A lithium battery soh estimation method based on federated learning and client multi-dimensional scoring mechanism

Abstract

The application discloses a lithium battery SOH estimation method based on a federal learning and a client multi-dimensional scoring mechanism, and the method comprises the following steps: a battery client constructs a local feature data set, designs a deep learning network model based on the format of the feature data set, and is used for a federal server and all clients; relevant parameters are initialized; a multi-dimensional scoring mechanism is constructed by fusing a dynamic loss, a data distribution deviation, a historical contribution degree and gradient consistency, and then a target function is constructed to maximize the total score of all clients; a genetic algorithm is used to solve the maximum fitness function and the corresponding optimal selection scheme of the client, the client is selected according to the scheme to perform local training, parameter uploading and global aggregation, and the cycle iteration is continued until the model parameters converge. The application proposes a client multi-dimensional scoring mechanism, evaluates the improvement potential of each client on the global model performance, optimizes the selection of participating clients, and improves the estimation accuracy and convergence speed of the lithium battery SOH under the federal learning framework.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium batteries, and particularly relates to a lithium battery SOH estimation method based on a federated learning and a client multi-dimensional scoring mechanism. BACKGROUND

[0002] Accurate estimation of the state of health (SOH) of lithium batteries is a core technology for ensuring the safe operation of electric vehicles, energy storage systems, mobile terminals, etc., and directly affects battery life prediction, fault warning, and device endurance assessment. However, the SOH of the battery has high intrinsic complexity and nonlinearity, and its estimation accuracy is heavily dependent on a large amount and variety of historical operation data.

[0003] Traditional centralized SOH estimation methods upload all terminal raw battery data (such as voltage, current, temperature) to a central server for processing and analysis. This method faces data privacy and security issues. Battery operation data contains sensitive information such as device operating conditions and user habits. For example, through battery raw operation data, user behavior can be traced, and directly collecting all client battery operation data and storing it centrally poses a risk of data privacy leakage. To address the above issues, federated learning technology provides a promising solution. Using the "data does not move, model moves" feature of federated learning, multiple battery device clients collect and store data locally, and only share model parameters rather than raw data, thereby collaborating to train a global model while protecting data privacy.

[0004] However, existing lithium battery SOH estimation methods under the federated learning framework still have the following problems: (1) Clients with severe data distribution bias may produce low-quality gradients, which can reduce the accuracy of the global model. (2) In each iteration of federated learning, the model gradient update direction calculated by different clients based on local data may differ significantly. Clients with inconsistent gradients participate in aggregation, and their updates cancel out the effective contributions of other clients, causing the model update direction of local training to deviate from the global optimal solution, slowing down the model convergence speed. (3) Simultaneous participation of all battery clients in training consumes a large amount of resources and makes it difficult to ensure a significant improvement in accuracy. Therefore, under the federated learning framework, it is necessary to build a battery client selection system that can evaluate the potential of clients participating in training to improve the performance of the global model. Based on this, the most valuable subset of clients is intelligently selected to participate in each round of training, which is a key technical breakthrough direction for optimizing resource utilization efficiency, accelerating model convergence, improving federated learning effectiveness, and thus improving SOH estimation accuracy. SUMMARY

[0005] To overcome the deficiencies of the above-mentioned existing SOH estimation methods based on federated learning, the present application proposes a lithium battery SOH estimation method based on federated learning and a client multi-dimensional scoring mechanism.

[0006] The technical solutions of the present application are as follows:

[0007] S1: Each battery client participating in training independently completes data acquisition locally, and constructs a local feature data set based on charge-discharge operation data; the feature data set includes a voltage range feature matrix, a current range feature matrix, and a temperature range feature matrix; based on the format of the feature data set, a deep learning network model structure is designed and applied to the federated server end and all clients; the deep learning network model takes the voltage range feature matrix, the current range feature matrix, and the temperature range feature matrix as input, and takes the battery SOH of each charge-discharge cycle as output;

[0008] S2: Initialize related parameters, set the model parameters and average gradient of the federated server end as the global model parameters and global average gradient, set the number of clients, which is used for subsequent model parameter aggregation and updating, and client selection optimization;

[0009] S3: The federated server distributes the global model parameters to all clients, and each client performs local training based on local battery data; after training is completed, each client records its own model parameters as initial data for subsequent optimization, and uploads its own model parameters to the federated server; after collecting the model parameters of each client, the federated server aggregates the parameters of all clients to update the global model parameters;

[0010] S4: Construct a target function for client optimization selection, the specific steps including: defining a binary decision variable representing the participation state of the client; establishing a multi-dimensional performance scoring mechanism that integrates dynamic loss, data distribution deviation, historical contribution, and gradient consistency to evaluate the potential of each client to improve the performance of the global model; finally, taking the maximization of the total performance score of the selected clients as the goal, combining the penalty function reflecting the lower limit of the number of selected clients and dynamic fairness constraints, to form the target function for client selection;

[0011] S5: Solve the target function for client optimization selection described in step S4 using a genetic algorithm, output the optimal client selection scheme, the specific steps including: encoding all binary decision variables representing the participation state of the client into a binary chromosome with a length equal to the total number of clients, where gene bit values 1 and 0 represent the activation and inactivation of the client respectively; after randomly generating an initial population, select the parent by fitness proportion, and perform single-point crossover and bit mutation operations; terminate the iterative calculation of the genetic algorithm when the maximum number of iterations is reached or the optimal fitness value of the target function is continuously stagnant, and decode the chromosome with the highest fitness value to obtain the optimal client selection scheme;

[0012] S6: The federal server issues the current global model parameters to all clients, and then selects the corresponding client according to the client optimal selection scheme obtained in S5, so that the client trains and updates the SOH model locally using the battery data. After the training is completed, the selected client records the updated model parameters and uploads the updated model parameters to the federal server. The federal server aggregates and updates the global model parameters after collecting all the selected client model parameters, and completes the current round of federal learning iteration;

[0013] S7: Steps S5 to S6 are iteratively executed until the global model parameters on the federal server converge.

[0014] In this scheme, the construction of the local feature data set and the design of the deep learning network model structure in step S1 are as follows:

[0015] S101: Each battery client independently completes data collection locally. The collected data includes voltage, current, temperature and other operating data in the battery charge and discharge cycle, and the battery SOH recorded in each charge and discharge cycle;

[0016] S102: Normalize the input variables and output variables, and divide the training set and the test set;

[0017] S103: Average segment the normalized voltage, current and temperature. The normalized voltage, current and temperature are divided into K segments using a fixed length equal division strategy, so that each segment contains the same data points, and the specific steps are as follows:

[0018]

[0019] Wherein, N D represents the number of data points of voltage, current or temperature, L D represents the number of data points contained in each segment, k represents the segment number, k = 1, 2, …, K; represents the starting index of segment k, represents the ending index of segment k, n represents the index number of data points, V k represents the voltage set of segment k, V n represents the voltage at index n, I k represents the current set of segment k, I n represents the current at index n, T k represents the temperature set of segment k, T n represents the temperature at index n;

[0020] S104: According to V k , I k , T k, construct the voltage difference feature, the current difference feature and the temperature difference feature, as shown below:

[0021]

[0022] wherein, respectively represent the voltage difference feature, the current difference feature and the temperature difference feature of the segment k, max(·) represents the maximum value in the set, and min(·) represents the minimum value in the set;

[0023] S105: The voltage difference feature, the current difference feature and the temperature difference feature constructed in step S104 are further constructed into a voltage difference feature matrix, a current difference feature matrix and a temperature difference feature matrix, respectively, as shown below:

[0024]

[0025]

[0026] wherein, respectively represent the voltage difference feature, the current difference feature and the temperature difference feature of the segment p, respectively represent the voltage difference feature, the current difference feature and the temperature difference feature of the segment q, respectively represent the absolute difference value of the pth and qth difference of voltage, current and temperature, respectively represent the voltage difference feature matrix, the current difference feature matrix and the temperature difference feature matrix;

[0027] S106: A plurality of one-dimensional convolution network models are constructed, the model input is the voltage difference feature matrix, the current difference feature matrix and the temperature difference feature matrix obtained in step S105, each row in the voltage difference feature matrix, the current difference feature matrix and the temperature difference feature matrix is input into the respective one-dimensional convolution neural network, then the extracted features are averaged according to voltage, current and temperature respectively, and finally feature splicing and SOH estimation are performed, as shown below:

[0028]

[0029] SOH = σ(w o · F fusion + b o )

[0030] wherein, Conv1D(·) represents one-dimensional convolution operation, and ReLU(·) represents an activation function, represents the mth row of the voltage difference feature matrix , represents the mth row of the current difference feature matrix , denotes the temperature range feature matrix the mth row of denotes the extracted features, denotes the extracted features, denotes the extracted features, respectively represent the average features of voltage, current, and temperature, F fusion denotes the fused features, Concatenate(·) denotes the fusion operation, σ(·) denotes the sigmoid activation function, w o , b o respectively represent the weights and bias terms of the output network, and SOH is the battery health state output by the network.

[0031] In this scheme, the target function of the client optimization selection constructed in step S4 is as follows:

[0032] S401: Define binary decision variable x i,t to represent the participation state of the client in the federated learning round, which is specifically as follows:

[0033]

[0034] where Γ represents the set of clients, t is the federated learning iteration round index, Epoch is the total number of rounds, x i,t = 1 indicates that the client i is selected to participate in federated learning training in the tth round of federated learning, and x i,t = 0 indicates that the client i is not selected in the tth round of federated learning.

[0035] S402: Based on the historical participation data of the client i, construct an evaluation function of the potential improvement of the client on the global model performance, including a dynamic loss function, a data distribution deviation function, a historical contribution function, and a gradient consistency function, reflecting the expected gain of the participation of the client in the tth round of federated learning on the global model performance, which is specifically as follows:

[0036]

[0037] where VF i,t represents the potential improvement function of the client i participating in the tth round of federated learning iteration on the global model performance, α1, α2, α3, and α4 are weight coefficients, represents the dynamic loss function of the client i before the tth round of federated learning iteration, represents the data distribution deviation function of the client i before the yth round of federated learning iteration, represents the historical contribution function of the client i before the tth round of federated learning iteration, denotes the gradient consistency function of the client i before the tth round of federated learning iteration;

[0038] the dynamic loss function Specifically as follows:

[0039]

[0040] wherein, W t-1 denotes the global model parameters aggregated in the t-1th round, W t-2 denotes the global model parameters aggregated in the t-2th round, Loss i denotes the loss value of the specified model parameters on the client i, denotes the local model parameters of the client j after participating in the federated learning training in the t-1th round, denotes the local model parameters of the client j after participating in the federated learning training in the t-2th round, n j denotes the local sample data volume of the client j, N t-1 denotes the total effective sample number in the t-1th round, N t-2 denotes the total effective sample number in the t-2th round, denotes the client set selected in the t-1th round of federated learning, denotes the client set selected in the t-2th round of federated learning;

[0041] the data distribution deviation function Specifically as follows:

[0042]

[0043]

[0044] wherein, w U , w I , w T respectively denote the weight of voltage, current, and temperature data distribution deviation, W UIT denotes the sum of all weight coefficients, W UIT = w U + w I + w T ; K D is the number of grades, EMD U , EMD I , EMD T respectively denote the voltage, current, and temperature data distribution deviation, P Ul and Q Ul respectively denote the proportion of the client voltage and the global voltage in the voltage grade l, P Il and Q Il respectively denote the proportion of the client current and the global current in the current grade l, PTl and Q Tl respectively represent the proportion of client temperature and global temperature at temperature level l, respectively represent the distance weight between adjacent levels of voltage, current and temperature;

[0045] the historical contribution function is specifically shown as follows,

[0046]

[0047] wherein γ represents a decay parameter of historical contribution; refers to the historical contribution of client i before the t-1th round of federated learning iteration; represents the newly added contribution of client i in the t-1th round of federated learning, which exists only when client i is selected to participate in the t-1th round of federated learning training, and is 0 when client i is not selected to participate in training; ||·||2 represents L2 norm, refers to the local model parameter of client i after the t-1th round of federated learning training;

[0048] the gradient consistency function is specifically shown as follows:

[0049]

[0050] wherein g t-1 refers to the global average gradient of the t-1th round of federated learning; refers to the effective gradient of client i participating in the t-1th round of federated learning training, represents the gradient of client i selected for training in the t-1th round of federated learning, represents the gradient of client i not selected for training in the t-1th round of federated learning, which is the historical gradient; refers to the effective gradient of client j participating in the t-1th round of federated learning training; is the number of clients in the set <·,·> represents the inner product of two gradients; ||·|| represents the norm; exp(·) represents the exponential function;

[0051] S403: Ensure aggregation effectiveness by setting the lower limit of the number of selected clients, and avoid resource unevenness by setting dynamic fairness constraints, which are specifically shown as follows:

[0052]

[0053] wherein x i,c represents the decision variable state of client i in the cth round of federated learning, K minVF represents the preset lower limit of the number of selected clients, R represents the rolling round window size of the dynamic fairness constraint, F min represents the minimum number of participations of the clients in the window, F max represents the maximum number of participations of the clients in the window;

[0054] S404: Based on the VF defined in S402, the potential function VF of the performance of the global model is defined based on the participation of the client i in the tth round of federated learning iteration i,t , and a penalty function is combined to reflect the lower limit of the number of selected clients and the dynamic fairness constraint, to construct a global objective function, as shown below:

[0055]

[0056] Wherein, f represents the global objective function, |Γ| represents the total number of all clients, and Φ represents the penalty function;

[0057] The penalty function Φ is specifically as follows,

[0058] Φ=Φ1+Φ2

[0059]

[0060] Wherein, Φ1 represents the penalty term for violating the lower limit of the number of selected clients; Φ2 represents the penalty term for violating the dynamic fairness constraint; ξ1 represents the penalty intensity coefficient for violating the lower limit of the number of selected clients, and ξ2 represents the penalty intensity coefficient for violating the fairness constraint; S i represents the total number of times that the client i is selected in the last R rounds; max(·) represents the maximum of the two;

[0061] In the scheme, the genetic algorithm is used to solve the objective function of the client optimization selection in step S5, and the optimal client selection scheme is output, and the specific steps are as follows:

[0062] S501: The client participation state decision variable is encoded into a binary chromosome, the chromosome length is equal to the total number of clients, and the gene bit i corresponds to the activation state of the client i, which is recorded as 1 when selected and 0 when not selected;

[0063] S502: The global objective function constructed in S404 is used as the fitness function;

[0064] S503: Randomly generate an initial population with a size of P num , and the gene bits of each chromosome are generated as 1 or 0 by independent random assignment;

[0065] S504: Calculate the fitness value of each individual in the population by adopting the fitness-based probability selection strategy, select the parent individual by roulette or tournament strategy, preferentially reserve the high-quality solution, then perform single-point crossover on the paired parents with a preset crossover probability, randomly select the cutting point to exchange the gene fragments to generate new offspring; finally, traverse each gene site of the offspring individual, and perform bit flip mutation with a low probability to introduce diversity;

[0066] S505: Terminate the genetic algorithm iteration when the maximum iteration number is reached or the continuous multiple generations of optimal fitness are insufficient, output the chromosome with the highest fitness in the past generations as the optimal solution, and obtain the optimal selection scheme of the client.

[0067] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0068] The present application can dynamically and intelligently select the client subset with the most contribution potential to the current global model optimization to participate in training, thereby effectively improving the estimation accuracy of the lithium battery SOH, the resource utilization efficiency and accelerating the model convergence speed under the federal learning framework. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0070] Figure 1 A lithium battery SOH estimation method based on federal learning and client multi-dimensional scoring mechanism is provided.

[0071] Figure 2 A flowchart for constructing the feature data set of the client and the model; DETAILED DESCRIPTION

[0072] In order to more clearly illustrate the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0073] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0074] In one specific embodiment, as shown in Figure 1 a lithium battery SOH estimation method based on federated learning and client multi-dimensional scoring mechanism, comprising the following steps:

[0075] S1: Each battery client participating in training independently completes data collection locally, and constructs a local feature data set based on charge and discharge operation data; the feature data set includes a voltage range feature matrix, a current range feature matrix and a temperature range feature matrix; based on the format of the feature data set, a deep learning network model structure is designed and applied to the federated server end and all clients; the deep learning network model takes the voltage range feature matrix, the current range feature matrix and the temperature range feature matrix as input, and takes the battery SOH of each charge and discharge cycle as output;

[0076] S2: Initialize related parameters, set the model parameters and average gradient of the federated server end as the global model parameters and global average gradient, set the number of clients, which is used for subsequent model parameter aggregation and update, client selection optimization;

[0077] S3: The federated server distributes the global model parameters to all clients, and each client performs local training based on local battery data; after training, each client records its own model parameters as initial data for subsequent optimization, and uploads its own model parameters to the federated server; after collecting the model parameters of each client, the federated server aggregates the parameters of all clients to update the global model parameters;

[0078] S4: Construct a target function for client optimization selection, the specific steps including: defining a binary decision variable representing the participation state of the client; establishing a client multi-dimensional performance scoring mechanism that integrates dynamic loss, data distribution deviation, historical contribution and gradient consistency to evaluate the potential of each client to improve the performance of the global model; finally, taking the maximization of the total score of the comprehensive performance of the selected clients as the goal, combining the penalty function reflecting the lower limit of the number of selected clients and dynamic fairness constraints, forming the target function of client selection;

[0079] S5: solving the target function of the client optimal selection in step S4 by using a genetic algorithm, outputting the client optimal selection scheme, the specific steps including: encoding all binary decision variables representing the participation state of the client into a binary chromosome with a length equal to the total number of clients, wherein the gene bit values 1 and 0 represent the activation and non-activation of the client respectively; after randomly generating an initial population, selecting the parent by fitness proportionate selection, and performing single-point crossover and bit mutation operation; when the maximum number of iterations is reached or the optimal fitness value of the target function is continuously stagnant, terminating the iterative calculation of the genetic algorithm, and decoding the chromosome with the highest fitness value to obtain the client optimal selection scheme;

[0080] S6: the federal server issues the current global model parameters to all clients, and then selects the corresponding clients according to the client optimal selection scheme obtained in S5, so that the selected clients train and update the SOH model locally using the battery data, and after the training is completed, the selected clients record the updated model parameters and upload the updated model parameters to the federal server; the federal server aggregates and updates the global model parameters after collecting all the selected client model parameters, completing the current round of federated learning iteration;

[0081] S7: iteratively performing steps S5 to S6 until the global model parameters on the federal server converge.

[0082] In this scheme, the construction of the local feature data set and the design of the deep learning network model structure in step S1 are as shown in the following table: Figure 2 The specific steps are as follows:

[0083] S101: each battery client independently completes data collection locally, and the collected data includes operating data such as voltage, current and temperature in the battery charge and discharge cycle, and the battery SOH recorded in each charge and discharge cycle;

[0084] S102: normalizing the input variables and output variables, and dividing the training set and the test set;

[0085] S103: performing average segmentation processing on the normalized voltage, current and temperature, and using a fixed length equal division strategy to divide the normalized voltage, current and temperature into K segments, so that each segment contains the same data points, which are as follows:

[0086]

[0087] Wherein, N D represents the number of data points of voltage, current or temperature, L D represents the number of data points contained in each segment, k represents the segment number, k = 1, 2, …, K; represents the starting index of segment k, denotes the end index of segment k, n denotes the index number of data points, V k denotes the voltage set of segment k, V n denotes the voltage of index n, I k denotes the current set of segment k, I n denotes the current of index n, T k denotes the temperature set of segment k, T n denotes the temperature of index n;

[0088] S104: V k , I k , T k obtained according to step S103 are constructed into voltage range features, current range features and temperature range features, as shown below:

[0089]

[0090] wherein, denote the voltage range feature, current range feature and temperature range feature of segment k respectively, max(·) denotes the maximum value in the set, and min(·) denotes the minimum value in the set;

[0091] S105: The voltage range features, current range features and temperature range features constructed in step S104 are further constructed into voltage range feature matrix, current range feature matrix and temperature range feature matrix respectively, as shown below:

[0092]

[0093] wherein, denote the voltage range feature, current range feature and temperature range feature of segment p respectively, denote the voltage range feature, current range feature and temperature range feature of segment q respectively, denote the absolute difference value of the pth and qth range of voltage, current and temperature respectively, denote the voltage range feature matrix, current range feature matrix and temperature range feature matrix respectively;

[0094] S106: A plurality of one-dimensional convolution network models are constructed, the model input is the voltage range feature matrix, current range feature matrix and temperature range feature matrix obtained in step S105, each row in the voltage range feature matrix, current range feature matrix and temperature range feature matrix is input into the respective one-dimensional convolution neural network, then the extracted features are averaged according to voltage, current and temperature respectively, and finally feature splicing and SOH estimation are performed, as shown below:

[0095]

[0096] SOH = σ(w o · F fusion + b o )

[0097] wherein Conv1D(·) represents a one-dimensional convolution operation, ReLU(·) represents an activation function, represents the mth row of the voltage range feature matrix represents the mth row of the current range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix represents the mth row of the temperature range feature matrix respectively represent the average features of voltage, current and temperature, F fusion represents the fusion feature, Concatenate(·) represents the fusion operation, σ(·) represents the sigmoid activation function, w o , b o respectively represent the weight and bias term of the output network, and SOH is the battery health state output by the network.

[0098] In the present scheme, the target function of the client optimization selection constructed in step S4 is as follows:

[0099] S401: Define a binary decision variable x i,t , which is used to represent the participation state of the client in the federated learning round, and is specifically as follows:

[0100]

[0101] wherein Γ represents the client set, t is the federated learning iteration round index, Epoch is the total round, x i,t = 1 indicates that the client i is selected to participate in the federated learning training in the tth round of federated learning, and x i,t = 0 indicates that the client i is not selected in the tth round of federated learning;

[0102] S402: Based on the historical participation data of the client i, a function for evaluating the potential of the client to improve the performance of the global model is constructed, including a dynamic loss function, a data distribution deviation function, a historical contribution function and a gradient consistency function, which reflects the expected gain of the participation of the client in the tth round of federated learning on the performance of the global model, and is specifically as follows:

[0103]

[0104] wherein, VF i,t represents the improvement potential function of the client i participating in the tth round of federated learning iteration on the global model performance, α1, α2, α3, α4 are weight coefficients, represents the dynamic loss function of the client i before the tth round of federated learning iteration, represents the data distribution deviation function of the client i before the tth round of federated learning iteration, represents the historical contribution degree function of the client i before the tth round of federated learning iteration, represents the gradient consistency function of the client i before the tth round of federated learning iteration;

[0105] The dynamic loss function is specifically as follows:

[0106]

[0107] wherein, W t-1 represents the global model parameter aggregated in the t-1th round, W t-2 represents the global model parameter aggregated in the t-2th round, Loss i (·) refers to the loss value of the specified model parameter on the client i, refers to the local model parameter of the client j after participating in the federated learning training in the t-1th round, refers to the local model parameter of the client j after participating in the federated learning training in the t-2th round, n j refers to the local sample data amount of the client j, N t-1 refers to the total effective sample number in the t-1th round, N t-2 refers to the total effective sample number in the t-2th round, refers to the client set selected in the t-1th round of federated learning, refers to the client set selected in the t-2th round of federated learning;

[0108] The data distribution deviation function is specifically as follows:

[0109]

[0110] wherein, w U , w I , w T respectively represent the weights of voltage, current, and temperature data distribution deviation, W UIT represents the sum of all weight coefficients, W UIT = w U + w I + w T ; K D is the number of grades, EMDU , EMD I , EMD T respectively represent the voltage, current, temperature data distribution deviation, P Ul and Q Ul respectively represent the proportion of client voltage and global voltage at voltage level l, P Il and Q Il respectively represent the proportion of client current and global current at current level l, P Tl and Q Tl respectively represent the proportion of client temperature and global temperature at temperature level l, respectively represent the distance weight between adjacent levels of voltage, current, temperature respectively;

[0111] The historical contribution degree function is as follows,

[0112]

[0113] wherein γ represents the decay parameter of historical contribution; refers to the historical contribution degree of client i before the t-1th round of federated learning iteration; represents the newly added contribution degree of client i in the t-1th round of federated learning, which exists only when client i is selected to participate in the t-1th round of federated learning training, and is 0 when client i is not selected to participate in training; ||·||2 represents L2 norm, refers to the local model parameter of client i after the t-1th round of federated learning training;

[0114] The gradient consistency function is as follows,

[0115]

[0116] wherein g t-1 refers to the global average gradient of the t-1th round of federated learning; refers to the effective gradient of client i participating in the t-1th round of federated learning training, represents the gradient of client i selected for training in the t-1th round of federated learning, represents the gradient of client i not selected for training in the t-1th round of federated learning, which is the historical gradient; refers to the effective gradient of client j participating in the t-1th round of federated learning training; is the number of clients in the set , <·,·> represents the inner product of two gradients; ||·|| represents the norm; exp(·) represents the exponential function;

[0117] S403: Ensure the effectiveness of aggregation by setting the lower limit of the number of selected clients, and avoid uneven distribution of resources by setting dynamic fairness constraints. The specific requirements are as follows:

[0118]

[0119] Where x i,c represents the decision variable state of client i in the cth round of federated learning, K min represents the preset lower limit of the number of selected clients, R represents the rolling round window size of the dynamic fairness constraint, F min represents the minimum number of participations of the client within the window, F max represents the maximum number of participations of the client within the window.

[0120] S404: Based on the potential function VF defined in S402 representing the improvement of the global model performance by client i participating in the tth round of federated learning iteration, and combined with the penalty function reflecting the lower limit of the number of selected clients and the dynamic fairness constraint, a global objective function is constructed, as shown below: i,t

[0121]

[0122] Where f represents the global objective function, |Γ| represents the total number of all clients, and Φ represents the penalty function.

[0123] The penalty function Φ is as follows,

[0124] Φ = Φ1 + Φ2

[0125]

[0126] Where Φ1 represents the penalty term for violating the lower limit of the number of selected clients; Φ2 represents the penalty term for violating the dynamic fairness constraint; ξ1 represents the penalty intensity coefficient for violating the lower limit of the number of selected clients, and ξ2 represents the penalty intensity coefficient for violating the fairness constraint; S i represents the total number of times that client i has been selected in the last R rounds; max(·) represents the maximum of the two;

[0127] In this scheme, the genetic algorithm is used to solve the objective function of client optimization selection in step S5, and the optimal client selection scheme is output. The specific steps are as follows:

[0128] S501: Encode the client participation state decision variable as a binary chromosome, with the chromosome length equal to the total number of clients. The gene bit i corresponds to the activation state of client i, and is recorded as 1 when selected and 0 when not selected.

[0129] S502: Use the global objective function constructed in S404 as the fitness function;​

[0130] S503: Randomly generate an initial population of size P, and generate 1 or 0 for each gene locus of each chromosome by independent random assignment; num

[0131] S504: Calculate the fitness value of each individual in the population using a fitness-based probability selection strategy, select parent individuals by roulette or tournament strategy, preferentially retain high-quality solutions, then perform single-point crossover on paired parents with a preset crossover probability, randomly select a cutting point to exchange gene fragments to generate new offspring; finally, traverse each gene locus of the offspring individuals, and perform bit flip mutation with a low probability to introduce diversity;

[0132] S505: Terminate the genetic algorithm iteration when the maximum number of iterations is reached or the optimal fitness is not improved for consecutive generations, output the chromosome with the highest fitness of all generations as the optimal solution, and obtain the optimal selection scheme for the client.​

Claims

1. A lithium battery SOH estimation method based on federated learning and client multi-dimensional scoring mechanism, characterized in that, The specific steps are as follows: S1: Each battery client participating in training independently completes data acquisition locally, and constructs a local feature data set based on the charge and discharge operation data; the feature data set includes a voltage range feature matrix, a current range feature matrix, and a temperature range feature matrix; Based on the format of the feature data set, a deep learning network model structure is designed and applied to the federated server end and all clients; the deep learning network model takes the voltage range feature matrix, the current range feature matrix, and the temperature range feature matrix as input, and takes the battery SOH of each charge and discharge cycle as output; S2: Initialize the relevant parameters, set the model parameters and average gradient of the federated server end as the global model parameters and global average gradient, set the number of clients, which is used for subsequent model parameter aggregation and update, client selection optimization; S3: The federated server distributes the global model parameters to all clients, and each client performs local training based on local battery data; After training is completed, each client records its own model parameters as initial data for subsequent optimization, and uploads its own model parameters to the federated server; after collecting the model parameters of each client, the federated server aggregates the parameters of all clients to update the global model parameters; S4: Construct a target function for client optimization selection, the specific steps including: defining a binary decision variable representing the participation state of the client; establishing a multi-dimensional performance scoring mechanism that integrates dynamic loss, data distribution deviation, historical contribution, and gradient consistency to evaluate the potential of each client to improve the performance of the global model; finally, maximize the sum of the comprehensive performance scores of the selected clients, and combine the penalty function reflecting the lower limit of the number of selected clients and dynamic fairness constraints to form the target function for client selection; S5: Solve the target function for client optimization selection in step S4 using a genetic algorithm to output the optimal client selection scheme, the specific steps including: encoding all binary decision variables representing the participation state of the client into a binary chromosome with a length equal to the total number of clients, where gene bit values 1 and 0 represent client activation and inactivation, respectively; after randomly generating an initial population, select parents by fitness proportion, and perform single-point crossover and bit mutation operations; terminate the iterative calculation of the genetic algorithm when the maximum number of iterations is reached or the optimal fitness value of the target function is continuously stagnant, and decode the chromosome with the highest fitness value to obtain the optimal client selection scheme; S6: The federated server issues the current global model parameters to all clients, then selects the corresponding clients according to the optimal client selection scheme obtained in S5, so that they train and update the SOH model locally using battery data; after training is completed, the selected clients record their updated model parameters and upload them to the federated server; after collecting the model parameters of all selected clients, the federated server aggregates and updates the global model parameters to complete this round of federated learning iteration; S7: Iteratively execute steps S5 to S6 until the global model parameters on the federated server converge.

2. The lithium battery SOH estimation method based on federated learning and client multi-dimensional scoring mechanism according to claim 1, characterized in that, The constructing local feature data set and designing deep learning network model structure in step S1 are specifically as follows: S101: Each battery client independently completes data acquisition locally, and the collected data includes voltage, current, temperature operation data in battery charge and discharge cycles and battery SOH recorded in each charge and discharge cycle; S102: Normalize the input variables and output variables, and divide the training set and the test set; S103: The normalized voltage, current and temperature are processed by average segmentation. The normalized voltage, current and temperature are respectively divided into segments by fixed length equalization strategy, so that each segment contains the same data points, as shown below: wherein, represents the number of data points of voltage, current or temperature, represents the number of data points contained in each segment, represents the segment number, ; represents the start index of segment , represents the end index of segment , represents the index number of data points, represents the voltage set of segment , represents the voltage of index , represents the current set of segment , represents the current of index , represents the temperature set of segment , represents the temperature of index ; S104: constructing the voltage difference feature, the current difference feature and the temperature difference feature according to the obtained in step S103 , , , the voltage difference feature, the current difference feature and the temperature difference feature are constructed as follows: wherein, , , respectively represent the voltage range feature, the current range feature, the temperature range feature of the segment , represents the maximum value in the set, represents the minimum value in the set; S105: The voltage range feature, the current range feature and the temperature range feature constructed in step S104 are further constructed into a voltage range feature matrix, a current range feature matrix and a temperature range feature matrix, respectively, and are specifically as follows: wherein, , , respectively represent the voltage difference feature, the current difference feature, the temperature difference feature of the segment , , , respectively represent the voltage difference feature, the current difference feature, the temperature difference feature of the segment , , , respectively represent the absolute difference value of the first segment and the second segment difference of voltage, current, temperature, , , respectively represent the voltage difference feature matrix, the current difference feature matrix, the temperature difference feature matrix; S106: A plurality of one-dimensional convolution network models are constructed, the model input is the voltage range feature matrix, the current range feature matrix and the temperature range feature matrix obtained in step S105, each row in the voltage range feature matrix, the current range feature matrix and the temperature range feature matrix is input into a respective one-dimensional convolution neural network, then the extracted features are averaged according to voltage, current and temperature respectively, and finally feature splicing and SOH estimation are performed, and are specifically as follows: wherein, represents a one-dimensional convolution operation, represents an activation function, represents a voltage difference feature matrix of the th row, represents a current difference feature matrix of the th row, represents a temperature difference feature matrix of the th row, represents extracted features, represents extracted features, represents extracted features, , , respectively represent average features of voltage, current, and temperature, represents a fusion feature, represents a fusion operation, represents a sigmoid activation function, , respectively represent weights and bias terms of the output network, and SOH is the battery state of health output by the network.

3. The lithium battery SOH estimation method based on federated learning and client multi-dimensional scoring mechanism according to claim 1, characterized in that, The constructing client optimization selection objective function in step S4 is specifically as follows: S401: Define binary decision variables for characterizing the participation status of the client in a federated learning round, as follows: wherein, represents a set of clients, is an index of a federated learning iteration round, is a total number of rounds, represents a client is selected to participate in federated learning training in the federated learning round, represents a client is not selected in the federated learning round; S402: Client-based Based on historical participation data, a function is constructed to evaluate the client's potential to improve the performance of the global model. This function includes a dynamic loss function, a data distribution bias function, a historical contribution function, and a gradient consistency function, reflecting its participation in the [missing information - likely a specific event or process]. The expected performance gain of each round of federated learning on the global model is as follows: wherein, representing a client participating in the round federated learning iteration, , , is a weight coefficient, representing a client before the round federated learning iteration, representing a client before the round federated learning iteration, representing a client before the round federated learning iteration, representing a client before the round federated learning iteration, The dynamic loss function , as follows: wherein, denotes the global model parameters of the round of federated learning, denotes the global model parameters of the round of federated learning, denotes the loss value of the model parameters specified on the client side, denotes the local model parameters of the client after participating in the federated learning training in the round, denotes the local model parameters of the client after participating in the federated learning training in the round, denotes the local sample data volume of the client , denotes the total effective sample number in the round, denotes the total effective sample number in the round, denotes the client set selected in the round of federated learning, denotes the client set selected in the round of federated learning; the data distribution bias function as follows: wherein, , , respectively represent the weight of voltage, current, temperature data distribution deviation, represent the sum of all weight coefficients, ; is the number of grades, , , respectively represent the voltage, current, temperature data distribution deviation, and respectively represent the proportion of client voltage and global voltage in voltage grade , and respectively represent the proportion of client current and global current in current grade , and respectively represent the proportion of client temperature and global temperature in temperature grade , respectively represent the distance weight between adjacent grades of voltage, current, temperature respectively. The historical contribution function As shown below, wherein, is a decay parameter representing historical contribution; is a client historical contribution degree before the th round of federated learning iteration; is a client newly added contribution degree in the th round of federated learning, which exists only when the client is selected to participate in the th round of federated learning training, and is 0 when the client is not selected to participate in the training; is an L2 norm, is a client local model parameter after the th round of federated learning training; The gradient consistency function , as follows: in, It refers to the first The global average gradient of the round of federated learning; Refers to the client Participating in the Effective gradients for federated learning training Indicates client In the The gradient selected for training in each round of federated learning. Indicates client In the The gradients of the unselected trainees in each round of federated learning are historical gradients. Refers to the client Participating in the Effective gradients for round-fed learning training; It is a set The number of clients in the middle, This represents the inner product of two gradients; Represents the norm; Represents an exponential function; S403: The lower limit of the number of selected clients is set to ensure the effectiveness of aggregation, and the dynamic fairness constraint is set to avoid uneven distribution of resources, and the specific requirements are as follows: wherein, represents the first client of the round federated learning decision variable state, represents the preset lower limit of the number of selected clients, represents the rolling round window size of the dynamic fairness constraint, represents the minimum number of participations of the client within the window, represents the maximum number of participations of the client within the window; S404: defining a representation client based on S402 participate in the round of federated learning iteration , and combined with the penalty function reflecting the lower limit of the number of selected clients and dynamic fairness constraints, the global objective function is constructed, as follows: wherein, represents a global objective function, represents a sum of all client numbers, represents a penalty function; The penalty function , as shown below, wherein, represents a penalty term for violating the lower bound of the number of selected clients; represents a penalty term for violating the dynamic fairness constraint; represents a penalty strength coefficient for violating the lower bound of the number of selected clients, represents a penalty strength coefficient for violating the fairness constraint; represents the client in the last total number of times the client was selected in the round; represents taking the maximum of the two.

4. The lithium battery SOH estimation method based on federated learning and client multi-dimensional scoring mechanism according to claim 1, characterized in that, The genetic algorithm is used to solve the client optimization selection objective function in step S5, and the optimal client selection scheme is output, and the specific steps are as follows: S501: encode the client engagement state decision variable into a binary chromosome, with chromosome length equal to the total number of clients, and gene bits corresponding to the activation state of the client, 1 if selected, 0 if not selected; corresponding to the activation state of the client, 1 if selected, 0 if not selected; S502: The global objective function constructed in S404 is used as the fitness function; S503: randomly generate an initial population with a size of of 1 or 0 for each gene locus of the chromosome; S504: A probability selection strategy based on fitness is used to calculate the fitness value of each individual in the population, parent individuals are selected through roulette or tournament strategy, high-quality solutions are preferentially retained, then paired parents are executed with a preset crossover probability to perform single-point crossover to generate new offspring by randomly selecting a cutting point to exchange gene fragments; finally, each gene site of the offspring individuals is traversed to perform bit flip mutation with a low probability to introduce diversity; S505: When the maximum number of iterations is reached or the optimal fitness of continuous generations is insufficiently improved, the genetic algorithm iteration is terminated, the chromosome with the highest fitness in the history is output as the optimal solution, and the optimal client selection scheme is obtained.

Images