Hydraulic turbine cavitation acoustic signal identification method based on big data machine learning

Patent metadata

Field	Value
Publication number	US-11840998-B2
Application number	US-202217859244-A
Country	US
Kind code	B2
Filing date	Jul 7, 2022
Priority date	Jul 8, 2021
Publication date	Dec 12, 2023
Grant date	Dec 12, 2023

Abstract

The present invention provides a hydraulic turbine cavitation acoustic signal identification method based on big data machine learning. According to the method, time sequence clustering based on multiple operating conditions under the multi-output condition of the hydraulic turbine set is performed by utilizing an neural network, characteristic quantities of the hydraulic turbine set under a steady condition in a healthy state is screened; a random forest algorithm is introduced to perform feature screening of multiple measuring points under steady-state operation of the hydraulic turbine set, optimal feature measuring points and optimal feature subsets are extracted, finally a health state prediction model is constructed by using gated recurrent units; whether incipient cavitation is present in the equipment is judged. The present invention can effectively identify the occurrence of incipient cavitation in the hydraulic turbine set, reducing unnecessary shutdown of the equipment and prolonging the service life.

First claim

The invention claimed is:

1. A hydraulic turbine cavitation acoustic signal identification method based on big data learning, comprising the following steps:
S 1 , obtaining latest acoustic signal time sequence data of each measuring point in real time through measuring points arranged on a hydraulic turbine set, and partitioning the acoustic signal time sequence data of each measuring point into multiple normalized acoustic signal subsequences, wherein a latest recorded acoustic signal subsequence of each measuring point is used as a real-time signal subsequence;
S 2 , inputting the multiple normalized acoustic signal subsequences of all measuring points obtained in S 1 into a self-organizing maps (SOM) neural network, clustering the multiple normalized acoustic signal subsequences into multiple clusters according to a corresponding operating condition of the hydraulic turbine set, and then dividing each of the multiple clusters into a steady-state cluster and an unsteady-state cluster according to a signal fluctuation degree of the multiple normalized acoustic signal subsequences in each of the multiple clusters;
S 3 , traversing distribution of the real-time signal subsequences of all measuring points in the multiple clusters; if a number of the real-time signal subsequences contained in the steady-state cluster is not lower than a minimum number threshold, it is judged that the hydraulic turbine is in a steady condition and an incipient cavitation warning process proceeds according to S 4 -S 8 ; otherwise, the current incipient cavitation warning process is interrupted;
S 4 , performing feature screening on the real-time signal subsequences contained in the steady-state cluster by a random forest (RF) algorithm, and extracting optimal feature measuring points which can sensitively reflect changes in the corresponding operating condition of the hydraulic turbine set and optimal feature subsets of each of the optimal feature measuring points;
S 5 , normalizing the optimal feature subsets of each of the optimal feature measuring points and calculating information entropy, and with the information entropy as an input, predicting a future trend of the hydraulic turbine set in a healthy state by using a health state prediction model constructed based on multilayer gate recurrent units (GRUs) to obtain predictive information entropy of the acoustic signal of each of the optimal feature measuring points in a next predictive step;
S 6 , obtaining acoustic signal time sequence data actually acquired from each of the optimal feature measuring points on the hydraulic turbine set in the next predictive step and calculating actual information entropy, and calculating a dynamic tolerance of each of the optimal feature measuring points from the predictive information entropy and the actual information entropy;
S 7 , based on a current output condition of the hydraulic turbine set, obtaining acoustic signal information entropy with incipient cavitation present) of the hydraulic turbine set in the next predictive step through prediction using a pre-constructed SOM network, and calculating a dynamic tolerance alarm threshold of each of the optimal feature measuring points from the predictive information entropy and the acoustic signal information entropy with the incipient cavitation present; and
S 8 , comparing a sum of the dynamic tolerances of all the optimal feature measuring points with a sum of the dynamic tolerance alarm thresholds based on a threshold method, and judging whether the incipient cavitation occurs to the hydraulic turbine set; if yes, an incipient cavitation warning is given; otherwise, no incipient cavitation warning is given.

2. The hydraulic turbine cavitation acoustic signal identification method of claim 1 , wherein in S 1 , the method of partitioning the acoustic signal time sequence data of each measuring point into multiple acoustic signal subsequences comprises the following steps:
S 11 , performing fixed-step sliding through a fixed-sized time window on the latest acoustic signal time sequence data of each measuring point, and extracting acoustic signal subsequences from the time window every time one step is slided by; and
S 12 , normalizing each of the acoustic signal subsequences extracted in S 11 to obtain the finally outputted multiple normalized acoustic signal subsequences.

3. The hydraulic turbine cavitation acoustic signal identification method of claim 1 , wherein the implementation method of S 2 comprises the following steps:
S 21 , inputting the multiple normalized acoustic signal subsequences of all measuring points obtained in S 1 as an input layer of the SOM neural network, so that the inputted multiple normalized acoustic signal subsequences are divided into different clusters through unsupervised learning clustering; and
S 22 , for each of the clusters clustered in S 21 , calculating multiple statistical values of data points in each of the multiple normalized acoustic signal subsequences, and then calculating a deviation of each of the statistical values of the different normalized acoustic signal subsequences in the same cluster; if the deviation of each of the statistical values corresponding to one cluster is less than a respective deviation threshold, such cluster is marked as the steady-state cluster; otherwise, such cluster is marked the unsteady-state cluster.

4. The hydraulic turbine cavitation acoustic signal identification method of claim 3 , wherein the multiple statistical values comprise the mean value, maximum value, minimum value and median of the data points in the multiple normalized acoustic signal subsequences, and the deviation is a variance.

5. The hydraulic turbine cavitation acoustic signal identification method of claim 1 , wherein the implementation method of S 4 comprises the following steps:
S 41 , performing a first disturbance on each real-time signal subsequence contained in the steady-state cluster based on a RF algorithm, and calculating a feature importance index Ψ k of each corresponding measuring point according to the results before and after the disturbance, wherein a calculation formula is as follows:

Ψ
k

=

1
B

⁢

∑

b
=
1

B

(

R
b
oob

-

R
bk
oob

)

where B represents a total number of the real-time signal subsequences contained in the steady-state cluster, R b oob represents a number of correctly classified out-of-bag (OOB) data of a decision-making tree before the first disturbance is performed on the bth real-time signal subsequence, and R bk oob represents a number of correctly classified OOB data of a decision-making tree after the first disturbance is performed on the bth real-time signal subsequence;
S 42 , based on the feature importance index of each measuring point obtained in S 41 , screening the optimal feature measuring points which can sensitively reflect the changes in the cor

Assignees

• Univ Zhejiang

Inventors

• Tong Zheming
• Xin Jiage
• Tong Shuiguang

Classifications

• F03B11/008 Primary

  Measuring or testing arrangements (in general G01) · CPC title

• F05B2260/80

  Diagnostics · CPC title

• F05B2270/305

  Tolerances · CPC title

• F05B2270/333

  Noise or sound levels · CPC title

• F05B2270/404

  active, predictive, or anticipative · CPC title