Junseok2/MachiningFM2.0

MachiningFM 2.0 — Physics-Guided Foundation Model for Machining

A pretrained foundation model for CNC machining sensor data, extended with a classical physics calibration layer for downstream tasks.

Architecture

Raw Machining Signal (force, vibration, AE, NC program, ...)
        │
        ▼
  MachiningFMV2 (CausalFusionTransformer, d_model=384)
        │
        ▼
  Latent Embedding
        │
        ▼
  Downstream Head (Ridge / MLP)
        │
        ▼
  Raw Prediction
        │
        ├── Taylor Tool-Life  (r_T = t / T_Taylor)
        ├── Kienzle Force     (F_measured / F_Kienzle)
        └── Cutting Energy    (E_c = ∫ F_c · V_c · dt)
        │
        ▼
  Physics Calibration (y_final = y_FM + α · g(physics))
        │
        ▼
   Final Prediction

Model Config (pretrained/machiningfm_v2_base.pt)

Parameter	Value
d_model	384
fusion_layers	6
num_heads	8
dropout	0.1
forecast_horizons	[64, 1280, 12800]
output_channels	3
trained steps	5,432
training loss	-1.94

Supported Input Modalities

Modality	Description
raw_waveform	High-rate sensor signals (force, vibration, AE)
spectral	FFT / STFT / CWT spectral features
cnc	CNC SEFC (Servo Error / Feed / Current) data
nc_tokens	NC program token sequences
image	Tool wear images
metadata_vector	Machining condition scalars

---

Files

MachiningFM2.0/
├── pretrained/
│   └── machiningfm_v2_base.pt       # 170MB — primary pretrained checkpoint
├── configs/
│   ├── model/base.yaml              # Model architecture config
│   ├── physics/default.yaml         # Generic steel/carbide parameters
│   └── physics/ti6al4v_carbide.yaml # Ti-6Al-4V parameters
└── README.md

Note: The large v1 pretraining checkpoint (7.4GB) is not included due to file size constraints. The v2 base checkpoint above was initialized from it and fine-tuned with the v2 architecture.

---

Usage

git clone https://github.com/junseokShim/MachiningFM.git
cd MachiningFM
pip install -e .

Download this checkpoint:

from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="Junseok2/MachiningFM2.0",
    filename="pretrained/machiningfm_v2_base.pt",
)

Load and encode:

import torch
from machiningfm.models.backbone import MachiningFMBackbone

backbone = MachiningFMBackbone(
    checkpoint_path=ckpt_path,
    backbone_mode="frozen",  # frozen | linear_probe | partial_finetune | full_finetune
)

batch = {"raw_waveform": torch.randn(1, 4096, 3)}
encoded = backbone.encode(batch)
embedding = encoded["embedding"]  # shape: (1, 384)

Tool wear regression with physics calibration:

from machiningfm.tasks.tool_wear import ToolWearRegressor
from machiningfm.physics.calibration import PhysicsCalibrator, PhysicsFeatures
from machiningfm.physics.taylor import TaylorParams, compute_tool_life_ratio

# Extract embeddings from backbone (offline)
# X_train, X_val, X_test: (N, 384) numpy arrays
# y_train, y_val, y_test: (N,) VB wear in mm

# Build physics features
params = TaylorParams(C=200.0, n=0.25)
pf = [
    PhysicsFeatures(tool_life_ratio=compute_tool_life_ratio(t, 250.0, 0.25, 0.125, params))
    for t in elapsed_times
]

# Fit with physics calibration
cal = PhysicsCalibrator(method="ridge")
reg = ToolWearRegressor(feature_dim=384, calibrator=cal)
reg.fit(X_train, y_train, pf_train, X_val, y_val, pf_val)

preds = reg.predict(X_test, pf_test)

---

Downstream Tasks

Task	Input	Output	Evaluation
A. Wear Regression	embedding	VB (mm)	MAE, RMSE, R²
B. Stage Classification	embedding	healthy/moderate/severe	Acc, Macro F1
C. RUL Prediction	embedding	remaining time (min)	MAE, RMSE, R²
D. Dimensional Compensation	wear + condition	offset (mm)	physics-derived interface

Wear Stage Thresholds (ISO 8688-1:1989)

• Healthy: VB < 0.1 mm
• Moderate: 0.1 ≤ VB < 0.2 mm
• Severe: VB ≥ 0.2 mm

---

Physics Models

Model	Status	Required Data
Taylor Tool-Life	Enabled	speed, feed, depth
Kienzle Force	Enabled	chip thickness, width
Cutting Energy	Enabled	force series, speed
Archard Wear	Disabled	F_N, L (not in standard datasets)
Usui Wear Rate	Disabled	cutting temperature (not in standard datasets)

Physics parameters are stored in YAML configs (see configs/physics/). All parameter sources are documented (literature / dataset_calibrated / manufacturer / user_defined).

---

Dataset

PHM Society Data Challenge 2010

• URL: https://www.phmsociety.org/competition/phm/10
• Sensor data: force (x/y/z), vibration (x/y/z), acoustic emission RMS
• Labels: flank wear VB (mm) per cut, per condition (real measurements)
• Citation: PHM Society (2010). PHM Data Challenge. Prognostics and Health Management Society.

Data Integrity Notes:

• PHM2010 does NOT contain dimensional accuracy measurements. Task D (dimensional compensation) outputs are physics-derived estimates, not experimental results.
• Archard/Usui models are disabled by default because PHM2010 lacks the required quantities.
• All splits use leave-one-condition-out to prevent temporal leakage.

---

Limitations

• Pretrained on proprietary CNC machining data — domain gap to new materials/machines is expected.
• PHM2010 benchmarks with synthetic data demonstrate code correctness, not production performance.
• Physics calibration provides marginal improvement on clean synthetic data; real benefit expected on noisy real-world data.
• Archard and Usui models require data not typically available in standard machining datasets.
• The 7.4GB v1 pretraining checkpoint is not included due to file size constraints.

---

Citation

If you use this model or framework, please cite:

@software{machiningfm2026,
  author = {Shim, Junseok},
  title  = {MachiningFM: Physics-Guided Foundation Model for Machining},
  year   = {2026},
  url    = {https://github.com/junseokShim/MachiningFM}
}

---

Source Code

• GitHub: https://github.com/junseokShim/MachiningFM
• License: MIT