An intelligent recognition method for cavitation phenomenon of water turbine based on combined spectrum feature extraction

Abstract

The application discloses a kind of intelligent recognition methods of water turbine cavitation phenomenon based on combination spectrum feature extraction, belong to signal recognition processing field.It includes collecting the noise data of water turbine runner model before and after cavitation occurs;Decompose the sound spectrum of each group of noise data and the special pressure pulsation sound spectrum, extract the index parameter representing the characteristics of the two sound spectrum and normalize, then map to the different element positions of row vector and column vector of matrix, form the characteristic vector containing bubble sound main characteristic ID, constitute the instant profile form matrix A representing the main physical sound state of bubble sound;The matrix C obtained by modifying the matrix A is input into the trained water turbine cavitation recognition model, and the cavitation discrimination result can be output.The method extracts the mixed sound spectrum of cannon sound spectrum and special pulsation spectrum in the collected water turbine primary cavitation data, inputs the mixed sound spectrum into the training model as the characteristic vector of cavitation recognition, and can realize the intelligent recognition of water turbine primary cavitation phenomenon by machine.

Description

Technical Field

[0001] This invention relates to a method for identifying cavitation phenomena in hydraulic turbines, specifically an intelligent method for identifying cavitation phenomena in hydraulic turbines based on combined spectrum feature extraction. Background Technology

[0002] Primary cavitation in a hydro turbine refers to the phenomenon where the local pressure within the liquid decreases to a critical value, causing a rapid increase in the number of gas nuclei and the onset of cavitation. It is accompanied by noise and vibration, leading to decreased power generation efficiency, reduced output, and intensified hydraulic vibration. This not only affects the service life of the hydro turbine but also threatens the safe operation of the hydropower station and the power grid.

[0003] Therefore, identifying turbine cavitation is crucial for the operational safety of hydropower stations and power grids. Currently, in practice, turbine cavitation is identified manually. This involves observing the vortex zone and flow patterns at the turbine runner's outlet using a transparent acrylic glass tailrace section. During the experiment, a stroboscopic instrument (an adjustable light source with alternating bright and dark illumination) is used. The stroboscopic instrument's frequency is adjusted to match or closely approximate the turbine's rotational speed, allowing the seemingly stationary or slowly rotating turbine blades to be clearly seen with the naked eye, thus revealing the cavitation status at the turbine runner's outlet. This method demands a high level of expertise from personnel, typically requiring at least ten years of experience to determine the presence of cavitation. This method is highly subjective, resulting in low accuracy and efficiency.

[0004] Existing technologies include methods for identifying cavitation acoustic signals from hydro turbines using big data learning. For example, patent CN113255848A discloses a method for identifying cavitation acoustic signals from hydro turbines based on big data learning. The technical solution involves: obtaining multiple neural network models based on big data learning; extracting time-series data of the turbine's acoustic signals; using a SOM neural network to perform time-series clustering based on various operating conditions under multiple output conditions of the turbine; screening feature quantities of the turbine under stable operating conditions in a healthy state; then introducing a random forest algorithm to screen features from multiple measurement points under stable operating conditions of the turbine; extracting the optimal feature measurement points and optimal feature subsets with high sensitivity to the prediction model; finally, using a gated cyclic unit to establish a health state prediction model; and using adaptive evaluation of the sum of dynamic tolerances of multiple measurement points to determine whether the equipment exhibits initial cavitation phenomena and to issue early warnings.

[0005] To study the characteristics of turbine cavitation noise, it is necessary to understand the features of turbine cavitation phenomena and accurately and comprehensively extract the information contained in the sample data. This requires combining the results of experimental data analysis with the principles of turbine cavitation itself, thus effectively distinguishing different stages of turbine cavitation and completing the diagnosis and identification of nascent turbine cavitation. The aforementioned patent's method of predicting future short-term stable operating conditions by screening characteristic quantities of the turbine unit under healthy conditions is lacking in both accuracy and efficiency in identifying nascent turbine cavitation phenomena.

[0006] When cavitation occurs in a water turbine, tiny bubbles are generated on the outer edge of the blades, producing polymorphic noise. Due to its atypical physical sound characteristics, current experimental techniques based on physical sound configurations cannot accurately identify whether cavitation has occurred in the model runner based on its own sound characteristics. Summary of the Invention

[0007] This invention aims to solve the aforementioned problems in existing turbine cavitation identification technologies and proposes an intelligent identification method for turbine cavitation phenomena based on combined spectrum feature extraction. This method can achieve intelligent identification of initial cavitation phenomena in turbines relatively quickly and accurately.

[0008] To achieve the above-mentioned objectives, the technical solution of the present invention is as follows:

[0009] A method for intelligent identification of cavitation phenomena in hydraulic turbines based on combined spectral feature extraction, characterized by the following steps:

[0010] S1. Collect noise data when cavitation occurs in the turbine runner model;

[0011] S2. Decompose the mixing signal in the noise data into several single-wave signals, and extract the first and second main frequency signals from them.

[0012] S3. Calculate the statistical characteristics of the first main frequency signal and the second main frequency signal respectively, obtain the relevant analysis index parameters and perform normalization processing;

[0013] S4. Map the normalized index parameters to different element positions of the row and column vectors of the matrix to form a PCSV vector containing the main feature ID of cavitation bubbles.

[0014] S5. Use the PCSV vector as a feature vector and input it into the neural network to train the turbine cavitation sound recognition model.

[0015] S6. Process the real-time collected turbine cavitation sound data and input it into the turbine cavitation sound recognition model, and output the cavitation recognition result.

[0016] Furthermore, step S5 includes:

[0017] S51. Assign different weights to the index parameters of different main frequency signals in the PCSV vector and perform weighted averaging. Then, arrange and combine the resulting new vectors to find the contour shape matrix B that best represents the main physical sound state of the noise data segment.

[0018] S52. Multiply each index parameter in the contour morphology matrix B by a correction value to obtain a corrected contour morphology matrix D containing several new vectors. This corrected contour morphology matrix D constitutes the turbine cavitation sound recognition model.

[0019] Furthermore, step S6 specifically includes:

[0020] S61. Collect real-time noise data of the turbine runner;

[0021] S62. Decompose the mixing signal in the instantaneous noise data into several single-wave signals, calculate the statistical characteristics of these single-wave signals respectively, obtain the relevant analysis index parameters and perform normalization processing.

[0022] S63. Map the normalized index parameters to different element positions of the row and column vectors of the matrix to form an instantaneous feature vector containing the main feature ID of the bubble sound.

[0023] S64. Assign different weights to the index parameters of different single-wave signals in the instantaneous feature vector and perform weighted averaging. Then rearrange and combine the resulting new vectors to find the instantaneous contour shape matrix A that best represents the main physical sound state of the noise data segment.

[0024] S65. Multiply each index parameter in the instant contour shape matrix A by a correction value to obtain the corrected contour shape matrix C containing several new vectors.

[0025] S66. Input the corrected contour shape matrix C into the pre-trained turbine cavitation sound recognition model and output the cavitation discrimination result.

[0026] Furthermore, if the modified profile shape matrix C is equal to the modified profile shape matrix D, the output result is that the turbine runner has cavitation; otherwise, the output result is that cavitation has not occurred.

[0027] Furthermore, the relevant analytical parameters include time-domain parameters, power spectral density parameters, and frequency-domain parameters; the time-domain analytical parameters include peak-to-peak value vpp, interquartile frequency probability PVpH, standard deviation St, kurtosis Ku, skewness Sk, and information entropy H; the frequency-domain parameters include centroid frequency PsdFc.

[0028] Furthermore, the weighting range for the index parameters belonging to the first main frequency signal is 0.6 to 0.8.

[0029] Furthermore, the zero-mean standardization method was used to normalize the relevant analytical index parameters.

[0030] Furthermore, the mixed signal in the noisy data is decomposed using a 0.97 confidence probability assignment function or a mixing sampling function.

[0031] Furthermore, the correction value is a discrimination coefficient or range, and the correction value is determined by combining empirical parameters and turbine cavitation test data.

[0032] In summary, the present invention has the following advantages:

[0033] 1. This method extracts feature data from the initial cavitation sound signal and pulsation signal of the water turbine and combines them to form a feature vector PCSV for judging the cavitation of the water turbine. This feature vector PCSV is different from conventional sound elements. It is a combination of a newly defined cannon sound spectrum and a special pulsation spectrum. The water turbine cavitation bubble sound recognition model obtained by inputting this feature vector into the neural network for training can improve the accuracy and efficiency of the machine in recognizing the initial cavitation phenomenon of the water turbine, and meet the 80% accuracy of intelligent judgment of cavitation by the machine.

[0034] 2. This method uses a brand-new and more accurate physical acoustic testing technology to replace the technology based on traditional physical acoustic configuration, avoiding the subjectivity and uncertainty of human testing, realizing the complete replacement of human labor by machines, greatly reducing the testing cost by more than 50%, bringing considerable economic benefits, realizing the digitalization and intelligentization of testing technology, and its scalability to other industries. Attached Figure Description

[0035] Figure 1 To collect a set of time-domain noise data of cavitation in the model rotor;

[0036] Figure 2 This is the cavitation identification and judgment process of the present invention. Detailed Implementation

[0037] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further clarifies the invention. Those skilled in the art should understand that the specific descriptions below are illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention. The terms "first," "second," etc., used in the specification, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, or apparatuses.

[0038] When cavitation does not occur in the turbine, the noise signal is mainly low-frequency, and its main components are water flow sound, electromagnetic noise and other background noise of the turbine. The power spectral density of its low-frequency components has a high amplitude, and it has the characteristics of large amplitude and low frequency. As cavitation occurs in the model runner, the amplitude of the mid-to-high frequency band tends to gradually increase.

[0039] like Figure 1 The image shows a set of noise data collected during cavitation of a model turbine runner. After analyzing the massive amount of cavitation noise data collected from the turbine runner, it was found that the sound state of the bubbles generated on the surface of the turbine runner blades has two characteristics: first, the sound pattern of the cavitation bubbles is similar to the sound spectrum of firecrackers; second, it is mixed with the special pressure pulsation sound spectrum unique to the turbine. That is, the characteristic sound state of cavitation is a distinct polymorphic mixed sound spectrum that includes both the sound spectrum of firecrackers and the special pulsation spectrum.

[0040] Since current experimental techniques cannot identify this signal, this invention proposes an intelligent identification method for turbine cavitation phenomena based on combined spectrum feature extraction. It is a polymorphic sound vector recognition (PSVFR) method that uses a newly defined turbine cavitation sound vector (PCSV) as a new tool for cavitation identification to help the machine intelligently judge the cavitation phenomenon of the model runner.

[0041] Specifically, this method includes the following steps:

[0042] Step 1: Collect noise data when cavitation occurs on the turbine runner model;

[0043] Step 2: Calculate the statistical characteristics of the noise data:

[0044] Since the cavitation sound of a water turbine is a mixed spectrum of firecracker-like noise and the special pressure pulsation sound of the water turbine, the sound spectrum of the bubbles in a large amount of collected noise data during water turbine cavitation is manually (by experts) decomposed to obtain relevant analytical index parameters representing the characteristics of the firecracker-like sound (first dominant frequency signal) and the special pressure pulsation sound (second dominant frequency signal). The specific operation is as follows:

[0045] First, the mixed signal in the acquired noise data is decomposed into several single-wave signals using a 0.97 confidence probability assignment function or a mixing sampling function. The first and second main frequency signals are then extracted. These two main frequency signals correspond to firecracker-like sounds and special pressure pulsation sounds, respectively.

[0046] Then, the statistical characteristics of the two single-wave signals are calculated respectively to obtain the correlation analysis parameters, including time-domain parameters, power spectral density parameters and frequency-domain parameters.

[0047] The time-domain analysis parameters include peak-to-peak value (Vpp); interquartile frequency probability (PVpH); mean (Mean); standard deviation (St); kurtosis (Ku); skewness (Sk); and information entropy (H).

[0048] Frequency domain analysis parameters include centroid frequency PsdFc; frequency standard deviation PsdRvf; and power spectrum amplitude mean PsdH (high frequency), PsdM (mid frequency), and PsdL (low frequency).

[0049] Since the spectral characteristics of cavitation noise signals in turbine models are sometimes not obvious, a power spectral density (PSD) parameter is introduced as a correction factor to prevent clutter interference.

[0050] In classification and recognition applications, frequency domain parameters, time domain parameters, and power spectral density parameters are used together as part of the feature vector to enrich the types of features and improve diagnostic accuracy.

[0051] Step 3: Normalization

[0052] Considering that min-max standardization, or deviation standardization, cannot eliminate the influence of variance on dimensions, this embodiment uses zero-mean standardization to normalize the parameters of the noisy data. Specifically, the mean (Mean) and standard deviation (St) of each category of indicators are calculated using the formula... = (X-Mean) / St calculates the normalized data for each type of indicator.

[0053] Step 4: Construct a cavitation sound recognition model for water turbines

[0054] Different weights were assigned to the index parameters of different main frequency signals, and a weighted average was performed to form parameters that represent the mixed bubble sound characteristics of cavitation in a water turbine. Among them, the weights of the index parameters representing the sound spectrum characteristics of firecrackers ranged from 0.6 to 0.8.

[0055] These extracted mixed cavitation sound characteristic parameters are selectively mapped to different row and column element positions of a matrix (the positions can be changed), forming a contour morphology matrix B, such as [Vpp, PVpH, std, ku, sk, IH, FC]. This contour morphology matrix B constitutes a machine-recognizable qualitative model of turbine cavitation sound, that is, it forms the qualitative elements for machine intelligent recognition.

[0056] Since the cavitation process of a hydro-turbine is a real-time dynamic process with complete randomness, it is impossible for the qualitative model of hydro-turbine cavitation sound to correspond one-to-one with the object model formed by the real-time collected hydro-turbine noise data. Therefore, in conjunction with human (expert) experience parameters, a correction coefficient or interval (such as 0 or 1, [1, 1, 2], etc.) is preset for each characteristic parameter (Vpp, PVpH, Ku, Sk, etc.) of the mixed sound in the contour morphology matrix B. These correction coefficients or intervals are multiplied by each PCSV vector in the qualitative model of hydro-turbine cavitation sound and then recombined to form a corrected contour morphology matrix D. This corrected contour morphology matrix D constitutes the hydro-turbine cavitation sound recognition model that the machine can use. This cavitation sound recognition model can be continuously updated through intelligent learning. As the amount of cavitation sound data increases, the preset discrimination coefficients or intervals will also be continuously adjusted.

[0057] Step 5: Identification of cavitation sound in water turbines

[0058] The possible cavitation bubbles collected in real time are compared with the identified cavitation bubbles. The specific process is as follows:

[0059] 1) Collect real-time noise data of the turbine runner.

[0060] 2) Decompose the mixed signal in the instantaneous noise data into several single-wave signals, calculate the statistical characteristics of these single-wave signals respectively, obtain the relevant analysis index parameters and perform normalization processing; the categories and normalization methods of the relevant analysis index parameters obtained here are the same as those in steps two and three above.

[0061] 3) Map the normalized index parameters to different element positions of the row and column vectors of the matrix to form an instantaneous feature vector containing the main feature ID of the bubble sound.

[0062] 4) Assign different weights to the index parameters of different single-wave signals in the instantaneous feature vector and perform weighted averaging. Then rearrange and combine the new vectors to find the instantaneous contour shape matrix A that best represents the main physical sound state of the noise data segment. The instantaneous contour shape matrix A constitutes the machine-recognizable turbine cavitation sound object model, forming the object element for machine intelligent recognition.

[0063] For each characteristic parameter (Vpp, PVpH, Ku, Sk, etc.) of the real-time acquired cavitation sound, a discrimination coefficient or interval (e.g., 0 or 1, [0.97, 3.2], etc.) is preset. These discrimination coefficients or intervals are multiplied by the PCSV feature vectors in the real-time acquired turbine cavitation sound object model, and then recombined into a new corrected profile morphology matrix C. This corrected profile morphology matrix C constitutes the machine-recognizable turbine cavitation sound discrimination tool model. This discrimination tool model can be continuously updated through intelligent learning, and the preset discrimination coefficients or intervals can be continuously adjusted as the time period of the real-time acquired cavitation sound data changes.

[0064] 5) Compare the modified contour morphology matrix C with the modified contour morphology matrix D. If they are equal, the turbine cavitation sound recognition model outputs the judgment result as "cavitation has occurred". If they are not equal, the output result is "cavitation has not occurred".

[0065] Although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for intelligent identification of cavitation phenomena in hydraulic turbines based on combined spectral feature extraction, characterized in that, Includes the following steps: S1. Collect noise data when cavitation occurs in the turbine runner model; S2. Decompose the mixed signal in the noise data into several single-wave signals, and extract the first main frequency signal corresponding to the firecracker sound and the second main frequency signal corresponding to the special pressure pulsation sound of the water turbine. S3. Calculate the statistical characteristics of the first main frequency signal and the second main frequency signal respectively, obtain the relevant analysis index parameters and perform normalization processing; S4. Map the normalized index parameters to different element positions of the row and column vectors of the matrix to form a PCSV vector containing the main feature ID of cavitation bubbles. S5. Assign different weights to the index parameters of different main frequency signals in the PCSV vector and perform weighted averaging. Then, arrange and combine the new vectors to find the contour shape matrix B that best represents the main physical sound state of the noise data segment. Multiply each index parameter in the contour shape matrix B by a correction value to obtain the corrected contour shape matrix D containing several new vectors. The corrected contour shape matrix D constitutes the turbine cavitation sound recognition model. S6. Process the real-time collected turbine cavitation sound data and input it into the turbine cavitation sound recognition model, and output the cavitation recognition result.

2. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1, characterized in that, Step S6 specifically includes: S61. Collect real-time noise data of the turbine runner; S62. Decompose the mixing signal in the instantaneous noise data into several single-wave signals, calculate the statistical characteristics of these single-wave signals respectively, obtain the relevant analysis index parameters and perform normalization processing. S63. Map the normalized index parameters to different element positions of the row and column vectors of the matrix to form an instantaneous feature vector containing the main feature ID of the bubble sound. S64. Assign different weights to the index parameters of different single-wave signals in the instantaneous feature vector and perform weighted averaging. Then rearrange and combine the resulting new vectors to find the instantaneous contour shape matrix A that best represents the main physical sound state of the noise data segment. S65. Multiply each index parameter in the instant contour shape matrix A by a correction value to obtain the corrected contour shape matrix C containing several new vectors. S66. Input the corrected contour shape matrix C into the pre-trained turbine cavitation sound recognition model and output the cavitation discrimination result.

3. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 2, characterized in that, If the modified profile shape matrix C is equal to the modified profile shape matrix D, the output result is that the turbine runner has cavitation; otherwise, the output result is that cavitation has not occurred.

4. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1 or 2, characterized in that, The relevant analysis parameters include time-domain parameters, power spectral density parameters, and frequency-domain parameters; the time-domain parameters include peak-to-peak value vpp, interquartile frequency probability PVpH, standard deviation St, kurtosis Ku, skewness Sk, and information entropy H; the frequency-domain parameters include centroid frequency PsdFc.

5. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1, characterized in that, The weighting range for the index parameters belonging to the first main frequency signal is 0.6 to 0.

8.

6. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1 or 2, characterized in that, The parameters of the relevant analysis indicators were normalized using the zero-mean standardization method.

7. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1 or 2, characterized in that, Use the 0.97 confidence probability assignment function or the mixing sampling function to decompose the mixed signal in the noisy data.

8. The intelligent identification method for cavitation phenomena in hydraulic turbines based on combined spectral feature extraction according to claim 1 or 2, characterized in that, The correction value is a discrimination coefficient or range, and the correction value is determined by combining empirical parameters and turbine cavitation test data.

Images