The case study we are about to deal with won the “Second best paper” award analyzing the topic of “Manufacturing and Industrial Process Design“. This is a case study that refers to a real machine learning project to support the solution of quality and process problems.
Furthermore, the Industry 4.0 paradigm allows the integration of physical systems with digital models. In this context, the systems that provide “knowledge base” to support production control and planning are undoubtedly included.
Case study introduction
The introduction of this project is the following: production lines can be stopped or encounter delays related to quality problems.
Quality problems always result into time that is dedicated to solving the problem. Measuring this time is inversely proportional to the operator’s experience, directly proportional to the complexity of the process and finally depends on the speed of detecting the problem.
The objective of the project illustrated in this case study was: how to support the detection and resolution of production problems?
The difficulty in performing this task lies in the fact that the solution should be able to support a dynamic process, whose events and actions change over time.
To reach this goal we decided to use a “Knowledge-based System“, a knowledge base to support the operator in planning the actions to solve problems, in order to reduce the time for trial-and-error procedures.
This KBS has been implemented through a Bayesian network, to support the action decision making process..
The case study is based on the automatic production of kitchen hoods with real-time control actions. The software platform was developed by NeXT, as provider of Industry 4.0 solutions.
The image above illustrates the architecture of the project, substantially realized and still running, in a production environment at one of our most important partners.
The Bayesian network is a graphical representation of cause-effect relationships; quality control detects the warnings and problems to be corrected using the actions suggested by the Bayesian network (Knowledge Base); finally, a Machine learning scheme offers the possibility of strengthening the efficiency in solving problems.
The system modeling process is defined by four phases: each phase concerns the development of a tool to support the decision-making workflow during the operation.
- Setting of quality control detection
- The Knowledge Base (KB) is used to search for causes through a Bayesian network
- The third tool (KB Weight definition) calculates the probability of each possible solution (probabilistic approach)
- The definition of action plans will be used to solve the related problem.
Quality control setting
Quality control (QC) is based on the monitoring of some numerical variables: each value can be measured manually or acquired automatically by the sensors.
An “out of control” (OOC) signal is generated during a quality control phase, when a sample value is not included in the interval.
A detected OOC is the result of one or more possible quality problems related to the production process: the Bayesian network supports the search for causes and solutions.
Definition of the knowledge base
The knowledge base was “built” through technical briefings and consultations with expert operators.
A Bayesian network model (model based on the cause) was used to represent all cause and effect relationships: each node can be a cause or an effect.
The relationship between two nodes is probabilistic.
Each node can have one of the three values: True, False, Unknown (during an Out of Control event): the effect is searched by analyzing each foil with the True status.
Definition of “weight factors”
A weight factor is assigned to each cause-effect relationship: this WF represents how much each “parent” influences the “child“
The Noisy-OR model has been implemented in order to resolve the probability between parent and child.
Definition of action plans
The case study layout includes 5 robots, 3 press brakes and 1 rotary machine. The production process has three stages: the production of the front shape, the production of the rear frame and the riveting of all the parts.
Below is an example of a detection/resolution problem after a sample inspection process (by an operator on a molding press).
The problem detection phase is shown on the left. The respective “Action plan” elaborated and proposed in relation to the knowledge base is shown on the left.
If the operator selects the proposed action plan and adds a positive feedback, the system increases the utility value of that plan, for that state of belief: support learning.
The following figure describes the success rate for the solution to the problem that was revealed, using the action plans proposed by the KBS tool.
This report is related to a 4-week production.
As you can see, the effect of using machine learning increases the success rate during the last sampling period.
In general, the test results show a reduction in the resolution time of quality problems, by about 50%.
- A project built on the knowledge base was described to support the decisions related to the action plans.
- A Bayesian network approach was applied;
- A KBS tool was developed and tested for the production of sheet metal parts.
- The test results show a reduction in the resolution time of quality problems, by about 50%.
As a future development, the proposed KBS tool will be implemented involving a Cloud Computing.
In goes without saying that NeXT is proud to have participated in this well-founded project.
All our know-how on software development was made available to this project, in particular in the creation of the Knowledge Base and the Bayesian Network and the implementation of the Machine learning method.
We thank our partners and all those who have collaborated on the award born from this experience … really 4.0!
Ascoltando: Terry Reid – “Seed of memory”
Lettura in corso: “Le città invisibili” by Italo Calvino