Federal learning-based multi-suite separation system performance collaborative optimization method and system

Abstract

The present application relates to the technical field of system optimization, and discloses a method and system for performance collaborative optimization of multiple sets of separation systems based on federated learning, which comprises the following steps: obtaining the radioactivity value of the current batch of raw material liquid of each set of At-211 separation system, analyzing the radioactivity value based on a deep neural network, and generating a working condition sensitive feature vector representing the influence of different batches of radioactivity value on the performance of each separation system; through the working condition sensitive feature vectors, working condition attribute factors, activity working condition vectors and collaborative optimization coefficients generated in each step, the influence of different batches of radioactivity value can be accurately mined, the correlation and difference of each system working condition can be determined, the fine adjustment of differentiated aggregation weight can be realized, the distribution out-generalization failure of cross-activity migration can be avoided, the process parameters can be ensured to adapt to the demand of different activity raw material liquid, the collaborative ability of multiple sets of At-211 separation system can be improved, and the separation efficiency and product purity can be improved.

Description

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 the present invention is achieved through the following technical solution:

[0006] A collaborative optimization method for the performance of multiple separation systems based on federated learning includes:

[0007] Step S1: Obtain the radioactivity value of the current batch of raw material 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.

[0008] Step S2: Analyze the working condition sensitive feature vectors based on the orthogonal decomposition neural network layer to obtain the working condition attribute factors that represent the significance of activity working condition characteristics.

[0009] Step S3: Based on federated learning, the working condition sensitive feature vector and working condition attribute factor are fused, and the radioactivity values ​​of each At-211 separation system are analyzed to generate an activity working condition vector representing the current working condition category of each At-211 separation system.

[0010] Step S4: Optimize the activity condition vector to obtain the collaborative optimization coefficients used to guide the differentiated aggregation of federated learning.

[0011] Furthermore, based on the analysis of radioactivity values ​​using a deep neural network, condition-sensitive feature vectors representing the impact of different batches of radioactivity values ​​on the performance of each separation system are generated, including:

[0012] The radioactivity values ​​of the current batch of raw material liquid in each At-211 separation system are input into a deep neural network. Based on its temporal attention mechanism, the dynamic change characteristics of the radioactivity values ​​are extracted to generate an activity perturbation vector representing the temporal fluctuation characteristics of the activity values.

[0013] The activity perturbation vector and radioactivity value are analyzed to generate a performance influence vector representing the magnitude of the influence of the activity perturbation on each performance index.

[0014] Furthermore, based on the analysis of radioactivity values ​​using deep neural networks, condition-sensitive feature vectors representing the impact of different batches of radioactivity values ​​on the performance of each separation system are generated, including:

[0015] The relationship between the activity perturbation vector and the performance impact vector is analyzed to generate a condition impact factor representing the degree of dynamic interaction between activity and performance;

[0016] By fusing radioactivity values, activity perturbation vectors, and operating condition influence factors, an operating condition sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system is generated.

[0017] Furthermore, by analyzing the condition-sensitive feature vectors based on the orthogonal decomposition neural network layers, condition attribute factors representing the significance of activity condition characteristics are obtained, including:

[0018] Orthogonally project the condition-sensitive feature vectors to obtain the condition projection coefficients representing the energy distribution in each orthogonal direction;

[0019] The working condition projection coefficients are transformed to generate a working condition activation response vector representing the characteristic state after nonlinear mapping.

[0020] The activation response vector of the working condition is calculated by a multi-head self-attention mechanism, and the contribution weight of the working condition component representing the importance of the working condition is generated.

[0021] Furthermore, based on the analysis of the operating condition sensitive feature vectors using the orthogonal decomposition neural network layers, operating condition attribute factors representing the significance of activity operating condition characteristics are obtained, including:

[0022] The working condition projection coefficients and the working condition component contribution weights are fused to generate the working condition ground state vector representing the main changes in the working condition.

[0023] Dominant feature analysis is performed on the ground state vector of the operating condition to generate operating condition attribute factors that represent the significance of activity operating condition characteristics.

[0024] Furthermore, by fusing the condition-sensitive feature vectors and condition attribute factors using federated learning, and analyzing the radioactivity values ​​of each At-211 separation system, an activity condition vector representing the current condition category of each At-211 separation system is generated, including:

[0025] The local fusion network of the federated learning client interacts with the condition-sensitive feature vector and condition attribute factor, and calculates the dynamic condition imprint representing the current condition status of each client by combining the current batch radioactivity value.

[0026] The dynamic operating condition imprints of each client are uploaded to the federated server, and the similarity between the dynamic operating condition imprints of different clients is calculated to generate an operating condition association matrix that represents the strength of the operating condition association between each client.

[0027] 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:

[0028] Aggregate analysis is performed on the working condition correlation matrix to generate federated consensus coefficients to guide federated optimization;

[0029] By fusing the dynamic operating condition imprints of each client with the federated consensus coefficient, an activity condition vector representing the current operating condition category of each At-211 separation system is generated.

[0030] Furthermore, the activity condition vector is optimized to obtain collaborative optimization coefficients used to guide differentiated aggregation in federated learning, including:

[0031] Based on the activity condition vectors of each At-211 separation system, the federated aggregation process under different cooperative weights is analyzed, and a condition response sensitivity representing the sensitivity of each system to the cooperative weights is generated.

[0032] Furthermore, the activity condition vector is optimized to obtain collaborative optimization coefficients used to guide differentiated aggregation in federated learning, which also include:

[0033] Constraint optimization is performed on the sensitivity to operating conditions to generate collaborative optimization coefficients to guide differentiated aggregation in federated learning.

[0034] Furthermore, a performance collaborative optimization system for multiple separate systems based on federated learning, applied to the aforementioned optimization method, includes:

[0035] 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.

[0036] 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.

[0037] The operating condition analysis unit is used to fuse the operating condition sensitive feature vector and operating condition attribute factor according to federated learning, and analyze the radioactivity value 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.

[0038] The collaborative optimization unit is used to optimize the activity condition vector to obtain collaborative optimization coefficients that guide the differentiated aggregation of federated learning.

[0039] In summary, the present invention has the following main beneficial effects:

[0040] By generating condition-sensitive feature vectors, which accurately represent the impact of different batches of radioactivity values ​​on the performance of each separation system, the effective process knowledge under similar activity conditions is fully explored, breaking the limitations of existing static polymerization strategies. By generating condition attribute factors, which highlight the significance of activity condition characteristics, the core differences between different activity conditions are accurately identified, avoiding polymerization deviations caused by unclear condition characteristics. By generating activity condition vectors, which clearly represent the current condition category of each At-211 separation system, the correlation strength and individual differences between systems are clarified, facilitating the later implementation of differentiated polymerization weight adjustment in federated learning and effectively avoiding distributional out-generalization failure during cross-activity migration. By generating collaborative optimization coefficients, the federated learning can be guided to achieve fine-grained adjustment of differentiated polymerization weights, ensuring that the issued process parameters accurately adapt to the separation requirements of feed liquids with different activities, 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

[0041] Figure 1 This is a flowchart illustrating the steps of the performance collaborative optimization method for multiple separation systems based on federated learning, as described in this invention.

[0042] Figure 2 This is a schematic diagram of the performance collaborative optimization system for multiple separation systems based on federated learning, as presented in this invention. Detailed Implementation

[0043] 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.

[0044] refer to Figure 1 and Figure 2 A collaborative optimization method for the performance of multiple separation systems based on federated learning includes:

[0045] Step S1: Obtain the radioactivity value of the current batch of raw material 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.

[0046] Step S2: Analyze the working condition sensitive feature vectors based on the orthogonal decomposition neural network layer to obtain the working condition attribute factors that represent the significance of activity working condition characteristics.

[0047] Step S3: Based on federated learning, the working condition sensitive feature vector and working condition attribute factor are fused, and the radioactivity values ​​of each At-211 separation system are analyzed to generate an activity working condition vector representing the current working condition category of each At-211 separation system.

[0048] Step S4: Optimize the activity condition vector to obtain the collaborative optimization coefficients used to guide the differentiated aggregation of federated learning.

[0049] In one embodiment, the radioactivity values ​​are analyzed based on a deep neural network to generate a condition-sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system, including:

[0050] The radioactivity values ​​of the current batch of feed liquid from each At-211 separation system are input into a deep neural network. Based on its temporal attention mechanism, the dynamic change features of the radioactivity values ​​are extracted to generate an activity perturbation vector representing the temporal fluctuation characteristics of the activity values. Specifically, the radioactivity values ​​of the current batch of feed liquid from each At-211 separation system are arranged in the order of collection time to form a temporal activity data matrix. The temporal activity data matrix is ​​input into a one-dimensional convolutional layer of the deep neural network to extract the local fluctuation features of the radioactivity values ​​at continuous time points and generate a temporal feature map. The temporal feature map is a three-dimensional feature map with the dimensions of batch number, time step, and number of feature channels.

[0051] The temporal feature map is input into the dynamic temporal attention module of the deep neural network. This module first calculates the coefficient of variation of the radioactivity value at each time step as the initial attention score.

[0052] Then, the temporal feature map is flattened along the feature channel dimension and input into a gated recurrent unit. The gated recurrent unit processes the data in the order of time steps, and outputs 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 and mapped to a correction coefficient vector with the same dimension as the number of feature channels. The 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 on each feature channel is 1, resulting in the attention weight matrix, whose dimensions are the time step and the number of feature channels.

[0053] The attention weight matrix is ​​multiplied element-wise with the temporal feature map, that is, the feature value of each feature channel at each time step is multiplied by the corresponding modified attention weight to obtain a weighted feature map. The dimension of the weighted feature map remains unchanged. The weighted feature map is then subjected to global average pooling along the new time step dimension, and an average value is obtained for each feature channel to form a one-dimensional feature vector with the dimension being the number of feature channels. The one-dimensional feature vector is then input into a fully connected layer for nonlinear transformation to generate an activity perturbation vector representing the temporal fluctuation characteristics of the activity value.

[0054] The activity perturbation vector and radioactivity value are analyzed to generate a performance impact vector representing the magnitude of the influence of the activity perturbation on each performance index. Specifically, this includes: adding 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 of N consecutive batches before the current batch, where the performance index data includes at least single batch yield, total separation time, and product concentration; using the sequence of each dimension of the joint input vector composed of N consecutive batches as independent variables and each performance index as dependent variable, calculating the Pearson correlation coefficient to obtain an initial correlation matrix. The number of rows in this matrix is ​​equal to the number of dimensions of the joint input vector, and the number of columns is equal to 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.

[0055] Calculate the mean value of the current radioactivity of each At-211 separation system, then compare the radioactivity of the current batch of raw material liquid with the mean value, calculate the percentage deviation between the two, normalize the percentage deviation to the 0-1 range, and obtain the activity reference adjustment coefficient.

[0056] Each element in the initial correlation matrix is ​​multiplied element-wise with the activity baseline 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.

[0057] 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. The method further includes:

[0058] The relationship between the activity perturbation vector and the performance impact vector is analyzed to generate a condition-related impact factor representing the degree of dynamic interaction between activity and performance. Specifically, this involves: obtaining the activity perturbation vector and performance impact vector for the N consecutive batches preceding the current batch; 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 its 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 taking the largest singular value as the basic interaction strength, which reflects the strongest interaction strength between activity perturbation and performance impact in historical data.

[0059] Calculate the mean of each column of 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 from the historical average pattern in terms of both activity perturbation and performance response.

[0060] 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 from smallest to largest, determine the percentile position of the current batch's 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.

[0061] 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 includes: 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; and multiplying the radioactivity value element-wise with the coefficient of variation vector to generate an activity modulation vector with the same dimensions as the activity perturbation vector.

[0062] The activity modulation vector is multiplied element by element by the operating condition influence factor to obtain the operating condition weighted modulation vector.

[0063] The operating condition weighted modulation vector and the activity perturbation vector are added element by element to obtain a new vector with the same dimension as the activity perturbation vector. This new vector is then normalized using the L2 norm so that the square root of the sum of the squares of each element is equal to 1, thereby generating an operating condition sensitive feature vector that represents the impact of different batches of radioactivity values ​​on the performance of each separation system.

[0064] 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.

[0065] In one embodiment, the working condition sensitive feature vector is analyzed based on the orthogonal decomposition neural network layer to obtain working condition attribute factors representing the significance of activity working condition characteristics, including:

[0066] Orthogonally project the condition-sensitive feature vectors to obtain the condition projection coefficients representing the energy distribution in each orthogonal direction. Specifically, this includes: collecting condition-sensitive feature vectors from all batches before the current batch, arranging these condition-sensitive feature vectors in batch order to construct a historical condition feature matrix. The rows of the historical condition feature matrix correspond to different batches, and the columns correspond to the dimensions of the condition-sensitive feature vectors. Perform singular value decomposition on the historical condition feature matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Simultaneously, calculate the sum of all singular values, sort the singular values ​​from largest to smallest, and accumulate them sequentially. When the accumulated sum first reaches 90% of the sum of singular values, record the number of accumulated singular values ​​at this time, denoted as m. Extract the column vectors in the right singular vector matrix corresponding to these m largest singular values. These column vectors are the initial orthogonal basis vectors. Arrange these m column vectors in the order of extraction to form the initial orthogonal basis matrix. Each column of the initial orthogonal basis matrix is ​​an initial orthogonal basis vector.

[0067] 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 with the condition-sensitive feature vector to obtain a set of values. Each value represents the projection length of the condition-sensitive feature vector in the 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.

[0068] Squaring each element in the projection coefficient vector yields the energy value in each orthogonal direction. Normalizing the energy value to the 0-1 interval gives the working condition projection coefficient representing the energy distribution in each orthogonal direction.

[0069] The working condition projection coefficients are transformed to generate a working condition activation response vector representing the characteristic state after nonlinear mapping. Specifically, this includes: collecting the working condition projection coefficients of all batches before the current batch, sorting these working condition projection coefficients in ascending order of value, dividing the sorted sequence into ten equal-length intervals, each interval containing the same number of working condition projection coefficients, recording the lower limit and upper limit of each interval, and obtaining a total of ten continuous and non-overlapping projection intervals.

[0070] Compare the current batch of working condition projection coefficients with ten projection intervals, find the projection interval to which the working condition projection coefficient belongs, calculate the difference between the current working condition projection coefficient and the lower limit of the projection interval, and then divide it by the interval length of the projection interval to obtain the relative position coefficient between 0 and 1.

[0071] Construct a ten-dimensional vector with an initial value of zero. 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 projection coefficient of the current working condition belongs, and keep the other nine dimensions as 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.

[0072] The working condition activation response vector is calculated through a multi-head self-attention mechanism to generate working condition component contribution weights that represent the importance of the working condition. Specifically, the number of attention heads is set to four, and the working condition activation response vector is mapped into four sub-vectors of equal dimensions through linear projection. Each sub-vector corresponds to the input of an attention head.

[0073] Within each attention head, a linear transformation is performed on each dimension of its sub-vector to generate two intermediate variables for that dimension, denoted as the query variable and the key variable, respectively. For any two dimensions in the sub-vector, the query variable and the key variable are multiplied to obtain the initial association scores for these two dimensions. The initial association scores are calculated pairwise for all dimensions within the sub-vector to obtain the association matrix. The number of rows and columns of the association matrix is ​​equal to the number of dimensions processed by the attention head.

[0074] The correlation matrix of each attention head is summed row by row to obtain the total correlation vector of each dimension with all other dimensions. Each element in the total correlation vector is divided 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. The initial contribution weights obtained from the four attention heads are then normalized to the interval between 0 and 1 to obtain the contribution weight of the condition component representing the importance of the condition.

[0075] In one embodiment, the analysis of 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 further includes:

[0076] The operating condition projection coefficient and the operating condition component contribution weight are fused to generate the 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 weight to obtain the energy important composite value; obtaining the operating condition projection coefficients of all batches before the current batch; calculating the average value of these operating condition projection coefficients; collecting all operating condition component contribution weights of the corresponding batch; calculating the average value of these operating condition component contribution weights; and multiplying the two average values ​​to obtain the historical average composite value.

[0077] Calculate the difference between the important composite value of energy and the historical average composite value. Divide the difference by the absolute value of the historical average composite value to obtain the relative rate of change. Construct a two-dimensional vector. Assign the projection coefficient of the current operating condition to the first dimension and the relative rate of change to the second dimension. Normalize the values ​​of the two-dimensional vector to the 0-1 interval. This is the operating condition ground state vector representing the main changes in the operating condition.

[0078] Dominant feature analysis is performed on the ground state vector of the operating condition to generate operating condition attribute factors that represent 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; and 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.

[0079] The larger of the operating condition proportion coefficient and the change rate proportion coefficient is used as the dominant factor to construct a two-dimensional vector. The first dimension is assigned the dominant factor, and the second dimension is assigned according to the dimension type. If the dominant factor comes from the operating condition proportion coefficient, it is assigned a value of 0, and if it comes from the change rate proportion coefficient, it is assigned a value of 1.

[0080] The modulus of a two-dimensional vector is calculated by taking the square root of the sum of the squares of the values ​​in the two dimensions and normalizing the result to the 0-1 range. This result is the condition attribute factor that represents the significance of the activity condition characteristics. The larger the condition attribute factor, the more significant the dominant characteristics of the condition.

[0081] 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.

[0082] In one embodiment, the operating condition sensitive feature vector and operating condition attribute factor are fused according to federated learning, and the radioactivity values ​​of each At-211 separation system are analyzed to generate an activity condition vector representing the current operating condition category of each At-211 separation system, including:

[0083] The local fusion network of the federated learning client interacts with the condition-sensitive feature vector and condition attribute factors, and calculates them in conjunction with the current batch radioactivity value 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 value are input into the local fusion network of each client. Locally on the client, each dimension of the condition-sensitive feature vector is multiplied by the condition attribute factor to obtain a weighted feature vector. At the same time, the current batch radioactivity value is multiplied element-wise by the weighted feature vector to obtain a modulation feature vector, thereby realizing the joint modulation of the original features by the attribute factors and the activity level.

[0084] The modulation feature vector is input into a fully connected layer with the number of neurons set to half 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. Finally, this vector is added element by element to the dimension-reduced feature vector to obtain the fused associated feature vector.

[0085] 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, resulting in a gating-adjusted vector. Each element in this vector is normalized to the 0-1 interval. The normalized vector is the dynamic operating condition imprint representing the current operating status of each client.

[0086] The dynamic working condition imprints of each client are uploaded to the federated server, and the similarity between the dynamic working condition imprints of different clients is calculated to generate a working condition association matrix representing the strength of the working condition association between clients. Specifically, under the federated learning framework, each client uploads the locally generated dynamic working condition imprint to the federated server. After receiving the dynamic working condition imprints of all clients, the server calculates the arithmetic mean of all elements of each client's dynamic working condition imprint and uses this mean as the baseline threshold of the dynamic working condition imprint. The baseline threshold reflects the average activation level of each dimension feature under the current working condition of the client.

[0087] For each client's dynamic operating 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, which represents the distribution of the dominant dimensions above average in the client's operating condition characteristics in the form of 0 and 1.

[0088] For any two clients, count the number of dimensions that are both 1 and the number of dimensions that are at least one 1 in the two binary pattern vectors. Divide the number of the former dimension by the number of the latter dimension and normalize the result to the 0-1 interval. This is the working condition correlation strength between the two clients. The greater the working condition correlation strength, the more similar the dominant feature distributions of the two clients are under the current working condition.

[0089] The working condition correlation strength between all clients is used as matrix elements. The rows and columns of the matrix correspond to each client and are arranged in the order of client number. All elements on the main diagonal of the matrix are assigned a value of 1. This symmetric matrix is ​​the working condition correlation matrix that represents the working condition correlation strength between each client.

[0090] In one embodiment, the method further includes fusing the condition-sensitive feature vector and condition attribute factors according to 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.

[0091] Aggregate analysis is performed on the working condition correlation matrix to generate federated consensus coefficients to guide federated optimization. Specifically, this includes: performing eigenvalue decomposition on the working condition correlation matrix to obtain a set of eigenvalues, the number of which is equal to the number of clients, and taking the largest eigenvalue as the overall consistency strength of all clients on the dominant working condition.

[0092] The order of the working condition correlation matrix is ​​the total number of clients participating in federated learning. The largest eigenvalue is divided by the total number of clients, and the result is normalized to the 0-1 interval to generate the federated consensus coefficient to guide federated optimization. The closer the federated consensus coefficient is to 1, the more concentrated the working condition correlation of each client is, and the higher the working condition consensus of the whole system is. The closer the federated consensus coefficient is to 0, the more dispersed the working conditions are among the clients and the greater the differences.

[0093] 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 the 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 by dimension to obtain a difference vector, and taking the absolute value of each element of the difference vector to obtain an absolute difference vector. Each dimension of the absolute difference vector reflects the degree of deviation of the client from the global average level in that feature dimension.

[0094] For each client, a modulation coefficient vector is constructed. Each dimension of the modulation coefficient vector is equal to 1 minus the element-wise square of the corresponding dimension of the absolute difference vector. The dynamic condition imprint of the client is multiplied element-wise with the modulation coefficient vector to obtain the modulated feature vector.

[0095] Each element in the modulated feature vector is normalized to the 0-1 interval, and then the modulated feature vector is the activity condition vector representing the current operating condition category of each At-211 separation system.

[0096] 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.

[0097] In one embodiment, the activity condition vector is optimized to obtain co-optimization coefficients used to guide differentiated aggregation in federated learning, including:

[0098] Based on the activity condition vectors of each At-211 separation system, the federated aggregation process under different cooperative weights is analyzed, and a condition response sensitivity representing the sensitivity of each system to the cooperative weights is generated. Specifically, a set of discrete cooperative weight values ​​is set, from 0.1 to 1.0, with a step size of 0.1, for a total of ten cooperative weight values. For each client, the cooperative weight value is used as the aggregation weight of that client in turn, and the remaining cooperative weight values ​​are evenly distributed to all other clients. Ten different weighted aggregation scenarios are simulated. 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.

[0099] 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.

[0100] The maximum value among the nine change values ​​of the client is selected as the response magnitude of the client to the change of the collaborative weight. The response magnitude reflects the maximum global aggregation fluctuation that the client's operating characteristics may cause under the weight change.

[0101] By normalizing the response amplitude to the 0-1 range, the sensitivity of the working condition response to the corresponding client can be obtained. The higher the sensitivity of the working condition response, the more sensitive the working condition characteristics of the system are to changes in the collaborative weights. In federated differential aggregation, more refined weight adjustment is required.

[0102] In one embodiment, optimizing the activity condition vector to obtain collaborative optimization coefficients for guiding federated learning differential aggregation further includes:

[0103] Constraint optimization of operating condition response sensitivity is performed to generate collaborative optimization coefficients to guide differentiated aggregation in federated learning. Specifically, this includes: calculating the sum of the operating condition response sensitivity of all clients, dividing the operating condition response sensitivity of each client by this sum to obtain the global sensitivity percentage of that client. The global sensitivity percentage is between 0 and 1, and the sum of the percentages of all clients is 1.

[0104] Obtain the allocable communication bandwidth for all clients in the current batch, normalize it to the 0-1 range to obtain the system resource constraint factor; multiply the global sensitivity ratio of each client by the system resource constraint factor to obtain the initial allocation coefficient. At this point, the sum of the initial allocation coefficients of all clients may not be equal to 1, so it needs to be adjusted. Then, a very small threshold of 0.01 is set, and all initial allocation coefficients less than this threshold are forced to be set to 0.01 to ensure that each client receives at least a small aggregate weight. Recalculate the sum of all the adjusted initial allocation coefficients, divide each initial allocation coefficient by this sum, and restore the sum of all initial allocation coefficients to 1. Repeat this process once to ensure the stability of the result.

[0105] Ultimately, the initial allocation coefficient for each client is the collaborative optimization coefficient. The collaborative optimization coefficient is used to guide the differentiated aggregation in federated learning. The larger the collaborative optimization coefficient, the greater the impact 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 the process parameters and effectively improving the separation effect of At-211 and impurities, thus realizing the collaborative optimization of the performance of multiple separation systems.

[0106] 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.

[0107] In one embodiment, a performance collaborative optimization system for multiple separate systems based on federated learning is applied to the above optimization method, including:

[0108] 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.

[0109] 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.

[0110] The operating condition analysis unit is used to fuse the operating condition sensitive feature vector and operating condition attribute factor according to federated learning, and analyze the radioactivity value 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.

[0111] The collaborative optimization unit is used to optimize the activity condition vector to obtain collaborative optimization coefficients that guide the differentiated aggregation of federated learning.

[0112] 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 using a deep neural network to generate a condition-sensitive feature vector representing the impact of different batch radioactivity values ​​on the performance of each separation system, including: The radioactivity values ​​of the current batch of raw material liquid in each At-211 separation system are input into a deep neural network. Based on its temporal attention mechanism, the dynamic change characteristics of the radioactivity values ​​are extracted to generate an activity perturbation vector representing the temporal fluctuation characteristics of the activity values. The activity perturbation vector and radioactivity value are analyzed to generate a performance impact vector representing the magnitude of the influence of the activity perturbation on each performance index. The relationship between the activity perturbation vector and the performance impact vector is analyzed to generate a condition impact factor representing the degree of dynamic interaction between activity and performance; By fusing radioactivity values, activity perturbation vectors, and operating condition influence factors, an operating condition sensitive feature vector representing the impact of different batches of radioactivity values ​​on the performance of each separation system is generated. Step S2 involves analyzing the condition-sensitive feature vectors based on the orthogonal decomposition neural network layers to obtain condition attribute factors representing the significance of activity condition characteristics, including: Orthogonally project the condition-sensitive feature vectors to obtain the condition projection coefficients representing the energy distribution in each orthogonal direction; The working condition projection coefficients are transformed to generate a working condition activation response vector representing the characteristic state after nonlinear mapping. The working condition activation response vector is calculated through a multi-head self-attention mechanism to generate the working condition component contribution weights representing the importance of the working condition; Step S3: Based on federated learning, the working condition sensitive feature vector and working condition attribute factor are fused, and the radioactivity values ​​of each At-211 separation system are analyzed to generate an activity working condition vector representing the current working condition category of each At-211 separation system. Step S4: Optimize the activity condition vector to obtain the collaborative optimization coefficients 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, Based on the analysis of the operating condition sensitive feature vectors using the orthogonal decomposition neural network layers, operating condition attribute factors representing the significance of activity operating condition characteristics are obtained, including: The working condition projection coefficients and the working condition component contribution weights are fused to generate the working condition ground state vector representing the main changes in the working condition. Dominant feature analysis is performed on the ground state vector of the operating condition to generate operating condition attribute factors that represent the significance of activity operating condition characteristics.

3. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 2, characterized in that, Federated learning is used to fuse condition-sensitive feature vectors and condition attribute factors, 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, including: The local fusion network of the federated learning client interacts with the condition-sensitive feature vector and condition attribute factor, and calculates the dynamic condition imprint representing the current condition status of each client by combining the current batch radioactivity value. The dynamic operating condition imprints of each client are uploaded to the federated server, and the similarity between the dynamic operating condition imprints of different clients is calculated to generate an operating condition association matrix that represents the strength of the operating condition association between each client.

4. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 3, characterized in that, Federated learning is used to fuse condition-sensitive feature vectors and condition attribute factors, 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 also includes: Aggregate analysis is performed on the working condition correlation matrix to generate federated consensus coefficients to guide federated optimization; By fusing the dynamic operating condition imprints of each client with the federated consensus coefficient, an activity condition vector representing the current operating condition category of each At-211 separation system is generated.

5. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 4, characterized in that, Optimizing the activity condition vector yields collaborative optimization coefficients used to guide differentiated aggregation in federated learning, including: Based on the activity condition vectors of each At-211 separation system, the federated aggregation process under different cooperative weights is analyzed, and a condition response sensitivity representing the sensitivity of each system to the cooperative weights is generated.

6. The method for collaborative performance optimization of multiple separation systems based on federated learning according to claim 5, characterized in that, Optimizing the activity condition vector yields collaborative optimization coefficients used to guide differentiated aggregation in federated learning, and also includes: Constraint optimization is performed on the sensitivity to operating conditions to generate collaborative optimization coefficients to guide differentiated aggregation in federated learning.

7. A performance collaborative optimization system for multiple separate systems based on federated learning, applied to the optimization method described in any one of claims 1-6, 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 factor according to federated learning, and analyze the radioactivity value 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 condition vector to obtain collaborative optimization coefficients that guide the differentiated aggregation of federated learning.

Images