CN-122021295-A - Acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning

CN122021295ACN 122021295 ACN122021295 ACN 122021295ACN-122021295-A

Abstract

The application discloses an acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning, which comprises the steps of S1, multi-source sound field data acquisition and preprocessing, S2, three-dimensional sound field feature extraction and reconstruction, S3, sound field state diagnosis and classification, S4, metamaterial deployment scheme optimization based on migration learning, namely performing domain adaptation on strategy library data and current sound field features by utilizing a joint distribution adaptation method based on the sound field space features and the sound field state tags, generating a metamaterial layout scheme adapted to the sound field state tags through an optimization algorithm, and S5, dynamic deployment and feedback optimization. According to the application, by deploying the sensor array and fusing the near-field acoustic hologram/beam forming technology, the three-dimensional spatial distribution (amplitude and phase) of the sound field can be reconstructed, deep features including spatial textures are extracted, the limitation of traditional single-point or simple array measurement is broken through, and a comprehensive and accurate physical field state information basis is provided for subsequent intelligent decision.

Inventors

SHI JIANBO
CAI XUAN
ZHANG YING
ZHANG CHI

Assignees

国网湖北省电力有限公司电力科学研究院
湖北前能电力技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. An acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning, which is characterized by comprising the following steps: S1, multi-source sound field data acquisition and preprocessing, namely arranging an acoustic sensor array in a target area, acquiring multi-channel sound pressure signals, and performing filtering, denoising and time-frequency conversion processing to obtain a sound field time-frequency characteristic matrix; S2, three-dimensional sound field feature extraction and reconstruction, namely processing the sound field time-frequency feature matrix based on a near-field sound holographic technology or a beam forming technology, reconstructing three-dimensional sound field distribution of a target area, and extracting sound field space features from the three-dimensional sound field distribution; s3, sound field state diagnosis and classification, namely inputting the sound field space features into a pre-trained machine learning model, classifying and identifying the sound field space features by the machine learning model, and outputting a sound field state label representing the current sound field state; S4, optimizing a metamaterial deployment scheme based on transfer learning, namely performing domain adaptation on strategy library data and current sound field characteristics by utilizing a joint distribution adaptation method based on the sound field space characteristics and the sound field state labels and combining strategy library constructed by historical deployment data, and generating a metamaterial layout scheme adapted to the sound field state labels through an optimization algorithm or reinforcement learning; And S5, dynamic deployment and feedback optimization, namely controlling the deployment of the adjustable metamaterial unit according to the metamaterial layout scheme, monitoring the deployed sound field response in real time to calculate performance indexes, and iteratively updating the machine learning model and the deployment strategy by using monitoring data.
2. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in S1, the multi-source sound field data acquisition and preprocessing comprises the following steps: s1.1, carrying out band-pass filtering on the acquired time domain signals of all channels, wherein the range of the band pass covers a target frequency band; S1.2, denoising the filtered signal by adopting a self-adaptive filtering or statistical characteristic-based background noise suppression method; s1.3, performing time-frequency analysis on the denoised time domain signal to obtain time-frequency distribution; S1.4, constructing a three-dimensional sound field time-frequency characteristic matrix according to time-frequency analysis results of all array elements.
3. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in S2, reconstructing three-dimensional sound field distribution based on near-field acoustic holographic technology comprises: performing two-dimensional space Fourier transform on the holographic sound pressure distribution to convert the holographic sound pressure distribution into a wave number domain; back-propagating the wavenumber spectrum from the holographic plane to a series of reconstruction planes using a back-propagation function; and carrying out two-dimensional space inverse Fourier transform on the wave number spectrum of each reconstruction surface to obtain complex sound pressure distribution on each reconstruction surface, and synthesizing the complex sound pressure distribution into three-dimensional sound pressure distribution of the target area.
4. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in S2, the reconstructing the three-dimensional sound field distribution based on the beam forming technique comprises: discretizing a target three-dimensional space into grid points, for each grid point, carrying out phase compensation and weighted summation on complex sound pressure of each channel according to theoretical time delay of sound wave propagation to each sensor, and calculating beam output energy of the grid point; And obtaining a three-dimensional sound field energy distribution map by carrying out space scanning on all grid points.
5. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in the step S3, the pre-trained machine learning model is a deep neural network-based multi-task learning model; The multi-task learning model comprises a shared feature encoder and a plurality of parallel task specific branches, wherein the shared feature encoder is used for extracting general feature representations from the sound field space features, and the task specific branches are respectively used for outputting classification results related to sound field types, sound source attributes and noise characteristics to jointly form the sound field state label.
6. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 5, wherein a joint loss function is adopted when the multi-task learning model is trained, the joint loss function is a weighted sum of loss functions of specific branches of each task, and parameters of a shared feature encoder and the specific branches of each task are updated simultaneously through a back propagation algorithm.
7. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in S3, the training or application process of the machine learning model comprises a migration learning step based on joint distribution adaptation, and the migration learning step is used for adapting the feature distribution of source domain data to the feature distribution of target domain data so as to improve the generalization capability of the model in the target domain.
8. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in S4, the generating a metamaterial layout scheme through reinforcement learning comprises: Modeling a metamaterial deployment process into a Markov decision process, wherein the state is the current sound field characteristic after domain adaptation, and the action is a metamaterial deployment vector; pre-training the reinforcement learning intelligent agent by utilizing strategy library data subjected to domain adaptation screening; The intelligent agent interacts with the real sound field environment, selects deployment actions according to the current state, updates the strategy according to the feedback after execution, and finally outputs the optimal deployment scheme.
9. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in the step S4, the metamaterial layout scheme is generated by an optimization algorithm, wherein in a feature space after domain adaptation, a plurality of historical cases most similar to the current sound field characteristics in a strategy library are searched, and corresponding deployment actions are obtained as candidate sets; Predicting expected performance of each candidate action under the current sound field by using a return prediction model; And taking the candidate action with the highest prediction performance as a starting point, performing local fine tuning optimization in an action space, and outputting an optimal deployment scheme.
10. The acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning according to claim 1, wherein in the step S5, the dynamic deployment and feedback optimization comprises the steps of acquiring new sound field data through an acoustic sensor array after executing the metamaterial layout scheme; Calculating a real-time performance index based on the new sound field data, and updating the current sound field state characteristics; And taking the deployment action, the real-time performance index and the updated state characteristics as feedback data for on-line fine tuning deployment strategy or periodically retraining the machine learning model.

Description

Acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning Technical Field The application belongs to the technical field of noise control, and particularly relates to an acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning. Background With the development of cities and the continuous increase of power demands, indoor substations are increasingly widely applied, and the problem of low-frequency noise pollution generated during operation of the indoor substations is also increasingly remarkable. The transformer substation noise mainly comes from magnetostriction effect generated by a transformer core under a power frequency alternating magnetic field, so that the noise presents a characteristic of 'discrete low-frequency line spectrum' taking 100Hz as a fundamental frequency, 200Hz and the like as harmonic waves. The low-frequency sound wave has a longer wavelength (such as about 3.4 meters in 100Hz sound wave wavelength), the traditional porous sound absorbing material is difficult to effectively absorb, and the traditional porous sound absorbing material is easy to penetrate through a building structure, so that the indoor sound environment is obviously disturbed. Particularly for indoor substations, strong acoustic standing waves are easy to form in the closed space, so that sound field distribution is very uneven, noise in certain positions is obviously amplified, and acoustic comfort and equipment operation environment are seriously affected. The existing noise reduction means such as a single resonance structure can only aim at a single frequency, and the adoption of multiple layers of metamaterial to respectively deal with different frequency bands can lead to the excessively thick structure, heavy weight and high cost, and the intelligent adaptability to specific sound field distribution is lacking. The current metamaterial deployment is dependent on experience or uniform coverage, and distribution differences of standing wave antinodes and nodes in an actual sound field are not fully considered, so that the material utilization rate is low, and the noise reduction effect is unbalanced. Although acoustic measurement and simulation technologies are mature, how to intelligently couple sound field diagnosis data with metamaterial deployment strategies to realize dynamic, accurate and self-adaptive efficient noise reduction is still a technical problem to be solved in the field. Therefore, it is necessary to provide a systematic method capable of realizing real-time diagnosis of a deep fusion sound field, intelligent decision of machine learning and precise regulation and control of adjustable metamaterials, so as to realize efficient, economical and self-adaptive suppression of specific low-frequency noise in a complex sound field environment. Disclosure of Invention The application provides an acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning, and aims to solve the problems that the prior art lacks intelligent adaptability to specific sound field distribution, the current metamaterial deployment is dependent on experience or uniform coverage, and the distribution difference of standing wave antinodes and nodes in an actual sound field is not fully considered. An acoustic metamaterial optimal deployment method based on sound field diagnosis and machine learning, the method comprising: S1, multi-source sound field data acquisition and preprocessing, namely arranging an acoustic sensor array in a target area, acquiring multi-channel sound pressure signals, and performing filtering, denoising and time-frequency conversion processing to obtain a sound field time-frequency characteristic matrix; S2, three-dimensional sound field feature extraction and reconstruction, namely processing the sound field time-frequency feature matrix based on a near-field sound holographic technology or a beam forming technology, reconstructing three-dimensional sound field distribution of a target area, and extracting sound field space features from the three-dimensional sound field distribution; s3, sound field state diagnosis and classification, namely inputting the sound field space features into a pre-trained machine learning model, classifying and identifying the sound field space features by the machine learning model, and outputting a sound field state label representing the current sound field state; S4, optimizing a metamaterial deployment scheme based on transfer learning, namely performing domain adaptation on strategy library data and current sound field characteristics by utilizing a joint distribution adaptation method based on the sound field space characteristics and the sound field state labels and combining strategy library constructed by historical deployment data, and generating a metamaterial layout scheme adapted to the sound field state labels through an optimization al