Stage 1: Unsupervised Learning – Discovering Structure in Raw Data
When most people think of AI, they picture models trained with labels: a dataset of “cat” and “dog” images, or sensor readings marked as “normal” or “fault.” That’s supervised learning.
But supervised learning assumes you already know what matters. In many real-world edge applications, you don’t. You may have gigabytes of sensor streams, audio feeds, or video frames with no labels at all. The first question is: what’s even in here?
That’s where unsupervised learning comes in.
How It Works
Unsupervised learning algorithms analyze data without labels, searching for natural structure. They don’t know the “right answer”, instead, they find relationships, similarities, and oddities in the data.
Some key methods:
-
Clustering
Algorithms like k-means or DBSCAN group similar data points. For example, a vibration sensor might reveal three natural clusters: idle, normal operation, and abnormal spikes, without anyone telling the system what those states mean. -
Dimensionality Reduction
Techniques like PCA (Principal Component Analysis) or t-SNE compress high-dimensional data (like a 100-feature sensor reading) into fewer dimensions, while keeping the essential relationships. This makes data easier to visualize and process, and often exposes hidden trends. -
Anomaly Detection
By learning what “normal” data looks like, unsupervised systems can spot outliers. For instance, a sound profile that doesn’t match the typical operating noise of a generator can be flagged as an early warning sign.
The Importance of Unsupervised Learning
On Digital View AI boards, unsupervised learning is not about running the final production model. It’s about data discovery and preparation:
-
Understanding unknown environments
A board installed on a jobsite may encounter conditions the developer didn’t anticipate. Unsupervised learning highlights patterns without needing pre-defined labels. -
Building better supervised models later
The clusters and anomalies discovered here become candidates for labeling. Instead of guessing what’s important, engineers can focus on real, data-driven structures. -
Reducing data overload
Not every data point needs to go to the cloud. Unsupervised methods can filter the flood, showing where the action really is.
Example: Worksite Video Feeds
Imagine deploying a Digital View board with a camera monitoring a construction site:
- The unsupervised algorithm clusters motion patterns into groups: “workers walking,” “vehicles moving,” “areas idle.”
- No one has told it what’s what, but it’s separated the data into meaningful categories.
- Anomalies stand out, like unexpected motion at night.
Later, in Stage 2 (Supervised Training), engineers add labels: “Cluster A = forklift activity,” “Cluster B = pedestrian traffic.” That labeled dataset is then used to train a task-specific model.
The Role of Stage 1
Unsupervised learning is the exploratory phase. It doesn’t deliver a finished product, but it shines a light on the unknowns in raw data. Without it, you risk building supervised models on shaky assumptions. With it, you start with real insight, patterns discovered before a single label is applied.
This is the foundation for everything that follows: labeling, training, optimization, and deployment.
