Application of Machine Learning techniques in the building of a predictive model for the manufacture of MDF boards.
This project is an example of the value of Artificial Intelligence in an industrial environment, both for the improvement of the production efficiency and of the final quality of the product.
In industrial production plants, controlling all the many variables that come into play in the
processes is critical and determines the quality of the end products. The control of these
variables isn’t an easy task and requires continuous adjustments by the plant’s operators. In this case, variations in factors like belt speed, temperature, humidity, cause frequent deviations from the set point that marks the optimal behaviour of the process. Control and prediction were critical, therefore.
The aim of the project was to model the behaviour of a delimited phase of the medium density fiberboards, so that their final quality meets the specifications previously defined and
increases the production efficiency. Both aims were achieved.
In this project we used for the data analysis and prediction of the production line’s behaviour. In the first phase we worked with historical data (drawing up of the theoretical model) to then, in the second phase, exploit real data and make predictions at the pace the manufacturing line requires.
First of all, we analysed a set of 60,000 records corresponding to 70 different variables (weights, pressures, speeds, levels, heights, etc.) collected over the course of one week. With this historical data, we determined which were the variables that, in the manufacturing phase the study concerns, were most determinant for the quality of the end product. For this it was necessary to apply a set of Big Data technologies that order, classify, clean and qualify the data. Next, we calculated the transfer function, which describes the process’s behaviour over the course of time.
Once we obtained this valuable new knowledge, we knew how things had been done, what the result was and what the factors were and how much they had affected said result. We could then start to determine the improvement criteria.
As a final step, we created a scorecard that provides all this information in real time. Next, after an intense job of testing and selecting the combination of most effective algorithms, applying Machine Learning technology, the graphs describing its prediction were entered on the scorecard, indicating how much a determined variable, identified as critical, should be adjusted to obtain a result very close to 100% reliability with respect to the one previously defined.
Once we had done this work, using our analysis and prediction engine, the model started to propose recommendations for taking decisions to improve the production efficiency and the end quality of the products. The implementation of this solution ends with the automation of the actions derived from the recommendations proposed by our tool. Put another way, the tool gets the line to self-regulate, obtaining better results in terms of quality and cost.
