Efficient sound propagation model adaptive selection method

Abstract

The invention provides an efficient sound propagation model adaptive selection method, which comprises the steps of decomposing data according to a sonar working state, sonar working parameters and environmental elements influencing sound propagation, calculating the distribution weight of each model based on a data-model dynamic adaptation mechanism provided by the invention, selecting an adaptive sound propagation model according to the weight, and selecting the sound propagation model according to the weight. Each sound propagation model can specially process different parts of different tasks or data so as to improve the overall calculation performance of the model. According to the calculation process, the parallel mode of the model is determined in combination with different calculation environments and the calculation power requirement characteristics of the selected model, codes are reconstructed and optimized according to the selected parallel mode, and the reconstructed model can be deployed to a single machine or a cluster for parallel processing so as to improve the calculation efficiency of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of sound propagation models, and mainly relates to an efficient method for adaptively selecting a sound propagation model. Background Art

[0002] The acoustic model adaptation technology is to adjust the acoustic model according to a small amount of test data, increase its matching degree with the test data, and thus improve the recognition performance of the system. In recent years, with the in-depth research on underwater sound field calculation, more and more requirements have been put forward for the calculation speed and accuracy of sound field theoretical calculation. Theoretically speaking, it has been very difficult to simply improve the calculation speed by modifying the model itself, and new ways must be found to solve this problem. Therefore, in order to improve the calculation efficiency, people have begun to try parallel computing based on hardware platforms.

[0003] Tan Yingwei, Chen Xiaoliang et al. provided an adaptive training method and system for an acoustic model, combined the advantages of LHT and KL divergence, and used adaptive data to retrain the neural network model to achieve the purpose of slowing down the mismatch between training data and scenario data, and ensured the improvement of the model's accuracy during the adaptation process.

[0004] Step S1, perform speech feature extraction, and use it as input to train and generate a seed model to obtain an objective function

[0005] First, extract Mel-scale filter bank features and Mel spectrum features from the original speech and adaptive speech data, and use the Mel spectrum as the feature of the original speech; among them, the adaptive data is recorded in a far-field scenario. Then, use the features of the original speech as input, and use negative cross-entropy as the objective function to train the parameters of the seed model

[0006] Step S2, adjust the network structure of the seed model and add a linear layer

[0007] After obtaining the seed model, the network structure of the seed model is adjusted, and a linear layer is directly added behind the first hidden layer. The linear hidden transformation can not only map the mismatched input vectors, but also utilize the discriminability of the hidden layer output.

[0008] Step S3, on the basis of the objective function, add a KL divergence regularization term

[0009] Use negative cross-entropy as the objective function, add a KL divergence regularization term, and delete the terms irrelevant to the model parameters.

[0010] Step S4, train the linear layer, and re-use the backpropagation algorithm to estimate the weights and biases of the hidden linear layer

[0011] After the network structure and objective function are determined, neural network learning is performed through the backpropagation algorithm. The backpropagation algorithm is based on the gradient descent method, and its learning process includes a forward propagation process and a backpropagation process. Among them, in the forward propagation stage, the Mel spectrum features of the voice signal are used as the training input to the neural network to obtain the excitation response; in the backpropagation stage, the error is obtained by calculating the difference between the excitation response and the target output corresponding to the training input, so as to obtain the response error. Then, the input excitation is multiplied by the response error to obtain the gradient of the weight. The gradient is multiplied by the training factor and negated and then added to the weight to complete the update of the weight. Then, the weights and offsets of the hidden linear layer are estimated according to the training results of the linear layer.

[0012] Step S5, training is completed, and the adaptive model is output

[0013] Due to the complex and changeable marine environment, different sonar functions, and different working frequency bands, a single sound propagation model is difficult to apply to the sound field calculation under different conditions. In order to adapt to different combat scenarios, marine environments, and sonar equipment, it is reasonable to combine multiple types of models for use. Currently, the manual selection method or the multi-condition decision method based on rules is usually adopted. For the former, it requires a high level of professional knowledge background of the propagation model user, and it is necessary to manually compare and verify the results generated by multiple models, with very low efficiency; for the latter, a fixed-threshold determination method is adopted, and the division of conditions is relatively rough, and it is impossible to achieve dynamic adjustment of specific application scenarios. On the other hand, currently commonly used sound field calculation models such as ray models, parabolic equations, and beam integration methods have high computational complexity and long calculation times, and it is difficult to meet the refined calculation requirements. Summary of the Invention

[0014] The present invention aims to solve the problem of dynamic adaptation of sound propagation models under different marine environmental conditions and different sonar working modes. For specific application scenarios, a data decomposition method for different modes is designed, and a cost function is constructed for each sub-dataset with the goal of optimizing the calculation speed and calculation accuracy to achieve the optimal adaptive selection of the sound propagation model. At the same time, aiming at the problems of long calculation time and low calculation efficiency in multi-model selection, the present invention proposes an efficient method for adaptive selection of sound propagation models. This parallel calculation method for sound propagation models in heterogeneous environments distributes the models to different devices for parallel calculation, effectively improving the calculation efficiency of the models.

[0015] The object of the present invention is achieved by the following technical solutions. An efficient method for adaptive selection of sound propagation models includes the following steps:

[0016] Step 1: Import parameters and perform unified standardization processing on the imported parameters;

[0017] Step 2: Send the data after standardization processing into the acoustic propagation model adaptive selection network. The acoustic propagation model adaptive selection network incorporates multiple acoustic propagation models, and the router determines which acoustic propagation model or models should take over the processing of the input data.

[0018] Step 3: Construct an acoustic propagation model and set a loss function to train the acoustic propagation model. The gradient of the loss function is used to adjust the parameters of the router and the acoustic propagation model to minimize the error between the predicted value and the actual label, enabling the model to adaptively select the output result of the acoustic propagation model under different environmental parameter conditions.

[0019] Step 4: According to the distributed parallel computing strategy for the input data, split the task data into multiple sub-datasets and assign each sub-dataset data to one or more acoustic propagation models.

[0020] Furthermore, the imported parameters include ocean environmental parameters and sonar parameters.

[0021] Furthermore, for the standardization processing, the specific steps are as follows:

[0022] (1) Data format conversion: The imported ocean environmental parameters and sonar parameters are uniformly converted to the NetCDF standard format. The converted data contains basic attributes such as task name, organizational unit name, sea area range, data type, and sonar type. Each type of data element needs to be uniformly standardized.

[0023] (2) Data deduplication: From standard data files from different sources by sea area, extract, load, and deduplicate all records of the same type of data and their basic attribute information respectively, and eliminate duplicate data that is repeatedly included in data from different sources.

[0024] (3) Data quality control: The methods for quality control of the imported environmental parameters include consistency checking and outlier checking.

[0025] (4) Standard dataset production: Screen and extract metadata from the integrated oceanographic, meteorological, and seabed data, and produce a standard dataset according to the diverse data application service requirements.

[0026] Furthermore, the consistency checking refers to a data quality inspection method used to verify whether the data is within its reasonable value range and whether the logical relationships between the data are reasonable; the outlier checking is used to identify abnormal data points in the dataset that are significantly different from most of the data.

[0027] Further, the router refers to a sparse gate network that receives input data and performs a series of learned non - linear transformations; the weights generated during the process represent the contribution degree of each sound propagation model to the current input; the weights are processed to form a probability distribution, and the probability distribution represents the probability that each sound propagation model is activated under given input parameters.

[0028] Further, the input data enters each sound propagation model after being selected by the sparse gate network. Each sound propagation model calculates the sound field according to its design and parameters, combined with the matching environmental parameters; the output of each sound propagation model is multiplied by its corresponding weight, and the weighted outputs are summed to form the final model output.

[0029] Further, the calculation method of the loss function is as follows:

[0030] Loss=w1E m +w2σ+w3△t (5)

[0031] Where, E m represents the root - mean - square error, σ represents the fitness of different groups of environments, △t represents the calculation time, and w1, w2, w3 are all coefficients specified in advance;

[0032] 1) The root - mean - square error E m

[0033] The root - mean - square error E m represents the error result between the model calculation and the measured value under the M - th group of data. F frequency points are calculated in each group of data, and the sound propagation model calculates N propagation loss values at different depths and distances for each frequency point Calculate the root - mean - square error E with the measured value TL in the data set i,m,f (i = 1,..., N, m = 1,..., M; f = 1,...F) m

[0034]

[0035] 2) The environmental fitness σ

[0036] Statistically calculate the root - mean - square error E of the propagation loss calculation of the sound propagation model under M groups of data m and assign a penalty factor δ for environmental fitness according to the magnitude of the root - mean - square error calculated for each group of data m as shown in the following formula:

[0037]

[0038] The fitness of the sound propagation model in M groups of environments is expressed as

[0039] 3) The calculation time tm

[0040] For N built-in models, the calculation duration of the sound propagation model is represented by △t n (1,..., N)

[0041]

[0042] where (△t n ) max represents the maximum duration of the calculations of the N models for the same set of data, and (△t n ) min represents the minimum duration of the calculations of the N models for the same set of data.

[0043] Furthermore, the distributed parallel computing strategy, in the generation process, includes splitting, communication, aggregation, and mapping;

[0044] Step 1: Splitting

[0045] Decompose the calculation into as many independent segments as possible. This step includes data splitting and model splitting;

[0046] (1) Data splitting

[0047] The data splitting is decomposed according to the specified depth, the specified sonar operating frequency, and the side scan at the specified grazing angle; each side scan records the hydrological data, ocean meteorological data, seabed sediment data, and seabed topography-related ocean environmental data related to sound propagation; traverse the sonar operating depth, sonar operating frequency band, sonar operating grazing angle, and the data measured in each azimuth to obtain multiple sub-task data sets;

[0048] (2) Model splitting

[0049] The model splitting is divided into two modes according to the hardware form combined with the model calculation characteristics, namely single program multiple data for a single machine and multiple programs multiple data for a cluster;

[0050] Step 2: Communication

[0051] Determine the amount of data transferred between sub-tasks, create a task dependency graph, represent tasks with nodes, and represent the communication volume with edges;

[0052] Step 3: Aggregation

[0053] Aggregate the tasks, and each aggregated group will be assigned to the same computing node;

[0054] Step 4: Mapping

[0055] For the application to be executed, flexibly assign / map the task groups generated in the aggregation step to available nodes according to the idle state of the hardware to form a mapping scheme;

[0056] Match the data of subtasks in different stages and the calculation processes of each model, and generate a final parallel computing strategy to reconstruct the sound propagation model.

[0057] The beneficial effects of the present invention are as follows: The distributed parallel computing strategy of data-model proposed by the present invention decomposes the data according to the sonar working state, sonar working parameters and environmental factors affecting sound propagation. Based on the data-model dynamic adaptation mechanism proposed by the present invention, the allocation weights of each model are calculated, and the appropriate sound propagation model is selected according to the weight size. Each sound propagation model can specifically process different tasks or different parts of the data to improve the overall computing performance of the model. This calculation process combines different computing environments and the computing power requirements of the selected model to determine the parallel mode of the model, and reconstructs and optimizes the code according to the selected parallel mode. The reconstructed model can be deployed to a single machine or a cluster for parallel processing to improve the computing efficiency of the model.

[0058] 1. Task specificity: The method of using a hybrid sound propagation model can effectively make full use of the advantages of multiple sound propagation models. Each sound propagation model can specifically process different tasks or different parts of the data, and has better universality under complex marine environmental conditions. Each sound propagation model can model different data distributions and patterns, thus significantly improving the accuracy and generalization ability of the model, so the model can better adapt to the complexity of the task.

[0059] 2. Flexibility: It can flexibly select and combine appropriate sound propagation models according to the computing requirements. The structure of the model allows dynamically selecting the activated sound propagation model according to the needs of the task to achieve flexible processing of the input data. This enables the model to adapt to different input distributions and task scenarios, improving the flexibility of the model.

[0060] 3. Efficiency: It adopts a distributed parallel strategy of data-model, splits the input data and the model according to the task requirements and the characteristics of the model to adapt to the parallel processing of a single machine or a cluster, and improves the computing efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those skilled in the art or ordinary technicians, without creative efforts, other drawings can be obtained according to these drawings.

[0062] Figure 1 It is the data-model dynamic adaptation workflow diagram of the present invention.

[0063] Figure 2Schematic diagram of the acoustic propagation adaptive network of the present invention. Detailed implementation manners

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] The present invention proposes an efficient method for adaptively selecting an acoustic propagation model, constructs a cost function with the goal of optimizing the model calculation speed and calculation accuracy, optimizes the network parameters through learning the correlation relationship between the input data and the acoustic propagation model, and improves the selection ability for different propagation models. Combining the distributed parallel computing strategy of data-model, determines the parallel mode of the module, and improves the data-model parallel computing ability under different requirements and different resources.

[0066] 1. Data-model dynamic adaptation mechanism

[0067] Uniform standardization processing is performed on the imported parameter data including environmental parameters, sonar parameters, etc. The data after standard processing is sent to the adaptive model selection network, which incorporates multiple acoustic propagation models. Its basic concept is to divide the task data into multiple sub-datasets according to the distributed parallel computing strategy for the input data, and assign one or more acoustic propagation models to each sub-dataset of data. Each acoustic propagation model can focus on processing this part of the input data, thereby improving the overall performance of the model.

[0068] The overall workflow diagram of the data-model dynamic adaptation mechanism is as Figure 1 shown:

[0069] The data-model dynamic adaptation method includes the following key steps.

[0070] Step 1: Parameter import

[0071] The imported parameters include ocean environmental parameters and sonar parameters. Among them, the ocean environmental parameters include sound speed profile (salinity, temperature, depth), ocean meteorological data (wind speed, rainfall, sea state), seabed sediment (coarse sand, fine sand, very fine sand, silty sand, sandy silt, silt, sand-silt-clay, clayey silt, silty clay), seabed topography (continental shelf, continental slope, seamount, trench, abyssal plain), mesoscale phenomena (mesoscale eddy, ocean front, ocean interior) and other data; sonar parameters include the working depth of the sonar, the working frequency of the sonar, the grazing angle of the sonar, etc. The imported parameters are divided into multiple sub-datasets according to the distributed parallel computing strategy of data-model, such as the example data 1 and data 2 in the above figure.

[0072] Step 2: Data standardization

[0073] The imported ocean environment parameters and sonar parameters are standardized, and the process includes data format conversion, data deduplication, data quality control, standard data set production, etc.

[0074] (1) Data format conversion

[0075] The imported ocean environment parameters and sonar parameters are uniformly converted to the NetCDF standard format. The converted data contains basic attributes such as task name, organization unit name, sea area range, data type, sonar type, etc. Each type of data element must be standardized. The metadata information of the standard data after format conversion should be extracted, supplemented, complete, and accurate.

[0076] (2) Data deduplication

[0077] According to the sea area, the same type of data records and their basic attribute information are fully extracted, loaded and deduplicated from standard data files from different sources, and duplicate data that are included multiple times in data from different sources are eliminated. Priority is given to retaining the original observation data with higher data accuracy and without interpolation processing.

[0078] (3) Data quality control

[0079] The quality control methods for imported environmental parameters include consistency test and outlier test.

[0080] 1)Consistency check

[0081] Consistency testing is a data quality testing method that is mainly used to verify whether the data is within its reasonable value range and whether the logical relationship between the data is reasonable. This test can help find data that is beyond the normal range, logically unreasonable or contradictory.

[0082] 2) Outlier test

[0083] The outlier test is used to identify abnormal data points in a data set that are significantly different from the majority of data. Commonly used methods include the Nair test method. The general calculation process is as follows:

[0084] a. Calculate the sample mean and the sample standard deviation s

[0085]

[0086] b. Calculate the outlier test statistic G

[0087] For the suspected outlier point x out ,calculate

[0088] c. Determine the critical value

[0089] According to the sample size n and the significance level α (usually 0.05 or 0.01), find the corresponding G from the critical value table of the Neyman test critical

[0090] d. Compare the test statistic with the critical value

[0091] If the calculated G is greater than G critical , then reject the null hypothesis and consider that x out is an outlier

[0092] (4) Standard data set production

[0093] Screen the integrated oceanographic, meteorological, and seabed data and extract metadata. According to the diverse data application service requirements, produce a standard data set

[0094] Step 3: Adaptive model selection

[0095] The acoustic propagation model adaptive selection network incorporates multiple acoustic propagation models, and each acoustic propagation model has its applicable calculation conditions. The core of this network, which determines which acoustic propagation model participates in calculating specific environmental parameters, is what the present invention calls the "router". The router consists of learned parameters and can be pre-trained. The router determines the weights assigned to each acoustic propagation model. During the training process, these expert and gating networks are trained simultaneously to optimize their performance and decision-making capabilities. Each acoustic propagation model processes a different subset of the training samples, focusing on a specific region of the input acoustic measurement space

[0096] (1) Router

[0097] The role of the router is to determine which acoustic propagation model should take over the processing of the input sample. The router is a sparse gating network that receives the input data and performs a series of learned non-linear transformations. This process generates a set of weights that represent the contribution degree of each acoustic propagation model to the current input. Usually, these weights are processed through functions such as softmax to ensure that they sum to 1, forming a probability distribution. Such a distribution represents the probability of each acoustic propagation model being activated under the given input parameters. Generally, the top K are selected as the models for the final calculation. For example, if there are three experts in the model, the output probabilities may be 0.5, 0.4, and 0.1, which means that the first acoustic propagation model contributes 50% to the processing of this data, the second acoustic propagation model contributes 40%, and the third acoustic propagation model contributes 10%. At this time, K can be selected as 2, and the output result y can be defined as follows

[0098] where x represents the input parameter, G(x)i represents the weight assigned to the i-th sound propagation model, E i (x) represents the i-th sound propagation model.

[0099] G(x) i = softmax(H(x) i ) (3)

[0100] H(x) i = (x * W g ) i + softplus((x * W noise ) i ) (4)

[0101] W g is a trainable parameter matrix that acts on the input parameter x, W noise represents adjustable Gaussian noise used to enhance the generalization ability of the training model. softplus(·) represents the activation function, and H(x) i keeps the top k.

[0102] (2) Sound propagation model

[0103] After being selected by the sparse gate network, the data enters each sound propagation model. Each sound propagation model calculates the sound field in combination with the matching environmental parameters according to its design and parameters. The output of each sound propagation model is multiplied by its corresponding weight, and the weighted outputs are summed to form the final model output. This weighted combination mechanism enables the model to adaptively select the output results of the sound propagation model under different environmental parameter conditions.

[0104] (3) Loss function

[0105] When constructing the loss function, the accuracy of model calculation, model applicability, and operation efficiency are weighed.

[0106] The value of this loss function is jointly determined by the root mean square error E m , the fitness σ of different groups of environments, and the calculation time △t

[0107] Loss = w1E m + w2σ + w3△t (5)

[0108] Here, w1, w2, and w3 are usually specified in advance and are specified according to the importance of different evaluation criteria for model selection. For example, in the combat planning stage, more attention is paid to the accuracy of model calculation, and the proportion of w1 can be appropriately increased; in the real-time analysis and real-time processing stage of combat, more attention is paid to the calculation efficiency of the model, and the proportion of w3 can be appropriately increased to comprehensively consider different evaluation factors to select the most suitable sound propagation model for the usage scenario.

[0109] 1) Root mean square error E m

[0110] Root mean square error E m Represents the error result between the model calculation and the measured value under the M - th group of data (environment). For each group of data, F frequency points are calculated, and the sound propagation model calculates a total of N propagation loss values at different depths and distances for each frequency point Calculate the root mean square error E with the measured value TL in the data set i,m,f (i = 1,..., N, m = 1,..., M; f = 1,...F) m

[0111]

[0112] 2) Environmental fitness σ

[0113] Statistically calculate the root mean square error E of the propagation loss calculation of the sound propagation model under M groups of data (environment) m , and assign a penalty factor δ for environmental fitness according to the magnitude of the root mean square error calculated for each group of data m , as shown in the following formula

[0114]

[0115] The fitness of the sound propagation model in M groups of environments can be expressed as

[0116] 3) Calculation time t m

[0117] For N built - in models, the calculation duration of the sound propagation model is denoted as △t n (1,..., N)

[0118]

[0119] Among them, (△t n ) max represents the maximum calculation duration of N models for the same group of data, and (△t n ) min represents the minimum calculation duration of N models for the same group of data

[0120] (4) Model training

[0121] The training of the model is carried out through the backpropagation algorithm at this stage. The gradient of the loss function is used to adjust the parameters of the sparse gate network and the sound propagation model to minimize the error between the predicted value and the actual label. This process is a key step in training the model weights, ensuring that the model can better adapt to the training data

[0122] In summary, the present invention proposes a "router" - a sparse gate network, which is used to determine which sound propagation model should take over the processing of the input sample. It receives input data and performs a series of learned non-linear transformations. After being selected by the sparse gate network, the data enters each sound propagation model. Each sound propagation model calculates the sound field according to its design and parameters, combined with the matching environmental parameters. The output of each sound propagation model is multiplied by its corresponding weight, and the weighted outputs are summed to form the final model output. This weighted combination mechanism enables the model to adaptively select the output results of the sound propagation models under different environmental parameter conditions.

[0123] 2. Distributed parallel computing strategy for data-model

[0124] The generation process of the distributed parallel computing strategy for data-model mainly includes segmentation, communication, aggregation, and mapping. The specific steps are as follows:

[0125] Step 1: Segmentation

[0126] Decompose the calculation into as many independent segments as possible. This step will reveal the parallelism in the algorithm. The granularity of the segments depends on different applications. Usually, the number of decomposed segments is one to two orders of magnitude more than the number of available computing nodes (CPU cores, CPUs, distributed computing nodes). This step mainly includes data segmentation and model segmentation.

[0127] (1) Data segmentation

[0128] Data segmentation can be decomposed according to the specified depth, specified sonar operating frequency, and side lines at the specified grazing angle.

[0129] Each side line records marine environmental data such as hydrological data, ocean meteorological data, seabed sediment data, and seabed topography related to sound propagation. Traversing the data measured at the sonar operating depth, sonar operating frequency band, sonar operating grazing angle, and each azimuth can obtain multiple sub-task data sets.

[0130] (2) Model segmentation

[0131] Model segmentation is mainly designed based on the hardware form combined with the model calculation characteristics, and can be mainly divided into two modes: single program multiple data for single machine and multiple programs multiple data for cluster.

[0132] 1) Single program multiple data mode

[0133] In the single program multiple data mode, all nodes (CPU cores, CPUs, or distributed computing nodes) of the execution platform run the same program, but they either apply the same operation to different data or execute different execution paths in the program. Its typical program structure is as follows:

[0134] Program initialization: This step usually deploys the program to a parallel platform and initializes the runtime system responsible for multi-thread or process communication and synchronization.

[0135] Obtain unique identifiers: The identifiers start counting from 0, enumerating the threads or processes in use. In some cases, the identifier can be a vector rather than a scalar. The lifecycle of the identifier is consistent with the thread or process it identifies. The identifier can also be persistent, i.e., it exists throughout the program, or is generated dynamically when needed.

[0136] Run the program: Execute the execution path with consistent unique IDs, which may include workload or data allocation, role diversification, etc.

[0137] Shut down the program: Shut down the threads or processes, and it may be necessary to merge some results to generate the final result.

[0138] The engineering model adopting the single program multiple data mode includes:

[0139] Sound field model, the single program multiple data mode can be adopted when performing regional multi-point calculations.

[0140] 2) Multiple programs multiple data mode

[0141] The multiple programs multiple data mode is mainly used in the following two situations:

[0142] a. The running environment is heterogeneous, and different executable files need to be deployed according to the node architecture.

[0143] b. The application program memory requirements are very strict, so it is necessary to reduce the program logic uploaded to each node to ensure basic components.

[0144] The above situations are covered by allowing different executable files that may come from different toolchains to be combined into an application program. Each computing node can run its own program logic to process its own data set, but may still need to follow the step sequence in 1). Programs based on CUDA belong to this mode, where a part of the program is binary code for the CPU host to execute, and the other part is binary code for the GPU coprocessor to execute.

[0145] The engineering model adopting the multiple programs multiple data mode includes:

[0146] Sound field model, the multiple programs multiple data mode can be adopted when performing single-point multi-directional calculations, with data parsing and organization executed on the CPU and model calculations executed on the GPU.

[0147] Step 2: Communication

[0148] Ideally, the segments decomposed from the previous step are completely independent. However, usually there are dependencies between subtasks, that is, the start of one subtask waits for the end of another subtask, and so on. Such dependencies may include data transmission. In this step, the amount of data to be transmitted between subtasks is determined. Combining the previous steps can create a task dependency graph, with nodes representing tasks and edges representing communication volume.

[0149] Step 3: Aggregation

[0150] Communication hinders parallel computing. One way to avoid communication is to aggregate tasks. Each aggregated group will be assigned to the same computing node to avoid intra-group communication. As a general rule of thumb, the number of groups should be one order of magnitude more than the number of available computing nodes at this time.

[0151] Step 4: Mapping

[0152] For the application to be executed, the task groups generated in the aggregation step are flexibly assigned / mapped to available nodes according to the idle state of the hardware to form a mapping scheme.

[0153] Match the data of subtasks in different stages and the calculation processes of each model to generate a final parallel computing strategy to reconstruct the sound propagation model.

[0154] In summary, decompose the data according to the sonar working state, sonar working parameters and environmental factors affecting sound propagation. Based on the data-model dynamic adaptation mechanism proposed in the present invention, calculate the allocation weights of each model, select the appropriate sound propagation model according to the weight size. Each sound propagation model can specifically process different tasks or different parts of the data to improve the overall computing performance of the model. This calculation process combines different computing environments and the computing power requirements of the selected model to determine the parallel mode of the model, and reconstruct and optimize the code according to the selected parallel mode. The reconstructed model can be deployed to a single machine or a cluster for parallel processing to improve the computing efficiency of the model.

[0155] Glossary of Related Technical Terms

[0156] (1) The Bellhop propagation model mentioned when introducing the technical background is a computer model used to simulate underwater sound propagation. It can simulate the processes of sound wave propagation, scattering, absorption and attenuation in water under different ocean environmental conditions, and is widely used in the fields of ocean acoustics, underwater communication, ocean exploration, etc. It is based on the Gaussian beam tracking method and can calculate the sound ray trajectory and sound field in a horizontally non-uniform environment.

[0157] (2) The NetCDF standard format is mentioned in the data format conversion. It is an abbreviation of Network Common Date Form and is a file format used for storing and distributing scientific data. It is widely used in fields such as meteorology, oceanography, climate science, and geophysics. The NetCDF file format is designed to efficiently store and access large amounts of numerical data while maintaining data portability and accessibility.

[0158] (3) The softmax function mentioned in the adaptive model selection is a commonly used activation function in machine learning and deep learning, especially when dealing with multi-class classification problems. It converts a real-valued vector into a probability distribution, such that each element in the vector lies between 0 and 1, and the sum of all elements is 1. This makes the softmax function very suitable for representing the probability that an event belongs to a certain class.

[0159] As described above, only the specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claimed rights.

Claims

1. An efficient method for adaptively selecting a sound propagation model, characterized in that: The steps include: Step 1: Import parameters and perform unified standardization on the imported parameters; Step 2: Send the standardized data to the sound propagation model adaptive selection network. The sound propagation model adaptive selection network has multiple sound propagation models built in, and determines which sound propagation model or models should take over the input data for processing through a router; Step 3: Build a sound propagation model and set a loss function to train the sound propagation model. The gradient of the loss function is used to adjust the parameters of the router and the sound propagation model to minimize the error between the predicted value and the actual label, so that the model can adaptively select the output result of the sound propagation model under different environmental parameter conditions. Step 4: Input data: According to the distributed parallel computing strategy, the task data is divided into multiple sub-data sets, and each sub-data set data is assigned one or more sound propagation models.

2. The efficient adaptive selection method of sound propagation model according to claim 1, characterized in that: The imported parameters include ocean environment parameters and sonar parameters.

3. The efficient adaptive selection method of sound propagation model according to claim 2, characterized in that: The specific steps of the standardization process are as follows: (1) Data format conversion: The imported ocean environment parameters and sonar parameters are uniformly converted to the NetCDF standard format. The converted data includes the task name, organization unit name, sea area range, data type, and basic attributes related to the sonar type. Each type of data element must be standardized; (2) Data deduplication: Extract, load and dedupe similar data records and their basic attribute information from standard data files from different sources according to the sea area, and remove duplicate data that is recorded multiple times in data from different sources; (3) Data quality control: The quality control methods for imported environmental parameters include consistency check and outlier check; (4) Production of standard data sets: Screening and extracting metadata from the integrated ocean hydrological, meteorological, and seafloor data, and producing standard data sets based on the diverse data application service needs.

4. The efficient adaptive selection method of sound propagation model according to claim 3, characterized in that: The consistency check refers to a data quality check method used to verify whether the data is within its reasonable value range and whether the logical relationship between the data is reasonable; the outlier check is used to identify abnormal data points in the data set that are significantly different from the majority of the data.

5. The efficient adaptive selection method for sound propagation model according to claim 4, characterized in that: The router is a sparse gate network that receives input data and performs a series of learned nonlinear transformations; the weights generated in the process represent the contribution of each sound propagation model to the current input; the weights are processed to form a probability distribution, which represents the The probability of each sound propagation model being activated under given input parameters.

6. The efficient adaptive selection method for sound propagation model according to claim 5, characterized in that: The input data enters each sound propagation model after being selected by the sparse gate network. Each sound propagation model calculates the sound field according to its design and parameters combined with the matching environmental parameters; the output of each sound propagation model is multiplied by its corresponding weight, and these weighted outputs are summed to form the final model output.

7. The efficient adaptive selection method for sound propagation model according to claim 6, characterized in that: The calculation method of the loss function is: Loss=w1Em+w2σ+w3△t (5) Among them, Em represents the root mean square error, σ represents the fitness of different groups of environments, △t represents the calculation time, and w1, w2, and w3 are all pre-specified coefficients; 1) Root mean square error Em The root mean square error Em represents the error between the model calculation and the measured value under the Mth group of data. F frequency points are calculated in each group of data. The sound propagation model calculates N propagation loss values ​​for each frequency point at different depths and distances. Calculate the root mean square error Em with the measured values ​​TLi,m,f (i=1,...,N,m=1,...,M;f=1,...F) in the data set 2) Environmental adaptability σ The root mean square error Em of the propagation loss calculation of the statistical sound propagation model under M groups of data is assigned a penalty factor δm of environmental fitness according to the root mean square error calculated for each group of data, as shown in the following formula: The fitness of the sound propagation model in the M group environment is expressed as 3) Calculation time tm For N built-in models, the sound propagation model calculation time is represented by △tn(1,...,N) Where (△t n ) max Indicates the maximum time for N models to calculate the same set of data, (△t n ) min Indicates the minimum time required to calculate N models for the same set of data.

8. The efficient adaptive selection method for sound propagation model according to claim 7, characterized in that: The distributed parallel computing strategy, the generation process includes segmentation, communication, aggregation and mapping; Step 1: Segmentation Break the computation into as many independent pieces as possible. This step includes data segmentation and model segmentation. (1) Data segmentation Data segmentation is performed by lateral line decomposition at a specified depth, specified sonar operating frequency, and specified grazing angle; each lateral line records the hydrological data, marine meteorological data, seabed bottom data, and seabed topography-related marine environmental data that affect sound propagation; multiple subtask data sets are obtained by traversing the sonar operating depth, sonar operating frequency band, sonar operating grazing angle, and data measured in various azimuths; (2) Model segmentation Model segmentation is divided into two modes: single program and multiple data for a single machine and multiple programs and multiple data for a cluster, based on the hardware form and the model calculation characteristics. Step 2: Communication Determine the amount of data transferred between subtasks and create a task dependency graph, with nodes representing tasks and edges representing communication volume; Step 3: Gather The tasks are clustered together and each clustered group will be assigned to the same computing node; Step 4: Mapping For the application to be executed, the task groups generated by the aggregation step are flexibly allocated / mapped to available nodes according to the idle state of the hardware to form a mapping scheme; The data of subtasks at different stages and the calculation process of each model are matched to generate the final parallel computing strategy to reconstruct the sound propagation model.

Images