Most detection starts by describing what’s wrong — a threshold to cross, a pattern to match. That works only for the faults you’ve already seen and named. EchoSignal starts from the other end: it learns what normal looks like across all your channels, and treats anything that departs from it as worth a look. Here is how that works, step by step.
Set thresholds and patterns for known faults. Reliable for what you’ve defined — blind to what you haven’t. The anomalies with no fixed pattern slip through, because no rule was written for them.
Model what normal looks like, then surface whatever departs from it — including the anomalies you’ve never seen and could not have specified in advance. The judgment of what each one means stays with your experts.
EchoSignal ingests multi-channel data — mixed analog and digital, different signal types together, in both steady-state and transient regimes. CSV in; nothing leaves your environment. The point of detection is rarely a single channel — it’s how the channels move together.
Rather than describing faults, EchoSignal builds a baseline of normal behaviour across those channels. Mature, signal-appropriate methods do the heavy lifting per channel — time–frequency methods such as wavelets for non-stationary vibration, current-signature methods for motor electrical signals, and others — combined into one model of normal. Different methods, working together.
Anything that departs from the learned normal is surfaced as worth a second look — without anyone having defined it in advance. No predefined rule has to exist for an anomaly to be caught. The system doesn’t decide it’s a fault; it decides it’s a departure, and raises it.
Every surfaced signal goes to a person — your domain expert. Which departures are real anomalies, which are noise, and what each one means: these are calls only those who know the system can make. EchoSignal puts their attention where it belongs, and never draws the conclusion for them.
Every judgment your experts make is remembered. Over time the system grows better at setting aside the familiar, so the same kinds of departures need less attention — freeing your experts to take on what’s genuinely new and harder. The loop tightens with use.
Anomalies can’t be defined in advance. So we don’t define them.
“Model normal, surface the deviations” makes no assumption about what kind of signal it’s looking at. That independence is exactly what lets the same engine move from one domain to the next — agricultural motors, grinding spindles, sensor arrays, acoustic detection. Each gets its own domain plugin; all share one engine.
General doesn’t mean generic.
Bring your own real signals and see what the baseline surfaces. That is the most direct way to judge whether this approach fits your systems.
