Storage battery real-time monitoring method based on edge calculation

Abstract

The invention discloses a storage battery real-time monitoring method based on edge calculation, and the method comprises the following steps: S1, collecting multi-source operation parameters of a storage battery, and constructing an operation parameter set; s2, processing the operation parameter set to generate a standardized state vector; s3, on the basis of a micro neural network model, generating graph structure reasoning characteristics are fused; s4, performing global search on a connection structure between the sub-network clusters by adopting a spider-now-monkey optimization algorithm, and updating the structure configuration of the micro neural network model; s5, on the basis of the graph structure reasoning features and the optimized sparse connection structure, forming a monitoring judgment result; s6, performing edge response control according to a monitoring judgment result, and generating a response record; and S7, based on the response record and the monitoring determination result, completing edge side dynamic optimization of the miniature neural network model. Sparse neural clusters and edge reasoning are adopted, and real-time monitoring and dynamic optimization control of the state of the storage battery are achieved.

Description

Technical Field

[0001] The present invention relates to the field of battery intelligent monitoring and edge computing technology, and in particular to a real-time battery monitoring method based on edge computing. Background Art

[0002] With the rapid development of energy storage systems, new energy vehicles, and distributed power supply equipment, batteries, as core energy carriers, have become key factors in ensuring system reliability due to their operational safety and state stability. In order to ensure stable battery performance, extend service life, and reduce operational risks, the industry has gradually focused on real-time monitoring and intelligent diagnosis of battery operating status. In existing technologies, common battery monitoring methods typically use a centralized architecture. By collecting operating parameters such as voltage, current, and temperature in the battery pack, the data is uploaded to a central server or cloud platform, and the central algorithm model completes status assessment and risk warning. However, this type of centralized processing method has certain bandwidth dependence, response delays, and fault isolation difficulties, which are not conducive to rapid on-site response and independent operation in edge deployment scenarios.

[0003] Some studies have begun deploying lightweight models on the edge nodes of battery management systems (BMS) to achieve local state perception and basic alarm functions. Despite this, current edge inference structures are mostly compressed fully connected networks or statically configured shallow models with limited expressive power, making it difficult to fully explore the nonlinear relationships and multidimensional coupling characteristics between operating parameters. In addition, to achieve model lightweighting, such methods often fix the network structure or use manually set sub-network partitioning strategies, lacking a structural adaptive optimization mechanism, resulting in unstable model performance under different operating conditions. In particular, when the battery operating state is affected by ambient temperature, current fluctuations, or aging trends, the fixed model structure may not be able to flexibly adapt to changes in feature distribution, thereby reducing state recognition accuracy.

[0004] Existing methods also generally ignore the structured correlations between operating parameters. In real-world scenarios, battery cells exhibit certain thermal and electrical coupling characteristics, and operating parameters often have a non-uniform distribution in physical space or electrical structure. Traditional methods often directly input all parameter vectors into the model without performing structural partitioning or relationship modeling, making it difficult to extract more discriminative regional features from the input data. Furthermore, when the connection structure is fixed, the information transmission path between different parameter subsets cannot be adjusted, limiting the model's ability to integrate multi-source state information.

[0005] Furthermore, in terms of neural network architecture design, there is currently a lack of structural search mechanisms tailored to edge deployment constraints. Some research has introduced swarm intelligence algorithms such as genetic algorithms and particle swarm optimization for structural optimization, but their adaptability to edge scenarios remains insufficient, and update strategies struggle to balance global search capabilities with local sensitivity adjustments. In actual deployments, battery system states are constantly changing, and the connection structure must be able to dynamically adjust based on field monitoring feedback to ensure the long-term effectiveness of the model.

[0006] In summary, existing battery monitoring technologies still have shortcomings in edge deployment, adaptive modeling, and state determination and response. Specifically, the model structure lacks inter-parameter partitioning logic and coupling modeling, lacks a sparse connection structure generation mechanism for the joint optimization of inference performance and computational constraints, and fails to effectively introduce a dynamic feedback-driven structural evolution strategy.

[0007] Therefore, how to provide a real-time battery monitoring method based on edge computing is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a real-time monitoring method for batteries based on edge computing. The present invention introduces a sparsely connected neural cluster structure micro neural network, and combines the spider monkey optimization algorithm to dynamically optimize the sparse connection structure between neural clusters. It describes in detail the entire process of multi-source operating parameter collection, graph structure reasoning feature construction, health status scoring and abnormal risk judgment, edge response control and model adaptive update, and has the advantages of lightweight deployment, fast response, high state recognition accuracy and sustainable structural optimization.

[0009] A real-time battery monitoring method based on edge computing according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect multi-source operating parameters of the battery and build an operating parameter set;

[0011] S2. Time synchronization and normalization are performed on the set of operating parameters to generate a standardized state vector;

[0012] S3. Input the standardized state vector into a micro-neural network model deployed on an edge computing node. The micro-neural network model adopts a sparsely connected neural cluster structure, which includes multiple sub-network clusters. The sparse connection mechanism is used for cross-cluster information interaction to generate graph structure reasoning features.

[0013] S4. Structural optimization of the sparse connection mechanism is performed, and the spider monkey optimization algorithm is used to perform a global search on the connection structure between sub-network clusters to update the structural configuration of the micro neural network model;

[0014] S5. Based on the graph structure reasoning features and the optimized sparse connection structure, the health status score and abnormal risk label are output to form the monitoring judgment result;

[0015] S6. Perform edge response control based on the monitoring judgment results. When the health status score is lower than the preset score threshold or the confidence of the abnormal risk label is higher than the preset risk threshold, generate a response record;

[0016] S7. Based on the response records and monitoring judgment results, send a model update request to the cloud platform according to the set cycle or trigger conditions, receive the neural cluster parameter or structure adjustment instructions issued by the cloud platform, and complete the edge-side dynamic optimization of the micro neural network model.

[0017] Optionally, the operating parameter set includes voltage parameters, current parameters, state of charge parameters, cell internal resistance parameters, temperature parameters, power and load parameters, discharge rate parameters and environmental parameters;

[0018] The voltage parameters include the voltage of a single cell, the total system voltage and the voltage fluctuation amplitude; the current parameters include the real-time operating current and the current change rate; the state of charge parameters include the current state of charge and the remaining power; the cell internal resistance parameters include the AC impedance of the single cell and the maximum and minimum difference between the internal resistance; the temperature parameters include the cell temperature, the casing temperature and the ambient temperature; the power and load parameters include the real-time load power and the power change rate; the discharge rate parameters include the current discharge rate, the maximum historical discharge rate and the rated rate ratio; the environmental parameters include the ambient humidity, air pressure and the external temperature fluctuation amplitude.

[0019] Optionally, the S3 specifically includes:

[0020] S31, normalize the state vector X=[x1,x2,...,x n ] is input to the micro neural network model deployed on the edge computing node, and the standardized state vector is used to describe the multi-source operating state of the battery at the current sampling moment;

[0021] S32. Based on the physical properties and functional classification of the operating parameters, the standardized state vector is divided into five input sub-vectors, corresponding to the voltage, current, thermal, capacity, and environmental feature subsets, respectively. Each input sub-vector is input into a corresponding sub-network cluster to form a sub-network cluster input set.

[0022] S33. Each sub-network cluster performs local feedforward inference on the input sub-vector, calculates the local state expression, and outputs a set of local state expressions:

[0023] H (j) =max(0,W (j) X(j) +b (j) );

[0024] Among them, H (j) is the local state expression vector of the jth sub-network cluster, W (j) is the weight matrix of the j-th sub-network cluster, X (j) is the jth feature subset of the input normalized state vector, b (j) is the bias vector corresponding to the j-th sub-network cluster, and j is the index number of the sub-network cluster;

[0025] S34, constructing a sparsely connected neural cluster structure for simulating the coupling relationship between different physical dimensions of the battery, setting a sparse connection control matrix M for limiting the information interaction relationship between each sub-network cluster, and outputting the sparse connection structure;

[0026] S35. Based on the local state expression set and sparse connection structure, perform cross-cluster feature fusion operation to generate graph structure reasoning features:

[0027]

[0028] Among them, H fusion is the graph structure reasoning feature, i is the index number of the sub-network cluster, M ij is the element in the sparse connection control matrix, H (i) is the local state expression of the ith sub-network cluster, and f(·) is the cross-cluster feature interaction function.

[0029] Optionally, the subnetwork cluster includes an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer receives an input sub-vector, and the input sub-vector corresponds to one of the voltage, current, heat, capacity or environment feature subsets; the first hidden layer is a fully connected neural network layer, which performs a first-layer linear transformation and nonlinear activation, and outputs a first-layer hidden state; the second hidden layer is a fully connected neural network layer, which receives the first-layer hidden state and outputs a local state expression.

[0030] Optionally, the sparse connection structure includes initializing a sparse connection control matrix M∈{0,1} to characterize the coupling relationship between different operating parameter subspaces of the battery 5×5 , where M ij Indicates whether the local state expression of the i-th sub-network cluster is allowed to be transferred to the j-th sub-network cluster to participate in feature fusion. If M ij =1, the transfer is allowed. If M ij = 0, transfer is not allowed. ij = 1, the output features of the i-th cluster are allowed to be transferred to the j-th cluster to participate in the fusion operation, and the output is a sparse connection structure.

[0031] Optionally, the S4 specifically includes:

[0032] S41, construct the initial candidate sparse connection structure set, set the individual encoding method of the spider monkey optimization algorithm, and convert each individual SMO (i) Encoded as a corresponding candidate connection structure;

[0033] S42. Initialize the individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r)}, the individual position corresponds to the value of the candidate sparse connection structure matrix in the search space;

[0034] S43, setting a search process of the spider monkey optimization algorithm, including four stages: subgroup formation, local search, global information sharing, and group reorganization, and performing iterative update operations on the candidate sparse connection structures in the individual groups;

[0035] S44. Define a fitness function to evaluate the prediction effect of the candidate sparse connection structure on the health status score and abnormal risk label:

[0036]

[0037] Among them, F(M (i) ) is the candidate sparse connection structure M (i) The fitness value, M (i) is the candidate sparse connection structure, N is the total number of training samples, t is the sample index, α is the weight coefficient of health status score loss, d is the number of dimensions of health status score, j is the regression dimension index, is the jth true health score value of the tth sample, To use the sparse connection structure M (i) The j-th health score prediction value generated, β is the weight coefficient of the abnormal risk label loss, C is the total number of categories of abnormal risk labels, and c is the classification category index. is the true label of the t-th sample for the c-th category, To use the sparse connection structure M (i) The predicted probability value of class c;

[0038] S45. Determine the optimal sparse connection structure M that minimizes the fitness function * , the optimal sparse connection structure M * Applied between sub-network clusters in the micro neural network model, it updates the generation path of graph structure reasoning features to improve the accuracy of battery health status scoring and anomaly identification, and meet the reasoning efficiency requirements of edge computing nodes.

[0039] Optionally, the S43 specifically includes:

[0040] S431, initialize individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r)}, forming the initial population;

[0041] S432, divide the initial population into G subgroups, set a local leader L in each subgroup g , used to guide individuals within the subgroup to perform local search operations;

[0042] S433. In each round of local search, update the position codes of individuals in the subpopulation:

[0043]

[0044] in, is the new sparse connection structure matrix generated after individual update, M (i) is the candidate sparse connection structure, γ is the local guide step coefficient, L g is the local leader, ∈ is the local perturbation amplitude coefficient, and randn() is the perturbation matrix generated under the standard normal distribution;

[0045] S434, after the preset algebraic cycle, collect the best individuals in all sub-populations and select the individual with the smallest fitness value as the global leader L global , and perform a reorganization operation on the entire population, and divide the new subgroups according to the following strategies: if the fitness difference is large, maintain the distribution diversity and randomly assign individuals to subgroups; if the fitness tends to converge, adopt the leader-guided grouping strategy and move towards L global convergence;

[0046] S435. Repeat the subgroup local search and global information sharing operations until the maximum iteration number S is met, and output the final optimized structure configuration.

[0047] Optionally, the S5 specifically includes:

[0048] S51, receiving the graph structure reasoning features output by the micro neural network model, and the graph structure reasoning features are transmitted through the optimal sparse connection structure M * The local state expressions of multiple sub-network clusters are fused to generate a global feature representation of the multi-dimensional battery operating state;

[0049] S52. Based on the graph structure reasoning features, a health status scoring module is constructed, and a linear mapping is performed on the input features to generate a health score vector;

[0050] S53. Simultaneously, an abnormal risk identification module is constructed to perform nonlinear mapping and Softmax normalization on the graph structure reasoning features to generate an abnormal risk label probability vector;

[0051] S54: The health score vector and the abnormal risk label probability vector are combined to form a monitoring judgment result, which is used to determine whether the current battery operating status meets the set safety and performance thresholds.

[0052] Optionally, the S6 specifically includes:

[0053] S61. Receive monitoring results and set a scoring threshold vector τ reg =[τ1,τ2,...,τ d ], used to determine whether the health status is lower than the set lower limit, where d is the number of dimensions of the health score vector, and the risk confidence threshold τ is set cls , used to determine whether there is a risk category above the confidence threshold in the abnormal risk label probability vector;

[0054] S62: Execute edge judgment logic on the health score vector. If there exists a dimension j where the health score vector is less than the score threshold vector, it is determined that the score is insufficient.

[0055] S63: Execute edge judgment logic on the abnormal risk label probability vector. If there is a category c that satisfies the abnormal risk label probability vector greater than the risk confidence threshold, it is determined to be a high-risk abnormality.

[0056] S64. When any of the above conditions is met, a response record R is generated, which includes a response timestamp, a current graph structure reasoning feature, a current monitoring judgment result, a trigger item type, and a current optimal sparse connection structure M. * .

[0057] Optionally, the S7 specifically includes:

[0058] S71. Set the model update trigger conditions, including generating a model update request when any of the following conditions is met: The cumulative number of response records of the edge node reaches the preset cache threshold N. cache , the duration of abnormal health status score exceeds the preset time threshold T err , the decline in model performance evaluation index exceeds the upper limit of tolerance δ perf ;

[0059] S72: When the update condition is triggered, the edge node sends a model update request to the cloud platform, which includes the snapshot of the current micro neural network model parameters, the optimal sparse connection structure M, and the like. * , the cumulative response record set, the latest graph structure reasoning feature sample subset and the corresponding monitoring judgment results;

[0060] S73. The cloud platform performs structure optimization and parameter retraining based on the received data to generate an updated sub-network cluster weight parameter set and a candidate sparse connection structure set, and performs fitness screening on the connection structures to obtain a new round of optimal sparse connection structures.

[0061] S74, sending the updated sub-network cluster parameters and the new round of optimal sparse connection structure to the edge nodes in the form of a model update package;

[0062] S75. After receiving the model update package, the edge node performs the following operations: replace the original sub-network cluster weight parameters with the updated sub-network cluster weight parameter set, replace the existing sparse connection structure with a new round of optimal sparse connection structure, re-establish the graph structure reasoning path and complete the structural reconstruction of the micro neural network model to form a dynamically optimized edge reasoning model.

[0063] The beneficial effects of the present invention are:

[0064] This invention builds a micro-neural network model suitable for battery status monitoring within an edge computing architecture. It employs a sub-network clustering strategy to physically partition operating parameters, and combines this with a sparse connection control mechanism to implement structured reasoning, effectively enhancing the model's capabilities in modeling multi-source parameter coupling. Compared to the prior art approach of directly inputting all operating parameters into a fixed-structure model, this invention can exploit the correlations between different status characteristics while maintaining a low computational load, improving the accuracy of identifying battery health status and abnormal risks.

[0065] In terms of structural optimization, the present invention introduces a spider monkey optimization algorithm to perform a global search for sparse connection structures between sub-network clusters. Through an adaptive update mechanism, dynamic adjustment of connection relationships is achieved, overcoming the problem that traditional structural optimization methods cannot continuously adapt to state evolution. At the same time, the fitness function uses the joint prediction error of health scores and risk labels as the optimization target, making the generated connection structure more closely aligned with actual monitoring needs. Compared to traditional methods that rely on manual configuration or static search strategies, the present invention significantly improves the adaptability and convergence efficiency of the structure while ensuring the accuracy of model inference.

[0066] Furthermore, this invention combines the output of graph structure reasoning features with edge nodes to determine health status scores and abnormal risk labels. When specific threshold conditions are met, response records are instantly generated, achieving a closed-loop control system of monitoring, feedback, and optimization. This closed-loop mechanism allows edge nodes to initiate model update requests to the cloud on demand. The cloud then completes model structure retraining and returns an optimization package, enabling the edge model to have long-term evolutionary and dynamic adaptability, improving the stability and robustness of the system in long-term operating environments.

[0067] In summary, the present invention realizes high-precision, low-latency, and sustainably optimized real-time battery monitoring capabilities on low-resource edge devices, which not only improves operational safety and O&M efficiency, but also provides a structural design and optimization method with promotional value for distributed intelligent monitoring of multi-parameter heterogeneous states. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flow chart of a real-time battery monitoring method based on edge computing proposed by the present invention;

[0070] Figure 2 This is a schematic diagram of the sparsely connected sub-network cluster structure of the micro neural network model in the real-time battery monitoring method based on edge computing proposed in the present invention;

[0071] Figure 3 This is a flowchart of the spider monkey optimization algorithm performing sub-group collaborative search and update in sparse connection structure optimization in the real-time battery monitoring method based on edge computing proposed by the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1-3 , a real-time battery monitoring method based on edge computing, comprising the following steps:

[0074] S1. Collect multi-source operating parameters of the battery and build an operating parameter set;

[0075] S2. Time synchronization and normalization are performed on the set of operating parameters to generate a standardized state vector;

[0076] S3. Input the standardized state vector into a micro-neural network model deployed on an edge computing node. The micro-neural network model adopts a sparsely connected neural cluster structure, which includes multiple sub-network clusters. The sparse connection mechanism is used for cross-cluster information interaction to generate graph structure reasoning features.

[0077] S4. Structural optimization of the sparse connection mechanism is performed, and the spider monkey optimization algorithm is used to perform a global search on the connection structure between sub-network clusters to update the structural configuration of the micro neural network model;

[0078] S5. Based on the graph structure reasoning features and the optimized sparse connection structure, the health status score and abnormal risk label are output to form the monitoring judgment result;

[0079] S6. Perform edge response control based on the monitoring judgment results. When the health status score is lower than the preset score threshold or the confidence of the abnormal risk label is higher than the preset risk threshold, generate a response record;

[0080] S7. Based on the response records and monitoring judgment results, send a model update request to the cloud platform according to the set cycle or trigger conditions, receive the neural cluster parameter or structure adjustment instructions issued by the cloud platform, and complete the edge-side dynamic optimization of the micro neural network model.

[0081] The present invention proposes a real-time battery monitoring method based on edge computing. By introducing a sparsely connected neural cluster structure micro-neural network, it realizes structured modeling and reasoning optimization of multi-source operating parameters. The spider monkey optimization algorithm is used to perform a global search for the connection structure between sub-network clusters, which improves the reasoning efficiency and adaptability of the model on edge nodes. Through the output of health status scores and abnormal risk labels, combined with the edge response mechanism and on-demand triggered model update strategy, real-time identification and dynamic optimization of the battery operating status are achieved. This method has the characteristics of high recognition accuracy, fast response speed, low resource usage and self-evolving structure, and is suitable for intelligent energy storage monitoring systems under complex working conditions.

[0082] In this embodiment, the operating parameter set includes voltage parameters, current parameters, state of charge parameters, cell internal resistance parameters, temperature parameters, power and load parameters, discharge rate parameters and environmental parameters;

[0083] The voltage parameters include the voltage of a single cell, the total system voltage and the voltage fluctuation amplitude; the current parameters include the real-time operating current and the current change rate; the state of charge parameters include the current state of charge and the remaining power; the cell internal resistance parameters include the AC impedance of the single cell and the maximum and minimum difference between the internal resistance; the temperature parameters include the cell temperature, the casing temperature and the ambient temperature; the power and load parameters include the real-time load power and the power change rate; the discharge rate parameters include the current discharge rate, the maximum historical discharge rate and the rated rate ratio; the environmental parameters include the ambient humidity, air pressure and the external temperature fluctuation amplitude.

[0084] This invention achieves comprehensive perception of the battery's operating status by constructing a multidimensional set of operating parameters encompassing voltage, current, state of charge, cell internal resistance, temperature, power and load, discharge rate, and environmental parameters. Each parameter is broken down into multiple sub-items, enhancing the richness and accuracy of the state feature representation. This provides high-quality input for subsequent graph modeling and health scoring, and strengthens the model's sensitivity to abnormal changes and its ability to identify them early.

[0085] In this embodiment, S3 specifically includes:

[0086] S31, normalize the state vector X=[x1,x2,...,x n ] is input to the micro neural network model deployed on the edge computing node, and the standardized state vector is used to describe the multi-source operating state of the battery at the current sampling moment;

[0087] S32. Based on the physical properties and functional classification of the operating parameters, the standardized state vector is divided into five input sub-vectors, corresponding to the voltage, current, thermal, capacity, and environmental feature subsets, respectively. Each input sub-vector is input into a corresponding sub-network cluster to form a sub-network cluster input set.

[0088] S33. Each sub-network cluster performs local feedforward inference on the input sub-vector, calculates the local state expression, and outputs a set of local state expressions:

[0089] H (j) =max(0,W (j) X (j) +b (j) );

[0090] Among them, H (j) is the local state expression vector of the jth sub-network cluster, W (j) is the weight matrix of the j-th sub-network cluster, X (j) is the jth feature subset of the input normalized state vector, b (j) is the bias vector corresponding to the j-th sub-network cluster, and j is the index number of the sub-network cluster;

[0091] S34, constructing a sparsely connected neural cluster structure for simulating the coupling relationship between different physical dimensions of the battery, setting a sparse connection control matrix M for limiting the information interaction relationship between each sub-network cluster, and outputting the sparse connection structure;

[0092] S35. Based on the local state expression set and sparse connection structure, perform cross-cluster feature fusion operation to generate graph structure reasoning features:

[0093]

[0094] Among them, H fusion is the graph structure reasoning feature, i is the index number of the sub-network cluster, M ij is the element in the sparse connection control matrix, H (i) is the local state expression of the ith sub-network cluster, and f(·) is the cross-cluster feature interaction function.

[0095] The present invention realizes lightweight reasoning of multi-dimensional state vectors on edge computing nodes by constructing a micro neural network model and introducing a sparsely connected sub-network cluster structure. The input features are divided into five categories: voltage, current, thermal, capacity and environment by using a dimensional modeling method, and are input into the corresponding sub-network clusters respectively, and the local state expression is extracted within each cluster. A sparse connection matrix is ​​further constructed to limit the information transfer relationship between the sub-network clusters, and a graph structure reasoning feature is generated by weighted feature fusion. Among them, the fusion operation is completed by a function combination in the form of a two-way interactive relationship between the five sub-network clusters. The expression is that the fusion feature is equal to the weighted sum of the function results between the connection weights between each pair of sub-clusters and their corresponding state expressions, which is used to extract structured interactive information between multi-dimensional states. This method significantly reduces the computational complexity of the model while improving the expressive power of state modeling, which helps to improve the accuracy of battery state recognition and the adaptability of edge deployment.

[0096] In this embodiment, the subnetwork cluster includes an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer receives an input sub-vector, and the input sub-vector corresponds to one of the voltage, current, heat, capacity or environment feature subsets; the first hidden layer is a fully connected neural network layer, which performs a first-layer linear transformation and nonlinear activation and outputs a first-layer hidden state; the second hidden layer is a fully connected neural network layer, which receives the first-layer hidden state and outputs a local state expression.

[0097] This invention utilizes a subnetwork cluster structure consisting of an input layer, two fully connected hidden layers, and an output layer to achieve deep feature extraction for specific subsets of operating parameters. By inputting voltage, current, thermal, capacity, or environmental features into corresponding subnetwork clusters, the model performs linear mapping and nonlinear activation within the local dimension to generate physically meaningful local state representations. This structure improves the model's ability to model coupling relationships within parameter subspaces while maintaining a lightweight network, helping to enhance the accuracy and stability of state recognition.

[0098] In this embodiment, the sparse connection structure includes initializing the sparse connection control matrix M∈{0,1} to describe the coupling relationship between different operating parameter subspaces of the battery. 5×5 , where M ij Indicates whether the local state expression of the i-th sub-network cluster is allowed to be transferred to the j-th sub-network cluster to participate in feature fusion. If M ij =1, the transfer is allowed. If M ij = 0, transfer is not allowed. ij = 1, the output features of the i-th cluster are allowed to be transferred to the j-th cluster to participate in the fusion operation, and the output is a sparse connection structure.

[0099] This paper constructs a sparse connection control matrix to indicate whether feature transfer and fusion are permitted between subnetwork clusters. The connection control matrix is ​​a binary matrix with five rows and five columns, with each element indicating whether an information path is established between two corresponding subnetwork clusters. Only when an element is set to 1 is the output feature of the previous cluster allowed to be transferred to the next cluster for fusion calculation. This mechanism effectively limits invalid connections and reduces redundant propagation paths, helping to improve the efficiency and stability of the graph structure inference process and enhancing the model's ability to model the coupling relationships between multidimensional battery parameters.

[0100] In this embodiment, the S4 specifically includes:

[0101] S41, construct the initial candidate sparse connection structure set, set the individual encoding method of the spider monkey optimization algorithm, and convert each individual SMO (i) Encoded as a corresponding candidate connection structure;

[0102] S42. Initialize the individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r)}, the individual position corresponds to the value of the candidate sparse connection structure matrix in the search space;

[0103] S43, setting a search process of the spider monkey optimization algorithm, including four stages: subgroup formation, local search, global information sharing, and group reorganization, and performing iterative update operations on the candidate sparse connection structures in the individual groups;

[0104] S44. Define a fitness function to evaluate the prediction effect of the candidate sparse connection structure on the health status score and abnormal risk label:

[0105]

[0106] Among them, F(M (i) ) is the candidate sparse connection structure M (i) The fitness value, M (i) is the candidate sparse connection structure, N is the total number of training samples, t is the sample index, α is the weight coefficient of health status score loss, d is the number of dimensions of health status score, j is the regression dimension index, is the jth true health score value of the tth sample, To use the sparse connection structure M (i) The j-th health score prediction value generated, β is the weight coefficient of the abnormal risk label loss, C is the total number of categories of abnormal risk labels, and c is the classification category index. is the true label of the t-th sample for the c-th category, To use the sparse connection structure M (i) The predicted probability value of class c;

[0107] S45. Determine the optimal sparse connection structure M that minimizes the fitness function * , the optimal sparse connection structure M * Applied between sub-network clusters in the micro neural network model, it updates the generation path of graph structure reasoning features to improve the accuracy of battery health status scoring and anomaly identification, and meet the reasoning efficiency requirements of edge computing nodes.

[0108] The present invention uses the spider monkey optimization algorithm to iteratively update the sparse connection structure, and searches for the optimal connection mode between sub-network clusters through a population collaboration mechanism. The fitness function comprehensively considers the health status score error and the abnormal risk label classification error, and constructs a loss evaluation model with two weighted terms. Among them, the health score part uses the mean square error form to measure the deviation between the predicted value and the true value, and the risk label part uses the cross entropy form to measure the difference between the classification probability and the true label distribution, and combines them to form a unified fitness evaluation index. During the optimization process, this function guides the algorithm to converge in the direction of taking into account both regression accuracy and classification accuracy, so that the generated sparse connection structure is more discriminative and stable. This method improves the expressiveness and task adaptability of the model structure, and effectively supports the demand for high-precision identification of battery health status under complex working conditions.

[0109] In this embodiment, the S43 specifically includes:

[0110] S431, initialize individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r)}, forming the initial population;

[0111] S432, divide the initial population into G subgroups, set a local leader L in each subgroup g , used to guide individuals within the subgroup to perform local search operations;

[0112] S433. In each round of local search, update the position codes of individuals in the subpopulation:

[0113]

[0114] in, is the new sparse connection structure matrix generated after individual update, M (i) is the candidate sparse connection structure, γ is the local guide step coefficient, L g is the local leader, ∈ is the local perturbation amplitude coefficient, and randn() is the perturbation matrix generated under the standard normal distribution;

[0115] S434, after the preset algebraic cycle, collect the best individuals in all sub-populations and select the individual with the smallest fitness value as the global leader L global , and perform a reorganization operation on the entire population, and divide the new subgroups according to the following strategies: if the fitness difference is large, maintain the distribution diversity and randomly assign individuals to subgroups; if the fitness tends to converge, adopt the leader-guided grouping strategy and move towards L global convergence;

[0116] S435. Repeat the subgroup local search and global information sharing operations until the maximum iteration number S is met, and output the final optimized structure configuration.

[0117] The present invention improves the global search capability of sparse connection structure optimization by introducing a sub-group collaborative search mechanism. The spider monkey optimization algorithm is used to divide the initial population into multiple sub-groups, and a local leader is set in each sub-group as a local guidance target. In the local search stage, the individual position is updated according to the local leader. The update formula is composed of the current individual position, the guidance offset term and the normal perturbation term, so that the individual has directional convergence ability and diversity perturbation ability. In the global stage, the optimal individual with the lowest fitness among all sub-groups is set as the global leader to guide a new round of sub-group reorganization and global guidance. This mechanism improves the diversity exploration ability and stability of the connection structure iterative process, and enhances the adaptability of the model structure to complex state distribution.

[0118] In this embodiment, the S5 specifically includes:

[0119] S51, receiving the graph structure reasoning features output by the micro neural network model, and the graph structure reasoning features are transmitted through the optimal sparse connection structure M * The local state expressions of multiple sub-network clusters are fused to generate a global feature representation of the multi-dimensional battery operating state;

[0120] S52. Based on the graph structure reasoning features, a health status scoring module is constructed, and a linear mapping is performed on the input features to generate a health score vector;

[0121] S53. Simultaneously, an abnormal risk identification module is constructed to perform nonlinear mapping and Softmax normalization on the graph structure reasoning features to generate an abnormal risk label probability vector;

[0122] S54: The health score vector and the abnormal risk label probability vector are combined to form a monitoring judgment result, which is used to determine whether the current battery operating status meets the set safety and performance thresholds.

[0123] This invention uses graph-structured reasoning features as input to construct a health status scoring module and an abnormal risk identification module, respectively, to comprehensively assess the battery's operating status. The health scoring module outputs a status score using a linear mapping, while the abnormal risk identification module outputs the probabilities of multiple risk labels through nonlinear transformation and probability normalization. The combination of these two modules generates monitoring and assessment results, helping to accurately identify potential anomalies and performance degradation, enhancing the system's ability to perceive multidimensional state changes and the reliability of its decision-making basis.

[0124] In this embodiment, S6 specifically includes:

[0125] S61. Receive monitoring results and set a scoring threshold vector τ reg =[τ1,τ2,...,τ d ], used to determine whether the health status is lower than the set lower limit, where d is the number of dimensions of the health score vector, and the risk confidence threshold τ is set cls , used to determine whether there is a risk category above the confidence threshold in the abnormal risk label probability vector;

[0126] S62: Execute edge judgment logic on the health score vector. If there exists a dimension j where the health score vector is less than the score threshold vector, it is determined that the score is insufficient.

[0127] S63: Execute edge judgment logic on the abnormal risk label probability vector. If there is a category c that satisfies the abnormal risk label probability vector greater than the risk confidence threshold, it is determined to be a high-risk abnormality.

[0128] S64. When any of the above conditions is met, a response record R is generated, which includes a response timestamp, a current graph structure reasoning feature, a current monitoring judgment result, a trigger item type, and a current optimal sparse connection structure M. * .

[0129] This invention implements edge-judgment logic on the health score vector and the abnormal risk label probability vector by setting scoring thresholds and risk confidence thresholds, enabling rapid assessment of battery status. When any indicator exceeds the risk assessment threshold, the system automatically generates a response record containing key features and structural information for subsequent control or model optimization processes. This mechanism improves the timeliness of abnormal response and the integrity of data feedback, enhancing the system's diagnostic closed-loop capabilities in actual operation.

[0130] In this embodiment, the S7 specifically includes:

[0131] S71. Set the model update trigger conditions, including generating a model update request when any of the following conditions is met: The cumulative number of response records of the edge node reaches the preset cache threshold N. cache, the duration of abnormal health status score exceeds the preset time threshold T err , the decline in model performance evaluation index exceeds the upper limit of tolerance δ perf ;

[0132] S72: When the update condition is triggered, the edge node sends a model update request to the cloud platform, which includes the snapshot of the current micro neural network model parameters, the optimal sparse connection structure M, and the like. * , the cumulative response record set, the latest graph structure reasoning feature sample subset and the corresponding monitoring judgment results;

[0133] S73. The cloud platform performs structure optimization and parameter retraining based on the received data to generate an updated sub-network cluster weight parameter set and a candidate sparse connection structure set, and performs fitness screening on the connection structures to obtain a new round of optimal sparse connection structures.

[0134] S74, sending the updated sub-network cluster parameters and the new round of optimal sparse connection structure to the edge nodes in the form of a model update package;

[0135] S75. After receiving the model update package, the edge node performs the following operations: replace the original sub-network cluster weight parameters with the updated sub-network cluster weight parameter set, replace the existing sparse connection structure with a new round of optimal sparse connection structure, re-establish the graph structure reasoning path and complete the structural reconstruction of the micro neural network model to form a dynamically optimized edge reasoning model.

[0136] The present invention supports the dynamic generation of model update requests during the operation of edge nodes based on the number of response records, the duration of scoring anomalies, and the degree of performance degradation by setting a multi-condition trigger mechanism. The cloud platform performs structural optimization and parameter retraining based on edge feedback data, and screens and generates a new round of optimal sparse connection structures. The updated model parameters and connection structure are sent to the edge nodes in the form of update packages to complete the reconstruction of the neural network structure and the adjustment of the inference path. This mechanism realizes the closed-loop optimization and adaptive evolution of the model in actual operation, improving the monitoring stability and model adaptability under long-term operation.

[0137] Example 1:

[0138] To verify the feasibility of the present invention, it was applied to the battery pack operation monitoring scenario at a power storage station. This station is equipped with multiple lithium iron phosphate battery modules. During operation, they are subject to periodic charging and discharging, large temperature fluctuations, and uneven cell aging. While traditional BMS systems have basic data collection capabilities, they have significant deficiencies in status judgment accuracy, abnormal response delays, and model maintenance flexibility. Especially during certain high-load operating periods, traditional models often experience health status score deviations and abnormality alarm delays, affecting battery operation safety and back-end energy scheduling strategies.

[0139] In this scenario, the real-time battery monitoring method based on edge computing proposed in the present invention is deployed to the edge control module in the energy storage system. The model input receives the battery operating parameters in real time, including single cell voltage, total system current, ambient temperature, state of charge (SOC), and cell internal resistance. After standardized preprocessing, the state vector is constructed and divided into five sub-vectors according to parameter categories, corresponding to voltage, current, thermal, capacity, and environmental categories. It is then input into the micro neural network model to complete the sub-network cluster structure reasoning, and the sparse connection structure is iterated under the action of the spider monkey optimization algorithm to improve the reasoning accuracy and operation efficiency.

[0140] The model output includes a health score vector and an abnormal risk label, based on which the edge node can make real-time judgments and generate response records. In a data comparison experiment that ran continuously for 30 days, compared with the traditional fixed structure model, the method of the present invention showed significant advantages in state recognition accuracy, response delay control, and model iteration flexibility. Specifically, in the battery cell fault detection scenario, the traditional model only has a recognition rate of 41.3% within 12 hours of the initial fault, while the model of the present invention can achieve a recognition accuracy of 92.8% within 3 hours. At the same time, under high-temperature operating conditions, the traditional model has a health score misjudgment rate of up to 22.5% due to the propagation of thermal parameter errors. However, this method dynamically adjusts the connection path of the thermal neural cluster through a sparse connection structure, reducing the misjudgment rate to 5.3%.

[0141] In addition, in terms of edge computing resource consumption, the traditional model structure has 3280 parameters and an average inference delay of about 72 milliseconds; after sparse connection optimization, the model of the present invention has compressed the number of parameters to 936 and reduced the inference delay to 28 milliseconds, effectively meeting the computing power limitations of edge devices. The cloud structure update cycle has been optimized from the original fixed adjustment every 30 days to on-demand triggering. The model structure was dynamically reconstructed 4 times within 30 days, with an average fitness improvement of 12.7%. The comprehensive response time of the final monitoring system was shortened from 6.2 seconds of the traditional solution to 1.4 seconds of the present invention, and the stability of the system operation was not affected.

[0142] The relevant comparative data are shown in the following table:

[0143] Table 1 Performance comparison data of the method of the present invention and the traditional method in actual monitoring tasks

[0144]

[0145] As can be seen from the above table, the present invention shows obvious advantages over traditional monitoring methods in multiple core performance indicators of battery operation status monitoring, especially in terms of status recognition accuracy, response efficiency, model resource consumption and structural adaptability, all of which have achieved key breakthroughs.

[0146] First, in terms of state recognition capability, the model of the present invention has an identification rate of 92.8% for abnormal battery states at the initial stage of a fault, which is much higher than the 41.3% of the traditional model, an improvement of more than 50 percentage points. This shows that the graph structure reasoning method based on sub-network cluster modeling and sparse connection structure optimization can more sensitively capture the early coupling characteristics between multi-dimensional parameters and effectively improve the system's ability to identify potential anomalies in advance. In addition, in high-temperature operation scenarios, the traditional model has a scoring misjudgment rate of up to 22.5% due to insufficient processing of thermal features. The model of the present invention reduces the misjudgment rate to 5.3% by guiding the thermal sub-network cluster to perform local reinforcement learning, maintaining higher stability and accuracy under extreme working conditions.

[0147] In terms of response efficiency, the average alarm response delay of the model of the present invention is 1.4 seconds, far lower than the 6.2 seconds of the traditional solution, reflecting its real-time reasoning advantage in edge deployment environments. Combined with the data of parameter compression rate exceeding 70% and average single reasoning time reduction of 61.1%, it also shows that the present invention achieves a dual improvement in computing power utilization and monitoring speed through the combination of lightweight model design and structural optimization algorithm. In particular, in energy storage systems with limited edge hardware resources, this structure has extremely high practicality.

[0148] It is worth mentioning that in terms of model structure adaptation, the present invention supports an on-demand triggering model update mechanism, achieving four structural optimizations based on feedback data during 30 consecutive days of operation. Traditional methods use fixed-cycle updates and are unable to respond to changes in system status. The average structural fitness of the present invention has increased by 12.7%, further demonstrating its ability to continuously evolve during long-term operation. In addition, the anomaly label false detection rate has been reduced from the traditional 18.4% to 4.6%, and no interruptions have occurred during system operation, demonstrating the improvement of the overall stability and security of the system.

[0149] Based on the above data, the present invention significantly improves the accuracy, response timeliness and system intelligence level of battery operation status monitoring, and provides a more efficient, safe and adaptive edge intelligent monitoring solution for energy storage systems.

[0150] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A real-time battery monitoring method based on edge computing, characterized in that: The steps include: S1. Collect multi-source operating parameters of the battery and build an operating parameter set; S2. Time synchronization and normalization are performed on the set of operating parameters to generate a standardized state vector; S3. Input the standardized state vector into a micro-neural network model deployed on an edge computing node. The micro-neural network model adopts a sparsely connected neural cluster structure, which includes multiple sub-network clusters. The sparse connection mechanism is used for cross-cluster information interaction to generate graph structure reasoning features. S4. Structural optimization of the sparse connection mechanism is performed, and the spider monkey optimization algorithm is used to perform a global search on the connection structure between sub-network clusters to update the structural configuration of the micro neural network model; S5. Based on the graph structure reasoning features and the optimized sparse connection structure, the health status score and abnormal risk label are output to form the monitoring judgment result; S6. Perform edge response control based on the monitoring judgment results. When the health status score is lower than the preset score threshold or the confidence of the abnormal risk label is higher than the preset risk threshold, generate a response record; S7. Based on the response records and monitoring judgment results, send a model update request to the cloud platform according to the set cycle or trigger conditions, receive the neural cluster parameter or structure adjustment instructions issued by the cloud platform, and complete the edge-side dynamic optimization of the micro neural network model.

2. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The operating parameter set includes voltage parameters, current parameters, state of charge parameters, cell internal resistance parameters, temperature parameters, power and load parameters, discharge rate parameters and environmental parameters.

3. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S3 specifically includes: S31, normalize the state vector X=[x1,x2,...,x n ] is input to the micro neural network model deployed on the edge computing node; S32. Based on the physical properties and functional classification of the operating parameters, the standardized state vector is divided into five input sub-vectors, corresponding to the voltage, current, thermal, capacity, and environmental feature subsets, respectively. Each input sub-vector is input into a corresponding sub-network cluster to form a sub-network cluster input set. S33. Each sub-network cluster performs local feedforward inference on the input sub-vector, calculates the local state expression, and outputs a set of local state expressions: H (j) =max(0,W (j) X (j) +b (j) ); Among them, H (j) is the local state expression vector of the jth sub-network cluster, W (j) is the weight matrix of the j-th sub-network cluster, X (j) is the jth feature subset of the input normalized state vector, b (j) is the bias vector corresponding to the j-th sub-network cluster, and j is the index number of the sub-network cluster; S34, constructing a sparsely connected neural cluster structure, setting a sparse connection control matrix M, and outputting the sparse connection structure; S35. Based on the local state expression set and sparse connection structure, perform cross-cluster feature fusion operation to generate graph structure reasoning features: Among them, H fusion is the graph structure reasoning feature, i is the index number of the sub-network cluster, M ij is the element in the sparse connection control matrix, H (i) is the local state expression of the ith sub-network cluster, and f(·) is the cross-cluster feature interaction function.

4. The method for real-time monitoring of batteries based on edge computing according to claim 3, characterized in that: The subnetwork cluster includes an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer receives an input subvector, which corresponds to one of the voltage, current, heat, capacity or environment feature subsets; the first hidden layer is a fully connected neural network layer, which performs a first-layer linear transformation and nonlinear activation and outputs a first-layer hidden state; the second hidden layer is a fully connected neural network layer, which receives the first-layer hidden state and outputs a local state expression.

5. The method for real-time monitoring of batteries based on edge computing according to claim 3, characterized in that: The sparse connection structure includes initializing the sparse connection control matrix M∈{0,1} to describe the coupling relationship between different operating parameter subspaces of the battery. 5×5 , where M ij Indicates whether the local state expression of the i-th sub-network cluster is allowed to be transferred to the j-th sub-network cluster to participate in feature fusion. If M ij =1, the transfer is allowed. If M ij = 0, transfer is not allowed. ij = 1, the output features of the i-th cluster are allowed to be transferred to the j-th cluster to participate in the fusion operation, and the output is a sparse connection structure.

6. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S4 specifically includes: S41, construct the initial candidate sparse connection structure set, set the individual encoding method of the spider monkey optimization algorithm, and convert each individual SMO (i) Encoded as a corresponding candidate connection structure; S42. Initialize the individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r) }, the individual position corresponds to the value of the candidate sparse connection structure matrix in the search space; S43, setting a search process of the spider monkey optimization algorithm, including four stages: subgroup formation, local search, global information sharing, and group reorganization, and performing iterative update operations on the candidate sparse connection structures in the individual groups; S44. Define a fitness function to evaluate the prediction effect of the candidate sparse connection structure on the health status score and abnormal risk label: Among them, F(M (i) ) is the candidate sparse connection structure M (i) The fitness value, M (i) is the candidate sparse connection structure, N is the total number of training samples, t is the sample index, α is the weight coefficient of health status score loss, d is the number of dimensions of health status score, j is the regression dimension index, is the jth true health score value of the tth sample, To use the sparse connection structure M (i) The j-th health score prediction value generated, β is the weight coefficient of the abnormal risk label loss, C is the total number of categories of abnormal risk labels, and c is the classification category index. is the true label of the t-th sample for the c-th category, To use the sparse connection structure M (i) The predicted probability value of class c; S45. Determine the optimal sparse connection structure M that minimizes the fitness function * , the optimal sparse connection structure M * Applied to the sub-network clusters in the micro neural network model, it updates the generation path of graph structure reasoning features.

7. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S43 specifically includes: S431, initialize individual position group P = {SMO (1) ,SMO (2) ,...,SMO (r) }, forming the initial population; S432, divide the initial population into G subgroups, set a local leader L in each subgroup g , used to guide individuals within the subgroup to perform local search operations; S433. In each round of local search, update the position codes of individuals in the subpopulation: in, is the new sparse connection structure matrix generated after individual update, M (i) is the candidate sparse connection structure, γ is the local guide step coefficient, L g is the local leader, ∈ is the local perturbation amplitude coefficient, and randn() is the perturbation matrix generated under the standard normal distribution; S434, after the preset algebraic cycle, collect the best individuals in all sub-populations and select the individual with the smallest fitness value as the global leader L global , and perform a reorganization operation on the entire population, and divide the new subgroups according to the following strategies: if the fitness difference is large, maintain the distribution diversity and randomly assign individuals to subgroups; if the fitness tends to converge, adopt the leader-guided grouping strategy and move towards L global convergence; S435. Repeat the subgroup local search and global information sharing operations until the maximum iteration number S is met, and output the final optimized structure configuration.

8. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S5 specifically includes: S51, receiving the graph structure reasoning features output by the micro neural network model, and the graph structure reasoning features are transmitted through the optimal sparse connection structure M * Fusion of multiple sub-network clusters to generate local state expressions; S52. Based on the graph structure reasoning features, a health status scoring module is constructed, and a linear mapping is performed on the input features to generate a health score vector; S53. Simultaneously, an abnormal risk identification module is constructed to perform nonlinear mapping and Softmax normalization on the graph structure reasoning features to generate an abnormal risk label probability vector; S54. The health score vector and the abnormal risk label probability vector are combined to form a monitoring judgment result.

9. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S6 specifically includes: S61. Receive monitoring results and set a scoring threshold vector τ reg =[τ1,τ2,...,τ d ], where d is the dimension of the health score vector, and the risk confidence threshold τ is set cls ; S62: Execute edge judgment logic on the health score vector. If there exists a dimension j where the health score vector is less than the score threshold vector, it is determined that the score is insufficient. S63: Execute edge judgment logic on the abnormal risk label probability vector. If there is a category c that satisfies the abnormal risk label probability vector greater than the risk confidence threshold, it is determined to be a high-risk abnormality. S64. When any of the above conditions is met, a response record R is generated, which includes a response timestamp, a current graph structure reasoning feature, a current monitoring judgment result, a trigger item type, and a current optimal sparse connection structure M. * .

10. The method for real-time monitoring of batteries based on edge computing according to claim 1, characterized in that: The S7 specifically includes: S71. Set the model update trigger conditions, including generating a model update request when any of the following conditions is met: The cumulative number of response records of the edge node reaches the preset cache threshold N. cache , the duration of abnormal health status score exceeds the preset time threshold T err , the decline in model performance evaluation index exceeds the upper limit of tolerance δ perf ; S72: When the update condition is triggered, the edge node sends a model update request to the cloud platform, which includes the snapshot of the current micro neural network model parameters, the optimal sparse connection structure M, and the like. * , the cumulative response record set, the latest graph structure reasoning feature sample subset and the corresponding monitoring judgment results; S73. The cloud platform performs structure optimization and parameter retraining based on the received data to generate an updated sub-network cluster weight parameter set and a candidate sparse connection structure set, and performs fitness screening on the connection structures to obtain a new round of optimal sparse connection structures. S74, sending the updated sub-network cluster parameters and the new round of optimal sparse connection structure to the edge nodes in the form of a model update package; S75. After receiving the model update package, the edge node performs the following operations: replace the original sub-network cluster weight parameters with the updated sub-network cluster weight parameter set, replace the existing sparse connection structure with a new round of optimal sparse connection structure, re-establish the graph structure reasoning path and complete the structural reconstruction of the micro neural network model to form a dynamically optimized edge reasoning model.

Images