Struct CusumDetector 

pub struct CusumDetector {
    pub k: f64,
    pub h: f64,
    pub s_hi: f64,
    pub s_lo: f64,
    pub target_mean: f64,
    pub target_std: f64,
}

CUSUM (Cumulative Sum) sequential change-point detector.

Detects a sustained shift in the mean of a metric stream. The algorithm maintains two accumulators:

• S_hi detects upward mean shifts (metric increasing).
• S_lo detects downward mean shifts (metric decreasing).

Parameters

• k — allowable slack, typically 0.5 * σ * shift_size_in_sigmas.
• h — decision threshold (number of standard deviations of the cumulative sum before an alarm is raised). A common choice is 4.0 – 5.0.

Example

use trustformers_debug::regression::statistical::{CusumDetector, ChangeDirection};

let mut cusum = CusumDetector::new(0.0, 1.0, 0.5, 4.0);
// feed values well above the target mean — should eventually alert
let mut alerted = false;
for _ in 0..20 {
    if let Some(a) = cusum.update(3.0) {
        assert_eq!(a.direction, ChangeDirection::Up);
        alerted = true;
        break;
    }
}
assert!(alerted, "CUSUM should have detected upward shift");

Fields

k: f64

Allowable slack (reference value).

h: f64

Decision threshold — alarm when S_hi > h or S_lo > h.

s_hi: f64

Upper cumulative sum.

s_lo: f64

Lower cumulative sum.

target_mean: f64

Target (in-control) process mean.

target_std: f64

Target (in-control) process standard deviation.

Implementations

impl CusumDetector

pub fn new(target_mean: f64, target_std: f64, k: f64, h: f64) -> Self

Creates a new CUSUM detector.

Arguments

• target_mean — in-control mean.
• target_std — in-control standard deviation (used to normalise inputs).
• k — slack parameter (0.5 is a common default for detecting 1σ shifts).
• h — threshold (4.0–5.0 gives low false-alarm rates for Gaussian inputs).

pub fn update(&mut self, value: f64) -> Option<CusumAlert>

Incorporates a new observation and returns an alert if a change point is detected.

The observation is first standardised as z = (value − target_mean) / target_std (unless target_std == 0, in which case the raw deviation is used).

After firing, the triggering accumulator is reset to zero so detection can resume. Callers that wish to track sustained changes should call reset manually instead.

pub fn reset(&mut self)

Resets both cumulative sums to zero without changing parameters.