A cetacean signal enhancement method based on learning differences
Through the learning-difference-based whale and porcine signal enhancement method, the characteristic differences between marine environmental noise and mammalian whistle signals are divided into low-frequency bands and high-frequency bands for processing. The signal enhancement is used to solve the problem of marine environment The difficulty in identifying target signals caused by noise complexity achieves more efficient signal denoising and enhancement effects.
Patent Information
- Application Number
- CN202410156179.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-04
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-02-04
AI Technical Summary
The noise of the marine environment is complex. Traditional methods ignore important information in the low-frequency part in the acoustic signal processing of marine mammals, making it difficult to identify target signals.
The whale and porcine signal enhancement method based on learning differences is used to base the characteristics of marine environmental noise and mammalian whistle signals, and it is divided into low-frequency bands and high-frequency bands for learning. The two-mask-DPCRN model is used for signal enhancement. Feature extraction and recovery for noise and signal are performed through mask1 and mask2.
Significantly improves signal denoising effect and signal enhancement performance, reduces interference between different frequency bands, improves signal-to-noise ratio, and ensures consistency and independence of each channel over the entire frequency band range.
Smart Images

Figure CN118016088B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cetacean signal enhancement, and is a cetacean signal enhancement method based on learning differences. Background Art
[0002] In the field of animal bioacoustics, the detection and recognition of marine mammal sounds remains a challenging task. However, the marine environment is not quiet, and the signals detected by underwater sonar may contain a variety of signal information. For example, the noise emitted by various noise sources such as waves, distant ships, ocean turbulence, and marine biological activities is affected by factors such as water bodies, sea surfaces, and seabeds during propagation, forming an extremely complex marine environmental noise field. Weak target signals and environmental noise interference will have a serious impact on the recognition of target signals. Therefore, it is of great significance and research value to process the signals received by sonar, reduce interference noise, and enhance target signals.
[0003] Since high-frequency sound waves decay rapidly with increasing propagation distance, the high-frequency part of the ocean background noise spectrum is very weak, while the low-frequency part is quite strong. Some spectrum bands account for a relatively small proportion relative to the sample frequency, but contain more information about the ocean environmental noise and mammal signals. However, the traditional indifferent learning method that targets the entire frequency band will ignore this detail because it pays more attention to high frequencies. Therefore, the present invention proposes a cetacean signal enhancement method based on learning differences. Instead of learning the characteristics of the entire frequency band, the cetacean signal is divided into two frequency bands to learn two different frequency characteristics based on the characteristic differences between the ocean environmental noise and the typical marine mammal whistle signals. At the same time, the independence and consistency of each channel with the entire frequency band are considered, which greatly improves the denoising effect and signal enhancement. Summary of the invention
[0004] In order to solve the above problems, the present invention proposes a cetacean signal enhancement method based on learning differences.
[0005] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0006] The present invention provides a cetacean signal enhancement method based on learning differences, and the present invention provides the following technical solutions:
[0007] A cetacean signal enhancement method based on learning differences, the method comprising the following steps:
[0008] Step 1: Build a dataset of cetacean sound signals;
[0009] Step 2: Build a two-mask-DPCRN model;
[0010] Step 3: Perform network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model;
[0011] Step 4: Enhance the cetacean sound signal based on the signal enhancement model.
[0012] Preferably, the step 1 is specifically:
[0013] The collected cetacean sound signal data is downsampled and the signal is smoothed in the time domain using the least squares smoothing filter to reduce the interference of short-time pulse signals on spectrum analysis;
[0014] The cetacean sound signals and the ocean environment noise are cut or spliced into equal lengths, and A / D sampling is performed, and then the analog signals are converted into digital signals through the process of quantization and encoding;
[0015] Each time a time segment is taken to calculate the short-time power spectrum frame by frame to form a two-dimensional time-frequency diagram.
[0016] Preferably, the two-dimensional time-frequency diagrams of the plurality of marine mammal sound signals and the marine environment noise are saved in the form of a collection, and the clean signal and the noise signal are mixed at a certain signal-to-noise ratio as a noisy sample data set; the clean signal is independently saved as a clean sample data set.
[0017] Preferably, the step 2 is specifically:
[0018] The time-frequency graph of the two-mask-DPCRN model extracts local patterns through five layers of encoding, and reduces the feature resolution to form a two-dimensional tensor of size L*F; the block size is C and the block jump size is The sequential input of length L is cut into blocks of S, where C and S correspond to the input lengths of intra-chunk RNN and inter-chunk RNN respectively.
[0019] Connect all the blocks into a three-dimensional tensor of size S*F*C; where S is the time axis, F is the frequency axis and C is the channel axis;
[0020] After the three-dimensional tensor passes through the DPRNN module, it is split in the F dimension into low-frequency band and high-frequency band;
[0021] Among them, the calculation formula of the learning target mask1 in the low frequency band is:
[0022] Where S and Y are clean samples and noisy samples respectively;
[0023] Among them, the calculation formula of the learning target mask2 in the high frequency band is:
[0024] where Y r and Y i are the real and imaginary parts of the noisy sample respectively; S i and S r are the real and imaginary parts of the clean sample respectively;
[0025] Mask1 is used in the low-frequency band to learn the low-frequency ocean environment noise, focusing more on the improvement of the signal-to-noise ratio and focusing on the spectral amplitude of noisy speech.
[0026] Mask2 is used in the high-frequency band to learn the cetacean signals in the high-frequency band, focusing more on the phase complement of the cetacean signals themselves, which has the effect of signal enhancement.
[0027] The output of the last layer is converted back to a sequential output of length L using the overlap-add method;
[0028] The decoder uses a transposed convolutional layer to restore the low-resolution features to the original size, where the structure of the decoder is symmetrical to that of the encoder, and there are skip connections between the encoder and decoder to pass detailed information.
[0029] Preferably, the step 3 is specifically:
[0030] The training set of cetacean acoustic signals obtained in step 1 is used as the input of the two-mask-DPCRN model, and the initial weight values and feature parameters of the two-mask-DPCRN model are assigned using the truncated Gaussian distribution method;
[0031] According to the learning target mask1 and the learning target mask2, the loss function values Loss1 and Loss2 of the two paths are calculated respectively, and the two loss function values Loss1 and Loss2 are added according to the weights to obtain the loss function value Loss Total ;
[0032] Loss Total The calculation formula is expressed as follows:
[0033] Loss Total =αLoss1+βLoss2
[0034] Among them, α and β are the weight coefficients corresponding to the two loss functions.
[0035] Preferably, the gradient descent method of back propagation is used, an appropriate learning rate is set, and the weight values and feature parameters of the two-mask-DPCRN model are fine-tuned so that the parameters of each node gradually approach the optimal solution. The loss function Loss Total minimize;
[0036] Save the trained model parameters and verify the model using the validation set. When the verification result shows that the signal-to-noise ratio is lower than 5 dB, adjust the network parameters or the learning rate to retrain. Otherwise, save the final trained weights and parameter values to the two-mask-DPCRN model to obtain the final signal enhancement model, which is used as the marine mammal acoustic signal enhancement network.
[0037] Preferably, the learning rate is set to 0.001-0.003.
[0038] A cetacean signal enhancement system based on learning differences, the system comprising:
[0039] A data set module, wherein the data set module constructs a data set of cetacean sound signals;
[0040] A model building module, wherein the model building module builds a two-mask-DPCRN model;
[0041] A model training module, wherein the model training module performs network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model;
[0042] A signal enhancement module is used to enhance the cetacean sound signal based on a signal enhancement model.
[0043] A computer-readable storage medium stores a computer program, which is executed by a processor to implement a cetacean signal enhancement method based on learning differences.
[0044] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a cetacean signal enhancement method based on learning differences is implemented.
[0045] The present invention has the following beneficial effects:
[0046] Compared with the prior art, the present invention has the following advantages:
[0047] The present invention provides a cetacean signal enhancement method based on learning differences. The collected marine mammal sound signals are filtered as a clean sample data set; the marine environmental noise and the clean signal are mixed at a certain signal-to-noise ratio as a noisy sample data set; according to the characteristic differences between the marine environmental noise and the typical marine mammal whistle signal, two learning target synchronous training networks are constructed, and the network is trained using the data set to obtain a network that improves the denoising effect and enhances the effective signal.
[0048] Instead of learning the characteristics of the entire frequency band, the entire spectrum is divided into two frequency bands to learn two different frequency characteristics based on the characteristic differences between ocean environmental noise and typical marine mammal whistle signals. The signal characteristics of different frequency bands are learned more specifically, so that the network can effectively suppress the low-frequency ocean environmental noise while enhancing the high-frequency cetacean signals.
[0049] Since the network considers both the independence and consistency of each channel with the entire frequency band, through independence each channel can independently adjust its transmission parameters to minimize interference between different frequency bands to achieve optimal performance; through consistency it can ensure that each channel maintains consistent transmission performance across the entire frequency band, thereby improving the overall capacity of the network. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0051] Figure 1 It is the overall flow chart of the present invention;
[0052] Figure 2 Two-mask-DPCRN model structure diagram. DETAILED DESCRIPTION
[0053] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.
[0055] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0056] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0057] The present invention is described in detail below in conjunction with specific embodiments. Specific embodiment one:
[0059] according to Figure 1 to Figure 2 As shown, the specific optimization technical solution adopted by the present invention to solve the above technical problems is: the present invention relates to a cetacean signal enhancement method based on learning differences.
[0060] Step 1: Build a dataset of cetacean sound signals;
[0061] The step 1 is specifically as follows:
[0062] The collected cetacean sound signal data is downsampled and the signal is smoothed in the time domain using the least squares smoothing filter to reduce the interference of short-time pulse signals on spectrum analysis;
[0063] The cetacean sound signals and the ocean environment noise are cut or spliced into equal lengths, and A / D sampling is performed, and then the analog signals are converted into digital signals through the process of quantization and encoding;
[0064] Each time a time segment is taken to calculate the short-time power spectrum frame by frame to form a two-dimensional time-frequency diagram.
[0065] The two-dimensional time-frequency diagrams of the multiple marine mammal sound signals and the marine environmental noise are saved in the form of a collection, and the clean signal and the noise signal are mixed at a certain signal-to-noise ratio as a noisy sample data set; the clean signal is independently used as a clean sample data set.
[0066] Step 2: Build a two-mask-DPCRN model;
[0067] The step 2 is specifically as follows:
[0068] The time-frequency graph of the two-mask-DPCRN model extracts local patterns through five layers of encoding, and reduces the feature resolution to form a two-dimensional tensor of size L*F; the block size is C and the block jump size is The sequential input of length L is cut into blocks of S, where C and S correspond to the input lengths of intra-chunk RNN and inter-chunk RNN respectively.
[0069] Connect all the blocks into a three-dimensional tensor of size S*F*C; where S is the time axis, F is the frequency axis and C is the channel axis;
[0070] After the three-dimensional tensor passes through the DPRNN module, it is split in the F dimension into low-frequency band and high-frequency band;
[0071] Among them, the calculation formula of the learning target mask1 in the low frequency band is:
[0072] Where S and Y are clean samples and noisy samples respectively;
[0073] Among them, the calculation formula of the learning target mask2 in the high frequency band is:
[0074] where Y r and Y i are the real and imaginary parts of the noisy sample respectively; S i and S r are the real and imaginary parts of the clean sample respectively;
[0075] The output of the last layer is converted back to a sequential output of length L using the overlap-add method;
[0076] The decoder uses a transposed convolutional layer to restore the low-resolution features to the original size, where the structure of the decoder is symmetrical to that of the encoder, and there are skip connections between the encoder and decoder to pass detailed information.
[0077] Step 3: Perform network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model;
[0078] The step 3 is specifically as follows:
[0079] The training set of cetacean acoustic signals obtained in step 1 is used as the input of the two-mask-DPCRN model, and the initial weight values and feature parameters of the two-mask-DPCRN model are assigned using the truncated Gaussian distribution method;
[0080] According to the learning target mask1 and the learning target mask2, the loss function values Loss1 and Loss2 of the two paths are calculated respectively, and the two loss function values Loss1 and Loss2 are added according to the weights to obtain the loss function value Loss Total ;
[0081] Loss Total The calculation formula is expressed as follows:
[0082] Loss Total =αLoss1+βLoss2
[0083] Among them, α and β are the weight coefficients corresponding to the two loss functions.
[0084] Use the back-propagation gradient descent method, set an appropriate learning rate, and fine-tune the weight values and feature parameters of the two-mask-DPCRN model so that the parameters of each node gradually approach the optimal solution. The loss function Loss Total minimize;
[0085] Save the trained model parameters and verify the model using the validation set. When the verification result shows that the signal-to-noise ratio is lower than 5 dB, adjust the network parameters or the learning rate to retrain. Otherwise, save the final trained weights and parameter values to the two-mask-DPCRN model to obtain the final signal enhancement model, which is used as the marine mammal acoustic signal enhancement network.
[0086] The forward propagation of data and the back propagation of errors are classic algorithms for calculating extreme values. The step size of each fine-tuning is determined by setting an appropriate learning rate. When the learning rate is too large, the calculation result will hover around the optimal solution and the optimal solution will never be obtained; when the learning rate is too small, the pace of each iterative update is too small, and the optimal solution cannot be found after multiple rounds of updates. The present invention sets the learning rate to 0.001-0.003.
[0087] Step 4: Enhance the cetacean sound signal based on the signal enhancement model.
[0088] The present invention provides a cetacean signal enhancement system based on learning differences, the system comprising:
[0089] A data set module, wherein the data set module constructs a data set of cetacean sound signals;
[0090] A model building module, wherein the model building module builds a two-mask-DPCRN model;
[0091] A model training module, wherein the model training module performs network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model;
[0092] A signal enhancement module is used to enhance the cetacean sound signal based on a signal enhancement model. Specific embodiment 2:
[0094] The difference between the second embodiment of the present application and the first embodiment is that:
[0095] The present invention takes the acoustic signal of cetaceans as an example. Figure 1 As shown, a cetacean signal enhancement method based on learning differences includes the following steps:
[0096] Step 1: Build a dataset.
[0097] The collected cetacean sound signal data is downsampled to improve the processing speed; the signal is smoothed in the time domain using the least squares smoothing filter to reduce the interference of short-time pulse signals on spectrum analysis;
[0098] As the duration of cetacean sound signals has a certain range, the data containing the signals are intercepted and invalid data are removed.
[0099] The marine mammal sound signals and the marine environmental noise are cut or spliced into equal lengths, and A / D sampling is performed, and then the analog signals are converted into digital signals through the process of quantization and encoding;
[0100] Take one time segment each time and calculate the short-time power spectrum frame by frame (frame refers to each time point) to form a two-dimensional time-frequency diagram;
[0101] The two-dimensional time-frequency diagrams of the multiple marine mammal sound signals and the marine environmental noise are saved in the form of a collection, and the clean signal and the noise signal are mixed at a certain signal-to-noise ratio as a noisy sample data set; the clean signal is independently used as a clean sample data set.
[0102] A part of the two sample data sets is randomly selected as the training set, and the rest is used as the validation set, with a ratio of 6:1.
[0103] Step 2: Construct a two-mask-DPCRN network.
[0104] The DPCRN model is a time-frequency domain speech enhancement model based on deep learning. It combines the local pattern modeling capability of CNN and the long-term modeling capability of DPRNN. It is formed by replacing the RNN in CRN with the DPRNN module. DPRNN uses two RNN blocks: intra-chunk RNN is used to model the spectrum of a single time frame, and inter-chunk RNN is used to model the change of spectrum over time.
[0105] The two-mask-DPCRN model is as follows Figure 2 As shown in the figure, the time-frequency graph is encoded in five layers to extract local patterns, and the feature resolution is reduced to form a two-dimensional tensor of size L*F; the block size is C and the block jump size is The sequential input of length L is cut into S blocks, where C and S correspond to the input lengths of intra-chunk RNN and inter-chunk RNN respectively. All blocks are then connected into a three-dimensional tensor of size S*F*C, where S is the time axis, F is the frequency axis, and C is the channel axis.
[0106] After the three-dimensional tensor passes through the DPRNN module, it is split in the F dimension into low-frequency band and high-frequency band.
[0107] Among them, the calculation formula of the learning target mask1 in the low frequency band is:
[0108] Where S and Y are clean samples and noisy samples respectively.
[0109] Among them, the calculation formula of the learning target mask2 in the high frequency band is:
[0110] where Y r and Y i are the real and imaginary parts of the noisy sample respectively; S i and S r are the real and imaginary parts of the clean sample respectively.
[0111] The output of the last layer is converted back to a sequential output of length L using the overlap-add method.
[0112] The decoder then uses a transposed convolutional layer to restore the low-resolution features to their original size, where the structure of the decoder is symmetrical to that of the encoder, and there are skip connections between the encoder and decoder to pass detailed information.
[0113] Step 3: Train and verify the network.
[0114] The training set of cetacean sound signals obtained in step 1 is used as the input of the two-mask-DPCRN model.
[0115] The two-mask-DPCRN model is assigned initial weight values and feature parameters using the truncated Gaussian distribution method;
[0116] According to the learning target mask1 and the learning target mask2, the loss function values Loss1 and Loss2 of the two paths are calculated respectively, and the two loss function values Loss1 and Loss2 are added according to the weights to obtain the loss function value Loss Total .
[0117] Loss Total Calculation formula: Loss Total =αLoss1+βLoss2.
[0118] Among them, α and β are the weight coefficients corresponding to the two loss functions.
[0119] Use the back-propagation gradient descent method, set an appropriate learning rate, and fine-tune the weight values and feature parameters of the two-mask-DPCRN model so that the parameters of each node gradually approach the optimal solution. The loss function Loss Total minimize.
[0120] Save the trained model parameters and verify the model using the validation set. When the verification result shows that the signal-to-noise ratio is lower than 5 dB, adjust the network parameters or the learning rate to retrain. Otherwise, save the final trained weights and parameter values to the two-mask-DPCRN model to obtain the final signal enhancement model, which is used as the marine mammal acoustic signal enhancement network.
[0121] The forward propagation of data and the back propagation of errors are classic algorithms for calculating extreme values. The step size of each fine-tuning is determined by setting an appropriate learning rate. When the learning rate is too large, the calculation result will hover around the optimal solution and the optimal solution will never be obtained; when the learning rate is too small, the pace of each iterative update is too small, and the optimal solution cannot be found after multiple rounds of updates. The present invention sets the learning rate to 0.001-0.003. Specific embodiment four:
[0123] The difference between the fourth embodiment of the present application and the third embodiment is that:
[0124] The present invention provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement a cetacean signal enhancement method based on learning differences. Specific embodiment five:
[0126] The difference between the fifth embodiment of the present application and the fourth embodiment is that:
[0127] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a cetacean signal enhancement method based on learning differences is implemented.
[0128] The present invention provides a cetacean signal enhancement method based on learning differences. The collected marine mammal sound signals are filtered as a clean sample data set; the marine environmental noise and the clean signal are mixed at a certain signal-to-noise ratio as a noisy sample data set; according to the characteristic differences between the marine environmental noise and the typical marine mammal whistle signal, two learning target synchronous training networks are constructed, and the network is trained using the data set to obtain a network that improves the denoising effect and enhances the effective signal.
[0129] The method is specifically as follows:
[0130] A cetacean signal enhancement method based on learning differences comprises the following steps:
[0131] Step 1: Build the dataset:
[0132] The collected marine mammal sound signals are filtered to remove invalid data; the marine mammal sound signals and marine environmental noise are cropped or spliced into equal lengths, and A / D sampling is performed, and then the analog signals are converted into digital signals through a quantization and encoding process, and then the short-time power spectrum is calculated frame by frame to form a two-dimensional time-frequency diagram; multiple marine environmental noises and clean signals are mixed at a certain signal-to-noise ratio as a noisy sample data set, and the marine mammal sound signals are used as a clean sample data set; a portion of the two sample data sets is randomly selected as a training set, and the rest is used as a validation set, with a ratio of 6:1.
[0133] Step 2: Construct a two-mask-DPCRN network:
[0134] The module of the DPCRN model was redesigned. According to the characteristics of the ocean environment noise, the cetacean signal was divided into two frequency bands to learn two different frequency characteristics, which were divided into low-frequency band and high-frequency band to obtain the two-mask-DPCRN model, and the initial weight value and characteristic parameters of the two-mask-DPCRN model were assigned according to the truncated Gaussian distribution method.
[0135] Mask1 is used in the low-frequency band to learn the low-frequency ocean environment noise, focusing more on the improvement of the signal-to-noise ratio and focusing on the spectral amplitude of noisy speech.
[0136] Mask2 is used in the high-frequency band to learn the cetacean signals in the high-frequency band, focusing more on the phase complement of the cetacean signals themselves, which has the effect of signal enhancement.
[0137] Step 3: Train and verify the network:
[0138] The training set described in step 1 is used as the input of the two-mask-DPCRN model; the two-mask-DPCRN model is trained by using the forward propagation of the training set data and the back propagation of the error; the weight and parameter values with the minimum overall error are obtained; the two-mask-DPCRN model is verified by using the verification set, and if the verification result improves the signal-to-noise ratio by less than 5dB, the network parameters are adjusted and retrained, otherwise the weights and parameter values obtained by the final training are saved in the two-mask-DPCRN model to obtain the final signal enhancement model.
[0139] In the description of this specification, the description of reference terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction. In addition, the terms "first" and "second" are used only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined. Any process or method description in the flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code including one or more executable instructions for implementing the steps of a custom logic function or process, and the scope of the preferred embodiment of the present invention includes other implementations, in which the functions may not be performed in the order shown or discussed, including in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by a person skilled in the art of the art to which the embodiments of the present invention belong. The logic and / or steps represented in the flowchart or otherwise described herein, for example, may be considered as a sequenced list of executable instructions for implementing the logic function, and may be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection having one or N wirings (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM).In addition, the computer readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, then editing, interpreting or processing in other suitable ways as necessary, and then storing it in a computer memory. It should be understood that the various parts of the present invention can be implemented with hardware, software, firmware, or a combination thereof. In the above-mentioned embodiment, N steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented with hardware, as in another embodiment, any one of the following technologies known in the art or their combination can be used to implement: a discrete logic circuit with a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit with a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.
[0140] The above is only a preferred implementation of a cetacean signal enhancement method based on learning differences. The protection scope of a cetacean signal enhancement method based on learning differences is not limited to the above embodiments. All technical solutions under this idea belong to the protection scope of the present invention. It should be pointed out that for those skilled in the art, several improvements and changes without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A cetacean signal enhancement method based on learning differences, characterized by: The method comprises the following steps: Step 1: Build a dataset of cetacean sound signals; The two-dimensional time-frequency diagrams of multiple cetacean sound signals and ocean environmental noise are saved in the form of a collection, and the cetacean sound signals and ocean environmental noise are mixed at a certain signal-to-noise ratio as a noisy sample data set; the cetacean sound signals are independently used as a clean sample data set; Step 2: Build a two-mask-DPCRN model; The step 2 is specifically as follows: The time-frequency graph of the two-mask-DPCRN model extracts local patterns through five layers of encoding, and reduces the feature resolution to form a two-dimensional tensor of size L*F; the block size is C and the block jump size is The sequential input of length L is cut into blocks of S, where C and S correspond to the input lengths of intra-chunk RNN and inter-chunk RNN respectively. Connect all the blocks into a three-dimensional tensor of size S*F*C; where S is the time axis, F is the frequency axis and C is the channel axis; After the three-dimensional tensor passes through the DPRNN module, it is split in the F dimension into low-frequency band and high-frequency band; Among them, the calculation formula of the learning target mask1 in the low frequency band is: Where S and Y are clean samples and noisy samples respectively; Among them, the calculation formula of the learning target mask2 in the high frequency band is: where Y r and Y i are the real and imaginary parts of the noisy sample respectively; S i and S r are the real and imaginary parts of the clean sample respectively; The output of the last layer is converted back to a sequential output of length L using the overlap-add method; The decoder uses a transposed convolutional layer to restore the low-resolution features to their original size. The structure of the decoder is symmetrical with that of the encoder, and there are skip connections between the encoder and decoder to pass detailed information. Step 3: Perform network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model; Step 4: Enhance the cetacean sound signal based on the signal enhancement model.
2. The method according to claim 1, characterized in that: The step 1 is specifically as follows: The collected cetacean sound signal data is downsampled and the signal is smoothed in the time domain using the least squares smoothing filter to reduce the interference of short-time pulse signals on spectrum analysis; The cetacean sound signals and the ocean environment noise are cut or spliced into equal lengths, and A / D sampling is performed, and then the analog signals are converted into digital signals through the process of quantization and encoding; Each time a time segment is taken to calculate the short-time power spectrum frame by frame to form a two-dimensional time-frequency diagram.
3. The method according to claim 1, characterized in that: Mask1 is used in the low-frequency band to learn the low-frequency ocean environment noise, and mask2 is used in the high-frequency band to learn the high-frequency whale signal.
4. The method according to claim 1, characterized in that: The step 3 is specifically as follows: The training set of cetacean acoustic signals obtained in step 1 is used as the input of the two-mask-DPCRN model, and the initial weight values and feature parameters of the two-mask-DPCRN model are assigned using the truncated Gaussian distribution method; According to the learning target mask1 and the learning target mask2, the loss function values Loss1 and Loss2 of the two paths are calculated respectively, and the two loss function values Loss1 and Loss2 are added according to the weights to obtain the loss function value Loss Total ; Loss Total The calculation formula is expressed as follows: Loss Total =αLoss1+βLoss2 Among them, α and β are the weight coefficients corresponding to the two loss functions.
5. The method according to claim 4, characterized in that: Use the gradient descent method of back propagation, set an appropriate learning rate, fine-tune the weight values and feature parameters of the two-mask-DPCRN model, so that the parameters of each node gradually approach the optimal solution, and the loss function Loss Total minimize; Save the trained model parameters and use the validation set to validate the model. When the validation result shows that the signal-to-noise ratio is below 5 dB, adjust the network parameters or the learning rate to retrain. Otherwise, save the final training weights and parameter values to the two-mask-DPCRN model to obtain the final enhanced model, which is used as the marine mammal acoustic signal enhancement network. Set the learning rate to 0.001-0.
003.
6. A cetacean signal enhancement system based on learning differences, characterized by: The system comprises: A data set module, wherein the data set module constructs a data set of cetacean sound signals; The two-dimensional time-frequency diagrams of multiple cetacean sound signals and ocean environmental noise are saved in the form of a collection, and the cetacean sound signals and ocean environmental noise are mixed at a certain signal-to-noise ratio as a noisy sample data set; the cetacean sound signals are independently used as a clean sample data set; A model building module, wherein the model building module builds a two-mask-DPCRN model; The time-frequency graph of the two-mask-DPCRN model extracts local patterns through five layers of encoding, and reduces the feature resolution to form a two-dimensional tensor of size L*F; the block size is C and the block jump size is The sequential input of length L is cut into blocks of S, where C and S correspond to the input lengths of intra-chunk RNN and inter-chunk RNN respectively. Connect all the blocks into a three-dimensional tensor of size S*F*C; where S is the time axis, F is the frequency axis and C is the channel axis; After the three-dimensional tensor passes through the DPRNN module, it is split in the F dimension into low-frequency band and high-frequency band; Among them, the calculation formula of the learning target mask1 in the low frequency band is: Where S and Y are clean samples and noisy samples respectively; Among them, the calculation formula of the learning target mask2 in the high frequency band is: where Y r and Y i are the real and imaginary parts of the noisy sample respectively; S i and S r are the real and imaginary parts of the clean sample respectively; The output of the last layer is converted back to a sequential output of length L using the overlap-add method; The decoder uses a transposed convolutional layer to restore the low-resolution features to their original size. The structure of the decoder is symmetrical with that of the encoder, and there are skip connections between the encoder and decoder to pass detailed information. A model training module, wherein the model training module performs network training and verification on the two-mask-DPCRN model to obtain a signal enhancement model; A signal enhancement module is used to enhance the cetacean sound signal based on a signal enhancement model.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method according to any one of claims 1 to 5.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.
Citation Information
Patent Citations
Speech enhancement model calculation amount compression method based on recurrent neural network
CN115273874A
Convolutional recurrent neural network and speech enhancement method and device
CN115273883A
Underwater acoustic signal noise reduction method, system and device and storage medium
CN115359771A
Speech enhancement hybrid processing method based on convolutional recurrent network and WPE algorithm
CN115424627A
Speech recognition data enhancement method based on feature replacement and masking of spectrogram
CN115881112A
