A Noise Area Detection and Identification Method

Abstract

The present invention relates to the field of noise detection technology, and in particular to a noise area detection and identification method, comprising: collecting audio signals of construction waste disposal equipment when it is running in different environments, and performing short-time Fourier transform to obtain the power spectrum and phase spectrum of the audio signal; establishing a multi-level audio feature extractor to extract the multi-level audio features of the construction waste disposal equipment in the source domain and the target domain; establishing a noise discriminator to discriminate whether the audio of sound sources in various directions belongs to noise; combining the multi-level audio feature extractor, the sound source position identifier and the noise discriminator into a noise area detection model, and training the model through an optimization algorithm; inputting a target domain test sample set into the trained noise area detection model to predict the noise direction of the construction waste disposal equipment online. The present invention solves the problem that the noise source positioning method in the prior art has poor environmental adaptability and is difficult to accurately position multiple noise sources at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of noise detection, and in particular to a noise area detection and identification method. Background Art

[0002] In the construction waste management environment, due to the large number of construction waste disposal equipment, noise problems are often more complicated and may come from the operation of various equipment, such as crushers, loaders, excavators and other engineering equipment; these equipment will generate noise of different frequencies and intensities during operation, which in turn affects the surrounding environment; in order to effectively resolve disputes caused by noise, such as complaints within the enterprise, noise disputes between enterprises, or noise problems between enterprises and surrounding residential areas, it is particularly important to accurately measure and evaluate the source of noise; through detailed analysis and positioning of the noise source, the specific source of the noise and its impact on the surrounding environment can be clarified, and help to formulate corresponding noise control measures; therefore, it is very necessary to accurately locate the area where the noise is generated.

[0003] Traditional noise source localization methods are based on signal models and signal processing techniques; although these methods have made significant progress in this field, they perform poorly and have poor noise localization accuracy in common and complex scenarios such as those with many noise sources and reverberation; in the past few years, the potential of data-driven deep learning technology in dealing with these difficult scenarios has attracted increasing attention; therefore, noise source localization methods based on deep neural networks have gradually been proposed in recent years; and according to existing research, noise source localization methods based on deep neural networks are superior to traditional noise source localization methods based on signal processing.

[0004] Although deep neural networks perform well in noise source positioning, in the actual construction waste management environment, various noises are mixed with the sounds of normal life, and the working environment and location of construction waste disposal equipment are constantly changing. The noise source positioning deep model trained in one working environment becomes less effective when applied to different working environments, and the accuracy of noise recognition and positioning will be significantly reduced. Therefore, it is necessary to provide a noise area detection and identification method with strong environmental adaptability and the ability to accurately locate multiple noise sources. Summary of the invention

[0005] In view of the shortcomings of the existing methods, the present invention solves the problems of poor environmental adaptability of the noise source positioning method in the prior art and difficulty in accurately positioning multiple noise sources at the same time.

[0006] The technical solution adopted by the present invention is: a noise area detection and identification method comprises the following steps:

[0007] Step 1: collect the audio signal of the construction waste disposal equipment when it is running, and perform short-time Fourier transform to obtain the power spectrum and phase spectrum of the audio signal;

[0008] As a preferred implementation of the present invention, the audio signal is collected by a microphone array.

[0009] Step 2: construct a multi-level audio feature extractor to extract multi-level audio features of construction waste disposal equipment in the source domain and the target domain;

[0010] The multi-level audio feature extractor comprises: a first convolution layer and a multi-level convolution block; wherein,

[0011] The convolution kernel size of the first convolutional layer is 7×7, the number of channels is 16, the stride is 2, and the BN and ReLu functions are used;

[0012] The multi-level convolution block uses three convolution branches; among them, the first branch has 1 convolution layer, 1×1 convolution kernel size, and 512 channels; the second branch has 2 convolution layers, 1×1 convolution kernel size, 64 channels, 3×3 convolution kernel size, and 512 channels; the third branch has 3 convolution layers, 1×1 convolution kernel size, 64 channels, 5×5 convolution kernel size, 128 channels, and 5×5 convolution kernel size, 512 channels;

[0013] The feature maps of each channel in the 512 channels obtained from each branch are globally averaged pooled; the feature vectors of the three branches are concatenated, and the output feature vector size is 1×1536.

[0014] Step 3: Establish a sound source position identifier to predict the sound source position of the audio;

[0015] As a preferred embodiment of the present invention, the sound source location identifier is composed of three fully connected layers and a Sigmoid activation function, and the number of neurons in the three fully connected layers is 1024, 1024, and 360 respectively; the output is a 360-dimensional predicted probability ={ g 1. g 2. … g i …、 g 360};in, g i Corresponding to the plane 360 ​​dimension i The probability of a sound source appearing in a certain direction.

[0016] Step 4: Establish a noise discriminator to determine whether the audio from various sound sources is noise;

[0017] As a preferred embodiment of the present invention, the noise discriminator consists of three fully connected layers and a Sigmoid activation function, the number of neurons in the three fully connected layers is 2048, 2048, and 360 respectively; the output is a 360-dimensional predicted probability ={ u 1. u 2. … u i …、 u 360};in, u i Corresponding to the plane 360 ​​dimension i The probability that the sound source in a direction is noise.

[0018] Step 5: Combine the multi-level audio feature extractor, the sound source location identifier and the noise discriminator into a noise area detection model, and train the model through an optimization algorithm;

[0019] As a preferred implementation of the present invention, step five specifically includes:

[0020] Step 51: The multi-level audio feature extractor and the sound source position identifier form a first feedforward neural network, and the multi-level audio feature extractor and the noise discriminator form a second feedforward neural network;

[0021] Step 52, the multi-level audio feature extractor is used to extract source domain features, the sound source position identifier is used to identify the sound source position of the feature and make the prediction result correctly fit the real audio orientation; the noise discriminator is used to identify whether the feature belongs to noise; and the loss of the prediction result of the sound source position identifier is minimized. L C and the loss of the noise discriminator prediction results L D To optimize the model, the formula is:

[0022] ;

[0023] In the formula, L C is the mean absolute error loss of the sound source location identifier prediction results, L D is the mean absolute error loss of the noise discriminator prediction results, represents the mean absolute error value, represents the absolute value, and are the predicted probability of the sound source location identifier and the actual sound source location, and In the i The predicted probability of the noise discriminator and the actual noise result in the dimension, is the weight parameter;

[0024] Step 53: performing covariance alignment on the source domain and target domain features extracted by the multi-level audio feature extractor;

[0025] As a preferred embodiment of the present invention, the formula of covariance alignment loss is:

[0026] ;

[0027] In the formula, represents the Frobenius norm, represents the covariance calculation, and Represent the feature matrices of the source domain and the target domain respectively, Represents the feature dimension.

[0028] Step 54: Combine the optimization objective function loss and the covariance alignment loss to obtain a total loss function, and use the total loss function to perform back propagation optimization;

[0029] As a preferred embodiment of the present invention, the formula of the total loss function is:

[0030]

[0031] In the formula, Weighing parameters for the loss function, L s To optimize the objective function loss, is the covariance alignment loss.

[0032] Step 6: Input the target domain test sample set into the trained noise area detection model to predict the noise direction of the construction waste disposal equipment online.

[0033] As a preferred implementation of the present invention, step six specifically includes:

[0034] By combining the prediction results of the sound source location identifier and the noise discriminator, a prediction threshold is set , will be above the prediction threshold The prediction results As the output of noise detection identification;

[0035] ;

[0036] In the formula, and Respectively expressed in i The prediction probability of the sound source location identifier and noise discriminator in the dimensional direction, For the iThe noise source in the dimensional direction, is the noise above the prediction threshold.

[0037] As a preferred embodiment of the present invention, a noise region detection and identification system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a noise region detection and identification method.

[0038] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and the computer program code implements a noise region detection and identification method when executed by a processor.

[0039] Beneficial effects of the present invention:

[0040] 1. Deep transfer learning technology was introduced to establish a noise area detection model for construction waste disposal equipment. A multi-level audio feature extractor was used to extract multi-level audio features containing rich information, and a sound source location identifier and a noise discriminator were constructed to simultaneously predict the sound source locations of multiple audios in 360-dimensional orientations and determine whether the audio is noise, overcoming the problem that traditional methods are difficult to locate multiple noise sources;

[0041] 2. Covariance alignment is introduced to transfer the noise recognition knowledge in one environment to the noise recognition in different environments, which improves the adaptability of the noise area detection model of construction waste disposal equipment in different environments;

[0042] 3. Able to accurately locate the noise sources of multiple construction waste disposal equipment at the same time;

[0043] 4. It has high accuracy and wide application range when used for noise detection and identification of construction waste disposal equipment in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of a noise area detection and identification method provided by an embodiment of the present invention;

[0045] Figure 2 It is a structural schematic diagram of a noise area detection and recognition training model provided by an embodiment of the present invention;

[0046] Figure 3 This is a method provided by an embodiment of the present invention. Figure 2 A schematic diagram showing the structure of the trained model being applied to the test sample set for verification;

[0047] Figure 4 This is a graph showing the detection performance verification results of the noise area detection model provided in an embodiment of the present invention on a test sample set. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0049] like Figure 1 As shown, a noise area detection and identification method includes the following steps:

[0050] Step 1: collect audio signals of construction waste disposal equipment in operation under different construction waste management environments, perform short-time Fourier transform, obtain the power spectrum and phase spectrum of the audio signals, and construct source domain and target domain sample sets;

[0051] Step 1 specifically includes:

[0052] A microphone array is used to collect audio signals from construction waste disposal equipment during operation, and the audio signals are divided into pieces of the same length, and then short-time Fourier transform is performed to obtain the power spectrum and phase spectrum of the audio signals while retaining the power and phase information; the power spectrum and phase spectrum are combined into two-channel samples as the input of the model. Different working times, changes in the position of construction waste disposal equipment, and different audio collection points result in different working environments. The source domain sample set and the target domain sample set are divided according to the different construction waste management environments. The target domain sample set is divided into a target domain training sample set and a target domain test sample set; the fully labeled source domain training sample set and the unlabeled target domain training sample set are used to train the noise area detection model. The target domain test sample set does not participate in the model training and is only used to verify the detection performance of the model.

[0053] As an example, a noise dataset in a construction waste management environment collected by a four-element cross microphone array is taken as an example; it includes noise data generated by the operation of construction waste disposal equipment and normal audio data in daily life; the experiment selects audio samples collected by 4 microphones in the four-element cross microphone array in the first working environment as the source domain training dataset; and 15,000 samples randomly selected from the samples collected in the second working environment are used as the target domain training sample set to train the model; finally, the remaining samples in the second working environment that did not participate in the model training are used as the target domain test sample set to verify the model.

[0054] The description of the dataset used is shown in Table 1:

[0055] Table 1 Sample sets of different noise sources in different working environments

[0056] Work Environment Total number of samples Sample with noise source 3 Sample with noise source 2 Samples with noise source 1 Normal noise-free sample 1 32658 6242 4649 6033 15734 2 27612 4651 5025 3986 13950

[0057] Among them, noise source 3, noise source 2, and noise source 1 represent three noise sources in the same scene.

[0058] Step 2: Establish a multi-level audio feature extractor to extract multi-level audio features of construction waste disposal equipment in the source domain and the target domain;

[0059] Step 2 specifically includes:

[0060] The multi-level audio feature extractor is used to map samples into a high-level feature space. The input is labeled source domain samples and unlabeled target domain samples, and the output is the high-level features of the audio samples.

[0061] The multi-level audio feature extractor consists of the first convolutional layer and multi-level convolutional blocks; the convolution kernel size of the first convolutional layer is 7×7, the number of channels is 16, the step size is 2, and the convolutional layer uses BN and ReLu functions;

[0062] The multi-level convolution block adopts three convolution branches, each branch adopts a different convolution kernel size and number of convolution layers; among them, the first branch has 1 convolution layer, 1×1 convolution kernel size, and 512 channels; the second branch has 2 convolution layers, the first layer has 1×1 convolution kernel size, 64 channels, the second layer has 3×3 convolution kernel size, and 512 channels; the third branch has 3 convolution layers, the first layer has 1×1 convolution kernel size, 64 channels, the second layer has 5×5 convolution kernel size, 128 channels, and the third layer has 5×5 convolution kernel size, and 512 channels.

[0063] The feature maps of each channel in the 512 channels obtained by each branch are globally averaged pooled, so that the size of the feature vector obtained by each branch is 1×512; then the feature vectors obtained by the three branches are concatenated, and finally the size of the feature vector output by the multi-level audio feature extractor is 1×1536.

[0064] It should be noted that the input samples are sample data collected by the four microphones in the microphone array. The signals collected by each microphone are transformed through short-time Fourier transform to obtain two-channel samples combining power spectrum and phase spectrum. Therefore, a group of samples collected by a microphone array is 8-channel samples.

[0065] Step 3: Establish a sound source position identifier to predict the sound source position of the audio;

[0066] Step three specifically includes:

[0067] The sound source location identifier consists of three fully connected layers and a Sigmoid activation function. The number of neurons in the three fully connected layers is 1024, 1024, and 360 respectively. Its input is the high-level features extracted by the multi-level audio feature extractor, and its output is a 360-dimensional prediction probability. ={ g 1. g 2. … gi …、 g 360}; g i Corresponding to the plane 360 ​​dimension i The probability of a sound source appearing in each direction corresponds to the probability of a sound source appearing in each direction of the 360-dimensional plane.

[0068] It should be noted that the output 360-dimensional orientation prediction probability may correspond to the sound source positions of multiple or single audio samples.

[0069] Step 4: Establish a noise discriminator to determine whether the audio from various sound sources is noise;

[0070] Step 4: Specifically include:

[0071] The noise discriminator consists of three fully connected layers and a Sigmoid activation function. The number of neurons in the three fully connected layers is 2048, 2048, and 360 respectively. Its input is the high-level features extracted by the multi-level audio feature extractor, and its output is a 360-dimensional prediction probability. ={ u 1. u 2. … u i …、 u 360};in, u i Corresponding to the plane 360 ​​dimension i The probability that the sound source in each direction is noise corresponds to the probability that the sound source in each direction of the 360-dimensional plane is noise.

[0072] Step 5: Combine the multi-level audio feature extractor, the sound source location identifier and the noise discriminator into a noise area detection model, and train the model through an optimization algorithm;

[0073] Step 5 specifically includes:

[0074] Step 51: The multi-level audio feature extractor and the sound source position identifier form a first feedforward neural network, and the multi-level audio feature extractor and the noise discriminator form a second feedforward neural network, and the two feedforward neural networks are respectively used to obtain the sound source position and to discriminate whether the audio at each sound source position is noise;

[0075] Step 52, the multi-level audio feature extractor is used to extract source domain features, the sound source position identifier is used to identify the sound source position of the feature and make the prediction result correctly fit the real audio orientation; the noise discriminator is used to identify whether the feature belongs to noise, and make the prediction result correctly fit the real noise discrimination result; thereby minimizing the loss of the prediction result of the sound source position identifier LC and the loss of the noise discriminator prediction results L D To optimize the model, the formula is:

[0076] ;

[0077] In the formula, L C is the mean absolute error loss of the sound source location identifier prediction results, L D is the mean absolute error loss of the noise discriminator prediction results, represents the mean absolute error loss function, represents the absolute value, and are the predicted probability of the sound source location identifier and the actual sound source location, and In the i The predicted probability of the noise discriminator and the actual noise result in the dimension, is a weight parameter whose value depends on its distance to the nearest sound source.

[0078] Step 53: In order to transfer the feature knowledge learned in the labeled source domain to the unlabeled target domain, the source domain and target domain features extracted by the multi-level audio feature extractor are covariance aligned, that is, the CORAL loss is used. To align the feature distribution of the source domain and the target domain, the formula for the covariance alignment loss is:

[0079] ;

[0080] In the formula, represents the Frobenius norm, represents the covariance calculation, and Represent the feature matrices of the source domain and the target domain respectively, Represents the feature dimension.

[0081] Step 54: Combine optimization objective function L s Loss and covariance alignment loss get the total loss function The model is back-propagated and optimized, supervised training is performed on the model through the labeled source domain, and the model is trained to learn transferable noise recognition knowledge by combining the labeled source domain and the unlabeled target domain:

[0082] ;

[0083] In the formula, For the loss function, the parameters are weighed. The value is 0.1. L s To optimize the objective function loss, is the covariance alignment loss.

[0084] In this embodiment, the number of iterative training of the noise area detection model is 500, and the adaptive moment estimation algorithm is used as the optimization algorithm. Figure 2 The training model shown is back-propagation optimized, the learning rate is set to 0.001, and the training stops when the number of iterative training reaches the set value.

[0085] Step 6: Input the target domain test sample set into the trained noise area detection model to predict the noise direction of the construction waste disposal equipment online.

[0086] Step 6 specifically includes:

[0087] The trained noise region detection model includes a multi-level audio feature extractor, a sound source location identifier, and a noise discriminator. Figure 3 shown.

[0088] The final prediction result of the online prediction of the noise direction of construction waste disposal equipment is obtained by combining the prediction results of the sound source location identifier and the noise discriminator, and setting a prediction threshold Its value is 0.5, which will be above the prediction threshold The prediction results As the output of noise detection recognition:

[0089] ;

[0090] In the formula, and Respectively expressed in i The prediction probability of the sound source location identifier and noise discriminator in the dimensional direction, For the i The noise source in the dimensional direction, is the noise above the prediction threshold.

[0091] The present invention uses a microphone array to collect audio signals of construction waste disposal equipment when it is running in different environments and performs short-time Fourier transform to obtain power spectrum and phase spectrum of the audio signals, and constructs source domain and target domain sample sets; establishes a multi-level audio feature extractor to extract multi-level audio features of construction waste disposal equipment in the source domain and the target domain; establishes a sound source position identifier to predict the sound source position of the audio; establishes a noise discriminator to discriminate whether the audio of sound sources in various directions belongs to noise; combines the multi-level audio feature extractor, the sound source position identifier and the noise discriminator into a noise area detection model, and trains the model through an optimization algorithm; inputs a target domain test sample set into the trained noise area detection model, and predicts the noise direction of the construction waste disposal equipment online.

[0092] It should be noted that if the prediction error of the noise source orientation is within 5°, the prediction result is considered correct. The detection results are shown in Table 2;

[0093] Table 2 Test results

[0094] Total number of samples Sample with noise source 3 Sample with noise source 2 Samples with noise source 1 Normal noise-free sample Number of test samples 27612 4651 5025 3986 13950 Accurate number of samples tested 25489 4052 4632 3748 13057 Detection accuracy 92.31% 87.12% 92.18% 94.03% 93.60%

[0095] The noise area detection accuracy results of the method of the present invention for the test sample set under the second working environment are shown in Table 2; it can be seen that the overall detection accuracy of the method of the present invention for various types of audio has reached 92.31%, and it can accurately distinguish normal noise-free audio samples from noisy audio samples; in the experimental test, due to the absence of interference from multiple equipment noises, the detection accuracy of the sample with noise source 1 is the highest, reaching 94.03%; the detection accuracy of the sample with noise source 2 is 92.18%, which is slightly lower than the detection accuracy of the sample with noise source 1, and the detection accuracy of the sample with noise source 3 is lower than that of other types of samples. This is due to the mutual interference between the noises of multiple devices, which has a certain impact on the detection results, but its accuracy has also reached 87.12%.

[0096] The noise region detection performance of the method of the present invention on the test sample set is as follows: Figure 4 As shown in the figure, the AUC value in the ROC curve is an indicator used to measure the classification ability of the model. Its value range is between 0 and 1. The larger the value, the better the classification performance of the model. Figure 4 It can be seen that the method of the present invention has good performance in noise detection, and the AUC value reaches 0.959; the noise area detection results of the test sample set under the second working environment reflect that the method of the present invention can accurately detect multiple noise sources of construction waste disposal equipment at the same time and adapt to different working environments.

[0097] In summary, the method of the present invention establishes a noise area detection model by introducing deep neural networks and transfer learning technology. It can extract multi-level audio features containing rich information and accurately identify the sound source position of the audio and whether it is noise. The model has high detection accuracy for multiple noise sources of construction waste disposal equipment. By aligning the covariance distribution of noise data set features under different working environments, the adaptability of the noise area detection model is significantly improved, and it can be effectively applied to the detection and identification of multiple noise sources of construction waste disposal equipment under different working environments.

[0098] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A noise area detection and identification method, characterized in that: The following steps are involved: Step 1: collect the audio signal of the construction waste disposal equipment when it is running, and perform short-time Fourier transform to obtain the power spectrum and phase spectrum of the audio signal; Step 2: construct a multi-level audio feature extractor to extract multi-level audio features of construction waste disposal equipment in the source domain and the target domain; The multi-level audio feature extractor comprises: a first convolution layer and a multi-level convolution block; wherein, The convolution kernel size of the first convolutional layer is 7×7, the number of channels is 16, the stride is 2, and the BN and ReLu functions are used; The multi-level convolution block uses three convolution branches; among them, the first branch has 1 convolution layer, 1×1 convolution kernel size, and 512 channels; the second branch has 2 convolution layers, 1×1 convolution kernel size, 64 channels, 3×3 convolution kernel size, and 512 channels; the third branch has 3 convolution layers, 1×1 convolution kernel size, 64 channels, 5×5 convolution kernel size, 128 channels, and 5×5 convolution kernel size, 512 channels; Perform global average pooling on the feature maps of each of the 512 channels obtained from each branch; concatenate the feature vectors of the three branches, and output a feature vector of size 1×1536; Step 3: Establish a sound source position identifier to predict the sound source position of the audio; Step 4: Establish a noise discriminator to determine whether the audio from various sound sources is noise; Step 5: Combine the multi-level audio feature extractor, the sound source location identifier and the noise discriminator into a noise area detection model, and train the model through an optimization algorithm; Step 6: Input the target domain test sample set into the trained noise area detection model to predict the noise direction of the construction waste disposal equipment online.

2. The noise area detection and identification method according to claim 1, characterized in that: The sound source location identifier consists of three fully connected layers and a Sigmoid activation function. The number of neurons in the three fully connected layers is 1024, 1024, and 360 respectively; the output is a 360-dimensional predicted probability ={ g 1. g 2. … g i …、 g 360 }; in, g i Corresponding to the plane 360 ​​dimension i The probability of a sound source appearing in a certain direction.

3. The noise area detection and identification method according to claim 1, characterized in that: The noise discriminator consists of three fully connected layers and a Sigmoid activation function. The number of neurons in the three fully connected layers is 2048, 2048, and 360 respectively; the output is a 360-dimensional prediction probability ={ u 1. u 2. … u i …、 u 360 }; in, u i Corresponding to the plane 360 ​​dimension i The probability that the sound source in a direction is noise.

4. The noise area detection and identification method according to claim 1, characterized in that: Step 5 specifically includes: Step 51: The multi-level audio feature extractor and the sound source position identifier form a first feedforward neural network, and the multi-level audio feature extractor and the noise discriminator form a second feedforward neural network; Step 52, the multi-level audio feature extractor is used to extract source domain features, the sound source position identifier is used to identify the sound source position of the feature and make the prediction result correctly fit the real audio orientation; the noise discriminator is used to identify whether the feature belongs to noise; and the loss of the prediction result of the sound source position identifier is minimized. L C and the loss of the noise discriminator prediction results L D To optimize the model, the formula is: ； In the formula, L C is the mean absolute error loss of the sound source location identifier prediction results, L D is the mean absolute error loss of the noise discriminator prediction results, represents the mean absolute error value, represents the absolute value, and are the predicted probability of the sound source location identifier and the actual sound source location, and In the i The predicted probability of the noise discriminator and the actual noise result in the dimension, is the weight parameter; Step 53: performing covariance alignment on the source domain and target domain features extracted by the multi-level audio feature extractor; Step 54: Combine the optimization objective function loss and the covariance alignment loss to obtain a total loss function, and use the total loss function to perform back propagation optimization.

5. The noise area detection and identification method according to claim 4, characterized in that: The formula for covariance alignment loss is: ； In the formula, represents the Frobenius norm, represents the covariance calculation, and Represent the feature matrices of the source domain and the target domain respectively, Represents the feature dimension.

6. The noise area detection and identification method according to claim 5, characterized in that: The formula of the total loss function is: ； In the formula, Weighing parameters for the loss function, L s To optimize the objective function loss, is the covariance alignment loss.

7. The noise area detection and identification method according to claim 6, characterized in that: Step 6 specifically includes: Combine the prediction results of the sound source location identifier and the noise discriminator to set the prediction threshold , will be above the prediction threshold The prediction results As the output of noise detection recognition, the formula is: ； In the formula, and Respectively expressed in i The prediction probability of the sound source location identifier and noise discriminator in the direction, For the i The noise source in the dimensional direction, is the noise above the prediction threshold.

8. The noise area detection and identification method according to claim 1, characterized in that: The audio signal is collected by a microphone array.

9. Noise area detection and recognition system, characterized in that: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the noise area detection and identification method according to any one of claims 1 to 8.

10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the noise area detection and identification method according to any one of claims 1 to 8 is implemented.

Images