Forest fire audio recognition method based on BP neural network, medium and product

Abstract

The invention discloses a forest fire audio recognition method based on a BP neural network, a medium and a product in the technical field of forest fire recognition. The method comprises the following steps: acquiring forest fire audio data to be detected; extracting a Mel-frequency cepstrum coefficient (MFCC) based on the frequency domain features of the forest fire audio data to be detected; inputting the Mel-frequency cepstral coefficient MFCC into a pre-trained forest fire audio recognition model based on a BP neural network, and outputting a forest fire category; wherein the forest fire category comprises a forest fire and a non-forest fire. Compared with a traditional visual monitoring means, the method can cover a wider forest area, in the same monitoring range, in addition, model training based on the BP neural network can efficiently process a large amount of audio data and rapidly output an identification result, the efficiency and accuracy of forest fire identification are further improved, and the method is suitable for popularization and application. The problems of limited coverage range, high cost and low accuracy in the existing forest fire identification technology are solved.

Description

Technical Field

[0001] The invention relates to a forest fire audio recognition method, a medium and a product based on a BP neural network, and belongs to the technical field of forest fire recognition. Background Art

[0002] As an indispensable "oxygen factory" and natural regulator on the earth, forests are vital to the human living environment. However, in recent years, due to the continuous expansion of human production activities and the frequent occurrence of natural disasters, the forest area is decreasing sharply. Among these destructive factors, the harm of forest fires is particularly prominent. Forest fires are highly destructive, sudden and easy to spread. Once they occur, they are often difficult to extinguish quickly, causing huge harm and losses to the ecosystem and human society. In order to meet this challenge, forest fire identification technology came into being, aiming to provide the possibility of early detection of forest fires and timely fire-fighting measures, thereby minimizing the losses caused by fires. Research in the field of forest fire identification technology has important practical application value.

[0003] The occurrence of forest fires usually goes through three stages: pyrolysis non-combustion stage, gas combustion stage and charcoal combustion stage. In the pyrolysis non-combustion stage, combustible materials will produce a lot of smoke under the action of external heat sources, which makes it possible to detect fires at an early stage. Based on this feature, the existing forest fire detection technology mainly collects forest environment data by installing equipment such as smoke sensors, flame sensors and temperature and humidity sensors. In addition, remote sensing technologies such as resolution spectrometers are used to extract smoke and burning traces from the fire scene and identify the fire point. On the basis of smoke identification, hybrid pixel decomposition and sub-pixel positioning technology are also used to determine the scope of forest fire smoke. At the same time, with the development of computer science and technology, technologies such as image processing, computer vision, and artificial intelligence have also been gradually applied to the field of fire identification, further improving the accuracy and efficiency of fire detection. The above technologies mainly collect images through surveillance cameras and use image processing and pattern recognition methods for analysis to determine whether a fire has occurred.

[0004] Although the existing forest fire identification technology has made some progress, it still has many defects. First, traditional smoke detection technology is sensitive to external environmental interference, and the types of recognition scenes are limited, resulting in high false alarm and missed detection rates. Secondly, although the fire detection technology based on image processing has achieved certain results, the detection method is relatively single, the response time is long, the reliability is low, and it cannot provide effective location information for subsequent emergency fire handling. In addition, the deployment cost of the above technology is high, and it is difficult to widely cover all forest areas. Although the fire prediction and detection method based on deep learning has powerful data classification, recognition and regression prediction capabilities, it requires a lot of data support, and when smoke data is missing or the data source is limited, the recognition ability will be significantly weakened. Therefore, the existing technology still needs to be further improved and perfected to better meet the challenges of forest fires.

[0005] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgement or any form of suggestion that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to solve the problems of limited coverage, high cost and low accuracy in the existing forest fire identification technology.

[0007] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0008] In a first aspect, the present invention provides a forest fire audio recognition method based on a BP neural network, comprising:

[0009] Obtain forest fire audio data to be detected;

[0010] Extracting Mel-frequency cepstral coefficients (MFCC) based on the frequency domain characteristics of the forest fire audio data to be detected;

[0011] Input the Mel frequency cepstral coefficient MFCC into a pre-trained forest fire audio recognition model based on BP neural network, and output the forest fire category;

[0012] Among them, the forest fire categories include forest fires and non-forest fires.

[0013] Furthermore, before the Mel-frequency cepstral coefficients MFCC are input into the pre-trained forest fire audio recognition model based on the BP neural network, the method further includes marking the audio samples belonging to forest fires in the forest fire audio data to be detected as 1, and marking the audio samples belonging to non-forest fires as 0.

[0014] Furthermore, the training method of the forest fire audio recognition model based on the BP neural network includes:

[0015] Based on the frequency domain characteristics of the acoustic forest monitoring dataset after random sampling, the Mel frequency cepstral coefficients (MFCC) are extracted.

[0016] The Mel-frequency cepstral coefficients (MFCCs) are used as input features and fed into the forest fire audio recognition model. The output is the prediction result that the audio sample belongs to the forest fire category.

[0017] The difference between the predicted result and the true label is calculated using the mean square error loss function to obtain a loss value;

[0018] The weight of the forest fire audio recognition model based on the BP neural network is updated using a gradient descent algorithm until the loss value is less than a preset allowable range.

[0019] Furthermore, the method for acquiring the frequency domain features of the randomly sampled acoustic forest monitoring data set includes:

[0020] Perform Fourier transform on the randomly sampled acoustic forest monitoring dataset to obtain the frequency domain signal;

[0021] The frequency domain features in the frequency domain signal are extracted, including mean, standard deviation, skewness, kurtosis, maximum, minimum, peak, root mean square, amplitude factor, waveform factor, pulse factor, crest factor and energy.

[0022] Furthermore, based on the frequency domain characteristics of the randomly sampled acoustic forest monitoring dataset, the Mel frequency cepstral coefficients MFCC are extracted, including:

[0023] Pre-emphasize the randomly sampled acoustic forest monitoring data set to obtain a pre-emphasized time domain signal;

[0024] After performing framing and windowing operations on the pre-emphasized time domain signal, a fast Fourier transform is performed to obtain the spectrum energy of each frame of the frequency domain signal;

[0025] The spectral energy of each frame of frequency domain signal is multiplied by a set of Mel filters, and then subjected to discrete cosine transform to obtain the Mel frequency cepstrum coefficient MFCC.

[0026] Furthermore, the spectral energy of each frame of the frequency domain signal is multiplied by a group of Mel filters, and then subjected to discrete cosine transformation to obtain Mel frequency cepstral coefficients MFCC, wherein the relationship between the Mel frequency cepstral coefficients MFCC and the frequency of each frame of the frequency domain signal is expressed as:

[0027] (1);

[0028] In the formula, Indicates the frequency of each frame of frequency domain signal The corresponding value on the Mel scale, where represents the frequency of each frame of the frequency domain signal The frequency unit is Hertz, Represents the base 10 logarithmic function.

[0029] Further, the performing framing and windowing operations on the pre-emphasized time domain signal includes:

[0030] The pre-emphasized time domain signal is divided into a plurality of short time frame signals, and each short time frame signal is multiplied by a Hamming window function.

[0031] Furthermore, the training method of the forest fire audio recognition model based on the BP neural network includes:

[0032] The extracted Mel-frequency cepstral coefficients (MFCCs) are used as input features and fed into the constructed forest fire audio recognition model, and the output is the prediction result that the audio sample belongs to the forest fire category.

[0033] The difference between the predicted result and the true label is calculated using the mean square error loss function to obtain a loss value;

[0034] The weight of the forest fire audio recognition model based on the BP neural network is updated using a gradient descent algorithm until the loss value is less than a preset allowable range.

[0035] Furthermore, the mean square error loss function is expressed as:

[0036] (2);

[0037] In the formula, represents the loss value, represents the total number of audio samples in the acoustic forest monitoring dataset after random sampling, represents the first audio samples, represents the first The predicted value of audio samples, represents the first The actual value of audio samples, Represents weight.

[0038] In a second aspect, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, wherein the computer program / instruction, when executed by a processor, implements the steps of the method described in the first aspect.

[0039] In a third aspect, the present invention provides a computer program product, comprising a computer program / instruction, characterized in that when the computer program / instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The present invention effectively broadens the coverage of audio recognition by acquiring the audio data of forest fires to be detected and using a pre-trained BP neural network model for recognition. Compared with traditional visual monitoring methods, the present invention can cover a wider range of forest areas. Within the same monitoring range, the number of monitoring nodes required for the present invention to recognize forest fire audio data is significantly reduced, thereby reducing overall construction and maintenance costs. In addition, model training based on BP neural networks can efficiently process large amounts of audio data and quickly output recognition results, further improving the efficiency and accuracy of fire recognition, and solving the problems of limited coverage, high cost, and low accuracy in existing forest fire recognition technologies.

[0042] 2. The present invention does not rely on the lighting intensity or other light sources at the monitoring site. Even in the case of insufficient lighting or visual monitoring failure, the forest fire audio recognition method based on the BP neural network can still accurately identify the fire category, providing strong support for early warning and rapid response to forest fires. The present invention also extracts Mel frequency cepstral coefficients MFCC through frequency domain features for model training, further improving the sensitivity and recognition accuracy of the forest fire audio recognition model based on the BP neural network to forest fire audio data.

[0043] 3. The present invention processes the randomly sampled acoustic forest monitoring data set, including marking forest fire audio samples as 1 and non-forest fire audio samples as 0, so that the forest fire audio recognition model based on BP neural network can accurately distinguish forest fires from non-forest fires, and maintain a high recognition rate even in complex forest environments. This efficient and accurate fire recognition capability provides strong support for early warning and rapid response to forest fires.

[0044] 4. During the training process of the forest fire audio recognition model based on the BP neural network, the present invention calculates the difference between the predicted result and the true label through the mean square error loss function, and continuously updates the model weight through the gradient descent algorithm until the loss value reaches the preset range. This not only optimizes the parameter configuration of the forest fire audio recognition model based on the BP neural network, but also significantly enhances the generalization performance and robustness of the forest fire audio recognition model based on the BP neural network. Even in the face of audio samples that have never appeared in the training set or complex and changeable forest environments, the model can still maintain stable recognition performance, providing more reliable technical support for forest fire monitoring and prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The figure is a schematic diagram of the overall architecture of the forest fire audio recognition model provided by an embodiment of the present invention;

[0046] Figure 2 FIG. 1 is a schematic diagram of a time domain signal waveform of audio data provided by an embodiment of the present invention;

[0047] Figure 3 The figure shows a schematic diagram of a frequency domain signal waveform obtained by Fourier transform FFT of audio data in the FSC22 data set provided by an embodiment of the present invention;

[0048] Figure 4 FIG. 1 is a flow chart of a forest fire audio recognition method based on a BP neural network provided by an embodiment of the present invention;

[0049] Figure 5 The figure shows a basic node schematic diagram of a BP neural network provided by an embodiment of the present invention;

[0050] Figure 6 The figure shows a schematic diagram of scattering distribution comparison of MFCC coefficients of audio data and frequency domain signal verification results provided by an embodiment of the present invention;

[0051] Figure 7 It is a schematic diagram showing the comparison of the maximum value and average value of the frequency domain characteristics of audio data reflected by the MFCC coefficient and the frequency domain signal provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0053] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0054] Example 1

[0055] like Figure 1 As shown, this embodiment introduces a forest fire audio recognition method based on BP neural network, including:

[0056] Get the forest fire audio data to be detected.

[0057] The forest fire audio data to be detected includes sounds from a forest fire scene or sounds from a non-forest fire scene.

[0058] Based on the frequency domain characteristics of the forest fire audio data to be detected, Mel-frequency cepstral coefficients MFCC are extracted.

[0059] The Mel-frequency cepstral coefficients MFCC are input into a pre-trained forest fire audio recognition model based on a BP neural network to output the forest fire category.

[0060] BP neural network is a multi-layer feedforward neural network that can learn the complex mapping relationship between input and output through back propagation algorithm training. The basic nodes of BP neural network are as follows: Figure 5 As shown, Figure 5 As shown, w represents the weight of each node, and A represents the specific value of each node.

[0061] The forest fire audio recognition model based on BP neural network has powerful nonlinear mapping and self-learning capabilities. It can extract useful feature information from complex audio data to achieve accurate classification. Accurate classification results can reduce false alarms and missed alarms and improve the efficiency and accuracy of fire monitoring.

[0062] The forest fire categories include forest fires and non-forest fires.

[0063] The classification results output by the forest fire audio recognition model based on BP neural network include forest fire and non-forest fire, which can be directly used in practical applications such as fire warning, monitoring and emergency response.

[0064] The randomly sampled acoustic forest monitoring dataset is used to ensure the diversity and representativeness of the audio data. The Mel-frequency cepstral coefficient (MFCC) can reflect the spectral characteristics of the audio signal and the rhythmic characteristics of the speech.

[0065] The present invention extracts Mel-frequency cepstral coefficients MFCC through frequency domain features to train a forest fire audio recognition model based on a BP neural network, thereby further improving the sensitivity and recognition accuracy of the forest fire audio recognition model based on the BP neural network to forest fire audio data.

[0066] The hazards of forest fires mainly include devastating damage to forest resources, destroying the balance between forest ecosystems and human living environment, endangering people's lives and safety, and causing serious property and resource losses to the country and the people. Real-time monitoring and timely warning of fires have become particularly important. Audio monitoring technology has unique characteristics compared to other types of monitoring methods. On the one hand, audio has a wide coverage range, requires fewer nodes, and naturally has lower costs. On the other hand, audio is not easily blocked and is not affected by the lighting intensity and other light sources at the monitoring site. There are basically no blind spots in the monitoring area.

[0067] Therefore, the method provided by the present invention has positive significance for real-time monitoring and timely early warning of fires, provides effective suggestions, and helps to reduce the harm caused by forest fires to a certain extent.

[0068] Example 2

[0069] like Figure 4 As shown, based on the same inventive concept as Example 1, this embodiment introduces a forest fire audio recognition method based on BP neural network, which specifically includes the following steps:

[0070] Step 1: Obtain the forest fire audio data to be detected.

[0071] Step 1.1: Get the FSC22 dataset.

[0072] This implementation uses the FSC22 dataset containing audio data from forests. The FSC22 dataset is a new benchmark dataset for acoustic-based forest monitoring. The sampling rate of the FSC22 dataset is standardized to 44,100 Hz sampling rate, 16-bit depth, and stereo channel configuration. The FSC22 dataset includes 2025 5-second audio clips from the online audio database Freesound. In the FSC22 dataset, all sound events are divided into 6 main categories, which are further subdivided into 34 subcategories. Each audio clip is manually annotated and verified to ensure the quality of the dataset.

[0073] Step 1.2: Analyze the time domain features of the audio data in the FSC22 dataset.

[0074] In this implementation, we first analyze the time domain characteristics of the audio data in the FSC22 dataset, such as Figure 2As shown, the schematic diagram of the time domain signal waveform of audio data describes the time domain signal waveform of some audio samples, wherein the sampling time of a single audio sample is 10 seconds and the sampling frequency FS is 441001 Hz. It can be seen from the time domain signal waveform of the audio sample that the waveform of the time domain signal is singular, and there is no obvious other type of waveform, indicating that the signal-to-noise ratio of the FSC22 data set used in this embodiment is relatively high.

[0075] Step 1.3: Perform Fourier transform on the audio data in the FSC22 dataset to analyze the frequency domain features.

[0076] In this implementation, the purpose of analyzing frequency domain features is to extract spectral information for analysis, such as the amplitude or phase of different frequencies in the spectral information. The spectral characteristics of bioelectric signals change under different experimental conditions, thereby encoding task-related details and clarifying the signal generation mechanism. Any bioelectric signal can be represented as a combination of sinusoidal oscillations of different frequencies. Therefore, frequency domain analysis helps to calculate spectral information, which is crucial for extracting reliable features. The method provided by the present invention can clearly, accurately and comprehensively depict the target forest area through targeted feature extraction based on frequency domain features, showing strong anti-interference ability and significantly improving target positioning and extraction efficiency.

[0077] The advantages of frequency domain features are:

[0078] (1) It is more feasible to identify signal changes or patterns rather than just observing time domain information;

[0079] (2) Provide stronger discrimination capabilities than time-domain features, which depend on the nature of the signal;

[0080] (3) Since only the spectral information of certain frequencies is used, a low-dimensional representation is obtained;

[0081] (4) Due to the fast and efficient spectral information calculation algorithm, the computational cost is reduced;

[0082] (5) It is not only possible to utilize the spectrum information of a single frequency, but also the spectrum information of a frequency band, which may be more convenient in some applications;

[0083] (6) Ability to pre-select frequencies for calculating spectral information for feature extraction based on the characteristics of the signal and application.

[0084] In this embodiment, the audio data in the FSC22 data set is converted into frequency domain signals through Fourier transform, which provides another perspective for analyzing and observing frequency domain signals.

[0085] When the digitizer samples the waveform, the audio samples are converted to discrete values. Because of the transformation, the Fourier transform cannot be directly applied to these data. In this embodiment, the discrete Fourier transform DFT is used instead of the fast Fourier transform FFT to produce a discrete form of the frequency domain signal.

[0086] The frequency domain signal obtained from the time domain signal by Fourier transform FFT in the FSC22 data set is as follows: Figure 3 As shown, the sampling time of a single audio sample is 0.5 seconds, and the 13 frequency domain features obtained are: mean, standard deviation, skewness, kurtosis, maximum, minimum, peak, root mean square, amplitude factor, shape factor, impulse factor, crest factor and energy.

[0087] Step 1.4: Analyze the sensitivity of Mel frequency cepstral coefficients MFCC to sound waves of different frequencies.

[0088] Mel-frequency cepstral coefficients (MFCCs) are widely used in speech recognition and speaker identification. Studies have shown that the human ear is particularly sensitive to sound waves with frequencies between 200Hz and 5000Hz, and that frequencies between 200Hz and 5000Hz significantly affect speech clarity. The masking effect states that loud sounds mask low-loud sounds and affect auditory perception. Low-frequency sounds travel farther on the basilar membrane of the cochlea, making them suitable as speech features. Mel-frequency cepstral coefficients (MFCCs) use auditory models without assumptions or restrictions, are applicable to a variety of signal types, and exhibit enhanced robustness, outperforming linear prediction cepstral coefficients (LPCCs) based on vocal cord models. Mel-frequency cepstral coefficients (MFCCs) are developed based on a nonlinear human ear auditory model, have a logarithmic relationship with signals of different frequencies, and can maintain strong recognition performance even at low signal-to-noise ratios.

[0089] Step 1.5: Analyze the evaluation metrics.

[0090] This embodiment uses the correlation coefficient To evaluate the degree of fit between the forest fire audio recognition model based on BP neural network and the data, although the correlation coefficient is usually applied to linear regression analysis, in the scenario provided by the present invention, the correlation coefficient It can also measure the proportion of variance in the dependent variable explained by the forest fire audio recognition model based on the BP neural network, the correlation coefficient The value range is from 0 to 1.

[0091] Among them, when the correlation coefficient The closer it is to 1, the better the fit between the forest fire audio recognition model based on the BP neural network and the data is, and it can explain most of the variability in the dependent variable. The closer it is to 0, the lower the fitting effect of the forest fire audio recognition model based on the BP neural network on the data, and the less able it is to effectively explain the variability in the dependent variable.

[0092] Step 2: input the Mel frequency cepstral coefficient MFCC into a pre-trained forest fire audio recognition model based on BP neural network, and output the forest fire category;

[0093] The forest fire categories include forest fires and non-forest fires.

[0094] The training method of the forest fire audio recognition model based on the BP neural network comprises:

[0095] Based on the frequency domain characteristics of the acoustic forest monitoring dataset after random sampling, the Mel frequency cepstral coefficients (MFCC) are extracted.

[0096] The Mel-frequency cepstral coefficients (MFCCs) are used as input features and fed into the forest fire audio recognition model. The output is the prediction result that the audio sample belongs to the forest fire category.

[0097] The difference between the predicted result and the true label is calculated using the mean square error loss function to obtain a loss value;

[0098] Among them, the mean square error loss function is expressed as:

[0099] (1);

[0100] In the formula, represents the loss value, represents the total number of audio samples in the acoustic forest monitoring dataset after random sampling, represents the first audio samples, represents the first The predicted value of audio samples, represents the first The actual value of audio samples, Represents weight.

[0101] The weight of the forest fire audio recognition model based on the BP neural network is updated using a gradient descent algorithm until the loss value is less than a preset allowable range.

[0102] In this embodiment, before the Mel frequency cepstral coefficient MFCC is input into the pre-trained forest fire audio recognition model based on the BP neural network, it also includes marking the audio samples belonging to forest fires in the forest fire audio data to be detected as 1, and marking the audio samples belonging to non-forest fires as 0, and dividing them into a training set and a test set in a ratio of 7:3.

[0103] In this embodiment, the method for acquiring the frequency domain features of the randomly sampled acoustic forest monitoring data set includes:

[0104] Perform Fourier transform on the training set to obtain the frequency domain signal;

[0105] The frequency domain features in the frequency domain signal are extracted, including mean, standard deviation, skewness, kurtosis, maximum, minimum, peak, root mean square, amplitude factor, waveform factor, pulse factor, crest factor and energy.

[0106] In this embodiment, based on the frequency domain characteristics of the acoustic forest monitoring data set after random sampling, the Mel frequency cepstral coefficients MFCC are extracted, including:

[0107] In this embodiment, due to the non-stationary but short-term stationary characteristics of the speech signal, it needs to be divided into frames for processing. Each frame contains N audio samples, such as 256 or 512 audio samples, and lasts for about 20 to 30 milliseconds. To ensure smooth transition between frames, an overlapping area is set between adjacent frames, containing M audio samples, which is about half or one third of the total number of audio samples N per frame.

[0108] In this embodiment, the sampling frequency of the audio data is set to 44100 Hz, the frame length is 17640, and the frame offset is 8820.

[0109] Pre-emphasize the randomly sampled acoustic forest monitoring data set to enhance the high-frequency time domain signal and reduce the low-frequency time domain signal to obtain the pre-emphasized time domain signal;

[0110] After performing framing and windowing operations on the pre-emphasized time domain signal, a fast Fourier transform is performed to obtain the spectrum energy of each frame of the frequency domain signal;

[0111] In this embodiment, performing framing and windowing operations on the pre-emphasized time domain signal includes dividing the pre-emphasized time domain signal into a plurality of short time frame signals, and multiplying each short time frame signal by a Hamming window function in order to reduce spectrum leakage.

[0112] The purpose of pre-emphasis is to enhance high-frequency time domain signals, reduce low-frequency time domain signals, make the signal spectrum flatter, and ensure consistent signal-to-noise ratio across the entire frequency range from low to high frequencies. It also compensates for the suppression of high-frequency components in the speech signal by the sound source, thereby eliminating the influence of vocal cords and lips during production. In addition, pre-emphasis can also highlight high-frequency resonance peaks.

[0113] The spectral energy of each frame of frequency domain signal is multiplied by a set of Mel filters, and then subjected to discrete cosine transform DCT to obtain the Mel frequency cepstral coefficient MFCC. The relationship between the Mel frequency cepstral coefficient MFCC and the frequency of each frame of frequency domain signal is expressed as:

[0114] (2);

[0115] In the formula, Indicates the frequency of each frame of frequency domain signal The corresponding value on the Mel scale, where represents the frequency of each frame of the frequency domain signal The frequency unit is Hertz, Represents the base 10 logarithmic function.

[0116] Discrete cosine transform DCT is mainly used for data or image compression, and can convert signals from the spatial domain to the frequency domain, and has excellent decorrelation performance. In this embodiment, discrete cosine transform DCT is used to convert frame frequency domain signals into Mel-frequency cepstral coefficients MFCC.

[0117] Step 3: Use the correlation coefficient As evaluation indicators, frequency domain features and Mel-frequency cepstral coefficients (MFCC) in the test set were extracted as input features, and a forest fire audio recognition model based on BP neural network was trained for verification.

[0118] The frequency domain features are divided according to the sampling frequency of the audio samples, and the Mel frequency cepstral coefficients MFCC are frame-lengthed and frame-shifted according to the number of samples in each frame of the frequency domain signal to ensure that the audio data lengths of the two input feature sets are consistent.

[0119] In this embodiment, 936 samples are used for frequency domain feature verification, and 846 samples are used for Mel frequency cepstral coefficient MFCC verification. The ratio of audio samples belonging to forest fire and non-forest fire is 3:1. In frequency domain feature verification, the training set has 700 audio samples and the test set has 236 audio samples; in Mel frequency cepstral coefficient MFCC verification, the training set has 624 audio samples and the test set has 240 audio samples. Each group of audio samples is tested 1000 times. Among them, Figure 6 The figure shows a schematic diagram of scattering distribution comparison of MFCC coefficients of audio data and frequency domain signal verification results provided by an embodiment of the present invention; Figure 7It is a schematic diagram showing the comparison of the maximum value and average value of the frequency domain characteristics of audio data reflected by the MFCC coefficient and the frequency domain signal provided by an embodiment of the present invention.

[0120] like Figure 6 As shown, using Mel-frequency cepstral coefficients (MFCCs) as input features yields significantly better and more stable performance without occasional increased errors.

[0121] The results of randomized experiments are as follows Figure 7 As shown in Figure 7, the Mel frequency cepstral coefficients MFCC can obviously produce better and more stable performance without occasional increased errors. As shown in Figure 7, red represents the frequency domain signal, blue represents the MFCC coefficients, and both the maximum and average values ​​are described. The two columns on the left are the maximum values, and the two columns on the right are the average values. By comparing the maximum value and the average value, it can be concluded that the Mel frequency cepstral coefficients MFCC can produce significantly superior performance.

[0122] In summary, using Mel-frequency cepstral coefficients MFCC as input features to train a forest fire audio recognition model based on a BP neural network can produce significantly superior performance.

[0123] Example 3

[0124] With the same inventive concept as other embodiments, this embodiment introduces a computer-readable storage medium on which computer instructions are stored, characterized in that when the computer instructions are executed by a processor, the steps of the method of the above-mentioned embodiment 1 or 2 are implemented.

[0125] Example 4

[0126] Similar to the inventive concept of other embodiments, this embodiment introduces a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the method of the above-mentioned embodiment 1 or 2 are implemented.

[0127] In summary, the present invention effectively broadens the coverage of audio recognition by acquiring the audio data of forest fires to be detected and using a pre-trained BP neural network model for recognition. Compared with traditional visual monitoring methods, the present invention can cover a wider range of forest areas. Within the same monitoring range, the number of monitoring nodes required for the present invention to recognize forest fire audio data is significantly reduced, thereby reducing overall construction and maintenance costs. In addition, model training based on BP neural networks can efficiently process large amounts of audio data and quickly output recognition results, further improving the efficiency and accuracy of fire recognition, and solving the problems of limited coverage, high cost, and low accuracy in existing forest fire recognition technologies.

[0128] The present invention does not rely on the lighting intensity or other light sources at the monitoring site. Even in the case of insufficient lighting or visual monitoring failure, the forest fire audio recognition method based on the BP neural network can still accurately identify the fire category, providing strong support for early warning and rapid response to forest fires. The present invention also combines frequency domain features and Mel frequency cepstral coefficients MFCC for model training, further improving the sensitivity and recognition accuracy of the forest fire audio recognition model based on the BP neural network to forest fire audio data.

[0129] The present invention processes the randomly sampled acoustic forest monitoring data set, including marking forest fire audio samples as 1 and non-forest fire audio samples as 0, so that the forest fire audio recognition model based on BP neural network can accurately distinguish forest fires from non-forest fires, and maintain a high recognition rate even in complex forest environments. This efficient and accurate fire recognition capability provides strong support for early warning and rapid response to forest fires.

[0130] During the training process of the forest fire audio recognition model based on the BP neural network, the present invention calculates the difference between the predicted result and the true label through the mean square error loss function, and continuously updates the model weight through the gradient descent algorithm until the loss value reaches the preset range. This not only optimizes the parameter configuration of the forest fire audio recognition model based on the BP neural network, but also significantly enhances the generalization performance and robustness of the forest fire audio recognition model based on the BP neural network. Even when faced with audio samples that have never appeared in the training set or complex and changeable forest environments, the model can still maintain stable recognition performance, providing more reliable technical support for forest fire monitoring and prevention and control.

[0131] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0133] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0134] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0135] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A forest fire audio recognition method based on BP neural network, characterized in that: include: Obtain forest fire audio data to be detected; Extracting Mel-frequency cepstral coefficients (MFCC) based on the frequency domain characteristics of the forest fire audio data to be detected; Input the Mel frequency cepstral coefficient MFCC into a pre-trained forest fire audio recognition model based on BP neural network, and output the forest fire category; Among them, the forest fire categories include forest fires and non-forest fires.

2. The forest fire audio recognition method based on BP neural network according to claim 1 is characterized in that: Before inputting the Mel-frequency cepstral coefficients MFCC into the pre-trained forest fire audio recognition model based on the BP neural network, the method also includes marking the audio samples belonging to forest fires in the forest fire audio data to be detected as 1, and marking the audio samples belonging to non-forest fires as 0.

3. The forest fire audio recognition method based on BP neural network according to claim 1 is characterized in that: The training method of the forest fire audio recognition model based on the BP neural network comprises: Based on the frequency domain characteristics of the acoustic forest monitoring dataset after random sampling, the Mel frequency cepstral coefficients (MFCC) are extracted. The Mel-frequency cepstral coefficients (MFCCs) are used as input features and fed into the forest fire audio recognition model. The output is the prediction result that the audio sample belongs to the forest fire category. The difference between the predicted result and the true label is calculated using the mean square error loss function to obtain a loss value; The weight of the forest fire audio recognition model based on the BP neural network is updated using a gradient descent algorithm until the loss value is less than a preset allowable range.

4. The forest fire audio recognition method based on BP neural network according to claim 3 is characterized in that: The method for acquiring the frequency domain features of the randomly sampled acoustic forest monitoring data set comprises: Perform Fourier transform on the randomly sampled acoustic forest monitoring dataset to obtain the frequency domain signal; The frequency domain features in the frequency domain signal are extracted, including mean, standard deviation, skewness, kurtosis, maximum, minimum, peak, root mean square, amplitude factor, waveform factor, pulse factor, crest factor and energy.

5. The forest fire audio recognition method based on BP neural network according to claim 3 is characterized in that: Based on the frequency domain features of the randomly sampled acoustic forest monitoring dataset, the Mel frequency cepstral coefficients (MFCCs) are extracted, including: Pre-emphasize the randomly sampled acoustic forest monitoring data set to obtain a pre-emphasized time domain signal; After performing framing and windowing operations on the pre-emphasized time domain signal, a fast Fourier transform is performed to obtain the spectrum energy of each frame of the frequency domain signal; The spectral energy of each frame of frequency domain signal is multiplied by a set of Mel filters, and then subjected to discrete cosine transform to obtain the Mel frequency cepstrum coefficient MFCC.

6. The forest fire audio recognition method based on BP neural network according to claim 5 is characterized in that: The spectral energy of each frame of frequency domain signal is multiplied by a group of Mel filters, and then subjected to discrete cosine transformation to obtain Mel frequency cepstral coefficients MFCC, wherein the relationship between the Mel frequency cepstral coefficients MFCC and the frequency of each frame of frequency domain signal is expressed as: （1）； In the formula, Indicates the frequency of each frame of frequency domain signal The corresponding value on the Mel scale, where represents the frequency of each frame of the frequency domain signal The frequency unit is Hertz, Represents the base 10 logarithmic function.

7. The forest fire audio recognition method based on BP neural network according to claim 5 is characterized in that: The framing and windowing operations are performed on the pre-emphasized time domain signal, comprising: The pre-emphasized time domain signal is divided into a plurality of short time frame signals, and each short time frame signal is multiplied by a Hamming window function.

8. The forest fire audio recognition method based on BP neural network according to claim 3 is characterized in that: The mean square error loss function is expressed as: （2）； In the formula, represents the loss value, represents the total number of audio samples in the acoustic forest monitoring dataset after random sampling, represents the first audio samples, represents the first The predicted value of audio samples, represents the first The actual value of audio samples, Represents weight.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images