SMTNet: Hierarchical cavitation intensity recognition based on sub-main transfer network

Yu Sha, Johannes Faber, Shuiping Gou, Бо Лю, Wei Li, Stefan Schramm, Horst Stoecker, T. Steckenreiter, Domagoj Vnucec, Nadine Wetzstein, Andreas Widl, Kai Zhou

TL;DR

This paper introduces SMTNet, a novel hierarchical deep learning framework for recognizing cavitation intensity in valve acoustics. By structuring the recognition process from coarse to fine, SMTNet achieves state-of-the-art accuracy, outperforming existing methods, especially in challenging scenarios with limited data and noisy environments.

What Problem Does This Paper Address?

Accurate machinery health monitoring is crucial for smart manufacturing. Cavitation, a phenomenon where bubbles form and collapse in fluids, can cause significant damage to valves, pumps, and pipes, leading to reduced efficiency and potential failures. While deep learning has shown promise in cavitation detection, current methods struggle with hierarchical classification problems where some cavitation states are more difficult to distinguish than others, or when classes naturally form a hierarchy. Existing approaches often treat all classes equally, failing to leverage the inherent structure of cavitation intensity levels. This paper addresses this limitation by proposing a framework that mimics how human experts diagnose cavitation, moving from a broad classification to progressively finer distinctions.

Methods

The core of this work is the Sub-Main Transfer Network (SMTNet), a two-step hierarchical classification model.

  • Hierarchical Structure: SMTNet employs a "sub-main" architecture. The "sub" network first classifies signals into two broad categories: cavitation or non-cavitation. The "main" network then takes the identified cavitation signals and further classifies them into specific intensity levels (e.g., "choked flow," "constant cavitation," "incipient cavitation"). This hierarchical approach allows for more nuanced recognition.
  • 1-D Double Hierarchical Residual Block (1-D DHRB): To effectively extract features from acoustic signals, the authors introduce a novel residual block. This block uses large convolution kernels and a hierarchical structure with two different-sized convolutional layers to capture sensitive frequency-domain features.
  • Sliding Window with Fast Fourier Transform (Swin-FFT): To combat the "few-shot" problem (limited training data), a data augmentation technique is proposed. Swin-FFT splits time-domain acoustic signals into segments using a sliding window and then transforms these segments into the frequency domain using FFT. This helps the model learn robust features even with scarce data.
  • Hierarchical Multi-Label Tree: This structure is used to embed the semantic relationships between different cavitation states into the deep learning model, guiding the hierarchical classification process.
  • Experience Filtering: This mechanism leverages the output of the "sub" network (cavitation detection) to refine the input for the "main" network, improving the accuracy of the intensity recognition by filtering out non-cavitation signals.

Datasets

The proposed SMTNet was evaluated on three distinct datasets provided by SAMSON AG:

  • Dataset 1: Two datasets without noise.
  • Dataset 2: Two datasets without noise.
  • Dataset 3: One dataset with real-world noise, making it a more challenging benchmark.

These datasets contain acoustic signals from valve cavitation under various operating conditions.

Key Insights and Ideas

  • Mimicking Human Expertise: The hierarchical approach of SMTNet is inspired by how human engineers diagnose complex issues, moving from general to specific. This suggests that incorporating domain knowledge about the problem's structure can significantly improve AI model performance.
  • The Power of Frequency Domain Features: The 1-D DHRB's success in capturing sensitive frequency-domain features highlights the importance of analyzing signals beyond their raw time-series representation for tasks like cavitation detection.
  • Data Augmentation for Scarcity: Swin-FFT proves effective in mitigating the impact of limited training data, a common challenge in industrial applications. This technique allows for better generalization.
  • Hierarchical Structure for Nuance: The two-step classification (cavitation vs. non-cavitation, then intensity levels) is more effective than a single-stage multi-class classifier, especially for distinguishing subtle differences between cavitation states.

Key Results

SMTNet demonstrated impressive performance across all three datasets:

  • Cavitation Detection Accuracy: Achieved as high as 97.02% (Dataset 1), 97.64% (Dataset 2), and 100% (Dataset 3).
  • Cavitation Intensity Recognition Accuracy: Reached 95.32% (Dataset 1), 97.16% (Dataset 2), and 100% (Dataset 3).
  • Robustness to Noise: Dataset 3, with real-world noise, still yielded perfect 100% accuracy for both detection and intensity recognition, showcasing the robustness of SMTNet and its data augmentation.
  • Performance Across Frequencies: SMTNet maintained high accuracy even when using downsampled acoustic signals (e.g., mobile phone frequencies), indicating its potential for deployment on less sophisticated hardware.
  • Superiority over Baselines: SMTNet consistently outperformed other methods, including SVM, Decision Trees, 1D CNN, and XGBoost, particularly in the challenging "incipient cavitation" state.

Limitations and Future Work

The authors acknowledge two primary limitations:

  • Incipient Cavitation Accuracy: The proposed method still shows lower accuracy for the "incipient cavitation" state compared to other states.
  • Two-Step Framework: The current approach is a two-step hierarchical detection framework.

Future work will focus on:

  • Modifying the method to improve the accuracy of incipient cavitation recognition.
  • Exploring a one-step hierarchical detection framework to potentially streamline the process.

Summary updated

Abstract

With the rapid development of smart manufacturing, data-driven machinery health management has been of growing attention. In situations where some classes are more difficult to be distinguished compared to others and where classes might be organised in a hierarchy of categories, current DL methods can not work well. In this study, a novel hierarchical cavitation intensity recognition framework using Sub-Main Transfer Network, termed SMTNet, is proposed to classify acoustic signals of valve cavitation. SMTNet model outputs multiple predictions ordered from coarse to fine along a network corresponding to a hierarchy of target cavitation states. Firstly, a data augmentation method based on Sliding Window with Fast Fourier Transform (Swin-FFT) is developed to solve few-shot problem. Secondly, a 1-D double hierarchical residual block (1-D DHRB) is presented to capture sensitive features of the frequency domain valve acoustic signals. Thirdly, hierarchical multi-label tree is proposed to assist the embedding of the semantic structure of target cavitation states into SMTNet. Fourthly, experience filtering mechanism is proposed to fully learn a prior knowledge of cavitation detection model. Finally, SMTNet has been evaluated on two cavitation datasets without noise (Dataset 1 and Dataset 2), and one cavitation dataset with real noise (Dataset 3) provided by SAMSON AG (Frankfurt). The prediction accurcies of SMTNet for cavitation intensity recognition are as high as 95.32%, 97.16% and 100%, respectively. At the same time, the testing accuracies of SMTNet for cavitation detection are as high as 97.02%, 97.64% and 100%. In addition, SMTNet has also been tested for different frequencies of samples and has achieved excellent results of the highest frequency of samples of mobile phones.
BibTeX
@misc{sha2022smtnet,
  author = {Yu Sha and Johannes Faber and Shuiping Gou and Бо Лю and Wei Li and Stefan Schramm and Horst Stoecker and T. Steckenreiter and Domagoj Vnucec and Nadine Wetzstein and Andreas Widl and Kai Zhou},
  title = {{SMTNet: Hierarchical cavitation intensity recognition based on sub-main transfer network}},
  year = {2022},
  eprint = {2203.01429},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.01429},
}
Figure