Dynamic Estimation Method of Squid Resources Based on Bayesian Deep Inversion Network

Abstract

The present invention discloses a method for dynamically estimating squid resources based on a Bayesian deep inversion network, comprising the following steps: S1. forming a preprocessed multi-source data set; S2. constructing a Bayesian deep inversion network model; S3. training the Bayesian deep inversion network model by using the preprocessed multi-source data set to form a set of trained model parameters; S4. obtaining the dynamic estimation result of squid resources and its uncertainty quantification index; S5. post-processing the dynamic estimation result of squid resources and the uncertainty quantification index to form a dynamic estimation report of squid resources with spatio-temporal distribution characteristics; S6. docking the dynamic estimation report of squid resources with a fishery management system for dynamic monitoring and real-time decision-making of squid resources. The present invention significantly reduces the false prediction rate in the low-confidence prediction area and improves the prediction stability in the high-confidence area, providing more reliable support data for the scheduling of fishery administration vessels and the planning of fishing paths.

Description

Technical Field

[0001] The present invention relates to the technical field of squid resources, and particularly relates to a dynamic estimation method for squid resources based on a Bayesian deep inversion network. Background Art

[0002] With the continuous increase in the pressure of global fishery resources monitoring and management, intelligent estimation technologies based on marine environmental data and the activity characteristics of target populations have gradually become an important supporting means for marine fishery management. As a high-value, widely distributed, and highly mobile economic marine organism, the dynamic assessment of squid resources is of great significance for ensuring sustainable fishing and avoiding overexploitation of resources. Currently, the estimation and prediction of squid resources mainly rely on two types of technical paths: one is the method based on acoustic detection and sampling statistics; the other is the regression prediction method based on traditional neural networks or machine learning models.

[0003] In the first type of method, acoustic detection usually combines multi-beam echosounders and underwater trawling means to scan the sea area and calculate the squid distribution density accordingly. However, this type of method is limited by the equipment deployment range, operating sea conditions, and seabed topography, making it difficult to achieve wide-area, high-frequency, and real-time resource monitoring, and it cannot effectively reflect the behavior migration and distribution change trends of squid populations under the influence of complex environmental variables. More importantly, the acoustic data itself has a large amount of noise and is severely interfered by the temperature-salinity structure, which is extremely likely to cause distortion of the resource estimation results and lacks a reliable uncertainty characterization method.

[0004] Although the second type of estimation method based on deep learning can improve the estimation accuracy of squid resources to a certain extent, there are several prominent problems in the current technology: First, the models generally adopt a static training paradigm and fail to effectively combine the temporal linkage characteristics between squid behavior and the marine environment, resulting in a lag in response to dynamic evolution patterns; second, there is a lack of an effective prior knowledge fusion mechanism, the generalization ability of the models is weak, and they are extremely likely to produce incorrect fittings for noise or marginal scenarios; third, most deep models belong to the "black box" structure and lack the ability to quantify the uncertainty of the output results, which limits the application credibility of the prediction results in actual fishery scheduling and risk management. In addition, the existing methods usually simply splice the marine environmental parameters and squid behavior parameters, ignoring the heterogeneous coupling characteristics between the two, and it is difficult to fully extract the interaction mechanism between potential behavior driving factors and key environmental variables.

[0005] Therefore, there is an urgent need for a more robust, interpretable, and practical intelligent estimation method to solve this problem. Summary of the Invention

[0006] An object of the present invention is to propose a dynamic estimation method for squid resources based on a Bayesian deep inversion network. The present invention significantly reduces the false prediction rate in the low-confidence prediction region and improves the prediction stability in the high-confidence region, providing more reliable support data for fishery administration vessel scheduling and fishing path planning.

[0007] A dynamic estimation method for squid resources based on a Bayesian deep inversion network according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect a real-time multi-source data set containing marine environmental parameters and squid biological activity parameters, and preprocess the real-time multi-source data set to form a preprocessed multi-source data set;

[0009] S2. Construct a Bayesian deep inversion network model;

[0010] S3. Use the preprocessed multi-source data set to train the Bayesian deep inversion network model. During the training process, use the dynamic prior update module to update the dynamic prior data in real time, and optimize the model parameters by maximizing the likelihood function and minimizing the prediction error to form a set of trained model parameters;

[0011] S4. Input the real-time collected multi-source data set into the trained Bayesian deep inversion network model, and use the dynamic prior update module to update the prior data of the real-time multi-source data set to obtain the dynamic estimation result of squid resources and its uncertainty quantification index;

[0012] S5. Post-process the dynamic estimation result of squid resources and the uncertainty quantification index to form a dynamic estimation report of squid resources with spatio-temporal distribution characteristics;

[0013] S6. Interface the dynamic estimation report of squid resources with the fishery management system for dynamic monitoring and real-time decision-making of squid resources.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Collect real-time multi-source data containing marine environmental parameters and squid biological activity parameters from marine observation platforms, satellite remote sensing devices, and underwater sensors to construct a real-time multi-source data set :

[0016] ;

[0017] wherein, represents the th original data record, represents the total number of real-time collected data, is the acquisition timestamp of the th data, is the ocean environmental parameter vector of the th record, representing seawater temperature , salinity , chlorophyll concentration , sea current velocity , is the squid biological activity parameter vector of the th record, representing the distribution position coordinates and the biological density index ;

[0018] S12. Perform time synchronization processing on the real-time multi-source data set , map the multi-source data from different devices and different time scales to the standard time series set uniformly, align the time of all multi-source data with a unified time step, and construct a multi-source data set after time synchronization, so that various parameters have a consistent time reference benchmark;

[0019] S13. Perform spatial interpolation processing on the multi-source data set after time synchronization. For the areas with missing measurements in the geographical space, estimate and fill in the missing spatial sites based on the adjacent values and their spatial position relationships at the observed positions, and form a spatially completed multi-source data set, so that the ocean environmental parameters and the squid biological activity parameters have complete spatial coverage in the target area;

[0020] S14. Perform missing value filling processing on the spatially completed multi-source data set. For the ocean environmental parameter or squid biological activity parameter vector that still has non-spatial dimension missing values, use the sliding time window method to fill in the data based on the historical observed values at adjacent times, and the filled data forms a multi-source data set after missing value processing;

[0021] S15. Remove outliers from the multi-source data set after missing value processing, calculate the historical mean and standard deviation of each parameter, and judge whether the current observed value of the multi-source data set after missing value processing is within the preset reasonable interval range. If the deviation degree exceeds the set threshold, mark this record as abnormal data and remove it. Finally, a preprocessed multi-source data set is formed.

[0022] Optionally, the S2 includes the following steps:

[0023] S21. Construct a Bayesian deep inversion network model , the Bayesian deep inversion network model includes an input layer, a dual-channel feature extraction structure, a fusion nesting layer, an output layer, and a dynamic prior model. The input layer receives the preprocessed multi-source data set , and is used to learn the dynamic distribution behavior characteristics of squid resources under different ocean environmental states;

[0024] S22. Construct a dual-channel feature extraction structure for the Bayesian depth inversion network model. The dual-channel feature extraction structures are the marine environment channel and the squid behavior channel , which are used to simulate the response mechanism of squid to environmental stimuli;

[0025] S23. Set up a fusion nested layer to perform feature-level connection operations on the feature vectors of the marine environment channel and the squid behavior channel to obtain a fused feature representation ;

[0026] S24. Set the squid resource estimation output output by the output layer of the Bayesian depth inversion network model:

[0027] ;

[0028] Among them, is the predicted squid spatial position coordinate, is the predicted squid density estimate, which is used to dynamically estimate the spatial distribution and quantity status of squid resources;

[0029] S25. Construct a dynamic prior model , and based on historical marine environmental parameters and squid biological activity parameters, construct a dynamic prior model with time weights:

[0030] ;

[0031] Among them, represents the squid resource dynamic prior data generated by the dynamic prior model for the current time , are the marine environmental parameters and squid biological activity parameters at the historical th moment respectively, is the time decay weight, is the statistical feature extraction function;

[0032] S27. Perform Bayesian posterior modeling on the squid resource dynamic prior data and the fused feature representation to calculate the posterior output distribution of the squid resource dynamic estimation:

[0033] ;

[0034] Among them, represents the Bayesian posterior distribution of the squid resource estimation output given the fused feature representation , represents the fused feature representation when the squid resource estimation output Likelihood function

[0035] Optionally, a residual structure is constructed in the marine environment channel to extract stable background change features:

[0036] ;

[0037] wherein is the environmental feature representation of the th layer, representing the implicit response feature of the squid to the marine environmental parameters, represents the environmental feature representation of the th layer, and are weights and biases, is the Sigmoid function representing the attention weight, and f is the activation function;

[0038] An attention mechanism is introduced in the squid behavior channel to enhance the recognition ability of the features of the squid aggregation area:

[0039] ;

[0040] wherein is the Sigmoid function representing the attention weight, is the Hadamard product, is the attention mapping parameter of the th layer, is the output feature vector of the th layer of the squid behavior channel, representing the spatial behavior features of individual or group squids, is the output feature vector of the th layer of the squid behavior channel.

[0041] Optionally, S3 includes the following steps:

[0042] S31. Use the preprocessed multi-source dataset to train the Bayesian depth inversion network model . The training samples are composed of multiple input and output pairs, where each input vector is composed of the concatenation of marine environmental parameters and squid biological activity parameters, and the output label is composed of the corresponding true distribution position coordinates and density values of the squids;

[0043] S32. During the training process, construct a Bayesian log-likelihood loss function containing dynamic prior constraints;

[0044] S33. Before the start of each training batch, call the dynamic prior model, based on the time step before the current training Update the prior data of squid resource dynamics with historical marine environmental parameters and squid biological activity parameters at a certain time point. The updated result of the prior data of squid resource dynamics is the prior data of squid resource dynamics at the current time step, which is used to participate in the calculation of the posterior probability in the training loss, so that the training process continuously reflects the temporal evolution characteristics of the marine environment and the dynamic changes of squid behavior.

[0045] S34. In each training iteration, calculate the Bayesian log-likelihood loss function using the input and corresponding output labels of the current batch of training samples, and perform gradient descent optimization on all trainable parameters in the Bayesian deep inversion network model based on the Bayesian log-likelihood loss function. The parameter update process uses the learning rate as a control factor to gradually correct the model weights according to the current gradient direction.

[0046] S35. Continuously execute the training iteration until the Bayesian log-likelihood loss function converges on the validation set or reaches the preset upper limit of the number of training epochs. After training is completed, a set of trained model parameters is formed.

[0047] Optionally, the Bayesian log-likelihood loss function consists of two parts: the first part is the log-likelihood term of the posterior distribution, which is used to evaluate the matching degree of the model output results under the guidance of the dynamic prior; the second part is the regularization term, which is used to constrain the magnitude of the network weight parameters to prevent the model from overfitting in the marine multi-source data. The minimization objective of the Bayesian log-likelihood loss function is to maximize the credibility of the estimation results.

[0048] Optionally, S4 includes the following steps:

[0049] S41. Feed the input vector in the real-time multi-source dataset into the input layer of the trained Bayesian deep inversion network model, and process it sequentially through the marine environment channel, the squid behavior channel, and the fusion nested layer to generate a fused feature representation. The fused feature representation comprehensively reflects the non-linear coupling relationship between the current marine environment changes and the individual behavior characteristics of squids.

[0050] S42. Call the dynamic prior model to update the prior distribution at the current moment. The update method is based on the historical marine environmental parameters and squid biological activity parameters recorded in multiple consecutive time steps before the current time point. The historical parameters are assigned decreasing weights in chronological order, and the squid resource dynamic prior data corresponding to the current time point is calculated. The squid resource dynamic prior data is used to express the empirical distribution trend of squid resources in the current environmental state.

[0051] S43. Use the fused feature representation and the squid resource dynamic prior data together in the Bayesian posterior inference process. According to the Bayesian principle, combine the likelihood information under the current feature conditions with the squid resource dynamic prior data to estimate the posterior distribution of the squid resource, including the distribution position and density prediction value of the squid resource at the current moment.

[0052] S44. Calculate the uncertainty quantification index corresponding to the current prediction result based on the posterior distribution of the squid resource. The uncertainty index includes the variances of the predicted position coordinates in three spatial dimensions and the volatility variance of the squid density estimation result.

[0053] S45. Output the dynamic estimation result of the squid resource at the current moment, including the predicted position coordinates and density values, and simultaneously output the corresponding uncertainty quantification index.

[0054] Optionally, the S5 includes the following steps:

[0055] S51. Structurally organize the dynamic estimation result of the squid resource output at the current moment and its corresponding uncertainty quantification index :

[0056] ;

[0057] Among them, , , respectively represent the variance estimation values of the three spatial dimension components of the predicted position under the Bayesian posterior distribution, represents the variance of the squid density prediction value at the current time step;

[0058] Arrange the squid spatial position coordinates in the dynamic estimation result of the squid resource and the squid density estimation value at time step t according to the time series and geographical region dimensions to construct a squid resource estimation data table;

[0059] S52. Introduce uncertainty information based on the squid resource estimation data table, establish a confidence label for each prediction cell, and divide the credibility levels.

[0060] S53. Based on the sorted predicted squid spatial position coordinates and squid density estimation values, combine with the sea area grid division rules to conduct regional aggregation statistics on the squid resource, calculate the average squid density, density change rate, and confidence level distribution per unit time within each sub-region, and form a spatial aggregation feature at the regional level.

[0061] S54. Construct the structure of the dynamic estimation report of the squid resource. The report includes the following information fields:

[0062] Time label: Current prediction time step ;

[0063] Location label: Spatial grid number of the predicted sea area;

[0064] Prediction result: Spatial position coordinates of squid , Estimated squid density ;

[0065] Uncertainty quantification index: Index set composed of position coordinate variance and density variance ;

[0066] Confidence level: High confidence / Medium confidence / Low confidence classification label;

[0067] Spatial aggregation features: Regional average density value, regional density volatility index, local density gradient trend index

[0068] Optionally, the confidence level classification rule is:

[0069] High-confidence area: If , Marked as high confidence;

[0070] Medium-confidence area: If , Marked as medium confidence;

[0071] Low-confidence area: If , Marked as low confidence;

[0072] Among them, and are the uncertainty interval division thresholds.

[0073] The beneficial effects of the present invention are:

[0074] (1) The present invention constructs a dual-channel structure of the marine environment channel and the squid behavior channel, respectively extracts independent feature representations of physical parameters such as temperature, salinity, and flow velocity and squid distribution coordinates and density behavior parameters, and performs feature cascading in the nested fusion layer to effectively capture the non-linear response characteristics of environmental changes to squid aggregation behavior.

[0075] (2) The present invention designs a dynamic prior generation module, which performs weighted fusion on environmental data and behavior parameters within the past K time steps based on a time decay function to generate a prior estimation distribution at time t, and performs Bayesian posterior inference with the current fusion features, enabling the model to make estimates by combining historical trends at each prediction time point, reflecting the inertia and delay characteristics of the marine ecosystem, and significantly enhancing the adaptability of the model to complex background changes in the ocean.

[0076] (3) The present invention obtains the posterior distribution variance of the predicted position and density of squid through the Bayesian model structure, and defines confidence level labels based on this to label the credibility of each predicted grid cell. In addition, thresholds δ1 and δ2 are set based on the variance aggregation index to achieve regional-level prediction credibility screening and aggregation statistics support, significantly reducing the false alarm rate in low-confidence prediction regions and improving the prediction stability in high-confidence regions, providing more reliable support data for fishery administration vessel scheduling and fishing path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0078] Figure 1 is a flowchart of a method for dynamically estimating squid resources based on a Bayesian deep inversion network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0080] Refer to Figure 1 , a method for dynamically estimating squid resources based on a Bayesian deep inversion network, includes the following steps:

[0081] S1. Collect a real-time multi-source data set containing marine environmental parameters and squid biological activity parameters, and preprocess the real-time multi-source data set to form a preprocessed multi-source data set;

[0082] S2. Construct a Bayesian deep inversion network model;

[0083] S3. Use the preprocessed multi-source data set to train the Bayesian deep inversion network model. During the training process, use the dynamic prior update module to update the dynamic prior data in real time, and optimize the model parameters by maximizing the likelihood function and minimizing the prediction error to form a set of trained model parameters;

[0084] S4. Input the real-time collected multi-source data set into the trained Bayesian deep inversion network model, and use the dynamic prior update module to update the prior data of the real-time multi-source data set to obtain the dynamic estimation result of squid resources and its uncertainty quantification index;

[0085] S5. Post-process the dynamic estimation result of squid resources and the uncertainty quantification index to form a dynamic estimation report of squid resources with spatio-temporal distribution characteristics;

[0086] S6. Interface the dynamic estimation report of squid resources with the fishery management system to conduct dynamic monitoring and real-time decision-making on squid resources.

[0087] In this embodiment, S1 includes the following steps:

[0088] S11. Collect real-time multi-source data containing marine environmental parameters and squid biological activity parameters from marine observation platforms, satellite remote sensing devices, and underwater sensors, and construct a real-time multi-source data set :

[0089] ;

[0090] Among them, represents the th original data record, represents the total number of real-time collected data, is the acquisition timestamp of the th data, is the vector of marine environmental parameters of the th record, representing seawater temperature , salinity , chlorophyll concentration , sea current velocity is the vector of squid biological activity parameters of the th record, representing the distribution position coordinates and the biological density index ;

[0091] S12. Perform time synchronization processing on the real-time multi-source data set to map multi-source data from different devices and different time scales to a standard time series set, align the time of all multi-source data with a unified time step, and construct a time-synchronized multi-source data set so that various parameters have a consistent time reference benchmark;

[0092] S13. Perform spatial interpolation processing on the time-synchronized multi-source data set. For areas with missing measurements in the geographical space, interpolate and estimate the missing spatial sites based on the adjacent values and their spatial position relationships at the observed positions to form a spatially completed multi-source data set, so that the marine environmental parameters and squid biological activity parameters have complete spatial coverage within the target area;

[0093] S14. Perform missing value filling processing on the spatially completed multi-source data set. For the vectors of marine environmental parameters or squid biological activity parameters that still have non-spatial dimension missing values, use the sliding time window method to fill the data based on the historical observed values at adjacent times, and the filled data forms a multi-source data set after missing value processing;

[0094] S15. Remove outliers from the multi-source dataset after missing value processing, calculate the historical mean and standard deviation of each parameter, and determine whether the current observation value of the multi-source dataset after missing value processing is within the preset reasonable range. If the deviation exceeds the set threshold, mark the record as abnormal data and remove it, finally forming a preprocessed multi-source dataset 。

[0095] In this embodiment, S2 includes the following steps:

[0096] S21. Construct a Bayesian deep inversion network model , the Bayesian deep inversion network model includes an input layer, a dual-channel feature extraction structure, a fusion nesting layer, an output layer, and a dynamic prior model. The input layer receives the preprocessed multi-source dataset , and is used to learn the dynamic distribution behavior characteristics of squid resources under different ocean environmental states;

[0097] S22. Construct the dual-channel feature extraction structure of the Bayesian deep inversion network model. The dual-channel feature extraction structures are respectively the ocean environment channel and the squid behavior channel , and are used to simulate the response mechanism of squid to environmental stimuli;

[0098] S23. Set the fusion nesting layer, perform a feature-level connection operation on the feature vectors of the ocean environment channel and the squid behavior channel to obtain a fused feature representation ;

[0099] S24. Set the squid resource estimation output output by the output layer of the Bayesian deep inversion network model :

[0100] ;

[0101] Among them, is the predicted squid spatial position coordinate, is the predicted squid density estimate, and is used to dynamically estimate the spatial distribution and quantity status of squid resources;

[0102] S25. Construct a dynamic prior model , and construct a dynamic prior model with time weights based on historical ocean environmental parameters and squid biological activity parameters:

[0103] ;

[0104] Among them, represents the squid resource dynamic prior data generated by the dynamic prior model for the current moment , are respectively the historical Ocean environmental parameters and squid biological activity parameters at a given moment, is the time decay weight, is the statistical feature extraction function;

[0105] S27. Use the squid resource dynamic prior data and the fused feature representation to perform Bayesian posterior modeling and calculate the posterior output distribution of squid resource dynamic estimation:

[0106] ;

[0107] where, represents the Bayesian posterior distribution of the squid resource estimation output under the condition given by the fused feature representation , represents the likelihood function of the fused feature representation when the squid resource estimation output is given.

[0108] In this embodiment, a residual structure is constructed in the ocean environmental channel to extract stable background change features:

[0109] ;

[0110] where, is the environmental feature representation of the th layer, representing the implicit response feature of squid to ocean environmental parameters, represents the environmental feature representation of the th layer, and are the weight and bias, is the Sigmoid function representing the attention weight, and f is the activation function;

[0111] An attention mechanism is introduced in the squid behavior channel to enhance the recognition ability of squid aggregation area features:

[0112] ;

[0113] where, is the Sigmoid function representing the attention weight, is the Hadamard product, is the attention mapping parameter of the th layer, is the output feature vector of the th layer of the squid behavior channel, representing the spatial behavior features of squid individuals or groups, is the Output feature vector of the squid behavior channel.

[0114] In this embodiment, S3 includes the following steps:

[0115] S31. Use the preprocessed multi-source dataset to train the Bayesian deep inversion network model The training samples consist of multiple input-output pairs, where each input vector is composed of the concatenation of marine environmental parameters and squid biological activity parameters, and the output label consists of the corresponding true distribution position coordinates and density values of the squid;

[0116] S32. During the training process, construct a Bayesian log-likelihood loss function containing dynamic prior constraints;

[0117] S33. Before the start of each training batch, call the dynamic prior model to update the dynamic prior data of squid resources based on the historical marine environmental parameters and squid biological activity parameters at the previous time points before the current training time step. The updated result of the dynamic prior data of squid resources is the dynamic prior data of squid resources at the current time step, which is used to participate in the calculation of the posterior probability in the training loss, so that the training process continuously reflects the temporal evolution characteristics of the marine environment and the dynamic changes of squid behavior;

[0118] S34. In each training iteration, calculate the Bayesian log-likelihood loss function using the input of the current batch of training samples and the corresponding output labels, and perform gradient descent optimization on all trainable parameters in the Bayesian deep inversion network model based on the Bayesian log-likelihood loss function. The parameter update process uses the learning rate as the control factor to gradually correct the model weights according to the current gradient direction;

[0119] S35. Continuously execute the training iteration until the Bayesian log-likelihood loss function converges on the validation set or reaches the preset upper limit of the number of training epochs. After the training is completed, a set of trained model parameters is formed.

[0120] In this embodiment, the Bayesian log-likelihood loss function consists of two parts: the first part is the log-likelihood term of the posterior distribution, which is used to evaluate the matching degree of the model output results under the guidance of the dynamic prior; the second part is the regularization term, which is used to constrain the magnitude of the network weight parameters to prevent the model from overfitting in the marine multi-source data. The minimization objective of the Bayesian log-likelihood loss function is to maximize the credibility of the estimation results.

[0121] In this embodiment, S4 includes the following steps:

[0122] S41. Feed the input vector in the real-time multi-source dataset into the input layer of the trained Bayesian depth inversion network model, and process it sequentially through the ocean environment channel, squid behavior channel, and fusion nested layer to generate a fused feature representation. The fused feature representation comprehensively reflects the non-linear coupling relationship between the current ocean environment changes and the individual behavior characteristics of squids;

[0123] S42. Call the dynamic prior model to update the prior distribution at the current moment. The update method is based on the historical ocean environment parameters and squid biological activity parameters recorded in multiple consecutive time steps before the current time point. The historical parameters are assigned decreasing weights in chronological order, and the dynamic prior data of squid resources corresponding to the current time point is calculated. The dynamic prior data of squid resources is used to express the empirical distribution trend of squid resources in the current environmental state;

[0124] S43. Use the fused feature representation and the dynamic prior data of squid resources together in the Bayesian posterior inference process. According to the Bayesian principle, combine the likelihood information under the current feature conditions with the dynamic prior data of squid resources to estimate the posterior distribution of the squid resource distribution position and density prediction value at the current moment;

[0125] S44. Calculate the uncertainty quantification index corresponding to the current prediction result based on the posterior distribution of squid resources. The uncertainty index includes the variances of the predicted position coordinates in three spatial dimensions, and the volatility variance of the squid density estimation result;

[0126] S45. Output the dynamic estimation result of squid resources at the current moment, including the predicted position coordinates and density values, and simultaneously output the corresponding uncertainty quantification index.

[0127] In this embodiment, S5 includes the following steps:

[0128] S51. Structurally organize the dynamic estimation result of squid resources output at the current moment and its corresponding uncertainty quantification index :

[0129] ;

[0130] Among them, , , respectively represent the variance estimation values of the three spatial dimension components of the predicted position under the Bayesian posterior distribution, represents the variance of the squid density prediction value at the current time step;

[0131] The squid spatial position coordinates in the dynamic estimation result of squid resources and the squid density estimation value at time step t Construct a squid resource estimation data table according to the time series and geographical region dimensions;

[0132] S52. Introduce uncertainty information based on the squid resource estimation data table, establish confidence level labels for each prediction cell, and divide the credibility levels;

[0133] S53. Based on the sorted predicted squid spatial position coordinates and squid density estimates, combine the sea area grid division rules to conduct regional aggregation statistics on the squid resources, calculate the average squid density, density change rate, and confidence level distribution within each sub-region per unit time, and form the spatial aggregation characteristics at the regional level;

[0134] The implementation process of the spatial aggregation characteristics at the regional level is based on the completed dynamic estimation results of squid resources and uncertainty quantification indicators, and specifically includes the following implementation methods:

[0135] Combine the output squid spatial position coordinates with the squid density estimates , classify all prediction results according to the three-dimensional grid spatial positions of the sea areas where they are located. The three-dimensional grid division rules of the sea areas are preset according to fixed longitude, latitude, and depth resolutions, divided into multiple spatial cells, and each cell corresponds to a spatial index number.

[0136] Based on the multiple prediction points belonging to each spatial cell, organize them in the time dimension according to the unit time window. Under each time window, count the squid density estimates of all prediction points in each grid cell and calculate its regional average density value to characterize the resource abundance characteristics of the region at the current time point.

[0137] Differentiate adjacent time steps in the order of time windows, extract the squid density changes in the same spatial cell within the unit time interval, and then calculate the density change rate of the cell to reflect the dynamic trend of the squid resources in the region.

[0138] And combine the uncertainty quantification index set , based on the variance weighting method of each prediction position or directly using the maximum variance value, conduct aggregation analysis on the confidence levels of multiple prediction points in each grid cell. Through the normalization and superposition of uncertainty indicators, calculate the average uncertainty level of the region, and accordingly mark the confidence level distribution at the regional level, and form the spatial distribution mapping of high-confidence regions, medium-confidence regions, and low-confidence regions.

[0139] Finally, output the results as regional statistical characteristics, constituting the spatial aggregation characteristic items of each spatial grid cell under the current time window, including: regional average squid density, regional squid density change rate, regional confidence level distribution.

[0140] S54. Construct the structure of the dynamic estimation report for squid resources. The report includes the following information fields:

[0141] Time label: The current predicted time step ;

[0142] Location label: The spatial grid number of the predicted sea area

[0143] Prediction result: The spatial position coordinates of squid , and the estimated squid density ;

[0144] Uncertainty quantification index: The index set composed of the variance of position coordinates and the variance of density ;

[0145] Confidence level: High confidence / Medium confidence / Low confidence classification label

[0146] Spatial aggregation characteristics: Regional average density value, regional density volatility index, local density gradient trend index

[0147] In this embodiment, the rules for dividing the confidence level are as follows:

[0148] High-confidence area: If , it is marked as high confidence;

[0149] Medium-confidence area: If , it is marked as medium confidence;

[0150] Low-confidence area: If , it is marked as low confidence;

[0151] Among them, and are the threshold values for dividing the uncertainty interval.

[0152] Example 1: At 4:12 am on October 3, 2024, in the waters east of Island A, the on-duty staff at the Marine Resources Information Center noticed an abnormal phenomenon in the monitoring system: The real-time observation data deployed on the "Island A - 07" intelligent buoy platform showed that the temperature of the near-bottom seawater suddenly dropped from 22.4°C to 20.7°C, the salinity increased from 32.8‰ to 33.5‰, and at the same time, the sea current speed fluctuated violently for a short time to reach 1.93 m / s. This abnormal phenomenon lasted for about 18 minutes. By retrieving the historical dynamic estimation reports of squid resources within 30 kilometers of this sea area through the dispatching platform, it was found that within the past 72 hours, the predicted squid density in this area showed a stable fluctuation trend, with an average value of about 6.2 kg / km². However, starting from 4:30 am, the underwater acoustic sensors deployed on the "Island A - 07" buoy began to continuously detect abnormal high-frequency echoes, and it was initially judged that there was a squid aggregation.

[0153] Since traditional estimation models (based on time-series LSTM) usually require more than 6 hours of historical data playback for estimation and prediction, and the prediction does not distinguish uncertainty, the duty officer decided to enable the Bayesian deep inversion resource estimation system deployed on the server of "Island A - Resource Regulation Node 02" of the present invention for a real-time dynamic inversion.

[0154] At 4:46 in the early morning, the system began to receive linkage data packets sent by the buoy of "Island A - 07", the submersible buoy of "Island B - 03", and the remote sensing data processing terminal of "Island C - 02". The system quickly processed the three types of data:

[0155] A total of 8,923 acquisition records were processed in this round of estimation;

[0156] Each record contains a timestamp, seawater temperature, salinity, chlorophyll concentration, sea current velocity, as well as underwater acoustic estimation density and initial coordinate points;

[0157] The processed data formed a complete time-synchronized sample set, with a missing rate of only 1.8%, and 273 outliers were removed.

[0158] The system completed the real-time update operation of the dynamic prior model at 4:52 in the early morning: Select historical estimation records in the same sea area within the past 180 minutes (a total of 36 time points), calculate the prior density distribution by exponential decay weighting. After the Bayesian inversion model fused the current input features and the dynamic prior, it completed the grid-level resource inversion of the entire sea area in only 32 seconds.

[0159] At this time, the system generated a "Squid Resource Dynamic Inversion Prediction Report" in the background, which mentioned:

[0160] Prediction time: 2024-10-03 04:52;

[0161] Target area: 123.107° east longitude, 29.744° north latitude, depth -38.2 meters;

[0162] Predicted squid density value: 9.73 kg / km²;

[0163] Coordinate space prediction variance: (σx = 5.3m, σy = 6.1m, σz = 3.8m);

[0164] Density variance: 0.44 kg / km²;

[0165] Confidence level: High confidence.

[0166] According to the automatic confidence differentiation strategy, the system has highlighted an area of approximately 19.4 square kilometers in the southeast of the line connecting Island A and Island B as a high-confidence prediction area, and recommends deploying working boats to this area to carry out the "quick exploration and slow encirclement" net operation. Immediately push this report to the dispatching terminal of the Municipal Bureau, and send a work suggestion notice to fishing boat No. "32089".

[0167] The fishing boat set off from the port at 5:30 in the morning and arrived at the target area at 7:20. A total of 3 rounds of on-site trawling operations were carried out, with each round lasting 30 minutes. The final operation statistics are as follows:

[0168] The first round of trawling: The total amount of squid caught was 714 kg, the net opening coverage area was 7.1 km², and the measured density was 10.04 kg / km²;

[0169] The second round of trawling: The measured density was 9.87 kg / km²;

[0170] The third round of trawling: The measured density was 10.23 kg / km²;

[0171] The average measured density of the 3 rounds: 10.05 kg / km²;

[0172] The deviation from the predicted value of 9.73 kg / km² of the present invention is only 3.18%, and the confidence interval completely contains the measured value, verifying a high degree of reliability.

[0173] As a comparison, at the same time, the Municipal Bureau also enabled the still-running traditional LSTM resource estimation system to make a prediction for this area. The predicted density it returned was 6.2 kg / km², and the position deviation was about 2.7 kilometers. It was not marked as a fishing hot spot. If relying on the traditional system, the operation on that day would completely bypass the high-density area, and the prediction effect was far inferior to that of the system of the present invention.

[0174] The system also automatically generated a prediction reliability distribution map, a regional heat overlay map, a volatility assessment map, and a work suggestion layer during this operation, and archived them into the "Resource Decision-making Assistance Platform" of the Municipal Bureau.

[0175] In addition, a systematic evaluation was carried out on all the forecast data in the sea area of Island A from September 30th to October 5th:

[0176]

[0177] The comparison data fully verified the significant advantages of the system of the present invention in terms of dynamic response speed, prediction space accuracy, density estimation error control, uncertainty modeling, and visualization output. Especially in the context of complex environmental disturbances and rapid ocean changes, this system demonstrated stable, reliable, and practical characteristics.

[0178] The present invention constructs a dual-channel structure of a marine environment channel and a squid behavior channel, separately extracts independent feature representations of physical parameters such as temperature, salinity, and flow velocity and behavior parameters such as squid distribution coordinates and density, and performs feature concatenation in a nested fusion layer to effectively capture the non-linear response characteristics of environmental changes to squid aggregation behavior.

[0179] The present invention designs a dynamic prior generation module, which performs weighted fusion on environmental data and behavior parameters within the past K time steps based on a time decay function to generate a prior estimation distribution at time t, and performs Bayesian posterior inference with the current fusion features, enabling the model to make an estimation by combining historical trends at each prediction time point, reflecting the inertia and delay characteristics of the marine ecosystem, and significantly enhancing the adaptability of the model to complex background changes in the ocean.

[0180] The present invention obtains the posterior distribution variance of the predicted position and density of squid through a Bayesian model structure, and accordingly defines a confidence level label for annotating the credibility of each predicted grid cell. In addition, based on the variance aggregation index, thresholds δ1 and δ2 are set to achieve regional-level prediction credibility screening and aggregation statistics support, significantly reducing the false alarm rate in low-confidence prediction regions and improving the prediction stability in high-confidence regions, providing more reliable support data for fishery administration vessel scheduling and fishing path planning.

[0181] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for estimating squid resources dynamics based on a Bayesian deep inversion network, characterized in that: The steps include: S1. Collecting a real-time multi-source data set including marine environment parameters and squid biological activity parameters, preprocessing the real-time multi-source data set to form a preprocessed multi-source data set; S2. Construct a Bayesian deep inversion network model; The S2 comprises the following steps: S21. Constructing a Bayesian deep inversion network model The Bayesian deep inversion network model includes an input layer, a dual-channel feature extraction structure, a fusion nesting layer, an output layer, and a dynamic prior model. The input layer receives the preprocessed multi-source data set. , used to study the dynamic distribution behavior characteristics of squid resources under different marine environmental conditions; S22. Construct a dual-channel feature extraction structure for the Bayesian depth inversion network model. The dual-channel feature extraction structures are the ocean environment channel Channel with Squid Behavior , used to simulate the response mechanism of squid to environmental stimuli; S23. Set a fusion nesting layer to perform feature-level connection operations on the feature vectors of the ocean environment channel and the squid behavior channel to obtain a fusion feature representation. ; S24. Set the squid resource estimation output of the output layer of the Bayesian deep inversion network model ; S25. Constructing a dynamic prior model , based on historical marine environmental parameters and squid biological activity parameters, a dynamic prior model with time weight is constructed: ； in, Represents the current moment generated by the dynamic prior model Dynamic prior data of squid resources, The history The ocean environment parameters and squid biological activity parameters at each moment, is the time decay weight, Extraction function for statistical features; S27. Dynamic prior data of squid resources and fusion feature representation Conduct Bayesian posterior modeling to calculate the posterior output distribution of squid resource dynamics estimates; S3. Use the preprocessed multi-source data set to train the Bayesian deep inversion network model, use the dynamic prior update module to update the dynamic prior data in real time during the training process, and optimize the model parameters by maximizing the likelihood function and minimizing the prediction error to form a set of trained model parameters; S4. Input the multi-source data set collected in real time into the trained Bayesian deep inversion network model, and use the dynamic prior update module to update the prior data of the real-time multi-source data set to obtain the dynamic estimation results of squid resources and their uncertainty quantitative indicators; S5. Post-process the squid resource dynamic estimation results and uncertainty quantitative indicators to form a squid resource dynamic estimation report with temporal and spatial distribution characteristics; S6. Interface the squid resource dynamic estimation report with the fishery management system to conduct dynamic monitoring of squid resources and real-time decision-making.

2. The method for estimating squid resources dynamics based on a Bayesian deep inversion network according to claim 1, characterized in that: The S1 comprises the following steps: S11. Collect real-time multi-source data including marine environmental parameters and squid biological activity parameters from ocean observation platforms, satellite remote sensing devices and underwater sensors to build a real-time multi-source data set : ； in, Indicates Original data records, Indicates the total number of real-time collected data. For the The collection timestamp of the data. For the The ocean environment parameter vectors recorded in ,salinity , chlorophyll concentration , current speed , For the The squid biological activity parameter vectors recorded in the 100 records represent the distribution location coordinates Biomass density index ; S12. Real-time multi-source datasets Perform time synchronization processing, uniformly map multi-source data from different devices and different time scales to a standard time series set, use a unified time step to time align all multi-source data, and construct a time-synchronized multi-source data set so that all parameters have a consistent time reference benchmark; S13. Perform spatial interpolation processing on the time-synchronized multi-source datasets. For areas with missing measurements in geographic space, interpolation estimation is performed based on the neighboring values ​​of the observed locations and their spatial positional relationships to fill in the missing spatial locations, thereby forming a spatially completed multi-source dataset, so that the marine environmental parameters and squid biological activity parameters have complete spatial coverage in the target area; S14. Fill missing values ​​in the spatially completed multi-source dataset. For the marine environmental parameters or squid biological activity parameter vectors that are still missing in non-spatial dimensions, use the sliding time window method to fill in the data based on the historical observations at the nearest moment. The filled data form the multi-source dataset after missing values ​​are processed. S15. Remove outliers from the multi-source data set after missing values ​​processing, calculate the historical mean and standard deviation of each parameter, and determine whether the current observation value of the multi-source data set after missing values ​​processing is within the preset reasonable range. If the deviation exceeds the set threshold, the record is marked as abnormal data and removed, and finally the pre-processed multi-source data set is formed. .

3. The method for dynamic estimation of squid resources based on Bayesian deep inversion network according to claim 2 is characterized in that: In the marine environment channel The residual structure is constructed in the process to extract stable background change features: ； in, For the Layer environmental feature representation, which represents the implicit response characteristics of squid to marine environmental parameters, Indicates Layer environment feature representation, and are weights and biases, The Sigmoid function represents the attention weight, and f is the activation function; In the Squid Behavior Channel The attention mechanism is introduced to enhance the recognition ability of squid gathering area features: ； in, The Sigmoid function represents the attention weight, is the Hadamard product, For the Layer attention map parameters, For the The output feature vector of the squid behavior channel represents the spatial behavior characteristics of individual or group squids. For the Output feature vector of the layer squid behavior channel.

4. The method for estimating squid resources dynamics based on Bayesian deep inversion network according to claim 3 is characterized in that: The S3 comprises the following steps: S31. Using preprocessed multi-source datasets Bayesian Deep Inversion Network Model Training is performed, and the training samples are composed of multiple input and output pairs, where each input vector is composed of the concatenation of marine environmental parameters and squid biological activity parameters, and the output label is composed of the corresponding squid real distribution position coordinates and density values; S32. During the training process, construct a Bayesian log-likelihood loss function including a dynamic prior constraint; S33. Before each training batch starts, the dynamic prior model is called to calculate the prior model based on the current training time step. The historical ocean environment parameters and squid biological activity parameters at each time point are used to update the squid resource dynamic prior data. The updated squid resource dynamic prior data is the squid resource dynamic prior data at the current time step, which is used to participate in the calculation of the posterior probability in the training loss, so that the training process continues to reflect the temporal evolution characteristics of the ocean environment and the dynamic changes of squid behavior. S34. In each training iteration, the Bayesian log-likelihood loss function is calculated using the training sample input of the current batch and the corresponding output label, and all trainable parameters in the Bayesian deep inversion network model are optimized by gradient descent based on the Bayesian log-likelihood loss function. The parameter update process uses the learning rate as the control factor, and the model weights are gradually corrected according to the current gradient direction; S35. Continue to perform training iterations until the Bayesian log-likelihood loss function becomes stable on the validation set or reaches the preset upper limit of the number of training rounds. After the training is completed, a set of trained model parameters is formed.

5. The method for estimating squid resources dynamics based on Bayesian deep inversion network according to claim 4 is characterized in that: The Bayesian log-likelihood loss function consists of two parts: the first part is the log-likelihood term of the posterior distribution, which is used to evaluate the degree of matching of the model output results under the guidance of dynamic priors; the second part is the regularization term, which is used to constrain the amplitude of the network weight parameters to prevent the model from overfitting in marine multi-source data. The minimization goal of the Bayesian log-likelihood loss function is to maximize the credibility of the estimation results.

6. The method for estimating squid resources dynamics based on Bayesian deep inversion network according to claim 5 is characterized in that: The S4 comprises the following steps: S41. The input vector in the real-time multi-source data set is sent to the input layer of the trained Bayesian deep inversion network model, and processed in sequence through the ocean environment channel, the squid behavior channel and the fusion nested layer to generate a fused feature representation, which comprehensively reflects the nonlinear coupling relationship between the current ocean environment changes and the squid individual behavior characteristics; S42. Call the dynamic prior model to update the prior distribution at the current time. The update method is based on the historical marine environmental parameters and squid biological activity parameters recorded in multiple consecutive time steps before the current time point. The historical parameters are assigned decreasing weights in chronological order. The dynamic prior data of squid resources corresponding to the current time point is calculated. The dynamic prior data of squid resources is used to express the empirical distribution trend of squid resources under the current environmental state; S43. The fusion feature representation and the squid resource dynamic prior data are used together in the Bayesian posterior inference process, and the posterior distribution of the squid resource distribution position and density prediction value of the squid resource at the current moment is estimated according to the Bayesian principle in combination with the likelihood information under the current feature conditions and the squid resource dynamic prior data; S44. Calculate the uncertainty quantification index corresponding to the current prediction result based on the posterior distribution of squid resources, the uncertainty index includes the variance of the predicted position coordinates in three spatial dimensions and the volatility variance of the squid density estimation result; S45. Output the dynamic estimation result of squid resources at the current moment, including the predicted position coordinates and density value, and simultaneously output the corresponding uncertainty quantitative index.

7. The method for estimating squid resources dynamics based on Bayesian deep inversion network according to claim 6 is characterized in that: The S5 comprises the following steps: S51. Dynamic estimation result of squid resources output at the current moment The corresponding uncertainty quantitative index Structuring: ； in, , , They represent the variance estimates of the three spatial dimension components of the predicted position under the Bayesian posterior distribution, Represents the variance of the squid density prediction at the current time step; The squid spatial position coordinates in the squid resource dynamic estimation result and the estimated squid density at time step t Construct a squid resource estimation data table based on time series and geographical area dimensions; S52. Introducing uncertainty information based on the squid resource estimation data table, establishing a confidence label for each prediction cell, and dividing the confidence levels; S53. Based on the sorted predicted squid spatial position coordinates and squid density estimation, combined with the sea area grid division rules, regional aggregation statistics of squid resources are performed, and the average squid density, density change rate and confidence level distribution per unit time in each sub-area are calculated to form regional-level spatial aggregation characteristics; S54. Construct a dynamic estimation report structure for squid resources. The report includes the following information fields: Time label: current prediction time step ; Location tag: spatial grid number of the predicted sea area; Prediction results: Squid spatial position coordinates , Squid density estimation ; Uncertainty quantification index: a set of indicators consisting of position coordinate variance and density variance ; Confidence level: high confidence / medium confidence / low confidence classification label; Spatial aggregation characteristics: regional average density value, regional density volatility index, local density gradient trend index.

8. The method for estimating squid resources dynamics based on Bayesian deep inversion network according to claim 7 is characterized in that: The credibility level classification rules are as follows: High confidence area: If , marked as high confidence; Medium Trusted Area: If , marked as medium confidence; Low confidence zone: If , marked as low confidence; in, and Threshold the uncertainty interval.

Images