Why Machine Learning Matters
As we get closer to Gartner’s projected 21 billion connected things by 2020, we have to seriously consider amount of data being generated by all of these sensors, devices and equipment. In fact, the technology industry cannot keep up even as the demand for data scientists reaches new heights. While privacy and security concerns receive much of the attention, our basic ability to process this information requires a better approach: enter machine learning.
As a subfield of Artificial Intelligence (AI), machine learning utilizes software algorithms which enable machines to learn a model from given information and to generalize that model to new information without being explicitly programmed to handle this new information. This means that a machine can – through experience – improve the way it analyzes, decides and acts upon data it receives. This learning ability is critical for analyzing data from the Internet of Things (IoT) because the machine learning algorithms can generalize and learn from new data without the supervision of a data scientist.
How Machines Learn
One of the most common types of machine learning used for IoT applications is called supervised machine learning. In supervised learning we begin with training dataset which includes known outcomes called labels.
For example, if you want to identify handwritten characters in a foreign language – say Japanese – a supervised algorithm will take thousands of pictures of hand-written characters along with labels containing the meaning in English. The algorithm will learn the connection between these characters and the associated meaning in English, and apply that relationship to classify known characters in images it’s never seen before. This is how Google can translate images.
In the construction or manufacturing industries these labels might represent the machine “health status” (good/bad), “performance status” (on/off) or some other information on the state of the machine. This labeled training dataset is necessary for supervised algorithms to learn from and deliver high-accuracy results, however, this approach is laborious, expensive and time-consuming. Additionally, there is the problem that a model trained on labeled data is inherently less generalizable – even when based on a large labeled data set – because the training process relies on specific training examples.
Furthermore, in IoT we often lack these training datasets making it nearly impossible to use supervised algorithms and must rely on techniques which don’t need training, known as unsupervised machine learning, which are often considered less capable, less accurate due to model-assumptions and less adaptive. So how do we build an unsupervised learning algorithm that doesn’t need training BUT is comparable in accuracy, performance and adaptability to the supervised algorithms?
Unsupervised Self-Learning Algorithms
Supervised machine learning models are data-driven and often the preferred choice for predictive IoT analytics models. In many practical settings, the learned model improves as the amount of data available to train the model grows and the training time increases.
Our new unsupervised self-learning algorithms are doing the same thing, but are training themselves on the real-time data that becomes available during day-to-day operations. When combined with a set of proprietary heuristics and rules, these self-learning algorithms are capable of getting better and better at understanding what is normal and what is not for a given machine or fleet of machines over time without any user supervision.
The self-learning algorithm is similar to leaving an intelligent device alone in a large house and allowing it to move around freely to try different things such as opening doors to many rooms, going up into the attic or down into the crawl-space, re-charging from different power sockets at different floors, and learning all it needs from its own experience without any supervision. The longer the device does this and the better it understands the surrounding environment, the more it is able to figure out what is normal and what is not, what is dangerous and what is not, and what is best to do next.
Self-Learning Maintenance Algorithm (SeLMA)
At Alchemy IoT, we apply our self-learning algorithms to the device health, uptime, failure detection and prevention to improve predictive and prescriptive maintenance programs and enable clients to derive more value from their assets. We call this solution SeLMA or Self-Learning Maintenance Algorithm. By enabling client’s machines to learn about their own behavior and experiences without supervision we empower them to not only keep their machines healthy and running, but to avoid costly time-consuming consulting engagements as well.
The main advantage that SeLMA offers is its pervasive, universal technology that can be applied equally well to a car, a tractor, a power transformer, or a hard disk drive. As long as the data is available, we know our algorithms can learn ‘normal’ behavior and report on observed anomalies. Instead of months of laborious efforts by trained data scientists searching for the best predictive model, we are offering an algorithm that can learn on its own. And, in the case when the data is readily available, SeLMA could be trained offline in batch analysis mode to accelerate its learning.
At Alchemy IoT, we’re opening a new chapter in IoT asset intelligence and equipment maintenance. Contact us to learn more about SeLMA, our Clarity application, or to see a demo of our solutions, we’d love to hear from you!