Federal learning-based performance collaborative optimization method and system for multiple sets of separation systems

Abstract

The invention relates to the technical field of system optimization, and discloses a federated learning-based performance collaborative optimization method and system for multiple sets of separation systems, and the method comprises the steps: obtaining the radioactivity value of the current batch of raw material liquid of each set of At-211 separation system, and carrying out the analysis of the radioactivity value based on a deep neural network, generating a working condition sensitive feature vector representing the influence of different batches of radioactive activity values on the performance of each separation system; through the working condition sensitive feature vector, the working condition attribute factor, the activity working condition vector and the collaborative optimization coefficient which are sequentially generated in each step, the influence of different batches of radioactive activity values can be accurately mined, the working condition association and difference of each system are determined, the differential aggregation weight fine adjustment is realized, the out-of-distribution generalization failure of cross-activity migration is avoided, and the system safety is improved. It is ensured that technological parameters adapt to the requirements of raw material solutions with different activities, the cooperative capacity of multiple At-211 separation systems is improved, and the separation efficiency and the product purity are improved.

Description

A method and system for performance collaborative optimization of multiple separation systems based on federated learning Technical Field

[0001] This invention relates to the field of system optimization technology, specifically to a method and system for collaborative performance optimization of multiple separate systems based on federated learning. Background Technology

[0002] Currently, when optimizing multiple At-211 separation systems collaboratively, a federated learning approach—local training, parameter uploading, aggregation update, and parameter distribution—is often used to achieve coordinated adjustment of process parameters across multiple systems, thereby optimizing the separation efficiency and product purity of each system.

[0003] However, the above-mentioned collaborative optimization method still has the following defects: When multiple At-211 separation systems process feed liquids with different radioactivity, the existing federated learning method generally adopts a static polymerization strategy, which fails to perform differentiated polymerization based on the key feature of feed liquid activity. Since the optimal separation parameters corresponding to different activities are clustered in the feature space, the static polymerization strategy cannot mine and learn the effective process knowledge under the same activity conditions from a global perspective, resulting in out-of-distribution generalization failure when migrating across activities. This leads to the issued process parameters not being able to adapt to the separation requirements of feed liquids with different activities, resulting in incomplete separation of At-211 from impurities and affecting the effect of collaborative optimization of the performance of different At-211 separation systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for collaborative performance optimization of multiple separation systems based on federated learning, thus solving the aforementioned problems.

[0005] The above-mentioned technical objective of this invention is achieved through the following technical solution: a method for collaborative optimization of the performance of multiple separation systems based on federated learning, comprising: step S1, obtaining the radioactivity value of the current batch of feed liquid in each At-211 separation system, analyzing the radioactivity value based on a deep neural network, and generating a condition-sensitive feature vector representing the influence of different batches of radioactivity value on the performance of each separation system; step S2, analyzing the condition-sensitive feature vector according to the orthogonal decomposition neural network layer to obtain a condition attribute factor representing the significance of activity condition characteristics; step S3, fusing the condition-sensitive feature vector and the condition attribute factor according to federated learning, and analyzing the radioactivity value of each At-211 separation system to generate an activity condition vector representing the current condition category of each At-211 separation system; step S4, optimizing the activity condition vector to obtain a collaborative optimization coefficient used to guide the differentiated aggregation of federated learning.

[0006] Furthermore, based on the analysis of radioactivity values ​​using a deep neural network, a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system is generated. This includes: inputting the radioactivity values ​​of the current batch of feed liquid in each At-211 separation system into the deep neural network, extracting the dynamic change characteristics of the radioactivity values ​​based on its temporal attention mechanism, and generating an activity perturbation vector representing the temporal fluctuation characteristics of the activity values; analyzing the activity perturbation vector and the radioactivity values ​​to generate a performance impact vector representing the magnitude of the impact of the activity perturbation on each performance index.

[0007] Furthermore, based on the analysis of radioactivity values ​​using deep neural networks, a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system is generated. This also includes: analyzing the relationship between the activity perturbation vector and the performance impact vector to generate a condition impact factor representing the degree of dynamic interaction between activity and performance; and fusing the radioactivity values, activity perturbation vector, and condition impact factor to generate a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system.

[0008] Furthermore, based on the analysis of the working condition sensitive feature vector by the orthogonal decomposition neural network layer, the working condition attribute factors representing the significance of the activity working condition characteristics are obtained, including: orthogonally projecting the working condition sensitive feature vector to obtain the working condition projection coefficients representing the energy distribution in each orthogonal direction; transforming the working condition projection coefficients to generate the working condition activation response vector representing the characteristic state after nonlinear mapping; and calculating the working condition activation response vector through a multi-head self-attention mechanism to generate the working condition component contribution weights representing the importance of the working condition.

[0009] Furthermore, based on the analysis of the working condition sensitive feature vector by the orthogonal decomposition neural network layer, the working condition attribute factor representing the significance of the activity working condition characteristics is obtained. It also includes: fusing the working condition projection coefficient with the working condition component contribution weight to generate the working condition ground state vector representing the main changes in the working condition; and performing dominant feature analysis on the working condition ground state vector to generate the working condition attribute factor representing the significance of the activity working condition characteristics.

[0010] Furthermore, based on federated learning, the condition-sensitive feature vector and condition attribute factors are fused, and the radioactivity values ​​of each At-211 separation system are analyzed to generate an activity condition vector representing the current condition category of each At-211 separation system. This includes: interacting with the condition-sensitive feature vector and condition attribute factors based on the local fusion network of the federated learning client, and calculating in conjunction with the current batch radioactivity values ​​to generate a dynamic condition imprint representing the current condition status of each client; uploading the dynamic condition imprints of each client to the federated server, and calculating the similarity between the dynamic condition imprints of different clients to generate a condition association matrix representing the strength of the condition association between clients.

[0011] Furthermore, based on federated learning, the system fuses the condition-sensitive feature vectors and condition attribute factors, analyzes the radioactivity values ​​of each At-211 separation system, and generates an activity condition vector representing the current condition category of each At-211 separation system. This also includes: performing aggregation analysis on the condition correlation matrix to generate a federated consensus coefficient to guide federated optimization; and fusing the dynamic condition imprints of each client with the federated consensus coefficient to generate an activity condition vector representing the current condition category of each At-211 separation system.

[0012] Furthermore, the activity condition vectors are optimized to obtain the collaborative optimization coefficients used to guide the differentiated aggregation of federated learning. This includes: analyzing the federated aggregation process under different collaborative weights based on the activity condition vectors of each At-211 separation system, and generating condition response sensitivity that represents the sensitivity of each system to the collaborative weights.

[0013] Furthermore, the activity condition vector is optimized to obtain the co-optimization coefficients used to guide the differentiated aggregation of federated learning. This also includes: constraining and optimizing the sensitivity of the condition response to generate the co-optimization coefficients used to guide the differentiated aggregation of federated learning.

[0014] Furthermore, a performance collaborative optimization system for multiple separation systems based on federated learning, applied to the aforementioned optimization method, includes: an impact analysis unit, used to obtain the radioactivity value of the current batch of feed liquid in each At-211 separation system, analyze the radioactivity value based on a deep neural network, and generate a condition-sensitive feature vector representing the impact of different batches of radioactivity value on the performance of each separation system; a feature analysis unit, used to analyze the condition-sensitive feature vector based on the orthogonal decomposition neural network layer, and obtain condition attribute factors representing the significance of activity condition characteristics; a condition analysis unit, used to fuse the condition-sensitive feature vector and condition attribute factors based on federated learning, and analyze the radioactivity value of each At-211 separation system, generating an activity condition vector representing the current condition category of each At-211 separation system; and a collaborative optimization unit, used to optimize the activity condition vector, and obtain collaborative optimization coefficients to guide the differentiated aggregation of federated learning.

[0015] In summary, the present invention has the following main advantages: By generating condition-sensitive feature vectors, these vectors can accurately represent the impact of different batches of radioactivity values ​​on the performance of each separation system, fully exploring effective process knowledge under similar activity conditions, breaking the limitations of existing static polymerization strategies. By generating condition attribute factors, these factors can highlight the significance of activity condition characteristics, accurately identify the core differences between different activity conditions, and avoid polymerization deviations caused by unclear condition characteristics. By generating activity condition vectors, these vectors can clearly represent the current condition category of each At-211 separation system, clarifying the correlation strength and individual differences between systems, facilitating the later implementation of differentiated polymerization weight adjustment in federated learning, effectively avoiding distributional out-generalization failure during cross-activity migration. By generating collaborative optimization coefficients, these coefficients can guide federated learning to achieve fine-grained adjustment of differentiated polymerization weights, ensuring that the issued process parameters accurately adapt to the separation requirements of different activity feed solutions, improving the thoroughness of At-211 separation from impurities, and comprehensively optimizing the performance coordination level of multiple At-211 separation systems. Attached Figure Description

[0016] Figure 1 is a flowchart of the performance collaborative optimization method for multiple separation systems based on federated learning according to the present invention; Figure 2 is a schematic diagram of the performance collaborative optimization system for multiple separation systems based on federated learning according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Referring to Figures 1 and 2, the collaborative optimization method for the performance of multiple separation systems based on federated learning includes: Step S1, obtaining the radioactivity value of the current batch of feed liquid in each At-211 separation system, analyzing the radioactivity value based on a deep neural network, and generating a condition-sensitive feature vector representing the impact of different batches of radioactivity value on the performance of each separation system; Step S2, analyzing the condition-sensitive feature vector based on the orthogonal decomposition neural network layer to obtain the condition attribute factor representing the significance of activity condition characteristics; Step S3, fusing the condition-sensitive feature vector and the condition attribute factor based on federated learning, and analyzing the radioactivity value of each At-211 separation system to generate an activity condition vector representing the current condition category of each At-211 separation system; Step S4, optimizing the activity condition vector to obtain the collaborative optimization coefficient used to guide the differentiated aggregation of federated learning.

[0019] In one embodiment, the analysis of radioactivity values ​​based on a deep neural network generates a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system. This includes: inputting the radioactivity values ​​of the current batch of feed liquid from each At-211 separation system into the deep neural network; extracting the dynamic change features of the radioactivity values ​​based on its temporal attention mechanism; and generating an activity perturbation vector representing the temporal fluctuation characteristics of the activity values. Specifically, this includes: arranging the radioactivity values ​​of the current batch of feed liquid from each At-211 separation system in the order of acquisition time to form a temporal activity data matrix; inputting the temporal activity data matrix into a one-dimensional convolutional layer of the deep neural network; extracting the local fluctuation features of the radioactivity values ​​at continuous time points; generating a temporal feature map, which is a three-dimensional feature map with dimensions including the number of batches, time step, and number of feature channels; inputting the temporal feature map into the dynamic temporal attention module of the deep neural network, which first calculates the coefficient of variation of the radioactivity values ​​at each time step as the initial attention score; then flattening the temporal feature map along the feature channel dimension and inputting it into a gated loop unit for gating. The recurrent unit processes data sequentially according to time steps, outputting a hidden state at each time step. This hidden state aggregates the temporal dependency information from the starting time step to the current time step. The hidden state of each time step is input into a fully connected layer, which maps it to a correction coefficient vector with the same dimension as the number of feature channels. This correction coefficient vector is added element-wise to the initial attention score of the corresponding time step to obtain the corrected attention weight. The corrected attention weight is normalized along the time step dimension so that the sum of the weights of all time steps for each feature channel is 1, resulting in an attention weight matrix with dimensions of time step and number of feature channels. The attention weight matrix is ​​then multiplied element-wise with the temporal feature map, i.e., the eigenvalue of each time step and each feature channel is multiplied by the corresponding corrected attention weight, resulting in a weighted feature map with the dimension unchanged. The weighted feature map is then global average pooled along the new time step dimension, resulting in an average value for each feature channel, forming a one-dimensional feature vector with the dimension of the number of feature channels. This one-dimensional feature vector is input into a fully connected layer for nonlinear transformation, generating an activity perturbation vector representing the temporal fluctuation characteristics of the activity value.

[0020] Analyzing the activity perturbation vector and radioactivity values, a performance impact vector representing the magnitude of the activity perturbation's influence on each performance index is generated. Specifically, this includes: appending the radioactivity value to the end of the activity perturbation vector to form a joint input vector; collecting the joint input vectors and corresponding performance index data from N consecutive batches preceding the current batch, with performance index data including at least single-batch yield, total separation time, and product concentration; using the sequence of each dimension of the joint input vector formed by the N consecutive batches as independent variables and each performance index as the dependent variable, calculating the Pearson correlation coefficient to obtain an initial correlation matrix. The number of rows in this matrix equals the number of dimensions of the joint input vector, and the number of columns equals the number of performance indices. Each element in the matrix represents the degree of linear correlation between the corresponding perturbation dimension and the corresponding performance index; calculating... The mean value of the current radioactivity of each At-211 separation system is calculated. Then, the radioactivity of the current batch of raw material liquid is compared with the mean value, and the percentage deviation between the two is calculated. This percentage deviation is normalized to the 0-1 interval to obtain the activity benchmark adjustment coefficient. Each element in the initial correlation matrix is ​​multiplied element-wise with the activity benchmark adjustment coefficient to obtain the corrected correlation matrix. The corrected correlation matrix is ​​then summed along the row direction to obtain a new vector. The dimension of the new vector is the same as the number of performance indicators. The new vector is then normalized using the L2 norm so that the square root of the sum of squares of each element is equal to 1. This generates a performance influence vector representing the magnitude of the influence of the activity perturbation on each performance indicator. Each element value in the performance influence vector represents the influence weight of the current activity perturbation on the corresponding performance indicator.

[0021] In one embodiment, the analysis of radioactivity values ​​based on a deep neural network generates a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system. This further includes analyzing the relationship between the activity perturbation vector and the performance impact vector to generate a condition impact factor representing the degree of dynamic interaction between activity and performance. Specifically, this includes: obtaining the activity perturbation vector and performance impact vector for the previous N consecutive batches; arranging these vectors in batch order to form a historical activity perturbation matrix and a historical performance impact matrix, where each matrix's rows correspond to samples from different batches, and columns correspond to various dimensions of different vectors; performing mean-reduction processing on the historical activity perturbation matrix and the historical performance impact matrix, i.e., subtracting the mean of its column from each element to obtain two centered matrices; calculating the cross-covariance matrix of these two matrices; performing singular value decomposition on the cross-covariance matrix; and using the largest singular value as the basic interaction strength, which reflects the strongest interaction strength between activity perturbation and performance impact in historical data; and calculating... Calculate the mean values ​​of each column in the historical activity perturbation matrix and the historical performance influence matrix to obtain the historical activity perturbation mean vector and the historical performance influence mean vector. Calculate the Mahalanobis distance between the current batch activity perturbation vector and the historical activity perturbation mean vector. Similarly, calculate the Mahalanobis distance between the current performance influence vector and the historical performance influence mean vector. Add the two Mahalanobis distances and divide by 2 to obtain the comprehensive deviation coefficient. The comprehensive deviation coefficient represents the overall deviation of the current operating condition relative to the historical average pattern in terms of both activity perturbation and performance response. Multiply the basic interaction strength by the comprehensive deviation coefficient, divide by the arithmetic mean of the two, and normalize the result to the 0-1 interval to obtain the dynamic interaction coefficient. Collect the radioactivity values ​​of all N batches, sort them in ascending order of value, determine the percentile position of the current batch activity value in the historical ranking, use this percentile value as the activity weight, multiply the dynamic interaction coefficient by the activity weight, and normalize the result to the 0-1 interval to obtain the operating condition influence factor representing the degree of dynamic interaction between activity and performance.

[0022] The radioactivity value, activity perturbation vector, and operating condition influence factor are fused to generate an operating condition sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system. Specifically, this involves: calculating the standard deviation of the activity perturbation vector; dividing the standard deviation by the mean of each dimension of the activity perturbation vector to obtain the coefficient of variation vector; multiplying the radioactivity value element-wise with the coefficient of variation vector to generate an activity modulation vector with the same dimension as the activity perturbation vector; multiplying the activity modulation vector element-wise with the operating condition influence factor to obtain an operating condition weighted modulation vector; and adding the operating condition weighted modulation vector element-wise with the activity perturbation vector to obtain a new vector with the same dimension as the activity perturbation vector. This new vector is then subjected to L2 norm normalization so that the square root of the sum of squares of each element equals 1, thus generating the operating condition sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system.

[0023] By combining deep neural networks with temporal attention mechanisms, the dynamic change characteristics of radioactivity values ​​are accurately extracted, generating an activity perturbation vector. This vector is then fused with performance impact vectors and operating condition impact factors to obtain an operating condition sensitive feature vector. This enables precise analysis of the impact of radioactivity values ​​on different batches, thereby uncovering process knowledge under different activity conditions. This avoids distributional out-generalization failures during cross-activity migration, making subsequent federated learning aggregation strategies more targeted. It ensures that the issued process parameters are adapted to the separation requirements of feedstock solutions with different activities, improves the thoroughness of At-211 separation from impurities, and optimizes the performance synergy of multiple At-211 separation systems.

[0024] In one embodiment, the working condition sensitive feature vector is analyzed according to the orthogonal decomposition neural network layer to obtain the working condition attribute factor representing the significance of activity working condition characteristics. This includes: orthogonally projecting the working condition sensitive feature vector to obtain the working condition projection coefficient representing the energy distribution in each orthogonal direction. Specifically, this includes: collecting the working condition sensitive feature vectors of all batches before the current batch, arranging these working condition sensitive feature vectors in batch order to construct a historical working condition feature matrix. The rows of the historical working condition feature matrix correspond to different batches, and the columns correspond to the various dimensions of the working condition sensitive feature vector. Singular value decomposition is performed on the historical working condition feature matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. At the same time, the sum of all singular values ​​is calculated, and the singular values ​​are sorted from largest to smallest and accumulated sequentially. When the accumulated sum first reaches 90% of the sum of singular values, the number of accumulated singular values ​​at this time is recorded. Let m be the column vectors corresponding to the m largest singular values ​​in the right singular vector matrix. These column vectors are the initial orthogonal basis vectors. Arrange these m column vectors in the order they were extracted to form the initial orthogonal basis matrix. Each column of the initial orthogonal basis matrix is ​​an initial orthogonal basis vector. Perform an inner product operation between the current batch of condition-sensitive feature vectors and the initial orthogonal basis matrix. That is, multiply each column of the basis matrix by the condition-sensitive feature vector to obtain a set of values. Each value represents the projection length of the condition-sensitive feature vector in that orthogonal direction. Combine these projection lengths in the order of the initial orthogonal basis vectors to form a projection coefficient vector. The dimension of the projection coefficient vector is equal to the number of initial orthogonal basis vectors. Square each element in the projection coefficient vector to obtain the energy value in each orthogonal direction. Normalize the energy value to the 0-1 interval to obtain the condition projection coefficients representing the energy distribution in each orthogonal direction.

[0025] The working condition projection coefficients are transformed to generate a working condition activation response vector representing the characteristic state after nonlinear mapping. Specifically, this involves: collecting the working condition projection coefficients from all previous batches; sorting these coefficients in ascending order of value; dividing the sorted sequence into ten equal-length intervals, each containing the same number of working condition projection coefficients; recording the lower and upper limits of each interval; obtaining ten consecutive and non-overlapping projection intervals; comparing the working condition projection coefficients of the current batch with the ten projection intervals; identifying the projection interval to which the current working condition projection coefficient belongs; and calculating the difference between the current working condition projection coefficient and the lower limit of its corresponding projection interval. Divide by the length of the projection interval to obtain the relative position coefficient between 0 and 1; construct a ten-dimensional vector with an initial value of zero, where the ten dimensions of the ten-dimensional vector correspond to the ten projection intervals. Assign the relative position coefficient to the dimension corresponding to the projection interval to which the current working condition projection coefficient belongs, while keeping the other nine dimensions at zero. Perform an exponential operation on the ten-dimensional vector, that is, calculate with the natural constant e as the base and each element value in the ten-dimensional vector as the exponent to obtain a set of working condition activation values. After normalizing each working condition activation value to the 0-1 interval, arrange them in the original order to generate a working condition activation response vector representing the characteristic state after nonlinear mapping.

[0026] A multi-head self-attention mechanism is used to calculate the working condition activation response vector, generating working condition component contribution weights representing the importance of the working conditions. Specifically, this involves: setting the number of attention heads to four; mapping the working condition activation response vector to four sub-vectors of equal dimensions through linear projection, with each sub-vector corresponding to the input of one attention head; within each attention head, performing a linear transformation on each dimension of its sub-vector to generate two intermediate variables for that dimension, denoted as the query variable and the key variable; for any two dimensions in the sub-vector, multiplying the query variable and the key variable to obtain the initial association scores for these two dimensions. Calculate the initial correlation score for each pair of all dimensions within the sub-vector to obtain the correlation matrix. The number of rows and columns of the correlation matrix is ​​equal to the number of dimensions processed by the attention head. Sum the correlation matrix of each attention head row by row to obtain the total correlation vector of each dimension with all other dimensions. Divide each element in the total correlation vector by the sum of all elements in the total correlation vector to make the sum of the weights of each dimension within the attention head equal to 1, thus obtaining the initial contribution weight of each dimension within the attention head. Normalize the initial contribution weights obtained from the four attention heads to the interval between 0 and 1 to obtain the contribution weight of the condition component representing the importance of the condition.

[0027] In one embodiment, the analysis of the operating condition sensitive feature vector based on the orthogonal decomposition neural network layer yields the operating condition attribute factor representing the significance of activity operating condition characteristics. The method further includes fusing the operating condition projection coefficient with the operating condition component contribution weights to generate an operating condition ground state vector representing the main changes in the operating condition. Specifically, this includes: multiplying the operating condition projection coefficient by the operating condition component contribution weights to obtain the energy important composite value; obtaining the operating condition projection coefficients of all batches prior to the current batch; calculating the average of these operating condition projection coefficients; simultaneously collecting all operating condition component contribution weights for the corresponding batch; calculating the average of these operating condition component contribution weights; multiplying the two averages to obtain the historical average composite value; calculating the difference between the energy important composite value and the historical average composite value; dividing the difference by the absolute value of the historical average composite value to obtain the relative rate of change; constructing a two-dimensional vector, with the first dimension assigned the current operating condition projection coefficient and the second dimension assigned the relative rate of change; and normalizing the values ​​of this two-dimensional vector to the 0-1 interval, which is the operating condition ground state vector representing the main changes in the operating condition.

[0028] Dominant feature analysis is performed on the ground state vector of the operating condition to generate operating condition attribute factors representing the significance of activity operating condition characteristics. Specifically, this includes: calculating the sum of the absolute values ​​of the operating condition projection coefficient and the relative rate of change in the ground state vector of the operating condition; dividing the operating condition projection coefficient by the sum of the absolute values ​​to obtain the operating condition proportion coefficient; dividing the absolute value of the relative rate of change by the sum of the absolute values ​​to obtain the rate of change proportion coefficient; using the larger value between the operating condition proportion coefficient and the rate of change proportion coefficient as the dominant factor, constructing a two-dimensional vector, assigning the dominant factor to the first dimension, and assigning a value to the second dimension according to the dimension type: if the dominant factor comes from the operating condition proportion coefficient, it is assigned a value of 0; if it comes from the rate of change proportion coefficient, it is assigned a value of 1; calculating the modulus of the two-dimensional vector, i.e., calculating the square root of the sum of the squares of the values ​​of the two dimensions, and normalizing the calculation result to the 0-1 interval, which is the operating condition attribute factor representing the significance of activity operating condition characteristics. The larger the operating condition attribute factor, the more significant the dominant feature of the operating condition.

[0029] By performing in-depth analysis of the operating condition sensitive feature vector through orthogonal decomposition neural network layers, operating condition projection coefficients, operating condition activation response vectors, and operating condition component contribution weights are generated sequentially. Finally, operating condition attribute factors representing the significance of activity operating condition characteristics are obtained. This can accurately discover the core features of similar activity operating conditions, enhance the identification of dominant features of activity operating conditions, avoid distribution out-generalization failure during cross-activity migration, ensure that process parameters are adapted to the needs of feed liquids with different activities, improve the thoroughness of At-211 separation from impurities, optimize the synergistic performance of multiple At-211 separation systems, and improve separation efficiency and product purity.

[0030] In one embodiment, the process involves fusing condition-sensitive feature vectors and condition attribute factors using federated learning, analyzing the radioactivity values ​​of each At-211 separation system, and generating an activity condition vector representing the current condition category of each At-211 separation system. This includes: interacting with the condition-sensitive feature vectors and condition attribute factors using the local fusion network of the federated learning client, and calculating based on the current batch radioactivity values ​​to generate a dynamic condition imprint representing the current condition status of each client. Specifically, under the federated learning framework, each At-211 separation system is used as a client, and the condition-sensitive feature vector, condition attribute factors, and current batch radioactivity values ​​are input into the local fusion network of each client. Locally, each dimension of the condition-sensitive feature vector is multiplied by the condition attribute factors to obtain a weighted feature vector. Simultaneously, the current batch radioactivity values ​​are multiplied element-wise by the weighted feature vector to obtain a modulation feature vector, thereby realizing the attribute factors. The original features are jointly modulated with the activity level. The modulated feature vector is input into a fully connected layer with half the number of neurons as the dimension of the condition-sensitive feature vector. The output is a dimension-reduced feature vector. The dimension-reduced feature vector is then multiplied by itself to obtain a square matrix with the number of rows and columns equal to the dimension of the dimension-reduced feature vector. The matrix is ​​then summed row by row to obtain a vector with the same dimension as the dimension-reduced feature vector. This vector is then added element by element to the dimension-reduced feature vector to obtain the fused associated feature vector. The associated feature vector is input into a gating unit. The gating unit first calculates the mean of the associated feature vector and compares each element with the mean. Elements greater than the mean are multiplied by 1.2, elements less than the mean are multiplied by 0.8, and elements equal to the mean are multiplied by 1 to obtain the gated vector. Each element in this vector is normalized to the 0-1 interval. The normalized vector is the dynamic condition imprint representing the current condition status of each client.

[0031] The dynamic work condition imprints of each client are uploaded to the federated server, and the similarity between the dynamic work condition imprints of different clients is calculated to generate a work condition association matrix representing the strength of work condition associations between clients. Specifically, under the federated learning framework, each client uploads its locally generated dynamic work condition imprint to the federated server. After receiving the dynamic work condition imprints of all clients, the server calculates the arithmetic mean of all elements of each client's dynamic work condition imprint, and uses this mean as the baseline threshold for that dynamic work condition imprint. The baseline threshold reflects the average activation level of each dimension feature under the current work condition state of the client. For each client's dynamic work condition imprint, the element value of each dimension is compared with the baseline threshold: if the element value is ≥ the baseline threshold, the dimension is assigned a value of 1; if the element value is < the baseline threshold, the dimension is assigned a value of 0. This yields a binary pattern vector for each client, representing the dominant dimension distribution above average in the client's operating condition characteristics using 0s and 1s. For any two clients, the number of dimensions where both are 1 and the number of dimensions where at least one is 1 are counted. The former is divided by the latter, and the result is normalized to the 0-1 interval. This represents the operating condition correlation strength between the two clients. A stronger correlation strength indicates a more similar distribution of dominant characteristics between the two clients under the current operating condition. The pairwise correlation strengths between all clients are used as matrix elements, with rows and columns corresponding to each client, arranged in client number order. All elements on the main diagonal of the matrix are assigned a value of 1. This symmetric matrix represents the operating condition correlation strength between clients.

[0032] In one embodiment, the federated learning process fuses the condition-sensitive feature vector and condition attribute factors, analyzes the radioactivity values ​​of each At-211 separation system, and generates an activity condition vector representing the current condition category of each At-211 separation system. The process also includes: performing aggregation analysis on the condition correlation matrix to generate a federated consensus coefficient to guide federated optimization. Specifically, this involves: performing eigenvalue decomposition on the condition correlation matrix to obtain a set of eigenvalues, where the number of eigenvalues ​​equals the number of clients; using the largest eigenvalue as the overall consistency strength of all clients on the dominant condition; the order of the condition correlation matrix is ​​the total number of clients participating in federated learning; dividing the largest eigenvalue by the total number of clients and normalizing the result to the 0-1 range to generate a federated consensus coefficient to guide federated optimization. The closer the federated consensus coefficient is to 1, the more concentrated the condition correlation among clients, and the higher the overall condition consensus of the system; the closer the federated consensus coefficient is to 0, the more dispersed the conditions among clients, and the greater the differences.

[0033] The dynamic operating condition imprints of each client are fused with the federated consensus coefficients to generate an activity condition vector representing the current operating condition category of each At-211 separation system. Specifically, this involves: the server collecting dynamic operating condition imprints uploaded by all clients, calculating the arithmetic mean of the dynamic operating condition imprints of all clients according to their corresponding dimensions to obtain a global baseline imprint vector; for each client, subtracting its dynamic operating condition imprint vector from the global baseline imprint vector dimension-wise to obtain a difference vector, and taking the absolute value of each element of the difference vector to obtain an absolute difference vector, where each dimension of the absolute difference vector reflects the degree of deviation of the client from the global average level in that feature dimension; for each client, constructing a modulation coefficient vector, where each dimension of the modulation coefficient vector equals 1 minus the element-wise square of the corresponding dimension of the absolute difference vector; multiplying the client's dynamic operating condition imprint with the modulation coefficient vector element-wise to obtain a modulated feature vector; normalizing each element in the modulated feature vector to the 0-1 interval, and then this modulated feature vector becomes the activity condition vector representing the current operating condition category of each At-211 separation system.

[0034] By deeply integrating the condition-sensitive feature vector and condition attribute factors through federated learning, and combining the radioactivity values ​​of each At-211 separation system, dynamic condition imprints, condition correlation matrices, and federated consensus coefficients are generated sequentially, ultimately yielding an activity condition vector. This further uncovers effective process knowledge under similar activity conditions, clarifies the correlation strength and differences between the operating conditions of each system, and ensures that process parameters are adapted to the separation requirements of feed liquids with different activities.

[0035] In one embodiment, the activity condition vector is optimized to obtain the collaborative optimization coefficients used to guide the differentiated aggregation of federated learning. This includes: analyzing the federated aggregation process under different collaborative weights based on the activity condition vectors of each At-211 separation system, and generating a condition response sensitivity representing the sensitivity of each system to the collaborative weights. Specifically, this includes: setting a set of discrete collaborative weight values, from 0.1 to 1.0, with a step size of 0.1, for a total of ten collaborative weight values. For each client, this collaborative weight value is used as the aggregation weight for that client, while the remaining collaborative weight values ​​are equally distributed to all other clients, simulating ten different weighted aggregation scenarios. In each scenario, the weighted sum of the activity condition vectors of all clients is calculated to obtain the global aggregation vector corresponding to that scenario. For each client... The global aggregation vectors obtained under two adjacent collaborative weight values ​​(e.g., 0.1 and 0.2, 0.2 and 0.3, and so on) are subtracted dimension by dimension to obtain a global difference vector. The absolute values ​​of all elements of the global difference vector are then summed to obtain the change magnitude at that step size. This operation is repeated to obtain a total of nine change magnitudes. The maximum value among the nine change magnitudes of the client is selected as the response magnitude of the client to the change in collaborative weight. The response magnitude reflects the maximum global aggregation fluctuation that the client's operating characteristics may cause under the weight change. The response magnitude is normalized to the 0-1 interval to obtain the operating condition response sensitivity of the client. The larger the operating condition response sensitivity, the more sensitive the system's operating characteristics are to the change in collaborative weight, and more refined weight adjustment is required in federated differential aggregation.

[0036] In one embodiment, optimizing the activity condition vector to obtain co-optimization coefficients for guiding differentiated aggregation in federated learning further includes: constraining and optimizing the condition response sensitivity to generate co-optimization coefficients for guiding differentiated aggregation in federated learning. Specifically, this includes: calculating the sum of the condition response sensitivities of all clients; dividing the condition response sensitivity of each client by this sum to obtain the global sensitivity percentage of that client, where the global sensitivity percentage is between 0 and 1, and the sum of the percentages of all clients is 1; obtaining the allocatable communication bandwidth for all clients in the current batch, normalizing it to the 0-1 range to obtain the system resource constraint factor; and multiplying the global sensitivity percentage of each client by the system resource constraint factor to obtain the initial allocation coefficients. At this point, the sum of the initial allocation coefficients of all clients may not be equal to 1, requiring further adjustments. Adjustments were made, and a very small threshold of 0.01 was set. All initial allocation coefficients smaller than this threshold were forcibly set to 0.01 to ensure that each client receives at least a small aggregation weight. The sum of all initial allocation coefficients after adjustment was recalculated, and each initial allocation coefficient was divided by this sum to restore the sum of all initial allocation coefficients to 1. This process was repeated once to ensure the stability of the results. Finally, the initial allocation coefficient of each client is the co-optimization coefficient. The co-optimization coefficient is used to guide the differentiated aggregation in federated learning. The larger the co-optimization coefficient, the greater the influence of the client on the global parameters, so that the aggregated global parameters can more accurately adapt to the current batch of raw material activity conditions of each separation system, thereby improving the adaptability of process parameter distribution, effectively improving the separation effect of At-211 and impurities, and realizing the co-optimization of the performance of multiple separation systems.

[0037] By optimizing the activity condition vector, the following parameters are generated sequentially: condition response sensitivity, global sensitivity percentage, system resource constraint factor, and initial allocation coefficient. This yields collaborative optimization coefficients to guide differentiated aggregation in federated learning. This approach can accurately match the differences in operating conditions among various At-211 separation systems, enabling fine-tuning of aggregation weights. It fully leverages the process knowledge of similar activity conditions, avoids distributional out-generalization failures across activity migrations, ensures that the global parameters after aggregation are adapted to the separation requirements of feed liquids with different activities, improves the adaptability of process parameter distribution, and optimizes the performance coordination level of multiple At-211 separation systems.

[0038] In one embodiment, a performance collaborative optimization system for multiple separation systems based on federated learning is applied to the aforementioned optimization method. This system includes: an impact analysis unit, used to obtain the radioactivity value of the current batch of feed liquid in each At-211 separation system, analyze the radioactivity value based on a deep neural network, and generate a condition-sensitive feature vector representing the impact of different batches of radioactivity value on the performance of each separation system; a feature analysis unit, used to analyze the condition-sensitive feature vector based on the orthogonal decomposition neural network layer to obtain a condition attribute factor representing the significance of activity condition characteristics; a condition analysis unit, used to fuse the condition-sensitive feature vector and the condition attribute factor based on federated learning, and analyze the radioactivity value of each At-211 separation system to generate an activity condition vector representing the current condition category of each At-211 separation system; and a collaborative optimization unit, used to optimize the activity condition vector to obtain collaborative optimization coefficients used to guide differentiated aggregation in federated learning.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for collaborative performance optimization of multiple separate systems based on federated learning, characterized in that, include: Step S1: Obtain the radioactivity value of the current batch of feed liquid in each At-211 separation system, analyze the radioactivity value based on a deep neural network, and generate a condition-sensitive feature vector representing the impact of different batches of radioactivity value on the performance of each separation system; Step S2: Analyze the condition-sensitive feature vector according to the orthogonal decomposition neural network layer to obtain the condition attribute factor representing the significance of activity condition characteristics; Step S3: Fusion of the condition-sensitive feature vector and the condition attribute factor according to federated learning, and analyze the radioactivity value of each At-211 separation system to generate an activity condition vector representing the current condition category of each At-211 separation system; Step S4: Optimize the activity condition vector to obtain the collaborative optimization coefficient used to guide the differentiated aggregation of federated learning.

2. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 1, characterized in that, The analysis of radioactivity values ​​using deep neural networks generates condition-sensitive feature vectors representing the impact of different batches of radioactivity values ​​on the performance of each separation system. This includes: inputting the radioactivity values ​​of the current batch of feed liquid in each At-211 separation system into the deep neural network; extracting the dynamic change characteristics of the radioactivity values ​​based on its temporal attention mechanism; generating an activity perturbation vector representing the temporal fluctuation characteristics of the activity values; and analyzing the activity perturbation vector and the radioactivity values ​​to generate a performance impact vector representing the magnitude of the impact of the activity perturbation on each performance index.

3. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 2, characterized in that, The analysis of radioactivity values ​​based on deep neural networks generates condition-sensitive feature vectors representing the impact of different batches of radioactivity values ​​on the performance of each separation system. It also includes: analyzing the relationship between activity perturbation vectors and performance impact vectors to generate a condition impact factor representing the degree of dynamic interaction between activity and performance; and fusing radioactivity values, activity perturbation vectors, and condition impact factors to generate condition-sensitive feature vectors representing the impact of different batches of radioactivity values ​​on the performance of each separation system.

4. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 3, characterized in that, Based on the analysis of the working condition sensitive feature vector by the orthogonal decomposition neural network layer, the working condition attribute factors representing the significance of the activity working condition characteristics are obtained, including: orthogonally projecting the working condition sensitive feature vector to obtain the working condition projection coefficients representing the energy distribution in each orthogonal direction; transforming the working condition projection coefficients to generate the working condition activation response vector representing the characteristic state after nonlinear mapping; and calculating the working condition activation response vector through a multi-head self-attention mechanism to generate the working condition component contribution weights representing the importance of the working condition.

5. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 4, characterized in that, The analysis of the operating condition sensitive feature vector based on the orthogonal decomposition neural network layer yields the operating condition attribute factor representing the significance of activity operating condition characteristics. It also includes: fusing the operating condition projection coefficient with the operating condition component contribution weight to generate the operating condition ground state vector representing the main changes in the operating condition; and performing dominant feature analysis on the operating condition ground state vector to generate the operating condition attribute factor representing the significance of activity operating condition characteristics.

6. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 5, characterized in that, The federated learning process fuses condition-sensitive feature vectors and condition attribute factors, analyzes the radioactivity values ​​of each At-211 separation system, and generates an activity condition vector representing the current condition category of each At-211 separation system. This includes: interacting with the condition-sensitive feature vectors and condition attribute factors through the local fusion network of the federated learning client, and calculating the current batch radioactivity values ​​to generate a dynamic condition imprint representing the current condition status of each client; uploading the dynamic condition imprints of each client to the federated server, calculating the similarity between the dynamic condition imprints of different clients, and generating a condition association matrix representing the strength of condition association between clients.

7. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 6, characterized in that, The federated learning approach fuses condition-sensitive feature vectors and condition attribute factors, analyzes the radioactivity values ​​of each At-211 separation system, and generates an activity condition vector representing the current condition category of each At-211 separation system. It also includes: performing aggregation analysis on the condition correlation matrix to generate a federated consensus coefficient to guide federated optimization; and fusing the dynamic condition imprints of each client with the federated consensus coefficient to generate an activity condition vector representing the current condition category of each At-211 separation system.

8. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 7, characterized in that, The activity condition vectors are optimized to obtain the collaborative optimization coefficients used to guide the differentiated aggregation of federated learning. This includes: analyzing the federated aggregation process under different collaborative weights based on the activity condition vectors of each At-211 separation system, and generating condition response sensitivity that represents the sensitivity of each system to the collaborative weights.

9. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 8, characterized in that, The activity condition vector is optimized to obtain the co-optimization coefficients used to guide the differentiated aggregation of federated learning. It also includes: constraining the sensitivity of the condition response to generate the co-optimization coefficients used to guide the differentiated aggregation of federated learning.

10. A performance collaborative optimization system for multiple separate systems based on federated learning, applied in the optimization method described in any one of claims 1-9, characterized in that, include: The impact analysis unit is used to obtain the radioactivity value of the current batch of feed liquid in each At-211 separation system. Based on a deep neural network, the radioactivity value is analyzed to generate a condition-sensitive feature vector representing the impact of different batches of radioactivity value on the performance of each separation system. The feature analysis unit is used to analyze the condition-sensitive feature vectors based on the orthogonal decomposition neural network layers to obtain condition attribute factors that represent the significance of activity condition characteristics. The operating condition analysis unit is used to fuse the operating condition sensitive feature vector and operating condition attribute factors according to federated learning, and analyze the radioactivity values ​​of each At-211 separation system to generate an activity operating condition vector representing the current operating condition category of each At-211 separation system; the collaborative optimization unit is used to optimize the activity operating condition vector to obtain collaborative optimization coefficients used to guide the differentiated aggregation of federated learning.

Images