Artificial intelligence and machine learning: the revolution of quality control, inspections, and fraud detection

Fundamentals of image recognition for machine learning

Machine-learning (ML) algorithms use massive amounts of data from a specific domain to optimize processes for more efficient outcomes. ML systems work by training themselves to recognize patterns and correlations connecting data points. Doing this requires robust algorithms, large datasets of relevant information, a narrow domain, and a concrete goal.

In the eyes of some of the top ML companies and researchers, deep learning is a breakthrough technology that can be applied to revolutionize hundreds of industries. Much of the abstract work has been done, and now it is time to use ML and turn it into sustainable business models.

For image recognition, it is crucial to correctly identify all images to distinguish the ones relevant to the dataset. The more examples of the correct pictures that an ML-based system is exposed to, the more accurately it can pick up patterns and identify differences in the field.

ML algorithms for defect detection on circuit boards

While automated visual inspection of PCBs has advanced considerably in recent years, several studies show that deep learning outperforms traditional machine-based classification and feature-extraction algorithms.^[1]

In most PCB manufacturing facilities, defects are initially detected by an automatic inspection (AOI) machine. Then a quality inspection engineer verifies each flagged PCB.

Unfortunately, the AOI machine can erroneously classify many boards as defective because of a scratch, small hole, or the presence of nanoparticles such as dust, paper fragments, or tiny air bubbles. Additionally, skilled engineers still make errors during inspection. Therefore, it is imperative that adequately trained ML systems be used to improve the accuracy of quality control procedures.

Last year, researchers from Yuan Ze University, Taiwan, published a study in which they achieved a defect-detection accuracy of 98.79% in PCBs using you-only-look-once (YOLO) convolutional neural networks.^[2]

Compared with other simple classifiers, YOLO is widely used in practice. It is a simple, unified object-detection model and can be trained directly using full images. Fast YOLO is the fastest general-purpose object detector. YOLOv2 provides the best tradeoff between real-time speed and accuracy for object detection compared with other detection systems across various detection datasets.

According to their paper, the team used 11,000 images, a network of 24 convolutional layers, and two fully connected layers. To train the Tiny YOLOv2 algorithm, the researchers used the Keras framework^[3] running on an Nvidia TITAN V GPU.

Fraud detection, counterfeit products, and component performance

In 2019, the Organization for Economic Cooperation and Development (OECD) reported that counterfeit and pirated goods made up 3.3% of global trade volumes. This percentage is rising every year. The country most affected by counterfeiting is the United States, whose brands or patents were connected to 24% of the fake products seized.^[4]

Electronics is one of the industries most affected by counterfeit products. As the price of critical components increases and there are disruptions in the supply chain, many fakes enter the market to fill the void. Making things more difficult, most of the counterfeit components are almost impossible to distinguish from the original parts.

Counterfeit detection on finished goods, especially on consumer products, can be achieved by visual inspection. However, it would take a well-trained eye and knowledge to distinguish between a fake and authentic product.

The advantage of image recognition in this regard is enormous. ML algorithms can analyze millions of images and detect even the smallest of inconsistencies and anomalies in shape, color, texture, and size.

When it comes to electronic components, it becomes more difficult to use just visual inspection. The market is plagued with counterfeit units and old stock. Using counterfeit or secondhand parts could result in equipment and property damages, product recalls, and potential liabilities.

In addition to image recognition, other methods and ML algorithms are necessary to improve accuracy when detecting counterfeit or modified parts. In the past few years, blockchain technology has quickly become one of the key solutions out there to protect the authenticity of products and components and secure the supply chain.

For example, IBM is mixing blockchain with imaging and artificial intelligence to verify the authenticity of products. Its Crypto Anchor Verifier^[5] uses ML, neural networks, and video analytics to evaluate properties of liquids and identify the color, saturation, viscosity, and other chemical properties, all as a means of better determining whether a product is fake or genuine. Additionally, it can be trained to recognize electric patterns in electronic components to distinguish new parts from previous generations of the same SKU.

The role of AI and ML in the supply chain is just starting

Recently, the disruptions of the global supply chain by the pandemic have forced many suppliers and manufacturers to look elsewhere for products and components. Using the latest technologies and ML can help avoid counterfeit parts and perform faster, more efficient, and more accurate quality control.

As image recognition and ML accuracy are constantly improving and getting faster, many industries are beginning to use them for quality control and counterfeit detection.

More industries are expected to move their operations to the cloud, increasing the availability of training data. As the capabilities of AI and ML increase exponentially, there is potential for these solutions to soon help those industries mentioned earlier improve their products’ quality and performance.