A method for detecting electric fish signals based on mel-frequency cepstrum and multi-layer perception

Abstract

The application discloses a kind of electric fish signal detection methods based on mel cepstrum and multilayer perception machine, belong to electric fish signal detection field, comprising: first, the sound emission signal generated by underwater discharge is collected, pre-processing is carried out by pre-emphasis, framing and windowing to improve signal-to-noise ratio;Subsequently, the mel frequency cepstrum coefficient feature of signal is extracted, and is optimized by first-order and second-order difference calculation, and is fused to form a feature parameter set with static and dynamic characteristics;Finally, the feature parameter and detection voltage value are input into multilayer perception machine neural network for nonlinear fitting, and the high-precision electric signal strength is reconstructed.The application effectively utilizes the sound emission signal feature, and significantly improves the sensitivity and confidence of underwater electric fish signal detection.

Description

Technical Field

[0001] This invention belongs to the field of electric fish signal detection, and particularly relates to a method for electric fish signal detection based on Mel-Cepstral Interpretation and Multilayer Perceptron. Background Technology

[0002] In the field of underwater ecological environment monitoring, electrofishing is strictly regulated due to its devastating impact on fishery resources. Electrofishing devices store energy through capacitors or inductors and release high-voltage, high-current pulses into the water, generating strong acoustic emission signals when discharged. Currently, the detection of electrofishing signals mainly relies on fixed detection devices. These devices typically use metal electrode plates to directly detect changes in the underwater electric field and combine them with analog-to-digital converters to convert the collected analog voltage signals into digital signals before uploading them to a server for analysis. This traditional voltage detection-based method constitutes the main technological foundation for current electrofishing signal detection.

[0003] However, the aforementioned existing methods have significant limitations. First, the high-frequency pulse signals emitted by electrofishing machines attenuate rapidly in water and are highly susceptible to underwater clutter, environmental noise, and water flow disturbances, resulting in low sensitivity and accuracy for fixed-electrode detection methods, making accurate signal capture difficult. Second, this method is ineffective against moving electrofishing vessels and lacks the ability to track moving targets. Furthermore, the directly acquired voltage signal has a low signal-to-noise ratio and fails to fully utilize the characteristic information contained in the acoustic emission signals accompanying the discharge process, leading to insufficient reliability and confidence of the entire system in complex underwater environments. Therefore, there is an urgent need in the field for a new electrofishing signal detection scheme that can integrate multimodal information and improve signal recognition accuracy and anti-interference capabilities. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for detecting electric fish signals based on Mel-Cepstral Interpretation and Multilayer Perceptron, comprising:

[0005] Acquire underwater discharge acoustic emission signals;

[0006] Based on the acoustic emission signal, noise and invalid signal segments are eliminated through preprocessing to obtain a clear acoustic emission signal;

[0007] Based on the clear acoustic emission signal, extract the characteristic parameters of the Mel frequency cepstral coefficients;

[0008] Based on the Mel frequency cepstral coefficient characteristic parameters, optimized characteristic parameters containing static and dynamic characteristics are obtained through differential optimization operations;

[0009] The optimized feature parameters and the detected voltage value are input into the multilayer perceptron model, and the fitted electrical signal intensity is output.

[0010] Optionally, the preprocessing operation includes:

[0011] Based on the acoustic emission signal, the high-frequency signal component is compensated by a pre-emphasis operation, wherein the pre-emphasis operation is implemented using a first-order high-pass filter, and the pre-emphasis coefficient in its transfer function ranges from 0.9 to 1.0.

[0012] Based on the pre-emphasized signal, the signal is divided into multiple approximately steady-state segments with lengths between 10 and 30 milliseconds through a framing operation, where there is overlap between adjacent segments;

[0013] Based on the framed signal, the discontinuous parts of the signal are smoothed by windowing, where the window function is a Hamming window.

[0014] Optionally, in the framing operation, the total number of frames is calculated based on the total signal length, frame length, and frame shift, where the frame shift is the overlapping portion between adjacent frames.

[0015] Optionally, in the windowing operation, the expression for the Hamming window is the multiplication of the window function value and the signal value, in order to correct the non-periodicity of the signal.

[0016] Optionally, the extraction of Mel frequency cepstral coefficient feature parameters includes:

[0017] Based on the clear acoustic emission signal, the signal is converted to the frequency domain using Fourier transform;

[0018] Based on the frequency domain signal, filtering is performed through a set of Mel filters, where the center frequency of the Mel filter is set based on the nonlinear relationship between the Mel frequency and the actual frequency;

[0019] Based on the filtered signal, calculate the energy in each Mel filter and take the logarithm of the energy;

[0020] Based on the energy obtained after taking the logarithm, the characteristic parameters of the Mel frequency cepstral coefficients are obtained through discrete cosine transform.

[0021] Optionally, the original dimension of the Mel frequency cepstral coefficient feature parameters is set to 32 dimensions, and the first 16 dimensions are taken as feature quantities, constructing a 16-dimensional feature based on the first-order and second-order differences. With 16 dimensions Furthermore, these are combined to form seven feature structures of 16-dimensional, 32-dimensional, and 48-dimensional dimensions for adjustable feature representation.

[0022] Optionally, the difference optimization operation includes:

[0023] Based on the Mel frequency cepstral coefficient characteristic parameters, calculate the first-order difference parameters to describe the relationship between two adjacent frames of signal;

[0024] Based on the first-order difference parameters, the second-order difference parameters are calculated to describe the relationship between three adjacent frames of signals.

[0025] The Mel frequency cepstral coefficient feature parameters, first-order difference parameters, and second-order difference parameters are horizontally concatenated to obtain combined feature parameters of multiple dimensions.

[0026] Optionally, the multilayer perceptron model includes an input layer, an output layer, and hidden layers, wherein the number of hidden layers is 11, the number of neurons in each layer is 500, and the activation function is a linear rectified function.

[0027] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0028] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] This invention effectively overcomes the limitations of traditional electrofishing detection techniques by fusing acoustic emission signal analysis with machine learning models. The method utilizes Mel-frequency cepstral coefficients to accurately extract underwater discharge acoustic signature features, and combines differential optimization to fully capture the dynamic characteristics of the signal, significantly improving feature discrimination capabilities in complex underwater environments. Furthermore, it employs a multilayer perceptron to nonlinearly fuse and reconstruct acoustic features and voltage signals, greatly enhancing the detection sensitivity and anti-interference capabilities for weak signals. Compared to traditional fixed voltage detection schemes, this method not only significantly improves the accuracy and reliability of signal detection but also effectively adapts to monitoring scenarios involving mobile electrofishing vessels, providing more reliable technical support for maritime patrol and law enforcement. Attached Figure Description

[0031] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0032] Figure 1 This is a general design block diagram of an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of a gun-type microphone interference tube according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the acoustic emission signal interception according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the first stage of discharge in the time and frequency domains according to an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the second stage of discharge in the time and frequency domains according to an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram of the time and frequency domains of the third stage of discharge according to an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of frame division according to an embodiment of the present invention;

[0039] Figure 8 This is a flowchart of the MFCC feature parameter extraction process according to an embodiment of the present invention;

[0040] Figure 9 This is a three-dimensional diagram of the MFCC feature parameters according to an embodiment of the present invention;

[0041] Figure 10 This is a schematic diagram of the energy intensity variation in the Mel domain according to an embodiment of the present invention;

[0042] Figure 11 This is a three-dimensional differential schematic diagram of MFCC according to an embodiment of the present invention;

[0043] Figure 12 This is a flowchart illustrating the feature parameter splicing process of an embodiment of the present invention;

[0044] Figure 13 This is a schematic diagram illustrating the neuron computation principle of an embodiment of the present invention;

[0045] Figure 14 This is a schematic diagram of the MLP network architecture according to an embodiment of the present invention. Detailed Implementation

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0048] Example 1

[0049] This embodiment provides a method for detecting electric fish signals based on Mel-frequency cepstral spectrometry and multilayer perceptron, including:

[0050] Acquire underwater discharge acoustic emission signals;

[0051] Based on the acoustic emission signal, noise and invalid signal segments are eliminated through preprocessing to obtain a clear acoustic emission signal;

[0052] Based on the clear acoustic emission signal, extract the characteristic parameters of the Mel frequency cepstral coefficients;

[0053] Based on the Mel frequency cepstral coefficient characteristic parameters, optimized characteristic parameters containing static and dynamic characteristics are obtained through differential optimization operations;

[0054] The optimized feature parameters and the detected voltage value are input into the multilayer perceptron model, and the fitted electrical signal intensity is output.

[0055] This embodiment proposes an algorithm for detecting electrofishing signals based on Mel-Frequency Cepstral Coefficients (MFCC) and Multilayer Perceptron (MLP). This method delves into the mechanism of underwater discharge acoustic emission, extracts features of the acoustic emission signal from electrofishing equipment using MFCC technology, and utilizes an MLP network to perform nonlinear fitting between the original voltage detection signal from the ADC and the extracted feature intensity, thereby obtaining electrical signal data with significantly improved accuracy and effectively enhancing the sensitivity and confidence of voltage detection. The overall algorithm design block diagram is shown below. Figure 1 As shown.

[0056] The principle of an electrofishing device is to convert a low-voltage, high-current power supply into a high-voltage one, releasing a high-current, high-voltage pulse into the water, thereby killing fish within a certain range. In essence, high-voltage, high-current pulse discharge is energy compression. Energy is stored in capacitors or inductors, and then released through a discharge circuit onto a pulse power load in a very short time, generating pulse power, typically greater than 30W. When this high-voltage, high-current pulse discharge occurs underwater, it produces a strong explosive effect. The resulting extremely strong shock pressure wave propagates outward, forming a sound pulse. This is the basic process of underwater discharge generating sound waves.

[0057] Theoretical research revealed that acoustic emission technology can be used to assess the existence of underwater high-frequency pulse discharge phenomena in the current environment. During the experiment, a large amount of audio data of underwater discharge from electrofishing devices was collected, along with voltage amplitudes at different frequencies and pulse intensities. The relationship between parameter variations of various high-frequency pulse signals and the acoustic emission signals they generate was explored. Furthermore, a series of main morphological characteristics of underwater pulse discharge acoustic emission signals were summarized, laying a foundation for further analysis and acquisition of relevant characteristic parameters of acoustic emission signals during underwater discharge.

[0058] Due to the complex acoustic structure of the laboratory, environmental reverberation, delay, and the sound from equipment such as transformers can affect experimental data. This embodiment uses a shotgun microphone for measurement. These microphones utilize a specially designed acoustic tube to reduce the influence of sound from other directions on the signal. The frequency response is 3Hz-50kHz, omnidirectional, with a sensitivity of 34mV / Pa and a peak input of 140dB. An independent sound card is used for A / D conversion, with a sampling rate of 24bit / 192kHz.

[0059] The acoustic detection device consists of two parts: an interference tube at the front and a pickup head and its associated circuitry at the rear. Figure 2 As shown. Its principle for reducing longitudinal acoustic signals is to utilize the side opening of the interference tube, ensuring that the distance difference between the sound waves entering the tube from the side and reaching the pickup head is exactly 1 / 2 wavelength, thus causing sound wave cancellation and reducing off-axis sensitivity. Using this model of microphone allows for better directional pickup of the desired experimental acoustic emission signals.

[0060] For the acoustic emission signal generated by underwater high-frequency pulsed discharge, this embodiment divides the underwater high-frequency pulsed discharge into three processes based on the changes in discharge intensity and acoustic emission power. Each process is captured as a 5000ms segment in this embodiment. Figure 3 As shown. The first process mainly involves the electrofishing device discharging at a fixed frequency, such as... Figure 4 As shown, the discharge power is relatively small, and the sound pressure level is also relatively low; in the second process, as the discharge intensity increases, the sound pressure level also increases accordingly. Figure 5 As shown, the crackling sounds produced by the electric arc gradually become more frequent; the discharge intensity reaches its peak in the third stage. Figure 6 As shown.

[0061] Noise interference and invalid signal segments exist during the acquisition of underwater discharge acoustic emission signals from electrofishing machines, which may degrade the model's performance in recognition tasks. To enhance the model's ability to perceive underwater discharge acoustic emission signals, preprocessing is necessary to eliminate noise and unnecessary information. Next, endpoint detection is performed on clear signals to remove invalid non-sound segments, retaining only the valid signal portions. Subsequently, frequency domain feature analysis is conducted to extract the frequency domain features of the acoustic emission signal, and these features are concatenated to obtain more comprehensive information, thereby more accurately reflecting the characteristics of underwater high-frequency pulsed discharge acoustic emission signals.

[0062] To reduce the noise level and make the acoustic emission signal clearer and the signal-to-noise ratio higher, the original signal needs to be preprocessed. These processes include a series of operations such as pre-emphasis, framing, and windowing.

[0063] During acoustic emission from a discharge, radiation causes energy loss in the signal at a rate of 6 dB per octave, resulting in lower energy in the high-frequency range. In this case, after FFT transformation, the dynamic range of the spectrum becomes larger, and the spectral values ​​decrease with increasing frequency. To compensate for the high-frequency signal components suppressed by the sound-emitting system, highlight high-frequency resonant peaks in the signal, and eliminate the influence of radiation on the high-frequency range, pre-emphasis can be used. Pre-emphasis compensates for this effect by attenuating the energy of low-frequency signals, thereby ensuring a consistent signal-to-noise ratio across the entire frequency band in the calculated spectrum.

[0064] This embodiment uses a first-order FIR high-pass filter to pre-emphasize the acoustic emission signal. Its transfer function is shown below.

[0065] ;

[0066] In the formula, This is the pre-emphasis coefficient. The value of this factor is typically between 0.9 and 1.0; in this embodiment, the pre-weighting coefficient is set to 0.97. Represents a complex variable used to describe the frequency response of a filter in the Z-domain.

[0067] To overcome the non-stationary characteristics of the signal, the time-varying discharge acoustic emission signal can be divided into multiple approximately steady-state segments of about 10 to 30 milliseconds in length using a framing method. The duration of each segment is called the frame length. To ensure a smooth transition between adjacent frames and avoid excessive abrupt changes, a method called frame shifting is used, which involves partially overlapping the signals between adjacent frames. A schematic diagram of framing is shown below. Figure 7 As shown.

[0068] Symbols Total length of speech. For frame length, This refers to the overlapping portion between frames. This is the frame shift (step size). Then the total number of frames... The calculation is as follows:

[0069] ;

[0070] because:

[0071] ;

[0072] Therefore:

[0073] ;

[0074] Since framing may truncate the acoustic emission signal, leading to spectral leakage, windowing is applied to the framed acoustic emission signal. This operation smooths out discontinuous portions of the acoustic emission signal, correcting its aperiodicity and reducing spectral leakage. The windowing operation can be expressed as follows:

[0075] ;

[0076] In the formula, n represents the index of the output signal, indicating the time point in the windowed signal y(n), and m represents the index of the input signal, indicating the time point in the original signal x(m). This indicates the acoustic emission signal after windowing. This represents the acoustic emission signal of underwater discharge. This represents the window function. Common window functions include rectangular windows, Hanning windows, and Hamming windows. In this embodiment, a Hamming window is selected, and its expression is shown below:

[0077] ;

[0078] in, The length of the window function.

[0079] Underwater discharge acoustic emission signals are time-domain signals, making them difficult to analyze directly from waveform. To better analyze these signals, they can be transformed to the frequency domain using methods such as Fourier transform to obtain a suitable frequency domain signal for analysis. MFCC (Medium Frequency Scale) is a characteristic parameter proposed based on the characteristics of human hearing, exhibiting excellent noise resistance and widely used in audio analysis. Human hearing perception of sound is non-linear; the pitch of a sound is not directly proportional to its actual frequency. MFCC simulates the way the human ear perceives sound using the Mel frequency scale, and its relationship with the actual frequency is given by the following formula:

[0080] ;

[0081] In the formula For Mel frequency, This is the actual frequency (in Hz).

[0082] MFCC characteristic parameters can reflect the static characteristics of the electroacoustic emission signal of an electrofishing machine. The characteristic parameter extraction process is as follows: Figure 8 As shown.

[0083] The core of MFCC lies in the Mel-frequency cepstral converter technique, which simulates the human hearing process by constructing a set of bandpass filters with triangular filtering characteristics located within a specific frequency band. The bandwidth of these filters is distributed within the Mel frequency range, meaning their spacing remains constant. This design allows MFCC to better simulate how the human ear perceives sounds at different frequencies. In the MFCC calculation process, the sound signal is first filtered by a set of Mel filters, then logarithmically calculated on the filtered signal, and finally, a discrete cosine transform is performed to obtain the final MFCC coefficients, which characterize the sound signal.

[0084] ;

[0085] In the formula This represents a bandpass filter, where k represents the filter number. Set the center frequency of the Mel filter, and , This represents the number of bandpass filters. This represents the center frequency of the next frequency band (m+1) of the m-th filter in the Mel filter bank. This represents the center frequency of the previous frequency band of the m-th filter in the Mel filter bank. In this embodiment, the number of bandpass filters is set to 32. The expression is as follows:

[0086] ;

[0087] In the formula yes inverse function, This indicates the window size for the Fourier transform (FFT), set to 25 milliseconds. The sampling frequency is 44100 in this embodiment. , This indicates the lowest frequency of the filter, set to 0Hz. Set the highest frequency of the filter to 8000. . The calculation formula is as follows:

[0088] ;

[0089] The input signal is processed through calculation to obtain its energy contained in each Mel filter, that is, the power value of the discrete power spectrum passing through the Mel filter, where the discrete power spectrum is represented by... This indicates that the Mel filter uses The calculation formula is as follows:

[0090] ;

[0091] In the formula , Used to indicate the first The frame acoustic transmission signal in the Mel domain The sum of energy across all frequency bands This represents the discrete power spectrum of the acoustic emission signal from the underwater discharge of an electric fishing device. The calculation formula is as follows:

[0092] ;

[0093] in, This represents the spectral values ​​of the signal in the i-th frame and k-band.

[0094] Then After taking the logarithm of the energy and performing a discrete cosine transform (DCT), the MFCC characteristic parameters are obtained, and the calculation formula is as follows:

[0095] ;

[0096] in, This indicates the number of Mel-scale filters. , For the feature parameter dimension of MFCC output, this embodiment takes... That is, outputting 32-dimensional MFCC feature parameters. Figure 9 The three-dimensional diagrams of MFCC parameters at different stages of underwater discharge show that the MFCC parameter values ​​of the underwater discharge acoustic emission signal of the electric fishing machine are different under different dimensions and different frame numbers, which can well characterize the features of underwater discharge acoustic emission signals with different power and frequency.

[0097] At the same time, such as Figure 10 As shown, the energy intensity of the underwater discharge acoustic emission signal in the Mel domain decreases and tends to 0 as the dimension of the Mel filter bank increases. Therefore, this chapter takes the signal of the first 16 dimensions as the characteristic quantity for analysis.

[0098] Difference methods are a commonly used technique in data processing, primarily for analyzing and processing time series data. They can be used to detect and eliminate trend components in data, reducing data fluctuations. Further first-order or higher-order difference operations on the MFCC data of underwater discharge acoustic emissions from electrofishing machines can eliminate linear trends, making the data more stable and suitable for subsequent analysis and pattern recognition.

[0099] The characteristic parameters are the first-order differences of MFCC, which can describe the relationship between two adjacent frames of the acoustic emission signal; The parameter is the second-order difference of MFCC, which describes the relationship between three adjacent frames of the underwater discharge acoustic emission signal of the electric fishing machine. and All reflect the dynamic characteristics of acoustic emission signals. Based on the MFCC characteristic parameters of the signal, using... The first and second differences are obtained using the following formulas:

[0100] ;

[0101] ;

[0102] In the formula For the first-order difference characteristic parameters of MFCC, use To indicate; For the second-order difference characteristic parameters of MFCC, use To indicate, This represents the j-th MFCC coefficient in the (i-1)-th frame. This represents the j-th MFCC coefficient in the (i+1)-th frame. This represents the j-th MFCC coefficient in frame i-2. This represents the j-th MFCC coefficient in frame i+2.

[0103] Figure 11 The figures show the first-order and second-order differential 3D plots of the underwater discharge acoustic emission signal from an electric fishing device. The parameter values ​​of the underwater discharge acoustic emission signal differ significantly in different dimensions.

[0104] Will , and The parameters were horizontally concatenated to obtain seven sets of characteristic parameters that could respectively characterize the static, dynamic, and hybrid characteristics of the underwater discharge acoustic emission signal of the electrofishing device, representing 16 dimensions. 16-dimensional 16-dimensional 32-dimensional 32-dimensional 32-dimensional And 48-dimensional The feature parameter concatenation process is as follows: Figure 12 The results are summarized in Table 1.

[0105] Table 1

[0106] Feature parameters Dimension Signal representation capability 16 Static features 16 Dynamic features 16 Dynamic features 32 Static and dynamic characteristics 32 Static and dynamic characteristics 32 Dynamic features 48 Static and dynamic characteristics

[0107] By using the differential method, the differences in signals are greatly reduced, and the static and dynamic characteristics of acoustic emission signals are obtained. At the same time, different combinations of feature parameters can be used to characterize different feature quantities, thereby improving the accuracy and specificity of feature parameter identification.

[0108] A multilayer perceptron (MLP) is a type of feedforward neural network where neurons are arranged in layers, with no feedback between layers. Each neuron receives and processes the output signals of all neurons in the previous layer, and then outputs these signals to the next layer of the MLP for independent computation. A neuron multiplies the output of the previous layer by its weights and sums the results as its input. Only when the input exceeds a neuron's threshold is the signal transmitted down the chain, processed by an activation function, and then used as the neuron's output. It is the output of the neurons in the previous layer. Through activation function The weighted sum of the inputs for the transformation is shown below:

[0109] ;

[0110] In the formula, The connection weights between the neurons in the previous layer and this neuron; This is the threshold of the neuron; For activation functions, the commonly used one is the Rectified Linear Array (ReLU), i.e. The computational principle of neurons is as follows: Figure 13 As shown:

[0111] The MLP neural network regression model maps the relationship between target distance, acoustic emission signal intensity, and detection voltage. After constructing the MLP neural network model, it needs to be trained and tested. Neural networks typically contain three basic elements: an input layer, an output layer, and hidden layers. Therefore, the training and testing datasets for the MLP neural network model are divided into input and output data. In this embodiment, the input data are the device detection voltage and the underwater discharge acoustic emission characteristic intensity, and the output data is the electrical signal intensity fitted by a single device. The collected data is shown in the figure. Min-Max normalization is used to map the data to... Used to remove the dimensions and units of measurement from data of different dimensions:

[0112] ;

[0113] In the formula For normalized data, The maximum value of the sample data. This represents the minimum value of the sample data.

[0114] Based on the normalized measurement results, the collected datasets were used for both training and testing. The dataset contained a total of 390 data sets, of which 350 sets were used as the training dataset to train the MLP neural network model; the other 30 sets served as the untrained test dataset for the MLP neural network model, and were not used in the training of the MLP neural network model, thereby verifying the predictive ability of the MLP neural network.

[0115] An MLP neural network was built using the PyTorch framework based on Python 3.9, with all computations performed on an Intel Core i5-12400F processor. Based on experimentally measured voiceprint feature intensity and detection voltage, the input and output layers of the constructed MLP neural network were initially set to a dimension of 1. Through gradual adjustments based on experimental data, the number of hidden layers was determined to be 11, with 500 neurons per layer. This completed the MLP neural network model used for mapping. The network architecture is as follows: Figure 14 As shown.

[0116] The number of training iterations primarily affects the loss function and training time during MLP neural network model training. The former represents the predictive ability of the neural network, while the latter represents the cost of training the neural network. The loss function gradually decreases with increasing training iterations. After the number of training iterations exceeds 900, the decrease in the loss function becomes slow, while the training time increases linearly with the number of training iterations. Considering both the loss function and training time, 900 training iterations were ultimately chosen.

[0117] After training, the MLP neural network model can reconstruct more accurate electrical signal data from given detected voltage values ​​and feature intensities, thereby improving the sensitivity and confidence of voltage detection in inspection robots. First, the predictive ability of the MLP neural network model is tested using training data. The input voltage and intensity values ​​from the collected training data are fed into the trained MLP neural network to reconstruct the corresponding fitted intensity values.

[0118] The root mean square error of the reconstructed training data was 12, and the maximum relative error was 1.10%, which is less than 2% and within a small range. This indicates that the MLP neural network can be used for electrical signal reconstruction and has good predictive ability.

[0119] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0120] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0121] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting electric fish signals based on Mel-frequency cepstral analysis and multilayer perceptron, characterized in that, include: Acquire underwater discharge acoustic emission signals; Based on the acoustic emission signal, noise and invalid signal segments are eliminated through preprocessing to obtain a clear acoustic emission signal; Based on the clear acoustic emission signal, extract the characteristic parameters of the Mel frequency cepstral coefficients; Based on the Mel frequency cepstral coefficient characteristic parameters, optimized characteristic parameters containing static and dynamic characteristics are obtained through differential optimization operations; The optimized feature parameters and the detected voltage value are input into the multilayer perceptron model, and the fitted electrical signal intensity is output. The extracted Mel frequency cepstral coefficient feature parameters include: Based on the clear acoustic emission signal, the signal is converted to the frequency domain using Fourier transform; Based on the frequency domain signal, filtering is performed through a set of Mel filters, where the center frequency of the Mel filter is set based on the nonlinear relationship between the Mel frequency and the actual frequency; Based on the filtered signal, calculate the energy in each Mel filter and take the logarithm of the energy; Based on the energy after taking the logarithm, the characteristic parameters of the Mel frequency cepstral coefficients are obtained through discrete cosine transform. The original dimension of the Mel frequency cepstral coefficient feature parameters is set to 32, and the first 16 dimensions are taken as feature quantities. A 16-dimensional feature is constructed based on the first-order and second-order differences. With 16 dimensions Furthermore, these are combined to form seven feature structures of 16-dimensional, 32-dimensional, and 48-dimensional dimensions for adjustable feature representation; Will , and The parameters were horizontally concatenated to obtain seven sets of characteristic parameters that could respectively characterize the static, dynamic, and hybrid characteristics of the underwater discharge acoustic emission signal of the electrofishing device, representing 16 dimensions. 16-dimensional 16-dimensional 32-dimensional 32-dimensional 32-dimensional And 48-dimensional ; First, the audio signal is filtered by a set of Mel filters. Then, logarithmic operations are performed on the filtered signal. Finally, discrete cosine transform is performed to obtain the final MFCC coefficients, which are used to characterize the features of the audio signal. ； In the formula This represents a Mel filter. It is the center frequency of the Mel filter, and , The number of Mel filters, This represents the center frequency of the next frequency band (m+1) of the m-th filter in the Mel filter bank; This represents the center frequency of the previous frequency band of the m-th filter in the Mel filter bank, and the number of Mel filters is set to 32. The expression is as follows: ； In the formula, yes inverse function, This indicates the window size for the Fourier transform (FFT), set to 25 milliseconds. This indicates the sampling frequency, using 44100. , This indicates the lowest frequency of the filter, set to 0Hz. Set the highest frequency of the filter to 8000. ; The calculation formula is as follows: ； The filtered signal is processed through calculation to obtain the energy contained in each Mel filter, i.e., the power value of the discrete power spectrum passing through the Mel filter, where the discrete power spectrum is represented by... This indicates that the Mel filter uses The calculation formula is as follows: ； In the formula , Used to indicate the first The frame acoustic transmission signal in the Mel domain The sum of energy across the frequency band of a Mel filter. This represents the discrete power spectrum of the acoustic emission signal from the underwater discharge of an electric fishing device. The calculation formula is as follows: ； in, This represents the spectral values ​​of the signal in the i-th frame and k-band; Will After taking the logarithm of the energy and performing a discrete cosine transform (DCT), the MFCC characteristic parameters are obtained, and the calculation formula is as follows: ； in, Indicates the number of Mel filters. , The feature parameter dimension of the MFCC output. That is, outputting 32-dimensional MFCC feature parameters.

2. The method according to claim 1, characterized in that, The preprocessing operations include: Based on the acoustic emission signal, the high-frequency signal component is compensated by a pre-emphasis operation, wherein the pre-emphasis operation is implemented using a first-order high-pass filter, and the pre-emphasis coefficient in its transfer function ranges from 0.9 to 1.

0. Based on the pre-emphasized signal, the signal is divided into multiple approximately steady-state segments with lengths between 10 and 30 milliseconds through a framing operation, where there is overlap between adjacent segments; Based on the framed signal, the discontinuous parts of the signal are smoothed by windowing, where the window function is a Hamming window.

3. The method according to claim 2, characterized in that, In the framing operation, the total number of frames is calculated based on the total signal length, frame length, and frame shift, where the frame shift is the overlapping part between adjacent frames.

4. The method according to claim 2, characterized in that, In the windowing operation, the expression for the Hamming window is the multiplication of the window function value and the signal value, in order to correct the non-periodicity of the signal.

5. The method according to claim 1, characterized in that, The difference optimization operation includes: Based on the Mel frequency cepstral coefficient characteristic parameters, calculate the first-order difference parameters to describe the relationship between two adjacent frames of signal; Based on the first-order difference parameters, the second-order difference parameters are calculated to describe the relationship between three adjacent frames of signals. The Mel frequency cepstral coefficient feature parameters, first-order difference parameters, and second-order difference parameters are horizontally concatenated to obtain combined feature parameters of multiple dimensions.

6. The method according to claim 1, characterized in that, The multilayer perceptron model includes an input layer, an output layer, and hidden layers, with 11 hidden layers and 500 neurons in each layer. The activation function is a linear rectified function.

7. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-6.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-6.

Images