When designing a printed circuit board (PCB) for a new application, every care is taken to make best possible use of the available space and to position components as close together as possible without risking a failure. So far, this process relies largely on the experience of the engineers, whose designs must then be tested in trials. A further complication is that results are not stringently documented, meaning that error-prone designs undergo repeat testing, which leads to increased costs.
The finished designs can then set high demands on manufacturing. For this reason, each completed PCB undergoes, at the very least, automated optical inspection (AOI). This uses image analysis techniques in order to determine that the PCB has been produced as per design and therefore does not have any technical defects. So far, however, this method generates a high false negative rate, i.e., a lot of fully functional PCBs are incorrectly classified as defective.
These supposedly defective PCBs must then be inspected once again by hand, either visually or by means of measuring equipment. In other words, an unacceptably high false negative rate means that non-defective PCBs are being rejected and then require reinspection, which in turn results in higher costs. On the other hand, if this rate is too low, the follow-up costs are high as a result of defective components entering the supply chain. It is difficult to achieve an ideal true positive/false negative rate based on human inspection, since human errors also enter into the equation.
Optimal selection based on self-learning techniques
A development of the Fraunhofer Institute for Applied Information Technology FIT shows what a future inspection process can look like. As in conventional AOI, a camera records images of the PCB. This improves the quality of the decisions made by the algorithms. Here, it is vital that the modules are provided with high-quality training data. Initially, the modules for machine learning and deep learning are fed with a good selection of data.
“The modular design means that we can harness several algorithms, which continually enhance their own performance. Data generated by ongoing automated inspection of components is fed back into the algorithm. This then provides the basis for a process of self-learning by the artificial intelligence module,” explains Timo Brune, project manager at Fraunhofer FIT. “This permanent feedback enhances the database and optimizes the true negative rate. Early estimates from industry indicate this could reduce the use of production resources by around 20 percent.”
Users can train the modules themselves, on the basis of their own process and manufacturing data. This means that companies retain control of their own data and are not required to send it to an external server, for example. The toolbox of algorithms can be applied to specific problems in any combination.
Intelligent design of new components
Once trained, the algorithms can also be used to design new PCBs. This ends the lengthy and costly procedure of trial and error whereby components are arranged on the board until the optimal configuration is found. Instead, the algorithm helps to predict which of the many possible variations will perform optimally.
Fraunhofer FIT's approach of using modular, self-improving algorithm platforms for design and quality control of printed circuit boards can also be beneficial for many other electrical systems. Also there, processes can be optimized in order to achieve significant savings in time and production costs.