Multi-modal and deep CNN (Convolutional Neural Network) offshore culture intelligent edge detection method and system

Abstract

The invention relates to the technical field of offshore culture intelligent monitoring, and discloses a multi-mode and deep CNN offshore culture intelligent edge detection method and system, and the method comprises the steps: obtaining local multi-depth water sample nutritive salt concentration data and nitrogen and phosphorus isotope composition data of each farm, and generating a local pollution source isotope fingerprint vector; adding Laplacian noise to the isotope fingerprint vector of each farm based on a differential privacy algorithm, and calculating a disturbance fingerprint vector meeting epsilon-differential privacy; combining the combined fingerprint feature space with a hydrological transmission model, and decomposing contribution proportions of different sources in nutritive salt at each observation point by using a federal Bayesian mixture model; and outputting a real-time contribution ratio and a confidence interval of each pollution source based on the trained pollution source identification model. According to the invention, accurate pollution source identification and contribution proportion quantification are realized on the premise of protecting the privacy of each party.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for nearshore aquaculture, and more specifically, to a multimodal and deep CNN intelligent edge detection method and system for nearshore aquaculture. Background Technology

[0002] In nearshore aquaculture areas, multiple farms share the same aquatic environment. Pollutants from each farm, along with surrounding land-based pollution and offshore inputs, create a complex and compound pollution situation. Accurately identifying the contribution ratio of each pollution source is crucial for developing scientific pollution control strategies. Traditional pollution source apportionment methods rely on isotope fingerprinting technology, which traces the source of pollutants by analyzing the nitrogen and phosphorus isotope composition of water. However, practical applications face the following technical challenges: isotope samples collected from a single farm lack sufficient spatiotemporal coverage, making it difficult to construct a complete regional pollution source isotope fingerprint spectral library; when each farm conducts independent source apportionment, it lacks characteristic data from neighboring pollution sources, leading to significant deviations in the calculation of source contribution ratios; furthermore, the raw isotope data contains commercially sensitive information such as farming density and feed formulations, and farms are unwilling to share this data due to commercial competition, creating data silos that severely restrict the effectiveness of collaborative regional pollution control. Summary of the Invention

[0003] This invention provides a multimodal and deep CNN-based intelligent edge detection method and system for nearshore aquaculture, addressing the obstacle of aquaculture farms' reluctance to share data due to concerns about the leakage of commercial information. It overcomes the source resolution bias caused by insufficient spatiotemporal coverage of data from a single aquaculture farm, achieving accurate pollution source identification and contribution ratio quantification while protecting the privacy of all parties.

[0004] The first aspect of this invention provides a multimodal and deep CNN-based intelligent edge detection method for nearshore aquaculture, comprising: Nutrient concentration data and nitrogen and phosphorus isotope composition data of water samples from various aquaculture farms at multiple depths were obtained. Wavelet transform was used to extract the spatiotemporal variation features of isotopes and generate local pollution source isotope fingerprint vectors. Based on the differential privacy algorithm, Laplace noise is added to the isotopic fingerprint vectors of each farm to calculate the perturbed fingerprint vectors that satisfy ε-differential privacy. The perturbation fingerprint vectors of each farm are aggregated using a secure multi-party computation protocol, and the regional isotopic fingerprint covariance matrix is ​​calculated in the ciphertext domain using homomorphic encryption technology. By combining the joint fingerprint feature space with the hydrological transport model, an enhanced source-sink transport matrix is ​​constructed, and the contribution ratio of different sources of nutrients in each observation point is decomposed using a federated Bayesian mixture model. Based on the trained pollution source identification model, real-time monitored nutrient concentration and isotope data are input, and the real-time contribution ratio and confidence interval of each pollution source are output.

[0005] Furthermore, the extraction of isotope spatiotemporal variation features using wavelet transform includes: The Morlet wavelet was used as the mother wavelet function to perform continuous wavelet transform on the standardized isotope time series data. The transformation coefficients are calculated at different scales and different translation positions to obtain the multi-scale decomposition results in the time-frequency domain. Statistical features are extracted from multi-scale wavelet coefficients, including energy distribution, peak position, and fluctuation amplitude at each scale. The extracted features are vectorized to construct the local pollution source isotopic fingerprint vector.

[0006] Furthermore, the process of adding Laplacian noise to the isotopic fingerprint vector based on the differential privacy algorithm includes: Calculate the global sensitivity of the isotopic fingerprint vector by taking twice the maximum change in the isotopic ratio. For each component of the isotopic fingerprint vector of each farm, noise values ​​are randomly sampled from a Laplace distribution with a scale parameter equal to the global sensitivity divided by the privacy budget. The sampled noise values ​​are added to the corresponding fingerprint vector components to generate a perturbed fingerprint vector that satisfies ε-differential privacy protection.

[0007] Furthermore, the process of calculating the regional isotopic fingerprint covariance matrix in the ciphertext domain using homomorphic encryption technology includes: Each farm uses the Paillier homomorphic encryption scheme to encrypt the perturbation fingerprint vector; By utilizing the homomorphic property of Paillier encryption, the elements of the covariance matrix are calculated in the ciphertext field through homomorphic multiplication. Through a secure multi-party computation protocol, all parties collaborate to complete the computation of the ciphertext field; The regional isotopic fingerprint covariance matrix is ​​obtained by using a threshold decryption method.

[0008] Furthermore, the process of constructing the enhanced source-sink transfer matrix includes: The original transport coefficient from each pollution source to the observation point is calculated based on spatial distance, average flow velocity, and diffusion coefficient. The original transmission coefficients are normalized to obtain a dimensionless transmission weight matrix. Each element in the transmission weight matrix represents the relative transmission contribution of the corresponding pollution source to the observation point, and the sum of the transmission weights of all pollution sources corresponding to each observation point is 1.

[0009] Furthermore, the training process of the federated Bayesian mixture model includes: A probabilistic mixture model based on a multivariate normal distribution is constructed, with model parameters including the contribution weights of each pollution source, the mean of the isotopic fingerprint, and the covariance. The model training uses the expectation-maximization algorithm, which consists of two alternating units: the E-step for calculating the posterior probability and the M-step for updating the model parameters. The negative log-likelihood function is used as the loss function, and training stops when the change in the loss function between two consecutive iterations is less than the convergence threshold.

[0010] Furthermore, the process of outputting the real-time contribution ratio and confidence interval of each pollution source includes: The real-time monitoring data is input into the trained model, and the posterior probability of each pollution source is calculated as the contribution ratio. The Markov chain Monte Carlo method is used to sample from the posterior distribution; Parameter sample sequences are generated using the Metropolis-Hastings algorithm; The confidence interval for the contribution ratio of each pollution source is calculated based on the sampling results.

[0011] Furthermore, it also includes: Vertical interpolation is performed on multi-depth sampled data to generate isotope distributions of continuous depth profiles; Calculate the coefficient of variation of isotope ratios at each depth layer to identify isotope layering structures; Based on the vertical gradient characteristics of isotopes, the sedimentation or floating trend of pollutants can be determined.

[0012] Furthermore, it also includes: Historical detection results are cached, and a fast query index based on a KD tree is established; When the Euclidean distance between the new data and the cached data is less than the similarity threshold, the cached result is returned directly.

[0013] A second aspect of the present invention provides a multimodal and deep CNN-based intelligent edge detection system for nearshore aquaculture, comprising: The data acquisition module is used to acquire multi-depth water sample data from various aquaculture farms; The privacy protection module is used to perform differential privacy perturbation on the isotope fingerprint vector; The secure computation module is used to aggregate and compute the covariance matrix in the ciphertext domain; The model training module is used to train federated Bayesian mixture models; The real-time inference module is used to output the contribution ratio of pollution sources and the confidence interval.

[0014] The beneficial effects of this invention are as follows: This invention protects isotopic fingerprint data using differential privacy technology. The added Laplace noise ensures that the original data meets the ε-differential privacy definition, thus overcoming the obstacle of farms being unwilling to share data due to concerns about the leakage of commercial information, and solving the data silo problem. Simultaneously, homomorphic encryption technology enables secure aggregation computation of multi-party data in the ciphertext domain. Each participant only needs to upload encrypted data without exposing the original isotopic features, further enhancing data security. The federated Bayesian mixture model fully utilizes the distributed isotopic data collected from various farms. By constructing a complete regional isotopic fingerprint feature space, it overcomes the source resolution bias caused by insufficient spatiotemporal coverage of data from a single farm, achieving accurate pollution source identification and contribution ratio quantification while protecting the privacy of all parties. Attached Figure Description

[0015] Figure 1 This is a flowchart of a multimodal and deep CNN intelligent edge detection method for nearshore aquaculture according to the present invention; Figure 2 This is a bar chart of nutrient concentration distribution in multi-depth water samples according to the present invention; Figure 3 This is a multi-scale energy distribution piecewise linear plot of nitrogen isotope wavelet coefficients according to the present invention; Figure 4 This is a scatter plot comparing the eigenvalues ​​before and after the differential privacy perturbation of this invention; Figure 5 This is a mixed map of the contribution weights of each pollution source and isotopic fingerprint features of the present invention; Figure 6 This is a bar chart of the real-time pollution source contribution ratio identification results (including confidence intervals) of the present invention; Figure 7 This is a line graph showing the training and convergence process of the EM algorithm of this invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0017] At least one embodiment of the present invention discloses a multimodal and deep CNN-based intelligent edge detection method for nearshore aquaculture, such as... Figure 1 As shown, it includes the following steps: Step 100: Obtain nutrient concentration data and nitrogen and phosphorus isotope composition data of local multi-depth water samples from each aquaculture farm, extract the spatiotemporal variation features of isotopes using wavelet transform, and generate local pollution source isotope fingerprint vectors.

[0018] Specifically, each aquaculture farm sets up multiple sampling points within its aquaculture area. Water samples are collected at three depths: surface (0-0.5m), middle layer (2-3m), and bottom layer (0.5m from the seabed). Nutrient concentrations, including ammonia nitrogen (NH4+), are determined from the collected water samples. + -N), nitrate (NO3) - -N), total nitrogen (TN), orthophosphate (PO4) 3- Nitrogen isotope ratios (δ¹²P) and total phosphorus (TP) concentrations were also measured using isotope mass spectrometry. 15 N) and phosphorus isotope ratio (δ) 18 O-PO4).

[0019] Furthermore, the effective measurement range for the aforementioned nutrient concentrations is as follows: ammonia nitrogen concentration mg / L, nitrate concentration mg / L, total nitrogen concentration mg / L, orthophosphate concentration mg / L, total phosphorus concentration mg / L; when the measured value is below the lower limit, it is recorded as half of the detection limit; when the measured value is above the upper limit, it needs to be diluted and re-measured; the effective measurement range of the aforementioned isotope ratios is: nitrogen isotope ratios phosphorus isotope ratio These ranges cover typical isotopic composition variations in nearshore aquaculture areas; when a measurement exceeds the valid range, it is identified as an outlier and removed, and resampling and measurement are performed.

[0020] The acquired nutrient concentration and isotope ratio data were preprocessed. Since nutrient concentration (in mg / L) and isotope ratio (in ‰) have different dimensions and numerical ranges, Z-score standardization was performed on each type of data to eliminate the influence of dimensional differences on subsequent feature extraction, converting the data into a standard distribution with a mean of 0 and a standard deviation of 1.

[0021] Wavelet transform was used to extract spatiotemporal variation features from the acquired standardized isotope time series data.

[0022] The input to the aforementioned wavelet transform algorithm is a time-series signal of isotope ratios:

[0023] in For each time point in the sampling time sequence, The total number of sampling times is given, and the output is a multi-scale wavelet coefficient matrix. .

[0024] in For each scale value of the selected scale sequence, This represents the total number of scales.

[0025] Furthermore, the aforementioned sampling time interval The value range is determined based on the time scale of the monitored target; for capturing intraday variation characteristics, Values Hours; for capturing seasonal changes, Values Days; Time series length The range of values ​​is The smaller value corresponds to short-term monitoring (such as hourly data over a month). Larger values ​​correspond to long-term monitoring (e.g., 2 years of weekly data). or daily data To ensure the effectiveness of wavelet transform, the time series length... Should meet ,in The maximum number of scale layers is used to ensure that the wavelet basis functions at the maximum scale can fully cover the low-frequency components of the signal.

[0026] The aforementioned formula for calculating wavelet transform is:

[0027] in, For the mother wavelet function, For scale parameters, For translation parameters, Represents complex conjugation. This is achieved through different scales. and different translation positions The transformation coefficients are calculated to obtain the multi-scale decomposition results in the time-frequency domain.

[0028] Furthermore, the aforementioned wavelet transform is implemented using discretization in actual calculations, converting continuous integrals into discrete summations: ,in The sampling time interval, The length of the time series. For time point index; translation parameters The value is , ,in The translation position index is used to translate the wavelet function along the time axis in steps of sampling intervals; for each scale ( (for scale index) and each translation position The wavelet coefficients were calculated. This forms a scale-time two-dimensional coefficient matrix.

[0029] Furthermore, statistical features are extracted from the multi-scale wavelet coefficients, including energy distribution, peak position, and fluctuation amplitude at each scale; the extracted features are then vectorized to construct a local pollution source isotopic fingerprint vector. ,in Indicates the first A breeding farm, For feature dimensions.

[0030] Furthermore, the formulas for calculating the energy values ​​at each of the aforementioned scales are as follows: ,in For translation position index, The length of the time series indicates the length of the first time series. The sum of the squared moduli of all wavelet coefficients at each scale reflects the total energy at that scale; the formula for calculating the peak position is... , indicating the first The time position index corresponding to the maximum modulus of wavelet coefficients at each scale; the formula for calculating the fluctuation amplitude is... , indicating the first The range of wavelet coefficient modulus variation at each scale; the formula for calculating the cross-scale energy ratio is: , ,in For scale indexing, The total number of scales represents the energy ratio between adjacent scales, reflecting the energy distribution characteristics across different scales.

[0031] It should be noted that the mother wavelet function is the Morlet wavelet, and its expression is:

[0032] in The center frequency is typically set to 6.

[0033] Furthermore, the aforementioned scale sequence The selection is based on the sampling period of the isotopic time series data, and the specific value is [value to be filled in]. ,in The sampling time interval (unit: days) Scale number of layers The range of values ​​is To cover multi-timescale features ranging from intraday to seasonal variations.

[0034] Furthermore, the aforementioned feature dimensions The quantity of statistical features of wavelet coefficients is determined by the number of features, specifically including: energy values ​​at each scale ( (number), peak position () Individual, fluctuation range ( (number) and cross-scale energy ratio (number) (number of features), therefore feature dimension When the number of scale layers At that time, feature dimension .

[0035] Furthermore, the aforementioned center frequency The theoretical basis for the value of 6 is as follows: Center frequency A value of 6 ensures that the Morlet wavelet exhibits good localization properties in both the time and frequency domains, satisfying the admissibility condition. ,in For the Fourier transform of the wavelet function, when At that time, the product of the time-frequency uncertainty of the wavelet function reaches close to the theoretical lower limit. The optimal value.

[0036] In this embodiment of the application, in order to enhance the spatiotemporal representativeness of isotopic fingerprints, the following steps are also included in addition to step 100: Step 101: Perform vertical interpolation on the multi-depth sampling data to generate the isotope distribution of the continuous depth profile.

[0037] Furthermore, the vertical interpolation in step 101 above adopts the cubic spline interpolation method, and the specific implementation steps are as follows: Let the three sampling depths be respectively (surface layer), (Middle level) (Bottom layer), the corresponding isotope ratio is , , ; Construct a piecewise cubic polynomial:

[0038] in Represent two interpolation intervals; by boundary conditions and continuity conditions , Solving for coefficients For depth range any depth within The interpolated isotope ratios are calculated using cubic polynomials within the corresponding intervals. This generates a continuous depth profile.

[0039] Furthermore, the boundary conditions for the aforementioned cubic spline interpolation adopt natural boundary conditions, that is, the second derivative is zero at both endpoints: and Combine internal nodes The continuity condition at the point forms a condition about the coefficient. A system of linear equations, where For the interpolation interval index, specifically, for the first interval ,have For the second interval ,have By solving the tridiagonal linear equation system Obtain all coefficients, where the coefficient matrix is... It is a tridiagonal matrix, with vectors on the right side. From the function value at the node The boundary conditions are determined; the cubic spline interpolation method ensures that the interpolation function has a continuous second derivative throughout the entire depth range, thus guaranteeing the smoothness of the isotope profile.

[0040] Step 102: Calculate the coefficient of variation of isotope ratios at each depth to identify the isotope layering structure.

[0041] Step 103: Based on the vertical gradient characteristics of isotopes, determine the sedimentation or floating trend of pollutants.

[0042] Furthermore, the formula for calculating the coefficient of variation in step 102 above is as follows: ,in The standard deviation of the isotope ratios at each depth layer. The average isotope ratio for each depth layer; when the coefficient of variation At that time, a significant isotopic stratification structure was determined, indicating that the pollution source characteristics of different water depths differed.

[0043] Furthermore, the formula for calculating the vertical gradient in step 103 above is as follows: ,in The ratio of the underlying isotopes. The surface isotope ratio, Total water depth (unit: m); when When it is determined that the pollutants have a clear settling trend, When it is determined that pollutants have a clear upward trend, It is determined that pollutants are evenly distributed in the water body.

[0044] Step 200: Add Laplace noise to the isotopic fingerprint vectors of each farm based on the differential privacy algorithm, and calculate the perturbed fingerprint vector that satisfies ε-differential privacy.

[0045] The input to the aforementioned differential privacy algorithm is the isotopic fingerprint vector of local pollution sources for each farm. and privacy budget parameters The output is satisfied. - Differential privacy-preserving perturbation fingerprint vector .

[0046] The aforementioned differential privacy algorithm first calculates the global sensitivity of the isotopic fingerprint vector:

[0047] in and This is the fingerprint vector of adjacent datasets. Then, the isotopic fingerprint vector for each farm is... Add Laplace noise to each component:

[0048] in, The scale parameter is represented as The noise value is obtained by randomly sampling from the Laplace distribution.

[0049] Furthermore, the aforementioned Laplace distribution The probability density function is:

[0050] in The specific implementation method for random sampling of an exponential function is as follows: First, generate samples that follow a uniform distribution. random numbers Then, the Laplace random variable is calculated using the inverse transform sampling method. ,in For a sign function, when When the value is 1, The value is -1; the generated random variable Obey the scale parameter is The Laplace distribution is used as noise and added to the corresponding components of the isotopic fingerprint vector.

[0051] Furthermore, the generated perturbation fingerprint vector satisfy - Differential privacy definition, that is, for any two adjacent datasets and (Different by only one record), and any output subset ,satisfy:

[0052] in This is a differential privacy algorithm that ensures attackers cannot infer the original isotopic features from the perturbed data.

[0053] It should be noted that global sensitivity The range of variation in historical isotope data is determined by analyzing the range of variation, usually by taking twice the maximum range of variation in the isotope ratio.

[0054] Furthermore, the aforementioned privacy budget parameters The range of values ​​is The smaller one Value (e.g.) It offers stronger privacy protection but introduces more noise, and a larger... Value (e.g.) This approach offers weaker privacy protection while maintaining high data utility; in this embodiment, considering both privacy protection requirements and source resolution accuracy requirements, the preferred approach is... .

[0055] Furthermore, the aforementioned global sensitivity The theoretical basis for taking twice the maximum variation range of the isotope ratio is as follows: According to differential privacy theory, global sensitivity needs to cover the maximum difference between adjacent datasets. Considering that each component of the isotope fingerprint vector is measured using the L1 norm, and that in practical applications, when a single farm leaves or joins a dataset, the change in the fingerprint vector usually does not exceed the historical maximum variation range, a coefficient of 2 is taken as a safety redundancy to ensure sufficient privacy protection margin; specifically, for nitrogen isotope ratios... Its maximum variation range is usually The corresponding sensitivity is For phosphorus isotope ratios Its maximum variation range is usually The corresponding sensitivity is .

[0056] Step 300: Aggregate the perturbation fingerprint vectors of each farm using a secure multi-party computation protocol, and calculate the regional isotopic fingerprint covariance matrix in the ciphertext domain using homomorphic encryption technology.

[0057] Each farm uses the Paillier homomorphic encryption scheme to encrypt the perturbation fingerprint vector:

[0058] in, It is the product of two large prime numbers. yes generator, It is a random number.

[0059] Furthermore, the parameter generation method for the aforementioned Paillier encryption scheme is as follows: first, select two large prime numbers. and Their bit lengths are all 1024 bits, satisfying ,calculate Generator The selection method is as follows The selection of generators satisfies the generator condition and simplifies the calculation; random numbers From the interval Uniformly and randomly selected from the middle, and satisfying To ensure encryption security; the public key is The private key is ,in This represents the least common multiple.

[0060] Furthermore, leveraging the homomorphic property of Paillier encryption, the elements of the covariance matrix are computed in the ciphertext domain:

[0061] in To determine the number of participating farms, Index for farms, Represents the homomorphic multiplication operation in the ciphertext field.

[0062] Furthermore, the aforementioned homomorphic multiplication operation in the ciphertext field The specific implementation method is as follows: for two ciphertexts and The result of its homomorphic multiplication is Homomorphic multiplication operations satisfy ,in This represents the decryption operation; using the properties of homomorphic addition, the elements of the covariance matrix are calculated. At that time, among them To determine the number of participating farms, To index the farms, first, for each farm... Calculate the product Then, the contributions of all farms are aggregated through continuous homomorphic addition of ciphertext fields (i.e., ciphertext multiplication), and finally divided by after decryption. The covariance value is obtained.

[0063] Furthermore, through a secure multi-party computation protocol, all parties collaborate to complete the computation of the encrypted domain, and the regional isotopic fingerprint covariance matrix is ​​ultimately obtained by a trusted third party or through a threshold decryption method. .

[0064] Furthermore, the specific implementation method of the aforementioned threshold decryption method is as follows: The private key... Divided into Individual shares ,in This refers to the share of private keys held by each participating party. The total number of participants is distributed to [the relevant parties]. Each participant satisfies any Individual shares ( The private key can be reconstructed using the Shamir secret sharing scheme.

[0065] Specifically, construct polynomial of degree , where the coefficient The coefficients are polynomials and randomly selected; calculate the share. , ,in Index for participants; During decryption, any Each participant contributes its share, and the private key is recovered through Lagrange interpolation:

[0066] in For the set of participants providing shares, and For participating parties index, The recovered private key is used to decrypt the ciphertext covariance matrix. The decryption formula is as follows: ,in .

[0067] In this embodiment of the application, in order to improve computational efficiency, the following steps are also included in addition to step 300: Step 301: Perform block processing on the high-dimensional fingerprint vector, decomposing the covariance matrix calculation into multiple sub-matrix calculation tasks.

[0068] Furthermore, the specific implementation method of the block processing in step 301 above is as follows: Let the fingerprint vector dimension be... The block size is Segment the fingerprint vector into The nth subvector, the nth The number of subvectors is ,in Subvector index; covariance matrix Accordingly, it was divided into Submatrix blocks Each submatrix for A dimensional matrix, where and The subvector index represents the first... The subvector and the first The covariance between subvectors; in practical applications, the block size. The value is usually taken as to To balance computational complexity and parallelism.

[0069] Step 302: Employ a parallel computing architecture to simultaneously process the ciphertext field operations of multiple submatrices.

[0070] Step 400: Combine the joint fingerprint feature space with the hydrological transport model to construct an enhanced source-sink transport matrix, and use a federated Bayesian mixture model to decompose the contribution ratio of different sources of nutrients at each observation point.

[0071] Based on the covariance matrix of regional isotopic fingerprints Construct a joint fingerprint feature space and generate a source-sink transport matrix by combining it with a hydrological transport model. .

[0072] Furthermore, the aforementioned method for constructing the joint fingerprint feature space is as follows: Covariance matrix of regional isotopic fingerprints Perform eigenvalue decomposition superscript Indicates transpose. It is an eigenvalue diagonal matrix. For each eigenvalue, The eigenvector matrix; select the first... The eigenvectors corresponding to the largest eigenvalues ​​form a dimension-reduced matrix. ,in The dimension after dimensionality reduction. The choice satisfies the cumulative variance contribution rate ,in For feature value index, The total number of eigenvalues; projecting the original fingerprint vector into a reduced-dimensional space yields the joint fingerprint features. The joint fingerprint feature space retains the main variation information of fingerprint data from each farm, while reducing the data dimensionality.

[0073] Furthermore, the aforementioned eigenvalue decomposition requires the covariance matrix to be... For a symmetric positive definite matrix, the eigenvalues ​​satisfy... Descending order; dimensionality after dimensionality reduction The range of values ​​is The lower boundary Ensure that at least two principal components are retained to capture the fundamental variation patterns of the data, upper bound. To avoid increased computational complexity due to excessively high dimensionality; eigenvector matrix Satisfying orthogonality constraints ,in It is the identity matrix, ensuring that the inner product remains invariant during the projection transformation.

[0074] Furthermore, we first calculate the original transmission coefficients:

[0075] in, Indicates from the source To the observation point The original transmission coefficient, For spatial distance, The average flow velocity, Let be the diffusion coefficient. The original transport coefficient has the following dimensions: .

[0076] Furthermore, the aforementioned spatial distance The calculation method is as follows: assuming the pollution source The geographic coordinates are Observation point The geographic coordinates are Using the WGS84 coordinate system, the latitude and longitude coordinates are first converted to projected coordinates (using UTM projection), and then the Euclidean distance is calculated. The unit is meters; for areas with a large span (over 10 kilometers), considering the influence of the Earth's curvature, the Haversine formula is used to calculate the great circle distance:

[0077] in km is the Earth's radius. Latitude Longitude.

[0078] To eliminate the influence of dimensions on subsequent calculations, the original transmission coefficients are normalized to obtain a dimensionless transmission weight matrix:

[0079] in For pollution source indexing, This represents the total number of pollution sources. (Normalized) Dimensionless numerical values ​​represent pollution sources. For observation points The relative transmission contribution satisfies ,in For pollution source indexing.

[0080] Constructing a federated Bayesian mixture model:

[0081] in, For observation point The nutrient concentration vector, For source Contribution weight, and Sources The mean and covariance of isotopic fingerprints.

[0082] The input to the aforementioned federated Bayesian mixture model is the nutrient concentration vector at each observation point: And the corresponding isotopes form a vector superscript This represents the transpose, and the output is the contribution weight of each pollution source. ,in The contribution weight values ​​of each pollution source, and the isotopic fingerprint parameters of each source. ,in For pollution source indexing, This represents the total number of pollution sources.

[0083] The aforementioned federated Bayesian mixture model was trained using the expectation-maximization algorithm.

[0084] The input to the aforementioned expectation-maximization algorithm is the observed dataset. (Including nutrient concentration vectors and isotopic composition vectors at each observation point) and initial parameter estimates. The output is the optimized model parameters. .

[0085] Furthermore, the aforementioned initial parameter estimates The method for obtaining it is: contribution weight Initialize to uniform distribution This indicates that the initial contributions from each pollution source are equal; isotopic fingerprint mean Initialization is performed using the K-means clustering algorithm; specifically, a vector is formed from the isotopes of all observation points. conduct Clustering, where Let the isotope composition vectors be defined for each observation point. For the index of observation points, For the total number of observation points, the first... Cluster centers as sources initial mean Covariance matrix Initialize to the identity matrix multiplied by a constant ,in The total variance of the observed data. The mean of the observed data, It is an identity matrix.

[0086] The aforementioned expectation-maximization algorithm consists of two alternating units: the E-step (expectation step) and the M-step (maximization step). In the E-step, based on the current parameter estimates... Calculate the posterior probability:

[0087] in For pollution source indexing, The total number of pollution sources, Indicates the observation point The nutrients come from pollution sources The posterior probability.

[0088] In the M-step, the model parameters are updated based on the posterior probabilities calculated in the E-step:

[0089]

[0090]

[0091] in For the index of observation points, The total number of observation points, indicated by the superscript. Indicates transpose. This represents the number of iterations.

[0092] Furthermore, the aforementioned contribution weights The following constraints need to be met: Normalization constraints Nonnegativity constraints , ,in For pollution source indexing, The total number of pollution sources; during the M-step update process, the normalization constraint is updated through the formula. Automatic satisfaction, because of the posterior probability satisfy Therefore, there is ,in For the index of observation points, The total number of observation points; the nonnegativity constraint is determined by the nonnegativity of the posterior probability. ensure.

[0093] Furthermore, the aforementioned covariance matrix It needs to satisfy the positive definiteness constraint, that is, for any non-zero vector ,satisfy ; During the M-step update process, to ensure the positive definiteness of the covariance matrix, the updated covariance matrix is... Add regularization terms The regularization coefficient , The identity matrix is ​​used; regularization ensures that the minimum eigenvalue of the covariance matrix is ​​not less than 1. This ensures the numerical stability and invertibility of the matrix.

[0094] The data transfer relationship in the aforementioned federated Bayesian mixture model is as follows: nutrient concentration and isotope data at each observation point first pass through the source-sink transfer matrix. The inverse mapping yields the equivalent concentration of the source space. Then, in the E-step, the posterior probability of each source is calculated. The posterior probability of each source is used to update the model parameters in the M-step. The E-step and M-step are executed iteratively until convergence.

[0095] The aforementioned federated Bayesian mixture model uses the negative log-likelihood function as the loss function:

[0096] in For the index of observation points, The total number of observation points. For pollution source indexing, This represents the total number of pollution sources.

[0097] Furthermore, the aforementioned loss function The time dimension is represented as follows: when the training data contains observations at multiple time points, the different time points are represented as follows: The observation dataset is denoted as ,in For data from each observation point at each time point, For the index of observation points, For time points The number of observation points; the extended spatiotemporal loss function is expressed as ,in For time point indexing, The total number of time points. For time points The weighting coefficients satisfy the normalization condition. and Weighting coefficients The method for determining the value is as follows: For equally spaced sampling data, uniform weighting is used. For scenarios that require emphasizing recent data, an exponentially decaying weighting method is used. ,in Represents an exponential function. For time point index, attenuation coefficient The range of values ​​is Larger Values ​​assign higher weight to recent data; pollution source contribution weights It can be time-varying at different points in time, reflecting the dynamic changes in the contribution of pollution sources.

[0098] Furthermore, the training process minimizes the loss function using the EM algorithm, and the convergence threshold is reached when the change in the loss function between two consecutive iterations is less than the convergence threshold. Stop training when the maximum number of iterations is reached. The event will end at that time.

[0099] It should be noted that, This indicates the total number of pollution sources, including farms, land-based sources, and offshore sources.

[0100] Furthermore, the total number of the aforementioned pollution sources The range of values ​​is The lower boundary Ensure that at least the main pollution source types (such as farms, land-based sources, and offshore inputs) can be distinguished, and the upper limit... To avoid instability in model parameter estimation due to an excessive number of pollution sources; total number of pollution sources Total number of observation points The relationship needs to satisfy To ensure sufficient observational data to support reliable estimation of model parameters; when At that time, it is necessary to reduce the number of pollution sources or merge similar pollution source categories.

[0101] Furthermore, the aforementioned average flow velocity The range of values ​​is determined based on the hydrological conditions of the nearshore aquaculture area. For nearshore waters dominated by tides, the range is [range missing]. m / s, where the smaller value corresponds to the period of weak tidal current and the larger value corresponds to the period of strong tidal current; in practical applications, real-time flow velocity data is continuously monitored and obtained by an acoustic Doppler current profiler (ADCP) deployed on site, and the average value within the monitoring period is taken as the model input parameter.

[0102] Furthermore, the aforementioned diffusion coefficient The value of is determined based on turbulent diffusion theory, and its calculation formula is: , in The dimensionless diffusion coefficient has a range of values ​​of . , Characteristic water depth (unit: m). For nearshore aquaculture areas, characteristic water depth... Usually in Within the range of m; considering the above parameters comprehensively, the diffusion coefficient The typical range of values ​​is m² / s.

[0103] Furthermore, the aforementioned convergence threshold The selection criteria are as follows: the convergence threshold corresponds to a relative change in the loss function of less than 0.0001%, at which point the model parameters have reached a numerically stable state, and further iterations have negligible improvement on parameter estimation; the maximum number of iterations... The setting is based on actual training experience. Under normal circumstances, the EM algorithm usually converges within 100-200 iterations. Setting 500 as the upper limit is to prevent infinite loops in extreme cases, while ensuring the robustness of the algorithm.

[0104] Step 500: Based on the trained pollution source identification model, input the real-time monitored nutrient concentration and isotope data, and output the real-time contribution ratio and confidence interval of each pollution source.

[0105] The real-time collected nutrient concentration and isotopic composition data of the water samples were subjected to the same Z-score standardization process as in step 100. The mean and standard deviation parameters saved during the training phase were used to standardize the real-time data to obtain the standardized data. and .

[0106] Standardized data and Input the data into the model trained in step 400 to calculate the posterior probability of each pollution source:

[0107] in, This indicates that the pollution originated from the first One source, For pollution source indexing, The total number of pollution sources, In order to be at the source of pollution Likelihood probability of observed data under given conditions.

[0108] The aforementioned likelihood probability The calculation is based on the multivariate normal distribution, and its formula is:

[0109] in, To represent an exponential function, superscript Indicates transpose. For the combined observation vector, For vector dimensions, As a source of pollution The composite mean vector, To synthesize the covariance matrix, This represents the determinant of the covariance matrix. The likelihood probability formula is used to calculate the probability of a given pollution source. Under the given parameter conditions, the probability density of the current nutrient concentration and isotopic composition was observed.

[0110] Furthermore, the aforementioned vector dimension The calculation method is as follows: nutrient concentration vector The dimension is 5 (including ammonia nitrogen, nitrate nitrogen, total nitrogen, orthophosphate, and total phosphorus), and the isotopic composition vector The dimension is 2 (including and Therefore, the combined observation vector The dimension is ; Comprehensive covariance matrix for A 1D symmetric positive definite matrix, its determinant It must be greater than zero to ensure the validity of the probability density function; when When determining that the covariance matrix is ​​close to singular, a regularization term needs to be added. To ensure numerical stability.

[0111] The contribution ratio of each pollution source is calculated based on posterior probability. .

[0112] The input to the aforementioned Bayesian inference algorithm is the trained model parameters. Real-time monitoring data and confidence level (Typically 95%), the output is the confidence interval of the contribution ratio of each pollution source. .

[0113] The aforementioned Bayesian inference algorithm uses the Markov Chain Monte Carlo (MCMC) method to sample from the posterior distribution. Specifically, it utilizes the Metropolis-Hastings algorithm to generate... Parameter samples Each sample corresponds to a set of contribution ratios. .

[0114] Furthermore, the specific implementation steps of the aforementioned Metropolis-Hastings algorithm are as follows: Step 1: Initialization: Setting initial parameter samples That is, using the optimal parameters obtained from training as the starting point; Step 2: Proposal Generation: In the... In the next iteration, from the proposal distribution Mid-sampling candidate parameters The proposed distribution adopts a multivariate normal distribution. The proposed covariance matrix , The empirical covariance matrix of the posterior distribution; Step 3: Calculate the probability of acceptance: Calculate the probability of acceptance:

[0115] Because a symmetric proposal distribution is used, the acceptance probability simplifies to:

[0116] Step 4: Acceptance Judgment: Generate uniformly random numbers ,like Then accept candidate parameters Otherwise, keep the current parameters. ; Step 5: Repeat steps 2 through 4 until generated. One sample.

[0117] Furthermore, for each pollution source , The sampled contribution ratio sequence Sort in ascending order, and take the [number]th [item]. Quantiles as a confidence lower bound Take the first Quantiles as the upper confidence bound , thus obtaining the confidence interval .

[0118] Furthermore, the aforementioned MCMC sampling quantity The value of needs to balance computational efficiency and statistical reliability. According to the convergence theory of the Monte Carlo method, the sample size should satisfy . To ensure the stability of the posterior distribution estimation; in this embodiment, a setting is provided. , including the former One sample was used as the burn-in period to eliminate the influence of initial values, and then... A sample size is used for confidence interval calculation; with the above sampling size settings, the Monte Carlo standard error of the contribution ratio estimate is typically less than [value missing]. .

[0119] Furthermore, the aforementioned confidence level Besides the usual value of 95% (i.e.) In addition, when more stringent statistical assurance is required, 99% can be used (i.e., In this case, the confidence interval is wider but the reliability is higher; in applications requiring rapid decision-making, 90% can also be used (i.e., At this point, the confidence interval is narrower but there is higher uncertainty; this implementation prefers a 95% confidence level to balance statistical reliability and interval accuracy.

[0120] In this embodiment of the application, in order to improve the response speed of real-time detection, the following steps are further included in addition to step 500: Step 501: Cache historical detection results and build a fast query index.

[0121] Furthermore, the fast query index in step 501 above is implemented using a KD-tree (K-dimensional tree) data structure. The specific construction method is as follows: the input data vector of historical detection... As nodes in a KD-tree, The number of historical records is used as the reference. The recursive process of constructing a KD-tree is as follows: select the dimension with the largest variance as the splitting dimension, split the dataset according to the median of that dimension, and recursively perform the above two steps on the left and right subsets until the number of data points contained in the leaf nodes is less than the threshold. The established KD-tree supports... Perform nearest neighbor lookup within time complexity, and transform the new data vector. The system compares the distance to historical data stored in the KD-tree to quickly locate the most similar historical data; simultaneously, it maintains a hash table to store the detection results (contribution ratio and confidence interval) corresponding to each historical input vector, thus achieving... Time complexity of result query.

[0122] Step 502: When the similarity between the new data and the cached data exceeds the threshold, return the cached result directly.

[0123] Furthermore, the similarity in step 502 above uses Euclidean distance as a metric; The calculation formula is ,in For dimensional indexing, For data dimensions, For the calculated distance value, For the standardized vector of the newly collected data, A standardized vector for cached data; similarity threshold. The range of values ​​is ,when The system determines if a result is similar and returns a cached result. In this implementation, considering both response speed and detection accuracy, the preferred method is... The similarity threshold corresponds to the average deviation between new data and cached data in each dimension not exceeding 1 standard deviation.

[0124] A nearshore aquaculture area in a county of a certain city in a certain province includes three sea cucumber farms (farm A, farm B, and farm C), one land-based sewage outlet, and an offshore pollution source, all of which affect the water quality of the area. Each farm needs to identify the contribution ratio of each pollution source to the aquaculture area in order to develop a scientific pollution control strategy. However, due to commercial competition, the farms are unwilling to share raw isotope data containing sensitive information such as stocking density and feed formulation.

[0125] Each aquaculture farm collected water samples within its aquaculture area to determine nutrient concentrations and isotope ratios. Taking aquaculture farm A as an example, the multi-depth water sample data collected during a certain observation period in July 2024 are shown in Table 1.

[0126] Table 1. Nutrient concentration and isotopic composition data for farm A:

[0127] After Z-score normalization of the acquired data, Morlet wavelet transform (center frequency) was applied. Scale number of layers Extracting the spatiotemporal variation characteristics of isotopes. Using nitrogen isotopes... For example, the statistical characteristics of wavelet coefficients calculated at 6 scales are shown in Table 2.

[0128] Table 2. Statistical characteristics of nitrogen isotope wavelet coefficients of farm A:

[0129] By combining the wavelet features of phosphorus isotopes, a local pollution source isotopic fingerprint vector for farm A is constructed. Feature Dimension .

[0130] Figure 2 The distribution of five nutrient concentrations at different depths (surface, middle, and bottom) of farm A is shown.

[0131] Figure 3 The distribution of energy features extracted by wavelet transform at different scales is shown.

[0132] Set privacy budget parameters Global sensitivity (Based on the maximum variation of nitrogen isotopes) (twice as high). Laplace noise is added to each component of the isotopic fingerprint vector of farm A. The scale parameter of the Laplace distribution... Table 3 shows the perturbation process of some characteristic components.

[0133] Table 3. Differential privacy perturbation process (partial features) of farm A:

[0134] Farm A obtains the perturbation fingerprint vector through the above perturbation process. Farms B and C each performed the same differential privacy perturbation operation on their respective isotopic fingerprint vectors to obtain perturbed fingerprint vectors. and .

[0135] Figure 4 The changes in eigenvalues ​​before and after adding Laplacian noise to isotopic fingerprint vectors using a differential privacy algorithm are demonstrated.

[0136] Each farm used the Paillier homomorphic encryption scheme to encrypt the perturbation fingerprint vector (key length 2048 bits) and uploaded it to the collaborative computing platform. The platform calculated the regional isotopic fingerprint covariance matrix in the ciphertext domain. Each element is decrypted using a threshold method (set). One participant, threshold Obtain the plaintext covariance matrix.

[0137] Nutrient concentration and isotopic composition data from 18 observation points within the region in July 2024 were collected as a training dataset. This dataset was then combined with hydrological transport model parameters (mean flow velocity). m / s, diffusion coefficient (m² / s), construct the source-sink transport matrix. Train using a federated Bayesian mixture model, total number of pollution sources. (Aquaculture farms A, B, and C; land-based sewage outlets; offshore input), total number of observation points After 168 iterations, it converged, and the loss function changed from its initial value. Down to The contribution weights of each pollution source and the mean parameters of the isotopic fingerprint obtained from the training are shown in Table 4.

[0138] Table 4 shows the contribution weights of each pollution source and the mean parameters of the isotopic fingerprint obtained during training:

[0139] Figure 5 The contribution weights of each pollution source (bar chart) and the mean isotopic fingerprint (line chart) obtained from training the federated Bayesian mixture model are displayed.

[0140] Figure 7 The convergence curve of the loss function of the Expectation Maximization (EM) algorithm is shown.

[0141] In August 2024, new water sample data were collected from regional monitoring points, and the nutrient concentration and isotopic composition are shown in Table 5.

[0142] Table 5 Water quality data from real-time monitoring points:

[0143] The monitoring data is input into the trained model, and then sampled using MCMC (Monitored Multi-Channel Model). The contribution ratio and 95% confidence interval of each pollution source were calculated from 1000 samples during the burn-in period. The results are shown in Table 6.

[0144] Table 6. Results of pollution source contribution ratio identification at monitoring point M1:

[0145] Figure 6 The contribution ratio of each pollution source at monitoring point M1 and its 95% confidence interval are shown.

[0146] According to the identification results, the nutrient pollution at monitoring point M1 mainly originated from aquaculture farm A (contributing 32%) and aquaculture farm B (contributing 29%), with land-based sewage outlets contributing 18%, aquaculture farm C contributing 16%, and offshore input contributing a relatively small amount (5%). These results provide a scientific basis for coordinated regional pollution control, and each aquaculture farm can formulate corresponding emission reduction measures based on its own contribution ratio.

[0147] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multimodal and deep CNN-based intelligent edge detection method for nearshore aquaculture, characterized in that, Includes the following steps: Nutrient concentration data and nitrogen and phosphorus isotope composition data of water samples from various aquaculture farms at multiple depths were obtained. Wavelet transform was used to extract the spatiotemporal variation features of isotopes and generate local pollution source isotope fingerprint vectors. Based on the differential privacy algorithm, Laplace noise is added to the isotopic fingerprint vectors of each farm to calculate the perturbed fingerprint vectors that satisfy ε-differential privacy. The perturbation fingerprint vectors of each farm are aggregated using a secure multi-party computation protocol, and the regional isotopic fingerprint covariance matrix is ​​calculated in the ciphertext domain using homomorphic encryption technology. By combining the joint fingerprint feature space with the hydrological transport model, an enhanced source-sink transport matrix is ​​constructed, and the contribution ratio of different sources of nutrients in each observation point is decomposed using a federated Bayesian mixture model. Based on the trained pollution source identification model, real-time monitored nutrient concentration and isotope data are input, and the real-time contribution ratio and confidence interval of each pollution source are output.

2. The multimodal and deep CNN nearshore aquaculture intelligent edge detection method according to claim 1, characterized in that, The extraction of isotope spatiotemporal variation features using wavelet transform includes: The Morlet wavelet was used as the mother wavelet function to perform continuous wavelet transform on the standardized isotope time series data. The transformation coefficients are calculated at different scales and different translation positions to obtain the multi-scale decomposition results in the time-frequency domain. Statistical features are extracted from multi-scale wavelet coefficients, including energy distribution, peak position, and fluctuation amplitude at each scale. The extracted features are vectorized to construct the local pollution source isotopic fingerprint vector.

3. The multimodal and deep CNN nearshore aquaculture intelligent edge detection method according to claim 1, characterized in that, The process of adding Laplacian noise to isotopic fingerprint vectors based on differential privacy algorithms includes: Calculate the global sensitivity of the isotopic fingerprint vector by taking twice the maximum change in the isotopic ratio. For each component of the isotopic fingerprint vector of each farm, noise values ​​are randomly sampled from a Laplace distribution with a scale parameter equal to the global sensitivity divided by the privacy budget. The sampled noise values ​​are added to the corresponding fingerprint vector components to generate a perturbed fingerprint vector that satisfies ε-differential privacy protection.

4. The multimodal and deep CNN nearshore aquaculture intelligent edge detection method according to claim 1, characterized in that, The process of calculating the regional isotopic fingerprint covariance matrix in the ciphertext domain using homomorphic encryption technology includes: Each farm uses the Paillier homomorphic encryption scheme to encrypt the perturbation fingerprint vector; By utilizing the homomorphic property of Paillier encryption, the elements of the covariance matrix are calculated in the ciphertext field through homomorphic multiplication. Through a secure multi-party computation protocol, all parties collaborate to complete the computation of the ciphertext field; The regional isotopic fingerprint covariance matrix is ​​obtained by using a threshold decryption method.

5. The multimodal and deep CNN nearshore aquaculture intelligent edge detection method according to claim 1, characterized in that, The process of constructing the enhanced source-sink transfer matrix includes: The original transport coefficient from each pollution source to the observation point is calculated based on spatial distance, average flow velocity, and diffusion coefficient. The original transmission coefficients are normalized to obtain a dimensionless transmission weight matrix. Each element in the transmission weight matrix represents the relative transmission contribution of the corresponding pollution source to the observation point, and the sum of the transmission weights of all pollution sources corresponding to each observation point is 1.

6. The multimodal and deep CNN nearshore aquaculture intelligent edge detection method according to claim 1, characterized in that, The training process of the federated Bayesian mixture model includes: A probabilistic mixture model based on a multivariate normal distribution is constructed, with model parameters including the contribution weights of each pollution source, the mean of the isotopic fingerprint, and the covariance. The model training uses the expectation-maximization algorithm, which consists of two alternating units: the E-step for calculating the posterior probability and the M-step for updating the model parameters. The negative log-likelihood function is used as the loss function, and training stops when the change in the loss function between two consecutive iterations is less than the convergence threshold.

7. The multimodal and deep CNN intelligent edge detection method for nearshore aquaculture according to claim 1, characterized in that, The process of outputting the real-time contribution ratio and confidence interval of each pollution source includes: The real-time monitoring data is input into the trained model, and the posterior probability of each pollution source is calculated as the contribution ratio. The Markov chain Monte Carlo method is used to sample from the posterior distribution; Parameter sample sequences are generated using the Metropolis-Hastings algorithm; The confidence interval for the contribution ratio of each pollution source is calculated based on the sampling results.

8. The multimodal and deep CNN intelligent edge detection method for nearshore aquaculture according to any one of claims 1 to 7, characterized in that, Also includes: Vertical interpolation is performed on multi-depth sampled data to generate isotope distributions of continuous depth profiles; Calculate the coefficient of variation of isotope ratios at each depth layer to identify isotope layering structures; Based on the vertical gradient characteristics of isotopes, the sedimentation or floating trend of pollutants can be determined.

9. The multimodal and deep CNN intelligent edge detection method for nearshore aquaculture according to any one of claims 1 to 7, characterized in that, Also includes: Historical detection results are cached, and a fast query index based on a KD tree is established; When the Euclidean distance between the new data and the cached data is less than the similarity threshold, the cached result is returned directly.

10. A multimodal and deep CNN intelligent edge detection system for nearshore aquaculture, used to execute the multimodal and deep CNN intelligent edge detection method for nearshore aquaculture as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-depth water sample data from various aquaculture farms; The privacy protection module is used to perform differential privacy perturbation on the isotope fingerprint vector; The secure computation module is used to aggregate and compute the covariance matrix in the ciphertext domain; The model training module is used to train federated Bayesian mixture models; The real-time inference module is used to output the contribution ratio of pollution sources and the confidence interval.

Images