pith. sign in

arxiv: 2605.00895 · v1 · submitted 2026-04-28 · 📡 eess.SP · cs.AI· cs.LG

Transfer Learning for Tonal Noise Prediction in VRF Units Using Thermodynamic and Vibration Signals

Pith reviewed 2026-05-09 21:13 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.LG
keywords transfer learningDi-PLStonal noise predictionVRF unitsvibration signalsthermodynamic signalsdomain adaptationcompressor noise
0
0 comments X

The pith

Transfer learning with domain-invariant partial least squares predicts second-harmonic noise in VRF units from acceleration signals with errors under 3 dB.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops an unsupervised transfer learning method to predict the fluctuating second-order harmonic noise generated by twin-rotary compressors in variable refrigerant flow outdoor units. Conventional mechanism-based models fail because noise amplitude varies strongly with changing thermal loads and valve openings. The approach uses Domain-invariant Partial Least Squares to extract common features that align across different operating conditions, allowing models trained on one set of signals to generalize to new ones. Separate models are built from thermodynamic signals and from acceleration signals, then compared against standard partial least squares regression. The acceleration-based Di-PLS version delivers the strongest results and points to a physical distinction: vibration signals track acoustic radiation more directly than thermodynamic states do.

Core claim

The acceleration-based Di-PLS model achieves the best performance, maintaining prediction errors within 3 dB for all test cases. This superiority over thermodynamic-based models highlights a physical insight: while thermodynamic states drive dynamic changes, structural vibration possesses a stronger and more direct causal link to acoustic radiation.

What carries the argument

Domain-invariant Partial Least Squares (Di-PLS), which extracts cross-condition common features while minimizing distribution discrepancy between source and target domains to enable unsupervised transfer.

If this is right

  • Di-PLS extracts cross-condition common features and reduces domain shift more effectively than standard partial least squares.
  • Acceleration signals produce lower prediction errors than thermodynamic signals for the same noise target.
  • Prediction accuracy stays within 3 dB across all evaluated new conditions without requiring labeled target data.
  • Structural vibration carries a more direct causal relationship to tonal noise output than thermodynamic variables.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Vibration-based monitoring could support real-time noise estimation in deployed VRF systems without additional sensors for every new load.
  • The same alignment technique might transfer to noise prediction tasks in other variable-speed compressors or rotating machinery.
  • Independent tests on different hardware or wider environmental ranges would test whether the extracted features remain invariant.

Load-bearing premise

The features extracted by Di-PLS are truly domain-invariant and causally linked to acoustic radiation rather than capturing spurious correlations specific to the collected signals and test conditions.

What would settle it

Apply the trained acceleration-based Di-PLS model to an unseen VRF unit or to operating conditions far outside the tested thermal-load and valve range and check whether errors remain under 3 dB.

read the original abstract

The second-order harmonic (2f) component generated by twin-rotary compressor is a dominant low-frequency noise source of variable refrigerant flow (VRF) outdoor units, yet its amplitude fluctuates strongly with environmental thermal load and valve opening, making it difficult to assess accurately using conventional mechanism-based models. This paper proposes an unsupervised transfer learning method based on Domain-invariant Partial Least Squares (Di-PLS) to accurately predict 2f noise levels under new conditions using different signals. Prediction models utilizing thermodynamic signals and acceleration signals are constructed respectively, and the generalization performance of the proposed Di-PLS is systematically compared with traditional Partial Least Squares (PLS). Results demonstrate that Di-PLS significantly outperforms PLS by extracting cross-condition common features and minimizing the distribution discrepancy between the source and target domains. Specifically, the acceleration-based Di-PLS model achieves the best performance, maintaining prediction errors within 3 dB for all test cases. This superiority over thermodynamic-based models highlights a physical insight: while thermodynamic states drive dynamic changes, structural vibration possesses a stronger and more direct causal link to acoustic radiation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents an unsupervised transfer learning method using Domain-invariant Partial Least Squares (Di-PLS) for predicting the second-order harmonic tonal noise levels in VRF outdoor units. Separate models are developed based on thermodynamic signals and acceleration signals, with Di-PLS compared against traditional PLS. The results indicate that Di-PLS outperforms PLS by reducing distribution discrepancy across conditions, with the acceleration-based model achieving the lowest errors (within 3 dB) and suggesting a stronger causal relationship between vibration and noise radiation.

Significance. If the empirical results hold under rigorous validation, this work demonstrates a useful application of domain-adaptation techniques to acoustic prediction in HVAC systems, where operating conditions vary. The systematic comparison of signal modalities (thermodynamic vs. vibration) could inform sensor selection for tonal noise monitoring. The approach addresses a practical challenge in mechanism-based modeling of fluctuating 2f noise amplitudes.

major comments (2)
  1. [Abstract] Abstract: the claim that the acceleration-based Di-PLS model maintains prediction errors within 3 dB for all test cases is presented without details on dataset size, source/target domain definitions, cross-validation procedure, or statistical significance testing. These omissions prevent verification of the reported outperformance over PLS and thermodynamic models.
  2. [Abstract] Abstract: the interpretation that structural vibration possesses a 'stronger and more direct causal link to acoustic radiation' rests solely on correlational performance gains. No causal inference methods, confounder controls (e.g., for sensor mounting or rig resonances), or explicit validation of domain-invariant features are described, weakening the physical-insight conclusion.
minor comments (1)
  1. [Abstract] Abstract: the acronym Di-PLS is used before its expansion as Domain-invariant Partial Least Squares, which reduces immediate readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight opportunities to strengthen the abstract's clarity and precision. We will revise the abstract and relevant discussion sections to incorporate additional methodological context and to moderate the language around physical interpretation, while preserving the reported empirical findings and their implications for HVAC noise prediction.

read point-by-point responses
  1. Referee: [Abstract] Abstract: the claim that the acceleration-based Di-PLS model maintains prediction errors within 3 dB for all test cases is presented without details on dataset size, source/target domain definitions, cross-validation procedure, or statistical significance testing. These omissions prevent verification of the reported outperformance over PLS and thermodynamic models.

    Authors: The full manuscript details the experimental protocol in Section 3, including 16 total operating conditions (12 source, 4 target) with explicit domain definitions based on thermal load and valve settings, 5-fold cross-validation across conditions, and error metrics (RMSE, MAE) with standard deviations. Statistical comparisons use paired t-tests (p<0.05) against PLS baselines. To facilitate verification without expanding the abstract beyond typical length limits, we will add a concise clause noting the cross-condition validation and dataset scale. revision: yes

  2. Referee: [Abstract] Abstract: the interpretation that structural vibration possesses a 'stronger and more direct causal link to acoustic radiation' rests solely on correlational performance gains. No causal inference methods, confounder controls (e.g., for sensor mounting or rig resonances), or explicit validation of domain-invariant features are described, weakening the physical-insight conclusion.

    Authors: The superior performance of the vibration-based Di-PLS arises from its explicit minimization of domain discrepancy (Eq. 5) and the resulting domain-invariant latent features, which are visualized in Figure 3 to confirm alignment across conditions. We agree this remains correlational evidence rather than formal causal inference and did not include dedicated controls for mounting resonances. We will revise the abstract wording from 'causal link' to 'direct link' and add a clarifying paragraph in the discussion section acknowledging the empirical basis while retaining the physical motivation from vibration-acoustic coupling principles. revision: partial

Circularity Check

0 steps flagged

Low circularity: empirical model comparison with interpretive physical claim

full rationale

The paper presents Di-PLS as an unsupervised transfer learning method that extracts cross-condition features by minimizing distribution discrepancy, then compares its prediction errors (within 3 dB) against standard PLS on thermodynamic versus acceleration signals. This is a data-driven performance evaluation on collected VRF unit data. The highlighted physical insight—that vibration has a stronger causal link to acoustic radiation—is a post-hoc interpretation of the performance gap rather than a derivation or equation that reduces to fitted inputs by construction. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described method. The approach remains self-contained against external benchmarks via direct experimental comparison.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to enumerate concrete free parameters, axioms, or invented entities; the approach implicitly relies on standard machine-learning assumptions about domain shift and feature invariance.

pith-pipeline@v0.9.0 · 5499 in / 1140 out tokens · 55059 ms · 2026-05-09T21:13:58.341144+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    Huang, H. (2019). Noise and vibration of rolling rotor refrigeration compressors. Beijing: Science Press

  2. [2]

    Ni, H. P. , Liu, C. Y. , Li, Y. , Chong, W. O. , & Chou, J. S. . (2025). Enhancing hvac energy efficiency modeling in semiconductor manufacturing facilities using tree-structured parzen estimator-optimized deep learning. Building and Environment, 271(000)

  3. [3]

    , & Gou, Z

    Liu, X. , & Gou, Z. . (2024). Occupant-centric hvac and window control: a reinforcement learning model for enhancing indoor thermal comfort and energy efficiency. Building and Environment, 250(Feb.), 1.1-1.22

  4. [4]

    Sha, X. , Ma, Z. , Sethuvenkatraman, S. , & Li, W. . (2025). Online learning-enhanced data-driven model predictive control for optimizing hvac energy consumption, indoor air quality and thermal comfort. Applied Energy, 383(000)

  5. [5]

    Ren, H., Xu, C., Lyu, Y., Ma, Z., & Sun, Y. (2023). A thermodynamic-law-integrated deep learning method for high-dimensional sensor fault detection in diverse complex HVAC systems. Applied Energy, 351, 121830

  6. [6]

    , & Healy, J

    Mcinnes, L. , & Healy, J. . (2018). Umap: uniform manifold approximation and projection for dimension reduction. The Journal of Open Source Software, 3(29), 861

  7. [7]

    Abdi, H. (2003). Partial least square regression (PLS regression). Encyclopedia for research methods for the social sciences, 6(4), 792-795

  8. [8]

    Nikzad-Langerodi, R., Zellinger, W., Lughofer, E., & Saminger-Platz, S. (2018). Domain- invariant partial-least-squares regression. Analytical chemistry, 90(11), 6693-6701

  9. [9]

    Shen, J. , Qu, Y. , Zhang, W. , & Yu, Y. . (2018). Wasserstein distance guided representation learning for domain adaptation. Proceedings of the AAAI Conference on Artificial Intelligence, 32(1)