Hero section background image

Detecting anomalies using robots with a new method of unsupervised learning AI

Detecting anomalies using robots with a new method of unsupervised learning AI

When operating large production facilities, such as power plants, factories, and refineries, there is a requirement that all equipment runs in a safe, secure, and efficient manner. Facility operators conduct routine inspection rounds multiple times a day to ensure this requirement is met. The process is laborious, sparse, costly, and prone to subjective biases.

With recent advances in robotics and automation, multiple industries are shifting to use these technologies to conduct inspections in a completely unmanned process. To meet the demand, new AI systems are being designed to mimic what an operator would assess while conducting their inspections. 

Lloyd Windrim

Designing the right AI systems can be a challenge. Consider the following safety hazards: a water leak could be a signal of corrosion; an unknown parked vehicle may indicate a security threat, or a toolbox in a walkway could lead to a tripping incident. Anomalies present themselves in a variety of shapes and forms, making it very difficult to characterize them within AI systems.

Unsupervised learning algorithms do not require labels to learn. This enables our system to detect anomalies it has never seen before.

State-of-the-art supervised machine algorithms rely on a pre-defined library of anomalies and a large number of training examples to help them characterize the anomaly into an algorithm. The nature of many anomalies is that they are a rare occurrence. The endless variety of anomalies and their potential sparsity create a near-impossible task for a supervised learning algorithm to accurately detect anomalies. 

Abyss Solutions has developed a technology for anomaly detection that uses a novel unsupervised machine learning algorithm at its core. This means that our algorithms only need examples of normal situations to learn from (not anomalies). Once our team defines what is normal about the data, we can detect abnormalities without ever having witnessed them before. Unlike change detection, algorithms commonly found in security applications, our algorithms learn to be robust to the natural changes in the environment from which the data was captured. This means that our system triggers alarms for anomalies, and not every time the wind blows. Abyss has the capability to deploy this technology to CCTV cameras, IoT devices, or even robots such as “Spot.”

Detecting anomalies using robots with a new method of unsupervised learning AI

Growing with Abyss Solutions