Quality has become one of the key differentiators among products and various brands all around the world. This makes them possible to preserve the brand reputation. The main reason behind this is the application of industrial automation and digitalization that connects the whole production operations. This reality inspires more companies to explore advanced technologies to improve their quality processes. The recent technology in this area is machine learning. Machine learning enables systems the ability to automatically learn and improve from experience without being explicitly programmed. Artificial intelligence and machine learning to have made a huge impact on human interaction. Even self-driving cars and driverless buses are on their way to being rolled out into the mainstream.
The empowerment of machine learning in the mobile industry is both exciting and overwhelming. Many mobile phone manufacturers have started to apply the technology to their mobile phone production operations. They use machine learning software to discover the design and production issues in real time. It can strengthen quality control on the line. Sametime can streamline their issue, response to deliver the latest products on demanding scheduled timelines.
Combining the principles of machine learning with computer vision technology can develop more advanced sensors and process control schemes. In streamline development and production area, a solution combines easy-to-deploy inspection stations with intelligent software that helps engineering and operations teams discover, fix, and monitor issues on the assembly line. Thus, it made possible to accelerate issue discovery, failure analysis, and corrective actions during the hardware development process.
Steps they have gone through
A popular mobile brand explained the steps have gone through when they first implemented machine learning and artificial intelligence in their production process. They first identified a handful of mobile phone assembly states. It highlighted all the key components of the phone as it was built. With this information in hand, they built and deployed inspection stations. That station consisted of cameras, tunable lighting, and customized fixtures. They set this in less than three weeks.
Cameras are a crucial factor in these inspection stations. They built a 20- megapixel Flir camera having no Intellectual property in it. Then integrated it with the software to work as a test station on the line. As they are trying this for the first time, they built a low-cost station. The main purpose of the cameras is taking pictures and scanning bar codes. The same time intelligent software does the analysis in real time to give a pass/fail result. As soon as the images were collected, they were uploaded to the manufacturer’s database. This makes it possible for their engineers all around the world access the data through the web application of the manufacturer. Considering the traditional industrial vision systems, we can say applications were incredibly specific there. Previously, the data were unavailable to the team because it was trapped on the local machines.
Initialization with PCB
One of the initial areas of its mobile phone production operations to which the mobile phone brand applied machine learning software was on the phones’ PCBs (printed circuit boards). The industry standard for PCB inspection is Automatic Optical Inspection (AOI) systems. The image of a PCB is compared to a digital CAD file to make sure each part is present. A limitation of AOI is that it cannot find new defects or damage that happened to PCB. Modern miniaturized devices may have undergone additional assembly steps. AOI cannot analyze that in PCBs.
When machine learning algorithm applied for the first time on PCB, dark blotches were detected on the subset of the PCBs. After examination, the engineers found out that the blotches were aligned with areas that connect two or more inner layers of the PCB(buried vias). Engineers investigated further and found that the boards were thicker than specified. This created serious problems with their critical tolerance stack-ups (calculations of the maximum, and minimum distance between two features or parts). As a source of the problem, this was very difficult to track down with AOI. Using the data from machine learning software, engineers found out that these PCBs were from the same vendor. So they worked with the supplier to rectify the problem.
Algorithms that are self-programmed
Once the first 30 units from the build were completed, engineers can use machine-learning algorithms to find new defects. Those defects weren’t previously aware of. After finding out the defect, subsequent units were set up to test for the same failure mode. Defects rates and trends are calculated in real-time after sorting failures automatically. There are a lot of things that a traditional industrial vision system cannot do. The machine-learning methods enable each machine learning software app(monitor) to learn the difference between a typical and an anomalous unit. Thus, possible to set up tests that can find unforeseen defects automatically.
Intelligent-software app
The monitor is described as a software app. This allows engineers to set up a test that runs only in the cloud. Monitor validates that it’s catching the things the engineers want to identify. Then a drop-down selection will push the self-programmed algorithm into the product’s recipe at the edge. The training of the algorithm takes place in the cloud with computing done at the edge. This occurs on a normal graphics card in an off-the-shelf computer.
A key feature of this prominent manufacturer’s machine learning technology is nothing but the self-programmed algorithm capability. This removes the need for a company to employ data scientists to apply machine learning to their production operations. The software is designed in a way such that an engineer provides the expertise looking at a part to determine what is defective. It is possible for engineers to sort or filter images by key parameters or places along the line by visiting a web app. They can tell the system where to look. Then the algorithms run and return a stack rank. This makes available the stack rank from most anomalous to the least anomalous one. Finally, the software user can then draw a threshold between what’s bad or good. They can also give a name like shifted part, titled switch, etc.
Quality and customer satisfaction
It is possible to find known as well as unknown issues with machine learning algorithms. If the test is conducted with a good product, then it will be easy to get the result that the client is looking for. The client’s main motive is customer satisfaction through the quality of the product(mobile phone) they manufacture. Because nobody wants a failing example to set up a test. If done so then the result will be a total failure. Another important feature is that this particular machine learning system only requires 30 images to start with. Whereas another manufacturer’s machine learning system may require more than ten thousand images.