Context: the project archive (KXKM_Batterie_Parallelator-main) had
no git history locally; a fresh repository is needed to host it on
git.saillant.cc (electron/KXKM_Batterie_Parallelator).

Approach: initialize a new repo on branch main, stage the archive
content, and harden .gitignore before the first commit.

Changes:
- Import the full project tree: firmware/, firmware-idf/,
  firmware-rs/, iosApp/, kxkm-bmu-app/, kxkm-api/, hardware/,
  docs/, specs/, scripts/, models/, tests/
- Keep project dotfiles tracked despite the trailing '.*' ignore
  rule: .github/, .claude/, .superpowers/, .gitattributes,
  .markdownlint.json
- Extend .gitignore: firmware/src/credentials.h (local secrets,
  template kept), kxkm-bmu-app/**/build/ (66 MB compiled iOS
  framework), .remember/ (session data)

Impact: the project can now be maintained on the self-hosted Gitea
forge with a clean, secret-free initial history.

2026-07-04 12:32:28 +02:00

2.5 KiB

Raw Permalink Blame History

scripts/ml

ML pipeline scripts for BMU advisory intelligence (SOH/RUL).

Safety rule:

Outputs are advisory only.
Firmware protections remain authoritative.

Main workflow

Feature extraction

python3 scripts/ml/extract_features.py --input models/consolidated.parquet --output models/features_v2.parquet --window 60

Feature adaptation

python3 scripts/ml/adapt_features.py --input models/features_v2.parquet --output models/features_adapted_v2.parquet

SOH model training

python3 scripts/ml/train_fpnn.py --input models/features_adapted_v2.parquet --output-dir models --epochs 50 --hidden 64 --degree 2 --soh-mode capacity --test-device k-led1 --val-ratio 0.1

Quantization

python3 scripts/ml/quantize_tflite.py --model models/fpnn_soh.pt --features models/features_adapted_v2.parquet --output models/fpnn_soh_v2_quantized.onnx --backend onnxrt

Final metrics

python3 scripts/ml/finalize_phase2_metrics.py --features models/features_adapted_v2.parquet --quantized models/fpnn_soh_v2_quantized.onnx --rul-model models/rul_sambamixer.pt --train-log phase2_fpnn_train_v2.log --output models/phase2_metrics.json

Remote execution

Use orchestrator:

scripts/ml/remote_kxkm_ai_pipeline.sh check
scripts/ml/remote_kxkm_ai_pipeline.sh bootstrap-container
scripts/ml/remote_kxkm_ai_pipeline.sh discover-dataset
scripts/ml/remote_kxkm_ai_pipeline.sh run-container

Artifacts

models/features_v2.parquet
models/features_adapted_v2.parquet
models/fpnn_soh_v2_quantized.onnx
models/phase2_metrics.json

Quantization iteration

Advanced quantization iteration:

python3 scripts/ml/quantize_tflite.py \
  --model models/fpnn_soh.pt \
  --features models/features_adapted_v2.parquet \
  --output models/fpnn_soh_v2_quantized.onnx \
  --backend onnxrt \
  --quant-format qdq \
  --calib-strategy stratified \
  --calib-samples 500 \
  --percentile-clip 1 99

Bootstrap verification now checks pandas, numpy, torch, onnxruntime, pyarrow, and onnx.

Latest verified remote status on kxkm-ai:

Technical unblock: done.
Remaining blocker: quantized quality gate on the latest remote dataset split.

Promoted baseline on kxkm-ai

Winning configuration promoted on 2026-03-30 16:17 UTC:

--quant-format qdq
--calib-strategy stratified
--calib-samples 500
--percentile-clip 1 99

Promoted remote metrics:

float32 MAPE: 7.7289%
quantized MAPE: 10.7734%
degradation: +3.0446 pp
quantized size: 15.99 KB
overall_gate_pass=true

2.5 KiB Raw Permalink Blame History