Machine Vision System with Deep Learning for inspection of metallurgical defects

We currently are in the process of implementing a new system for the inspection of metallurgical defects, which will go one step further than the current one we already have with artificial vision. We seek to ensure that all of the parts we deliver strictly meet the quality requirements established by our customers. Our goal is total quality, ‘zero defects’.

Deep learning, being a part of machine learning, is not an abstract technology. It is a tool that helps to improve our industrial activity. With the right support of our partners, it is an exciting and realistic project. We are going to achieve the standardisation of inspection criteria, which will help us to improve our customer service, as well as, through greater automation, the reduction of delivery times for our parts.

Machine Vision vs. Deep Learning

According to the Automated Imaging Association (AIA), machine vision encompasses all industrial and non-industrial applications, in which a combination of hardware and software provides operational guidance to devices in performing their functions, based on image capture and processing. 

Deep learning is a set of machine learning algorithms that attempt to model high-level abstractions in data, using computational architectures that support multiple and iterative nonlinear transformations of data expressed in matrix or tensor form.

For some time, we have used a system for inspection of defects in parts with artificial vision, based on measurement. By projecting a shadow of the part that was captured with a camera, the tolerance was measured and compared to a pattern. Through the comparison, it was detected whether a part was bad or good, based on how it deviated from that pattern.  

Currently we are making a qualitative leap in the incorporation of new Industry 4.0 technologies, by implementing an artificial vision installation with deep learning, which detects surface defects with an advanced comparison system and continuous & autonomous learning. It consists of two cameras that take a series of photographs of the part and an algorithm, which compares these photographs with other standard photographs that we have previously fed into the machine with its defects. 

With these improvements, we move from a system based on a mathematical operation, that is, from measuring and comparing against a standard, to detecting the surface effect with a more advanced and complex criterion. The machine is able to differentiate whether the part is or not ok, whether it is recoverable or not and what defect it has, and to learn and improve continuously. 

Phases of the project

The project consists of 3 phases: 

1. Commissioning
   The commissioning or training of the machine is done with a catalog of defects. In our case, we started with 300 photographs.
2. Training
   The machine has to be trained. Parts are passed to it and it carries out the inspection itself. As soon as it finds a defect that is not in the catalog, these new photographs are incorporated into the database, which means, that it is continuously learning.
3. ‘Continuous’ learning
   This happens when the machine finds a part it doesn’t know exactly whether it is ‘good or bad’. In the same machine software, the operator will be able to decide or define himself.
   While the machine is already working, if something comes up (i.e. if some part or some new or doubtful situation comes up), the operator or the quality technician can or should define if the part is OK or not, if it is recoverable or not, and the machine incorporates it into its system.

Deep learning and human intervention: The importance of people and partners

The human factor is key in this project. The training is carried out by technical staff from Ecrimesa’s Quality Department. We start with a database of 200 images, and for a month, the technicians analyse the images processed by the machine. In the case of finding errors that the system has not detected, this ‘defect’ is incorporated into the catalog.  

This project has been led by Guillermo Alonso and Emilio Álvarez, Quality Managers of Ecrimesa Group, together with BSP Systems and Siali. 

BSP Systems is a leading supplier of industrial automation, with customers throughout Spain and in countries such as Germany, Hungary, Romania, Russia, China and Mexico. Siali is a startup that has developed the artificial vision software and learning system, winners of the CaixaBank Entrepreneur XXI Award.

How do we improve metallurgical quality compliance?

The implementation of new technologies in Ecrimesa Group is always oriented to a tangible improvement for our customers. The artificial vision system with deep learning for inspection of metallurgical parts, implies zero defects.

Uniformity in quality

With this new system, we ensure that the inspection criteria will always be the same and the quality standard will be met without exceptions. In addition, the key indicators will not be conditioned by the decision of a quality operator, but are automated, so that we guarantee that 100% of the parts strictly comply with the quality requirements set by the customer.

Reduced inspection cycle time

We go from the current 16 hours per cycle to one person every 40 minutes going for 5 minutes to load the machine with new parts, which will mean a significant improvement in productivity.

Improved customer service

Standardization of inspection criteria will ensure that all delivered parts meet the customer’s quality requirements and automation will reduce part delivery times.