Presentazione%20di%20PowerPoint

1
DEVELOPMENT OF A METHOD FOR RELIABLE AND LOW COST
PREDICTIVE MAINTENANCE Jacopo Cassina
2
Agenda

1. Aims of the work
2. The PROMISE Project
3. Consumer Goods Scenario
4. Used tool
5. Methodology
6. Merloni Termo Sanitari application
7. Comparison with another algorithm
8. Results and Further Development

3
Aims

• This paper will present a methodology, which can
  assist technician and researchers during the
  development of a predictive maintenance
  algorithm, based on soft computing techniques,
  into the consumer goods scenario.
• It has been developed, improved and tested within
  a research and two application packages of an
  European project called PROMISE.

4
PROMISE

• PROduct lifecycle Management and Information
  Tracking using Smart Embedded Systems.
• The Promise aim develop a new PLM tool and new
  PLM methodologies, also for consumer goods.
• The PROMISE RD
• Data and information management and modelling
• Smart wireless embedded systems
• Predictive maintenance
• Design for X
• End Of Life planning
• Adaptive production management

Data Management tools
Decision Support System Tools
5
Consumer Goods Scenario

• Business requirements
• Attention to costs of
• the development of the algorithm
• The sensors
• The computational power
• Transmission of data
• Simple product

Soft computing

• Easy to use
• Short training
• Could train itself

• Robust - Adaptable
• Can analyze easily lots of parameters
• Can model rules and particular conditions

6
Short overview on the used Tool

• The proposed soft computing methodology is the
  following
• Inside a Fuzzy environment
• we will use a neural network
• to train an expert system
• Then the Rules of the expert system will be used
  to predict the residual life of the product.
• This approach could exploit the advantages of all
  the techniques, reducing the weaknesses.
• Exist dedicated hardware for fuzzy expert systems

7
Methodology

• To achieve the algorithm a methodology has been
  developed and followed.
• It aims to exploit the peculiarities of the
  scenario and of the used tool, reducing the
  complexity and the costs of the experiments and
  of the whole development.
• Eight steps will compose the methodology
• definition of the monitored breakdowns
• definition of the sub-system to be controlled
• selection of the variables to be controlled for
  each sub-system
• analysis of the whole product and selection of
  the minimum number of variables and sensors
• design of the experiments
• experimentation
• training of the algorithm
• test and validation of the algorithm

8
Merloni Termo Sanitari Application

• First application of the methodology and of the
  tool.
• Aim
• achieve a reliable predictive maintenance
  algorithm for a boiler produced by MTS.
• First step selection of the failures that has to
  be analyzed.
• The selected failures, till now, are
• The domestic hot water service failure
• The flame turn off
• The burning efficiency reduction
• The failure of the water pumps
• Second step Definition of the corresponding
  Sub-Systems.
• The domestic hot water Heat Exchanger
• The flame sensor - The burner
• The burner
• The Water Pump

9
Sub-System DHW heat exchanger

• FAILURE limestone on the plates decrease the
  heat exchange capacity
• CAUSES limestone contained in the water

10
3 step Selection of the controlled variables

• Measurable variables by boiler control board
• Domestic Hot water temp (San-Out)
• Primary circuit flow temp (P-In)
• Primary circuit return temp (P-Out)
• Burned power
• Additional measured variables
• DHW tapping flow rate
• Heating circuit pressure

Sensitivity analysis with these other variables
11
7 step Training of the FES
Antecedent / Consequents    Antecedent / Consequents    Antecedent / Consequents    Antecedent / Consequents    Antecedent / Consequents    Antecedent / Consequents   
P-Out P-IN Out-San Gas AGING weights
29,40 40,20 29,20 4953,1 5,00 1,00
33,30 45,30 32,80 4973,757 5,00 1,00
38,50 52,60 37,50 4953,1 5,00 1,00
39,90 54,70 38,80 5035,729 5,00 1,00
49,50 54,90 51,40 1606,643 5,00 1,00
54,20 72,50 41,50 5115,858 50,00 1,00
54,70 73,20 41,90 5063,799 50,00 1,00
54,80 73,30 42,10 5063,799 50,00 1,00
54,90 73,30 42,10 5063,799 50,00 1,00
55,10 73,50 42,50 5032,563 50,00 1,00
34,30 42,30 23,90 4973,757 100,00 1,00
34,90 43,00 24,10 5004,743 100,00 1,00
36,10 44,40 24,60 4953,1 100,00 1,00
36,60 45,00 24,80 4973,757 100,00 1,00
37,20 45,70 25,10 4953,1 100,00 1,00

• 3 different products
• A new Heat Exchanger
• A half aged Heat Exchanger
• An old, broken Heat Exchanger
• For each 3 experiments using different hot water
  target temperature.

12
Training Data Sets
13
8 step test and validation

• The algorithm has been tested and validated on
  some data of aged boilers and a set of data
  coming from an accelerated aging test
  (acceleration 8X ).
• Data recorded for 1 day a week.
• Sample rate 30 sec.
• It started about one year ago, and is still
  ongoing the boiler still works well.
• The algorithm analyzes each set of antecedents
  and provide an estimation of the aging.
• Then the final result is a moving average of 1000
  estimations.

Antecedent / Consequents   Antecedent / Consequents   Antecedent / Consequents   Antecedent / Consequents   Antecedent / Consequents   Antecedent / Consequents  
P-Out P-IN Sec-OUT Gas AGING Date
50 70 48 5132 19,50741 24-giu-05
46 66 41 5170 36,70522 15-lug-05
56 75 48 5095 42,09512 15-set-05
57 77 47 5132 48,99102 15-ott-05
58 73 51 5123 54,78653 15-nov-05
57 74 48 5023 62,18932 15 dec 05
58 76 50 5132 66,65374 15-gen-05
58 64 50 1620 72,37012 14-feb-06
14
Comparison with another ES

• Previously an expert System has been trained by
  MTS human Experts.
• It has been compared with the self training fuzzy
  expert system we used.

4 months 32 real months
15
Conclusions and Further Development

• Conclusions
• A methodology for the development of soft
  computing predictive maintenance algorithms has
  been proposed
• The first tests has been done
• Till now, on simple products and sub-systems,
  works well and required few data for training
• Further Development
• Make a comparison with neural networks
• Improve the training with more data
• Complete the testing analyzing the accelerated
  aging test till the breakdown of the boiler.
• Make a sensitivity analysis using also other
  sensors data
• Use the methodology on other and more complex
  product inside the PROMISE Project (even beyond
  consumer good scenario)

16
Thanks for your kind attention.

• Ing. Jacopo Cassina
•  
• e-mail jacopo.cassina_at_polimi.it
• Tel [phone]
• Fax [phone]
• Skype jacopo.cassina

17
Soft Computing Techniques

• Soft computing concerns the integration of
  different techniques, such as expert system,
  fuzzy logic, neural network and genetic
  algorithms, aimed to build machine intelligence.
• The guiding principle of soft computing is
• Exploit the tolerance for imprecision,
  uncertainty and partial truth to achieve
  tractability, robustness and low solution cost
  (Jin).

18
Predictive Maintenance

• Based on the degradation monitoring, diagnosis
  and prognosis.
• Generates over costs
• If the product is replaced earlier than needed
• If a the failure has not been predicted
• For sensors, data recording and analysis.
• Many papers have been written about maintenance
  of plants or complex and expensive machines, as
  far as we know few on consumer goods

19
Methodology

1. definition of the monitored breakdowns
2. definition of the sub-system to be controlled
3. selection of the variables to be controlled for
   each sub-system
4. analysis of the whole product and selection of
   the minimum number of variables and sensors
5. design of the experiments
6. experimentation
7. training of the algorithm
8. test and validation of the algorithm