A woman frantic after the death of her husband who turned to software that was able to learn from the conversations and electronic footprint her husband had left to create new virtual communications with him. The more Martha entered data, in the form of photographs and videos, the more complete the response of the system, to the extent that the written messages gave way to telephone conversations and finally to a bionic body, a living -although artificial- reproduction of her husband.
The episode, titled “Be right back”, became a milestone in the series and demonstrated the level of development of the many applications of Artificial Intelligence, the process that gets machines to carry out tasks traditionally done by human beings in an autonomous way.
The constant search for disruptive technology-based solutions has led Dominion to explore the possibilities offered by this analysis and data processing tool together with other benchmark players in the field of Industry 4.0. One of the forms that Artificial Intelligence takes, and in which Dominion is specialised, is Deep Learning, a subset of Machine Learning algorithms based Artificial Neural Networks that, as Arkaitz Etxezarreta, Data Analytics Manager of Dominion explains, “need a great quantity of data and computing power to bring out all their potential”.
This tool is applicable, among other things, to industrial processes such as predictive maintenance, and particularly to “quality inspections, both of surfaces and dimensions”, a field in which Dominion works in a number of research projects together with other technology partners.
The main objective is to transfer the domain knowledge of the people who carry out these quality inspections to a system that allows them to be made automatically, based on a specific solution for the capture and processing of images through this or other techniques.
Arkaitz Etxezarreta
| Data Analytics Manager
The application of a Deep Learning solution to the quality inspection process is complex but can exponentially improve its efficiency. First, it requires the capture and labelling of “thousands of images of different types of defects in the part being inspected”. This procedure, one of the most important in the whole process but sometimes taken for granted, is done using an image capture system that requires a highly detailed configuration in which several aspects come into play, such as the positioning of the system, the lighting and the shadows generated, the kind of light, the size of the defects to be detected, etc. All these images “serve to train a Deep Learning algorithm so that it can extract and generalize the characteristics of each defect in an autonomous way. Once configured, and with the solution applied in the plant, the inspection is done automatically, thereby maximizing levels of reliability."
To date, Dominion has worked with Artificial Vicomtech, a technology centre specialised in Artificial Vision and a world leader in the manufacture of steel components, in an R&D project designed to detect defects in sheet metal of different types through Deep Learning techniques.
The results of the pilot phase are encouraging for the future of this application, giving a correct percentage of detecting defects of over 80%. The handicap, as Mr Etxezarreta explains, continues to be the collection and cataloguing of the different types of defects in large quantities, as the system is not yet capable of differentiating between them with the required level of accuracy. “The key to success in using this type of technology lies in the amount of data available to train the algorithm”, he adds.
Nevertheless, its potential is enormous. As Arkaitz says, the technology can be a basic “support tool” on a production line, where operatives have limited time to make checks, so they find that they can only check one sample of the parts produced. A system equipped with artificial vision can check 100% of the production thanks to a technology such as Deep Learning. For example, it can check that a vehicle bumper (fender) contains all the components it should have before the operative starts the final assembly process. As well as lightening the operative’s workload, the control system makes it possible to locate possible errors or assembly defects early in the process, thus avoiding the need to locate the defect when the part has already been completed and assembled.
As well as in quality control, Machine Learning offers great possibilities in predictive maintenance, in which Dominion provides end-to-end services that place it in a leadership position. Thanks to its expertise in data analysis and its multidisciplinary teams, Dominion can configure specific algorithms to identify patterns of defective functioning in a range of industrial processes and thus optimize maintenance tasks. “We collect data from sensors, machine states, etc., both current and historic, to construct a model that allows us to know the state of a production asset before reaching a failure situation. Using this information, we can configure an early warning system that tells us when behaviours are detected that are similar to those leading to a failure in the past”. These techniques also allow us to shorten the so-called “diagnosis time” and maximize the efficiency of this kind of task in any sector, from energy to metalworking.
“Our value proposition is that we can provide an end-to-end service that ranges from the acquisition, installation and configuration of systems through capturing data and images to the training and adjustment of the algorithm that is required, including recent Deep Learning techniques, and their deployment in a production plant. Dominion is completely market-oriented to support its clients, increase the efficiency of their processes and, basically, solve their everyday problems”, concludes Dominion’s Data Analytics Manager.
