Eine fiktive Kühlwasserpumpe sendet Vibrations- und Temperaturwerte. Das System lernt zuerst den Normalzustand und berechnet dann Anomaliewert, Gesundheitsindex und Restlebensdauer. Injizieren Sie einen Fehler wie Lagerverschleiß und beobachten Sie, wie früh die Warnung kommt und wie breit das Unsicherheitsband ist.
A fictional cooling-water pump streams vibration and temperature values. The system first learns the normal state, then computes anomaly score, health index and remaining useful life. Inject a fault such as bearing wear and watch how early the warning comes and how wide the uncertainty band is.
Centrifugal cooling-water pump (fictional asset) · simulated vibration, temperature and process sensors · anomaly score, health index and remaining useful life computed live
A centrifugal pump is simulated as a set of coupled sensor streams (about four samples per second; each sample represents 15 minutes of operation, so one real second ≈ one operating hour at 1×). Each signal follows a mean-reverting random walk around its setpoint; slow natural wear and any injected fault add physically plausible offsets: bearing wear raises the high-frequency vibration band and the bearing temperature, imbalance raises the 1× rpm band and overall RMS, cavitation raises broadband vibration while flow, discharge pressure and motor current fall and become noisier.
The model layer learns a baseline from the first 120 samples (30 h) (mean and standard deviation per signal), converts every smoothed reading into a
z-score and compresses the largest deviation into an anomaly score a = 1 − exp(−max(0, |z|max − 2.5) / 4). The
health index is 100 · (1 − ā) with ā an exponentially smoothed anomaly. For the remaining useful life a least-squares line is fitted to the
health index over the last 100 samples (25 h) and extrapolated to the failure threshold of 30; the slope's standard error from the fit residuals gives the
uncertainty band (slope ± 1.96 SE). A trend is only reported when the decline is statistically significant (|slope| > 2 SE).
The recommendation combines status and RUL (no action · plan within · stop within) and names the most likely fault by simple pattern matching on
which bands are elevated: HF + temperature → bearing, 1× dominant → imbalance, broadband + flow/pressure drop → cavitation.
Everything you see is synthetic and computed locally in your browser — no sensors, no server, no customer data. The asset, its tag and all values are fictional; the random sequence is seeded, so a reload reproduces the same run. In a plant, the same pipeline would run on a historian or edge gateway and feed the CMMS work-order queue.