CN-122017628-A - Battery early warning method, system and medium based on unsupervised learning

CN122017628ACN 122017628 ACN122017628 ACN 122017628ACN-122017628-A

Abstract

The invention discloses a battery early warning method, system and medium based on unsupervised learning, and relates to the technical field of battery safety management. The method comprises the steps of synchronously collecting voltage, current and temperature data of each module in a battery pack, constructing a three-dimensional tensor comprising time and module dimensions after time alignment and standardization pretreatment, inputting the tensor into a pre-trained variable self-encoder model, mapping the data to a potential space by an encoder to obtain hidden variables, reconstructing an ideal voltage representing the health state of the battery by a decoder, calculating reconstruction errors of measured voltage and reconstruction voltage of each module, generating a dynamic threshold value based on group statistical characteristics of all module reconstruction errors, and judging abnormality and triggering early warning when the reconstruction errors of specific modules continuously exceed the threshold value in a continuous time window. The invention has the advantages of no need of fault samples, high sensitivity, low false alarm rate and high reliability, and can effectively capture the early weak faults of the battery.

Inventors

LI LEI
ZANG HAITAO
LIU PENG
ZHANG ZHAOSHENG
WANG SIRUI
YANG YULING

Assignees

北京理工大学

Dates

Publication Date: 20260512
Application Date: 20260224

Claims (9)

1. The battery early warning method based on the unsupervised learning is characterized by comprising the following steps of: S1, data acquisition and synchronization, namely synchronously acquiring voltage data, total current data and temperature data of a plurality of modules in a battery pack, and performing time alignment processing on all the data to form a time-synchronous multichannel data sequence; S2, constructing a space-time characteristic tensor, namely preprocessing the multichannel data sequence, and constructing a three-dimensional tensor simultaneously comprising a time dimension and a module number dimension by utilizing a sliding window method; S3, unsupervised voltage reconstruction, namely inputting the three-dimensional tensor into a pre-trained variational self-encoder model, mapping input data into a potential space by an encoder of the model to obtain hidden variables, and reconstructing ideal voltage data representing the health state of the battery by a decoder of the model based on the hidden variables; s4, calculating a reconstruction error, namely calculating the difference between an original voltage sequence and a reconstruction voltage sequence of each module to obtain a corresponding reconstruction error; s5, generating a dynamic threshold and judging abnormality, namely calculating group statistical characteristics according to the reconstruction errors of all current modules, generating a dynamic threshold based on the group statistical characteristics, and judging that an abnormality exists in a certain module and triggering early warning when the reconstruction error of the module continuously exceeds the dynamic threshold in a plurality of continuous time windows.
2. The method for battery early warning based on unsupervised learning according to claim 1, wherein the time alignment process in S1 includes assigning uniform time stamps to all collected data and filling the data with delay or loss by interpolation.
3. The battery early warning method based on unsupervised learning according to claim 1, wherein the encoder of the variation self-encoder model in S3 is implemented by a long-short-term memory network or a one-dimensional convolutional neural network.
4. The battery early warning method based on the unsupervised learning according to claim 1, wherein the hidden variable in S3 is obtained by re-parameterization, specifically, the hidden variable is obtained by linearly combining a probability distribution parameter outputted by an encoder with an external noise variable sampled from a standard normal distribution.
5. The battery early warning method based on the unsupervised learning according to claim 1, wherein the output layer of the decoder of the variational self-encoder model in S3 adopts a linear activation function.
6. The method of claim 1, wherein the variational self-encoder model is trained using historical operating data of the battery in a normal state of health.
7. The battery early warning method based on the unsupervised learning according to claim 1, wherein the generation of the dynamic threshold in S5 is specifically that a mean value and a standard deviation of all module reconstruction errors are calculated, and a preset multiple of the mean value and the standard deviation is added as the dynamic threshold.
8. A battery early warning system based on unsupervised learning for implementing the battery early warning method based on unsupervised learning as claimed in any one of claims 1 to 7, comprising: the data acquisition and preprocessing module is configured to synchronously acquire operation data of a plurality of modules in the battery pack, perform time alignment and standardization preprocessing on the data, and construct a three-dimensional tensor comprising a time dimension and a module number dimension; the model reasoning and reconstruction module is internally provided with a pre-trained variational self-encoder model and is configured to receive the three-dimensional tensor and reconstruct ideal voltage data representing the health state of the battery, and the reconstruction error of each module is calculated; The abnormality judgment and alarm module is configured to generate a dynamic threshold value based on the reconstruction errors of all the modules, and trigger an early warning signal when the reconstruction errors of the specific modules are judged to continuously exceed the dynamic threshold value.
9. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements a battery pre-warning method based on unsupervised learning as claimed in any one of claims 1 to 7.

Description

Battery early warning method, system and medium based on unsupervised learning Technical Field The invention relates to the technical field of battery safety management, in particular to a battery early warning method, system and medium based on unsupervised learning. Background The power battery is used as a core component of the new energy automobile, and the health state of the power battery is directly related to the endurance mileage, reliability and safety of the whole automobile. In practical application, the power battery pack is formed by connecting hundreds or thousands of battery cells in series and parallel, and due to the differences of manufacturing process, working environment, use load and the like, the performance attenuation among the battery cells is difficult to keep synchronous, so that the inconsistency of parameters such as voltage, internal resistance and the like is caused. This inconsistency, especially the deterioration of voltage consistency, is an early key sign of serious safety hazards such as battery performance degradation and thermal runaway. Therefore, accurate monitoring and early warning of battery consistency are important issues in the field of battery safety management. Currently, the industry is faced with significant challenges in battery state of health assessment and early warning. First, many prior art techniques rely too much on a single dimension of battery capacity for health assessment. However, capacity fade is a relatively slow macroscopic process, and dynamic performance differences caused by early faults such as micro-short circuit and lithium precipitation in the battery cannot be sensitively reflected, so that it is difficult to comprehensively and accurately judge the real aging state of the battery. Secondly, although the voltage inconsistency is a key index for representing battery abnormality, the prior art lacks a method capable of effectively and quantitatively analyzing and utilizing the coupling inconsistency, particularly, key features are accurately extracted from massive operation data, working condition interference is overcome, and stable monitoring of weak voltage deviation is realized, so that the current technical difficulty is still realized. To address the above challenges, various technical solutions have been proposed in the industry, but all have limitations: patent CN113960484a proposes a health diagnostic method based on a linear fit of monomer differential pressure. The method evaluates by counting the voltage differences and fitting their linear trend. However, early battery failures tend to exhibit complex nonlinear characteristics over the voltage sequence, which are difficult to capture effectively with linear models, resulting in insufficient sensitivity. Meanwhile, the adopted fixed threshold filtering mode can misreject effective fault information, and simple algebraic correction fails to model the dynamic coupling relation of multidimensional parameters such as voltage, current and the like. Patent CN120910721a proposes an evaluation framework based on causal inference. In order to realize causal inference, the method needs to highly aggregate dynamic time sequence data into static indexes, so that the complete track of the evolution of parameters such as voltage and the like along with time is thoroughly lost, and early signs of battery safety risk are just hidden in the time sequence details, so that the requirement of real-time fine monitoring cannot be met. Patent CN121106303a proposes an energy management method based on a fusion of a transducer reinforcement learning and a battery mechanism model. The bottleneck of the method is that the computational complexity of the transducer model is high, and the computational force requirement of the vehicle-gauge real-time decision is difficult to meet. Meanwhile, the performance of the system is highly dependent on the parameter precision of a built-in battery model, the parameters are difficult to accurately calibrate in real time in practical application, new risks are introduced due to parameter mismatch, and uncertainty exists in generalization capability of an offline training strategy. In summary, the prior art solution presents a stepwise limitation when dealing with the challenges of battery safety management, from insufficient sensitivity of simple linear method, to time sequence information loss of advanced statistical framework, to deployment difficulty and generalization risk of complex intelligent model. These limitations together reveal a contradiction between the current state of the art and the high standard requirements of vehicle safety management for early warning sensitivity, state estimation finesse and algorithm real-time reliability. Therefore, there is an urgent need in the art for a new early-stage safety pre-warning scheme for a battery that can be highly sensitive, robust, and easy to deploy. Disclosure of Invention The invention aims to provide a