A multi-modal and deep CNN offshore aquaculture intelligent detection edge method and system

Abstract

The application relates to the technical field of offshore aquaculture intelligent monitoring, and discloses a multi-modal and deep CNN offshore aquaculture intelligent detection edge method and system, wherein the method comprises the following steps: acquiring multi-depth water sample nutrient salt concentration data and nitrogen and phosphorus isotope composition data of each local aquaculture farm, and generating local pollution source isotope fingerprint vectors; adding Laplace noise to the isotope fingerprint vectors of each aquaculture farm based on a differential privacy algorithm, and calculating disturbance fingerprint vectors satisfying epsilon-differential privacy; combining a joint fingerprint characteristic space and a hydrological transmission model, and decomposing the contribution proportion of different sources in the nutrient salt of each observation point by using a federal Bayesian mixture model; and outputting the real-time contribution proportion and confidence interval of each pollution source based on a trained pollution source identification model. The application realizes accurate pollution source identification and contribution proportion quantification under the premise of protecting the privacy of each party.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore aquaculture intelligent monitoring, more particularly, it relates to a multi-modal and deep CNN offshore aquaculture intelligent detection edge method and system. BACKGROUND

[0002] In the offshore aquaculture area, multiple farms share the same water environment, and the pollutants generated by each farm and the surrounding land pollution, far sea input and other pollution sources superimpose each other, forming a complex compound pollution situation. Accurate identification of the contribution proportion of each pollution source is crucial for formulating scientific pollution prevention and control strategies. The traditional pollution source analysis method relies on isotope fingerprint technology, which traces the source of pollutants by analyzing the nitrogen and phosphorus isotope composition in the water body. However, in practical application, there are the following technical problems: the isotope samples collected by a single farm are insufficient in terms of spatial and temporal coverage, making it difficult to build a complete regional pollution source isotope fingerprint library; when each farm independently analyzes the source, there is a lack of characteristic data of neighboring pollution sources, resulting in large deviation in source contribution proportion calculation; at the same time, the original isotope data contains commercial sensitive information such as aquaculture density and feed formula, and each farm is reluctant to share data due to commercial competition, forming a data island, which seriously restricts the effectiveness of regional pollution collaborative governance. SUMMARY

[0003] The present application provides a multi-modal and deep CNN offshore aquaculture intelligent detection edge method and system, which solves the problem of data sharing reluctance due to fear of commercial information leakage in related technologies. The deviation in source analysis caused by insufficient spatial and temporal coverage of single farm data realizes the technical problem of accurate pollution source identification and contribution proportion quantification under the premise of protecting the privacy of each party.

[0004] The present application provides a multi-modal and deep CNN offshore aquaculture intelligent detection edge method, comprising:

[0005] Obtain the multi-depth water sample nutrient salt concentration data and nitrogen and phosphorus isotope composition data of each farm, extract the isotope spatiotemporal variation characteristics using wavelet transform, and generate local pollution source isotope fingerprint vectors;

[0006] Add Laplace noise to the isotope fingerprint vectors of each farm based on the differential privacy algorithm, and calculate the perturbed fingerprint vectors that satisfy the epsilon-differential privacy;

[0007] Aggregate the perturbed fingerprint vectors of each farm using a secure multi-party computation protocol, and calculate the regional isotope fingerprint covariance matrix in the ciphertext domain through homomorphic encryption technology;

[0008] The joint fingerprint feature space is combined with the hydrological transport model to construct an enhanced source-sink transport matrix, and a federal Bayesian mixture model is used to decompose the contribution proportion of different sources of nutrients at each observation point.

[0009] Based on the trained pollution source identification model, the real-time monitoring of nutrient salt concentration and isotope data are input, and the real-time contribution proportion and confidence interval of each pollution source are output.

[0010] Further, the wavelet transform is used to extract the spatial and temporal variation characteristics of the isotope, which includes:

[0011] The Morlet wavelet is used as the mother wavelet function for continuous wavelet transform of the normalized isotope time series data;

[0012] The transform coefficients are calculated at different scales and different translation positions to obtain the multi-scale decomposition results in the time-frequency domain;

[0013] Statistical feature extraction is performed on the multi-scale wavelet coefficients, including energy distribution, peak position and fluctuation amplitude at each scale;

[0014] The extracted feature vectors are vectorized to construct local pollution source isotope fingerprint vectors.

[0015] Further, the process of adding Laplace noise to the isotope fingerprint vector based on differential privacy algorithm includes:

[0016] The global sensitivity of the isotope fingerprint vector is calculated, and the maximum change amplitude of the isotope ratio is taken as 2 times;

[0017] For each component of the isotope fingerprint vector of each farm, a noise value is randomly sampled from a Laplace distribution with a scale parameter equal to the global sensitivity divided by the privacy budget;

[0018] The sampled noise value is added to the corresponding fingerprint vector component to generate a perturbed fingerprint vector that satisfies the ε-differential privacy protection.

[0019] Further, the process of calculating the regional isotope fingerprint covariance matrix in the ciphertext domain through homomorphic encryption technology includes:

[0020] Each farm uses the Paillier homomorphic encryption scheme to encrypt the perturbed fingerprint vector;

[0021] Using the homomorphic property of Paillier encryption, the covariance matrix elements are calculated in the ciphertext domain through homomorphic multiplication operation;

[0022] Through a secure multi-party computation protocol, all parties cooperatively complete the ciphertext domain calculation;

[0023] The regional isotope fingerprint covariance matrix is obtained by threshold decryption.

[0024] Further, the process of constructing the enhanced source-sink transport matrix comprises:

[0025] Calculating the original transport coefficient from each pollution source to the observation point based on spatial distance, average flow velocity and diffusion coefficient;

[0026] Normalizing the original transport coefficient to obtain a dimensionless transport weight matrix;

[0027] Each element in the transport weight matrix represents the relative transport contribution of the corresponding pollution source to the observation point, and the sum of the transport weights of all pollution sources corresponding to each observation point is 1.

[0028] Further, the training process of the federal Bayesian mixture model comprises:

[0029] Constructing a probability mixture model based on multivariate normal distribution, and the model parameters include the contribution weight of each pollution source, the mean and covariance of the isotope fingerprint;

[0030] Using the expectation maximization algorithm for model training, including two alternately executed units of E-step calculating posterior probability and M-step updating model parameters;

[0031] Using negative log-likelihood function as loss function, and stopping training when the change of loss function of adjacent two iterations is less than convergence threshold.

[0032] Further, the process of outputting the real-time contribution proportion and confidence interval of each pollution source comprises:

[0033] Inputting real-time monitoring data into the trained model to calculate the posterior probability of each pollution source as the contribution proportion;

[0034] Using Markov chain Monte Carlo method to sample from the posterior distribution;

[0035] Generating a sequence of parameter samples by Metropolis-Hastings algorithm;

[0036] Based on the sampling results, calculating the confidence interval of the contribution proportion of each pollution source.

[0037] Further, it further comprises:

[0038] Vertically interpolating the multi-depth sampling data to generate the isotope distribution of the continuous depth profile;

[0039] Calculating the coefficient of variation of isotope ratio of each depth layer to identify the isotope stratification structure;

[0040] Based on the vertical gradient characteristics of the isotope, judging the settlement or floating trend of the pollutant.

[0041] Furthermore, it also includes:

[0042] Historical detection results are cached, and a fast query index based on a KD tree is established;

[0043] When the Euclidean distance between the new data and the cached data is less than the similarity threshold, the cached result is returned directly.

[0044] A second aspect of the present invention provides a multimodal and deep CNN-based intelligent edge detection system for nearshore aquaculture, comprising:

[0045] The data acquisition module is used to acquire multi-depth water sample data from various aquaculture farms;

[0046] The privacy protection module is used to perform differential privacy perturbation on the isotope fingerprint vector;

[0047] The secure computation module is used to aggregate and compute the covariance matrix in the ciphertext domain;

[0048] The model training module is used to train federated Bayesian mixture models;

[0049] The real-time inference module is used to output the contribution ratio of pollution sources and the confidence interval.

[0050] 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

[0051] Figure 1 This is a flowchart of a multimodal and deep CNN intelligent edge detection method for nearshore aquaculture according to the present invention;

[0052] Figure 2 This is a bar chart of nutrient concentration distribution in multi-depth water samples according to the present invention;

[0053] Figure 3 This is a multi-scale energy distribution piecewise linear plot of nitrogen isotope wavelet coefficients according to the present invention;

[0054] Figure 4 This is a scatter plot comparing the eigenvalues ​​before and after the differential privacy perturbation of this invention;

[0055] Figure 5 is the mixed graph of pollution source contribution weight and isotope fingerprint characteristics of the present application;

[0056] Figure 6 is the bar chart of real-time pollution source contribution proportion identification result (including confidence interval) of the present application;

[0057] Figure 7 is the broken line chart of EM algorithm training convergence process of the present application. DETAILED DESCRIPTION

[0058] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate. Also, it should be understood that some features described with respect to one example can be combined in other examples.

[0059] At least one embodiment of the present application discloses a multi-modal and deep CNN offshore aquaculture intelligent detection edge method, as shown in Figure 1 The method comprises the following steps:

[0060] Step 100: Obtain the multi-depth water sample nutrient salt concentration data and nitrogen and phosphorus isotope composition data of each local aquaculture farm, extract the isotope temporal and spatial variation characteristics by using wavelet transform, and generate the local pollution source isotope fingerprint vector.

[0061] Specifically, each aquaculture farm sets multiple sampling points in its aquaculture area, and water samples are collected at three depths of surface layer (0-0.5m), middle layer (2-3m), and bottom layer (0.5m from the seabed) at each sampling point. The collected water samples are subjected to nutrient salt concentration determination, including ammonia nitrogen (NH4 + -N), nitrate nitrogen (NO3 - -N), total nitrogen (TN), orthophosphate (PO4 3- -P), and total phosphorus (TP) concentration. At the same time, isotope mass spectrometer is used to determine nitrogen isotope ratio (δ 15 N) and phosphorus isotope ratio (δ 18 O-PO4).

[0062] Further, the effective measurement range of the aforementioned nutrient salt concentration is:

[0063] Ammonia nitrogen concentration mg / L, nitrate nitrogen concentration mg / L, total nitrogen concentration mg / L, orthophosphate concentration mg / L, total phosphorus concentration mg / L; when the measured value is lower than the lower limit, it is recorded as half of the detection limit value, and when the measured value is higher than the upper limit, it needs to be diluted and re-measured; the effective measurement range of the aforementioned isotope ratio is: nitrogen isotope ratio , phosphorus isotope ratio These ranges cover the typical isotope composition variation range in offshore aquaculture areas; when the measured value exceeds the effective range, it is determined as an abnormal value and is rejected, and resampling is determined.

[0064] The obtained nutrient salt concentration data and isotope ratio data are pre-processed. Since the nutrient salt concentration (unit: mg / L) and the isotope ratio (unit: ‰) have different dimensions and numerical ranges, in order to eliminate the influence of dimension difference on subsequent feature extraction, each type of data is respectively processed by Z-score standardization, and the data is converted to a standard distribution with a mean of 0 and a standard deviation of 1.

[0065] The wavelet transform is used to extract the spatial and temporal variation characteristics of the obtained standardized isotope time series data.

[0066] The input of the aforementioned wavelet transform algorithm is the isotope ratio time series signal:

[0067]

[0068] Wherein is the sequence of each time point of the sampling time, is the total number of sampling times, and the output is a multi-scale wavelet coefficient matrix .

[0069] Wherein is the sequence of each scale value of the selected scale, is the total number of scales.

[0070] Further, the value range of the aforementioned sampling time interval is determined according to the time scale of the monitoring target. For capturing the characteristics of the diurnal variation, the value is hours; for capturing the characteristics of seasonal variation, the value is days; the value range of the time series length is , wherein the smaller value corresponds to short-term monitoring (such as 1 month of hourly data, ), and the larger value corresponds to long-term monitoring (such as 2 years of weekly data, or daily data ); in order to ensure the effectiveness of the wavelet transform, the time series length should satisfy wherein is the maximum scale number, to ensure that the wavelet basis function of the maximum scale can completely cover the low-frequency components of the signal.

[0071] The calculation formula of the foregoing wavelet transform is:

[0072]

[0073] wherein, is the mother wavelet function, is the scale parameter, is the translation parameter, denotes the complex conjugate. By calculating the transform coefficients at different scales and different translation positions , the multi-scale decomposition result in the time-frequency domain is obtained.

[0074] Further, the foregoing wavelet transform is realized by discretization in actual calculation, and the continuous integral is converted into discrete summation: wherein, is the sampling time interval, is the length of the time series, is the time point index; the value of the translation parameter is , wherein, is the translation position index, so that the wavelet function is translated on the time axis with a sampling interval as the step size; for each scale (scale index) and each translation position , the wavelet coefficient is calculated, forming a scale-time two-dimensional coefficient matrix.

[0075] Further, statistical feature extraction is performed on the multi-scale wavelet coefficients, including the energy distribution, peak position and fluctuation amplitude of each scale; after the extracted features are vectorized, the local pollution source isotope fingerprint vector is constructed, wherein denotes the th farm, is the feature dimension.

[0076] Further, the calculation formula of the foregoing energy value of each scale is wherein, is the translation position index, is the length of the time series, and the sum of the modulus squares of all wavelet coefficients on the th scale is denoted, reflecting the total energy of the scale; the calculation formula of the peak position is , which denotes the th scale.the time position index corresponding to the maximum value of the wavelet coefficient modulus on the scale; the calculation formula of the fluctuation amplitude is , represents the variation range of the wavelet coefficient modulus on the scale ; the calculation formula of the cross-scale energy ratio is , , wherein is the scale index, is the total number of scales, represents the energy ratio value between adjacent scales, and reflects the distribution characteristics of energy among different scales.

[0077] It should be noted that the mother wavelet function is selected as Morlet wavelet, and its expression is:

[0078]

[0079] , wherein is the center frequency, which is usually taken as 6.

[0080] Further, the selection of the aforementioned scale sequence is based on the sampling period of the isotope time series data, and the specific value is , wherein is the sampling time interval (unit: day), , and the value range of the scale number is to cover the multi-time scale characteristics from the intra-day variation to the seasonal variation.

[0081] Further, the aforementioned feature dimension is determined by the number of wavelet coefficient statistical features, specifically including: the energy value of each scale (N ), the peak position (N ), the fluctuation amplitude (N ), and the cross-scale energy ratio (N ), so the feature dimension , when the scale number , the feature dimension .

[0082] Further, the theoretical basis for the value of the aforementioned center frequency being 6 is:

[0083] The value of the center frequency being 6 makes the Morlet wavelet have good localization characteristics in both time domain and frequency domain, and satisfies the admissibility condition , wherein is the Fourier transform of the wavelet function, and when , the time-frequency uncertainty product of the wavelet function reaches the optimal value close to the theoretical lower limit .

[0084] In the embodiments of the present application, in order to enhance the spatial and temporal representativeness of the isotope fingerprint, the following steps are further included on the basis of step 100:

[0085] Step 101: vertically interpolating the multi-depth sampling data to generate an isotope distribution of a continuous depth profile.

[0086] Further, the vertical interpolation in the aforementioned step 101 adopts a cubic spline interpolation method, and the specific implementation steps are as follows:

[0087] Suppose three sampling depths are (surface layer), (mid-layer), (bottom layer), and the corresponding isotope ratios are , , ;

[0088] Construct a piecewise cubic polynomial:

[0089]

[0090] wherein denotes two interpolation intervals; the coefficients are solved through boundary conditions and continuity conditions ; for any depth in the depth range , the interpolated isotope ratio is calculated through the cubic polynomial of the corresponding interval to generate a continuous depth profile.

[0091] Further, the boundary conditions of the aforementioned cubic spline interpolation adopt natural boundary conditions, i.e., the second-order derivative is zero at both ends: and ; combined with the continuity conditions at the inner nodes , a linear equation group about the coefficients is formed, wherein is the interpolation interval index, specifically, for the first interval , there is ; for the second interval , there is ; all the coefficients are obtained by solving the three-diagonal linear equation group , wherein the coefficient matrix is a three-diagonal matrix, and the right end vector is determined by the function values at the nodes and the boundary conditions; the cubic spline interpolation method ensures that the interpolation function has a continuous second-order derivative in the entire depth range, thereby ensuring the smoothness of the isotope profile.

[0092] Step 102: Calculate the coefficient of variation of isotopic ratios of each depth layer, and identify the isotopic stratification structure.

[0093] Step 103: Determine the settling or floating trend of pollutants based on the vertical gradient characteristics of isotopes.

[0094] Further, the coefficient of variation calculation formula in the aforementioned step 102 is wherein is the standard deviation of isotopic ratios of each depth layer, is the average value of isotopic ratios of each depth layer; when the coefficient of variation is greater than 1, it is determined that there is a significant isotopic stratification structure, indicating that the pollution source characteristics of different depth water layers are different.

[0095] Further, the vertical gradient calculation formula in the aforementioned step 103 is wherein is the isotopic ratio of the bottom layer, is the isotopic ratio of the surface layer, is the total water depth (unit: m); when , it is determined that the pollutants have a significant settling trend, when , it is determined that the pollutants have a significant floating trend, and when , it is determined that the pollutants are uniformly distributed in the water body.

[0096] Step 200: Add Laplace noise to the isotopic fingerprint vector of each farm based on the differential privacy algorithm, and calculate the perturbed fingerprint vector that satisfies ε-differential privacy.

[0097] The input of the aforementioned differential privacy algorithm is the local pollution source isotopic fingerprint vector of each farm and the privacy budget parameter , and the output is the perturbed fingerprint vector that satisfies - differential privacy protection .

[0098] The aforementioned differential privacy algorithm first calculates the global sensitivity of the isotopic fingerprint vector:

[0099]

[0100] wherein and are the fingerprint vectors of adjacent data sets. Then, Laplace noise is added to each component of the isotopic fingerprint vector of each farm :

[0101]

[0102] wherein, represents a Laplace distribution with scale parameter , and the noise value is obtained by random sampling from the Laplace distribution.

[0103] Further, the probability density function of the aforementioned Laplace distribution is:

[0104]

[0105] wherein represents an exponential function, and the specific implementation method of random sampling is as follows: first, a random number satisfying a uniform distribution is generated, and then a Laplace random variable is calculated by an inverse transform sampling method, wherein is a sign function, taking a value of 1 when , and taking a value of -1 when ; the generated random variable satisfies a Laplace distribution with a scale parameter , and is added as noise to the corresponding component of the isotope fingerprint vector.

[0106] Further, the generated perturbed fingerprint vector satisfies a differential privacy definition, that is, for any two adjacent data sets and (differing by only one record), and any output subset , the following is satisfied:

[0107]

[0108] wherein is a differential privacy algorithm, ensuring that an attacker cannot infer the original isotope feature from the perturbed data.

[0109] It should be noted that the global sensitivity is determined by analyzing the change range of historical isotope data, and is usually taken as twice the maximum change amplitude of the isotope ratio.

[0110] Further, the value range of the aforementioned privacy budget parameter is , wherein a smaller value (such as ) provides stronger privacy protection but introduces larger noise, and a larger value (such as ) provides weaker privacy protection but maintains higher data utility; in the present embodiment, the privacy protection requirement and the source resolution accuracy requirement are comprehensively considered, and is preferred.

[0111] Further, the aforementioned global sensitivity The theoretical basis for taking 2 times the maximum change range of the isotope ratio is that, according to the differential privacy theory, the global sensitivity needs to cover the maximum difference between adjacent data sets, considering that the L1 norm is used to measure each component of the isotope fingerprint vector, and in actual application, when a single farm exits or joins the data set, the change of the fingerprint vector is usually not more than the historical maximum change range, in order to ensure sufficient privacy protection margin, the 2 times coefficient is taken as a safety redundancy; specifically, for the nitrogen isotope ratio , the maximum change range is usually , and the corresponding sensitivity is ; for the phosphorus isotope ratio , the maximum change range is usually , and the corresponding sensitivity is .

[0112] Step 300: Use a secure multi-party computation protocol to aggregate the disturbance fingerprint vectors of each farm, and calculate the regional isotope fingerprint covariance matrix in the ciphertext domain through homomorphic encryption technology.

[0113] Each farm uses the Paillier homomorphic encryption scheme to encrypt the disturbance fingerprint vector:

[0114]

[0115] wherein, is the product of two large prime numbers, is the generator of , and is a random number.

[0116] Further, the parameter generation method of the aforementioned Paillier encryption scheme is as follows: first, select two large prime numbers and , both of which have a bit length of 1024 bits, and satisfy , calculate ; the selection method of the generator is , and the selection of the generator satisfies the generator condition and simplifies the calculation; the random number is uniformly randomly selected from the interval , and satisfies to ensure the security of encryption; the public key is , and the private key is , wherein represents the least common multiple.

[0117] Further, the homomorphic property of Paillier encryption is used to calculate the covariance matrix elements in the ciphertext domain:

[0118]

[0119] in To determine the number of participating farms, Index for farms, Represents the homomorphic multiplication operation in the ciphertext field.

[0120] 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.

[0121] 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. .

[0122] 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.

[0123] Specifically, construct polynomial of degree , where the coefficient The coefficients are polynomials and randomly selected; calculate the share. , ,in Index for participants;

[0124] During decryption, any A participant provides its share, and the private key is recovered by Lagrange interpolation:

[0125]

[0126] wherein is a set of participants providing shares, and is a participant index, ; the ciphertext covariance matrix is decrypted using the recovered private key, and the decryption formula is wherein .

[0127] In the embodiments of the present application, in order to improve the calculation efficiency, the following steps are further included on the basis of step 300:

[0128] Step 301: The high-dimensional fingerprint vector is processed by blocking, and the covariance matrix calculation is decomposed into multiple sub-matrix calculation tasks.

[0129] Further, the specific implementation method of the blocking in the aforementioned step 301 is: assuming that the fingerprint vector dimension is , the block size is , the fingerprint vector is segmented into sub-vectors, the th sub-vector is , wherein is a sub-vector index; the covariance matrix is correspondingly segmented into sub-matrix blocks , each sub-matrix is a dimensional matrix, wherein and are sub-vector indexes, and represent the covariance between the th sub-vector and the th sub-vector; in actual application, the block size is usually valued at to , so as to balance the calculation complexity and parallelism.

[0130] Step 302: The ciphertext domain operation of multiple sub-matrices is processed simultaneously by using a parallel computing architecture.

[0131] Step 400: The joint fingerprint feature space is combined with the hydrological transmission model to construct an enhanced source-sink transmission matrix, and the federal Bayesian mixture model is used to decompose the contribution proportion of different sources in the observed point nutrient salt.

[0132] Based on the regional isotope fingerprint covariance matrix , a joint fingerprint feature space is constructed, and a source-sink transmission matrix is generated by combining a hydrological transmission model .

[0133] Further, the construction method of the aforementioned joint fingerprint feature space is:

[0134] Eigenvalue decomposition is performed on the regional isotope fingerprint covariance matrix , where the superscript denotes transposition, is a diagonal matrix of eigenvalues, is each eigenvalue, is an eigenvector matrix; select the eigenvectors corresponding to the first largest eigenvalues to form the dimension reduction matrix , where is the dimension after dimension reduction, the selection of satisfies the cumulative variance contribution rate , where is the eigenvalue index, is the total number of eigenvalues; project the original fingerprint vector into the dimension reduction space to obtain the joint fingerprint feature The joint fingerprint feature space retains the main variation information of the fingerprint data of each farm, while reducing the data dimension.

[0135] Further, the aforementioned eigenvalue decomposition requires that the covariance matrix be a symmetric positive definite matrix, and the eigenvalues satisfy in descending order; the dimension after dimension reduction is in the range of , where the lower bound ensures that at least two principal components are retained to capture the basic variation pattern of the data, and the upper bound avoids an increase in computational complexity caused by too high a dimension; the eigenvector matrix satisfies the orthogonality constraint , where is an identity matrix, ensuring that the projection transformation preserves the inner product invariance.

[0136] Further, first calculate the original transmission coefficient:

[0137]

[0138] where, denotes the original transmission coefficient from the source to the observation point , is the spatial distance, is the average flow velocity, is the diffusion coefficient. The dimension of the original transmission coefficient is .

[0139] ​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:

[0140]

[0141] in km is the Earth's radius. Latitude Longitude.

[0142] To eliminate the influence of dimensions on subsequent calculations, the original transmission coefficients are normalized to obtain a dimensionless transmission weight matrix:

[0143]

[0144] 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.

[0145] Constructing a federated Bayesian mixture model:

[0146]

[0147] in, For observation point The nutrient concentration vector, For source Contribution weight, and Sources The mean and covariance of isotopic fingerprints.

[0148] The input to the aforementioned federated Bayesian mixture model is the nutrient concentration vector at each observation point:

[0149] And the corresponding isotopes form a vector superscript denotes transpose, output is the contribution weight of each pollution source wherein is the contribution weight value of each pollution source, and the isotope fingerprint parameter of each source wherein is the pollution source index, is the total number of pollution sources.

[0150] The training of the aforementioned federal Bayesian mixture model employs expectation-maximization algorithm.

[0151] The input of the aforementioned expectation-maximization algorithm is the observation dataset (including the nutrient salt concentration vector and the isotope composition vector of each observation point) and the initial parameter estimate value , and the output is the optimized model parameter .

[0152] Further, the method for obtaining the aforementioned initial parameter estimate value is: the contribution weight is initialized as uniform distribution , indicating that the initial contribution of each pollution source is equal; the isotope fingerprint mean value is initialized by K-means clustering algorithm, specifically, the isotope composition vectors of all observation points are clustered into clusters, wherein is the isotope composition vector of each observation point, is the observation point index, is the total number of observation points, and the th cluster center is taken as the initial mean value of the source ; the covariance matrix is initialized as a constant multiplied by the unit matrix , wherein is the total variance of the observation data, is the mean value of the observation data, is the unit matrix.

[0153] The aforementioned expectation-maximization algorithm includes two alternately executed units: E step (expectation step) and M step (maximization step):

[0154] In the E step, the posterior probability is calculated based on the current parameter estimate value :

[0155]

[0156] wherein is the pollution source index, is the total number of pollution sources, denotes the observation point The nutrients come from pollution sources The posterior probability.

[0157] In the M-step, the model parameters are updated based on the posterior probabilities calculated in the E-step:

[0158]

[0159]

[0160]

[0161] in For the observation point index, The total number of observation points, indicated by the superscript. Indicates transpose. This represents the number of iterations.

[0162] Furthermore, the aforementioned contribution weights The following constraints need to be met:

[0163] 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 observation point index, The total number of observation points; the nonnegativity constraint is determined by the nonnegativity of the posterior probability. ensure.

[0164] Furthermore, the aforementioned covariance matrix It needs to satisfy the positive definiteness constraint, that is, for any non-zero vector ,satisfy ;

[0165] 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.

[0166] The data transmission relationship of the aforementioned federal Bayesian mixture model is that the nutrient salt concentration and isotope data of each observation point are first mapped back by the source-sink transmission matrix to obtain the equivalent concentration in the source space, then the posterior probability of each source is calculated in the E step, and the posterior probability of each source is used to update the model parameters in the M step, and the E step and the M step are iteratively executed until convergence.

[0167] The aforementioned federal Bayesian mixture model uses a negative log-likelihood function as the loss function:

[0168]

[0169] wherein is the observation point index, is the total number of observation points, is the pollution source index, is the total number of pollution sources.

[0170] Further, the time dimension embodiment of the aforementioned loss function is that when the training data contains observations at multiple time points, the observation data sets at different time points are denoted as , wherein is the data of each observation point at each time point, is the observation point index, is the number of observation points at time point ; the extended spatiotemporal loss function is expressed as , wherein is the time point index, is the total number of time points, is the weight coefficient of time point , satisfying the normalization condition and ; the value of the weight coefficient is:

[0171] For equally spaced sampling data, a uniform weight is used; for scenarios that need to emphasize recent data, an exponentially decaying weight is used, wherein represents an exponential function, is the time point index, and the value range of the decay coefficient is , a larger value gives higher weight to recent data; the pollution source contribution weight may be time-varying at different time points, reflecting the dynamic change characteristics of the pollution source contribution.

[0172] ​Further, the training process minimizes the loss function by EM algorithm, and stops training when the loss function changes less than a convergence threshold between two adjacent iterations, or terminates when the maximum iteration number is reached.

[0173] It should be noted that, represents the total number of pollution sources, including each aquaculture farm, land source and open sea input source.

[0174] Further, the total number of the aforementioned pollution sources is in the range of , wherein the lower bound ensures that at least the main pollution source types (such as aquaculture farms, land sources, and open sea inputs) can be distinguished, and the upper bound avoids unstable estimation of model parameters caused by too many pollution sources; the relationship between the total number of pollution sources and the total number of observation points needs to satisfy to ensure that there is enough observation data to support reliable estimation of model parameters; when , the number of pollution sources needs to be reduced or similar pollution source categories need to be merged.

[0175] Further, the value range of the aforementioned average flow velocity is determined according to the hydrological conditions of the nearshore aquaculture area, and for the nearshore waters dominated by tides, the value range is m / s, wherein the smaller value corresponds to the weak tidal period, and the larger value corresponds to the strong tidal period; in actual application, real-time flow velocity data is obtained by continuous monitoring of the acoustic Doppler current profiler (ADCP) arranged on site, and the average value in the monitoring period is taken as the model input parameter.

[0176] Further, the value of the aforementioned diffusion coefficient is determined according to the turbulent diffusion theory, and its calculation formula is ,

[0177] wherein is the dimensionless diffusion coefficient, and its value range is , is the characteristic water depth (unit: m), and for the nearshore aquaculture area, the characteristic water depth is usually in the range of m; considering the above parameters comprehensively, the typical value range of the diffusion coefficient is m² / s.

[0178] Further, the aforementioned convergence threshold ​​The selection basis is that 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 numerical steady state, and further iterations have negligible improvements in parameter estimation; the maximum number of iterations The setting is based on actual training experience, and under normal circumstances, the EM algorithm can converge within 100-200 iterations, and setting 500 as the upper limit is to prevent infinite loops in extreme cases, while ensuring the robustness of the algorithm.

[0179] Step 500: Based on the completed pollution source identification model, input the real-time monitoring of nutrient salt concentration and isotope data, and output the real-time contribution ratio and confidence interval of each pollution source.

[0180] The real-time collected water sample nutrient salt concentration data and isotope composition data are subjected to the same Z-score standardization processing as step 100, and the mean and standard deviation parameters saved during the training phase are used to standardize the real-time data to obtain standardized data and .

[0181] The standardized data and are input into the model trained in step 400, and the posterior probability of each pollution source is calculated:

[0182]

[0183] wherein, represents pollution from the th source, is the pollution source index, is the total number of pollution sources, is the likelihood probability of observing data under the condition of the th pollution source.

[0184] The aforementioned likelihood probability is calculated based on a multivariate normal distribution, and its calculation formula is:

[0185]

[0186] wherein, represents the exponential function, and the superscript represents the transpose, is the combined observation vector, is the vector dimension, is the comprehensive mean vector of the th pollution source, is the comprehensive covariance matrix, represents the determinant of the covariance matrix. The likelihood probability formula is calculated under the given pollution source Under the given parameter conditions, the probability density of the current nutrient concentration and isotopic composition was observed.

[0187] 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 ;

[0188] 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.

[0189] The contribution ratio of each pollution source is calculated based on posterior probability. .

[0190] 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. .

[0191] 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. .

[0192] Furthermore, the specific implementation steps of the aforementioned Metropolis-Hastings algorithm are as follows:

[0193] Step 1: Initialization: Setting initial parameter samples That is, using the optimal parameters obtained from training as the starting point;

[0194] 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. where the proposal covariance matrix , is the empirical covariance matrix of the posterior distribution;

[0195] Third step: acceptance probability calculation:

[0196] Calculate acceptance probability:

[0197]

[0198] Since the symmetric proposal distribution is adopted, the acceptance probability is simplified as:

[0199]

[0200] Fourth step: acceptance decision: generate a uniform random number , if then accept the candidate parameter , otherwise keep the current parameter ;

[0201] Fifth step: repeat steps two to four until samples are generated.

[0202] Further, for each pollution source ,

[0203] Arrange the sampled contribution proportion sequence in ascending order, take the th quantile as the lower confidence limit , and take the th quantile as the upper confidence limit to obtain the confidence interval .

[0204] Further, the value of the aforementioned MCMC sampling number needs to balance the calculation efficiency and statistical reliability. According to the convergence theory of the Monte Carlo method, the sample number should satisfy to ensure the stability of the posterior distribution estimation; in the present embodiment, it is set that , the first samples are included as burn-in period to eliminate the influence of initial values, and the last samples are used for confidence interval calculation; under the above sampling number setting, the Monte Carlo standard error of the contribution proportion estimation is usually less than .

[0205] Further, the aforementioned confidence level is usually taken as 95% (i.e. ), and can be taken as 99% (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.

[0206] 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:

[0207] Step 501: Cache historical detection results and build a fast query index.

[0208] 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.

[0209] Step 502: When the similarity between the new data and the cached data exceeds the threshold, return the cached result directly.

[0210] Furthermore, the similarity in step 502 above uses Euclidean distance as a metric;

[0211] 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 an average deviation of no more than 1 standard deviation in each dimension between the new data and the cached data.

[0212] A certain coastal aquaculture area in a certain county of a certain city in a certain province contains 3 sea cucumber farms (Farm A, Farm B, and Farm C), 1 land-based sewage outlet, and a distant sea input source, which collectively affect the water quality of the area. Each farm needs to identify the contribution proportion of each pollution source to the aquaculture area in order to develop a scientific pollution prevention and control strategy, but each farm is reluctant to share raw isotope data containing sensitive information such as aquaculture density and feed formula due to commercial competition.

[0213] Each farm collects water samples within its aquaculture area and measures nutrient salt concentrations and isotope ratios. Taking Farm A as an example, the multi-depth water sample data collected in a certain observation period in July 2024 is shown in Table 1.

[0214] Table 1 Nutrient salt concentration and isotope composition data of Farm A:

[0215]

[0216] After Z-score standardization of the obtained data, Morlet wavelet transform (center frequency , scale number ) is used to extract the spatio-temporal variation characteristics of isotopes. Taking nitrogen isotope as an example, the statistical characteristics of wavelet coefficients calculated at 6 scales are shown in Table 2.

[0217] Table 2 Statistical characteristics of nitrogen isotope wavelet coefficients of Farm A:

[0218]

[0219] Combined with the wavelet features of phosphorus isotopes, the local pollution source isotope fingerprint vector of Farm A is constructed, with a feature dimension of .

[0220] Figure 2 The distribution of 5 kinds of nutrient salt concentrations at different depths (surface, middle layer, and bottom layer) of Farm A is shown.

[0221] Figure 3 The energy feature distribution extracted by wavelet transform at different scales is shown.

[0222] The privacy budget parameter is set to the global sensitivity (based on 2 times the maximum variation amplitude of nitrogen isotope). Laplace noise is added to each component of the isotope fingerprint vector of Farm A, and the scale parameter of the Laplace distribution is Table 3 shows the perturbation process of partial characteristic components.

[0223] Table 3: Differential privacy perturbation process of farm A (partial characteristics):

[0224]

[0225] Farm A obtains the perturbed fingerprint vector by the above perturbation process. Farm B and farm C perform the same differential privacy perturbation operation on their respective isotope fingerprint vectors to obtain perturbed fingerprint vectors .

[0226] Figure 4 The characteristic value changes before and after the differential privacy algorithm adds Laplace noise to the isotope fingerprint vector are shown.

[0227] Each farm uses the Paillier homomorphic encryption scheme to encrypt the perturbed fingerprint vector (key length 2048 bits) and uploads it to the collaborative computing platform. The platform calculates each element of the regional isotope fingerprint covariance matrix in the ciphertext domain, and obtains the plaintext covariance matrix through a threshold decryption method (set parties, threshold ).

[0228] Nutrient salt concentration data and isotope composition data of 18 observation points in the region in July 2024 are collected as training data sets. Combined with hydrological transport model parameters (average flow rate m / s, diffusion coefficient m² / s), a source-sink transport matrix is constructed. A federated Bayesian mixture model is used for training, with a total of pollution sources (farms A, B, and C, land-based sewage outlets, and distant sea inputs) and a total of observation points. After 168 iterations, the loss function decreases from the initial value to . The contribution weights of each pollution source and the mean parameters of the isotope fingerprint obtained by training are shown in Table 4.

[0229] Table 4: Contribution weights of each pollution source and mean parameters of isotope fingerprint obtained by training:

[0230]

[0231] Figure 5 The contribution weights of each pollution source (bar chart) and the mean of the isotope fingerprint (line chart) obtained by training the federated Bayesian mixture model are shown.

[0232] Figure 7The loss function convergence curve of the expectation maximization (EM) algorithm is shown.

[0233] In August 2024, the regional monitoring point collected new water sample data, and the nutrient salt concentration and isotope composition are shown in Table 5.

[0234] Table 5 Water quality data of real-time monitoring points:

[0235]

[0236] The monitoring data is input into the trained model, and the contribution ratio and 95% confidence interval of each pollution source are calculated by MCMC sampling (burn-in period of 1000 samples).

[0237] Table 6 Identification results of pollution source contribution ratio of monitoring point M1:

[0238]

[0239] Figure 6 The contribution ratio of each pollution source of monitoring point M1 and its 95% confidence interval are shown.

[0240] According to the identification results, the nutrient salt pollution of monitoring point M1 mainly comes from aquaculture farm A (contribution ratio of 32%) and aquaculture farm B (contribution ratio of 29%), land source sewage outlet contributes 18%, aquaculture farm C contributes 16%, and far sea input contributes less (5%). The results provide a scientific basis for regional pollution collaborative governance, and each aquaculture farm can develop corresponding emission reduction measures according to its own contribution ratio.

[0241] The above describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative and not limiting, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the 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 characteristics 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