A method and system for efficiently identifying thunder signals based on microphone array

Abstract

The application discloses a thunder sound signal efficient recognition method and system based on a microphone array, belongs to the thunder sound detection technical field, and comprises five steps of audio signal collection, audio signal screening, preprocessing, model construction and model application. The application adopts the microphone array to synchronously collect multi-channel environmental audio signals in real time, and improves the spatial expression capability of the thunder sound signals. By introducing a multi-channel screening mechanism based on the short-time frame average energy standard deviation, the mute section and the low dynamic background noise are effectively removed, and the effectiveness of the candidate signals is improved. Meanwhile, the application fuses the multi-channel features based on the energy and variance weighting, so that the discrimination capability and the anti-interference capability of the neural network model for the thunder sound audio events are enhanced. The method provided by the application is suitable for real-time thunder sound recognition and monitoring in a complex environment, and has good recognition performance and engineering application value.

Description

Technical Field

[0001] The present application belongs to the field of thunder detection technology, and specifically relates to a method and system for efficiently identifying thunder signals based on a microphone array. Background Art

[0002] Thunder, a unique natural sound, is crucial for accurate detection in weather warnings and disaster prevention and mitigation. Timely and accurate thunder detection can aid weather warning systems, providing information for forecasting severe weather like thunderstorms and heavy rain, and ultimately aiding disaster preparedness. Existing thunder recognition technologies primarily focus on the following methods:

[0003] 1. Methods based on short-term energy peak detection: This method relies on identifying the energy peak of thunder signals and is often used for preliminary detection of thunder events. However, in practice, this method is susceptible to interference from environmental noise. In the presence of strong winds, heavy rain, or other high-energy background noise, the reliability of energy peak detection decreases significantly, resulting in high false alarm and missed alarm rates in thunder recognition systems, reducing recognition accuracy.

[0004] 2. Spectral energy analysis: These methods aim to identify the spectral characteristics of thunder signals and classify them using the specific patterns in thunder's frequency distribution. However, in complex meteorological environments, the spectral characteristics of thunder may share similarities with the spectral characteristics of various natural and man-made noise sources. This is particularly true in the presence of significant background noise, such as wind and traffic noise. This significantly impacts the accuracy of spectral energy analysis, leading to insufficient robustness.

[0005] 3. Single-channel neural network-based approach: This approach attempts to utilize machine learning techniques, such as support vector machines (SVMs), to classify the features of thunder signals collected by a single microphone, aiming to achieve automatic thunder recognition. Although machine learning methods can improve recognition performance to a certain extent, single-channel neural network-based approaches still face inherent technical bottlenecks in the field of thunder recognition. Their main limitations are: (a) Insufficient utilization of spatial information: Single-channel approaches can only process audio signals collected by a single microphone and cannot effectively utilize the differential information of thunder signals in spatial propagation, making it difficult to effectively distinguish thunder from widely distributed ambient noise. (b) Susceptibility to single-channel signal quality limitations: In complex real-world noise environments, the quality of audio signals collected by a single-channel microphone is easily affected by ambient noise, resulting in reduced feature extraction reliability and, in turn, high false alarm and false negative rates. This performance bottleneck of single-channel thunder recognition systems is particularly prominent when multiple noise sources interfere simultaneously.

[0006] In summary, while the aforementioned existing technologies have made some progress in thunder detection and have been applied in real-world scenarios, these approaches still face significant limitations in coping with complex environmental noise, effectively utilizing spatial information, and processing massive amounts of data. These technical bottlenecks hinder the further development and performance improvement of thunder detection technology in practical applications. Therefore, overcoming the shortcomings of existing technologies is a pressing issue in the field of thunder detection technology. Summary of the Invention

[0007] Purpose of the invention: This application develops a method and system for efficiently identifying thunder signals based on a microphone array, aiming to solve the technical problem of low accuracy of thunder recognition in the prior art.

[0008] Technical Solution: In a first aspect, embodiments of the present application provide a method for efficiently identifying thunder signals based on a microphone array, comprising:

[0009] Get a sample audio signal;

[0010] determining a label type of the sample audio signal, where the label type includes a thunder signal and a non-thunder signal;

[0011] Performing data preprocessing on the sample audio signal based on logarithmic filter bank features to obtain a feature matrix;

[0012] A basic model is obtained based on a deep convolutional neural network, and the feature matrix is ​​used as the input of the basic model, and the label type corresponding to the feature matrix is ​​used as the output of the basic model. The basic model is trained to obtain a thunder signal recognition model;

[0013] An audio signal to be tested is obtained and input into the thunder signal recognition model to obtain a thunder signal recognition result of the audio signal to be tested.

[0014] In some embodiments, the audio signal is obtained based on a microphone array;

[0015] The microphone array has a first axis, a second axis and a third axis that are orthogonally arranged, and the intersection of the first axis, the second axis and the third axis is a zero point;

[0016] The microphone array comprises:

[0017] a first sampling point and a third sampling point symmetrically arranged about the zero point along the first axis;

[0018] a second sampling point and a fourth sampling point symmetrically arranged about the zero point along the second axis;

[0019] a fifth sampling point arranged along the third axis;

[0020] The first sampling point, the second sampling point, the third sampling point, the fourth sampling point and the fifth sampling point are at the same distance from the zero point;

[0021] The audio signal includes sub-signals acquired at the first sampling point, the second sampling point, the third sampling point, the fourth sampling point, and the fifth sampling point.

[0022] In some embodiments, the audio signal includes multiple sub-signals, and the step of determining the label type of the sample audio signal includes:

[0023] Obtaining a short-time frame average energy standard deviation of the sub-signal;

[0024] Determine a threshold value of the standard deviation of the short-time frame average energy;

[0025] In response to the maximum value of the short-time frame average energy standard deviation being greater than or equal to the threshold value of the short-time frame average energy standard deviation among the multiple sub-signals, the sample audio signal is determined to be a valid audio signal; otherwise, the sample audio signal is determined to be an invalid audio signal and deleted. This helps to quickly screen audio data and reduce computing resource consumption. After training, it can also improve resource utilization efficiency when used in engineering applications.

[0026] A thunder signal test is performed on the valid audio signal to determine whether the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal, thereby improving the accuracy of subsequent analysis.

[0027] In some embodiments, the step of obtaining the short-time frame average energy standard deviation of the sub-signal includes:

[0028] Performing windowed sampling on the sub-signal to obtain a first windowed signal;

[0029] Performing frame sampling on the first windowed signal to obtain a plurality of first frames of data;

[0030] Obtain a short-time energy average value of a plurality of first frames of data, wherein a characterization formula of the short-time energy average value includes:

[0031] ;

[0032] in, is the short-time energy average value; is the length of the first frame of data, which is used to represent the number of sampling points of the windowed sampling contained in the first frame of data; The sub-signal is The amplitude value of the sampling point, is the sampling point index; To be applied to The window function value of each sampling point;

[0033] The short-time energy average value is normalized, and the short-time frame average energy standard deviation is obtained based on the normalized short-time energy average value in the plurality of first frames of data. The characterization formula of the short-time frame average energy standard deviation includes:

[0034] ;

[0035] ;

[0036] in, is the standard deviation of the average energy of the short time frame; is the number of the first frame data; For the a short-time energy average value of the normalized first frame data; is the average of multiple normalized short-time energy averages.

[0037] In some embodiments, before obtaining the short-time frame average energy standard deviation of the sub-signal, the step of determining the label type of the sample audio signal further includes:

[0038] Acquire an audio time-frequency graph of the sub-signal based on short-time Fourier transform;

[0039] Obtaining a correlation coefficient calculated based on the audio time-frequency diagram;

[0040] Determine the adaptive threshold;

[0041] In response to the correlation coefficient being greater than the adaptive threshold, determining that the sample audio signal is a valid audio signal; otherwise, determining that the sample audio signal is an invalid audio signal;

[0042] A thunder signal test is performed on the valid audio signal to determine whether the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

[0043] In some embodiments, the audio signal includes a plurality of sub-signals, and the step of performing data preprocessing on the sample audio signal based on the log filter bank feature to obtain a feature matrix includes:

[0044] Performing windowed sampling on the sub-signal to obtain a second windowed signal;

[0045] Performing frame sampling on the second windowed signal to obtain a plurality of second frames of data;

[0046] Performing feature processing on the second frame data based on the logarithmic filter bank feature to obtain a feature matrix of the sub-signal;

[0047] The global average energy and the energy variance of the time dimension of the sub-signal are obtained based on the characteristic matrix of the sub-signal; wherein the characterization formulas of the global average energy and the energy variance of the time dimension include:

[0048] ;

[0049] in, For the The global average energy of the sub-signals; is the number of the second frame data; is the number of sampling points contained in the second frame data; For the The sub-signal The second frame of data A logarithmic filter feature of a Mel filter is used to characterize the feature matrix of the sub-signal;

[0050] ;

[0051] ;

[0052] in, For the The energy variance of each sub-signal in the time dimension; For the Among the sub-signals The average energy of the second frame data; For the The average energy value of the sub-signals;

[0053] A weighting coefficient of the sub-signal is obtained based on the global average energy of the sub-signal and the energy variance in the time dimension. The characterization formula of the weighting coefficient includes:

[0054] ;

[0055] in, For the The weighting coefficients of the sub-signals; For the The global average energy of the sub-signals; For the The energy variance of each sub-signal in the time dimension;

[0056] The feature matrices of the plurality of sub-signals are weighted and summed based on the weight coefficients of the sub-signals to obtain a fused feature matrix, which is the feature matrix of the sample audio signal. The characterization formula of the fused feature matrix includes:

[0057] ;

[0058] in, is the fusion feature matrix; For the The characteristic matrix of the sub-signals.

[0059] In some embodiments, the step of performing feature processing on the second frame data based on the logarithmic filter bank feature to obtain the feature matrix of the sub-signal includes:

[0060] Perform a fast Fourier transform on the second frame data to obtain a power spectrum of the second frame data, where a characterization formula of the power spectrum includes:

[0061] ;

[0062] ;

[0063] in, For the a power spectrum of the second frame data; For the The second frame data is in the The power at the frequency component; is the fast Fourier transform; is the square of the complex modulus; is the total number of the frequency components; is the second frame data;

[0064] The power spectrum is weighted filtered based on a Mel filter bank to obtain Mel filter bank energy, where the Mel filter bank includes multiple Mel filters, and the Mel filter bank energy includes output energies of the multiple Mel filters on the power spectrum. The representation formula of the output energy includes:

[0065] ;

[0066] in, For the Mel filter for the The output energy of the power spectrum of the second frame data; For the Mel filters on the frequency components Response from the is the number of the Mel filters;

[0067] Performing a logarithmic operation on the output energy of the Mel filter on the second frame of data to obtain a logarithmic filter bank feature of the second frame of data, wherein a characterization formula of the logarithmic filter bank feature includes:

[0068] ;

[0069] in, For the The second frame data Logarithmic filter characteristics of a Mel filter; To avoid zero values ​​in the logarithmic function, a smoothing constant is used;

[0070] A feature matrix of the sub-signal is obtained based on logarithmic filter bank features of a plurality of second frame data in the sub-signal. A characterization formula of the feature matrix of the sub-signal includes:

[0071] ;

[0072] in, For the The characteristic matrix of the sub-signals; for OK A matrix of columns.

[0073] In some embodiments, the thunder signal recognition model includes:

[0074] A plurality of sequentially connected convolution blocks, each of which includes a sequentially connected convolution layer, a normalization layer, an activation layer, and an average pooling layer;

[0075] The classification module includes a fully connected layer and an activation layer connected in sequence.

[0076] In some embodiments, the windowed sampling is performed based on a Hamming window.

[0077] In a second aspect, an embodiment of the present application further provides an efficient thunder signal recognition system based on a microphone array, comprising:

[0078] A data acquisition module, configured to acquire a sample audio signal;

[0079] a label classification module, the label classification module being configured to determine a label type of the sample audio signal, the label type including a thunder signal and a non-thunder signal;

[0080] A feature matrix module, configured to perform data preprocessing on the sample audio signal based on log filter bank characteristics to obtain a feature matrix;

[0081] a model training module, the model training module being configured to obtain a basic model based on a deep convolutional neural network, and using the feature matrix as input to the basic model and the label type corresponding to the feature matrix as output of the basic model, to perform model training on the basic model and obtain a thunder signal recognition model;

[0082] The result prediction module is used to obtain the audio signal to be tested, and input the thunder signal recognition model to obtain the thunder signal recognition result of the audio signal to be tested.

[0083] Beneficial effect: Compared with the prior art, the embodiment of the present application provides an efficient thunder signal recognition method based on a microphone array, which includes obtaining a sample audio signal for model training and determining the label type of the sample audio signal, the label type including thunder signal and non-thunder signal; performing data preprocessing on the sample audio signal through log filter group features to obtain a feature matrix of the sample audio signal; obtaining a basic model through a deep convolutional neural network, and using the feature matrix as the input of the basic model and the label corresponding to the feature matrix as the output of the basic model to train the basic model to obtain a thunder signal recognition model; obtaining an audio signal to be tested and inputting it into the thunder signal recognition model to obtain a thunder signal recognition result of the audio signal to be tested; the present application combines the data processing method of the log filter group features with the deep convolutional neural network model, which can effectively enhance the ability to capture the time-frequency characteristics of thunder signals and the anti-interference ability under low signal-to-noise ratio conditions, thereby improving the accuracy of thunder recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0085] Figure 1 A flowchart of the steps of the method for efficiently identifying thunder signals based on a microphone array provided in an embodiment of the present application;

[0086] Figure 2 A flowchart of the steps for determining the label type of a sample audio signal in the microphone array-based efficient thunder signal recognition method provided in an embodiment of the present application;

[0087] Figure 3 A flowchart of the steps for obtaining the short-time frame average energy standard deviation of sub-signals in the efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application;

[0088] Figure 4A flowchart of the steps for performing preliminary label type identification on sub-signals in the efficient thunder signal identification method based on a microphone array provided in an embodiment of the present application;

[0089] Figure 5 A flowchart of the steps for obtaining a feature matrix in the efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application;

[0090] Figure 6 A flowchart of the steps for obtaining a feature matrix of sub-signals in the efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application;

[0091] Figure 7 A module connection diagram of a microphone array-based thunder signal efficient recognition system provided in an embodiment of the present application;

[0092] Figure 8 This is the accuracy trend chart of the training set and the validation set;

[0093] Figure 9 This is the loss change trend chart of the training set and the validation set;

[0094] Figure 10 Evaluate the confusion matrix plot for the test set data;

[0095] Figure 11 A structural diagram of a microphone array in an efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application;

[0096] Figure 12 A flowchart of a method for efficiently identifying thunder signals based on a microphone array according to an embodiment of the present application;

[0097] Figure numerals: 10, data acquisition module; 20, label classification module; 30, feature matrix module; 40, model training module; 50, result prediction module. DETAILED DESCRIPTION

[0098] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0099] Thunder, a unique natural sound, is crucial for accurate detection in weather warnings and disaster prevention and mitigation. Timely and accurate thunder detection can aid weather warning systems, providing information for forecasting severe weather like thunderstorms and heavy rain, and ultimately aiding disaster preparedness. Existing thunder recognition technologies primarily focus on the following methods:

[0100] 1. Short-time energy peak detection-based methods: These methods rely on identifying the energy peaks of thunder signals and are commonly used for preliminary detection of thunder events. However, such methods are susceptible to environmental noise in practical applications. When there are strong winds, heavy rain, or other high-energy background noise in the environment, the reliability of energy peak detection decreases significantly, leading to high false positive and false negative rates for thunder recognition systems, reducing recognition accuracy.

[0101] 2. Frequency spectrum energy analysis-based methods: These methods aim to identify the frequency spectrum characteristics of thunder signals and use the specific patterns exhibited by thunder in frequency distribution for signal classification. However, in complex meteorological environments, the frequency spectrum characteristics of thunder may exhibit similarities with those of various natural or artificial noise sources, especially in the presence of significant background noise such as wind noise and traffic noise. The recognition accuracy of frequency spectrum energy analysis methods is greatly affected, and their robustness is insufficient.

[0102] 3. Single-channel neural network-based methods: These methods attempt to use machine learning techniques, such as Support Vector Machines (SVM), to classify the features of thunder signals collected by a single microphone, with the goal of achieving automatic thunder recognition. Although machine learning methods can improve recognition performance to some extent, single-channel neural network-based methods still face inherent technical bottlenecks in the field of thunder recognition. The main limitations are: (a) insufficient utilization of spatial information: single-channel methods can only process audio signals collected by a single microphone and cannot effectively utilize the differential information of thunder signals in spatial propagation, making it difficult to effectively distinguish thunder from spatially distributed environmental noise. (b) susceptible to single-channel signal quality: in actual complex noise environments, the quality of audio signals collected by a single-channel microphone is easily disturbed by environmental noise, leading to a decrease in feature extraction reliability and causing high false positive and false negative rates. In particular, in the presence of multiple source noise interference, the performance bottleneck of single-channel thunder recognition systems is particularly prominent.

[0103] In summary, although the above existing technologies have made certain progress in the field of thunder recognition and have been applied in practical scenarios, these technical methods still have limitations in dealing with complex environmental noise, effectively utilizing spatial information, and processing massive data, which restrict the further development and performance improvement of thunder recognition technology in practical applications. Therefore, how to overcome the shortcomings of existing technologies is a problem that needs to be solved in the field of thunder detection technology.

[0104] Therefore, the embodiment of the present application provides a thunder sound signal efficient identification method based on a microphone array, which comprises the following steps: acquiring sample audio signals for model training, and determining label types of the sample audio signals, wherein the label types comprise thunder sound signals and non-thunder sound signals; performing data preprocessing on the sample audio signals through a logarithmic filter bank feature to acquire a feature matrix of the sample audio signals; acquiring a basic model through a deep convolutional neural network, taking the feature matrix as an input of the basic model, taking labels corresponding to the feature matrix as an output of the basic model, performing model training on the basic model, and acquiring a thunder sound signal identification model; acquiring a to-be-tested audio signal, and inputting the to-be-tested audio signal into the thunder sound signal identification model to acquire a thunder sound signal identification result of the to-be-tested audio signal; and the data processing mode of the logarithmic filter bank feature and the deep convolutional neural network model are combined, the capture capability of the thunder sound signal time-frequency feature and the anti-interference capability under a low signal-to-noise ratio condition are effectively enhanced, and the accuracy of thunder sound identification is improved.

[0105] In some embodiments, referring to Figure 1 and Figure 12 , Figure 1 the step flowchart of the thunder sound signal efficient identification method based on the microphone array provided by the embodiment of the present application, Figure 12 the program flowchart of the thunder sound signal efficient identification method based on the microphone array provided by the embodiment of the present application, and the thunder sound signal efficient identification method based on the microphone array provided by the embodiment of the present application is specifically implemented through steps 100 to 500:

[0106] Step 100: acquiring sample audio signals.

[0107] In some embodiments, the microphone array is arranged to acquire audio signals, referring to Figure 11 , Figure 11 the structure diagram of the microphone array in the thunder sound signal efficient identification method based on the microphone array provided by the embodiment of the present application, the microphone array has a first axis, a second axis and a third axis arranged orthogonally, and the intersection of the first axis, the second axis and the third axis is a zero point; the microphone array comprises:

[0108] a first sampling point and a third sampling point arranged symmetrically about the zero point along the first axis;

[0109] a second sampling point and a fourth sampling point arranged symmetrically about the zero point along the second axis;

[0110] a fifth sampling point arranged along the third axis;

[0111] The distances from the first sampling point, the second sampling point, the third sampling point, the fourth sampling point and the fifth sampling point to the zero point are the same;

[0112] The audio signal includes sub-signals acquired at a first sampling point, a second sampling point, a third sampling point, a fourth sampling point, and a fifth sampling point.

[0113] Specifically, the present application sets microphones for collecting sub-signals at the first sampling point, the second sampling point, the third sampling point, the fourth sampling point and the fifth sampling point, and the microphones are omnidirectional condenser microphones with a frequency response range of 20-20kHz. Taking the zero point as the origin, the microphone at the first sampling point is mic_1, and the position is (1.55, 0, 0); the microphone at the second sampling point is mic_2, and the position is (0, 1.55, 0); the microphone at the third sampling point is mic_3, and the position is (-1.55, 0, 0); the microphone at the fourth sampling point is mic_4, and the position is (0, -1.55, 0); the microphone at the fifth sampling point is mic_5, and the position is (0, 0, 1.55); the dimension of each position is meter. The microphone array of the present application realizes a symmetrical distribution in three-dimensional space, and maximizes the capture of the spatial characteristics of thunder signals by forming balanced spatial coverage. Therefore, the baseline distance between adjacent microphones is = 2.19 meters (the microphone coordinate radius is 1.55 meters, which is the distance from the coordinate origin to the microphone. The distance between two microphones is called the baseline distance, which is the distance between the hypotenuse of an isosceles right triangle). This distance has been optimized to ensure that the array has sufficient spatial resolution to distinguish sound sources from different directions while avoiding the high-frequency aliasing and engineering difficulties that may be caused by an excessively long baseline. In addition, the 1.55-meter distance between the center microphone, mic_5 (the fifth sampling point), and the array's geometric center (0, 0, 0) effectively captures acoustic information in the vertical direction, further enhancing the array's spatial perception of thunder signals.

[0114] It can be understood that this application combines the compactness and efficiency of microphone arrays when collecting audio signals, and demonstrates excellent performance in capturing thunder propagation characteristics, such as propagation delay, phase difference, and amplitude change, which helps to improve the accuracy and stability of signal processing.

[0115] In some embodiments, when collecting audio signals, the sampling rate of the present application is 44100 Hz, the sampling duration is set to 5 seconds, the sampling file is an audio signal in .wav format, and it is continuously collected according to the file naming format of year-month-day-hour-minute-second. At the same time, GPS is used to calibrate the time, and the time error is controlled at the millisecond level. The audio stream file is named with the year-month-day-hour-minute-second timestamp. After the data collection is completed, in order to facilitate subsequent archiving and offline analysis, the collected multi-channel audio stream is encapsulated in .wav format and packaged into a .zip compressed file. Specifically, data is collected continuously for 24 hours a day, and each microphone generates a total of .wav audio files. For 5 microphones, .wav audio files. Based on this, the total amount of audio data per month is .wav audio files, each 432KB in size, resulting in 1,069GB of data requiring real-time processing per month. However, in the thunderstorm-prone summer, thunderstorm activity is typically concentrated within a few hours over several days, so not all collected audio signals are thunder data. Assuming a 10-hour recording time for thunder in a thunderstorm-prone month, 36,000 audio files would be collected, with a monthly sample size of 2,592,000, representing approximately 1.38% of the total data. This indicates that the vast majority of the collected audio data (98.62%) is non-thunder data, and this proportion can rise to 99.8% in months and regions with fewer thunderstorms. However, further analysis of the non-thunder data reveals that, across different seasons, stations, and non-thunder data, 50%-75% is silence or audio signals without significant energy fluctuations. Therefore, it is necessary to screen the collected audio samples to identify thunder signals.

[0116] Step 200: Determine the label type of the sample audio signal, where the label type includes thunder signal and non-thunder signal.

[0117] In some embodiments, the audio signal includes multiple sub-signals, which are audio data obtained from channels formed by multiple microphones. Figure 2 , Figure 2 This is a flowchart of the steps for determining the label type of a sample audio signal in the efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application. The method for determining the label type of a sample audio signal in the present application is implemented through steps 210 to 240:

[0118] Step 210: Obtain the short-time frame average energy standard deviation of the sub-signal.

[0119] In some embodiments, before obtaining the short-time frame average energy standard deviation of the sub-signal, the present application performs preliminary tag type identification on the sub-signal, see Figure 4 , Figure 4 This is a flowchart of the steps for performing preliminary tag type identification on sub-signals in the efficient thunder signal identification method based on a microphone array provided in an embodiment of the present application. The method for performing preliminary tag type identification on sub-signals in the present application is implemented through steps 201 to 205:

[0120] Step 201: Obtain an audio time-frequency diagram of a sub-signal based on short-time Fourier transform.

[0121] Step 202: Obtain the correlation coefficient calculated based on the audio time-frequency diagram of the standard thunder template.

[0122] Step 203: Determine an adaptive threshold.

[0123] In some embodiments, the adaptive threshold is set to 0.4.

[0124] Step 204: In response to the correlation coefficient being greater than the adaptive threshold, determine that the sample audio signal is a valid audio signal; otherwise, determine that the sample audio signal is an invalid audio signal.

[0125] Step 205: Perform a thunder signal test on the valid audio signal, and determine that the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

[0126] In some embodiments, see Figure 3 , Figure 3 This is a flowchart of the steps for obtaining the short-time frame average energy standard deviation of the sub-signal in the efficient thunder signal recognition method based on the microphone array provided in an embodiment of the present application. The method for obtaining the short-time frame average energy standard deviation of the sub-signal in the present application is specifically implemented through steps 211 to 214:

[0127] Step 211: Perform window sampling on the sub-signal to obtain a first windowed signal.

[0128] Step 212: Frame sampling is performed on the first windowed signal to obtain a plurality of first frame data.

[0129] In some embodiments, windowed sampling is performed based on a Hamming window, with a window length of 1024 sampling points and a frame shift of 512 sampling points. Specifically, considering that the speech signal has short-term stationarity in a short period of time (approximately 23 milliseconds), that is, its characteristics change relatively slowly in a short period of time, the present application sets the frame length to 23 milliseconds to ensure that the short-term characteristics of the speech signal can be effectively captured. In order to achieve a frame length of approximately 23 milliseconds, based on the 44.1 kHz sampling rate used by the device in this application, it is necessary to calculate the corresponding Hamming window length (i.e., frame length, in units of sampling points): Sampling point.

[0130] Step 213: Obtain the short-time energy average value of the plurality of first frame data. The characterization formula of the short-time energy average value includes:

[0131] ;

[0132] in, is the short-time energy average; is the length of the first frame of data, which is used to represent the number of sampling points of the windowed sampling contained in the first frame of data; For the sub-signal The amplitude value of the sampling point, The sampling point index ranges from 0 to ; To be applied to The window function value of each sampling point.

[0133] Step 214: normalize the short-time energy average value, and obtain the short-time frame average energy standard deviation based on the normalized short-time energy average value in the plurality of first frames of data. The characterization formula of the short-time frame average energy standard deviation includes:

[0134] ;

[0135] ;

[0136] in, is the standard deviation of the average energy of the short time frame; is the number of the first frame data; For the The short-time energy average value after normalization of the first frame data; is the average of multiple normalized short-time energy averages.

[0137] In some embodiments, the characterization formula for normalizing the short-time energy average value includes:

[0138] ;

[0139] in, is the normalized short-time energy average; It is the maximum value of the short-time energy average value in all the first frame data in the sub-signal.

[0140] Step 220: Determine a threshold value of the short-time frame average energy standard deviation.

[0141] In some embodiments, the present application performs statistical analysis on audio signals of existing thunder and a large number of station ambient noise samples (including silence, wind, rain, and traffic noise, etc.) to determine that the preset threshold is 0.00005.

[0142] Step 230: In response to the maximum value of the short-time frame average energy standard deviation being greater than or equal to the threshold value of the short-time frame average energy standard deviation among the multiple sub-signals, determine that the sample audio signal is a valid audio signal; otherwise, determine that the sample audio signal is an invalid audio signal.

[0143] Specifically, for each group of audio samples containing 5 channels, the short-time frame average energy standard deviation of each channel is calculated respectively, and the maximum value is taken as the overall energy feature of the sample and compared with the threshold: if the maximum value is less than the threshold, the entire group of audio is judged to be an invalid audio signal and is eliminated as a whole; otherwise, the group of audio samples is retained as a valid audio signal.

[0144] It is understandable that deleting invalid audio signals after determining them can help to quickly filter audio data and reduce computing resource consumption; after training, it can also improve resource utilization efficiency during engineering use.

[0145] Step 240: Perform a thunder signal test on the valid audio signal, and determine that the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

[0146] In some embodiments, after determining a valid audio signal, further manual inspection is performed to determine that the label type of the valid audio signal that meets the inspection requirements is a thunder signal, so as to improve the accuracy of subsequent analysis.

[0147] As can be understood, the short-time frame average energy standard deviation screening mechanism proposed in this application utilizes a 23ms Hamming window to effectively capture the short-time characteristics of audio signals, balancing time and frequency domain resolution. Based on a 44.1kHz sampling rate, this corresponds to a Hamming window length of 1024 samples. In conjunction with this screening mechanism, the frame shift is set to 512 samples (approximately 11.6ms), maintaining a 50% overlap. This setting reduces computational redundancy while ensuring smooth short-time energy calculation, minimizing the loss of sudden signal mutations, and optimizing the temporal continuity and feature extraction accuracy of thunder detection. Statistical analysis of the short-time frame average energy standard deviation values ​​of a large number of thunder and non-thunder audio samples revealed that the short-time frame average energy standard deviation values ​​of thunder audio signals are generally higher than those of background noise signals. Based on experimental data analysis, a threshold of 0.00005 was set to balance the thunder signal detection rate with the background noise false rejection rate. This threshold is determined based on the statistical characteristics of the short-time frame average energy standard deviation and experimental optimization. It aims to enhance the ability to distinguish the energy fluctuation characteristics of thunder signals from background noise signals, so as to balance thunder detection efficiency and resource conservation.

[0148] Step 300: Perform data preprocessing on the sample audio signal based on the logarithmic filter bank feature to obtain a feature matrix.

[0149] In some embodiments, see Figure 5 , Figure 5 This is a flowchart of the steps for obtaining a feature matrix in the method for efficient thunder signal recognition based on a microphone array provided in an embodiment of the present application. The method for obtaining a feature matrix in the present application is implemented through steps 310 to 360:

[0150] Step 310: Perform window sampling on the sub-signal to obtain a second windowed signal.

[0151] Step 320: Perform frame sampling on the second windowed signal to obtain a plurality of second frame data.

[0152] Step 330: Perform feature processing on the second frame data based on the logarithmic filter bank feature to obtain a feature matrix of the sub-signal.

[0153] In some embodiments, see Figure 6 , Figure 6 This is a flowchart of the steps for obtaining the characteristic matrix of sub-signals in the efficient thunder signal recognition method based on a microphone array provided in an embodiment of the present application. The method for obtaining the characteristic matrix of sub-signals in the present application is specifically implemented through steps 331 to 334:

[0154] Step 331: Perform a fast Fourier transform on the second frame data to obtain a power spectrum of the second frame data. The power spectrum characterization formula includes:

[0155] ;

[0156] ;

[0157] in, For the The power spectrum of the second frame data; For the The second frame data is in The power at the frequency component; is the fast Fourier transform; is the square of the complex modulus; is the total number of frequency components; The second frame data.

[0158] Step 332: Perform weighted filtering on the power spectrum based on the Mel filter bank to obtain Mel filter bank energy. The Mel filter bank includes multiple Mel filters. The Mel filter bank energy includes the output energy of the multiple Mel filters on the power spectrum. The representation formula of the output energy includes:

[0159] ;

[0160] in, For the Mel filter for the The output energy of the power spectrum of the second frame data; For the Mel filters on the frequency components Response from the is the number of Mel filters.

[0161] Step 333: Perform a logarithmic operation on the output energy of the Mel filter on the second frame data to obtain the logarithmic filter bank features of the second frame data. The representation formula of the logarithmic filter bank features includes:

[0162] ;

[0163] in, For the The second frame of data Logarithmic filter characteristics of a Mel filter; Smoothing constant to avoid zero values ​​in the logarithmic function.

[0164] Step 334: Obtain a feature matrix of the sub-signal based on the logarithmic filter bank features of the plurality of second frame data in the sub-signal. The representation formula of the feature matrix of the sub-signal includes:

[0165] ;

[0166] in, For the The characteristic matrix of the sub-signals; for OK A matrix of columns, is the number of the second frame data, is the number of sampling points contained in the second frame of data.

[0167] Step 340: Obtain the global average energy and the energy variance of the time dimension of the sub-signal based on the characteristic matrix of the sub-signal. The characterization formulas of the global average energy and the energy variance of the time dimension include:

[0168] ;

[0169] in, For the The global average energy of the sub-signals; is the number of the second frame data; is the number of sampling points contained in the second frame of data; For the The sub-signal The second frame of data The logarithmic filter features of the Mel filter are used to characterize the feature matrix of the sub-signal;

[0170] ;

[0171] ;

[0172] in, For the The energy variance of each sub-signal in the time dimension; For the Among the sub-signals The average energy of the second frame data; For the The average energy value of the sub-signals.

[0173] Step 350: Obtain the weighting coefficient of the sub-signal based on the global average energy and the energy variance in the time dimension of the sub-signal. The representation formula of the weighting coefficient includes:

[0174] ;

[0175] in, For the The weighting coefficients of the sub-signals; For the The global average energy of the sub-signals; For the The energy variance of the sub-signal in the time dimension.

[0176] Step 360: Perform weighted summation on the feature matrices of the multiple sub-signals based on the weighting coefficients of the sub-signals to obtain a fused feature matrix, which is the feature matrix of the sample audio signal. The representation formula of the fused feature matrix includes:

[0177] ;

[0178] in, is the fusion feature matrix; For the The characteristic matrix of the sub-signals.

[0179] Specifically, the dataset used in this example is derived from thunder audio data collected by microphone array devices deployed by the applicant in multiple cities. During the test, the ambient noise level was 40-50dB, covering a variety of weather conditions such as sunny days, rainy days and thunderstorms. A total of 433 groups of thunder samples and 947 groups of non-thunder samples were screened out (each sample contains a 5-channel .wav file with a duration of 5 seconds), which is 2165 thunder data and 4735 non-thunder sample data. The thunder samples were confirmed through manual annotation, and their characteristics are low-frequency roar and energy mutation; non-thunder samples include silence, wind, rain and traffic noise. The specific training process divides the dataset into training set, validation set and test set in a ratio of 7:1.5:1.5. After preprocessing, the audio signal is converted into logarithmic filter bank features to extract its representative feature information in the time and frequency domain. Specifically, the audio signal is first framed, and then the 64-dimensional Mel filter bank energy value is calculated on each frame, and its logarithm is taken to obtain a two-dimensional feature matrix with a dimension of (862, 64), where 862 represents the number of frames and 64 represents the number of filters. For the five-channel audio data of each sample, the above features are extracted respectively, and five independent (862, 64) feature matrices are generated accordingly. Taking into account the spatial anisotropy of thunder signals during propagation and the non-uniform impact of environmental noise on different channels, this application proposes an adaptive fusion strategy based on energy-variance weighting. This method calculates the product of the signal energy and feature variance of each channel feature as a weight index, giving higher weights to high-quality channels while weakening the contribution of channels contaminated by noise. This weighted fusion strategy is superior to traditional channel stacking or simple averaging methods, and can effectively retain multi-channel complementary information while reducing redundancy and noise interference, thereby reducing the computational complexity of the model. After fusion, a single (862, 64) feature matrix is ​​obtained as the input of the model. Compared to directly using three-dimensional feature tensor input, this method not only preserves the richness of multi-channel information but also significantly improves the model's robustness and generalization capabilities, especially in complex environments. Furthermore, this fusion mechanism, as a preprocessing step, keeps the model architecture compact, avoiding the overfitting problem that can result from complex multi-channel processing networks, and achieving a balance between computational efficiency and classification performance.

[0180] Step 400: Obtain a basic model based on a deep convolutional neural network, use the feature matrix as the input of the basic model, and the label type corresponding to the feature matrix as the output of the basic model, perform model training on the basic model, and obtain a thunder signal recognition model.

[0181] In some embodiments, see Figure 8 and Figure 9 , Figure 8 This is the accuracy trend chart of the training set and the validation set. Figure 9The following plots show the trend of loss changes for the training and validation sets. Both the training accuracy and validation accuracy increase rapidly within the first 20 epochs, then stabilize, ultimately settling at approximately 0.98, demonstrating that the model has a high fit for both the training and validation data. The training loss and validation loss decrease significantly within the first 10 epochs and then gradually converge to a low level (approximately 0.02), indicating that the model converged well during training and exhibited no significant overfitting. These results further validate the robustness and reliability of the proposed model for thunder recognition tasks and provide theoretical support for its application in practical applications.

[0182] In some embodiments, see Figure 10 , Figure 10 The confusion matrix plot for the test data was used to evaluate the trained automatic thunder recognition model. The model was fully validated on an independent test dataset. The test results were evaluated using the confusion matrix and various performance metrics. The model achieved 99% accuracy for non-thunder categories and 98% accuracy for thunder categories, demonstrating extremely high classification accuracy. Correspondingly, the false positive rate (false alarm rate) was only 1%, and the false negative rate (false negative rate) was only 2%, effectively balancing false alarms and false negatives. Further analysis of thunder category recognition performance showed precision, recall, and F1 scores all exceeding 0.98. These detailed metrics demonstrate the model's ability to accurately capture thunder events in complex backgrounds and effectively suppress non-thunder interference.

[0183] In some embodiments, the thunder signal recognition model includes:

[0184] Multiple convolution blocks connected sequentially, where the convolution block includes a sequentially connected convolution layer, a normalization layer, an activation layer, and an average pooling layer;

[0185] Classification module,The classification module includes sequentially connected fully connected layers and activation layers.

[0186] Step 500: Acquire an audio signal to be tested, input it into a thunder signal recognition model, and obtain a thunder signal recognition result of the audio signal to be tested.

[0187] In some embodiments, based on step 100, the audio signal to be tested is obtained, and then the audio signal is screened through step 200 to preliminarily remove non-thunder audio signals, and then preprocessed according to the method of step 300, and finally input into the model constructed in step 400 to further identify whether the collected audio signal is a thunder signal.

[0188] For example, in order to quantify the technical advantages of the present invention, the performance of the present model is compared with the prior art. The comparison results are detailed in Table 1. As shown in Table 1, the model of the present application is significantly better than the prior art in many key performance indicators. For example, compared with the best method based on filter combination in the literature [1] Wang Y, Yang J, Zhang Q, et al. Application of combined filtering in thunder recognition [J]. Remote Sensing, 2023, 15(2): 432. (overall accuracy rate of 93.18%), the overall recognition accuracy of the present application is about 98.5%, achieving an improvement of more than 5%. More importantly, the advantages of this application are more prominent in terms of segmentation and recognition capabilities: the accuracy of non-thunder categories is as high as 99%, which is much higher than the specificity (91%) of the CNN-based method in the literature [2] Zhang H, Yan B, Gu S, et al. Thunder signal detection viadeep learning[C] / / Journal of Physics: Conference Series. IOP Publishing,2021, 1828(1): 012023., indicating that the present invention has significantly enhanced suppression capabilities for non-thunder interference (such as wind, rain, traffic noise, etc.), and can greatly reduce the false alarm rate; the accuracy (>0.98) and recall rate (>0.98) of thunder categories, as well as the comprehensive performance index F1 score (>0.98), are also far higher than the thunder category accuracy reported in literature [2] (referred to as Sensitivity in literature [2], which is 0.79). It can be seen that this application can identify thunder events with extremely high accuracy and completeness, effectively reducing missed alarms.

[0189] Table 1 Comparison of the performance of the model of the present invention and the prior art

[0190]

[0191] In the table, “—” indicates that the corresponding literature did not calculate the measurement value of the evaluation indicator, that is, literature [1] and literature [2] did not calculate the indicators in the corresponding rows, so they are marked with “—” (there is no “—” in literature [2]).

[0192] This application achieves these excellent performances thanks to the effective combination of a proposed multi-channel audio feature extraction method and a deep convolutional neural network model. In particular, through multi-channel feature fusion and Mel filter bank energy analysis, the model enhances its ability to capture the time-frequency characteristics of thunder signals and its ability to resist interference under low signal-to-noise ratio conditions. These significant technological advances enable this invention to provide efficient and reliable technical support for real-time, accurate monitoring and early warning systems for thunderstorms, promising broad practical application prospects.

[0193] It can be understood that the embodiment of the present application provides an efficient method for identifying thunder signals based on a microphone array, which includes obtaining a sample audio signal for model training and determining the label type of the sample audio signal, the label type including thunder signal, non-thunder signal, and thunder signal; performing data preprocessing on the sample audio signal through log filter group features to obtain a feature matrix of the sample audio signal; obtaining a basic model through a deep convolutional neural network, and using the feature matrix as the input of the basic model, and the label corresponding to the feature matrix as the output of the basic model, training the basic model to obtain a thunder signal recognition model; obtaining an audio signal to be tested and inputting it into the thunder signal recognition model to obtain a thunder signal recognition result of the audio signal to be tested; the present application combines the data processing method of the log filter group features with the deep convolutional neural network model, which can effectively enhance the ability to capture the time-frequency characteristics of thunder signals and the anti-interference ability under low signal-to-noise ratio conditions, thereby improving the accuracy of thunder recognition.

[0194] Accordingly, the embodiment of the present application also provides a thunder signal efficient recognition system based on a microphone array, see Figure 7 , Figure 7 This is a module connection diagram of the microphone array-based thunder signal efficient recognition system provided in an embodiment of the present application. The microphone array-based thunder signal efficient recognition system provided in an embodiment of the present application includes:

[0195] The data acquisition module 10 is used to acquire a sample audio signal;

[0196] A label classification module 20 is used to determine a label type of the sample audio signal, where the label type includes a thunder signal and a non-thunder signal;

[0197] A feature matrix module 30 is used to perform data preprocessing on the sample audio signal based on the logarithmic filter bank characteristics to obtain a feature matrix;

[0198] The model training module 40 is used to obtain a basic model based on a deep convolutional neural network, and use the feature matrix as the input of the basic model and the label type corresponding to the feature matrix as the output of the basic model to train the basic model and obtain a thunder signal recognition model;

[0199] The result prediction module 50 is used to obtain the audio signal to be tested, and input the thunder signal recognition model to obtain the thunder signal recognition result of the audio signal to be tested.

[0200] In some embodiments, the tag classification module 20 is specifically configured to:

[0201] Obtain the short-time frame average energy standard deviation of the sub-signal;

[0202] Determine a threshold value of the standard deviation of the short-time frame average energy;

[0203] In response to a maximum value of the short-time frame average energy standard deviation being greater than or equal to a threshold value of the short-time frame average energy standard deviation among the multiple sub-signals, the sample audio signal is determined to be a valid audio signal; otherwise, the sample audio signal is determined to be an invalid audio signal;

[0204] A thunder signal test is performed on the valid audio signal to determine that the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

[0205] In some embodiments, the tag classification module 20 is specifically configured to:

[0206] Performing window sampling on the sub-signal to obtain a first windowed signal;

[0207] Performing frame sampling on the first windowed signal to obtain a plurality of first frame data;

[0208] Obtain the short-time energy average value of multiple first frame data. The characterization formula of the short-time energy average value includes:

[0209] ;

[0210] in, is the short-time energy average; is the length of the first frame of data, which is used to represent the number of sampling points of the windowed sampling contained in the first frame of data; For the sub-signal The amplitude value of the sampling point, is the sampling point index; To be applied to The window function value of each sampling point;

[0211] The short-time energy average value is normalized, and based on the normalized short-time energy average value, the short-time frame average energy standard deviation is obtained in the plurality of first frame data. The characterization formula of the short-time frame average energy standard deviation includes:

[0212] ;

[0213] ;

[0214] in, is the standard deviation of the average energy of the short time frame; is the number of the first frame data; For the The short-time energy average value after normalization of the first frame data; is the average of multiple normalized short-time energy averages.

[0215] In some embodiments, the tag classification module 20 is specifically configured to:

[0216] Obtain audio time-frequency diagram of sub-signal based on short-time Fourier transform;

[0217] Obtain the correlation coefficient of the standard thunder template calculation based on the audio time-frequency diagram;

[0218] Determine the adaptive threshold;

[0219] In response to the correlation coefficient being greater than the adaptive threshold, determining the sample audio signal to be a valid audio signal; otherwise, determining the sample audio signal to be an invalid audio signal;

[0220] A thunder signal test is performed on the valid audio signal to determine whether the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

[0221] In some embodiments, the feature matrix module 30 is specifically configured to:

[0222] Performing windowed sampling on the sub-signal to obtain a second windowed signal;

[0223] Performing frame sampling on the second windowed signal to obtain a plurality of second frame data;

[0224] Perform feature processing on the second frame data based on the logarithmic filter bank feature to obtain a feature matrix of the sub-signal;

[0225] The global average energy and the energy variance in the time dimension of the sub-signal are obtained based on the characteristic matrix of the sub-signal. The characterization formulas of the global average energy and the energy variance in the time dimension include:

[0226] ;

[0227] in, For the The global average energy of the sub-signals; is the number of the second frame data; is the number of sampling points contained in the second frame of data; For the The sub-signal The second frame of data The logarithmic filter features of the Mel filter are used to characterize the feature matrix of the sub-signal;

[0228] ;

[0229] ;

[0230] in, For the The energy variance of each sub-signal in the time dimension; For the Among the sub-signals The average energy of the second frame data; For the The average energy value of the sub-signals;

[0231] The weighting coefficient of the sub-signal is obtained based on the global average energy of the sub-signal and the energy variance in the time dimension. The characterization formula of the weighting coefficient includes:

[0232] ;

[0233] in, For the The weighting coefficients of the sub-signals; For the The global average energy of the sub-signals; For the The energy variance of each sub-signal in the time dimension;

[0234] The feature matrices of multiple sub-signals are weighted and summed based on the weighting coefficients of the sub-signals to obtain a fusion feature matrix, which is the feature matrix of the sample audio signal. The representation formula of the fusion feature matrix includes:

[0235] ;

[0236] in, is the fusion feature matrix; For the The characteristic matrix of the sub-signals.

[0237] In some embodiments, the feature matrix module 30 is specifically configured to:

[0238] Perform fast Fourier transform on the second frame data to obtain the power spectrum of the second frame data. The characterization formula of the power spectrum includes:

[0239] ;

[0240] ;

[0241] in, For the The power spectrum of the second frame data; For the The second frame data is in The power at the frequency component; is the fast Fourier transform; is the square of the complex modulus; is the total number of frequency components; is the second frame data;

[0242] The power spectrum is weighted filtered based on the Mel filter bank to obtain the Mel filter bank energy. The Mel filter bank includes multiple Mel filters. The Mel filter bank energy includes the output energy of the power spectrum of multiple Mel filters. The representation formula of the output energy includes:

[0243] ;

[0244] in, For the Mel filter for the The output energy of the power spectrum of the second frame data; For the Mel filters on the frequency components Response from the is the number of Mel filters;

[0245] The output energy of the Mel filter on the second frame data is logarithmically operated to obtain the logarithmic filter bank features of the second frame data. The characterization formula of the logarithmic filter bank features includes:

[0246] ;

[0247] in, For the The second frame of data Logarithmic filter characteristics of a Mel filter; To avoid zero values ​​in the logarithmic function, a smoothing constant is used;

[0248] A feature matrix of the sub-signal is obtained based on the logarithmic filter bank features of the plurality of second frame data in the sub-signal. The representation formula of the feature matrix of the sub-signal includes:

[0249] ;

[0250] in, For the The characteristic matrix of the sub-signals; for OK A matrix of columns.

[0251] The present application has provided a detailed introduction to an efficient method and system for identifying thunder signals based on a microphone array provided in an embodiment of the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only intended to help understand the method and core concept of the present application. At the same time, for those skilled in the art, according to the concept of the present application, there may be changes in the specific implementation methods and application scope. In summary, the contents of this specification should not be understood as limiting the present application.

Claims

1. An efficient thunder signal recognition method based on microphone array, characterized in that: include: Acquire a sample audio signal, where the audio signal includes a plurality of sub-signals; determining a label type of the sample audio signal, where the label type includes a thunder signal and a non-thunder signal; The step of determining the label type of the sample audio signal comprises: Obtaining a short-time frame average energy standard deviation of the sub-signal; the step of obtaining the short-time frame average energy standard deviation of the sub-signal includes: Performing window sampling on the sub-signal to obtain a first windowed signal; Performing frame sampling on the first windowed signal to obtain a plurality of first frames of data; Obtain a short-time energy average value of a plurality of first frames of data, wherein a characterization formula of the short-time energy average value includes: Among them, E frame is the short-time energy average value; N is the length of the first frame of data, which is used to represent the number of sampling points of the windowed sampling contained in the first frame of data; x[n] is the amplitude value of the sub-signal at the nth sampling point, and n is the sampling point index; w[n] is the window function value applied to the nth sampling point; The short-time energy average value is normalized, and the short-time frame average energy standard deviation is obtained based on the normalized short-time energy average value in the plurality of first frames of data. The characterization formula of the short-time frame average energy standard deviation includes: Wherein, STE_SD is the standard deviation of the short-time frame average energy; M is the number of the first frame data; E i is the normalized short-time energy average value of the i-th first frame data; μ is the average of multiple normalized short-time energy average values; Determine a threshold value of the standard deviation of the short-time frame average energy; In response to a maximum value of the short-time frame average energy standard deviation being greater than or equal to a threshold value of the short-time frame average energy standard deviation among the multiple sub-signals, determining that the sample audio signal is a valid audio signal; otherwise, determining that the sample audio signal is an invalid audio signal; Performing a thunder signal test on the valid audio signal, and determining that the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal; Performing data preprocessing on the sample audio signal based on logarithmic filter bank features to obtain a feature matrix; A basic model is obtained based on a deep convolutional neural network, and the feature matrix is ​​used as the input of the basic model, and the label type corresponding to the feature matrix is ​​used as the output of the basic model. The basic model is trained to obtain a thunder signal recognition model; An audio signal to be tested is obtained and input into the thunder signal recognition model to obtain a thunder signal recognition result of the audio signal to be tested.

2. The method for efficiently identifying thunder signals based on a microphone array according to claim 1, characterized in that: Acquiring the audio signal based on a microphone array; The microphone array has a first axis, a second axis and a third axis that are orthogonally arranged, and the intersection of the first axis, the second axis and the third axis is a zero point; The microphone array comprises: a first sampling point and a third sampling point symmetrically arranged about the zero point along the first axis; a second sampling point and a fourth sampling point symmetrically arranged about the zero point along the second axis; a fifth sampling point arranged along the third axis; The first sampling point, the second sampling point, the third sampling point, the fourth sampling point and the fifth sampling point are at the same distance from the zero point; The audio signal includes sub-signals acquired at the first sampling point, the second sampling point, the third sampling point, the fourth sampling point, and the fifth sampling point.

3. The method for efficiently identifying thunder signals based on a microphone array according to claim 1, characterized in that: Before obtaining the short-time frame average energy standard deviation of the sub-signal, the step of determining the label type of the sample audio signal further includes: Acquire an audio time-frequency diagram of the sub-signal based on short-time Fourier transform; Obtaining a correlation coefficient calculated based on the audio time-frequency diagram; Determine the adaptive threshold; In response to the correlation coefficient being greater than the adaptive threshold, determining that the sample audio signal is a valid audio signal; otherwise, determining that the sample audio signal is an invalid audio signal; A thunder signal test is performed on the valid audio signal to determine whether the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal.

4. The method for efficiently identifying thunder signals based on a microphone array according to claim 1, wherein: The audio signal includes a plurality of sub-signals, and the step of performing data preprocessing on the sample audio signal based on the logarithmic filter bank feature to obtain a feature matrix includes: Performing windowed sampling on the sub-signal to obtain a second windowed signal; Performing frame sampling on the second windowed signal to obtain a plurality of second frames of data; Performing feature processing on the second frame data based on the logarithmic filter bank feature to obtain a feature matrix of the sub-signal; The global average energy and the energy variance in the time dimension of the sub-signal are obtained based on the characteristic matrix of the sub-signal; wherein the characterization formula of the global average energy includes: Among them, E (c) is the global average energy of the cth sub-signal; T is the number of the second frame data; N is the number of sampling points contained in the second frame data; is the logarithmic filter feature of the j-th Mel filter in the t-th second frame data of the c-th sub-signal, used to characterize the feature matrix of the sub-signal; Among them, V (c) is the energy variance of the c-th sub-signal in the time dimension; is the average energy of the t-th second frame data in the c-th sub-signal; is the average energy value of the cth sub-signal; A weighting coefficient of the sub-signal is obtained based on the global average energy of the sub-signal and the energy variance in the time dimension. The characterization formula of the weighting coefficient includes: Among them, α (c) is the weighting coefficient of the cth sub-signal; E (l) is the global average energy of the lth sub-signal; V (l) is the energy variance of the lth sub-signal in the time dimension; The feature matrices of the plurality of sub-signals are weighted and summed based on the weight coefficients of the sub-signals to obtain a fused feature matrix, which is the feature matrix of the sample audio signal. The characterization formula of the fused feature matrix includes: Among them, X fused is the fusion feature matrix; X (c) is the characteristic matrix of the c-th sub-signal.

5. The method for efficiently identifying thunder signals based on a microphone array according to claim 4, characterized in that: The step of performing feature processing on the second frame data based on the logarithmic filter bank feature to obtain the feature matrix of the sub-signal includes: Perform a fast Fourier transform on the second frame data to obtain a power spectrum of the second frame data, where a characterization formula of the power spectrum includes: P t =[P t (1),P t (2),...,P t (k)]; P t (k)=|FFT(frame[t])[k]| 2 ,k=1,2,…,K; Among them, P t is the power spectrum of the t-th second frame data; P t (k) is the power of the t-th second frame data at the k-th frequency component; FFT is the fast Fourier transform; |·| 2 is the square of the complex modulus; K is the total number of the frequency components; frame is the second frame data; The power spectrum is weighted filtered based on a Mel filter bank to obtain Mel filter bank energy, where the Mel filter bank includes multiple Mel filters, and the Mel filter bank energy includes output energies of the multiple Mel filters on the power spectrum. The representation formula of the output energy includes: Among them, E t (j) is the output energy of the power spectrum of the j-th Mel filter for the t-th second frame data; H j (k) is the response of the h-th Mel filter at frequency component k; J is the number of Mel filters; Performing a logarithmic operation on the output energy of the Mel filter on the second frame of data to obtain a logarithmic filter bank feature of the second frame of data, wherein a characterization formula of the logarithmic filter bank feature includes: x tj =ln(E t (j)+∈); Among them, x tj is the logarithmic filter feature of the jth Mel filter in the tth second frame data; ∈ is a smoothing constant to avoid zero values ​​in the logarithmic function; A feature matrix of the sub-signal is obtained based on logarithmic filter bank features of a plurality of second frame data in the sub-signal. A characterization formula of the feature matrix of the sub-signal includes: Among them, X (c) is the characteristic matrix of the cth sub-signal; R T×N is a matrix with T rows and N columns.

6. The method for efficiently identifying thunder signals based on a microphone array according to claim 1, characterized in that: The thunder signal recognition model includes: A plurality of sequentially connected convolution blocks, each of which includes a sequentially connected convolution layer, a normalization layer, an activation layer, and an average pooling layer; The classification module includes a fully connected layer and an activation layer connected in sequence.

7. The method for efficiently identifying thunder signals based on a microphone array according to claim 1 or 4, characterized in that: Windowed sampling is performed based on a Hamming window.

8. An efficient thunder signal recognition system based on a microphone array, characterized in that: include: A data acquisition module (10), the data acquisition module (10) is used to acquire a sample audio signal, the audio signal including a plurality of sub-signals; a label classification module (20), the label classification module (20) being used to determine a label type of the sample audio signal, the label type including a thunder signal and a non-thunder signal; The step of determining the label type of the sample audio signal comprises: Obtaining a short-time frame average energy standard deviation of the sub-signal; the step of obtaining the short-time frame average energy standard deviation of the sub-signal includes: Performing window sampling on the sub-signal to obtain a first windowed signal; Performing frame sampling on the first windowed signal to obtain a plurality of first frames of data; Obtain a short-time energy average value of a plurality of first frames of data, wherein a characterization formula of the short-time energy average value includes: Among them, E frame is the short-time energy average value; N is the length of the first frame of data, which is used to represent the number of sampling points of the windowed sampling contained in the first frame of data; x[n] is the amplitude value of the sub-signal at the nth sampling point, and n is the sampling point index; w[n] is the window function value applied to the nth sampling point; The short-time energy average value is normalized, and the short-time frame average energy standard deviation is obtained based on the normalized short-time energy average value in the plurality of first frames of data. The characterization formula of the short-time frame average energy standard deviation includes: Wherein, STE_SD is the standard deviation of the short-time frame average energy; M is the number of the first frame data; E i is the normalized short-time energy average value of the i-th first frame data; μ is the average of multiple normalized short-time energy average values; Determine a threshold value of the standard deviation of the short-time frame average energy; In response to a maximum value of the short-time frame average energy standard deviation being greater than or equal to a threshold value of the short-time frame average energy standard deviation among the multiple sub-signals, determining that the sample audio signal is a valid audio signal; otherwise, determining that the sample audio signal is an invalid audio signal; Performing a thunder signal test on the valid audio signal, and determining that the label type of the valid audio signal that passes the test is a thunder signal, otherwise the label type is a non-thunder signal; A feature matrix module (30), the feature matrix module (30) is used to perform data preprocessing on the sample audio signal based on the logarithmic filter bank characteristics to obtain a feature matrix; A model training module (40) is used to obtain a basic model based on a deep convolutional neural network, and use the feature matrix as input to the basic model, and the label type corresponding to the feature matrix as output of the basic model, to perform model training on the basic model to obtain a thunder signal recognition model; A result prediction module (50) is used to obtain the audio signal to be tested, input the thunder signal recognition model, and obtain the thunder signal recognition result of the audio signal to be tested.

Images