Fix: CsI(Tl) non-linear response correction + detector calibration overhaul
Root cause of Am-241 misidentification: the Radiacode 103's CsI(Tl) crystal shifts low-energy peaks upward (59.5 keV → 71.6 keV for Am-241) due to non-proportional scintillation response. The model was trained on theoretical peak positions and couldn't match the shifted real peaks. Changes: - Add inverse CsI(Tl) non-linear correction to inference pipeline (radiacode_monitor.py, web/config.py, test_detection.py) E_apparent = E_true * (1 + 0.37 * exp(-E_true/100)) Corrects channel mapping so peaks appear at theoretical energies - Fix energy calibration: DetectorConfig now uses E = 0.33 + 2.97*ch with 1023 channels, matching the real detector (was energy_min=20, skip_first_channel=True, different channel width) - Add K-escape peaks for CsI(Tl) iodine X-ray escape (E - 28.5 keV) - Add asymmetric peak shapes for low-energy tails (< 200 keV) - Add log1p normalization in dataset and inference (replaces max-norm) - Add background-subtracted training mode (subtract_background flag) - Add low-signal augmentation (0.01-5 Bq activities, 30-300s durations) - Update docker-compose.yml: batch_size=32, duration=30-300s, CSI_NONLINEAR_ALPHA/BETA env vars for detect and web - Web dashboard: apply CsI correction to displayed spectra - Various UI fixes (Chart.js width, zoom/pan, isotope lines) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Jacquin Antoine
202
test_detection.py
Normal file
202
test_detection.py
Normal file
@ -0,0 +1,202 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Test isotope detection: vega_best vs vega_final, with/without background subtraction."""
|
||||
|
||||
import os
|
||||
import sys
|
||||
import json
|
||||
import numpy as np
|
||||
import torch
|
||||
from pathlib import Path
|
||||
|
||||
# Paths (container-mounted)
|
||||
MODELS_DIR = Path(os.environ.get("MODELS_DIR", "/models"))
|
||||
DATA_DIR = Path(os.environ.get("DATA_DIR", "/data"))
|
||||
VEGA_ML_PATH = Path(os.environ.get("VEGA_ML_PATH", "/models/vega_ml"))
|
||||
|
||||
# Add vega_ml to path
|
||||
sys.path.insert(0, str(VEGA_ML_PATH))
|
||||
from training.vega.model import VegaModel, VegaConfig
|
||||
from training.vega.isotope_index import IsotopeIndex
|
||||
|
||||
# Energy calibration
|
||||
ENERGY_OFFSET = 0.33
|
||||
ENERGY_SLOPE = 2.97
|
||||
THRESHOLD = 0.5
|
||||
|
||||
# CsI(Tl) non-linear response correction
|
||||
CSI_NONLINEAR_ALPHA = 0.37
|
||||
CSI_NONLINEAR_BETA = 100.0
|
||||
|
||||
|
||||
def correct_csilinear_energy(spectrum_rate, num_channels=1023):
|
||||
"""Apply inverse CsI(Tl) non-linear response correction.
|
||||
|
||||
Remaps channels so peaks appear at theoretical energy positions
|
||||
(matching training data), correcting for the detector's non-proportional
|
||||
scintillation response that shifts low-energy peaks upward.
|
||||
"""
|
||||
alpha = CSI_NONLINEAR_ALPHA
|
||||
beta = CSI_NONLINEAR_BETA
|
||||
|
||||
output_channels = np.arange(num_channels, dtype=np.float64)
|
||||
e_true = ENERGY_OFFSET + ENERGY_SLOPE * output_channels
|
||||
|
||||
# Forward model: E_apparent = E_true * (1 + alpha * exp(-E_true / beta))
|
||||
e_apparent = e_true * (1 + alpha * np.exp(-e_true / beta))
|
||||
|
||||
# Input channel where detector placed counts for this true energy
|
||||
source_channels = (e_apparent - ENERGY_OFFSET) / ENERGY_SLOPE
|
||||
source_channels = np.clip(source_channels, 0, num_channels - 1.001)
|
||||
|
||||
lower = np.floor(source_channels).astype(int)
|
||||
upper = np.minimum(lower + 1, num_channels - 1)
|
||||
frac = source_channels - lower
|
||||
|
||||
corrected = spectrum_rate[lower] * (1 - frac) + spectrum_rate[upper] * frac
|
||||
return corrected
|
||||
|
||||
|
||||
def load_model(model_path):
|
||||
"""Load a VegaModel checkpoint."""
|
||||
device = torch.device("cpu")
|
||||
checkpoint = torch.load(model_path, map_location=device, weights_only=False)
|
||||
config = VegaConfig(**checkpoint["model_config"])
|
||||
model = VegaModel(config)
|
||||
model.load_state_dict(checkpoint["model_state_dict"])
|
||||
model.eval()
|
||||
return model, config
|
||||
|
||||
|
||||
def run_inference(model, config, isotope_index, spectrum_rate, threshold=THRESHOLD,
|
||||
apply_correction=True):
|
||||
"""Run inference on a spectrum rate array (1023 channels)."""
|
||||
if spectrum_rate.max() <= 0:
|
||||
return []
|
||||
|
||||
# Apply CsI(Tl) non-linear correction so peaks match training data positions
|
||||
if apply_correction:
|
||||
spectrum_rate = correct_csilinear_energy(spectrum_rate)
|
||||
|
||||
log_spectrum = np.log1p(np.maximum(spectrum_rate, 0))
|
||||
max_val = log_spectrum.max()
|
||||
normalized = log_spectrum / max_val if max_val > 0 else log_spectrum
|
||||
tensor = torch.tensor(normalized, dtype=torch.float32).unsqueeze(0)
|
||||
|
||||
with torch.no_grad():
|
||||
logits, activities = model(tensor)
|
||||
|
||||
probs = torch.sigmoid(logits).numpy()[0]
|
||||
activities = activities.numpy()[0] * config.max_activity_bq
|
||||
|
||||
results = []
|
||||
for i in range(len(probs)):
|
||||
if probs[i] >= threshold:
|
||||
results.append({
|
||||
"isotope": isotope_index.index_to_name(i),
|
||||
"probability": float(probs[i]),
|
||||
"activity_bq": float(activities[i]),
|
||||
})
|
||||
return sorted(results, key=lambda x: -x["probability"])
|
||||
|
||||
|
||||
def main():
|
||||
# Load isotope index
|
||||
isotope_index = IsotopeIndex.load(MODELS_DIR / "vega_isotope_index.txt")
|
||||
print(f"Isotope index: {isotope_index.num_isotopes} isotopes\n")
|
||||
|
||||
# Load monitor state (real spectrum from detector)
|
||||
with open(DATA_DIR / "monitor_state.json") as f:
|
||||
state = json.load(f)
|
||||
|
||||
counts = np.array(state["counts"], dtype=np.float64)
|
||||
live_time = state["cumulated_live_time_s"]
|
||||
print(f"Spectre reel : {live_time:.0f}s live time, {counts.sum():.0f} coups, {len(counts)} canaux")
|
||||
print(f"CPS : {state['cps']:.2f}")
|
||||
|
||||
# Load background
|
||||
bg_data = np.load(DATA_DIR / "background_24h.npy", allow_pickle=True).item()
|
||||
bg_counts = bg_data["counts"].astype(np.float64)
|
||||
bg_live_time = float(bg_data["duration"])
|
||||
print(f"Background : {bg_live_time/3600:.1f}h, {bg_counts.sum():.0f} coups\n")
|
||||
|
||||
# Prepare spectra
|
||||
rate = counts[:1023] / live_time
|
||||
bg_rate = bg_counts[:1023] / bg_live_time
|
||||
net_rate = np.clip(rate - bg_rate, 0, None)
|
||||
|
||||
# Apply CsI correction to show peak positions
|
||||
corrected_rate = correct_csilinear_energy(rate)
|
||||
corrected_net = correct_csilinear_energy(net_rate)
|
||||
|
||||
print("=" * 70)
|
||||
print(f" Sans correction CsI:")
|
||||
print(f" Canal max (brut) : {rate.argmax():>4d} ({ENERGY_OFFSET + ENERGY_SLOPE * rate.argmax():.1f} keV)")
|
||||
print(f" Canal max (net) : {net_rate.argmax():>4d} ({ENERGY_OFFSET + ENERGY_SLOPE * net_rate.argmax():.1f} keV)")
|
||||
print(f" Avec correction CsI:")
|
||||
print(f" Canal max (brut) : {corrected_rate.argmax():>4d} ({ENERGY_OFFSET + ENERGY_SLOPE * corrected_rate.argmax():.1f} keV)")
|
||||
print(f" Canal max (net) : {corrected_net.argmax():>4d} ({ENERGY_OFFSET + ENERGY_SLOPE * corrected_net.argmax():.1f} keV)")
|
||||
print(f" Rate max (brut) : {rate.max():.2f} cps")
|
||||
print(f" Rate max (net) : {net_rate.max():.2f} cps")
|
||||
print("=" * 70)
|
||||
|
||||
# Am-241 should be at 59.5 keV → ch ~20
|
||||
print(f"\n Am-241 region (59.5 keV) apres correction CsI:")
|
||||
for ch in range(16, 26):
|
||||
e = ENERGY_OFFSET + ENERGY_SLOPE * ch
|
||||
print(f" ch {ch:3d} ({e:5.1f} keV): brut={corrected_rate[ch]:.5f} net={corrected_net[ch]:.5f}")
|
||||
|
||||
# Load both models
|
||||
models = {
|
||||
"vega_best": load_model(MODELS_DIR / "vega_best.pt"),
|
||||
"vega_final": load_model(MODELS_DIR / "vega_final.pt"),
|
||||
}
|
||||
|
||||
scenarios = {
|
||||
"brut (sans soustraction)": rate,
|
||||
"net (avec soustraction bg)": net_rate,
|
||||
}
|
||||
|
||||
for model_name, (model, config) in models.items():
|
||||
print(f"\n{'─' * 70}")
|
||||
print(f" Modele : {model_name}")
|
||||
print(f"{'─' * 70}")
|
||||
for scenario_name, spectrum in scenarios.items():
|
||||
print(f"\n Scenario : {scenario_name}")
|
||||
results = run_inference(model, config, isotope_index, spectrum)
|
||||
if results:
|
||||
print(f" {'Isotope':>10s} {'Probabilite':>12s} {'Activite (Bq)':>15s}")
|
||||
print(f" {'─'*10} {'─'*12} {'─'*15}")
|
||||
for r in results:
|
||||
print(f" {r['isotope']:>10s} {r['probability']*100:>11.1f}% {r['activity_bq']:>15.1f}")
|
||||
else:
|
||||
print(f" Aucun isotope detecte (seuil = {THRESHOLD})")
|
||||
|
||||
# Also show top-10 probabilities below threshold for context
|
||||
print(f"\n{'═' * 70}")
|
||||
print(" Top-10 probabilites (tous scenarios, meme sous le seuil)")
|
||||
print(f"{'═' * 70}")
|
||||
for model_name, (model, config) in models.items():
|
||||
print(f"\n Modele : {model_name}")
|
||||
for scenario_name, spectrum in scenarios.items():
|
||||
if spectrum.max() <= 0:
|
||||
continue
|
||||
# Apply CsI correction before inference
|
||||
corrected = correct_csilinear_energy(spectrum)
|
||||
log_spectrum = np.log1p(np.maximum(corrected, 0))
|
||||
max_val = log_spectrum.max()
|
||||
normalized = log_spectrum / max_val if max_val > 0 else log_spectrum
|
||||
tensor = torch.tensor(normalized, dtype=torch.float32).unsqueeze(0)
|
||||
with torch.no_grad():
|
||||
logits, _ = model(tensor)
|
||||
probs = torch.sigmoid(logits).numpy()[0]
|
||||
top10 = np.argsort(probs)[::-1][:10]
|
||||
print(f"\n {scenario_name} :")
|
||||
for idx in top10:
|
||||
name = isotope_index.index_to_name(idx)
|
||||
prob = probs[idx]
|
||||
marker = " *" if prob >= THRESHOLD else ""
|
||||
print(f" {name:>10s} : {prob*100:6.2f}%{marker}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user