Turboshaft engine centrifugal compressor noise source identification method based on neural network

Abstract

The invention provides a noise source identification method for a centrifugal compressor of a turboshaft engine based on a neural network, and the method achieves the determination and fault diagnosis of an aerodynamic noise source through the analysis of the spectrogram features generated in the operation process of the compressor. In the process of building the neural network model, the noise of the gas compressor is calculated by using a simulation method, and a sound signal-noise source model is built in combination with a CNN algorithm in the neural network. The technology is suitable for the fields of aero-engines, gas turbines and the like, has the advantages of being high in real-time performance, high in accuracy, good in expandability and the like, and provides an intelligent solution for rotating machinery noise control.

Description

Technical Field

[0001] The present invention relates to the technical field of aerodynamic noise analysis, and in particular to a method for identifying noise sources of a turboshaft engine centrifugal compressor based on a neural network. Background Art

[0002] The centrifugal compressor in a turboshaft engine is a device that increases gas pressure through mechanical compression and is widely used in energy, aerospace, industrial manufacturing, and other fields. However, the noise pollution caused by its high-speed operation urgently needs to be addressed. Aerodynamic noise, as a key characteristic reflecting the operating status of equipment, provides new possibilities for noise source identification through the construction of neural networks. Therefore, a neural network-based compressor noise source identification model provides a basis for equipment health monitoring, fault diagnosis, and noise reduction optimization, and has broad engineering application prospects. Summary of the Invention

[0003] The present invention aims to address the problems in the prior art by proposing a method for identifying noise sources in a turboshaft engine centrifugal compressor based on a neural network. The method can identify noise sources within a specific frequency range in real time based on experimental measurement results of existing turboshaft engine centrifugal compressors.

[0004] The present invention is achieved through the following technical solutions. The present invention proposes a method for identifying noise sources of a turboshaft engine centrifugal compressor based on a neural network. The method comprises the following steps:

[0005] Obtaining time-varying grid flow field parameters of a centrifugal compressor in a turboshaft engine under target operating conditions, wherein the grid flow field parameters include the center coordinates of each grid cell, an initial velocity vector, and flow field pressure data that varies with time;

[0006] Topologically expand the flow field grid of the centrifugal compressor in the turboshaft engine, construct an acoustic field calculation grid based on the spatial distribution of the test points and the grid scaling coefficient, and determine the size parameters of the acoustic field grid and the coordinate mapping relationship of the test points in the acoustic field grid;

[0007] Establishing a finite element model of the sound field, solving the acoustic pressure fluctuation equation for the grid flow field parameters, obtaining the time domain noise signal of each test point in the sound field grid, and converting the time domain noise signal into a frequency domain noise spectrum through fast Fourier transform;

[0008] Data preprocessing is performed on the above noise spectrum image. The preprocessed data is used as input to construct a CNN-based acoustic signal-noise source mapping model. The trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine.

[0009] Furthermore, the grid flow field parameters are preprocessed, specifically: data interpolation processing is performed on the original CFD calculation results, and a staggered grid design is adopted to ensure the spatial continuity of the flow field grid and the acoustic field grid.

[0010] Furthermore, the grid flow field parameters are solved by time-domain finite element method based on the sound pressure fluctuation equation to obtain the time-varying sound pressure signal of the preset monitoring point in the sound field calculation grid; the time-varying sound pressure signal is converted into a frequency-domain acoustic feature containing a 1 / 3 octave spectrum through fast Fourier transform.

[0011] Furthermore, the process of establishing the acoustic signal-noise source mapping model includes:

[0012] Obtain compressor noise sample data under different operating conditions through CFD simulation;

[0013] Convert the frequency domain noise spectrum into a two-dimensional spectrogram as the model input;

[0014] Build a deep neural network architecture consisting of convolutional layers, pooling layers, and fully connected layers;

[0015] The back-propagation algorithm is used to optimize the network weight parameters and establish the mapping relationship between acoustic features and noise sources.

[0016] Furthermore, the trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine, specifically:

[0017] The real-time collected acoustic signal is pre-processed to generate a standard spectrogram;

[0018] Automatically extract impeller rotation frequency features through the CNN neural network model;

[0019] Combined with the real-time operating parameters of speed and flow, the characteristic frequency components corresponding to different parts of the compressor are output.

[0020] Furthermore, the data preprocessing includes:

[0021] Label the noise type of the data and mark the noise source for each spectrogram;

[0022] Mark the peak position of the data and mark the peak coordinates of the noise source on the spectrogram.

[0023] Furthermore, a multi-task CNN model is constructed in the CNN-based acoustic signal-noise source mapping model, namely the CNN neural network model. The uses of this model include classifying noise types and locating peak heat maps; the above-mentioned large number of labeled spectrograms and peak positions are used as training data for constructing the CNN model; and the predicted classification results and positioning results are used as the prediction results of the CNN model to construct a related neural network model.

[0024] Furthermore, the centrifugal compressor in the turboshaft engine is used in the turboshaft engine, and is a single-stage centrifugal compressor with a wedge-shaped diffuser; the centrifugal impeller is composed of main blades and splitter blades, the centrifugal impeller is a moving blade, and the diffuser blade is a stationary blade.

[0025] The present invention has the following beneficial effects:

[0026] The present invention proposes a method for identifying the noise source of a turboshaft engine centrifugal compressor based on a neural network. By analyzing the acoustic spectrogram characteristics generated during the operation of the centrifugal compressor in the turboshaft engine, the aerodynamic noise source can be identified and fault diagnosis can be performed. During the construction of the neural network model, a simulation method is used to calculate the noise of the centrifugal compressor in the turboshaft engine, and a sound signal-noise source model is constructed in combination with the CNN algorithm in the neural network. During the analysis process, the acoustic signal is converted into a two-dimensional acoustic spectrogram and input into the established model. The deep neural network is used to automatically extract noise characteristics, effectively identifying complex noise sources such as impeller rotation and airflow separation. At the same time, combined with the equipment operating status parameters, data support is provided for health monitoring, noise reduction optimization, and surge warning of the centrifugal compressor in the turboshaft engine. This technology is applicable to fields such as aircraft engines and gas turbines. It has the advantages of strong real-time performance, high accuracy, and good scalability, and provides an intelligent solution for rotating machinery noise control. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the noise source identification method in the implementation process of the present invention;

[0028] Figure 2 This is the overall structure diagram of the centrifugal compressor in the turboshaft engine;

[0029] Figure 3 This is the time domain diagram of the simulated aerodynamic noise;

[0030] Figure 4 This is the frequency domain diagram of the simulated aerodynamic noise after Fourier transform;

[0031] Figure 5 Schematic diagram of the neural network aerodynamic noise execution process;

[0032] Figure 6 Flowchart of the iterative strategy during the neural network construction process;

[0033] Figure 7 This is a schematic diagram of the frequency domain diagram of the experimental aerodynamic noise and the identification results. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] This paper analyzes the acoustic spectrogram characteristics generated during the operation of a turboshaft engine's centrifugal compressor to identify aerodynamic noise sources and diagnose faults. During the neural network modeling process, simulation methods were used to calculate the turboshaft engine's centrifugal compressor noise, and the CNN algorithm within the neural network was used to construct an acoustic signal-noise source model.

[0036] During the analysis process, the acoustic signal is converted into a two-dimensional spectrogram and input into the established model. The deep neural network is used to automatically extract noise features, effectively identifying complex noise sources such as impeller rotation and airflow separation. At the same time, combined with the equipment operating status parameters, it provides data support for the health monitoring, noise reduction optimization and surge warning of the centrifugal compressor in the turboshaft engine.

[0037] Specifically, see Figure 1-Figure 7 The present invention proposes a method for identifying noise sources of a turboshaft engine centrifugal compressor based on a neural network, the method comprising the following steps:

[0038] Obtaining time-varying grid flow field parameters of a centrifugal compressor in a turboshaft engine under target operating conditions, wherein the grid flow field parameters include the center coordinates of each grid cell, an initial velocity vector, and flow field pressure data that varies with time;

[0039] Topologically expand the flow field grid of the centrifugal compressor in the turboshaft engine, construct an acoustic field calculation grid based on the spatial distribution of the test points and the grid scaling coefficient, and determine the size parameters of the acoustic field grid and the coordinate mapping relationship of the test points in the acoustic field grid;

[0040] Establishing a finite element model of the sound field, solving the acoustic pressure fluctuation equation for the grid flow field parameters, obtaining the time domain noise signal of each test point in the sound field grid, and converting the time domain noise signal into a frequency domain noise spectrum through fast Fourier transform;

[0041] Data preprocessing is performed on the above noise spectrum image. The preprocessed data is used as input to construct a CNN-based acoustic signal-noise source mapping model. The trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine.

[0042] This paper proposes a neural network-based compressor noise source identification method. By analyzing the acoustic spectrogram characteristics generated during compressor operation, it can identify aerodynamic noise sources and conduct fault diagnosis. During the neural network model construction process, a simulation method is used to calculate compressor noise, and the CNN algorithm in the neural network is combined to construct an acoustic signal-noise source model. This technology is suitable for fields such as aircraft engines and gas turbines. It has the advantages of strong real-time performance, high accuracy, and good scalability, providing an intelligent solution for rotating machinery noise control.

[0043] The centrifugal compressor in the turboshaft engine is used in the turboshaft engine and is a single-stage centrifugal compressor with a wedge-shaped diffuser; the centrifugal impeller is composed of main blades and splitter blades, the centrifugal impeller is a moving blade, and the diffuser blade is a stationary blade.

[0044] The grid flow field parameters are preprocessed, specifically: data interpolation processing is performed on the original CFD calculation results, and a staggered grid design is adopted to ensure the spatial continuity of the flow field grid and the acoustic field grid.

[0045] The analytical process of the sound field parameters specifically includes: performing a time-domain finite element solution on the grid flow field parameters based on the sound pressure fluctuation equation to obtain a time-varying sound pressure signal at a preset monitoring point in the sound field calculation grid; and converting the time-varying sound pressure signal into a frequency-domain acoustic feature containing a 1 / 3 octave spectrum through a fast Fourier transform.

[0046] The process of establishing the acoustic signal-noise source mapping model includes:

[0047] Obtain compressor noise sample data under different operating conditions through CFD simulation;

[0048] Convert the frequency domain noise spectrum into a two-dimensional spectrogram as the model input;

[0049] A deep neural network architecture consisting of convolutional layers, pooling layers, and fully connected layers is established; the convolutional layers are used to extract local features of the spectrogram, the pooling layers are used to reduce the feature dimensionality, and the fully connected layers are used to map the features to the output space of the classification and localization tasks.

[0050] The back-propagation algorithm is used to optimize the network weight parameters and establish the mapping relationship between acoustic features and noise sources.

[0051] The trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine, specifically;

[0052] The real-time collected acoustic signal is pre-processed to generate a standard spectrogram;

[0053] Automatically extract impeller rotation frequency features through the CNN neural network model;

[0054] Combined with the real-time operating parameters of speed and flow, the characteristic frequency components corresponding to different parts of the compressor are output.

[0055] The data preprocessing includes:

[0056] Label the noise type of the data and mark the noise source (such as impeller rotation, airflow separation, etc.) for each spectrogram;

[0057] Mark the peak position of the data and mark the peak coordinates (such as time and frequency) of the noise source on the spectrogram.

[0058] A multi-task CNN model is constructed within the CNN-based acoustic signal-noise source mapping model, also known as the CNN neural network model. This model's functions include classifying noise types and locating peak heatmaps. This model employs a multi-task learning architecture to simultaneously complete the noise type classification task and the noise source peak location localization task. The model input is a preprocessed two-dimensional spectrogram, and the output includes noise type classification results and a heatmap of noise source peak locations. This clarifies the model's task objectives (classification and localization) and provides guidance for subsequent network architecture design and implementation. The large number of annotated spectrograms and peak locations are used as training data for the CNN model. This data is then used to train the model, and the network weights are continuously adjusted through a backpropagation algorithm to gradually improve the model's classification accuracy and localization precision. The predicted classification and localization results are used as the predictions for the CNN model, and the relevant neural network model is constructed.

[0059] In this embodiment of the present invention, multi-task learning is implemented, where the classification task is optimized using the cross-entropy loss function, and the localization task is optimized using the mean squared error loss function. The loss functions of the two tasks are jointly optimized using a weighted summation method to ensure a balanced model for classification and localization.

[0060] The validation and application of the present invention's embodiments involve evaluating the model's performance on a validation set to ensure its generalization capabilities in classification and localization. In practical applications, the real-time acoustic signal is preprocessed into a standard spectrogram and then fed into the model, which automatically extracts noise features and outputs noise source identification results.

[0061] After establishing the above-mentioned CNN neural network model, noise source identification can be performed. This model targets the centrifugal compressor in the turboshaft engine. The relevant spectrum diagram obtained from the centrifugal compressor test is input into the model, and the frequencies corresponding to different components in the centrifugal compressor in the turboshaft engine can be identified.

[0062] In summary, the method described in the present invention converts acoustic signals into two-dimensional spectrograms and inputs them into the established model. It uses a deep neural network to automatically extract noise features, effectively identifying complex noise sources such as impeller rotation and airflow separation. At the same time, combined with equipment operating status parameters, it provides data support for health monitoring, noise reduction optimization, and surge warning of centrifugal compressors in turboshaft engines.

[0063] The present invention proposes a method for identifying the noise source of a turboshaft engine centrifugal compressor based on a neural network. By analyzing the acoustic spectrogram characteristics generated during the operation of the centrifugal compressor in the turboshaft engine, the aerodynamic noise source can be identified and fault diagnosis can be performed. During the construction of the neural network model, the noise of the centrifugal compressor in the turboshaft engine is calculated using a simulation method, and the CNN algorithm in the neural network is combined to construct an acoustic signal-noise source model. During the analysis process, the acoustic signal is converted into a two-dimensional acoustic spectrogram and input into the established model. The deep neural network is used to automatically extract noise characteristics, effectively identifying complex noise sources such as impeller rotation and airflow separation. At the same time, combined with the equipment operating status parameters, it provides data support for health monitoring, noise reduction optimization, and surge warning of the centrifugal compressor in the turboshaft engine. This model is suitable for fields such as aircraft engines and gas turbines. It has the advantages of strong real-time performance, high accuracy, and good scalability, and provides an intelligent solution for rotating machinery noise control.

[0064] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the definition of the claims.

Claims

1. A method for identifying noise sources in a turboshaft engine centrifugal compressor based on a neural network, characterized by: The method comprises the following steps: Obtaining time-varying grid flow field parameters of a centrifugal compressor in a turboshaft engine under target operating conditions, wherein the grid flow field parameters include the center coordinates of each grid cell, an initial velocity vector, and flow field pressure data that varies with time; Topologically expand the flow field grid of the centrifugal compressor in the turboshaft engine, construct an acoustic field calculation grid based on the spatial distribution of the test points and the grid scaling coefficient, and determine the size parameters of the acoustic field grid and the coordinate mapping relationship of the test points in the acoustic field grid; Establishing a finite element model of the sound field, solving the acoustic pressure fluctuation equation for the grid flow field parameters, obtaining the time domain noise signal of each test point in the sound field grid, and converting the time domain noise signal into a frequency domain noise spectrum through fast Fourier transform; Data preprocessing is performed on the above noise spectrum image. The preprocessed data is used as input to construct a CNN-based acoustic signal-noise source mapping model. The trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine.

2. The method according to claim 1, characterized in that The grid flow field parameters are preprocessed, specifically: data interpolation processing is performed on the original CFD calculation results, and a staggered grid design is adopted to ensure the spatial continuity of the flow field grid and the acoustic field grid.

3. The method according to claim 1, characterized in that Based on the sound pressure fluctuation equation, the grid flow field parameters are solved by time-domain finite element method to obtain the time-varying sound pressure signal of the preset monitoring point in the sound field calculation grid; the time-varying sound pressure signal is converted into a frequency-domain acoustic feature containing a 1 / 3 octave spectrum through fast Fourier transform.

4. The method according to claim 1, wherein The process of establishing the acoustic signal-noise source mapping model includes: Obtain compressor noise sample data under different operating conditions through CFD simulation; Convert the frequency domain noise spectrum into a two-dimensional spectrogram as the model input; Build a deep neural network architecture consisting of convolutional layers, pooling layers, and fully connected layers; The back-propagation algorithm is used to optimize the network weight parameters and establish the mapping relationship between acoustic features and noise sources.

5. The method according to claim 1, characterized in that The trained CNN neural network model is used to output the frequencies corresponding to different components of the centrifugal compressor in the turboshaft engine, specifically; The real-time collected acoustic signal is pre-processed to generate a standard spectrogram; Automatically extract impeller rotation frequency features through the CNN neural network model; Combined with the real-time operating parameters of speed and flow, the characteristic frequency components corresponding to different parts of the compressor are output.

6. The method according to claim 1, characterized in that The data preprocessing includes: Label the noise type of the data and mark the noise source for each spectrogram; Mark the peak position of the data and mark the peak coordinates of the noise source on the spectrogram.

7. The method according to claim 6, characterized in that A multi-task CNN model is constructed in the CNN-based acoustic signal-noise source mapping model, i.e., the CNN neural network model. The uses of this model include classifying noise types and locating peak heat maps. The large number of labeled spectrograms and peak positions mentioned above are used as training data for constructing the CNN model. The predicted classification results and positioning results are used as the prediction results of the CNN model construction to construct a related neural network model.

8. The method according to claim 1, characterized in that The centrifugal compressor in the turboshaft engine is used in the turboshaft engine and is a single-stage centrifugal compressor with a wedge-shaped diffuser; the centrifugal impeller is composed of main blades and splitter blades, the centrifugal impeller is a moving blade, and the diffuser blade is a stationary blade.

Images