I Will Diagnose Your Time-Series Data with Automated Frequency Band Scanning and State Change Detection

WHAT THIS IS

This is a diagnostic tool for single-channel time-series data. It automatically:
- Finds the frequency bands that contain structure
- Classifies each band into one of five types
- Detects abrupt changes in system behavior
- Compares data from different time periods

It does NOT require labeled training data, pre-defined templates, or a physical model. It uses four independent statistical metrics: coefficient of variation, sample entropy, envelope skewness, and recurrence rate.

Academic reference: Zenodo preprint DOI 10.5281/zenodo.22837806

WHAT YOU GET

A diagnostic report (PDF) with:
- Per-band metrics: CV, sample entropy, skewness, recurrence rate
- Per-band classification
- Detected transition timestamps
- Three-panel visualization (raw signal, CV over time, entropy over time)

A structured results table (CSV) containing all band metrics and labels.

A methodology note explaining the analysis flow, parameters, and how to read the results.

HOW IT WORKS — PLAIN ENGLISH

Think of your data as a radio signal. Traditional tools tell you "there is energy at 50-80 Hz." This tool tells you "the signal at 50-80 Hz is stable and structured" or "the signal at 50-80 Hz is changing in segments" or "that band is just random noise."

It also tells you the exact time when the behavior changed — for example: "at t=568 seconds, the system switched from stable to unstable."

FIVE CLASSIFICATION LABELS

- signal: stable, structured signal
- segmented: structured but changing in segments
- white_noise: flat, random noise
- nonstationary: random fluctuation
- signal_low: structure present, but data too short for full analysis

WHAT THIS TOOL CANNOT DO

- Cannot reconstruct the waveform shape
- Cannot locate a physical source
- Cannot predict the future
- Cannot process multi-channel data
- Cannot process data shorter than 500 points
- Cannot process data with a strong trend unless detrended first

WHAT YOU NEED TO PROVIDE

1. Your time-series data file
Format: CSV, TSV, TXT, or HDF5
Content: at least one numeric column (timestamp optional)
Length: 1000 points minimum (5000+ recommended)

2. Sampling rate
Examples: one sample per second, one per millisecond, LIGO 16384 Hz

3. Your goal
Examples:
- Find which frequency band contains a signal
- Detect the exact time of a state change
- Compare structure between two data periods

WHEN TO USE IT

- You have a long, single-channel time series
- You suspect the system behavior changed at some point
- You want to distinguish "power glitch" from "structural change"
- You need an automated first-pass screening before deeper analysis

WHEN NOT TO USE IT

- Real-time streaming
- Multi-sensor arrays
- You need physical explanation
- You need fault localization
- You need future prediction

PRICING

- Basic: one dataset, report + CSV (contact for quote)
- Standard: up to 3 datasets, comparative analysis, one Q&A round
- Advanced: batch analysis, custom parameters, source code license

FAQ

Q: How is this different from FFT / PSD?
A: PSD tells you "how much power at each frequency." This tool tells you "what kind of structure is at each frequency, and when it changes."

Q: Can it locate the fault?
A: No. It only has one-dimensional time-series data, no spatial information. It tells you "the structure changed," not "where it changed."

Q: What is the time precision?
A: Depends on window length. With a 1-second window, transition detection is accurate to about ±1 second.

Q: Can it handle my data?
A: First check: length ≥ 1000 points, no strong trend, and a clear change expected. If so, it should work. If not, the tool will tell you it cannot analyze.