Detection and Recognition
• 189x2 dimensional list storing 189 images of data set
• CNN – AlexNet
Case Study
• A Quality Control Application on a Smart Factory Prototype Using Deep
Learning Methods
“OK” and “NOT OK” data set
• 356 images in 3 colors: red, gray and black
• 3024x3024 pixels in RGB resized to 227x227 pixels to match CNN’s input image
dimensions
• ACF object detector algorithm
Case Study
• “NOT OK”
• “OK”
Thank You
https://msie4.ait.ac.th/ Together We Will Make Our Education Stronger
@MSIE4Thailand
MSIE 4.0 Channel
Curriculum Development
of Master’s Degree Program in
Industrial Engineering for Thailand Sustainable Smart Industry
