FRISA is a global leading manufacturer of seamless rolled rings and open die forgings. They offer a wide range of alloys and steels, which enables them to give service to a great variety of industrial markets in more than 40 countries including the United States, Canada, Sweden, Italy, UK, Hungary, China, Philippines, France, Germany, and Denmark.
For this project, Ensitech worked with the aerospace industry area to which they supply seamless rolled rings in nickel and titanium-based alloys for the aerospace turbine components.
Since this plant began operations in 2003, the pieces inspection processes had been carried out manually and this meant high costs and a great number of man-hours had to be invested just in the inspection of the pieces that are produced. Besides, this was done based on the criteria of each inspector and was subject to human errors.
With the purpose of speeding up these processes an lower the costs, the idea of automatizing it through an artificial intelligence solution was raised that not only allowed for the reduction of the time of the inspection process but, above all, to be much more precise and effective, given that, in the plant, between 150 and 200 complex rings were inspected per day.
This meant a heavy workload for the engineering team and it also made decision-making more difficult by depending on the criteria of the inspectors to decide which pieces would be left and which ones would be diverted.
“[The reason] why we stayed with Ensitech was the approach they have of getting close and offering solutions: it is not just like “I’ll make you a software that does this”, but “I will bring you a mathematician that understands your problem and we are going to put our feet in the mud to understand it”. It was the desire to help. […] The majority of suppliers are willing to sell and Ensitech is willing to help.” Commented David Guillén, Quality Supervisor of FRISA.
FRISA already had a tool that allowed it to manually define the critical measures required for the inspection process and, by expert judgment, the classification of the piece was determined. With the new tool developed and tested in Microsoft Azure and then implemented in AWS SageMaker.
The purpose of the tool is to align and center different part models from a three-dimensional point cloud scan. Once the scan is available, it is processed using the implemented model that defines a variable machining height, the most optimal centering area, and lastly, the layout, that is, if the part goes to the next forging process depending on its classification.
The project was divided into 2 stages: alignment and layout. Alignment is a process of classic optimization where the center of the point cloud is calculated allowing its correct visualization. In parallel, the data preparation and cleaning functions are executed, obtaining parameters such as centering areas and initial points for the alignment algorithm.
In the second stage, the data set for the training of a neuronal network is created. The training was carried out combining state-of-the-art architectures from Deep Learning for image recognition and applying transfer learning techniques. To reduce false negatives different neural network architectures (ResNet-34, ResNet-50, Inception, and DenseNet) and variations of characteristics were used.
The final result is a system that has 95% of precision in the identification of pieces with potential quality problems, with just 5% of false positives and no false negatives.
Thanks to the tool, FRISA achieved the optimization of its processes, reducing the costs and time of evaluation of the forgings and increasing the accuracy of the decisions.
As a reflection of the aforementioned benefits, which are summarized in more precise results, favorable numbers have been obtained in the reduction of time in the analysis of the operator to each of the pieces.
It should also be mentioned that for FRISA the knowledge and technology derived from this project have been used to apply them in other developments.
We want to hear from you. Reach out to us today.