How NOBU works
Most monitoring systems wait for something to go wrong. NOBU works differently — it actively tracks what it doesn't yet understand about your system, and problems surface in those gaps first.
NOBU vs. traditional approaches
| Traditional monitoring | NOBU | |
|---|---|---|
| Warning time | After threshold crossed | Before readings look unusual — scales with sampling rate |
| Setup | Weeks of baseline collection | Starts immediately, no training data required |
| False alarms | Frequent — static thresholds don't adapt | Significantly fewer — confidence reflects actual system state |
| Confidence | Binary: alarm or no alarm | Graded: "How certain am I that something is changing?" |
| Adaptation | Manual recalibration after changes | Continuously adapts to operating conditions |
| New failure modes | Can't detect what it wasn't trained for | Detects anything that increases system uncertainty |
Why the difference exists
Traditional: learn what's normal
Most systems build a model of "normal" from historical data, then alarm when readings deviate. This fails on new equipment, after maintenance, or for failure modes never seen before. It also means weeks of collecting baseline data before you get any value.
NOBU: track what's uncertain
NOBU maintains many simultaneous hypotheses about your system's current state, weighted by how well each explains incoming data. It doesn't need to know what "normal" looks like — it identifies emerging problems by finding where its own uncertainty is increasing.
Warning time scales with how often you sample. At one reading per minute, NOBU typically identifies emerging problems within a handful of readings — well before threshold-based alarms would trigger. The exact lead time depends on how fast the problem develops and your sampling rate.
On confidence
When NOBU reports 85% confidence, that number comes from tracking many possible states of your system simultaneously. It's not a black-box anomaly score — it's a probability grounded in how well each hypothesis explains what the sensors are actually reporting.
This means the confidence level is actionable: high confidence means act now, medium means increase monitoring, low means note and move on. Over time, as NOBU sees more of your system's behavior, its estimates become more precise.
Two ways to integrate
Agent (continuous monitoring)
Install the NOBU agent on a laptop, Raspberry Pi, or industrial PC connected to your sensors. It collects data, forwards it to NOBU, and you monitor results in your dashboard. Available for Linux, Windows, and macOS.
API (on-demand inference)
For Control and API Fuzzing services, call the REST API directly from your application. Send sensor data or system state, get back decisions with confidence levels. Billed by compute time.
Getting started
Choose a service
PowerSense for power quality, Predictive Maintenance for equipment health, or API services for control and security testing.
Install the agent
Download the agent for your platform, set your API key in config.yaml, and start the service. Takes about 5 minutes.
Open your dashboard
Watch NOBU learn your system in real time. First useful predictions typically appear within an hour of data collection.