Physical Operating-Condition-Guided Vision-Language Matching for Cavitation Diagnosis in Industrial Pipeline Systems

Yu Sha; Junqi Tao; Deyu Wang; Horst Stoecker; Domagoj Vnucec; Andreas Widlf; Kai Zhou

cs.AIOct 20, 2026

Physical Operating-Condition-Guided Vision-Language Matching for Cavitation Diagnosis in Industrial Pipeline Systems

Authors: Yu Sha, Junqi Tao, Deyu Wang, Horst Stoecker, Domagoj Vnucec, Andreas Widlf, Kai Zhou

Organizations: School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, 518172, China · Frankfurt Institute for Advanced Studies, Frankfurt am Main, 60438, Germany · Institut für Theoretische Physik, Goethe Universität Frankfurt, Frankfurt am Main, 60314, Germany · GSI Helmholtzzentrum für Schwerionenforschung GmbH, Darmstadt, 64291, Germany ·

Abstract

Cavitation diagnosis in industrial pipeline systems is challenging because acoustic responses vary substantially with operating conditions, while similar time-frequency patterns may correspond to different cavitation states. To address this issue, we propose a physical operating-condition-guided vision-language matching (POC-VLM) for cavitation diagnosis. POC-VLM first constructs physical operating-condition-aware semantic descriptions by explicitly incorporating sample-specific operating variables into candidate cavitation texts. A time-frequency-conditioned text adaptation module then dynamically aligns these physical semantics with the observed acoustic response, followed by global-local vision-language matching to capture both holistic and localized discriminative patterns. In parallel, a K-nearest neighbours-based non-parametric physical operating-condition prior estimates the physical plausibility of each candidate state and is fused with the vision-language semantic evidence at the decision level. Experiments on three real-world cavitation datasets show that POC-VLM consistently outperforms representative data-driven, multimodal and domain-specific diagnostic methods. Further analyses verify the complementary roles of vision-language semantic matching and physical operating-condition priors, while demonstrating robust performance under varying sampling rates, noise levels and observation durations.

Figures & tables

Figure 1: Overall framework of the proposed POC-VLM for cavitation diagnosis. It combines vision-language semantic matching with a physical operating-condition prior through decision-level evidence fusion.

Figure 2: Schematic view of the test rack at SAMSON AG.

Physical parameters	Types	Ranges	Tolerances
Flow rate	Krohne M950	0 $-$ $180\text{\,}{\mathrm{m}}^{3}\text{\,}{\mathrm{h}}^{-1}$	$\pm$ $0.9\text{\,}{\mathrm{m}}^{3}\text{\,}{\mathrm{h}}^{-1}$
Valve stroke	Sylvac S229	0 $-$ $50\text{\,}\mathrm{mm}$	$\pm$ $0.001\text{\,}\mathrm{mm}$
Upstream pressure	SAMSON 6054	0 $-$ $16\text{\,}\mathrm{bar}$	$\pm$ $0.09\text{\,}\mathrm{bar}$
Downstream pressure	SAMSON 6054	0 $-$ $16\text{\,}\mathrm{bar}$	$\pm$ $0.09\text{\,}\mathrm{bar}$
Medium temperature	SAMSON 5204	-20 $-$ $150\text{\,}\mathrm{\SIUnitSymbolCelsius}$	$\pm$ $0.45\text{\,}\mathrm{\SIUnitSymbolCelsius}$
Valve structure noise	Vallen VS45-H	20 $-$ $450\text{\,}\mathrm{kHz}$	$\pm$ $63\text{\,}\mathrm{dB}$

Table 1: Detailed information of various measurement sensors.

Dataset	Operation parameters
Dataset	Valve stroke ( $\mathrm{mm}$ )	Upstream pressure ( $\mathrm{bar}$ )	Temperature ( $\mathrm{\SIUnitSymbolCelsius}$ )
Cavitation-Short	[15, 13.5, 11.25, 7.5, 3.75, 1.5, 0.75]	[10, 9, 6, 4]	25–50
Cavitation-Long	[60, 55, 45, 30, 25, 15, 6]	[10, 6, 4]	23–52
Cavitation-Noise	15	10	32–39

Table 2: Details of three real-world cavitation datasets for valve operation with various upstream pressures.

Dataset	Training set				Testing set
	Cavitation			Non	Cavitation			Non
	Choked flow	Constant	Incipient	Non	Choked flow	Constant	Incipient	Non
Cavitation-Short	58( $\times$ 66)	75( $\times$ 66)	32( $\times$ 66)	121( $\times$ 66)	14( $\times$ 66)	18( $\times$ 66)	8( $\times$ 66)	30( $\times$ 66)
Cavitation-Long	117( $\times$ 555)	315( $\times$ 555)	55( $\times$ 555)	158( $\times$ 555)	31( $\times$ 555)	81( $\times$ 555)	9( $\times$ 555)	40( $\times$ 555)
Cavitation-Noise	32( $\times$ 555)	32( $\times$ 555)	32( $\times$ 555)	32( $\times$ 555)	8( $\times$ 555)	8( $\times$ 555)	8( $\times$ 555)	8( $\times$ 555)

Table 3: Details of the training and test sets.

(\cdot)

denotes the number after the sliding window (a window duration of

0.06\text{\,}\mathrm{s}

, an overlap ratio of 0.25 and 512 CWT scales).

Types	Methods	Cavitation-Short				Cavitation-Long				Cavitation-Noise
Types	Methods	Acc.	Pre.	Rec.	F1	Acc.	Pre.	Rec.	F1	Acc.	Pre.	Rec.	F1
Data	ResNet18	89.39	89.40	89.39	88.51	93.35	93.33	93.35	93.33	98.05	98.05	98.05	98.05
	ResNet34	89.55	89.71	89.55	88.79	93.38	93.30	93.38	93.32	98.06	98.06	98.06	98.06
	ResNet50	89.24	89.24	89.24	88.48	93.44	93.35	93.44	93.36	97.84	97.84	97.84	97.84
	DenseNet121	90.11	90.22	90.11	89.45	93.30	93.17	93.30	93.19	98.58	98.58	98.58	98.58
	DenseNet161	89.94	89.99	89.94	89.23	93.40	93.23	93.40	93.27	97.76	97.76	97.76	97.76

Table 4: Diagnostic performance comparison of different methods across real-world industrial cavitation datasets.

Types	Methods	Cavitation-Short				Cavitation-Long				Cavitation-Noise
Types	Methods	Acc.	Pre.	Rec.	F1	Acc.	Pre.	Rec.	F1	Acc.	Pre.	Rec.	F1
Data	LiftingNet	85.29	82.75	75.89	73.95	88.43	86.31	87.97	87.05	95.86	94.62	95.84	95.20
	MIPLCNet	86.57	83.38	77.71	75.00	89.14	87.06	88.18	87.55	96.57	95.78	96.61	96.18
	ResNet-APReLU	86.86	84.04	77.89	75.56	90.71	88.83	90.91	89.70	97.29	96.77	97.49	97.12
	LSTM-RDRN	87.71	85.42	78.50	76.72	91.14	89.36	91.55	90.32	98.71	98.40	98.53	98.46
Hybrid	HKG-ResNet34	89.71	90.54	78.02	76.74	92.44	88.55	92.21	90.16	99.06	99.06	99.06	99.06

Table 5: Diagnostic performance comparison with domain-specific methods across real-world cavitation datasets.

Figure 3: Normalized confusion matrices of PerceiverIO and POC-VLM across the three real-world cavitation datasets. (a-c) PerceiverIO on Cavitation-Short, Cavitation-Long and Cavitation-Noise, respectively. (d-f) POC-VLM on the corresponding datasets.

Model Variant	Accuracy	Precision	Recall	F1-score
w/o Vision-Language Pathway	0.9104 $\pm$ 0.0018	0.8985 $\pm$ 0.0048	0.8732 $\pm$ 0.0043	0.8800 $\pm$ 0.0050
w/o Physical Operating-Condition Prior	0.8859 $\pm$ 0.0049	0.8864 $\pm$ 0.0044	0.8243 $\pm$ 0.0049	0.8384 $\pm$ 0.0051
POC-VLM	0.9714 $\pm$ 0.0040	0.9694 $\pm$ 0.0058	0.9549 $\pm$ 0.0017	0.9608 $\pm$ 0.0038

Table 6: Ablation results of different model variants of POC-VLM on Cavitation-Short. The results are averaged over three independent runs.

Physical Prior	Textual Description	Accuracy	Precision	Recall	F1-score
✗	Static class	0.8316 $\pm$ 0.0031	0.8359 $\pm$ 0.0036	0.7248 $\pm$ 0.0032	0.7342 $\pm$ 0.0025
✗	Dynamic POC-aware	0.8859 $\pm$ 0.0049	0.8864 $\pm$ 0.0044	0.8243 $\pm$ 0.0049	0.8384 $\pm$ 0.0051
✓	Static class	0.9422 $\pm$ 0.0028	0.9232 $\pm$ 0.0027	0.9320 $\pm$ 0.0024	0.9272 $\pm$ 0.0035
✓	Dynamic POC-aware	0.9714 $\pm$ 0.0040	0.9694 $\pm$ 0.0058	0.9549 $\pm$ 0.0017	0.9608 $\pm$ 0.0038

Table 7: Ablation analysis of textual semantic representation and the physical operating-condition prior on Cavitation-Short. The results are averaged over three independent runs.

Figure 4: Ablation and sensitivity analysis of POC-VLM on Cavitation-Short. (a) Window duration, (b) window overlap ratio, (c) number of CWT scales, (d) CLIP visual patch size, (e) number of cross-attention heads and (f) number of P2Tformer layers. The results are averaged over three independent runs.

Figure 5: Analysis of physical operating conditions for POC-VLM on Cavitation-Short. (a) Combination analysis of the 255 non-empty combinations of eight physical operating-condition variables, (b) nearest-neighbour analysis and (c) physical-prior fusion weight analysis. The results are averaged over three independent runs.

Number	Physical Operating Conditions								Evaluation Metrics
Number	${X_{f}}$	$K_{v}$	$h$	$p_{u}$	$p_{d}$	$Q$	$p_{v}$	$T$	Accuracy	Precision	Recall	F1-score
1	✓								0.7297	0.6341	0.6312	0.6262
2	✓	✓							0.9143	0.9248	0.8701	0.8856
3	✓		✓	✓					0.9714	0.9694	0.9549	0.9608
4			✓	✓	✓	✓			0.9714	0.9694	0.9549	0.9608
5		✓	✓	✓	✓	✓			0.9714	0.9694	0.9549	0.9608

Table 8: Best-performing physical operating-condition variable combinations for different numbers of variables on Cavitation-Short.

Figure 6: Generalization and robustness analysis of POC-VLM. (a) Sampling-rate analysis and (b) noise-level analysis on Cavitation-Short, and (c) cross-duration analysis between Cavitation-Short and Cavitation-Long. The results are averaged over three independent runs.

Fault Diagnosis

Fault diagnosis identifies, localizes, and explains failures in engineered and computational systems using sensor data, logs, models, and domain knowledge.

See all 146 papers

Physical-Variable-Informed Deep Learning for Robust Diagnosis in Industrial Fluid Transport Systems
Oct 20, 2026Yu Sha, Junqi Tao, Dixin Zhou +4Fault Diagnosis Industrial Fault Diagnosis
School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, 518172, China. · Frankfurt Institute for Advanced Studies, Frankfurt am Main, 60438, Germany. · Key Laboratory of Quark & Lepton Physics and Institute of Particle Physics, Central China Normal University, Wuhan, 430079, China. +4
MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection
Jul 21, 2026Shuimu Chen, Jing Jin, Nan Su +5Industrial Anomaly Detection VLM Distillation
Tsinghua University · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · Shenzhen University
When Visual Signals Mislead: A Mechanistic Study of Attribute Hallucination in Vision-Language Models
Aug 11, 2026Yufei Zhang, Chenlu Zhan, Hongwei WangHallucination Detection in VLMs VLM Adaptation
Zhejiang University

Oct 20, 2026J/K move · Enter open · S save

Physical-Variable-Informed Deep Learning for Robust Diagnosis in Industrial Fluid Transport Systems

Yu Sha, Junqi Tao, Dixin Zhou, Horst Stoecker+3

School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, 518172, China. · Frankfurt Institute for Advanced Studies, Frankfurt am Main, 60438, Germany. · Key Laboratory of Quark & Lepton Physics and Institute of Particle Physics, Central China Normal University, Wuhan, 430079, China.+4

Fault Diagnosis Industrial Fault Diagnosis Physics-Informed ML

Open paper AskPDF arXiv

Industrial fluid transport systems are monitored through heterogeneous sensing streams, but robust fault diagnosis remains difficult because response signals vary with operating conditions and differ from physical variables in sampling rate, observation duration and data granularity. To address these challenges, we propose a physical-variable-informed deep learning framework (PhyNet) for robust fault diagnosis under real-world sampling heterogeneity. PhyNet treats physical variables as global operating constraints and response signals as local diagnostic evidence. Its physical-to-response cross-attention mechanism connects sequence-level physical-variable tokens with a learnable global response token, enabling interaction without strict temporal alignment. Across four benchmarks, including three real-world cavitation datasets and one public two-phase-flow dataset, PhyNet achieves the best performance across all evaluated metrics. Class-wise analyses, together with attention and masking analyses, show improved discrimination of ambiguous states and indicate that physical variables meaningfully constrain local response interpretation. These findings demonstrate a deployment-oriented approach to robust diagnosis in heterogeneous industrial environments.

Abstract

Fault Diagnosis

Explore similar work

Physical-Variable-Informed Deep Learning for Robust Diagnosis in Industrial Fluid Transport Systems

MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection

When Visual Signals Mislead: A Mechanistic Study of Attribute Hallucination in Vision-Language Models