Search

CN-122017602-A - Lithium battery health state prediction method and device based on dynamic coupling mechanism

CN122017602ACN 122017602 ACN122017602 ACN 122017602ACN-122017602-A

Abstract

The invention provides a lithium battery health state prediction method and device based on a dynamic coupling mechanism, wherein the method comprises the following steps: the multi-dimensional operation data of the lithium battery are collected through the BMS and preprocessed, the multi-scale characteristics are extracted, a weight self-adaptive enhancement prediction model is constructed, time sequence attention weighting is applied to the multi-scale characteristics, small sample learning and characteristic data enhancement coupling characteristic calibration branches and online increment learning parameter updating branches are embedded, data sufficiency, prediction errors and distribution change indexes are calculated in real time, and double-branch weights are adjusted in a self-adaptive mode through a dynamic weight fusion layer. And obtaining an initial SOH predicted value through weighted fusion. And outputting a final SOH predicted value through physical constraint built in the model and self-adaptive noise robust Kalman filtering error correction after optimization, and combining a threshold value to realize health assessment and early warning.

Inventors

  • JIANG LEI
  • WANG YEQIN
  • WANG HAO
  • Han Mingen
  • CHEN YONGHAO
  • GUO ZHIHANG
  • PENG TIAN
  • ZHANG CHU

Assignees

  • 淮阴工学院

Dates

Publication Date
20260512
Application Date
20260402

Claims (10)

  1. 1. The lithium battery health state prediction method based on the dynamic coupling mechanism is characterized by comprising the following steps of: step 1, collecting multi-dimensional operation data of a lithium battery through a battery management system BMS, preprocessing the data, and constructing a lithium battery health monitoring data set; step 2, extracting local time sequence features by adopting an Attention convolution converter Attention-ConvTransformer, extracting global time sequence features by combining with an improved time sequence global discriminant analysis T-GLDA, and horizontally splicing a local feature matrix and a global feature matrix to form a multi-scale feature matrix; Step 3, constructing a weight self-adaptive enhancement prediction model WA-TA-LightGRU-PC, embedding a small sample learning and characteristic data enhancement coupled characteristic calibration branch, an online increment learning parameter updating branch and a Light-weight multi-layer perceptron Light-MLP dynamic weight fusion layer into the weight self-adaptive enhancement prediction model, calculating data sufficiency, prediction errors and distribution change indexes in real time, dynamically outputting double-branch weights and weighting fusion to obtain an initial predicted value; Step 4, rationality constraint is carried out on the initial predicted value through a physical constraint layer built in the weight self-adaptive enhancement prediction model, error correction is carried out on the optimized self-adaptive noise robust Kalman filter ANR-KF, and a final health state SOH predicted value is output; and 5, evaluating the health grade of the lithium battery and generating a corresponding early warning signal according to the SOH predicted value of the health state and a preset threshold value.
  2. 2. The method of claim 1, wherein in step 1, the lithium battery multi-dimensional operational data comprises charge curve data, discharge curve data, environmental data, and cyclic aging data; The preprocessing comprises abnormal value elimination based on 3 sigma criterion, missing value filling based on K neighbor interpolation and maximum and minimum normalization processing, wherein the normalization formula is as follows: , Wherein, the As the raw data is to be processed, 、 Respectively minimum and maximum values of the data set characteristics, Normalized data; the data sets are proportionally divided into a training set, a verification set and a test set, wherein the training set is used for model parameter learning, the verification set is used for super-parameter optimization, and the test set is used for model performance evaluation.
  3. 3. The method according to claim 2, wherein step 2 comprises: the Attention-ConvTransformer extracts local time sequence characteristics, namely extracting local characteristics comprising voltage inflection points and a charging platform through 3 layers of convolution layers with convolution kernel sizes of 3 multiplied by 3,5 multiplied by 5 and 3 multiplied by 3, wherein the step sizes of the convolution layers are 1, the filling mode is same, and a 2-head Attention mechanism is introduced to dynamically adjust characteristic weights, and the formula is as follows: , Wherein, the , , , In order to convolve the output feature matrix, 、 、 In order for the weight matrix to be learnable, 、 、 Respectively a query matrix, a key matrix and a value matrix, which are obtained by linear transformation of convolution output characteristics, In order to pay attention to the head dimension, To query the transposed product of the matrix and the key matrix, In order for the scaling factor to be a factor, To activate the function, the weights are normalized, The local feature matrix is output; representing an attention calculation; Representing a convolution; The improved time sequence global discriminant analysis T-GLDA extracts global time sequence features, wherein a time weight factor is introduced to optimize intra-class and inter-class dispersion matrixes, and a global aging rule in a time sequence dimension is captured, wherein the formula is as follows: , , , , Wherein, the As a global average of the time sequence, For the time weight factor of the i-th sample, For the current moment of time, For the instant of acquisition of the ith sample, As a coefficient of the decay in time, For the feature vector of the i-th sample, In order to spread the matrix within the time series class, In order to spread the matrix among the time series classes, For the total number of aging stage categories, As the time weight of the c-th category, For the number of samples of category c, Is the eigenvector of the ith sample in class c, For the time series mean value of class c, superscript Representing matrix transposition, e representing natural constants, solving Corresponding feature vectors of the first 64 maximum feature values, projecting the original data to obtain a global feature matrix ; Feature fusion horizontal stitching And (3) with Forming a multi-scale feature matrix Dimension (dimension) 。
  4. 4. The method according to claim 3, wherein in step 3, the weight adaptive enhancement prediction model WA-TA-LightGRU-PC includes a time series attention layer, a feature calibration branch, a parameter update branch, and a Light-MLP dynamic weight fusion layer; The time sequence attention layer carries out time sequence attention weighting on the multi-scale feature matrix F, and highlights key time sequence features, and the formula is as follows: , Wherein, the To perform a time-sequential position encoding operation on the multi-scale feature matrix F, Is that And (3) with Is used to determine the matrix product of (c), Is that And (3) with The transposed matrix after the product is used, As a dimension of the features, The characteristic matrix after time sequence weighting is adopted; Feature calibration branch embedding ProtoNet small sample prototype learning and aging stage prototype feature construction based on real small sample The feature data enhancement is realized through feature disturbance and interpolation, the similarity between the input features and the prototype features is calculated, and the feature weights are dynamically adjusted, wherein the formula is as follows: , , , Wherein, the As a prototype feature of the class k aging stage, For a k-th aging stage small sample set Is used for the number of samples of (a), In order to be a sample in a small sample set, As a feature vector of the object set, In the form of an SOH tag, Is that The feature mapping result after passing through the time sequence attention layer, For sum operation; In order to enhance the post-feature matrix, In order to enhance the operation of the features, As the coefficient of the disturbance(s), To pair(s) Is used for the L2 normalization operation of (2), To pair(s) And (3) with Is a linear interpolation operation of (1); Obtaining SOH predicted value of characteristic calibration branch by 2-layer lightweight full-connection layer mapping ; The parameter updating branch adopts an online sequence updating mechanism, only updates the weight of an output layer in real time, and embeds micro Bayesian optimization to dynamically adjust the updating rate, wherein the formula is as follows: , , Wherein, the For the output layer weight matrix at time t, In order to update the rate of the data, Is the true SOH value at time t-1, The predicted SOH value of the branch is updated for the time-1 parameter, The feature matrix is weighted for the time sequence at time t, Is that Is used to determine the transposed matrix of (a), Updating the SOH predicted value of the branch for the parameter at the t-th moment; Light-MLP dynamic weight fusion layer input data sufficiency Prediction error And distribution variation Outputting the double-branch weight And Satisfies the following conditions Weighting to obtain an initial predicted value: , , Wherein, the For the mapping operation of the lightweight multi-layer perceptron, via Obtaining normalized characteristic calibration branch weight after activation And parameter update branch weights , Is the initial predicted value.
  5. 5. The method of claim 4, wherein in step 3, the data sufficiency is The product of the effective data duty ratio and the distribution uniformity score is calculated, and the formula is as follows: , The effective sample number is the number of qualified samples after pretreatment at the current moment, and the theoretical sample number of the whole life cycle is the total expected sample number in the whole life cycle of the lithium battery; prediction error For the absolute error of the normalized SOH prediction, the formula is: , Wherein, the The final SOH predicted value at the t-1 time; data distribution variation The distribution difference of the newly added data and the historical data is measured by Wasserstein distance, and the formula is as follows: , Wherein, the For a joint distribution set of newly added data distribution and historical data distribution, To be in a collection Is performed to find the operation that minimizes the target value, Is distributed for the combination of The following variables The desire is to be found that, For the newly added data samples, In order to sample the historical data, Is that And (3) with Is a L2 norm of (c).
  6. 6. The method according to claim 5, wherein in step 4, the constraint conditions of the physical constraint layer include SOH ε [0,1], SOH attenuation of single cycle of 0.5% or less, and capacity attenuation constraint based on SEI film growth mechanism: , Wherein, the As the rate of change of SOH over time, In order for the attenuation coefficient to be a factor, In order to charge and discharge the electric current, As a factor of the influence of the current, For activation energy, R is the gas constant, T is absolute temperature, Is Arrhenius factor; constraint conditions are fused into the model loss function through a regularization term: , Wherein, the As a function of the total loss, To predict SOH value And true SOH value Is a mean square error loss of (a), In order for the coefficient of balance to be present, Is a physical constraint loss.
  7. 7. The method of claim 6, wherein in step 4, the error correction is calculated by the formula: , , , , Wherein, the For the adaptive observation noise covariance matrix at time t, In order to adaptively update the coefficients, Is the observed noise covariance matrix at time t-1, To weight attention vector The operation of converting to a diagonal matrix, As a result of the initial predictive value, Is the error vector of the true value and the initial predicted value, Is the kalman gain at time t, Is the state covariance matrix at time t, In order to observe the matrix, In order to constrain the penalty coefficients, As the inverse of the matrix in brackets, As a result of the predicted value of SOH, E is an identity matrix, which is a predicted value after physical constraint.
  8. 8. The method of claim 7, wherein in step 5, the lithium battery health rating comprises: A health-care level, which is a health-care level, ); The attenuation level of the optical fiber is provided with an attenuation stage, ; A pre-warning stage, wherein the pre-warning stage, );; When (when) When the first-level early warning is sent out, when When sending out secondary early warning, when And when the system does not send out early warning, the early warning signal is output in two modes of audible and visual warning and remote pushing through the BMS system.
  9. 9. The lithium battery health state prediction device based on the dynamic coupling mechanism is characterized by comprising a memory, a processor, a sensor module, a data transmission module and an early warning module: The sensor module comprises a temperature sensor, a voltage sensor, a current sensor, a humidity sensor and a cycle counter, and is arranged on the surface of a lithium battery cell and inside a battery pack to collect multidimensional operation data in real time; The data transmission module adopts CAN bus and wireless Bluetooth dual-mode transmission to transmit the acquired data to the processor; the memory adopts an SD card and is used for storing a computer program capable of running on a processor, lithium battery running data, model parameters and historical prediction results; The processor employing a core single chip microcomputer supporting floating point operations, when running the computer program, performing the steps of the method of any one of claims 1 to 9; The early warning module comprises a buzzer, an LED indicator light and a wireless communication module and is used for receiving early warning signals output by the processor and executing corresponding early warning operation.
  10. 10. An electronic device comprising a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method of any of claims 1 to 8.

Description

Lithium battery health state prediction method and device based on dynamic coupling mechanism Technical Field The invention relates to the technical field of lithium battery health management, in particular to a lithium battery health state prediction method and device based on a dynamic coupling mechanism. Background In the rapid development of new energy industry, lithium batteries play an irreplaceable key role as core energy storage components. The lithium battery is widely applied to the scenes of new energy automobiles, energy storage power stations, portable electronic equipment and the like, and the health state of the lithium battery directly determines the running stability, the cruising ability and the use safety of the terminal equipment. The traditional lithium battery Health State of Health (SOH) prediction method mostly depends on electrochemical mechanism modeling or a single data driving model, the former requires complex internal reaction dynamics analysis, the problems of large parameter identification difficulty and weak adaptability to complex working conditions exist, and the latter depends on a large amount of full life cycle aging data, so that the calculation cost is high, the generalization capability is insufficient, and prediction errors are easily enlarged due to data noise. These problems not only reduce the accuracy and instantaneity of SOH prediction, but also limit the engineering application of the SOH prediction in small sample and dynamic working condition scenes. In the early technology, the SOH monitoring of the lithium battery is mostly dependent on off-line detection or simple threshold judgment, and a technical short plate with early warning lag and incapability of early judging attenuation trend exists. In addition, potential safety hazards such as lithium battery thermal runaway and capacity dip are focused on single threshold monitoring such as temperature and voltage in the traditional early warning model, comprehensive consideration of multidimensional influences such as charge-discharge multiplying power, circulation times and environmental humidity is lacked, nonlinearity and time-varying characteristics in the aging process of the battery are difficult to comprehensively capture, practicability and reliability of an early warning system are limited, and full-period guarantee cannot be provided for safe operation of equipment. With the deep fusion of big data and artificial intelligence technology, the modern lithium battery health state prediction method has made breakthrough progress. The current research and practice gradually turns to a prediction scheme driven by multi-source data fusion, multi-type sensors are arranged through a Battery Management System (BMS), multi-dimensional operation data such as voltage, current, temperature, cycle times and the like are collected in real time, and feature mining and model construction are carried out by combining advanced algorithms such as deep learning, machine learning and the like. The technical innovations remarkably improve the accuracy and adaptability of SOH prediction of the lithium battery, and provide firmer technical support for safe operation and maintenance of new energy equipment. However, the existing method still has the defects of prominent short plates, insufficient generalization capability in a small sample scene, greatly reduced prediction precision caused by data scarcity in a new battery model or an application scene of the small population, limited dynamic adaptability, difficulty in adapting to complex working conditions such as charge-discharge multiplying power fluctuation, environment temperature mutation and the like in real time, low super-parameter optimization efficiency, long model training period and easiness in sinking into local optimum, and insufficient stability of full life cycle prediction precision caused by the lack of a synergistic mechanism of data enhancement and small sample learning and incremental learning. Disclosure of Invention Aiming at the problems of poor generalization, weak dynamic adaptability, insufficient prediction rationality, large deployment difficulty, lack of a cooperative mechanism and the like of the prior lithium battery SOH prediction technology, the invention provides a lithium battery health state prediction method and device based on a dynamic coupling mechanism, and the integrated design of 'small sample + data enhanced coupling, incremental learning online updating, dynamic weight adjustment and physical constraint correction' is adopted to realize the accuracy, rationality, instantaneity and embedded deployment suitability of the lithium battery full life cycle SOH prediction, thereby providing technical guarantee for the safe and stable operation of the lithium battery. The lithium battery health state prediction method based on the dynamic coupling mechanism comprises the following steps: Step 1, acquiring multi-dimensional operation dat