Artificial intelligence, machine learning, deep learning.

These words represent the sound carpet of our times, for those who, like us, are involved in IT and related technologies. Internet, TV, specialized magazines are invaded by these concepts: they draw the future, for now with discrete and rough brush strokes, but we know what will be the framework that will emerge sooner or later.

A sound carpet also inside our offices: that’s where we want to aim, our extraordinary obsession, the artificial intelligence.

This is why we throw ourselves headfirst towards all the opportunities to develop innovative projects, even better if we work with the same customers, willing to put themselves to the test after being obviously satisfied by other “average” projects.

This is precisely the case of a project, A.I.O.C.A.P., on which we are still working, but is sufficiently advanced to be able to describe it.

If you have been following us for a long time, you know that our playground is the production data monitoring, in particular the production efficiency: from here all our software solutions, called OEE, Andon or Statistical Process Control, are launched are always based on data.

Through the Andon-OEE connection, for example, we are able to provide the customer with the production efficiency monitoring, both in real time (Andon) and in the period (OEE) and understand where are the problems that prevent the realization of a 100% OEE (i.e. maximum efficiency): quality, speed or machine downtime problems.

So, until now, we have been able to tell the customer:

you have a maintenance problem of the machines because they stop often;

you have a quality problem of the product because dents are present;

you have a slow flow problem because the line is badly organized or the objects are not within the reach of the operator.

Good. It’s a lot. Indeed very much.

And then, to close the loop of all this monitoring, the A.I.O.C.A.P project comes forward.

Let’s see what it is.

Automatic Intelligent Out of Control Action Proposal

The A.I.O.C.A.P. project (Automatic Intelligent Out of Action Action Proposal) can be defined as a machine learning system for decision support in solving quality problems.

What does this mean precisely?

It means that, in case of a quality problem, the System will be able to offer suggestions for the operator regarding their solution, in the form of an action plan.

Goal and advantages of the A.I.O.C.A.P. project

The goal of the A.I.O.CA.P. project is dual:

  1. To support the operator, in real time, in the decision to solve a quality problem encountered during production. The suggestion will be based on a probabilistic assessment of the current state of the process and on a utility model automatically learned by the system, based on the choices of the operator.
  2. To allow a knowledge transfer from human resources to the machine, by implementing a kind of machine learning. In this way another type of knowledge transfer will have benefits: the one from the most experienced problem-solving operators to the less experienced ones, thus improving the performance of each individual operator.

Regarding the advantages of a project like this one, we can say that the A.I.O.C.A.P. system will make possible two types of activities:

  1. In case of a quality problem, it proposes an ordered list of Action Plans, each including a series of actions necessary for the solution;
  2. Carries out simulations that allow the operator to understand in advance what are the consequences regarding other factors, each quality problem being included in the system registry.

The first step to implement an A.I.O.C.A.P. system is to create a base of probabilistic knowledge.

The probabilistic knowledge base

The base of knowledge, which is the foundation of the A.I.O.C.A.P. project, is built through human experience and will be applied in order to learn the decision model through the logic of reinforcement learning: each solution of the problem will reinforce the solution chosen for that specific problem and therefore, that solution will increase the probability of being chosen by the system, when the same problem occurs again.

The System is a proposal/suggestion: the human operator will be free to make changes. Changes that will enrich the knowledge base of the system.

The last one is fundamental for the construction of a causal model of the process through the use of a Bayesian Network Graph.

The following image shows an example of a Bayesian network realized for the A.I.O.C.A.P. project.

Bayesian Network Graph

Bayesian Network Graph

The Bayesian network will provide a probabilistic degree of belief in the current state of the process, following an event linked to one or more quality problems, detected automatically by the System.

The construction of the Bayesian Network Graph goes through three steps involving machine and maintenance experts:

  1. The first step involves the topology creation of the causal model of the process. All nodes (states and evidences) are listed from the causes to the effects. These nodes are then connected by defining the oriented arcs of the network.
  2. The second step is assigned to each line of the graph (arcs) relevance (weights) based on human experience.
  3. Finally, the test and validation of the estimations made by this network are carried out, through some simulations concerning quality problems.

A.I.O.C.A.P. in the field

Smart Portable Device

Smart Portable Device

In the field, the operator will have a new work tool. What we could call a portable smart device. It might be

a palmtop of course, like the one in the picture (which shows the Action plan proposed by the System to solve a body problem), but nothing prevents the System from being installed on smart glass, tablet or other.

In the field, the operator could find himself managing quality problems in different situations:

  • During a sample quality control. For example a visual inspection of the piece, in which the operator will check the presence of scratches, the amplitude of the angles, the state of the rivets, etc.
  • When a quality problem is registered through the Andon system. In this case the operator can operate directly from the line touch panel already provided by the Andon System.

Touch panel Andon e Smart Portable Device

Touch panel Andon e Smart Portable Device

In both cases, as soon as the operator has registered the quality problem encountered, the owner of the smart portable device (which could be the maintenance manager) will receive a push notification from the A.I.O.C.A.P. System APP, and at the same time the list of the most probable causes of that problem. At this point the operator can provide further tests to the system or continue clicking the «NeXT» button, thus viewing the sequence of actions, developed by the Intelligent Agent.

 

AIOCAP and the portable smart device

AIOCAP and the portable smart device

 

Machine learning

Every time the operator chooses or modifies one of the Action Plans proposed by the system, the System will be enriched by this choice: this is the moment in which the knowledge transfer takes place, from the operator to the machine, through machine learning.

Each choice of the operator strengthens the link between the probabilistic belief state and the chosen solution, for that specific problem: it is in practice a system designed to improve its performance over time.

The knowledge base can be expanded as the operator can create new action plans or edit the existing ones.

Adding and editing an action plan

Adding and editing an action plan

A system created in this way is clearly intended to refine more and more: the more experience it will manage to absorb, the more its proposals will have a higher degree of utility.

This will surely speed up the problem solving activity in case of a quality problem. But it will also be able to gain experience from the wisdom of the best and most skilled operators, thus being able to assign the maintenance task to even less experienced operators.

This is a project that goes in the direction of quality automation, meaning this type of automation that goes alongside the man, but does not replace him:  in fact the man is the fundamental resource that makes the final choice and therefore enriches the knowledge base.

It’s still early to talk about artificial intelligence, but we want to go towards this direction. Machine learning is a method to succeed, a tomorrow we hope to be near, to create projects of a real artificial intelligence.

The A.I.O.C.A.P. project is a step in that direction and, as they say,
every great journey begins with a first step.