Acoustic imaging method based on multi-task learning network

Abstract

The invention discloses an acoustic imaging method based on a multi-task learning network, and the method comprises the following steps: S1, taking an acoustic positioning task and a sound source quantification task as a feature regression problem, and directly extracting the features of an original signal as network input; s2, based on a position prediction branch of a time-frequency characteristic designed by a multi-task learning framework and an intensity prediction branch based on an amplitude characteristic, carrying out accurate modeling on a positioning and quantification task, and predicting sound source intensity by using the amplitude characteristic; and S3, combining the position prediction branch and the intensity prediction branch to design an Acoustic-Net model, learning the characteristics of amplitude and phase change by using the original audio data, and fully verifying the validity. According to the method, position and strength estimation with higher precision can be obtained, hardware resources can be saved, calculation loss can be reduced, the operation efficiency is higher, and the requirements of daily industrial production can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of acoustic imaging, and in particular to an acoustic imaging method based on a multi-task learning network. Background Art

[0002] In daily life, the human ear can identify and locate various sounds it hears. For example, when receiving sounds from the surroundings, people can quickly determine the direction and volume of the sound target. However, the original sound positioning function of the human ear is only to solve basic problems of life and survival, and the positioning accuracy is very limited. In order to locate and quantify sound signals more accurately, acoustic imaging technology based on microphone arrays has begun to attract people's attention.

[0003] Acoustic imaging refers to the use of microphone arrays to collect multi-channel audio data, using beamforming algorithms at specified frequencies to calculate the sound source distribution information on the scanning plane, and then fusing the sound source distribution information with the real-life image to quickly determine the spatial position and source of the sound source on the intuitive fused image. Since acoustic imaging technology can intuitively and quickly determine the spatial distribution of sound sources through thermal maps, it plays an important role in transportation, noise detection, industrial anomaly detection and other fields, and has a wide range of application value.

[0004] In recent years, with the powerful feature representation ability of deep learning, acoustic imaging algorithms based on deep networks have received widespread attention. For example, Reiter and Bell adopted the widely used AlexNet (Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks [J]. Communications of the ACM, 2017, 60 (6): 84-90) to identify the location of the sound source, and its input is the pre-imaging result obtained after the delayed sum algorithm. Allman et al. suggested that the sound source localization problem be regarded as a classification task and designed a deep convolutional network based on VGGNet (Simonyan K, Zisserman A. Very deep convolutional networks for large scale image recognition[J].arXiv preprint arXiv:1409.1556,2014) and Fast R-CNN(Girshick R.Fast r-cnn[C] / / Proceedings of the IEEE international conferenceon computer vision.2015:1440-1448) dual-module deep network. However, the above two methods can only estimate the location of the sound source and cannot provide intensity prediction. To deal with this problem, Kujawski et al. used ResNet(He K, Zhang X, Ren S, et al.Deep residual learning for image recognition[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition.2016:770-778) architecture to process the results of pre-imaging, and unified the sound source localization and quantization into a parameter regression problem, and successfully estimated the location and intensity of the sound source in the case of a single sound source. However, due to the complex network architecture, the above methods are difficult to integrate into systems with limited hardware resources, and they are often only applicable to simple scenarios with a single sound source. In recent years, some scholars have proposed neural network processing solutions for multiple sound sources (Feng L, Zan M, Huang L, et al. A double-step grid-free method for sound source identification using deep learning [J]. Applied Acoustics, 2022, 201: 109099), but such methods usually require more prior knowledge, such as the number of sound sources to be measured.

[0005] In summary, existing deep network-based methods are completely data-driven and use brute force stacking of modules to improve network performance, ignoring the rich information contained in acoustic domain knowledge, making them sensitive to data and lacking interpretability. In addition, such methods are limited by complex network architectures and are difficult to integrate into hardware systems with limited computing resources, making the algorithm unable to meet the needs of daily industrial production. Summary of the invention

[0006] The purpose of the present invention is to provide an acoustic imaging method based on a multi-task learning network to solve the above defects.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] The acoustic imaging method based on a multi-task learning network includes the following steps:

[0009] S1. The acoustic localization task and the sound source quantification task are regarded as feature regression problems, and the features of the original signal are directly extracted as network input;

[0010] S2, a position prediction branch based on time-frequency features and an intensity prediction branch based on amplitude features designed based on a multi-task learning framework to accurately model the localization and quantization tasks, that is, using time-frequency features to learn phase changes and using amplitude features to predict the intensity of the sound source;

[0011] S3. Combining the position prediction branch and the intensity prediction branch, we design a concise and effective overall model, Acoustic-Net. We use the original audio data to learn the characteristics of amplitude and phase changes, and fully verify the effectiveness of the network, thereby completing the entire process of acoustic imaging based on a multi-task learning network.

[0012] Preferably, in step S2, the position prediction branch of the time-frequency feature specifically operates as follows:

[0013] ① First, the original acoustic signal is short-time Fourier transformed to convert it into a time-frequency distribution diagram to utilize the rich phase information in the time-frequency representation;

[0014] Assuming the original signal is s(t), STFT converts it to the time-frequency domain through window function and Fourier transform, that is:

[0015]

[0016] Where s(n) is the input signal; g(nm) is the window function, which is a smooth time window used to intercept a part of the input signal at a specific time position; m is the center position of the sliding window; n is the time index, which represents the sampling point of the signal s(n); ω is the frequency variable, which represents the angular frequency in the frequency domain; e -jnω is a complex exponential used to convert signals from the time domain to the frequency domain; X(m,ω) is the result of STFT, which is the representation of the input signal s(n) at time n and frequency ω;

[0017] Then, the time-frequency plot of the input signal is defined as:

[0018] S(m,ω)=|X(m,ω)| 2 ;

[0019] ②. MobileOne is used as the backbone network. By combining reparameterization technology and deep separable convolution, structural simplification and improved computational efficiency are achieved in the training and inference stages.

[0020] Preferably, during the training phase, the MobileOne has a depthwise separable convolution architecture consisting of an upper and lower part, namely a depthwise convolution module in the upper part and a pointwise convolution module in the lower part;

[0021] The deep convolution module consists of three branches: the leftmost branch is a 1×1 convolution; the middle branch is a re-parameterized 3×3 convolution, i.e., k 3×3 convolutions; the right part is a skip connection containing a BN layer;

[0022] The point convolution module consists of two branches: the left branch is a re-parameterized 1×1 convolution, which is composed of k 1×1 convolutions; the right branch is a skip connection containing a BN layer.

[0023] Preferably, for the reparameterization of the deep convolution module, the process is specifically as follows:

[0024] Assuming that the input channel of the depth convolution part is d, the number of groups of the depth convolution is also d;

[0025] For the 1×1 convolution branch on the left, the 1×1 convolution and BN are merged into a 1×1 convolution with d grouping numbers using the reparameterization method based on the merging of the convolution layer and the BN layer. This convolution is recorded as A.

[0026] For the middle reparameterized 3×3 convolution branch, for each individual 3×3 convolution, the 3×3 convolution and its BN are merged into a 3×3 convolution with a group number of d using reparameterization method 1 (merging of convolution layer and BN layer), thereby obtaining k 3×3 convolutions; then, reparameterization method 2 (multi-branch addition) is used to merge these k 3×3 convolutions into one convolution, which is recorded as B;

[0027] For the skip connection on the right, we first add a convolution before the BN layer. The parameters of the convolution are set manually. It needs to satisfy that for any input x, the output after the convolution is also x. Such a convolution is called Diracdelta. With the Dirac delta convolution, we can use the reparameterization method based on the merging of the convolution layer and the BN layer to merge it with the BN layer to obtain a new 3×3 convolution, which is denoted as C.

[0028] Use multiple parallel branches to perform convolution operations, each branch contains different operations. This allows the model to learn more diverse feature representations during training and better adapt to changes in input data;

[0029] The final merging process merges the convolution kernel (weight matrix) and bias of each branch into a new equivalent convolution kernel and bias.

[0030] Preferably, in the inference phase, MobileOne converts all multi-branch structures into a standard convolution operation to simplify calculations and improve inference efficiency, so that the network can fully utilize the computational acceleration of the hardware optimization library during inference, thereby saving computing resources of the terminal device.

[0031] Preferably, in step S2, the intensity prediction branch based on amplitude features specifically operates as follows: the original acoustic signal is used as the network input, and a shallow one-dimensional convolution process is used to effectively estimate the sound source intensity to make full use of the amplitude information in the original signal; the shallow one-dimensional convolution process includes 3 one-dimensional convolution layers and 4 fully connected layers.

[0032] Preferably, in step S3, the Acoustic-Net model has the following specific working steps: given a training data pair (s(t), x gt ,y gt ,p gt ), s(t) represents the input signal, (x gt ,y gt ) represents the target position coordinates, p gt Represents the target sound intensity; first, the acoustic signal is subjected to STFT, which is converted into the corresponding time-frequency image and input into the position prediction branch to obtain the prediction of the sound source position At the same time, the original signal is used as the input of the intensity prediction branch to obtain the intensity estimation of the sound source.

[0033] Preferably, in step S3, the network loss function of the Acoustic-Net model includes prediction deviations of position and intensity;

[0034] Let the training set be the work, and the network loss function can be expressed as:

[0035]

[0036] In the formula, α is the coordination factor, and L 1 is the position estimation loss, L 2 is the intensity estimation loss; Γ represents a set of training data sets, each element of which contains the information of a training sample, including the acoustic signal s(t), the corresponding target position coordinate (x gt ,y gt ) and target sound intensity p gt ; It means averaging the losses of all training samples in the set Γ.

[0037] Position estimation loss L 1 , which is expressed as:

[0038]

[0039] Strength estimation loss L 2 , which is expressed as:

[0040]

[0041] Wherein, the scale change operation 20lg(·) converts the sound source intensity to sound pressure level; p ref Represents the reference sound pressure, generally set to 2×10 -5 ; Represents an estimate of the strength of a sound source.

[0042] Preferably, in step S3, the validity of the network is fully verified, and the verification method is: based on the mean error MDE of the position estimation and the mean absolute percentage error MAPE of the SPL estimation;

[0043] The calculation formulas of MDE and MAPE are:

[0044]

[0045] The beneficial effects of the present invention are:

[0046] (1) The acoustic imaging method based on the multi-task learning network of the present invention combines acoustic signal processing with convolutional neural network. Based on the irrelevance of the sound source position and intensity, two independent network branches are proposed based on the multi-task learning framework, so that the model can focus on the representation of different features in a biased manner, thereby better learning the position and intensity features and improving the generalization ability of the network.

[0047] (2) The acoustic imaging method of the present invention is based on a multi-task learning network. In order to save hardware resources and reduce computing loss, a position prediction branch based on the MobileOne backbone network and an intensity prediction branch based on a shallow one-dimensional convolution process are designed. These two lightweight and easy-to-infer network architectures are easy to integrate into the hardware system and have practical engineering application value.

[0048] (3) The acoustic imaging method of the present invention is based on a multi-task learning network. The acoustic network Acoustic-Net can obtain more accurate position and intensity estimation and has more efficient operation efficiency, demonstrating the great potential of the model in real-time acoustic imaging applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 It is a structural diagram of the feature extraction module MobileOne of the present invention;

[0051] Figure 3 Schematic diagram of the spatial position of the 56-channel microphone array of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further described below in conjunction with the embodiments. It should be noted that these are merely examples and illustrations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should be deemed to fall within the protection scope of the present invention.

[0053] Embodiment 1:

[0054] Figure 1 Flow chart of the method of the present invention. Figure 1 As shown, the acoustic imaging method based on the multi-task learning network includes the following steps:

[0055] S1. The acoustic localization task and sound source quantification task are regarded as feature regression problems, and the features of the original signal are directly extracted as network input.

[0056] S2. Two independent network branches are designed based on the multi-task learning framework, namely, the position prediction branch based on time-frequency features and the intensity prediction branch based on amplitude features, to accurately model the positioning and quantization tasks, that is, using time-frequency features to learn phase changes and using amplitude features to predict the intensity of the sound source.

[0057] (1) Position prediction branch based on time-frequency features:

[0058] ① In order to make full use of the potential phase information in the original signal, the position prediction branch first performs a short-time Fourier transform (STFT) on the original acoustic signal and converts it into a time-frequency distribution graph to utilize the rich phase information in the time-frequency representation.

[0059] Specifically, assuming the original signal is s(t), STFT converts it to the time-frequency domain through window function and Fourier transform, that is:

[0060]

[0061] Where s(n) is the input signal; g(nm) is the window function, which is a smooth time window used to intercept a part of the input signal at a specific time position; m is the center position of the sliding window; n is the time index, which represents the sampling point of the signal s(n); α is the frequency variable, which represents the angular frequency in the frequency domain; e -jnωis a complex exponential used to convert the signal from the time domain to the frequency domain; X(m,ω) is the result of STFT, which is the representation of the input signal s(n) at time n and frequency ω.

[0062] Then, the time-frequency plot of the input signal is defined as:

[0063] S(m,ω)=|X(m,ω)| 2 .

[0064] ② In order to save hardware resources, it is necessary to design a lightweight and easy-to-infer network architecture. Figure 2 is a structural diagram of the feature extraction module MobileOne of the present invention, as shown Figure 2 As shown in the figure, the method of the present invention uses MobileOne as the backbone network. MobileOne combines reparameterization technology with deep separable convolution to achieve structural simplification and improve computational efficiency in the training and reasoning stages. Its core architecture includes the following parts:

[0065] During the training phase, MobileOne's depthwise separable convolution architecture consists of two parts, the upper part is the depthwise convolution, and the lower part is the pointwise convolution. The depthwise convolution module consists of three branches. The leftmost branch is a 1×1 convolution; the middle branch is a re-parameterized 3×3 convolution, that is, k 3×3 convolutions; the right part is a jump connection containing a BN layer. The pointwise convolution module consists of two branches. The left branch is a re-parameterized 1×1 convolution, consisting of k 1×1 convolutions; the right branch is a jump connection containing a BN layer. MobileOne can extract information from multiple scales and different feature spaces through multi-branch and multi-size convolution kernels.

[0066] The reparameterization process can be described as:

[0067] Assuming that the input channel of the depth convolution part is d, the number of groups of the depth convolution is also d.

[0068] For the 1×1 convolution branch on the left, the 1×1 convolution and BN are merged into a 1×1 convolution with d grouping numbers using reparameterization method 1 (merging the convolution layer and the BN layer), and this convolution is denoted as A.

[0069] For the middle reparameterized 3×3 convolution branch, for each individual 3×3 convolution, the 3×3 convolution and its BN are merged into a 3×3 convolution with a grouping number of d using reparameterization method one (merging of convolution layer and BN layer), thereby obtaining k 3×3 convolutions; then reparameterization method two (multi-branch addition) is used to merge these k 3×3 convolutions into one convolution, which is denoted as B.

[0070] For the skip connection on the right, we first add a convolution before the BN layer. The parameters of the convolution are set manually. It needs to satisfy that for any input x, the output after the convolution is also x. This kind of convolution is called Diracdelta. With Dirac delta convolution, we can use the reparameterization method 1 (merging the convolution layer and the BN layer) to merge it with the BN layer to get a new 3×3 convolution, which is denoted as C.

[0071] Multiple parallel branches are used for convolution operations, each branch contains different operations. This design enables the model to learn more diverse feature representations during training and better adapt to changes in input data.

[0072] The final merging process will merge the convolution kernel (weight matrix) and bias of each branch into a new equivalent convolution kernel and bias. This is done through a simple addition operation, which greatly reduces the complexity of the calculation in the inference phase, and the expressiveness of the model will not be significantly affected.

[0073] The point convolution module consists of two branches. The left branch is a re-parameterized 1×1 convolution, which consists of k 1×1 convolutions; the right branch is a jump connection containing a BN layer. The main function of the point convolution is to mix the results of multiple channels together to form a new feature map, so that the network can capture cross-channel feature interactions.

[0074] After training is completed, multiple convolution branches are fused into an equivalent single convolution layer through reparameterization. That is, in the inference phase, all multi-branch structures are converted to a standard convolution operation to simplify calculations and improve inference efficiency. This allows the network to fully utilize the computational acceleration of hardware optimization libraries (such as cuDNN and MKL) during inference, thereby saving computing resources on the terminal device. By using different network structures during training and inference, efficient hardware computer performance optimization is achieved.

[0075] S22. Intensity prediction branch based on amplitude features: In order to make full use of the amplitude information in the original signal, the intensity prediction branch takes the original acoustic signal as the network input and uses a shallow one-dimensional convolution process to effectively estimate the intensity of the sound source. The process contains 3 one-dimensional convolution layers and 4 fully connected layers.

[0076] S3. Combining the position prediction branch and the intensity prediction branch, a concise and effective overall model Acoustic-Net is designed. Acoustic-Net aims to use raw audio data to learn the characteristics of amplitude and phase changes. The model has fewer network parameters, fast inference speed, and is easy to integrate into the hardware system. The effectiveness of the network is fully verified through the Acoustic-Net model and extensive experiments.

[0077] Combining the position prediction branch and the intensity prediction branch, a concise and effective overall model Acoustic-Net is designed. Acoustic-Net aims to use raw audio data to learn the characteristics of amplitude and phase changes. Given a training data pair (s(t), x gt ,y gt ,p gt ), s(t) represents the input signal, (x gt ,y gt ) represents the target position coordinates, p gt Represents the target sound intensity. Acoustic-Net first performs STFT on the acoustic signal, converts it into the corresponding time-frequency image and inputs it into the position prediction branch to obtain the prediction of the sound source position. At the same time, the original signal is used as the input of the intensity prediction branch to obtain the intensity estimation of the sound source.

[0078] The loss function of the Acoustic-Net network contains the prediction bias of position and intensity, aiming to reduce the difference between the network estimate and the true label. Let the training set be the work, and the network loss function can be expressed as:

[0079]

[0080] In the formula, α is the coordination factor, and L 1 is the position estimation loss, L 2 is the intensity estimation loss; Γ represents a set of training data sets, each element of which contains the information of a training sample, including the acoustic signal s(t), the corresponding target position coordinate (x gt ,y gt ) and target sound intensity p gt ; It means averaging the losses of all training samples in the set Γ by dividing by the total number of samples. This form ensures that the loss values ​​are comparable on training data sets of different sizes, thereby ensuring that each training sample has the same weight and influence when optimizing the model, thereby achieving a stable training process.

[0081] Among them, the position estimation loss L 1 , which is expressed as:

[0082]

[0083] Strength estimation loss L 2 , which is expressed as:

[0084]

[0085] Wherein, the scale change operation 20lg(·) converts the sound source intensity to sound pressure level (SPL); p ref Represents the reference sound pressure, generally set to 2×10 -5 ; Represents an estimate of the strength of a sound source.

[0086] The data set used in this experiment is a random white noise signal set generated by the acoustic software Acoular. The schematic diagram of the 56-microphone spiral array used to collect signals is shown in the figure. Figure 3 The properties of this dataset are shown in Table 1. The experimental dataset contains 4200 signal data, of which 2400 data are used as training sets, 800 as validation sets, and 1000 as test sets. The coordination factor α in Acoustic-Net is set to 100.

[0087] Table 1. Properties of the experimental dataset

[0088]

[0089]

[0090] In order to verify the effectiveness of the model, the Acoustic-Net of the method of the present invention is compared with the Acoustic-AlexNet (Reiter A, Bell M AL. A machine learning approach to identifying point source locations in photoacoustic data [C] / / Photons Plus Ultrasound: Imaging and Sensing 2017: volume 10064. SPIE, 2017: 504 509) and Acoustic-ResNet (Kujawski A, Herold G, Sarradj E. Adeep learning method for grid The free localization and quantification of sound sources [J]. The Journal of the Acoustical Society of America, 2019, 146 (3): EL225-EL231) is used for comparison. The mean error of the position estimation (Mean-Distance Error, MDE) and the mean absolute percentage error of the SPL estimation (Mean Absolute Percentage Errors, MAPE) are used to evaluate the algorithm performance, which are defined as follows:

[0091]

[0092] The performance index values ​​of different models for position and SPL estimation are compared, as shown in Table 2.

[0093] Table 2. MDE and MAPE index values ​​of different models

[0094]

[0095]

[0096] As can be seen from Table 2, the errors of the position and sound pressure intensity estimated by the Acoustic-Net of the method of the present invention are smaller than those of other models, and the required inference time is shorter, which has higher operating efficiency.

[0097] The acoustic imaging method based on the multi-task learning network of the present invention combines acoustic signal processing with convolutional neural network. Based on the irrelevance of sound source position and intensity, two independent network branches are proposed based on the multi-task learning framework, so that the model focuses on the representation of different features in a biased manner, thereby better learning position and intensity features and improving the generalization ability of the network.

[0098] The acoustic imaging method of the present invention is based on a multi-task learning network. In order to save hardware resources and reduce computing loss, a position prediction branch based on the MobileOne backbone network and an intensity prediction branch based on a shallow one-dimensional convolution process are designed. These two lightweight and easy-to-infer network architectures are easy to integrate into the hardware system and have practical engineering application value.

[0099] The acoustic imaging method of the present invention is based on a multi-task learning network. The acoustic network Acoustic-Net can obtain more accurate position and intensity estimation and has more efficient operation efficiency, demonstrating the great potential of the model in real-time acoustic imaging applications.

[0100] The above is an exemplary description of the invention. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as such non-substantial improvements are made by adopting the method concept and technical solution of the present invention, or the concept and technical solution of the invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. An acoustic imaging method based on a multi-task learning network, characterized in that: The following steps are involved: S1. The acoustic localization task and the sound source quantification task are regarded as feature regression problems, and the features of the original signal are directly extracted as network input; S2, a position prediction branch based on time-frequency features and an intensity prediction branch based on amplitude features designed based on a multi-task learning framework to accurately model the localization and quantization tasks, that is, using time-frequency features to learn phase changes and using amplitude features to predict the intensity of the sound source; S3. Combining the position prediction branch and the intensity prediction branch, we design a concise and effective overall model, Acoustic-Net. We use the original audio data to learn the characteristics of amplitude and phase changes, and fully verify the effectiveness of the network, thereby completing the entire process of acoustic imaging based on a multi-task learning network.

2. The acoustic imaging method based on a multi-task learning network according to claim 1, characterized in that: In step S2, the position prediction branch of the time-frequency feature is specifically operated as follows: ① First, the original acoustic signal is short-time Fourier transformed to convert it into a time-frequency distribution diagram to utilize the rich phase information in the time-frequency representation; Assuming the original signal is s(t), STFT converts it to the time-frequency domain through window function and Fourier transform, that is: Where s(n) is the input signal; g(nm) is the window function, which is a smooth time window used to intercept a part of the input signal at a specific time position; m is the center position of the sliding window; n is the time index, which represents the sampling point of the signal s(n); α is the frequency variable, which represents the angular frequency in the frequency domain; e -jnω is a complex exponential used to convert signals from the time domain to the frequency domain; X(m,ω) is the result of STFT, which is the representation of the input signal s(n) at time n and frequency ω; Then, the time-frequency plot of the input signal is defined as: S(m,ω)=|X(m,ω)| 2 ; ②. MobileOne is used as the backbone network. By combining reparameterization technology and deep separable convolution, structural simplification and improved computational efficiency are achieved in the training and inference stages.

3. The acoustic imaging method based on a multi-task learning network according to claim 2, characterized in that: During the training phase, the depthwise separable convolution architecture of MobileOne consists of two parts, namely, the depthwise convolution module in the upper part and the pointwise convolution module in the lower part; The deep convolution module consists of three branches: the leftmost branch is a 1×1 convolution; the middle branch is a re-parameterized 3×3 convolution, i.e., k 3×3 convolutions; the right part is a skip connection containing a BN layer; The point convolution module consists of two branches: the left branch is a re-parameterized 1×1 convolution, which is composed of k 1×1 convolutions; the right branch is a skip connection containing a BN layer.

4. The acoustic imaging method based on a multi-task learning network according to claim 3, characterized in that: For the reparameterization of the deep convolution module, the specific process is as follows: Assuming that the input channel of the depth convolution part is d, the number of groups of the depth convolution is also d; For the 1×1 convolution branch on the left, the 1×1 convolution and BN are merged into a 1×1 convolution with d grouping numbers using the reparameterization method based on the merging of the convolution layer and the BN layer. This convolution is recorded as A. For the middle reparameterized 3×3 convolution branch, for each individual 3×3 convolution, the 3×3 convolution and its BN are merged into a 3×3 convolution with a group number of d using reparameterization method 1 (merging of convolution layer and BN layer), thereby obtaining k 3×3 convolutions; then, reparameterization method 2 (multi-branch addition) is used to merge these k 3×3 convolutions into one convolution, which is recorded as B; For the skip connection on the right, we first add a convolution before the BN layer. The parameters of the convolution are set manually. It needs to satisfy that for any input x, the output after the convolution is also x. Such a convolution is called Dirac delta. With the Dirac delta convolution, we can use the reparameterization method based on the merging of the convolution layer and the BN layer to merge it with the BN layer to obtain a new 3×3 convolution, which is denoted as C. Use multiple parallel branches to perform convolution operations, each branch contains different operations, so that the model can learn more diverse feature representations during training and better adapt to changes in input data; The final merging process merges the convolution kernel and bias of each branch into a new equivalent convolution kernel and bias.

5. The acoustic imaging method based on a multi-task learning network according to claim 2, characterized in that: During the inference phase, MobileOne converts all multi-branch structures into a standard convolution operation to simplify calculations and improve inference efficiency, so that the network can fully utilize the computing acceleration of the hardware optimization library during inference, thereby saving computing resources on the terminal device.

6. The acoustic imaging method based on a multi-task learning network according to claim 1, characterized in that: In step S2, the intensity prediction branch based on amplitude features specifically operates as follows: The original acoustic signal is used as the network input, and a shallow one-dimensional convolution process is used to effectively estimate the sound source intensity to make full use of the amplitude information in the original signal; the shallow one-dimensional convolution process includes 3 one-dimensional convolution layers and 4 fully connected layers.

7. The acoustic imaging method based on a multi-task learning network according to claim 1, characterized in that: In step S3, the Acoustic-Net model has the following specific working steps: Given a training data pair (s(t),x gt ,y gt ,p gt ), s(t) represents the input signal, (x gt ,y gt ) represents the target position coordinates, p gt Represents the target sound intensity; first, the acoustic signal is subjected to STFT, which is converted into the corresponding time-frequency image and input into the position prediction branch to obtain the prediction of the sound source position At the same time, the original signal is used as the input of the intensity prediction branch to obtain the intensity estimation of the sound source.

8. The acoustic imaging method based on a multi-task learning network according to claim 1, characterized in that: In step S3, the network loss function of the Acoustic-Net model includes prediction deviations of position and intensity; Let the training set be the work, and the network loss function can be expressed as: Where α is the coordination factor, L1 is the position estimation loss, and L2 is the intensity estimation loss; Γ represents a set of training data sets, each element of which contains the information of a training sample, including the acoustic signal s(t), the corresponding target position coordinates (x gt ,y gt ) and target sound intensity p gt ; It means averaging the losses of all training samples in the set Γ; The position estimation loss L1 is expressed as: The strength estimation loss L2 is expressed as: Wherein, the scale change operation 20lg(·) converts the sound source intensity to sound pressure level; p ref Represents the reference sound pressure, generally set to 2×10 -5 ; Represents an estimate of the strength of a sound source.

9. The acoustic imaging method based on a multi-task learning network according to claim 1, characterized in that: In step S3, the validity of the network is fully verified by: the mean error MDE of the position estimation and the mean absolute percentage error MAPE of the SPL estimation; The calculation formulas of MDE and MAPE are:

Images