Crate struktura 

Time-series anomaly detection with no training data.

Struktura detects when the structure of a signal changes: detrended fluctuation analysis (DFA) for the correlation structure, plus a streaming monitor that calibrates itself on the first rows of your data.

Quick start

use struktura::{compare, is_degraded};

// Compare current readings against a known-good baseline
let result = compare(&normal_readings, &current_readings);
println!("{}", result); // "HEALTHY shift=+0.003" or "CRITICAL shift=-0.45"

// Or just ask: is this signal degraded compared to baseline?
if is_degraded(&normal_readings, &current_readings) {
    trigger_alert();
}

Domains

• space: spacecraft telemetry monitoring (reaction wheels, magnetometers, batteries)
• market: financial regime detection (trending / random walk / mean-reverting)
• text: writing rhythm analysis (human literary prose vs mechanical/AI)
• rhythm: event timing analysis (git commits, heartbeats, keystrokes)

Works in no_std environments (default-features = false). 85-112x faster than Python.

Modules

autopilot

Autonomic layer over the hybrid monitor: detect → decide → adapt → continue, without a human in the loop.

case

A case directory: saves a recording, its incidents, detector config, and a manifest so struktura replay can reproduce and diff the analysis.

changepoint

Changepoint detection: locates WHERE the structure changed.

classify

Signal classification — what kind of signal is this?

codegen

conformal

Conformal prediction: calibrated confidence for any detector.

context

Operating context: column typing and a context timeline that pairs each sample row with its active operating mode, command, and annotations.

evolve_real

Evolve against real labeled data: the bridge between the synthetic RED/BLUE loop (crate::redblue) and actual benchmark anomalies.

ffi

fingerprint

Structural fingerprint — the “DNA” of a signal.

genome

Genome structural analysis via DFA on GC-content windows.

incident

Incident records: groups temporally proximate alarms into inspectable incidents with per-channel evidence, context, and reconstruction state. IncidentBuilder folds AlarmReports and Events into Incidents (alarms within gap ticks join; farther out starts a new one), pulling in explanations via explain_alarm and context from a ContextTimeline.

market

Financial time series regime detection via DFA.

mfdfa

Multifractal DFA (MFDFA) — reveals signals with multiple scaling regimes.

monitor

Flight-grade streaming hybrid health monitor.

prognosis

Prognosis: time-to-threshold estimation from a health-metric trajectory.

redblue

RED/BLUE adversarial self-improvement.

replay

Re-run a detector on a saved case, diff the output against the saved incidents, and report what changed.

report

Human-readable investigation report from incident records. Designed to be printed to a terminal or written to a Markdown file.

rhythm

Rhythm analysis — DFA on inter-event timing sequences.

rover

Rover health monitoring — structural analysis for planetary exploration.

rover_flight

Fixed-size rover health monitor for flight computers.

smap_eval

Telemanom-protocol evaluation with a closed-form predictor.

space

Spacecraft health monitoring via DFA structural analysis.

telemetry_bench

Coupled-spacecraft telemetry benchmark comparing DFA to the standard telemetry fault taxonomy (packet loss, spike, stuck, drift, regime shift, mixed).

text

Text structural analysis via DFA on sentence-length sequences.

trend

Trend detection: is the signal’s structure drifting over time?

Structs

BaselineTracker

BootstrapCI

CompareResult

Result of comparing two signals.

DfaResult

Result of a DFA or ACR computation.

ShuffleProof

SlidingWindow

SplitHalfResult

StructuralLaw

Complete structural analysis of a time series.

Enums

HealthVerdict

Health verdict comparing current DFA alpha against a known baseline.

LawQuality

How confident the analysis is in the derived scaling exponent.

Functions

acr

Compute autocorrelation decay exponent.

analyze

Full structural analysis of a time series.

anomaly_scores

Per-window anomaly scores from sliding DFA.

bootstrap_alpha

Subsampling confidence interval for α (Politis–Romano style).

compare

Compare two signals and get a verdict: is the structure the same?

dfa

Compute the DFA scaling exponent of a time series.

dfa_box_sizes

Box sizes that dfa, dfa_into, dfa_fast_into and dfa_scratch use for a series of n samples: up to 12 distinct sizes, geometrically spaced from max(16, n / 50) to n / 4. Returns the sizes and how many of them are valid (0 when the range is empty).

dfa_fast_into

Prefix-sum DFA: identical boxes and mathematics to dfa_into, but the per-segment sums (Σy, Σj·y, Σy²) are O(1) prefix-difference lookups instead of an O(s) pass per segment. One O(n) pass builds the profile prefixes; each of the ≤12 box sizes then costs O(n/s) segments × O(1).

dfa_into

DFA with a caller-provided buffer, avoiding allocation on the hot path.

dfa_scratch

Allocation-free DFA: the same mathematics as dfa_into, with the cumulative profile written into a caller-owned slice. Runs on no_std targets with no heap at all; a stack array or a static mut buffer works.

dfa_short

DFA for short series (from about 24 samples), or None when the series cannot be measured (too short, constant, or fewer than 3 usable box sizes).

has_changed

Has the signal’s structure changed at all?

health_check

Compare current DFA alpha against a known healthy baseline.

is_degraded

Is the current signal structurally degraded compared to baseline?

prove_structure

sanitize

Filter out NaN and Inf values from a signal.

shuffle

Deterministic Fisher-Yates shuffle of a signal.

split_half_validate