sprint-8: EKF + adaptive tuner + HWID + SHA-256 audit hash-chain

- heading_ekf.py: 2-state Kalman filter fusing PGN 127250 heading and
  127251 ROT with shortest-arc innovation and symmetric covariance update
- adaptive_tuner.py: gradient-descent outer-loop Kp/Ki adjuster bounded
  to ±adaptive_max_deviation_pct; oscillation vs steady-state detection
- hwid.py: HMAC-SHA256 activation token (verify side); hwid_from_mac_words
  converts three Modbus uint16 MAC words to 12-char hex HWID
- audit.py: SHA-256 hash-chain -- each JSONL line carries prev_hash and
  line_hash; verify_chain() detects tampering, deletion, insertion
- firmware/system/hwid.h+cpp: esp_efuse_mac_get_default wrapper + FNV-32
  hash + "AA:BB:CC:DD:EE:FF" formatter
- modbus_registers.yaml + generated .h/.py: HWID_MAC_01/23/45 at
  input addrs 9/10/11 (three 16-bit words = 6-byte MAC)
- modbus_slave.cpp: INPUT_HWID_MAC_01/23/45 cases read eFuse MAC
- main.cpp: logs HWID string + FNV-32 hash at boot (activation traceability)
- tests: 72 new tests (audit signing, EKF, adaptive tuner, HWID) -- 398 total

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-05-20 03:07:27 -04:00
parent 5f9b445572
commit 45642fda0e
15 changed files with 1217 additions and 5 deletions
+169
View File
@@ -0,0 +1,169 @@
"""2-state heading EKF (Extended Kalman Filter) -- Sprint 8.
Fuses NMEA 2000 heading (PGN 127250) and rate-of-turn (PGN 127251) into a
smoothed, low-latency heading/ROT estimate for the outer PID loop.
State vector x = [heading_deg, rot_dps]
Process model (constant-ROT):
h_{k+1} = h_k + rot_k * dt
r_{k+1} = r_k (ROT modelled as random walk)
Measurements:
z_heading = h_k + v_h (v_h ~ N(0, R_h))
z_rot = r_k + v_r (v_r ~ N(0, R_r))
Angles are kept in the range [0, 360) for the state but the update step
works on signed shortest-arc differences to avoid wrap-around errors.
Usage::
ekf = HeadingEKF()
# On each 10 Hz tick:
ekf.predict(dt_s=0.1)
if new_heading_available:
ekf.update_heading(heading_deg, noise_deg=2.0)
if new_rot_available:
ekf.update_rot(rot_dps, noise_dps=1.0)
h, rot = ekf.heading_deg, ekf.rot_dps
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
@dataclass
class HeadingEKF:
"""2-state linear Kalman filter for heading and rate-of-turn.
Parameters
----------
heading_deg:
Initial heading estimate (degrees, 0-360).
rot_dps:
Initial rate-of-turn estimate (degrees per second, signed).
process_noise_heading:
Process noise variance for heading (deg²). Larger = trust model less.
process_noise_rot:
Process noise variance for ROT (deg²/s²).
"""
heading_deg: float = 0.0
rot_dps: float = 0.0
# Process noise (Q matrix diagonal)
process_noise_heading: float = 0.01 # deg²
process_noise_rot: float = 0.1 # (deg/s)²
# Covariance matrix P (2×2, stored as flat [p00, p01, p10, p11])
_P: list[float] = field(default_factory=lambda: [1.0, 0.0, 0.0, 1.0])
# ------------------------------------------------------------------
def predict(self, dt_s: float) -> None:
"""Propagate the state and covariance forward by ``dt_s`` seconds."""
h = self.heading_deg
r = self.rot_dps
# State transition: heading += rot * dt
self.heading_deg = (h + r * dt_s) % 360.0
# Jacobian F = [[1, dt], [0, 1]]
dt = dt_s
p00, p01, p10, p11 = self._P
# P = F P Fᵀ + Q
new_p00 = p00 + dt * p10 + dt * p01 + dt * dt * p11 + self.process_noise_heading
new_p01 = p01 + dt * p11
new_p10 = p10 + dt * p11
new_p11 = p11 + self.process_noise_rot
self._P = [new_p00, new_p01, new_p10, new_p11]
def update_heading(self, measured_deg: float, noise_deg: float = 2.0) -> None:
"""Kalman update step for a heading measurement.
Parameters
----------
measured_deg:
Raw heading from PGN 127250, degrees [0, 360).
noise_deg:
Standard deviation of the sensor noise (degrees). Variance = noise²
"""
R = noise_deg * noise_deg
p00, p01, p10, p11 = self._P
# Innovation (shortest arc)
innov = _shortest_arc(measured_deg, self.heading_deg)
# H = [1, 0] (observe heading only)
# S = H P Hᵀ + R = p00 + R
S = p00 + R
if S == 0:
return
# Kalman gain K = P Hᵀ / S → [k0, k1] = [p00/S, p10/S]
k0 = p00 / S
k1 = p10 / S
# Update state
self.heading_deg = (self.heading_deg + k0 * innov) % 360.0
self.rot_dps += k1 * innov
# Update P = (I - K H) P
self._P = [
(1 - k0) * p00,
(1 - k0) * p01,
p10 - k1 * p00,
p11 - k1 * p01,
]
def update_rot(self, measured_rot_dps: float, noise_dps: float = 1.0) -> None:
"""Kalman update step for a rate-of-turn measurement.
Parameters
----------
measured_rot_dps:
Raw ROT from PGN 127251, degrees per second (signed).
noise_dps:
Standard deviation of the ROT sensor noise (deg/s).
"""
R = noise_dps * noise_dps
p00, p01, p10, p11 = self._P
# Innovation
innov = measured_rot_dps - self.rot_dps
# H = [0, 1] (observe ROT only)
# S = p11 + R
S = p11 + R
if S == 0:
return
# Kalman gain: [k0, k1] = [p01/S, p11/S]
k0 = p01 / S
k1 = p11 / S
# Update state
self.heading_deg = (self.heading_deg + k0 * innov) % 360.0
self.rot_dps += k1 * innov
# Update P
self._P = [
p00 - k0 * p01,
(1 - k1) * p01,
p10 - k0 * p11,
(1 - k1) * p11,
]
@property
def covariance(self) -> tuple[float, float, float, float]:
"""Return (p00, p01, p10, p11) — the 2×2 covariance matrix."""
return tuple(self._P) # type: ignore[return-value]
def _shortest_arc(a: float, b: float) -> float:
"""Signed shortest-arc from ``b`` to ``a`` (degrees)."""
diff = (a - b + 180.0) % 360.0 - 180.0
return diff