Construction and recognition method for sound source recognition network of vehicles in highway tunnel

Abstract

The invention discloses a construction and recognition method for a sound source recognition network of vehicles in a highway tunnel, and the method comprises the steps: setting a distributed microphone array in the tunnel, and collecting a vehicle audio signal; a MobileNetV3 model based on the Mel frequency spectrum is constructed; the Mel spectrum feature extraction module is used for taking the Mel spectrum feature extracted by the Mel spectrum feature extraction module as input, taking a vehicle audio signal type as output, and training a MobileNetV3 model based on the Mel spectrum to obtain a sound recognition model; constructing a sound source localization model; according to the method, the MobileNetV3 model based on the Mel frequency spectrum is adopted for sound recognition, the sound source positioning model is constructed in a combined mode for sound source positioning, the MobileNetV3 model and the sound source positioning model work cooperatively, the feature description capability of abnormal sound of a vehicle is enhanced, the influence of tunnel echo and environment noise on recognition is effectively reduced, accidents in the tunnel are found in time and positioned accurately, and the method is suitable for popularization and application. The recognition and positioning accuracy of the vehicle sound source in the complex tunnel environment is improved, and the technical problem that in the prior art, the recognition precision of the accident sound in the tunnel is not high is solved.

Description

Technical Field

[0001] The present invention belongs to the fields of acoustics and artificial intelligence, and relates to an accident detection method, in particular to a method for constructing a sound source recognition network and identifying vehicles in a highway tunnel. Background Art

[0002] With the continuous development of transportation infrastructure and the widespread application of intelligent transportation systems, highway tunnels, as important transportation facilities, face increasing challenges in their safe operation. Timely detection and accurate location of accidents within tunnels are crucial for reducing secondary accidents and improving rescue efficiency. Traditional tunnel monitoring systems primarily rely on video surveillance technology, but in long tunnels, curved tunnels, or in poorly lit areas, single-use visual monitoring has significant limitations, making it incapable of timely detection and accurate location of accidents.

[0003] Traditional methods for locating vehicles in tunnels primarily rely on fixed sensor networks or video image analysis, such as infrared sensors and microwave radar. While these methods can provide some positioning information, they suffer from high installation costs, difficult maintenance, and limited positioning accuracy. Especially in the event of an accident, these methods often fail to quickly and accurately determine the accident location, delaying rescue efforts. Furthermore, traditional accident detection methods often rely on changes in traffic flow parameters and the identification of anomalies in video images, making real-time and accurate accident identification difficult in complex tunnel environments.

[0004] In the field of sound recognition, traditional sound feature extraction methods such as MFCC and spectral features are widely used. However, these methods perform poorly in the unique acoustic environment of tunnels and are unable to effectively distinguish between the sounds of normal vehicles and the abnormal sounds associated with accidents. While deep learning methods such as CNNs and RNNs have achieved significant results in sound recognition, their direct application to vehicle sound recognition and accident detection in tunnels still faces challenges such as insufficient data, difficulty in feature extraction, and interference from ambient noise.

[0005] To address the energy waste caused by the lack of dynamic control in tunnel lighting systems, research has employed vehicle localization methods based on GCC-Phat delay estimation and the Levenberg-Marquardt algorithm to solve the TDOA-AOA model. To promptly detect vehicle anomalies, a sound recognition method based on a modified MobileNetV3 model using Mel spectrum has been developed for real-time monitoring of vehicles traveling in tunnels. While these methods achieve a certain degree of vehicle localization and anomaly identification, they still suffer from low positioning accuracy and limited recognition accuracy, particularly in acoustically complex environments like tunnels.

[0006] Therefore, the present invention starts from the analysis of tunnel acoustic characteristics and vehicle sound features, combines deep learning technology, and proposes a new method for accident detection and positioning in tunnels based on sound source identification and positioning. The method aims to solve the problems existing in the existing technology, improve the accuracy of accident detection in tunnels and the precision of positioning, and provide strong support for safe tunnel operation. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for constructing a sound source identification network for vehicles in a highway tunnel, so as to solve the technical problem of low accuracy in accident sound identification and positioning in tunnels in the existing technology.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A method for constructing a sound source identification network for vehicles in a highway tunnel is performed according to the following steps:

[0010] Step 1: Set up a distributed microphone array in the tunnel to collect vehicle audio signals;

[0011] Step 2: Build a MobileNetV3 model based on Mel spectrum;

[0012] The Mel spectrum-based MobileNetV3 model includes a Mel spectrum feature extraction module and an improved MobileNetV3 classification module arranged in parallel;

[0013] The Mel spectrum feature extraction module is used to extract the Mel spectrum features of the vehicle audio signal;

[0014] The improved MobileNetV3 classification module is used to analyze the Mel features of vehicle audio signals and identify abnormal conditions;

[0015] Step 3: Using the Mel spectrum features extracted by the Mel spectrum feature extraction module as input and the vehicle audio signal type as output, the MobileNetV3 model based on Mel spectrum is trained to obtain a sound recognition model.

[0016] Step 4: construct a sound source localization model;

[0017] The sound source localization model includes a time delay estimation submodule, a TDOA-AOA model construction submodule and a parameter optimization submodule which are arranged in sequence;

[0018] The delay estimation submodule is used to calculate the sound delay information between each microphone in the microphone array based on the GCC-Phat algorithm;

[0019] The TDOA-AOA model building submodule is used to establish a mathematical model of the sound source position based on the sound delay information, and convert the sound delay information into spatial position information;

[0020] The parameter optimization submodule is used to perform nonlinear least squares optimization on the TDOA-AOA model construction submodule using the Levenberg-Marquardt algorithm to obtain the spatial position coordinates of the vehicle.

[0021] The present invention includes the following technical features:

[0022] The improved MobileNetV3 classification module includes a 3×3 convolutional layer, 11 Bneck modules, a global average pooling layer and a Hard-Swish activation function arranged in sequence;

[0023] Each of the Bneck modules includes a 1×1 dilated convolution, a 3×3 depth-separable convolution, a pooling operation, and a fully connected layer, which are connected in sequence.

[0024] The specific process of the sound source localization model is as follows:

[0025] Q 1, according to the microphone array layout information, obtain the spatial coordinates of each microphone (x i ,y i );

[0026] Q2, each microphone collects the vehicle audio signal and obtains the time t when the vehicle audio signal reaches each microphone i ;

[0027] Q 3, select microphone MIR1 as the reference microphone and calculate the TDOA measurement value Δdis i,1 and AOA measurement value α j ;

[0028]

[0029] e i1 =e i -e1=c(τ i -τ1)

[0030]

[0031] ΔDis=[Δdis 2,1 ,Δdis 3,1 ,Δdis 4,1 ,…,Δdis G,1 ] T

[0032] Dis=[dis1,dis2,…,dis G ]T

[0033] Dis1=[dis1,dis1,...,dis1] T

[0034] in:

[0035] x i and y i The coordinates of the microphone;

[0036] x1 and y1 coordinates of the reference microphone MIR1;

[0037] c is the speed of sound in tunnel air;

[0038] τ i is the noise error when the i-th microphone collects the vehicle audio signal;

[0039] t i is the time it takes for the sound wave to reach the i-th microphone from the vehicle; in the actual environment, t i Contains noise error τ that follows a zero-mean Gaussian distribution i ; i=1,2,…,G,

[0040] e i,1 The error in TDOA measurement caused by noise in the tunnel;

[0041] e i and e1 is the time measurement value τ i and τ1 cause the obedience to mean 0 and variance σ 2 Normally distributed error;

[0042] e α is the error in measurement;

[0043] Q4, TDOA measurement value Δdis obtained based on Q3 i,1 and AOA measurement value α j , establish the TDOA-AOA joint maximum likelihood estimation equation, and use the LM algorithm to solve the target position coordinates, specifically using the following formula;

[0044]

[0045] σ is the standard deviation of the measurement noise or the standard deviation of the estimation error.

[0046] A method for detecting a sound source of a vehicle in a highway tunnel comprises the following steps:

[0047] Step 1: collecting audio signals from moving vehicles and preprocessing them;

[0048] Step 2: Input the pre-processed audio signal of the moving vehicle obtained in Step 1 into the MobileNetV3 model based on Mel spectrum constructed in the method for constructing a sound source identification network for vehicles in highway tunnels, obtain a sound recognition result, and determine whether the sound recognition result contains any abnormality. If so, proceed to Step 3; if not, terminate the process.

[0049] The sound recognition results include the driving sound, tire blowout sound, scraping sound and collision sound during normal driving of the vehicle;

[0050] Step three: input the preprocessed audio signal of the moving vehicle obtained in step one into the sound source localization model constructed in the sound source identification network construction method for vehicles in the highway tunnel to obtain the spatial coordinates of the moving vehicle.

[0051] Compared with the prior art, the present invention has the following beneficial technical effects:

[0052] (I) The present invention adopts the MobileNetV3 model based on Mel spectrum for sound recognition, and combines it with the construction of a sound source localization model for sound source localization. The two work together to significantly enhance the feature description capability of abnormal vehicle sounds, effectively reduce the impact of tunnel echoes and environmental noise on recognition, and can effectively identify abnormal vehicle sounds, realize the timely discovery and precise positioning of accidents in tunnels, improve the recognition and positioning accuracy of vehicle sound sources in complex tunnel environments, and solve the technical problem of low accuracy in the recognition of accident sounds in tunnels in the existing technology. At the same time, it also provides reliable technical support for intelligent control of tunnel lighting and emergency response, effectively prevents the occurrence of secondary accidents, and optimizes energy utilization efficiency.

[0053] (II) The present invention collects diverse sound samples in different tunnel environments and vehicle driving conditions. At the same time, the collected sound signals are denoised and enhanced to simulate the acoustic characteristics of different locations in the tunnel. The existing real accident sound data and simulated accident sounds are combined to generate comprehensive acoustic features in an improved deep learning network, thereby increasing the diversity and representativeness of the samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is the overall framework diagram of the present invention.

[0055] Figure 2 This is an example of a flow chart for extracting Mel spectrum features in the embodiment.

[0056] Figure 3 1 is an example diagram of a Mel filter in an embodiment.

[0057] Figure 4 It is an example of the Mel feature spectrum image extracted in the embodiment.

[0058] Figure 5 2 is a structural diagram of the MobileNetV3 model based on Mel spectrum in the embodiment.

[0059] Figure 6 2 is a schematic diagram of a sound source localization model in an embodiment.

[0060] Figure 7 3. The figure shows a comparison of the accuracy, loss value and confusion matrix results of the present invention and the original network on the Mel feature map in the embodiment, wherein (a) is the training result diagram of the original MobileNetV3 network model, (b) is the confusion matrix diagram of the original MobileNetV3 network model, (c) is the training result diagram of the improved MobileNetV3 network model of the present invention, and (d) is the confusion matrix diagram of the improved MobileNetV3 network model of the present invention.

[0061] Figure 8 3. The figures are line graphs of accuracy and loss values ​​and confusion matrices of the present invention and the existing classification recognition network on the Mel feature graph in the embodiment, wherein (a) is a graph showing changes in accuracy of the MobileNetV3 network model and other comparison models on the training set, (b) a graph showing changes in loss values ​​of the MobileNetV3 network model and other comparison models on the training set, (c) a graph showing changes in accuracy of the MobileNetV3 network model and other comparison models on the validation set, and (d) a graph showing changes in loss values ​​of the MobileNetV3 network model and other comparison models on the validation set.

[0062] Figure 9 It is a bar chart comparing the accuracy, precision, and F1 score of the present invention and the existing classification and recognition network on the Mel feature graph in the embodiment.

[0063] Figure 10 3 is a comparison chart of the solution accuracy of the LM method used in the present invention and other solution methods on the sound source localization model in this embodiment.

[0064] Figure 11 This is a comparison diagram of the algorithm convergence of the LM method used in the present invention and other solution methods on the sound source localization model in this embodiment.

[0065] The specific contents of the present invention are further explained in detail below with reference to the embodiments. DETAILED DESCRIPTION

[0066] It should be noted that, unless otherwise specified, all components in the present invention are components known in the art.

[0067] Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent modifications made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0068] The present invention provides a method for constructing a sound source identification network for vehicles in a highway tunnel. The method is performed according to the following steps:

[0069] Step 1: Set up a distributed microphone array in the tunnel to collect vehicle audio signals;

[0070] Step 2: Build a MobileNetV3 model based on Mel spectrum;

[0071] The Mel spectrum-based MobileNetV3 model includes a parallel Mel spectrum feature extraction module and an improved MobileNetV3 classification module;

[0072] The Mel spectrum feature extraction module is used to extract the Mel spectrum features of the vehicle audio signal;

[0073] The improved MobileNetV3 classification module is used to analyze the Mel features of vehicle audio signals and identify abnormal conditions;

[0074] Step 3: Using the Mel spectrum features extracted by the Mel spectrum feature extraction module as input and the vehicle audio signal type as output, the MobileNetV3 model based on Mel spectrum is trained to obtain a sound recognition model.

[0075] Step 4: construct a sound source localization model;

[0076] The sound source localization model includes a time delay estimation submodule, a TDOA-AOA model construction submodule, and a parameter optimization submodule, which are set in sequence;

[0077] The delay estimation submodule is used to calculate the sound delay information between each microphone in the microphone array based on the GCC-Phat algorithm;

[0078] The TDOA-AOA model building submodule is used to establish a mathematical model of the sound source location based on the sound delay information and convert the sound delay information into spatial position information;

[0079] The parameter optimization submodule is used to perform nonlinear least squares optimization on the TDOA-AOA model construction submodule using the Levenberg-Marquardt algorithm to obtain the spatial position coordinates of the vehicle.

[0080] In the above technical solution, the time delay estimation submodule improves the estimation accuracy in a noisy environment; the parameter optimization submodule improves the positioning accuracy and realizes the accurate positioning of abnormal events in the tunnel.

[0081] The internal operation of the Mel spectrum feature extraction module is as follows: after preprocessing the vehicle audio signal, a frame signal is obtained. The frame signal is Fourier transformed to obtain frequency domain information. The square of the absolute value of the result is the signal energy spectrum. The energy spectrum is then dot-producted using a Mel filter bank. The logarithmic form of the filter bank output is taken, which is the logarithmic Mel spectrum, also known as the Mel spectrum.

[0082] The Mel spectrum-based MobileNetV3 model is used for sound recognition, and combined with the construction of a sound source localization model for sound source localization. The two work together to significantly enhance the feature description capability of abnormal vehicle sounds and effectively reduce the impact of tunnel echoes and environmental noise on recognition. It can effectively identify abnormal vehicle sounds, realize the timely discovery and precise positioning of accidents in tunnels, improve the recognition and positioning accuracy of vehicle sound sources in complex tunnel environments, and solve the technical problem of low accuracy in accident sound recognition in tunnels in existing technologies. At the same time, it also provides reliable technical support for intelligent control of tunnel lighting and emergency response, effectively prevents the occurrence of secondary accidents, and optimizes energy utilization efficiency.

[0083] Specifically, the improved MobileNetV3 classification module includes a 3×3 convolutional layer, 11 Bneck modules, a global average pooling layer, and a Hard-Swish activation function.

[0084] Each Bneck module consists of a sequentially connected 1×1 dilated convolution, a 3×3 depthwise separable convolution, a pooling operation, and a fully connected layer.

[0085] In the above technical solution, see Figure 5 The input is a 224×224×3 color image. It first performs preliminary feature extraction through a 3×3 convolutional layer. Then, deep feature learning is performed through 11 Bneck modules (which integrate the SE attention mechanism and the NAM normalized attention module). A global average pooling layer is then used to compress the feature map to a 1×1 size. Finally, a 1×1 convolution is performed for final feature integration. Finally, a Hard-Swish activation function is used to output the classification result. The Bneck modules balance computational efficiency and feature expression capabilities. The entire network design takes both computational efficiency and feature expression capabilities into consideration.

[0086] By introducing the Mish activation function to replace the traditional ReLU and Hard-Swish functions, its good smoothness is used to better process complex audio signals containing echoes and reverberations in tunnel environments, and it performs well in time-frequency domain feature extraction. By introducing a depthwise separable dilated convolution module with a dilation rate of 2 in the depthwise convolution part of MobileNetV3-Small, the receptive field range is expanded without increasing the number of parameters and computational complexity, thereby enhancing the model's ability to capture acoustic features with a larger time domain span and effectively identifying audio signals in tunnel environments with echo and delay characteristics. By replacing the SE module with the NAM attention mechanism in the 8th to 11th layer bneck modules and combining channel attention and spatial attention designs, it not only retains the ability to enhance channel features, but also introduces information interaction in the spatial dimension, enabling the model to better capture the time-frequency dependencies and local features of audio signals.

[0087] Specifically, the specific process of the sound source localization model is as follows:

[0088] Q 1, according to the microphone array layout information, obtain the spatial coordinates of each microphone (x i ,y i );

[0089] Q2, each microphone collects the vehicle audio signal and obtains the time t when the vehicle audio signal reaches each microphone i ;

[0090] Q 3, select microphone MIR1 as the reference microphone and calculate the TDOA measurement value Δdis i,1 and AOA measurement value α j ;

[0091]

[0092] e i1 =e i -e1=c(τ i -τ1)

[0093]

[0094] ΔDis=[Δdis 2,1 ,Δdis 3,1 ,Δdis 4,1 ,…,Δdis G,1 ] T

[0095] Dis=[dis1,dis2,...,dis G ] T

[0096] Dis1=[dis1,dis1,…,dis1]T

[0097] in:

[0098] x i and y i The coordinates of the microphone;

[0099] x1 and y1 coordinates of the reference microphone MIR1;

[0100] c is the speed of sound in tunnel air;

[0101] τ i is the noise error when the i-th microphone collects the vehicle audio signal;

[0102] t i is the time it takes for the sound wave to reach the i-th microphone from the vehicle; in the actual environment, t i Contains noise error τ that follows a zero-mean Gaussian distribution i ; i=1,2,…,G,

[0103] e i,1 The error in TDOA measurement caused by noise in the tunnel;

[0104] e i and e1 is the time measurement value τ i and τ1 cause the obedience to mean 0 and variance σ 2 Normally distributed error;

[0105] e α is the error in measurement;

[0106] Q4, TDOA measurement value Δdis obtained based on Q3 i,1 and AOA measurement value α j , establish the TDOA-AOA joint maximum likelihood estimation equation, and use the LM algorithm to solve the target position coordinates, specifically using the following formula;

[0107]

[0108] σ is the standard deviation of the measurement noise or the standard deviation of the estimation error.

[0109] In the above technical solution, e α Generally speaking, the mean is 0 and the variance is σ 2 The normal distribution of α is arctan((yy j ) / (xx j )), with a variance of σ 2 Normal distribution;

[0110] The present invention also provides a method for detecting the sound source of a vehicle in a highway tunnel, comprising the following steps:

[0111] Step 1: collecting audio signals from moving vehicles and preprocessing them;

[0112] Step 2: Input the pre-processed audio signal of the moving vehicle obtained in Step 1 into the Mel spectrum-based MobileNetV3 model constructed in the method for constructing a sound source identification network for vehicles in highway tunnels to obtain a sound recognition result. Then, determine whether the sound recognition result contains any anomalies. If so, proceed to Step 3; otherwise, terminate the process.

[0113] The sound recognition results include the vehicle's normal driving sound, tire blowout sound, scraping sound, and collision sound;

[0114] Step three: input the pre-processed audio signal of the moving vehicle obtained in step one into the sound source localization model constructed in the sound source identification network construction method for vehicles in the highway tunnel to obtain the spatial coordinates of the moving vehicle.

[0115] Actual measurement example 1:

[0116] Following the above technical solution, this test example presents a MobileNetV3 model based on Mel spectrum to verify the effectiveness of the Mish activation function and attention mechanism in tunnel sound event recognition, and compares them with commonly used sound event recognition networks: original MobileNetV3, ResNet50, VGG16, GoogLeNet, MobileNetV2, ShuffleNetV2, VGGish and YAMNet.

[0117] (1) Experimental environment

[0118] The system used in the experiment is Windows 10 X64 operating system, and the model training framework is PyTorch 2.3.1 deep learning framework. The computer processor is 11th Gen Intel(R) Core(TM) i5-1135G7@

[0119] 2.40GHz 2.42GHz, 16GB of RAM, and Python interpreter version 3.10.5.

[0120] (2) Experimental process

[0121] In the present invention, the image input dimension is standardized to 224×224 pixels, and all 1000 samples are used to form the data set; the batch size is set to 64; the learning rate adopts a small value of 0.0001 to ensure the stability of the optimization process; the number of training iterations is set to 50; and the loss function adopts the cross-entropy loss function.

[0122] (3) Experimental analysis

[0123] Figure 7 and Figure 8 and Figure 9 The results of various methods on three evaluation metrics, accuracy, precision, and F1 score, are presented. The proposed method performs best among lightweight networks, achieving an F1 score of 94.1%, only 1.8 percentage points lower than ResNet50, while reducing the number of model parameters and computational complexity by over 85%. This demonstrates that the proposed method's overall performance outperforms other lightweight networks, demonstrating superior results for identifying abnormal vehicle sounds in the complex acoustic environment of tunnels. It also validates the effectiveness of the Mish activation function and NAM attention mechanism in sound event recognition.

[0124] Actual measurement example 2:

[0125] Following the above technical solution, this experimental example presents a sound source localization model based on GCC-Phat time delay estimation and TDOA-AOA localization, solved using the Levenberg-Marquardt (LM) algorithm. To compare the effectiveness of the LM method for localizing the sound sources of vehicle accidents in tunnels, it was compared with the commonly used solutions: the Taylor method, the Chan algorithm, the Newton method, the Taylor-Chan algorithm, the Gauss-Newton method, and the weighted least squares method.

[0126] (1) Experimental environment Experimental simulation environment: Matlab2024b, Windows10 system.

[0127] (2) Experiment on the convergence performance of the sound source localization model solution algorithm; Common sound source localization model solution methods include: Taylor method, Chan algorithm, Newton method, Taylor-Chan algorithm, Gauss-Newton method, weighted least squares method, Levenberg-Marquardt algorithm, etc. In this positioning model, the positioning effects of different algorithms are as follows: Figure 10 As shown, the algorithm convergence performance is as follows Figure 11 As shown. Figure 10As can be seen, the LM algorithm exhibits the best performance, with an RMSE of 0.2684m. Its estimated points are most closely clustered around the true location (black asterisk), demonstrating good accuracy and stability. This is due to their ability to balance convergence speed and stability during the iteration process, effectively addressing the nonlinearity of the sound source localization model. The Gauss-Newton method and weighted least squares method are second only to the LM algorithm. The Taylor method and Newton method have errors of approximately 0.3889m and 0.421m, respectively, and their estimated point distribution is slightly more dispersed than the Gauss-Newton and LM algorithms. The Taylor-Chan algorithm (pink dots) exhibits a significant bias toward the lower left, forming a cluster that deviates significantly from the true location. This is due to the Chan algorithm's large initial estimation error, which the Taylor iterations are unable to correct. The Chan algorithm is barely visible in the figure, indicating that it is not suitable for this localization model.

[0128] Of the seven algorithms, the LM algorithm performed best, with its fitness value eventually stabilizing to approximately 0.001. The Gauss-Newton and weighted least squares methods also performed well, quickly converging to lower values. The Taylor and Newton methods performed moderately well, stabilizing around 0.01. The Taylor-Chan algorithm converged more slowly, requiring 45 iterations to achieve significant improvement. The Chan algorithm, however, was limited in accuracy due to limitations in its initial reference point selection method.

[0129] In summary, the LM algorithm is the best choice for hybrid positioning applications. It not only provides high accuracy but also maintains the reliability and stability of the algorithm. This shows that the proposed method can effectively improve the localization accuracy of vehicle accident sound sources in tunnels.

Claims

1. A method for constructing a sound source identification network for vehicles in a highway tunnel, characterized in that: The method proceeds as follows: Step 1: Set up a distributed microphone array in the tunnel to collect vehicle audio signals; Step 2: Build a MobileNetV3 model based on Mel spectrum; The Mel spectrum-based MobileNetV3 model includes a Mel spectrum feature extraction module and an improved MobileNetV3 classification module arranged in parallel; The Mel spectrum feature extraction module is used to extract the Mel spectrum features of the vehicle audio signal; The improved MobileNetV3 classification module is used to analyze the Mel features of vehicle audio signals and identify abnormal conditions; Step 3: Using the Mel spectrum features extracted by the Mel spectrum feature extraction module as input and the vehicle audio signal type as output, the MobileNetV3 model based on Mel spectrum is trained to obtain a sound recognition model. Step 4: construct a sound source localization model; The sound source localization model includes a time delay estimation submodule, a TDOA-AOA model construction submodule and a parameter optimization submodule which are arranged in sequence; The delay estimation submodule is used to calculate the sound delay information between each microphone in the microphone array based on the GCC-Phat algorithm; The TDOA-AOA model building submodule is used to establish a mathematical model of the sound source position based on the sound delay information, and convert the sound delay information into spatial position information; The parameter optimization submodule is used to perform nonlinear least squares optimization on the TDOA-AOA model construction submodule using the Levenberg-Marquardt algorithm to obtain the spatial position coordinates of the vehicle.

2. The method for constructing a sound source identification network for vehicles in a highway tunnel according to claim 1, characterized in that: The improved MobileNetV3 classification module includes a 3×3 convolutional layer, 11 Bneck modules, a global average pooling layer and a Hard-Swish activation function arranged in sequence; Each of the Bneck modules includes a 1×1 dilated convolution, a 3×3 depth-separable convolution, a pooling operation, and a fully connected layer, which are connected in sequence.

3. The method for constructing a sound source identification network for vehicles in a highway tunnel according to claim 1, characterized in that: The specific process of the sound source localization model is as follows: Q 1, according to the microphone array layout information, obtain the spatial coordinates of each microphone (x i ,y i ); Q2, each microphone collects the vehicle audio signal and obtains the time t when the vehicle audio signal reaches each microphone i ; Q 3, select microphone MIR1 as the reference microphone and calculate the TDOA measurement value Δdis i,1 and AOA measurement value α j ; by i,1 =e i −e1=c(τ i −τ1) ΔDis=[Δdis 2,1 ,Δdis 3,1 ,Δdis 4,1 ,...,Δdis G,1 ] T Dis=[dis1,dis2,...,dis G ] T Dis1=[dis1,dis1,...,dis1] T in: x i and y i The coordinates of the microphone; x1 and y1 coordinates of the reference microphone MIR1; c is the speed of sound in tunnel air; τ i is the noise error when the i-th microphone collects the vehicle audio signal; t i is the time it takes for the sound wave to reach the i-th microphone from the vehicle; in the actual environment, t i Contains noise error τ that follows a zero-mean Gaussian distribution i ; i=1,2,…,G, e i,1 The error in TDOA measurement caused by noise in the tunnel; e i and e1 is the time measurement value τ i and τ1 cause the obedience to mean 0 and variance σ 2 Normally distributed error; e α is the error in measurement; Q4, TDOA measurement value Δdis obtained based on Q3 i,1 and AOA measurement value α j , establish the TDOA-AOA joint maximum likelihood estimation equation, and use the LM algorithm to solve the target position coordinates, specifically using the following formula; σ is the standard deviation of the measurement noise or the standard deviation of the estimation error.

4. A method for detecting the sound source of a vehicle in a highway tunnel, characterized in that: The following steps are involved: Step 1: collecting audio signals from moving vehicles and preprocessing them; Step 2: Input the preprocessed audio signal of the moving vehicle obtained in Step 1 into the Mel spectrum-based MobileNetV3 model constructed in the method for constructing a sound source identification network for vehicles in a highway tunnel according to any one of Claims 1 to 3, obtain a sound recognition result, and determine whether the sound recognition result contains any abnormality. If so, proceed to Step 3; if not, terminate the process. The sound recognition results include the driving sound, tire blowout sound, scraping sound and collision sound during normal driving of the vehicle; Step three: input the preprocessed audio signal of the moving vehicle obtained in step one into the sound source localization model constructed in the sound source identification network construction method for vehicles in a highway tunnel according to any one of claims 1 to 3 to obtain the spatial coordinates of the moving vehicle.

Images