CN-122017647-A - Energy storage battery health state monitoring and managing method and system

CN122017647ACN 122017647 ACN122017647 ACN 122017647ACN-122017647-A

Abstract

The application relates to a method and a system for monitoring and managing the state of an energy storage battery, wherein the method comprises the steps of constructing a multidimensional physical feature tensor according to charge and discharge data of the battery, establishing a KAN network integrating a physical model, inputting the multidimensional physical feature tensor into the KAN network for joint pre-training, updating spline coefficients in the KAN network and physical parameters in the physical model, performing first-order Taylor expansion on the spline coefficients based on a total loss function of the joint pre-training, calculating physical sensitivity of each connection in the KAN network, performing structural pruning on the KAN network according to the physical sensitivity, screening effective input features in the multidimensional physical feature tensor according to a structural pruning result, and performing subspace fine-tuning on the structural pruned KAN network to obtain a sparse neural network model for estimating the state of the battery.

Inventors

HE QINGQING
TANG CHAOYANG
REN DAN
TANG CHAO
ZHOU KELIANG
YIN XIN
XIAO FAN

Assignees

武汉理工大学

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (8)

1. The energy storage battery health state monitoring and managing method is characterized by comprising the following steps: S1, constructing a multidimensional physical characteristic tensor according to charge and discharge data of a battery; S2, establishing a KAN network fused with a physical model, inputting the multidimensional physical feature tensor into the KAN network for joint pre-training, and synchronously updating spline coefficients in the KAN network and physical parameters in the physical model, wherein the physical model is a Verhulst model for describing battery capacity attenuation characteristics, and the expression is as follows: ; Wherein t is the number of charge and discharge cycles, u (t) is the battery capacity loss rate at the time t, K is the decay rate constant, M is the maximum theoretical capacity loss, k and M are embedded into the network as learnable physical parameters; S3, performing first-order Taylor expansion on the spline coefficients based on a total loss function of the joint pre-training, and calculating physical sensitivity of each connection in the KAN network; s4, structural pruning is carried out on the KAN network according to the physical sensitivity, and effective input features in the multidimensional physical feature tensor are screened according to the structural pruning result; S5, carrying out subspace fine adjustment on the KAN network after structure pruning to obtain a sparse neural network model for battery health state estimation, outputting a high-precision SOH estimated value through the model, synchronously outputting an aging rate constant and a maximum capacity loss parameter, and providing an additional physical diagnosis basis for quality analysis of a full life cycle of the battery; The total loss function comprises data fitting loss and physical residual loss, wherein the data fitting loss is a mean square error between a predicted value of the KAN network and a real battery health state value, and the physical residual loss is constructed according to a deviation between the Verhulst model and a predicted output of the KAN network: the total loss function L is: ; Wherein L is the total loss function, As the mean square error of the network predicted value and the true SOH value, In order for the physical residual error to be lost, The weight coefficient lost for the physical residual error, D is the sample batch size of the training data, m is the index of the sample data, The true capacity loss rate for the mth sample, The capacity loss rate predicted for the network for the mth sample, As a physical residual function.
2. The method of claim 1, wherein constructing a multi-dimensional physical feature tensor from charge and discharge data of the battery comprises extracting at least two of a cycle number, an ohmic internal resistance, a constant current charge duration, a constant voltage charge duration, a highest discharge process temperature and an average discharge process voltage from the charge and discharge data, and normalizing the extracted features to form the multi-dimensional physical feature tensor.
3. The method of claim 1, wherein connections between nodes of adjacent layers in the KAN network employ parameterized B-spline basis functions, and wherein an output that a j-th node of a first layer of the KAN network passes to an i-th node of a first +1 layer through the connection function is represented as: ; Wherein, the As a function of the connection between the j-th node of the first layer and the i-th node of the (l+1) -th layer, For the output value of the j-th node of the first layer, N is the total number of spline basis functions in the join function, N is the index of the B-spline basis function, B n is the N-th basis function in the join function, A learnable spline coefficient that is the nth basis function in the join function.
4. The method of claim 1, wherein the physical sensitivity is calculated as follows: ; Wherein, the For the learnable spline coefficients of the nth basis function in the join function, As a function of the total loss, The gradient of the spline coefficients for the total loss function, A physical sensitivity score for the connection between the jth node of the first layer and the i node of the (l+1) th layer.
5. The method of claim 4, wherein said structural pruning of the KAN network based on the physical sensitivity comprises: Calculating the average value of the physical sensitivity scores of all the connections of the current layer; determining a pruning threshold according to the product of the average value and the pruning intensity adjusting parameter; And generating a binary mask matrix, setting a mask corresponding to the connection with the physical sensitivity score lower than the pruning threshold value to be zero, and cutting off a corresponding signal transmission path.
6. The method of claim 5, wherein the filtering valid input features in the multidimensional physical feature tensor according to the result of structural pruning includes checking output connection states of feature nodes in the KAN network input layer, and in response to detecting that all output connections of any feature node are pruned, determining input features corresponding to any feature node as physical irrelevant features and rejecting.
7. The method of claim 1, wherein the subspace tuning of the KAN network after structural pruning comprises freezing a binary mask matrix generated by structural pruning, and performing fine tuning training on the retained spline coefficients and the physical parameters in the physical model under a fixed sparse topology.
8. The battery management system is characterized by comprising a data acquisition module, a central processing module and a storage module; The data acquisition module is used for acquiring voltage, current and temperature data of the battery; The neural network model trained by the method of any one of claims 1-7 is preset in the storage module; The central processing module is used for constructing a multidimensional physical characteristic tensor according to the data acquired by the data acquisition module and calling the neural network model to monitor the health state of the battery.

Description

Energy storage battery health state monitoring and managing method and system Technical Field The application relates to the technical field of battery monitoring, in particular to a method and a system for monitoring and managing the state of an energy storage battery. Background The lithium ion battery has been widely used in the fields of electric automobiles, energy storage power stations, portable electronic devices and the like because of the advantages of high energy density, long cycle life and the like. In the long-term use process of the battery, the battery is influenced by factors such as temperature fluctuation, change of charge-discharge multiplying power, electrochemical side reaction and the like, irreversible capacity attenuation and impedance increase can occur in the battery, and the available capacity of the battery is gradually reduced. Accurate estimation of the State of Health (SOH) of the battery is critical to ensuring safe operation of the powertrain system and to developing reasonable maintenance strategies. Currently, methods for estimating the state of health of a battery can be broadly classified into two major types, namely, a method based on an equivalent circuit model or an electrochemical mechanism model and a data driving method based on machine learning. Although the method based on the mechanism model has definite physical interpretation capability, the on-line identification process of model parameters has high calculation complexity and insufficient real-time performance, the data driving method obtains good prediction precision by means of fitting capability of the data driving method to a high-dimensional nonlinear relation, but the method lacks explicit constraint on a battery attenuation physical rule in the training process, and easily generates a prediction result against electrochemical common knowledge when facing noise interference or training data distribution deviation. In order to achieve both physical interpretability and data fitting capability, researchers have tried to introduce a growth model describing population growth saturation characteristics into battery degradation analysis so as to represent an S-shaped saturation attenuation trend of battery capacity presented along with cycle times, and use a model equation as a training process of soft constraint embedded neural network to construct a hybrid prediction framework with physical significance and nonlinear modeling capability. In addition, the novel neural network structure based on the Kerr Mo Ge Roche-Arnod representation theorem shows higher function approximation efficiency by configuring a learnable spline function at the edge of the network to replace the traditional fixed activation function, and provides a novel network carrier for the mixed framework. However, such network models that incorporate physical constraints often maintain a fully connected dense topology when built, often with parameter volumes on the order of tens of thousands or even higher, and there are practical difficulties in deploying such dense models directly for embedded Battery Management System (BMS) chips that have very limited computational power and storage resources. For contradiction between model and hardware resources, one intuitive way of handling is to perform structural pruning on the trained network to cut down the parameter scale. However, existing network pruning methods still face several practical difficulties when applied to hybrid predictive models that contain physical constraints. On the one hand, after the parameter cutting with a larger proportion is executed, abnormal fluctuation which is inconsistent with the electrochemical aging trend of the battery often occurs in the prediction curve of the model, for example, the non-physical phenomenon of capacity rebound and the like which violates the monotonic attenuation rule is presented in the later stage of the degradation, and the problem is more prominent along with the improvement of the compression rate, so that the reliability of the output of the model is seriously weakened. On the other hand, after the internal structure of the network is simplified, the existing scheme still executes complete acquisition, transmission and preprocessing processes on the data of all original sensor channels in the system operation stage, so that the hardware of the front end of the BMS bears unnecessary calculation power and storage expenditure in the links of data throughput, analog-to-digital conversion, feature calculation and the like, and the conduction of light-weight practical benefit to the whole machine level is restricted. The above-mentioned deficiencies make engineering deployment of neural network models with physical constraints on resource-constrained embedded battery management platforms still challenging. Disclosure of Invention In view of the above, the embodiment of the application provides a neural network model training method and a ba