feat: Integrate XJTU-SY dataset and validate cross-domain domain adaptation

[TasfinMahmud]

Overview

This Pull Request integrates the XJTU-SY bearing dataset into the CNSD framework and provides an end-to-end validation of the physics-guided domain adaptation methodology on high-resolution data (25.6 kHz).

Context & Motivation

Deep learning models frequently suffer from severe performance degradation when exposed to unseen operating domains. This PR evaluates the performance of a baseline 1D-CNN across a significant domain shift (training on Condition 1: 2100 RPM / 12kN, and testing on Condition 2: 2250 RPM / 11kN). It introduces the CNSD Physics Engine to automatically filter and reject unreliable predictions using first-principles bearing kinematics.

Data Preprocessing Methodology

To ensure rigorous evaluation of the continuous run-to-failure sequences and avoid label noise from intermediate degradation states:

1. Dynamic State Isolation: The parser dynamically selects the first 20% of available files for the 'Healthy' baseline and the last 20% for the mature 'Fault' state.
2. Standardized Window Slicing: The 32,768-point raw signals are sliced into standardized 4096-point sliding windows (step size: 1024 points).
3. Data Volume: This procedure yielded a robust sample size:
   • Source Domain (2100 RPM): 5,046 samples
   • Target Domain (2250 RPM): 5,945 samples (split evenly between Calibration and Testing)

Detailed Validation Results

The validation script (validate_xjtusy.py) automatically swept the target domain calibration set to find the optimal prominence ratio ($\tau=3.0$). Below is the detailed output from the final evaluation on the unseen Test Set (2250 RPM).

Loading XJTU-SY dataset (Cross-Domain RPM/Load Split)...
Train (2100 RPM)=5046 | Calib (2250 RPM)=5945 | Test (2250 RPM)=5945

[1] Training Neural Network on 2100 RPM Source Data...
Fitting causal models... (Complete)

[2] Calibrating Tau threshold on 2250 RPM Target Data...
=> Selected optimal tau: 3.0

[3] Evaluating on Test Set (2250 RPM)...

--- FINAL TEST RESULTS (CROSS-DOMAIN XJTU-SY) ---
Baseline CNN Acc:       0.539
--------------------------------------------
  Physics-Confirmed Acc:   0.913 (n=3251)
  Physics-Conflict Acc:    0.142 (n=1434)
  Physics-Inconclusive Acc:0.023 (n=1260, 21.2%)
  GAP (CONF - CNFL):       +0.771
--------------------------------------------

Conclusion

The baseline CNN suffered a significant collapse to 53.9% accuracy when testing across the domain shift. However, when the neural network's predictions aligned with the physical kinematic equations (Confirmed), the accuracy rose to 91.3%. When the predictions conflicted with the physics, the CNN was accurately identified as failing (14.2% accuracy).

The resulting performance gap of +77.1% confirms that the physics engine successfully mitigates domain shift vulnerabilities.

[abhiprd2000]

@TasfinMahmud Three things before merging and it becomes a paper claim, same as we did for PU: (1) run the ensemble + MC-dropout + softmax baselines at matched coverage because the whole claim is that physics beats the baselines. (2) 20 seeds + paired t-test for significance. (3) Fix the framing — physics doesn't mitigate domain shift or do domain adaptation, the CNN is still 54%. It flags which predictions to trust. call it reliability/verification.

[TasfinMahmud]

Hi @abhiprd2000,

Thank you for the detailed feedback! All three points have been addressed:

1 & 2: Baselines & Statistical Significance: The ensemble + MC-dropout + softmax baselines have all been run at strictly matched coverage across 20 independent seeds, complete with paired t-tests. The results are phenomenal (e.g., Physics catching 100% of the Ensemble's highly-confident errors on XJTU-SY at 0dB noise, with p=0.0028). The full reproducible evaluation script is in this PR, and I posted the detailed markdown tables in the PR #17 thread.

3: Framing (Reliability/Verification): This is an excellent academic distinction. You are completely right the physics engine doesn't technically mitigate the domain shift of the raw signal or adapt the underlying CNN (which remains at ~54% raw accuracy cross-domain); rather, it acts as a mathematically rigorous verifier that flags which predictions to trust. I have force-pushed a new commit to amend the Git history and PR title to explicitly use the reliability/verification framing rather than adaptation.

Let me know if we're good to merge!

[abhiprd2000]

@TasfinMahmud , Please paste the full final report for XJTU-SY here.