Something Doesn’t Seem Right

It has often been said, in one way or another, that the first step in solving a problem is recognizing that you have one. This makes a lot of sense intuitively — if no problem has been defined, how could a fix possibly be put in place? Since many problems can go undetected for a long time before they are noticed, it is no wonder that anomaly detection is rapidly rising in popularity, especially in the areas of industrial inspection, medical imaging, and logistics.

Anomaly detection is a process that involves identifying patterns in data that deviate from expected norms. This could involve, for example, inspecting images of a bridge to watch for small cracks or other issues that could become much bigger problems in the future if they are not addressed at an early stage. Similarly, keeping watch over industrial equipment allows maintenance to be performed before a failure occurs, potentially preventing downtime and the huge costs that come along with it.

There are two primary ways in which anomalies are detected at present by automated systems. Supervised machine learning classifiers can be trained to recognize the differences between normal and abnormal states, however, these models must be shown lots of examples of the faults that one wants to detect. These examples can be hard to come by, and moreover, one may not know all of the potential failure states of a system in advance. Unsupervised learning-based approaches can solve these issues, yet they are very computationally intensive, which quickly makes them impractical when hundreds or thousands of monitors are needed.

With today’s announcement from Edge Impulse, there is now another option. The team at Edge Impulse previously released an object detection algorithm called FOMO that takes a task that normally requires a large amount of computational resources and allows it to run on microcontrollers with just a few kilobytes of memory. They have now taken the next step with that architecture and have released FOMO-AD, and as you have probably already guessed from the name, this new model is tuned for anomaly detection.

The primary innovation in FOMO-AD involves swapping out the FOMO model’s classification head for an anomaly detector head. In particular, a Gaussian Mixture Model was included in the new architecture. This enables the model to find patterns that deviate from the training data, and thanks to the FOMO underpinnings, it is also able to locate the exact locations in an image where the abnormalities are.

As with FOMO, FOMO-AD is highly optimized so that it can run on a wide range of edge computing platforms. It is perfectly happy running on a low-power microcontroller with a minimal amount of memory, but it can also scale up to powerful GPU-based systems when necessary. FOMO-AD also solves the data collection issues associated with many existing systems — the training process is unsupervised, and no images of fault states are required.

Edge Impulse provided some examples of FOMO-AD in action. They demonstrated how defects in thermostatic valves can be found using this approach, which could have important implications in manufacturing environments. In another demonstration, it was shown how FOMO-AD can find cracks in concrete. This capability could be useful in inspections of buildings, bridges, and other infrastructure.

Extensive documentation is available if you want to see how FOMO-AD might be integrated into your own inspection processes.