A shrimp call recognition method based on periodic analysis and non-negative matrix factorization

Abstract

The present invention discloses a shrimp call recognition method based on periodic analysis non-negative matrix decomposition, and its field is the field of aquaculture technology. The present invention introduces a periodic analysis non-negative matrix decomposition algorithm with self-learning ability, decomposes the original acoustic data into a basis matrix and a coding matrix, extracts preliminary information about the time change of each sound source, can effectively separate natural environmental noise and biological calls, and has a strong adaptive ability to different signals, realizes unsupervised noise separation, improves the accuracy and timeliness of recognition, adapts to dynamic environmental noise, and provides strong data support for shrimp pond aquaculture management; has self-learning ability, can cluster different basis matrices by introducing sparse constrained non-negative matrix decomposition technology, effectively recognize and distinguish the vocal characteristics of shrimp, can meet the needs of shrimp pond aquaculture, and does not need to retrain the model when facing various types of biological sound separation in various environments. It is efficient, fast and accurate, and has a wide range of application scenarios.

Description

Technical Field

[0001] The invention relates to the technical field of aquaculture, and in particular to a shrimp call recognition method based on periodic analysis non-negative matrix decomposition. Background Art

[0002] Under natural environmental conditions, the collected acoustic signals are usually mixed with environmental noise and human noise. In the monitoring process of shrimp ponds, the sound of bubbles in the water, the sounds of human daily life, the calls of other animals, and even the passing of distant airplanes can become noise interference. Under the combined effect of many factors, the shrimp pond monitoring signal appears as a non-stationary signal. The removal of this type of non-stationary noise has always been a difficult problem in acoustic signal processing. At present, many common denoising methods such as spectral subtraction, Wiener filter, minimum mean square error (MMSE), multi-layer adaptive spectral filtering (MLSA), maximum a posteriori probability estimation (MAPA), generalized maximum a posteriori probability estimation (GMAPA), etc., although they can effectively denoise in certain specific environments, cannot effectively separate noise when facing non-stationary noise.

[0003] In recent years, the traditional NMF (non-negative matrix factorization) method has shown strong performance in noise separation. NMF (non-negative matrix factorization) extracts potential signal and noise components by decomposing mixed signals into basis matrices and coefficient matrices. However, NMF (non-negative matrix factorization) has a significant disadvantage: it requires a large amount of supervised training data to achieve the best results. Specifically, when using NMF (non-negative matrix factorization) for denoising, the original data of the research object must be collected in advance, and the pure signal source data must be separated as training data to train the model. In the natural environment, noise interference is ubiquitous. Collecting raw data and then separating pure signal sources is costly, time-consuming, and difficult. This premise is difficult to achieve under realistic conditions, and it is impossible to effectively identify shrimp calls, thereby failing to help farmers objectively judge the health of shrimps, and failing to accurately formulate and implement breeding plans, which reduces breeding efficiency and yield. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a shrimp call recognition method based on periodic analysis non-negative matrix decomposition, which solves the problem that the prior art cannot effectively separate shrimp call signals and noise signals, and the separation cost is high, the time is long, and the difficulty is great, and the shrimp calls cannot be effectively identified, thereby failing to help farmers objectively judge the health status of shrimps, thereby reducing the breeding efficiency and production.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A shrimp call recognition method based on periodic analysis non-negative matrix decomposition is provided, which comprises the following steps:

[0007] S1. Preset relevant parameters;

[0008] S2, obtaining original shrimp call data;

[0009] S3, performing time cascade processing on the original shrimp call data to obtain a cascade matrix; S4, performing NMF decomposition on the cascade matrix to obtain a base matrix and a coding matrix;

[0010] S5, performing periodic analysis on the coding matrix and processing it using sparse NMF clustering to obtain clustering results of different periodic signals and corresponding clustering labels;

[0011] S6, according to each clustering result and clustering label of the coding matrix, reorganize the label of the basis matrix corresponding to the coding matrix, obtain the updated basis matrix, reconstruct the original shrimp call data according to different periodic signals, obtain the noise result and the shrimp call result, and complete the recognition of the shrimp call;

[0012] S7. Reconstruct the original shrimp call data according to the clustering results in S5 and the basis matrix updated in S6 to obtain a noise result and a shrimp call result.

[0013] Furthermore, the relevant parameters in S1 include the number of clusters, the number of time frames, the number of basis matrices, the number of NMF update iterations and the sparsity parameter; the original shrimp call data in S2 is represented by a time-frequency representation matrix.

[0014] Furthermore, the specific process of S3 is:

[0015] Determine whether the time-frequency representation matrix corresponding to the original shrimp call data contains negative values, if so, return to S1; otherwise, normalize the time-frequency representation matrix to obtain a normalized time-frequency representation matrix;

[0016] The frequency dimension size and time dimension size of the normalized time-frequency representation matrix are obtained, and based on the number of time frames, the data in the normalized time-frequency representation matrix are subjected to time cascade processing by translating and superimposing the normalized time-frequency representation matrix in the row direction to obtain a cascade matrix; wherein the calculation formula for time cascade processing is:

[0017]

[0018] in, Indicates the number of cascades, represents the summation function, Represents the normalized time-frequency representation matrix The position in The corresponding value, Represents a cascade matrix.

[0019] Further, S4 includes the following steps:

[0020] S4-1, globally scaling the cascade matrix to obtain a processed cascade matrix;

[0021] S4-2, decomposing the processed cascade matrix according to the number of base matrices to obtain an initial base matrix and an initial coding matrix;

[0022] S4-3, initializing the initial base matrix and the initial coding matrix to obtain an initialized base matrix and an initialized coding matrix;

[0023] S4-4, according to the sparsity parameter, using projection to adjust the sparsity corresponding to the initialized base matrix and the sparsity corresponding to the initialized encoding matrix to obtain an adjusted base matrix and an adjusted encoding matrix;

[0024] S4-5, using the gradient descent method to update the adjusted base matrix and the adjusted encoding matrix, and calculate the corresponding root mean square error;

[0025] S4-6, repeat S4-5 until the number of NMF update iterations is reached, and obtain each optimized basis matrix, optimized encoding matrix and corresponding root mean square error;

[0026] S4-7, selecting a group with the smallest root mean square error as the final optimized base matrix and the final optimized encoding matrix;

[0027] S4-8. Determine whether the final optimized base matrix and the final optimized coding matrix meet the first requirement; if so, use the optimized base matrix and the optimized coding matrix as the base matrix and the coding matrix; otherwise, return to S4-5.

[0028] Further, S5 includes the following steps:

[0029] S5-1, using Fourier transform to extract the periodic characteristics of the coding matrix;

[0030] S5-2. Calculate the significance of periodic features;

[0031] S5-3, screening the periodic features according to the significance of the periodic features to obtain screened periodic features;

[0032] S5-4. Use sparse NMF clustering to cluster the filtered periodic features to obtain clustering results and corresponding clustering labels.

[0033] Further, S5-1 includes the following steps:

[0034] S5-1-1, performing discrete Fourier transform on each column of data in the encoding matrix to obtain spectrum information corresponding to each column of data;

[0035] S5-1-2. Calculate the mean of each column of the encoding matrix based on the spectrum information;

[0036] S5-1-3, subtract the corresponding mean value from each column of the encoding matrix, and perform discrete Fourier transform to obtain the initial periodic characteristics;

[0037] S5-1-4. Retain the initial periodic characteristics of the positive frequency and calculate the corresponding amplitude to obtain the periodic characteristics.

[0038] Further, S6 includes the following steps:

[0039] S6-1. Calculate the mean of the difference between the peak value and the median value of each clustering result of the encoding matrix;

[0040] S6-2, sorting the means in descending order to obtain the corresponding sorting results;

[0041] S6-3. According to the sorting results and the clustering labels corresponding to the clustering results, the labels of the base matrix corresponding to the encoding matrix are reorganized to obtain an updated base matrix.

[0042] Further, S7 includes the following steps:

[0043] S7-1, reconstructing the original shrimp call data according to each clustering result and the updated basis matrix to obtain a noise reconstructed signal matrix and a shrimp call reconstructed signal matrix;

[0044] S7-2, dividing each column of the noise reconstruction signal matrix and the shrimp call reconstruction signal matrix according to the number of time frames, to obtain corresponding noise time frame sub-matrices and shrimp call time frame sub-matrices;

[0045] S7-3, calculating the average value of each noise time frame sub-matrix and each shrimp call time frame sub-matrix along the time dimension to obtain a noise result and a shrimp call result.

[0046] The beneficial effects of the present invention are as follows: the present invention decomposes the original acoustic data into a basis matrix and a coding matrix by introducing a periodic analysis non-negative matrix decomposition algorithm with self-learning ability, extracts preliminary information about the time change of each sound source, can effectively separate natural environmental noise and biological calls, and has a strong adaptive ability to different signals, realizes unsupervised noise separation, improves the accuracy and timeliness of recognition, adapts to dynamic environmental noise, and provides strong data support for shrimp pond farming management; the present invention has self-learning ability, can cluster different basis matrices by introducing sparse constrained non-negative matrix decomposition technology, effectively identify and distinguish the vocal characteristics of shrimps, can not only meet the needs of shrimp pond farming, but also can be directly applied to the analysis of other biological sounds, and does not need to retrain the model when separating various types of biological sounds in a variety of environments. It is efficient, fast and accurate, and has a wide range of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 Schematic diagram of data flow of the present invention

[0049] Figure 3 It is a schematic diagram of the processing result of the measured shrimp pond data of the present invention;

[0050] Figure 4 This is the fitting curve of shrimp calls and dissolved oxygen in water. DETAILED DESCRIPTION

[0051] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0052] like Figure 1 and Figure 2 As shown, a shrimp call recognition method based on periodic analysis non-negative matrix decomposition includes the following steps:

[0053] S1. Preset relevant parameters;

[0054] The relevant parameters in S1 include the number of clusters, the number of time frames, the number of basis matrices, the number of NMF update iterations, and the sparsity parameter. Among them, the number of clusters is used for unsupervised separation of sound source groups with different periodicities; the number of time frames is used to determine the degree of overlap of time frames in the cascade process; the number of basis matrices is used to control the number of basis vectors in the NMF decomposition; the number of NMF update iterations is used to determine the number of optimization steps in the decomposition process; the sparsity parameter is used to control the sparsity of the basis matrix and the encoding matrix. The relevant parameters can be set according to the actual data situation and the requirements for the separation effect. For example, the purpose of clustering is to separate biological signals from other signals, and all signals except biological signals are regarded as noise signals, and then the number of clusters (sound source data) is set to 2.

[0055] S2, obtaining original shrimp call data;

[0056] The original shrimp call data in S2 is represented by a time-frequency representation matrix; if the original shrimp call data is not a time-frequency representation matrix, the data needs to be Fourier transformed and converted into a time-frequency representation matrix. In addition, users can perform pre-processing operations such as filtering on the original shrimp call data according to their needs.

[0057] S3, performing time cascade processing on the original shrimp call data to obtain a cascade matrix;

[0058] The specific process of S3 is:

[0059] Determine whether the time-frequency representation matrix corresponding to the original shrimp call data contains negative values. If so, return to S1; otherwise, normalize the time-frequency representation matrix, normalize the maximum value in the time-frequency representation matrix to 1, and obtain the normalized time-frequency representation matrix;

[0060] Get the frequency dimension size of the normalized time-frequency representation matrix and the time dimension size , and based on the number of time frames, the data in the normalized time-frequency representation matrix is ​​temporally cascaded by translating and superimposing the normalized time-frequency representation matrix in the row direction. Let the number of cascades be , then a size of The new matrix is ​​the cascade matrix, which can ensure the time continuity of the signal. The calculation formula for the time cascade processing is:

[0061]

[0062] in, Indicates the number of cascades, represents the summation function, Represents the normalized time-frequency representation matrix The position in The corresponding value, Represents a cascade matrix.

[0063] S4, performing NMF decomposition on the cascade matrix to obtain a base matrix and a coding matrix;

[0064] Processing time-frequency matrices using non-negative matrix factorization with sparsity constraints , for the time-frequency matrix Decompose and get the basis matrix and the encoding matrix . The introduction of sparse constraints can control the basis matrix and the encoding matrix The sparsity of Approximately , to obtain the group with the smallest root mean square error (RMSE) as the optimal result. The corresponding calculation formula is:

[0065]

[0066] in, Represents the time-frequency matrix No. Line Column data, Represents the basis matrix No. Line Column data, Represents the encoding matrix No. Line Column data, represents the summation function, Represents a constant.

[0067] The sparsity constraint is: If the user provides a sparsity parameter, the algorithm will adjust the basis matrix by projection The sparsity of the encoding matrix. The sparsity control is achieved by non-negative projection of the elements in the matrix to the specified Euclidean norm target; if the sparsity parameter is not provided by the user, a null value is used.

[0068] S4 includes the following steps:

[0069] S4-1, globally scaling the cascade matrix to obtain a processed cascade matrix;

[0070] S4-2, decomposing the processed cascade matrix according to the number of base matrices to obtain an initial base matrix and an initial coding matrix;

[0071] S4-3, initializing the initial base matrix and the initial coding matrix to obtain an initialized base matrix and an initialized coding matrix; the base matrix and the encoding matrix is the target decomposition matrix, whose size is determined by the input data. If no initial value is provided, the basis matrix is ​​randomly initialized and the encoding matrix , and the encoding matrix Normalization is performed to keep the Euclidean norm of its rows fixed to ensure numerical stability.

[0072] S4-4, according to the sparsity parameter, using projection to adjust the sparsity corresponding to the initialized base matrix and the sparsity corresponding to the initialized encoding matrix to obtain an adjusted base matrix and an adjusted encoding matrix;

[0073] S4-5, using the gradient descent method to update the adjusted base matrix and the adjusted encoding matrix, and calculate the corresponding root mean square error;

[0074] S4-6, repeat S4-5 until the number of NMF update iterations is reached, and obtain the optimized base matrix, optimized encoding matrix and corresponding error root mean square; in each iteration, calculate the corresponding gradient, update the matrix along the negative gradient direction, and finally ensure that the updated matrix satisfies the sparsity constraint. During the iteration process, dynamically adjust the step size to ensure that the objective function (error root mean square) can be reduced.

[0075] S4-7, selecting a group with the smallest root mean square error as the final optimized base matrix and the final optimized encoding matrix;

[0076] S4-8. Determine whether the final optimized base matrix and the final optimized coding matrix meet the first requirement; if so, use the optimized base matrix and the optimized coding matrix as the base matrix and the coding matrix; otherwise, return to S4-5.

[0077] Among them, the first requirement is that the product of the optimized base matrix and the optimized coding matrix is ​​approximately a time-frequency matrix, and needs to meet the sparsity requirements set by the user, such as sparsity parameters.

[0078] S5, performing periodic analysis on the coding matrix and processing it using sparse NMF clustering to obtain clustering results and corresponding clustering labels;

[0079] S5 includes the following steps:

[0080] S5-1, using Fourier transform to extract the periodic characteristics of the coding matrix;

[0081] S5-1 includes the following steps:

[0082] S5-1-1, performing discrete Fourier transform on each column of data in the encoding matrix to obtain spectrum information corresponding to each column of data;

[0083] S5-1-2. Calculate the mean of each column of the encoding matrix based on the spectrum information;

[0084] S5-1-3, subtract the corresponding mean value from each column of the encoding matrix, and perform discrete Fourier transform to obtain the initial periodic characteristics;

[0085] S5-1-4. Retain the initial periodic characteristics of the positive frequency and calculate the corresponding amplitude to obtain the periodic characteristics.

[0086] The corresponding calculation formula is:

[0087]

[0088] in, represents the periodic matrix, i.e. the periodic characteristics, represents a periodic analytical function, Indicates the initial periodic characteristics of positive frequencies.

[0089] S5-2. Calculate the significance of the periodic feature; where significance refers to the difference between the maximum amplitude and the median of the spectrum of the periodic feature, which is used to measure the significance of the periodic feature of the column of data. The larger the difference, the stronger the periodicity of the column of data.

[0090] S5-3, screening the periodic features according to their significance, that is, selecting the periodic features whose significance is greater than a preset threshold, to obtain the screened periodic features;

[0091] S5-4, cluster the filtered periodic features using sparse NMF clustering to obtain clustering results and corresponding clustering labels. The calculation formula corresponding to the clustering is:

[0092]

[0093] in, represents the clustering result, represents the minimum function, represents the transposed matrix of a periodic matrix, represents the basis matrix obtained by each non-negative matrix decomposition in the clustering step, represents the sparse penalty factor, represents the norm 1, Represents the distance metric function.

[0094] S6. According to the clustering results and clustering labels of the encoding matrix, the labels of the base matrix corresponding to the encoding matrix are reorganized to obtain an updated base matrix;

[0095] S6 includes the following steps:

[0096] S6-1. Calculate the mean of the difference between the peak value and the median value of each clustering result of the encoding matrix;

[0097] S6-2, sorting the means in descending order to obtain the corresponding sorting results;

[0098] S6-3. According to the sorting results and the clustering labels corresponding to the clustering results, the labels of the base matrix corresponding to the encoding matrix are reorganized to obtain an updated base matrix.

[0099] S7. Reconstruct the original shrimp call data according to the clustering results in S5 and the basis matrix updated in S6 to obtain a noise result and a shrimp call result.

[0100] S7 includes the following steps:

[0101] S7-1. According to each clustering result and the updated basis matrix, the original shrimp call data is reconstructed to obtain the noise reconstruction signal matrix and the shrimp call reconstruction signal matrix. The corresponding calculation formula is:

[0102]

[0103] in, The reconstructed shrimp is called the signal matrix, represents the reconstructed noise signal matrix, The clustering result label is the basis matrix of shrimp. The basis matrix indicating that the clustering result label is noise Represents the encoding matrix of shrimp call signal, Represents the encoding matrix of the noise signal.

[0104] S7-2, dividing each column of the noise reconstruction signal matrix and the shrimp call reconstruction signal matrix according to the number of time frames, to obtain corresponding noise time frame sub-matrices and shrimp call time frame sub-matrices;

[0105] S7-3, calculate the average value of each noise time frame sub-matrix and each shrimp call time frame sub-matrix along the time dimension to obtain the noise result and the shrimp call result. The corresponding calculation formula is:

[0106]

[0107] in, Indicates the average value calculation result, that is, the noise result or the shrimp call result, Indicates The average value of the sub-matrices of the time frames.

[0108] In one embodiment of the present invention, the original shrimp call data of a pond is obtained, and the recognition method proposed by the present invention is used to process it, and the result obtained is as follows: Figure 3 As shown, preliminary information about the time changes of each sound source is extracted, and the sparse constrained non-negative matrix factorization technology is introduced to cluster different basis matrices, effectively identify and distinguish the vocal characteristics of shrimps, and perfectly separate natural environmental noise and shrimp calls.

[0109] The results of shrimp call recognition can be analyzed based on the significant differences in the vocal characteristics of shrimp when they are hypoxic, eating or sick. Figure 4 As shown, taking dissolved oxygen concentration as an example, Figure 1 As shown in the figure, the relationship between the intensity of shrimp calls and the dissolved oxygen concentration is shown. By analyzing the fitting curve of shrimp calls and dissolved oxygen concentration, it can be found that when the dissolved oxygen concentration increases, the intensity of shrimp calls increases significantly, which is consistent with the trend of dissolved oxygen concentration. Farmers can set the dissolved oxygen concentration threshold according to actual breeding needs as an early warning mechanism. When the dissolved oxygen concentration is lower than the threshold, a warning is issued to guide farmers to dynamically adjust the oxygenation time and dosage, thereby improving breeding efficiency and yield.

[0110] In summary, the present invention introduces a periodic analysis non-negative matrix decomposition algorithm with self-learning ability, decomposes the original acoustic data into a basis matrix and a coding matrix, extracts preliminary information about the time change of each sound source, can effectively separate natural environmental noise and biological calls, and has a strong adaptive ability to different signals, realizes unsupervised noise separation, improves the accuracy and timeliness of recognition, adapts to dynamic environmental noise, and provides strong data support for shrimp pond farming management; the present invention has self-learning ability, can cluster different basis matrices by introducing sparse constrained non-negative matrix decomposition technology, effectively identify and distinguish the vocal characteristics of shrimps, can not only meet the needs of shrimp pond farming, but also can be directly applied to the analysis of other biological sounds, and does not need to retrain the model when separating various types of biological sounds in a variety of environments. It is efficient, fast and accurate, and has a wide range of application scenarios.

Claims

1. A shrimp call recognition method based on periodic analysis non-negative matrix decomposition, characterized by: The following steps are involved: S1. Preset relevant parameters; S2, obtaining original shrimp call data; S3, performing time cascade processing on the original shrimp call data to obtain a cascade matrix; S4, performing NMF decomposition on the cascade matrix to obtain a base matrix and a coding matrix; S5, performing periodic analysis on the coding matrix and processing it using sparse NMF clustering to obtain corresponding clustering results and corresponding clustering labels; S6. According to the clustering results and clustering labels of the encoding matrix, the labels of the base matrix corresponding to the encoding matrix are reorganized to obtain an updated base matrix; S7, reconstructing the original shrimp call data according to the clustering results in S5 and the basis matrix updated in S6 to obtain a noise result and a shrimp call result; The relevant parameters in S1 include the number of clusters, the number of time frames, the number of base matrices, the number of NMF update iterations and the sparsity parameter; the original shrimp call data in S2 is represented by a time-frequency representation matrix; The specific process of S3 is as follows: Determine whether the time-frequency representation matrix corresponding to the original shrimp call data contains negative values, if so, return to S1; otherwise, normalize the time-frequency representation matrix to obtain a normalized time-frequency representation matrix; The frequency dimension size and time dimension size of the normalized time-frequency representation matrix are obtained, and based on the number of time frames, the data in the normalized time-frequency representation matrix are subjected to time cascade processing by translating and superimposing the normalized time-frequency representation matrix in the row direction to obtain a cascade matrix; wherein the calculation formula for time cascade processing is: in, Indicates the number of cascades, represents the summation function, Represents the normalized time-frequency representation matrix The position in The corresponding value, Represents a cascade matrix.

2. The shrimp call recognition method based on periodic analysis non-negative matrix decomposition according to claim 1 is characterized in that: The S4 comprises the following steps: S4-1, globally scaling the cascade matrix to obtain a processed cascade matrix; S4-2, decomposing the processed cascade matrix according to the number of base matrices to obtain an initial base matrix and an initial coding matrix; S4-3, initializing the initial base matrix and the initial coding matrix to obtain an initialized base matrix and an initialized coding matrix; S4-4, according to the sparsity parameter, using projection to adjust the sparsity corresponding to the initialized base matrix and the sparsity corresponding to the initialized encoding matrix to obtain an adjusted base matrix and an adjusted encoding matrix; S4-5, using the gradient descent method to update the adjusted base matrix and the adjusted encoding matrix, and calculate the corresponding root mean square error; S4-6, repeat S4-5 until the number of NMF update iterations is reached, and obtain each optimized basis matrix, optimized encoding matrix and corresponding root mean square error; S4-7, selecting a group with the smallest root mean square error as the final optimized base matrix and the final optimized encoding matrix; S4-8. Determine whether the final optimized base matrix and the final optimized coding matrix meet the first requirement; if so, use the optimized base matrix and the optimized coding matrix as the base matrix and the coding matrix; otherwise, return to S4-5.

3. The shrimp call recognition method based on periodic analysis non-negative matrix decomposition according to claim 1 is characterized in that: The S5 comprises the following steps: S5-1, using Fourier transform to extract the periodic characteristics of the coding matrix; S5-2. Calculate the significance of periodic features; S5-3, screening the periodic features according to the significance of the periodic features to obtain screened periodic features; S5-4. Use sparse NMF clustering to cluster the filtered periodic features to obtain clustering results and corresponding clustering labels.

4. The shrimp call recognition method based on periodic analysis non-negative matrix decomposition according to claim 3 is characterized in that: The S5-1 comprises the following steps: S5-1-1, performing discrete Fourier transform on each column of data in the encoding matrix to obtain spectrum information corresponding to each column of data; S5-1-2. Calculate the mean of each column of the encoding matrix based on the spectrum information; S5-1-3, subtract the corresponding mean value from each column of the encoding matrix, and perform discrete Fourier transform to obtain the initial periodic characteristics; S5-1-4. Retain the initial periodic characteristics of the positive frequency and calculate the corresponding amplitude to obtain the periodic characteristics.

5. The shrimp call recognition method based on periodic analysis non-negative matrix decomposition according to claim 3 is characterized in that: The S6 comprises the following steps: S6-1. Calculate the mean of the difference between the peak value and the median value of each clustering result of the encoding matrix; S6-2, sorting the means in descending order to obtain the corresponding sorting results; S6-3. According to the sorting results and the clustering labels corresponding to the clustering results, the labels of the base matrix corresponding to the encoding matrix are reorganized to obtain an updated base matrix.

6. The shrimp call recognition method based on periodic analysis non-negative matrix decomposition according to claim 3 is characterized by: The S7 comprises the following steps: S7-1, reconstructing the original shrimp call data according to each clustering result and the updated basis matrix to obtain a noise reconstructed signal matrix and a shrimp call reconstructed signal matrix; S7-2, dividing each column of the noise reconstruction signal matrix and the shrimp call reconstruction signal matrix according to the number of time frames, to obtain corresponding noise time frame sub-matrices and shrimp call time frame sub-matrices; S7-3, calculating the average value of each noise time frame sub-matrix and each shrimp call time frame sub-matrix along the time dimension to obtain a noise result and a shrimp call result.

Images