知几 · MurSense

• One question in, a verdict out. "How's the bridge doing today?" returns a health verdict with an alert level, the reasoning behind it, and the specific channels to look at — computed in seconds (cache hit) to under two minutes (full recompute), not hours.
• Compresses the whole workflow into one conversation. Data → analysis → diagnosis → plain-language interpretation, in a single exchange — instead of hand-running scripts and reading raw spectra.
• Standard engineering output. PSD spectra, modal frequencies, strain RMS, five-level color-coded alerts with explicit thresholds, and a one-click archivable Word report.
• Built for non-specialists, not just signal-processing experts. The whole experience — dashboard, assistant, reports — runs in Chinese, English, or French, switchable in one click.

The main workspace: one question box, one-click example prompts for every capability, and a "Load Identification" card explicitly marked Planned — not something the product claims to do today.

MurSense is built on a strict division of labor: the language model understands the question, decides which analysis tool to call, and explains the result in plain language. It never writes analysis code on the fly and never touches the raw waveform directly. The actual computation — PSD via Welch's method plus peak detection, sign-based preprocessing to keep modal peaks identifiable under heavy traffic transients, a fully unit-tested signal chain (detrending, filtering, calibration) — runs in a separate, deterministic algorithm package. A conclusion on screen is never the model "imagining" a plausible-sounding number; it's read off a real computed result.

Alert	Relative change η	Meaning
Green	η ≤ 1%	Normal, no action needed
Blue	1% < η ≤ 2%	Minor change, routine observation
Yellow	2% < η ≤ 4%	Worth watching, increase monitoring frequency
Orange	4% < η ≤ 8%	Clear anomaly, on-site check recommended
Red	η > 8%	Significant anomaly, immediate investigation recommended

Five-level alerting, based on the relative change η of a structural feature against its own baseline — never a hardcoded absolute threshold pulled from thin air.

Real sensor data → validated analysis code → the AI agent selects and runs the right tool → a structured diagnosis with figures and a report.

• Built-in dataset is a real recording, not synthetic sample data — distributed acoustic sensing waveforms from an actual in-service railway bridge: 1,034 channels along the span, 100 Hz sampling, continuous.
• Every figure on screen is computed at question time. PSD maps, modal-frequency picks, RMS distributions, spatial anomaly maps — validated algorithms run on the raw recording live; nothing in the demo is a canned image.
• The tool-call chip above every analytical answer names the exact function that ran (check_bridge_health, run_bridge_analysis) — a visible, checkable link between what's on screen and the code that produced it.

A live answer: the tool that ran, a full health assessment with baseline-compared metrics, and a spatial deviation map — computed on the real bridge recording, not staged.