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

Yu Sha; Junqi Tao; Dixin Zhou; Horst Stoecker; Domagoj Vnucec; Andreas Widl; Kai Zhou

cs.LGOct 20, 2026

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

Authors: Yu Sha, Junqi Tao, Dixin Zhou, Horst Stoecker, Domagoj Vnucec, Andreas Widl, 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. · Key Laboratory of Quark & Lepton Physics and Institute of Particle Physics, Central China Normal University, Wuhan, 430079, China. · School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, 518172, China. ·

Abstract

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.

Figures & tables

Figure 1: Overview of PhyNet architecture . a , Schematic of an industrial fluid transport system monitoring setup. High-frequency response measurements and routinely recorded physical variables are acquired by the corresponding sensors. These are synchronously processed through a signal conditioning module and a high-precision data acquisition system to form paired response signal-physical variable observations for diagnosis. b , Architecture of PhyNet. Physical Variables Embedding summarizes each variable sequence as a sequence-wise variate token. Response Embedding embeds non-overlapping patches as response tokens and generates a learnable global response token. Physical-to-Response Cross-Attention injects physical variate tokens into the global response token. Response Self-Attention propagates the updated global context to all response tokens. The attention blocks are stacked for $L$ layers.

Types	Methods	Cavitation-Short				Cavitation-Long				Cavitation-Noise				VTPFR
Types	Methods	Acc.	Pre.	Rec.	F1	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	87.30	87.24	87.30	87.22
	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	88.89	88.89	88.89	88.63
	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	84.66	84.56	84.66	84.36
	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	88.36	88.89	88.36	87.93
	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	89.95	89.93	89.95	89.82

Table 1: Comparison of model performance across four datasets.

Figure 2: Class-wise error patterns and ROC analysis on representative datasets . a, b , Confusion matrices of PerceiverIO and PhyNet on Cavitation-Long, respectively. c, d , Receiver operating characteristic (ROC) curves of PerceiverIO and PhyNet on Cavitation-Long, respectively. e, f , Confusion matrices of PerceiverIO and PhyNet on VTPFR, respectively. g, h , ROC curves of PerceiverIO and PhyNet on VTPFR, respectively. The confusion matrices visualize class-wise prediction patterns and misclassification at the chosen decision threshold, whereas the ROC curves illustrate the relationship between the true positive rate and the false positive rate across thresholds. The area under each ROC curve corresponds to the AUC value, which provides a threshold-independent measure of discriminative performance.

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 2: Comparison with domain-specific fault intensity diagnosis methods.

Figure 3: Physical-response cross-attention and its association with diagnostic performance . a , b , Layer-wise mean cross-attention weights from the global response token to physical-variable tokens for representative high-confidence prototype samples (a) and low-margin boundary samples (b) from the four flow states. High-confidence prototype samples denote correctly classified examples with high prediction confidence, whereas low-margin boundary samples denote correctly classified examples with small margins between the top two predicted classes. c , Class-averaged physical-variable attention across the four flow states. d , Accuracy under random masking of physical-variable combinations. Random masking of different numbers of physical-variable combinations is equivalent to sampling physical-variable subsets of different sizes. e , Physical-variable importance across the whole test set of Cavitation-Long. f , Accuracy under single-variable masking. The accuracy drop is defined relative to the 96.59 $\%$ accuracy of the PhyNet (full model) on Cavitation-Long. The reported values are averaged over three independent model-training runs ( $n=3$ ), and the error bars denote one standard deviation.

Figure 4: Local response evidence under physical variables . a-d , Representative spectrograms from choked-flow cavitation (a), constant cavitation (b), incipient cavitation (c) and non-cavitation (d). For each row, the left panel shows the input spectrogram, while the middle and right panels show the gradient-based localization maps produced by the complete PhyNet and PhyNet without physical-to-response cross-attention, respectively. All localization maps are displayed using the same colour scale. Technical details of the localization procedure are provided in the Supplementary Methods.

Variants	Acc.	Pre.	Rec.	F1
w/o physical variables	79.01	78.48	79.01	78.47
w/o global response token	85.71	86.15	85.71	85.29
w/o cross-attention	87.88	88.21	87.88	88.03
w/o self-attention propagation	87.95	87.35	87.95	87.60
naive fusion	83.76	84.21	83.76	82.85
PhyNet	96.59	96.57	96.59	96.55

Table 3: Asymmetric fusion ablation on Cavitation-Long.

Figure 5: Robustness analysis of PhyNet under various conditions . a-c , Sensitivity of PhyNet to time-frequency preprocessing on Cavitation-Short, including window duration (a), window overlap ratio (b) and number of continuous wavelet transform (CWT) scales (c). d-f , Sensitivity of PhyNet to architectural hyperparameters on Cavitation-Short, including patch size (d), number of attention heads (e) and number of encoder layers (f). g , Performance of PhyNet across different sampling frequencies on Cavitation-Short. h , Accuracy of PhyNet under different signal-to-noise ratios on Cavitation-Short. i , Performance comparison between PhyNet and the strongest hybrid baseline (PerceiverIO) during cross-duration transfer between Cavitation-Short and Cavitation-Long (Training set

\to

Test set: Short

\to

Long and Long

\to

Short). Panels a-h show the performance of PhyNet under isolated perturbations of a single factor, while the remaining settings are fixed at their default values. The reported values are averaged over three independent model-training runs (

n=3

), and the error bars denote one standard deviation. In addition, the results for the remaining metrics are provided in the Supplementary Information.

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

DSPR: Dual-Stream Physics-Residual Networks for Trustworthy Industrial Time Series Forecasting
Apr 8, 2026Yeran Zhang, Pengwei Yang, Guoqing Wang +1Time Series Forecasting Multivariate Time Series Forecasting
School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, China · Research Center, East Hope Group Co., Ltd, Shanghai, China
YOTOnet: Zero-Shot Cross-Domain Fault Diagnosis via Domain-Conditioned Mixture of Experts
May 6, 2026Zesen Wang, Zihao Wu, Yue Hu +2Cross-Domain Transfer Learning Bearing Fault Diagnosis
School of Mechanical and Power Engineering, East China University of Science and Technology, Shanghai, China
S2S-FDD: Bridging Industrial Time Series and Natural Language for Explainable Zero-shot Fault Diagnosis
Mar 9, 2026Baoxue Li, Chunhui ZhaoFault Diagnosis Industrial Fault Diagnosis
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China · School of Information and Electrical Engineering, Zhejiang University City College, Hangzhou, China

Apr 8, 2026J/K move · Enter open · S save

DSPR: Dual-Stream Physics-Residual Networks for Trustworthy Industrial Time Series Forecasting

Yeran Zhang, Pengwei Yang, Guoqing Wang, Tianyu Li

School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, China · Research Center, East Hope Group Co., Ltd, Shanghai, China

Time Series Forecasting Multivariate Time Series Forecasting Non-Stationary Time Series Forecasting Physics-Informed ML Explainable Time Series Forecasting Dynamic Graph Learning

Open paper AskPDF arXiv

Accurate forecasting of industrial time series requires balancing predictive accuracy with physical plausibility under non-stationary operating conditions. Existing data-driven models often achieve strong statistical performance but struggle to respect regime-dependent interaction structures and transport delays inherent in real-world systems. To address this challenge, we propose DSPR (Dual-Stream Physics-Residual Networks), a forecasting framework that explicitly decouples stable temporal patterns from regime-dependent residual dynamics. The first stream models the statistical temporal evolution of individual variables. The second stream focuses on residual dynamics through two key mechanisms: an Adaptive Window module that estimates flow-dependent transport delays, and a Physics-Guided Dynamic Graph that incorporates physical priors to learn time-varying interaction structures while suppressing spurious correlations. Experiments on four industrial benchmarks spanning heterogeneous regimes demonstrate that DSPR consistently improves forecasting accuracy and robustness under regime shifts while maintaining strong physical plausibility. It achieves state-of-the-art predictive performance, with Mean Conservation Accuracy exceeding 99% and Total Variation Ratio reaching up to 97.2%. Beyond forecasting, the learned interaction structures and adaptive lags provide interpretable insights that are consistent with known domain mechanisms, such as flow-dependent transport delays and wind-to-power scaling behaviors. These results suggest that architectural decoupling with physics-consistent inductive biases offers an effective path toward trustworthy industrial time-series forecasting. Furthermore, DSPR's demonstrated robust performance in long-term industrial deployment bridges the gap between advanced forecasting models and trustworthy autonomous control systems.

Abstract

Fault Diagnosis

Explore similar work

DSPR: Dual-Stream Physics-Residual Networks for Trustworthy Industrial Time Series Forecasting

YOTOnet: Zero-Shot Cross-Domain Fault Diagnosis via Domain-Conditioned Mixture of Experts

S2S-FDD: Bridging Industrial Time Series and Natural Language for Explainable Zero-shot Fault Diagnosis