Mastering Anomaly Detection in Manufacturing: Key Benefits, Best Practices, and Use Cases for Implementation

Sponsored by

Anomaly detection is an important deep learning tool for industrial quality inspection. Mateusz Barteczko, manager of computer vision application engineering at Zebra Technologies, recently hosted a webinar with A3 to dive deeper into this topic. The webinar expands upon the concept of anomaly detection, exploring key benefits and best practices while sharing use cases across industries. Below we’ll explore the fundamental insights within the presentation.

You can watch the full webinar, "Anomaly Detection in Industrial Manufacturing: Best Practices for Successful Implementation," for free here.

Anomaly Detection in Machine Vision Applications

Unlike the deep learning models that are trained on a labeled dataset — referred to as supervised learning — anomaly detection is considered unsupervised learning. Therefore, it can identify defects without prior knowledge of them by learning from a dataset of “okay” samples. With anomaly detection, says Barteczko, the algorithm can discover hidden patterns of data without human input.

Benefits of Anomaly Detection

There are many advantages to utilizing anomaly detection within your production line. For instance, the quality inspection training process with anomaly detection is typically faster than supervised methods of inspection because it only needs images of defect-free products to collect data and enable learning. This can allow the manufacturer to deploy the model sooner and adapt quickly to changes in the product or process.

Zebra offers model training on a local computer with a GPU or on a cloud service with its Zebra Aurora Vision Deep Learning™ software. Offering a set of deep learning-based tools, this software improves the quality and operational efficiency of existing machine vision solutions. “It’s fairly easy to manage your datasets properly when you use our Aurora Vision software with its user-friendly editor for training deep learning models,” says Barteczko.

Best Practices for Training Anomaly Detection Models

Quality data is key to successful anomaly detection. For achieving accuracy and reliability, preparing a well-constructed dataset is imperative. Diversity within the dataset is a key factor to the model’s effectiveness. This means, offering images with variations in lighting, orientation, and from multiple batches of product. “In some processes,” says Barteczko, “It's also a good practice to capture images at multiple moments during the production process. For example, instead of installing your camera and capturing 50 images in one minute, take the first 10 images, then wait a few minutes and capture the next 10 images and repeat this a few times.”

Being sure the model does not overfit to the sample is an important consideration in your data preparation, too — a point Barteczko elaborates on in his presentation. He strongly suggests also including “not okay” samples as a final step before you deploy to see how well the model recognizes potential defects.

Related Blog Articles

Robots Revolutionize Grocery Shopping: The Future of Automation in North American Stores

Read

Jumpstart Your AI Projects: Explore Top AI Applications at the AI & Smart Automation Conference

Read

Barteczko also cautions against mixing training and test datasets. The test set should not include images used in training so that you can accurately evaluate whether the model can generalize new data.

A Real-World Application of Anomaly Detection

Anomaly detection can be utilized in a variety of industries on a multitude of products. In this webinar, Barteczko offers use cases for climbing rope, terminal block, car interior panel, thermoform cap, and sushi box inspection. The samples offer real-world examples of the need for a diverse training set. They also highlight products with low variability while spotlighting the circumstances in which anomaly detection was the most effective tool to find defects.

Of the use cases, the sushi box served as an example of the increased challenges of quality inspection in food packaging. In production samples like this, no two products are the same, and variations in size, shape, and color are to be expected. In this example, there are many objects within the sushi box including chopsticks, ginger, wasabi, sushi, and more, adding to the variables challenging the inspection process.

In this instance, the model was able to correctly detect a defect of a foreign object that was hiding some of the rolls. It correctly did not flag other differences based on placement because the training dataset included images with multiple variations of object orientation. Challenging products like this with increased acceptable variations require much more data in the training set to be able to effectively distinguish anomaly from variable.

To learn more about each use case, you can watch the webinar on demand now.

About Zebra

Zebra helps organizations empower their frontline, ensuring that everyone and everything is visible, connected and fully optimized. The company's portfolio spans software to innovations in robotics, machine vision, automation and digital decisioning, scanning, track-and-trace, and mobile computing solutions.