An embedded platform-oriented underwater acoustic signal modal decomposition acceleration implementation method

Abstract

The application relates to the field of underwater acoustic signal processing and embedded computing technology, in particular to an underwater acoustic signal mode decomposition acceleration implementation method for an embedded platform, which comprises the following steps: in an initialization stage, a fixed work area is pre-applied, and feature sequences in a two-dimensional or multi-dimensional underwater acoustic spectrum matrix are processed in sequence; in a collection mode decomposition process, temporary mode components obtained through each decomposition are accumulated in a mode accumulation buffer area in real time, so that all collection times and all mode intermediate results are avoided from being saved; after mode screening and reconstruction of a current sequence are completed, the reconstruction result is written back to an output matrix in real time, and the work area is reused to process a next sequence. The method can reduce peak memory occupation, reduce dynamic memory application, and improve the operation efficiency and time delay stability of an embedded underwater acoustic processing platform.

Description

Technical Field

[0001] The embodiments of this application relate to the fields of underwater acoustic signal processing and embedded computing technology, and in particular to a method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms and a computer-readable storage medium. Background Technology

[0002] Empirical Mode Decomposition (EMD), as an adaptive signal processing method, can decompose nonlinear and non-stationary signals into multiple intrinsic mode functions (IMFs). It is widely used for local feature extraction, transient component analysis, and non-stationary feature separation in underwater acoustic signals. To alleviate the mode aliasing problem, improved methods such as Ensemble Empirical Mode Decomposition (EEMD) enhance the robustness of the decomposition by introducing multiple noise perturbations and result averaging.

[0003] However, in practical engineering applications such as underwater acoustic array signal processing and sonar echo analysis, the data to be processed is often represented as two-dimensional or multi-dimensional spectral matrices (such as time-frequency spectra and spatial spectra). Traditional implementations, when performing modal decomposition on each dimension of the matrix, typically store all intrinsic mode functions, intermediate residuals, noise perturbation results, and ensemble average temporary results for all sequences. When the matrix dimension is high or the number of ensemble decompositions is large, the storage scale of intermediate variables expands dramatically, leading to excessively high peak memory usage. This places enormous memory pressure on resource-constrained embedded underwater acoustic processing platforms (such as sonar equipment, unmanned platforms, and buoy nodes), increases data migration overhead, reduces cache hit rate, and severely restricts the real-time or near-real-time deployment capabilities of the algorithm.

[0004] Furthermore, in traditional mode decomposition, intermediate variables such as the extremum point array, envelope array, and spline interpolation working region require frequent dynamic memory allocation and deallocation, which easily leads to memory fragmentation and uncontrollable fluctuations in processing latency. Simultaneously, due to differences in noise levels and modal complexity among different underwater acoustic input sequences, traditional adaptive sieving stopping criteria result in varying iteration counts for each sequence, further exacerbating latency instability.

[0005] Therefore, how to effectively reduce peak memory usage, decrease dynamic memory allocation, and improve the stability of processing latency and the adaptability of engineering deployment while ensuring the decomposition and reconstruction effects are the technical problems that urgently need to be solved. Summary of the Invention

[0006] In view of this, embodiments of this application propose an accelerated implementation method for underwater acoustic signal mode decomposition for embedded platforms. This method aims to solve at least one of the technical problems of existing empirical mode decomposition and its ensemble extension methods on embedded underwater acoustic processing platforms, namely, high peak memory consumption, frequent dynamic memory requests, and unstable processing latency, which make it difficult to meet real-time or near-real-time processing requirements. This method can significantly reduce peak memory consumption, reduce the number of dynamic memory requests, and improve the operating efficiency, processing latency stability, and engineering deployment adaptability of the embedded underwater acoustic processing platform.

[0007] To achieve the above objectives, embodiments of this application propose a method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms, the method comprising the following steps: The feature matrix of the underwater acoustic signal to be processed is obtained, and during the program initialization phase, a fixed working area for mode decomposition is determined based on preset parameters. The size of the fixed working area remains constant. The preset parameters include sequence length, maximum mode order, ensemble average degree, and the size of temporary variables required for envelope interpolation. Extract the current feature sequence sequentially from the underwater acoustic signal feature matrix and write it into the input sequence buffer of the fixed working area; Within a fixed working area, mode decomposition is performed on the current feature sequence, and the temporary mode components obtained from each decomposition are immediately accumulated into the mode accumulation buffer within the fixed working area. After completing the mode decomposition of the current feature sequence, the accumulated results in the mode accumulation buffer are averaged to obtain the intrinsic mode functions of the current sequence. The intrinsic mode functions are filtered, and the filtered mode components are reconstructed. The reconstruction results are written back to the output matrix, and the fixed working area is reused to process the next feature sequence.

[0008] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the above-described method for accelerating the underwater acoustic signal mode decomposition for embedded platforms.

[0009] The embodiments of this application propose an accelerated implementation method for underwater acoustic signal mode decomposition for embedded platforms. First, a fixed working area for mode decomposition is pre-allocated, and feature sequences are extracted sequentially. Then, the temporary modal components obtained from each decomposition are immediately accumulated into the mode accumulation buffer. Next, after the decomposition is completed, the accumulated results are averaged to obtain the intrinsic mode functions. The selected modal components are reconstructed and written back to the output matrix in real time. The working area is then reused to process the next sequence.

[0010] In this application, firstly, compared to traditional methods that require storing intermediate results for all sequences, all modal orders, and all set exponents, this application processes sequences sequentially and accumulates temporary modal data in real time. It only needs to store the modal accumulation result of the currently processed sequence within a fixed working area, without needing to store any intermediate modal data for other sequences. Therefore, it only needs to maintain one fixed working area, significantly reducing the scale of modality-related storage. Secondly, compared to traditional methods that require repeated dynamic allocation and release of temporary memory such as extreme point arrays, envelope arrays, and spline interpolation working areas when processing each sequence, this application allocates its working area once during the initialization phase, and does not re-allocate it during subsequent sequence processing. Furthermore, dynamic memory allocation and release avoid memory fragmentation, thereby reducing dynamic memory requests, eliminating memory fragmentation, and improving operational stability. Third, by reusing a fixed working area, this application eliminates latency fluctuations caused by frequent dynamic memory requests and releases, making the processing time for each sequence tend to be constant, improving the stability and predictability of processing latency, which is beneficial for real-time or near-real-time system deployment. Fourth, by writing back the reconstruction results immediately, this application avoids the operation of temporarily storing intermediate results in an additional large-capacity buffer before batch writing, thereby reducing data movement and cache access overhead, and improving the overall processing speed on the embedded platform.

[0011] Based on this, this application can solve at least one of the technical problems of existing empirical mode decomposition and its set extension methods on embedded underwater acoustic processing platforms, namely high peak memory usage, frequent dynamic memory requests, and unstable processing latency, which make it difficult to meet real-time or near-real-time processing requirements. In this way, it can significantly reduce peak memory usage, reduce the number of dynamic memory requests, and improve the operating efficiency, processing latency stability, and engineering deployment adaptability of the embedded underwater acoustic processing platform.

[0012] Optionally, the feature matrix of the underwater acoustic signal to be processed The expression is as follows: ; in, This represents the number of one-dimensional sequences to be decomposed. Indicates the length of each sequence; Indicates the sequence number. Indicates the sampling point number within the sequence; The underwater acoustic signal feature matrix is ​​stored in one-dimensional contiguous memory, and the underwater acoustic signal feature matrix, output matrix, and sequence buffer within the fixed working area adopt a consistent column-major contiguous storage layout. Matrix elements... linear address It is expressed as follows: .

[0013] Optionally, the fixed working area for mode decomposition includes: an input sequence buffer, a residual buffer, a current mode buffer, a mode accumulation buffer, an upper envelope buffer, a lower envelope buffer, an envelope mean buffer, an extreme point location buffer, a mode statistical contribution buffer, a sorting index buffer, and a reconstructed sequence buffer.

[0014] Optionally, the extreme point location buffer includes a local maximum point location buffer and a local minimum point location buffer; fixed working area The expression is as follows: ; in, This is the buffer for the current input sequence. As a residual buffer, For the current modal buffer, For modal accumulation buffer, For the upper envelope buffer, For the lower envelope buffer, As a buffer for the envelope mean, This serves as a buffer zone for the location of local maxima. This serves as a buffer zone for the location of local minima. As a buffer for modal statistical contribution, For sorting index buffer, To reconstruct the sequence buffer.

[0015] Optionally, within a fixed working area, mode decomposition is performed on the current feature sequence, and the temporary mode components obtained from each decomposition are immediately accumulated into the mode accumulation buffer within the fixed working area, including: If the current feature sequence is the th The expression for a one-dimensional underwater acoustic feature sequence is as follows: ; Based on ensemble empirical mode decomposition or complementary ensemble empirical mode decomposition, mode decomposition is performed on the current feature sequence, and after each mode decomposition, for the ... For each of the first-order modal components, perform the following accumulation operation: ; in, Represents the first decomposition obtained in a single step. Temporal modal components, Indicates the first A first-order modal accumulation buffer.

[0016] Optionally, the maximum modal order is set to: ; in, This represents the floor operation; The number of screening iterations for each mode is set to a fixed number or a preset upper limit.

[0017] Optionally, the intrinsic mode functions are screened, including: calculating the statistical contribution of each order of intrinsic mode functions; sorting the intrinsic mode functions of each order according to their statistical contribution, and selecting the set of modes whose cumulative contribution reaches a preset threshold.

[0018] Optionally, the selected modal components are reconstructed to write the reconstruction results back to the output matrix, including: The modal components in the modality set are directly accumulated into the reconstruction buffer, as expressed below: ; in, Represents a set of modes. ; Represents the reconstructed sequence. Indicates the first The intrinsic mode function of order.

[0019] Optionally, reusing the fixed work area to process the next feature sequence includes: after processing the current sequence, clearing or overwriting each buffer in the fixed work area for use in processing the next feature sequence. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0021] Figure 1 This is a flowchart of a method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms provided in this application; Figure 2 This is an overall flowchart of a method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms provided in this application; Figure 3 This is a comparison diagram of full storage method and low memory storage method in underwater acoustic signal mode decomposition provided in this application; Figure 4 This is a schematic diagram of a fixed working area multiplexing structure for underwater acoustic signal mode decomposition provided in this application; Figure 5This is a schematic diagram of the real-time accumulation of temporary modes and real-time writing back of underwater acoustic reconstruction results provided in this application; Figure 6 This is a comparison diagram of the modal decomposition results of underwater acoustic signals on an embedded platform and a computer, as provided in this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0023] Empirical Mode Decomposition (EMD), as an adaptive signal processing method, can decompose nonlinear and non-stationary signals into multiple intrinsic mode functions (IMFs). It is widely used for local feature extraction, transient component analysis, and non-stationary feature separation in underwater acoustic signals. To alleviate the mode aliasing problem, improved methods such as Ensemble Empirical Mode Decomposition (EEMD) enhance the robustness of the decomposition by introducing multiple noise perturbations and result averaging.

[0024] However, in practical engineering applications such as underwater acoustic array signal processing and sonar echo analysis, the data to be processed is often represented as two-dimensional or multi-dimensional spectral matrices (such as time-frequency spectra and spatial spectra). Traditional implementations, when performing modal decomposition on each dimension of the matrix, typically store all intrinsic mode functions, intermediate residuals, noise perturbation results, and ensemble average temporary results for all sequences. When the matrix dimension is high or the number of ensemble decompositions is large, the storage scale of intermediate variables expands dramatically, leading to excessively high peak memory usage. This places enormous memory pressure on resource-constrained embedded underwater acoustic processing platforms (such as sonar equipment, unmanned platforms, and buoy nodes), increases data migration overhead, reduces cache hit rate, and severely restricts the real-time or near-real-time deployment capabilities of the algorithm.

[0025] Furthermore, in traditional mode decomposition, intermediate variables such as the extremum point array, envelope array, and spline interpolation working region require frequent dynamic memory allocation and deallocation, which easily leads to memory fragmentation and uncontrollable fluctuations in processing latency. Simultaneously, due to differences in noise levels and modal complexity among different underwater acoustic input sequences, traditional adaptive sieving stopping criteria result in varying iteration counts for each sequence, further exacerbating latency instability.

[0026] Therefore, how to effectively reduce peak memory usage, decrease dynamic memory allocation, and improve the stability of processing latency and the adaptability of engineering deployment while ensuring the decomposition and reconstruction effects are the technical problems that urgently need to be solved.

[0027] In view of this, embodiments of this application propose an accelerated implementation method for underwater acoustic signal mode decomposition for embedded platforms. This method aims to solve at least one technical problem existing in empirical mode decomposition and its ensemble extension methods on embedded underwater acoustic processing platforms, such as high peak memory usage, frequent dynamic memory requests, and unstable processing latency, which makes it difficult to meet real-time or near-real-time processing requirements. This method can significantly reduce peak memory usage, reduce the number of dynamic memory requests, and improve the operating efficiency, processing latency stability, and engineering deployment adaptability of the embedded underwater acoustic processing platform.

[0028] One embodiment of this application proposes a method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms, applied to embedded underwater acoustic processing devices. These embedded underwater acoustic processing devices can be, for example, sonar devices, unmanned platforms, or buoy nodes. This embodiment and subsequent embodiments use embedded underwater acoustic processing devices as examples for illustration. The implementation details of the proposed method for accelerating underwater acoustic signal mode decomposition for embedded platforms are described below. These details are provided for ease of understanding and are not essential for implementing this solution.

[0029] The specific process of the underwater acoustic signal mode decomposition acceleration method for embedded platforms proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain the feature matrix of the underwater acoustic signal to be processed, and determine the fixed working area for mode decomposition based on preset parameters during the program initialization phase.

[0030] The size of the fixed working area remains constant; preset parameters include sequence length, maximum modal order, ensemble average degree, and the size of temporary variables required for envelope interpolation.

[0031] In practical applications, the size of the fixed working area will be adjusted accordingly as preset parameters such as sequence length and maximum modal order change, but it can be smaller than that of traditional storage methods. For details, please refer to the following practical application examples, such as embedded underwater acoustic processing applications.

[0032] It should be noted that the fixed size of the fixed workspace here refers to the memory pre-allocation being completed all at once according to preset parameters during the program initialization phase. In essence, the fixed workspace provided in this application actually sacrifices a small amount of flexibility, namely the inability to dynamically expand, in exchange for controllable memory usage, stable latency, and improved operational reliability.

[0033] Specifically, the first step is to obtain the feature matrix of the underwater acoustic signal to be processed. This matrix can be a two-dimensional or multi-dimensional spectral matrix; for example, in broadband spatial spectrum processing scenarios, this matrix can be of size [missing information]. Two-dimensional spectral matrix ;in, This represents the number of one-dimensional sequences to be decomposed. This represents the length of each sequence. In broadband spatial spectrum processing scenarios, It can represent the number of glancing angles. It can represent the number of frequency points.

[0034] For example, to obtain the feature matrix of the underwater acoustic signal to be processed: ; in, This represents the number of one-dimensional sequences to be decomposed. This represents the length of each sequence. In broadband spatial spectrum processing scenarios, It can represent the number of glancing angles. It can represent the number of frequency points; Indicates the sequence number. Indicates the sampling point number within the sequence; To optimize memory access efficiency, the underwater acoustic signal feature matrix is ​​preferably stored using a one-dimensional contiguous memory method. For example, taking column-major storage as an example, matrix elements... linear address for: .

[0035] The specific storage order of the underwater acoustic signal feature matrix in this application can be row-major or column-major, but the input matrix, output reconstruction matrix and sequence buffer in the fixed working area are required to adopt a consistent column-major contiguous storage layout to reduce the overhead of data format conversion and cache access.

[0036] During program initialization, a fixed-size working area is pre-allocated according to preset parameters for modality decomposition operations on all subsequent sequences. This fixed working area is not released throughout the entire processing cycle, thus avoiding frequent dynamic memory allocation and deallocation. The preset parameters include at least: sequence length. Maximum modal order , average number of sets The size of the temporary variables required for envelope interpolation. For example, the size of the temporary variables required for envelope interpolation can refer to the length and number of temporary arrays required for envelope interpolation.

[0037] The fixed working area used for mode decomposition includes: input sequence buffer, residual buffer, current mode buffer, mode accumulation buffer, upper envelope buffer, lower envelope buffer, envelope mean buffer, extreme point location buffer, mode statistical contribution buffer, sorting index buffer, and reconstructed sequence buffer. The extreme point location buffer can include local maximum point location buffer and local minimum point location buffer.

[0038] For example, fixed work area Specifically, it contains multiple sub-buffers for storing different intermediate variables, and its structure can be represented as follows: ; in, This is the buffer for the current input sequence. As a residual buffer, This is the current modality buffer, also known as the current intrinsic modality function buffer. For modal accumulation buffer, For the upper envelope buffer, For the lower envelope buffer, As a buffer for the envelope mean, This serves as a buffer zone for the location of local maxima. This serves as a buffer zone for the location of local minima. As a buffer for modal statistical contribution, For sorting index buffer, To reconstruct the sequence buffer.

[0039] Based on preset parameters, the fixed work area can be calculated. The total memory required. That is, it can be determined during program initialization based on the sequence length. Maximum modal order , average number of sets The size of the temporary variables required for envelope interpolation is determined by pre-allocating a fixed working area of ​​a fixed size for underwater acoustic signal mode decomposition.

[0040] Specifically, fixed work area The total memory size is obtained by adding the following parts: The first part is the modal accumulation buffer. The memory occupied is equal to the sequence length. Multiply by the maximum modal order Multiply by the number of bytes occupied by the data type; The second part is the memory occupied by the single-sequence class buffer, including the input sequence buffer. Residual buffer Current intrinsic mode function buffer Upper envelope buffer Lower envelope buffer Envelope mean buffer Reconstructing the sequence buffer The size of each single-sequence buffer is the sequence length. Multiplying by the number of bytes occupied by the data type, the total size of this part is... Multiply by the number of single-sequence buffers and then multiply by the number of bytes occupied by the data type; The third part is the memory occupied by the extreme point location buffer, including the local maximum point location buffer. and local minimum point location buffer The size of each buffer is equal to the maximum number of extreme points multiplied by the number of bytes occupied by the index type, where the maximum number of extreme points does not exceed the sequence length. The index type is unsigned integer, and each index occupies 4 bytes; Part Four is the modal statistical contribution buffer. The memory occupied is equal to the maximum modal order. Multiply by the number of bytes occupied by the data type; The fifth part is the sorting index buffer. The memory occupied is equal to the maximum modal order. Multiply by the number of bytes occupied by the index type; Part VI specifies the size of the temporary variables required for envelope interpolation, including the node array, slope array, right-hand side term array, and second derivative array, which is approximately the length of the sequence. Multiply by 4 and then by the number of bytes occupied by the data type.

[0041] The sum of the above six parts constitutes the fixed work area. The total memory size. Each data type occupies 8 bytes when it is a double-precision floating-point number, and 4 bytes when it is a single-precision floating-point number.

[0042] Taking the 401x151 two-dimensional broadband spatial spectrum matrix of the embodiment in the specification as an example, the sequence length N=151, and the maximum modal order The data type uses double-precision floating-point, each occupying 8 bytes. The maximum number of extreme points is 75 (approximately half the sequence length). The index type is unsigned integer, each occupying 4 bytes. The size of the cubic spline interpolation temporary variable is approximately 151 multiplied by 4 and then multiplied by 8, equaling 4832 bytes. Substituting the above parts into the calculation, we can obtain the fixed working area. The overall size is controlled within the range of several kilobytes (KB) to tens of KB. Combining the memory requirements of the input and output matrices, the overall storage requirement of this application is approximately 1 megabyte (MB), which reduces the overall memory usage by about 75% to 80% compared to the traditional full-size storage method; in terms of modality-related storage scale, from Approximately reduced It equals 401 times.

[0043] It should be noted that the following buffers are used: The current input sequence buffer stores a single sequence to be processed extracted from the feature matrix. The residual buffer stores the residual signal after each sieving step. The current intrinsic mode function (IMF) buffer stores a single IMF component obtained from a single decomposition. The mode accumulation buffer accumulates temporary mode components of the same order obtained from each decomposition step during the ensemble decomposition process. The upper and lower envelope buffers store the upper and lower envelopes obtained through extreme point interpolation. The envelope mean buffer stores the average value of the upper and lower envelopes. The extreme point location buffer records the locations of local maxima and minima. The mode statistical contribution buffer stores the energy contribution of each order IMF. The sorting index buffer sorts the mode contributions. The reconstructed sequence buffer stores the reconstruction results of the filtered mode components.

[0044] The aforementioned workspace is reused across different sequences. After the current sequence is processed, the memory is not released. Instead, the next sequence is processed by clearing, overwriting, or resetting the index.

[0045] Understandably, traditional methods require repeated dynamic allocation and deallocation of temporary memory, such as extreme point arrays, envelope arrays, and spline interpolation working areas, when processing each sequence. This application allocates a fixed-size working area once during program initialization and reuses this working area in the processing of all subsequent sequences, eliminating the need for new dynamic memory allocation and fundamentally eliminating the memory fragmentation problem caused by frequent dynamic memory allocation and deallocation. Simultaneously, it avoids the risk of failure on every allocation attempt, making the program more stable and reliable, especially suitable for embedded underwater acoustic processing devices that require long-term uninterrupted operation.

[0046] Step 102: Extract the current feature sequence from the underwater acoustic signal feature matrix in sequence and write it into the input sequence buffer of the fixed working area.

[0047] For the Extracting one-dimensional underwater acoustic feature sequences: ; It is understandable that the sequence data is written to the input sequence buffer in the fixed work area. After that, all mode decomposition operations for this sequence will be performed within this fixed working area, such as extreme point detection, envelope construction, envelope mean calculation, sieving iteration, and residual update.

[0048] The temporary variables in the above process all reuse the same set of buffers and do not store intermediate results for a long time.

[0049] Step 103: Within the fixed working area, perform mode decomposition on the current feature sequence and immediately accumulate the temporary mode components obtained from each decomposition into the mode accumulation buffer within the fixed working area.

[0050] Within a fixed working area, mode decomposition is performed on the current sequence. The mode decomposition method here can be empirical mode decomposition or its ensemble extensions, such as ensemble empirical mode decomposition or complementary ensemble empirical mode decomposition.

[0051] For example, when using ensemble empirical mode decomposition, a predetermined number of noisy decomposition steps are performed. Each decomposition process is completed within a fixed working area and includes sub-steps such as extreme point detection, envelope construction, envelope mean calculation, sieving iteration, and residual update. Temporary variables generated by all these sub-steps reuse the same set of buffers within the working area and are not stored permanently.

[0052] For example, in complementary ensemble empirical mode decomposition, a pair of white noises with the same amplitude but opposite signs are generated each time, and positive and negative perturbation input sequences are constructed and decomposition is performed respectively; the temporary mode components of the same order obtained from the positive and negative decompositions are accumulated into the mode accumulation buffer; after the decomposition of all ensemble degrees is completed, the accumulated result is divided by twice the ensemble average degree to obtain the final intrinsic mode function.

[0053] To clearly illustrate the algorithmic foundation upon which this application is based, the basic processes of Ensemble Empirical Mode Decomposition (EEMD) and Complementary Ensemble Empirical Mode Decomposition (CEEMD) employed in this application are briefly described below as a reference for existing technologies: The basic process of EEMD includes: S1, determining the number of times N_E Gaussian white noise is added and the amplitude of the white noise; S2, performing EMD decomposition on the signal with Gaussian white noise added each time to obtain a set of intrinsic mode functions (IMFs); S3, repeating S1 and S2, adding a new white noise sequence each time; S4, summing and averaging the IMF components of the same order obtained each time to obtain the final intrinsic mode functions. Since the mean of Gaussian white noise is 0, the noise cancels each other out after integration and averaging.

[0054] The basic process of CEEMD includes: processing the raw signal Add a pair of white noises with the same amplitude but opposite signs to construct positive and negative disturbance inputs; perform EMD decomposition on the positive and negative disturbance inputs respectively; after completing the decomposition of all sets, two sets of eigenmode functions are obtained; the final eigenmode function is obtained by averaging the positive and negative sets.

[0055] When using EEMD, complete After the first decomposition with added noise, the second... The first-order eigenmode function is: ; When using CEEMD, a pair of white noises with the same amplitude but opposite signs is generated each time to construct positive and negative perturbation input sequences, and decomposition is performed separately. The temporary mode components of the same order obtained from the positive and negative decompositions are accumulated to... In the middle; in the completion After the set decomposition, the th The first-order eigenmode function is: .

[0056] It should be noted that the inventive point of this application is that, based on the above-mentioned EEMD or CEEMD basic process, a memory management architecture that uses fixed work area reuse, sequential processing, real-time accumulation of temporary modes, and real-time writing back of reconstruction results reduces the peak memory usage on the embedded platform and improves the stability of processing latency, without changing the mathematical decomposition principle of EEMD or CEEMD itself.

[0057] It is worth mentioning that, in this application, the temporary modal components obtained from each noisy decomposition are... Instead of being saved to a separate array, it is directly accumulated into the modality accumulation buffer corresponding to the current sequence. And after each modality decomposition, for the ... For each of the first-order modal components, perform the following accumulation operation: Specifically, for the first For the first modal components, perform an accumulation operation: ; in, Represents the first decomposition obtained in a single step. Temporal modal components, Indicates the first A first-order modal accumulation buffer.

[0058] After the accumulation is complete, the temporary modal components The buffer used can then be overwritten and reused in the next decomposition process. In this way, there is no need to save intermediate results for all set counts, all sequences, and all modal orders, which greatly saves memory space.

[0059] Step 104: After completing the mode decomposition of the current feature sequence, average the accumulated results in the mode accumulation buffer to obtain the intrinsic mode function of the current sequence.

[0060] After the preset number of decompositions and accumulations are completed, We perform an average to obtain the th digit of the current sequence. The intrinsic mode function of order.

[0061] After the set decomposition has been completed for the preset number of times, the accumulated results in the modal accumulation buffer can be averaged, that is, averaged... By averaging, we can obtain the first digit of the current sequence. The order intrinsic mode function. For example, for the th order intrinsic mode function. The first mode, whose final IMF is: ; in, This represents the average frequency of a set, which can be a preset frequency.

[0062] To control processing latency and improve time predictability, the maximum modal order... Based on sequence length Configure it, for example, set it to: ; in, This represents the floor operation; It should be noted that when the sequence length is At that time, based on the principle of signal binary division, the order of the IMF that can be extracted by empirical mode decomposition theoretically does not exceed [a certain value]. The subtraction of 1 is to remove residual trend terms, hence the maximum modal order is set above. The formula. For example, if ,but .

[0063] For example, the number of screening iterations for each mode can be set to a fixed number of 3 to 10 times, preferably 5 times.

[0064] Meanwhile, the number of screening iterations for each mode no longer uses an adaptive stopping criterion, but is set to a fixed number or a preset upper limit. This effectively reduces the fluctuations in computational overhead caused by differences in noise levels, local fluctuations, and modal complexity between different underwater acoustic input sequences, thereby improving the predictability of processing time. The preset upper limit can be a fixed number of screening iterations.

[0065] Step 105: Filter the intrinsic mode functions and reconstruct the filtered mode components to write the reconstruction results back to the output matrix, and reuse the fixed working area to process the next feature sequence.

[0066] After performing mode decomposition on the current sequence, we obtain Individual eigenmode functions Then, the modality screening and reconstruction stage begins. Among these, .

[0067] In one possible embodiment, the intrinsic mode functions are screened by: calculating the statistical contribution of each order of intrinsic mode functions; sorting the intrinsic mode functions according to their statistical contribution, and selecting the set of modes whose cumulative contribution reaches a preset threshold.

[0068] For example, the preset threshold can be a pre-configured value, such as 0.5.

[0069] Specifically, firstly, the statistical contribution of each intrinsic mode function (IMF) is calculated and stored in a contribution buffer. Then, using a sorting index buffer, all IMFs are sorted from largest to smallest contribution. Next, based on a preset contribution rate threshold (i.e., a preset threshold), the top Q IMFs whose cumulative contribution rate reaches the threshold are selected to form the mode set for reconstruction.

[0070] For example, when the statistical contribution is the energy contribution rate of each order intrinsic mode function, the th The formulas for calculating the energy contribution rate of an IMF include: ; ; in, Indicates the first The signal energy of the eigenmode function of order 1; This represents the total modal order of the current sequence.

[0071] According to the energy contribution rate of each eigenmode function The intrinsic mode functions (EMFs) of each order are sorted from largest to smallest, and the set of modes whose cumulative contribution rate reaches a preset threshold is selected. A larger preset threshold results in higher reconstruction accuracy but also a larger number of modes involved in the reconstruction; a smaller preset threshold results in faster reconstruction but may lose some detailed features. This application uses a preset threshold of 0.5 for illustrative purposes.

[0072] The modal components in the modality set are directly accumulated into the reconstruction buffer, as expressed below: ; in, Represents a set of modes. ; Represents the reconstructed sequence. Indicates the first The intrinsic mode function of order.

[0073] After reconstruction is complete, the reconstruction results are immediately written back to the output matrix. The corresponding position: ; In one possible embodiment, reusing a fixed work area to process the next feature sequence includes: after processing the current sequence, clearing or overwriting each buffer in the fixed work area for use in processing the next feature sequence.

[0074] Specifically, after the write-back operation is completed, all buffers corresponding to the current sequence are cleared or overwritten to prepare for processing the next (the first) entry. (Item) feature sequence can be reused.

[0075] Through the above steps, the storage size for modality-related functions in the traditional method is: This application reduces it to Furthermore, here This indicates the size of the fixed work area, which is much smaller than .

[0076] Understandably, compared with existing technologies, this application reduces intermediate mode data storage, lowers peak memory usage and dynamic memory allocation by using fixed work area reuse, sequential mode decomposition, real-time accumulation of temporary modes, and real-time writing back of reconstruction results. This improves the operating efficiency, stability, and latency predictability of the embedded underwater acoustic processing platform and is suitable for underwater acoustic signal mode decomposition and feature reconstruction under resource-constrained conditions.

[0077] The overall low-memory acceleration process in this embodiment can be found in the following reference. Figure 2 The diagram shows the overall flowchart of an accelerated implementation method for underwater acoustic signal mode decomposition for embedded platforms, which is divided into three processing stages.

[0078] The first stage is the construction of the underwater acoustic signal feature matrix input and fixed working area stage. It receives the two-dimensional or multi-dimensional underwater acoustic signal feature matrix from the external input and completes the unified storage layout of the feature matrix using a one-dimensional continuous memory method. At the same time, in the program initialization stage, based on preset parameters such as sequence length, maximum modal order, ensemble average degree and envelope interpolation temporary variable size, the memory allocation of the fixed working area and the sub-buffer division are completed in one go, providing a reusable fixed memory space for the subsequent processing of the full sequence.

[0079] The second stage is the low-memory mode decomposition stage of underwater acoustic signals, which extracts individual feature sequences sequentially from the stored feature matrix and sends them to a fixed working area for mode decomposition. During each noise perturbation process of the ensemble mode decomposition, the temporary modal components of each order obtained from a single decomposition are immediately accumulated to the corresponding order of the mode accumulation buffer. There is no need to save the independent intermediate results of each decomposition. All intermediate operations such as extreme point detection, envelope fitting, and sieving iteration are completed in the reuse buffer of the fixed working area, thereby significantly reducing the peak memory usage of mode correlation.

[0080] The third stage is the underwater acoustic signal mode screening and reconstruction stage and the real-time write-back stage. After the decomposition of all set times for the current sequence is completed, the effective modes are screened and determined based on the statistical contribution of each mode. The selected target mode components are directly accumulated and reconstructed to obtain the mode reconstruction result of the current sequence. The reconstruction result is then written back to the corresponding position of the output matrix in real time. After the processing of the current sequence is completed, the buffers in the fixed working area are cleared or overwritten and reset. The process returns to the sequence-by-sequence extraction stage to continue processing the next feature sequence until the mode decomposition and reconstruction output of all feature sequences in the feature matrix are completed.

[0081] Next, this application will be described using a specific embedded underwater acoustic processing application as an example.

[0082] Assume the input is a Two-dimensional broadband spatial spectrum matrix ( Complementary ensemble empirical mode decomposition was employed, with an ensemble average degree of 50, a total decomposition degree of 100 for both positive and negative perturbations, and a maximum mode order of [missing value]. .

[0083] First, the two-dimensional spectral matrix Stored in a contiguous block of memory. If a column-major layout is used, then the... The first angle, the second The storage locations corresponding to each frequency point are: This method enables the input matrix, output matrix, and intermediate processing to maintain a consistent data access convention, reducing data format conversion and additional data transfer.

[0084] like Figure 3 As shown in (a), traditional full-scale storage methods typically require saving the entire angle sequence, all modal orders, and temporary modal results during set decomposition. The ensemble average degree is 50, the total degree of complementary positive and negative perturbation decomposition is 100, and the maximum mode order is... Taking double-precision data type as an example, if all intrinsic mode functions of the entire sequence are stored, the mode-dependent storage size is: .

[0085] If all IMFs of all sequences are stored, the corresponding memory is approximately 401*6*151*8 bytes ≈ 3.70MB; if all temporary modal results in the set decomposition process are further stored, the modality-related storage scale corresponds to approximately 369.6 megabytes (MB) of memory.

[0086] In contrast, this application only saves the current sequence and a fixed working area, such as Figure 3 As shown in (b), in the low-memory storage method, the modal accumulation buffer size corresponds to approximately 9.4 kilobytes (KB) of memory for the current sequence. Further combining the input sequence buffer, residual buffer, upper envelope buffer, lower envelope buffer, statistical sorting buffer, reconstruction buffer, and spline interpolation working area, the overall size of the fixed working area can still be controlled within tens of KB. Including the input and output matrices, the overall storage requirement of this application is approximately 1 MB.

[0087] Therefore, compared with the traditional implementation that stores the final intrinsic mode functions of the entire sequence, the overall memory usage of this application can be reduced by approximately 75% to 80%; in terms of mode-related storage scale, this application is... Reduced to It decreased by approximately 401 times.

[0088] Next, during program initialization, a fixed working area for modal decomposition is allocated all at once. This working area includes an input sequence buffer, a residual buffer, a current modality buffer, a modality accumulation buffer, an upper envelope buffer, a lower envelope buffer, an extreme point buffer, a spline interpolation working area, a statistical sorting buffer, and a reconstruction buffer. This working area is reused between different angle sequences. After the current sequence is processed, the memory is not released but is reused for the next sequence by clearing, overwriting, or resetting the index, thereby reducing dynamic memory allocation, memory fragmentation, and processing latency fluctuations.

[0089] During processing, the program extracts the frequency sequence line by line according to the angular dimension. For the first... From each angle, the frequency sequence in the two-dimensional spectral matrix is ​​extracted into the current input buffer and then subjected to logarithmic domain transformation.

[0090] Then on Perform complementary ensemble empirical mode decomposition. In this example, the noise intensity can be set to 0.2, and the ensemble averaging round can be set to 50. During each ensemble averaging round, the program generates a white noise sequence and constructs positive and negative perturbation input sequences for empirical mode decomposition.

[0091] The modal results obtained from each temporary decomposition are not saved separately, but are directly accumulated into the modal accumulation buffer of the current sequence. After accumulation, the temporary modal results can be overwritten by the next decomposition. After all sets are decomposed, the accumulated results are averaged to obtain the eigenmode functions corresponding to the current sequence.

[0092] The process is as follows Figure 4 As shown: After the program starts, it first allocates a fixed-size fixed work area once, and initializes the input sequence buffer within it. Current mode buffer Residual buffer Modal accumulation buffer Envelope and extreme point working area, statistics and sorting buffer, and reconstruction buffer. Subsequently, the system acquires the two-dimensional spectral matrix to be processed. And extract the first line line by line. Write a one-dimensional sequence .in, Figure 4 The envelope and extreme point working area includes an upper envelope buffer, a lower envelope buffer, an envelope mean buffer, and an extreme point location buffer; the statistics and sorting buffer includes a modal statistics contribution buffer and a sorting index buffer.

[0093] During the decomposition calculation phase, the current sequence starts from... Load, calculate, and then write. Simultaneously update During this process, The data will be fed into the envelope and extreme point working area for extreme point extraction and envelope fitting, and the fitted results will be sent back to... Accumulate. Meanwhile, The data is also simultaneously sent to the statistics and sorting buffer for subsequent intrinsic mode function (IMF) filtering. Once all mode decompositions of the current sequence are complete, the filtering results from the statistics and sorting buffer are used to select the IMF from the data. Extract valid modes from the reconstructed buffer. The sequence is then cumulatively reconstructed to generate a reconstructed sequence of the current sequence. Finally, the reconstruction result is obtained from... Write back to the output matrix The OK.

[0094] After the current sequence is processed, the system will clear or overwrite all intermediate data in the fixed working area to seamlessly and continuously process the next feature sequence, thereby achieving pipelined processing with extremely low memory usage.

[0095] To improve the processing latency stability on the embedded platform, this embodiment uses a fixed number of sieving operations for each mode. Preferably, the number of sieving operations can be set to 5. The node array, slope array, right-hand term array, and second derivative array required for upper and lower envelope interpolation are all obtained from the same continuous working area and reused between different sequences. When there are insufficient extreme points or spline interpolation fails, linear interpolation is used as a degradation method to ensure that the processing flow is not interrupted.

[0096] Specifically, to improve the processing latency stability on the embedded platform, this application adopts the following degradation processing method in the envelope construction stage: when the number of local maxima or local minima of the current sequence is less than 3, it is determined that there are insufficient extrema; at this time, the cubic spline interpolation step is skipped, and the upper and lower envelopes are constructed directly by linearly connecting adjacent extrema. Specifically, for the upper envelope, linear interpolation is used between two adjacent maxima; for the lower envelope, linear interpolation is used between two adjacent minima; at the end of the sequence, the value of the most recent extrema is extended outward as the end envelope value; if numerical anomalies occur during the cubic spline interpolation process, such as matrix singularity or coefficient overflow, the above linear interpolation method is also switched to perform degradation processing; the number of sieving iterations for each mode is set to a fixed number or a preset upper limit, and the adaptive stopping criterion is no longer used to eliminate latency fluctuations caused by differences in noise levels of different sequences.

[0097] After completing the mode decomposition of the current sequence, the program calculates the statistical contribution of each intrinsic mode function and sorts them from largest to smallest contribution. Then, it selects several intrinsic mode functions whose cumulative contribution rate reaches a preset threshold to participate in the reconstruction. For example, when the contribution rate threshold is 0.5, it selects the set of modes with a statistical contribution rate of 50%. .

[0098] The reconstruction process is as follows: ; After reconstruction, immediately write the result back to the output two-dimensional matrix: .

[0099] After the write-back is completed, the input buffer, modality accumulation buffer, statistical filtering buffer, and reconstruction buffer corresponding to the current sequence are all cleared or overwritten and reused for the next sequence.

[0100] Therefore, the entire processing only saves the mode decomposition results of the current sequence, and not all the intrinsic mode functions corresponding to all angles. The process of real-time accumulation of temporary modes and real-time writing back of reconstruction results is as follows: Figure 5 As shown.

[0101] To verify the modal decomposition and reconstruction effect of this application on an embedded platform, this embodiment further compares the low-memory acceleration implementation results on the embedded platform with the reference implementation results on the computer. The results are as follows: Figure 6 As shown. Among them, Figure 6 (a) in the figure represents the broadband spatial spectrum reconstruction result of the underwater acoustic signal obtained by the computer using the same input data and the same mode decomposition parameters. Figure 6 (b) in the figure represents the broadband spatial spectrum reconstruction result of the underwater acoustic signal obtained by the embedded platform using the method described in this application. Figure 6 It can be seen that the reconstructed broadband spatial spectrum obtained by the embedded platform is consistent with the reference result on the computer in terms of the main spectral peak positions, frequency distribution range, target energy accumulation region, and background structure. This indicates that this application can still maintain the validity of the underwater acoustic signal mode decomposition and reconstruction results without saving all sequences, all set times, and all intermediate results of intrinsic mode functions. In particular, within the agreed angular and frequency dimensions, the embedded platform results can reproduce the main high-energy stripes and spatial spectral peak structures in the computer results, indicating that the low-memory acceleration implementation method does not change the basic output characteristics of mode decomposition and reconstruction.

[0102] This embodiment demonstrates that this application can complete mode decomposition, mode filtering, and reconstruction output of a two-dimensional spectral matrix on an embedded platform. Compared with the traditional full-data storage method, this application significantly reduces peak memory usage, decreases the number of dynamic memory allocations and data migrations, improves program stability and controllable processing latency, and is suitable for broadband spatial spectrum processing and mode decomposition and reconstruction applications under conditions of limited memory and computing power.

[0103] The embodiments of this application propose an accelerated implementation method for underwater acoustic signal mode decomposition for embedded platforms. First, a fixed working area for mode decomposition is pre-allocated. Feature sequences are extracted sequentially. Then, the temporary modal components obtained from each decomposition are accumulated into the mode accumulation buffer in real time. Next, after the decomposition is completed, the accumulated results are averaged to obtain the intrinsic mode functions. The selected modal components are reconstructed and written back to the output matrix in real time. The working area is reused to process the next sequence.

[0104] Compared with the prior art, this application has the following technical advantages: First, compared with traditional methods that require storing intermediate results of all sequences, all modal orders, and all set times, this application processes sequences one by one and accumulates temporary modalities in real time. It only needs to store the modal accumulation result of the currently processed sequence in a fixed working area, without storing any intermediate modal data of other sequences. Thus, it only needs to maintain a fixed working area, which greatly reduces the scale of modality-related storage.

[0105] Second, compared with traditional methods that require repeated dynamic allocation and release of temporary memory such as extreme point arrays, envelope arrays, and spline interpolation working areas when processing each sequence, the working area of ​​this application is allocated once during the initialization phase. In subsequent sequence processing, no dynamic memory allocation and release are performed, which avoids the generation of memory fragmentation, thereby reducing dynamic memory allocation, eliminating memory fragmentation, and improving operational stability.

[0106] Third, by reusing a fixed work area, this application eliminates latency fluctuations caused by frequent dynamic memory allocation and deallocation, making the processing time for each sequence tend to be constant, improving the stability and predictability of processing latency, and facilitating the deployment of real-time or near-real-time systems.

[0107] Fourth, this application avoids the operation of temporarily storing intermediate results in an additional large-capacity buffer and then writing them in batches by writing back the reconstruction results in real time, thereby reducing data movement and cache access overhead and improving the overall processing speed on the embedded platform.

[0108] Based on this, this application can solve at least one of the technical problems of existing empirical mode decomposition and its set extension methods on embedded underwater acoustic processing platforms, namely high peak memory usage, frequent dynamic memory requests, and unstable processing latency, which make it difficult to meet real-time or near-real-time processing requirements. In this way, it can significantly reduce peak memory usage, reduce the number of dynamic memory requests, and improve the operating efficiency, processing latency stability, and engineering deployment adaptability of the embedded underwater acoustic processing platform.

[0109] The steps described above are for clarity only. In implementation, they can be combined into one step, or some steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0110] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a method for accelerating the implementation of underwater acoustic signal mode decomposition for an embedded platform as described in the above method embodiments.

[0111] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0112] Those skilled in the art will understand that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms, characterized in that, The method includes: The feature matrix of the underwater acoustic signal to be processed is obtained, and during the program initialization phase, a fixed working area for mode decomposition is determined based on preset parameters. The size of the fixed working area remains constant. The preset parameters include sequence length, maximum mode order, ensemble average degree, and the size of temporary variables required for envelope interpolation. Extract the current feature sequence sequentially from the underwater acoustic signal feature matrix and write it into the input sequence buffer of the fixed working area; Within a fixed working area, mode decomposition is performed on the current feature sequence, and the temporary mode components obtained from each decomposition are immediately accumulated into the mode accumulation buffer within the fixed working area. After completing the mode decomposition of the current feature sequence, the accumulated results in the mode accumulation buffer are averaged to obtain the intrinsic mode functions of the current sequence. The intrinsic mode functions are filtered, and the filtered mode components are reconstructed. The reconstruction results are written back to the output matrix, and the fixed working area is reused to process the next feature sequence.

2. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 1, characterized in that, underwater acoustic signal feature matrix to be processed The expression is as follows: ； in, This represents the number of one-dimensional sequences to be decomposed. Indicates the length of each sequence; Indicates the sequence number. Indicates the sampling point number within the sequence; The underwater acoustic signal feature matrix is ​​stored in one-dimensional contiguous memory, and the underwater acoustic signal feature matrix, output matrix, and sequence buffer within the fixed working area adopt a consistent column-major contiguous storage layout. Matrix elements... linear address It is expressed as follows: 。 3. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 1, characterized in that, The fixed working area used for mode decomposition includes: input sequence buffer, residual buffer, current mode buffer, mode accumulation buffer, upper envelope buffer, lower envelope buffer, envelope mean buffer, extreme point location buffer, mode statistical contribution buffer, sorting index buffer, and reconstructed sequence buffer.

4. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 3, characterized in that, The extreme point location buffer includes the local maximum point location buffer and the local minimum point location buffer; Fixed work area The expression is as follows: ； in, This is the buffer for the current input sequence. As a residual buffer, For the current modal buffer, For modal accumulation buffer, For the upper envelope buffer, For the lower envelope buffer, As a buffer for the envelope mean, This serves as a buffer zone for the location of local maxima. This serves as a buffer zone for the location of local minima. As a buffer for modal statistical contribution, For sorting index buffer, To reconstruct the sequence buffer.

5. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 1, characterized in that, The step of performing mode decomposition on the current feature sequence within a fixed working area, and immediately accumulating the temporary mode components obtained from each decomposition into the mode accumulation buffer within the fixed working area, includes: If the current feature sequence is the th The expression for a one-dimensional underwater acoustic feature sequence is as follows: ； in, The length of each sequence; Represents the characteristic matrix of underwater acoustic signals Matrix elements in; Based on ensemble empirical mode decomposition or complementary ensemble empirical mode decomposition, mode decomposition is performed on the current feature sequence, and after each mode decomposition, for the ... For each of the first-order modal components, perform the following accumulation operation: ； in, Represents the first decomposition obtained in a single step. Temporal modal components, Indicates the first A first-order modal accumulation buffer.

6. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 5, characterized in that, When sequence length At that time, the maximum modal order Set to: ； in, This represents the floor operation; The number of screening iterations for each mode is set to a fixed number.

7. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 1, characterized in that, The filtering of intrinsic mode functions includes: Calculate the statistical contribution of each order of intrinsic mode function; Based on the statistical contribution of each order of intrinsic mode functions, the intrinsic mode functions are sorted, and the set of modes whose cumulative contribution reaches a preset threshold is selected.

8. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 7, characterized in that, The step of reconstructing the selected modal components and writing the reconstruction result back to the output matrix includes: The modal components in the modality set are directly accumulated into the reconstruction buffer, as expressed below: ； in, Represents a set of modes. ; Represents the reconstructed sequence. Indicates the first The intrinsic mode function of order.

9. The method for accelerating the implementation of underwater acoustic signal mode decomposition for embedded platforms according to claim 1, characterized in that, Reuse the fixed working area to process the next feature sequence, including: After the current sequence has been processed, each buffer in the fixed working area is cleared or overwritten to prepare for the processing of the next feature sequence.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement the underwater acoustic signal mode decomposition acceleration method for embedded platforms as described in any one of claims 1 to 9.

Images