CN-122017575-A - Method and system for estimating battery health state based on bidirectional attention fusion

CN122017575ACN 122017575 ACN122017575 ACN 122017575ACN-122017575-A

Abstract

The invention provides a battery health state estimation method and system based on bidirectional attention fusion in the technical field of battery management and state estimation, and the method comprises the steps of S1, collecting a large amount of charge and discharge time sequence data and priori information, extracting local features and global features from each charge time sequence data, S2, constructing natural language description based on each priori information, local features and global features, constructing a data set based on each natural language description and charge time sequence data, S3, creating a battery health state estimation model based on a semantic embedding module, a time sequence embedding module, an explicit alignment module, a bidirectional attention fusion module and a prediction module, S4, training the battery health state estimation model through the data set, and S5, estimating the battery health state through the trained battery health state estimation model. The method has the advantage that the accuracy, the robustness and the practicability of SOH estimation are greatly improved.

Inventors

CHEN WENPING
LIANG QIHUI
YE YING

Assignees

福建星云软件技术有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. A battery state of health estimation method based on bidirectional attention fusion is characterized by comprising the following steps: Step S1, collecting a large amount of charging time sequence data of batteries in the charging process of different multiplying powers and prior information of the batteries, performing step division and statistics on each charging time sequence data, and extracting local features and global features; s2, constructing natural language description based on the prior information, the local features and the global features, and constructing a data set based on the natural language description and charging time sequence data; step S3, a battery health state estimation model is established based on the semantic embedding module, the time sequence embedding module, the explicit alignment module, the bidirectional attention fusion module and the prediction module, and a loss function of the battery health state estimation model is set; s4, training a battery health state estimation model through the data set and the loss function; And S5, estimating the state of health of the battery through the trained battery state of health estimation model.
2. The method for estimating a battery state of health based on bidirectional attention fusion as set forth in claim 1, wherein in the step S1, the charging time series data includes at least a charging voltage, a charging current, a charging temperature, and an SOC; the prior information at least comprises battery rated capacity, battery rated voltage, battery cycle times, charge and discharge depth and service environment temperature The local features and the global features comprise a charging duration, a voltage interval, a temperature interval and an SOC interval.
3. The method for estimating battery state of health based on bi-directional attention fusion as set forth in claim 1, wherein in step S3, the semantic embedding module is configured to extract a semantic embedding vector from a natural language description through a large language model; the time sequence embedding module is used for extracting time sequence embedding vectors from the charge and discharge time sequence data through a time sequence model; The explicit alignment module is used for performing explicit alignment on the semantic embedded vector and the time sequence embedded vector in a potential space through a contrast learning algorithm; the bidirectional attention fusion module is used for executing bidirectional cross-modal attention fusion on the semantic embedded vector and the time sequence embedded vector which are subjected to explicit alignment to obtain a comprehensive embedded vector; The prediction module is used for outputting a battery state of health estimation result according to the comprehensive embedded vector.
4. The method for estimating a battery state of health based on bi-directional attention fusion as set forth in claim 1, wherein in said step S3, the formula of said loss function is: ; ; Wherein, the Representing the total loss value of the loss function; Representing a loss of alignment; Representing SOH supervision loss; Representing hyper-parameters for balancing alignment loss and SOH supervision loss, D () representing cosine similarity function, stopgrad () representing stop gradient; Representing a characteristic representation of the time sequence embedded vector after passing through the predictor; representing a feature representation of the semantic embedded vector after passing through the predictor; Representing a characteristic representation of the time sequence embedded vector projected by the projector; representing a feature representation of the semantic embedded vector after it is projected by the projector.
5. The method for estimating a battery state of health based on bidirectional attention fusion as set forth in claim 1, wherein in the step S3, the fusion process of the bidirectional attention fusion module is as follows: Taking the explicitly aligned semantic embedded vector as a Query1, taking the time sequence embedded vector as a Key1 and a Value1, and fusing the Query1, the Key1 and the Value1 to obtain a fused representation h1; taking the time sequence embedded vector after explicit alignment as Query2, taking the semantic embedded vector as Key2 and Value2, and fusing the Query2, key2 and Value2 to obtain fusion representation h2; setting a learning parameter alpha as a dynamic weight, and fusing the fusion representation h1 and the fusion representation h2 to obtain a comprehensive embedded vector: complex embedding vector = α+h1+ (1- α) h2.
6. A battery state of health estimation system based on bidirectional attention fusion is characterized by comprising the following modules: The data acquisition module is used for acquiring a large number of charging time sequence data of the battery and prior information of the battery in the charging process of different multiplying powers, performing step division and statistics on each charging time sequence data, and extracting local features and global features; The data set construction module is used for constructing natural language descriptions based on the prior information, the local features and the global features and constructing a data set based on the natural language descriptions and charging time sequence data; a battery state of health estimation model creation module for creating a battery state of health estimation model based on the semantic embedding module, the time sequence embedding module, the explicit alignment module, the bidirectional attention fusion module, and the prediction module, setting a loss function of the battery state of health estimation model; The battery state of health estimation model training module is used for training the battery state of health estimation model through the data set and the loss function; And the battery state of health estimation module is used for estimating the state of health of the battery through the trained battery state of health estimation model.
7. The system for estimating a battery state of health based on bi-directional attention fusion of claim 6, wherein said charging time series data comprises at least a charging voltage, a charging current, a charging temperature and an SOC; the prior information at least comprises battery rated capacity, battery rated voltage, battery cycle times, charge and discharge depth and service environment temperature The local features and the global features comprise a charging duration, a voltage interval, a temperature interval and an SOC interval.
8. The battery state of health estimation system based on bi-directional attention fusion of claim 6, wherein in said battery state of health estimation model creation module, said semantic embedding module is configured to extract semantic embedding vectors from natural language descriptions through a large language model; the time sequence embedding module is used for extracting time sequence embedding vectors from the charge and discharge time sequence data through a time sequence model; The explicit alignment module is used for performing explicit alignment on the semantic embedded vector and the time sequence embedded vector in a potential space through a contrast learning algorithm; the bidirectional attention fusion module is used for executing bidirectional cross-modal attention fusion on the semantic embedded vector and the time sequence embedded vector which are subjected to explicit alignment to obtain a comprehensive embedded vector; The prediction module is used for outputting a battery state of health estimation result according to the comprehensive embedded vector.
9. The system for estimating battery state of health based on bi-directional attention fusion of claim 6, wherein in said model for estimating battery state of health creation module, the formula of said loss function is: ; ; Wherein, the Representing the total loss value of the loss function; Representing a loss of alignment; Representing SOH supervision loss; Representing hyper-parameters for balancing alignment loss and SOH supervision loss, D () representing cosine similarity function, stopgrad () representing stop gradient; Representing a characteristic representation of the time sequence embedded vector after passing through the predictor; representing a feature representation of the semantic embedded vector after passing through the predictor; Representing a characteristic representation of the time sequence embedded vector projected by the projector; representing a feature representation of the semantic embedded vector after it is projected by the projector.
10. The battery state of health estimation system based on bi-directional attention fusion of claim 6, wherein in the battery state of health estimation model creation module, the fusion process of the bi-directional attention fusion module is: Taking the explicitly aligned semantic embedded vector as a Query1, taking the time sequence embedded vector as a Key1 and a Value1, and fusing the Query1, the Key1 and the Value1 to obtain a fused representation h1; taking the time sequence embedded vector after explicit alignment as Query2, taking the semantic embedded vector as Key2 and Value2, and fusing the Query2, key2 and Value2 to obtain fusion representation h2; setting a learning parameter alpha as a dynamic weight, and fusing the fusion representation h1 and the fusion representation h2 to obtain a comprehensive embedded vector: complex embedding vector = α+h1+ (1- α) h2.

Description

Method and system for estimating battery health state based on bidirectional attention fusion Technical Field The invention relates to the technical field of battery management and state estimation, in particular to a battery health state estimation method and system based on bidirectional attention fusion. Background With the wide application of electric automobiles and energy storage systems, the safety and reliability of lithium batteries are becoming increasingly an industry key issue. The State of Health (SOH) is used as a core index for measuring the degradation degree of the battery performance, and the accurate estimation of the State of Health (SOH) has important values for battery life prediction, operation and maintenance policy optimization and safety early warning. At present, the SOH estimation method can be mainly classified into two types, namely a traditional modeling method based on time series data and a semantic analysis method based on a large language model (Large Language Model, LLM). The first type of method mainly builds an estimation model by means of a cyclic neural network (RNN), a long-term short-term memory network (LSTM) or a transducer according to voltage, current, temperature, state of Charge (SOC) and other signals acquired in the charging and discharging processes of the battery. The method can effectively capture dynamic change characteristics in the battery operation process, such as an evolution rule of a charge-discharge curve, but usually does not fully consider static background information such as rated parameters (such as nominal capacity and rated voltage), user charging habits, environmental temperature and the like, so that an estimation result has a certain limitation in the aspect of comprehensiveness. The second category of methods is the emerging semantic modeling approach based on large language models in recent years. The method converts rated parameters, using habits and statistical characteristics of the charging process of the battery into natural language description, and further generates semantic embedded vectors by using LLM so as to fuse unstructured priori knowledge. For example, the rated capacity, the number of cycles, the typical charging pattern, etc. of the battery may be entered into the LLM in the form of text prompts to extract high-level semantic features. Although the method can introduce rich context information, the method often cannot be fully and deeply fused with time sequence features, and obvious modal differences exist between the two types of features, so that the precision and the practical application effect of SOH prediction are restricted. The existing method still faces a plurality of key challenges in achieving multi-modal information fusion. Firstly, a modal alignment mechanism is imperfect, a traditional fusion mode such as a unidirectional cross attention mechanism generally uses a time sequence feature as a query vector, LLM semantic features as key value pairs, and the lack of explicit alignment constraint leads to differences in semantic space distribution of different modalities, and affects the consistency and effectiveness of feature fusion. Secondly, the fusion mechanism is single, the existing method depends on a unidirectional attention mechanism, bidirectional interaction of time sequence information and semantic information is difficult to realize, complementary relation between the time sequence information and the semantic information cannot be fully mined, and for example, dependence and reverse influence of time sequence dynamic on semantic background cannot be modeled at the same time. In addition, the system robustness is insufficient, when partial modal data is missing or the quality is poor (such as sensor failure and incomplete user input information), the existing method is difficult to maintain reliable SOH estimation capability, and the applicability of the method in a real complex scene is limited. Therefore, how to provide a battery state of health estimation method and system based on bidirectional attention fusion, so as to improve accuracy, robustness and practicability of SOH estimation, is a technical problem to be solved urgently. Disclosure of Invention The invention aims to solve the technical problem of providing a battery health state estimation method and system based on bidirectional attention fusion, which can improve the accuracy, the robustness and the practicability of SOH estimation. In a first aspect, the present invention provides a battery state of health estimation method based on bidirectional attention fusion, including the steps of: Step S1, collecting a large amount of charging time sequence data of batteries in the charging process of different multiplying powers and prior information of the batteries, performing step division and statistics on each charging time sequence data, and extracting local features and global features; s2, constructing natural language descrip