Centrifugal Pump Cavitation Detection Using Multi-Sensor Vibration AI

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing cavitation identification methods for centrifugal pumps are inaccurate and labor-intensive, relying on coarse-grained labels and manual feature extraction, which hinders precise characterization of cavitation states and real-time monitoring.

Innovation Solution

A cavitation state identification method using an adaptive one-dimensional convolutional neural network that processes raw vibration signals from multiple sensors, combined with high-speed imaging to quantify cavitation intensity, enabling real-time and end-to-end diagnosis.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual feature extraction and classification algorithms are used for cavitation diagnosis, then the method is interpretable and requires less computational resources, but the identification accuracy is low and cannot achieve real-time monitoring

Engineering Contradiction:

Improvecavitation identification accuracyVSAvoidmodel complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces manual feature extraction and traditional classification algorithms with a deep learning-based automatic feature learning system. The convolutional neural network automatically learns fault features from raw vibration signals through multi-level nonlinear mapping, eliminating the need for manual feature engineering while achieving higher identification accuracy and real-time monitoring capability.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The deep learning model performs self-learning of cavitation fault features from the vibration data without requiring manual intervention for feature extraction. The system automatically adapts to different cavitation states through training on labeled data, enabling it to self-optimize its diagnostic capability.

Inventive Principle:

Principle #25Self-service

2Measurement precision

If coarse-grained cavitation labels are used for training, then the labeling process is simple and fast, but the identification results are rough and cannot accurately characterize cavitation development degree

Engineering Contradiction:

Improvecavitation state identification precisionVSAvoidlabeling efficiency

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent performs preliminary action by using high-speed imaging to capture cavitation bubbles and visually assess cavitation development degree before creating training labels. This preliminary visualization step enables the creation of fine-grained labels that accurately represent different cavitation stages, which then train the model to achieve precise identification results.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If two-dimensional convolutional neural network is used with manually extracted features converted to images, then the model can leverage image processing capabilities, but the spatial correlation of vibration signals is destroyed and fault information is lost

Engineering Contradiction:

Improvefeature learning capabilityVSAvoidspatial correlation information

Core Design Contradiction:

Adaptability or versatilityVSLoss of information

Solution Approach 1:

Instead of converting one-dimensional vibration signals to two-dimensional images (the conventional approach), the patent inverts the approach by using a one-dimensional convolutional neural network that processes the raw time-series vibration signals directly. This preserves the temporal and spatial correlation inherent in the vibration data while still leveraging convolutional operations for effective feature extraction.

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS12492935B2Cavitation state identification method driven by vibration data of fluid machinery

Publication Date: 2025.12.09 ZHEJIANG UNIV

AI summary

A cavitation state identification method driven by vibration data of fluid machinery is disclosed in the present invention, and belongs to the field of big data learning models. According to the present invention, an adaptive neural network is trained by means of a cavitation data set to form a cavitation state identification model, such that vibration signal sequences can be collected online by all vibration sensors arranged at different positions of a target centrifugal pump, the collected vibration signal sequences are input into the cavitation state identification model obtained by training, and a current real-time cavitation intensity of the target centrifugal pump is predicted online. Moreover, the cavitation intensity predicted in the present invention can use a more detailed quantitative label, such that fine-grained prediction about a cavitation development degree is achieved.