3. IGIT & CMBIT
A SYSTEM INTEGRATOR WORKING IN FOLLOWING AREAS:
• ELECTRICAL WORK
• IT INFRASTRUCTURE
• SOFTWARE DEVELOPMENT
Currently seeking its way to move from providing services to
delivering innovative products of its own
A PRODUCTION COMPANY INTERESTED IN IMPLEMENTING A
SOLID INDUSTRY 4.0 SOLUTIONS FOR PIPE BENDING INDUSTRY
5. THE PROCESS
THE ROBOTIZED PROCESS CELL DESIGNED TO WORK AS A
FLEXIBLE PIPE BENDING CENTRE.
• Cost-efficient, flexible robotic station for lot-size-one
manufacturing of bent metal elements.
• The currently used station consists of
• an industrial robot
• automated band saw
• bending machine.
• Costs five times less than a dedicated machine
• But it has its limitations...
6. THE CHALLENGES
Traceability
Repeatability
Lot-size-one
Lack of warehouse optimization
Lack of skilled workforce
Missing pipe identification
Careless and inaccurate pipe placement on feed rack
8. THE SOLUTION
• Extension to process cell, allowing for:
• Reduction of error-rate caused by operator
• Optimalization of production with given criteria
• Production time
• Material waste
9. THE HARDWARE
The robotic arm has been equipped with improved, robust
gripper containing also:
• 3D Camera
• Raspberry Pi4 microcomputer
11. MATERIAL RECOGNITION
APPROACH: TRANSFER LEARNING USING RESNET34
• What is transfer learning?
• Why Resnet34?
• Convolutional Neural Networks
RESULTS:
Confusion matrix
Predicted class
Black steel Brushed stainless
steel
Polished stainless
steel
Raw
aluminium
Actualclass
Black steel 10
Brushed stainless
steel
8 2
Polished stainless
steel
6 4
Raw aluminium 10
Metrics Black steel Brushed stainless
steel
Polished stainless
steel
Raw aluminium
TP 10 8 4 10
TN 30 24 28 30
FP 0 6 2 0
FN 0 2 6 0
Precision 1.0 0,57 0,66 1.0
Accuracy 0,8
12. MATERIAL GRASPING
APPROACH: CAMERA BASED POSITION AND ORIENTATION
RECOGNITION
• Image preprocessing
• Estimation of the shape
• Identification of the grasping point
0,00%
20,00%
40,00%
60,00%
80,00%
100,00%
Boundaries identified Pipe catched Pipe catched in desired part
SHAPE RECOGNITION AND CATCH - SUMMARY
Straight pipe Pre-bent pipe
RESULTS:
13. MATERIAL WASTE REDUCTION
APPROACH: REINFORCEMENT LEARNING
• Basic concepts: Agent, Environment & State, Action, Reward
• Building an example data generator to validate important
scenarios
• Assumptions:
• State must always be the same size
• State can't be too big
• Rewards must support optimization goals
RESULTS: EFFECTS ARE VISIBLE BUT STILL HAS SOME ROOM
FOR IMPROVEMENTS:
• Always cutting all available pipes
• Sometimes chooses an invalid pipe (too short for processing)
15. MIDIH COMPONENTS INTEGRATED
FOLLOWING COMPONENTS FROM APACHE STACK HAVE BEEN
IMPLEMENTED:
• NiFi+MiNiFi
Data acquisition from camera and orchestration
• Kafka
Middleware layer providing message-based
communication
• Cassandra
The data persistence component
• Zeppelin
A multipurpose notebook serving as an integrated
development environment
19. TECHNICAL KPIs
ACCURACY OF THE LENGHT MEASUREMENT: 5CM
ACCURACY OF MATERIAL IDENTIFICATION: 80%
ACCURACY OF GRASPING: 80%
20. BUSINESS KPIs
• NEW POTENTIAL CUSTOMERS OF IGIT: 3
• NEW POTENTIAL CUSTOMERS OF CMBIT: 3
• NEW PRODUCTS IN CMBIT OFFER PROCESSED
AUTOMATICALLY : 20
• REDUCTION OF THE MANUFACTURING ERRORS: 5%
• REDUCTION OF THE UNUSABLE WASTE: 35%
• REDUCTION OF THE PROCESSING TIME: 25%
21. EXPLOITATION PLAN
CONTINUE OF THE DEVELOPMENT OF THE AI BASED
TECHNOLOGY
• Reach TRL9
• Build a product around the solution
• Implement the results to different domains (Welding)
EARN THE CUSTOMERS ATTENTION THORUGH MARKETING
STRATEGY
• Association with MIDIH and H2020
• Demonstrate the technology through electronic media (
YouTube, remote demonstrations)