Intelligent sorting challenges the ceramic industry
Macs Tech, a company that studies neural networks and parallel calculation systems, has turned its attention to automatic ceramic tile sorting as part of its work in the field of quality control in sectors where conventional algorithm techniques have proved inadequate.
Its aim is to create a system that is capable of classifying tiles effectively, objectively and repeatedly, with sufficient rapidness and low costs and the ability to adapt autonomously to changes in materials.
Macs Tech with its experience in developing a system for sorting agglomerate tiles, Emilceramica (Fiorano) with its expertise in the field of tile production, TVI with its colour image acquisition technology and Cromenco with its know-how in the development of parallel calculation software, together formed a consortium which received EC funding for a project called ASPECT (Automatic Selector Processing for Ceramic Tiles).
The ultimate aim of the project was to develop a demonstration model of the proposed technology, which was installed at Emilceramica.
Emilceramica was interested in the visual tile inspection sector for some time. Apart from the apparently still distant objective of automating the sorting phase, it felt the need to introduce a degree of objectivity into shade inspection. The concept of shades of colour, and in particular that of different or sufficiently similar shades has many ramifications from customer service (complete batches, co-ordinating with special pieces, etc.) to process control and optimisation of warehouse batches.
Projects that focused exclusively on defect detection were not considered adequate since shade inspection is also required, so an automatic tile selector, which could not distinguish shades, would not be a particularly useful machine. On the other hand, although a shade selector would not be able to entirely automate the selection process because of the need to detect defects, it would nonetheless be able to be used off-line for handling all colour tone problems.
One particularly important competitive factor is the ability to supply distributors and tile-layers with a selected product, which after installation produces a pleasant overall effect in accordance with common aesthetic tastes. The problem derives from the following key factors:
A series of uncontrollable process parameters makes it impossible to obtain a uniform finished product (or a finished product that varies in a controlled manner). This means it is necessary to perform selection in order to obtain batches of "uniform" tiles, where the concept of uniformity is currently based upon the skill of the sorting personnel. The term "uniformity" does not, in reality, correctly express the aim of tile sorting. It would be more accurate to use the term "matching" in keeping with the tendency of current ceramic production to imitate natural materials and reproduce "natural" non-uniformity in tiles. The term "tone analyser" traditionally used for automatic sorting machines would appear to be equally inadequate, since it can only really be applied to plain coloured tiles which have all but disappeared from production.
The problem tackled by ASPECT was to sort ceramic tiles according to their aesthetic matching. ASPECT aims to demonstrate Macs Tech technology by means of a device that:
A system capable of objectively evaluating the matching parameters, which can be electronically archived and easily compared at any time, would make rematching possible and cost-effective.
ASPECT is based on a simple-to-use neural software classification system called Max: all that has to be done is to show the required shade sample to it and Max learns. After the operator has set the desired sorting accuracy, Max will indicate which tiles are "out of tone" and/or when it is necessary to change shade when the product deviates from the current shade.
In order to artificially reproduce the behaviour of a human operator in sorting heterogeneous tiles into aesthetically acceptable subsets, it is necessary to extract a large quantity of information from the image of the tile and then make a decision, attributing the correct weight to each of the extracted characteristics.
After the initial learning phase, the complex human vision system is capable of extracting only the relevant characteristics from the image, amplifying them and ignoring others. In other words, the eye-brain system works as a kind of intelligent filter, extracting and considering some characteristics, which will be used, in the subsequent sorting operation. Max seeks to reproduce human abilities through the use of neural network techniques for extracting the tile´s characteristics.
Sorting, which is typically a fuzzy operation given the lack of a precise boundary between a tile belonging or not belonging to a subset, is therefore performed on the basis of the extracted characteristic and the weight assigned to each. During the learning phase, the decision-maker "learns" from the sample which characteristics it must take account of, thereby forming a selection model or criterion in a similar manner to the human sorting operator.
Max is now capable of making a choice in the same way as its human counterpart, but it has the additional capability of being able to repeat it after a period of time with exactly the same outcome.
The result of the project was a prototype tone analyser with some major simplifications compared to the solutions currently available on the market. For example, the special acquisition technique allows the systems to be installed without having to make mechanical or electrical modifications to the sorting line. This results in lower indirect costs and means that the system can be moved when required. The hardware system used in ASPECT consists of two subsystems, one being lighting and image acquisition, the other processing.
Mechanically the two systems are housed in separate structures, only the first one is specific, as it has to be positioned on the sorting line. A commercially available industrial cabinet equipped with air-conditioning is used for the computer system.
A shade selection scale divided into the following levels was used:
To allow the equipment to be tested during development, a "dummy line" was installed at the Macs Tech facility. This was a roundabout line with the same characteristics as the normal belt conveyors used in the ceramic industry. The tests were conducted in intermediate conditions between industrial and laboratorial.
Only a small number of tiles were analysed (a few dozen per article) for logistic reasons, whereas the operating conditions (speed of the line with all its irregularities, vibrations etc.) were very similar to real conditions. Other typical conditions of ceramic factories (dust, high temperatures, etc) were not simulated, but were taken into account during design.
All the levels in the above scale were examined in the laboratory tests. The system proved capable of distinguishing between all shades in all cases. Tests on a real production line are currently in progress; some problems connected with normal running of the line have already been found and solved and tests have been performed at the two highest levels with positive results although for the time being with a small number of shades.
Therefore, the system has given the expected results, but it still has to be engineered to operate normally on production lines and validated with a larger number of shades and articles. Contacts are being made with other ceramic tile manufacturers in order to verify the system with different products.
Automated sorting systems would bring numerous benefits to the entire sector with major economic advantages. A fully automatic sorting system would be able to guarantee product quality, increase plant efficiency and reduce fixed and periodic investments.
The results obtained by ASPECT in shade selection, although extremely significant, are not sufficient to form a basis for the construction of an automatic sorting system because a tile defect detection capability is also required.
A further project called OUTLINE is currently being implemented to meet this requirement and the first results will soon be available to complement those obtained with ASPECT.