Intelligent networked new energy vehicle power battery soh prediction method based on bidirectional constraint koopman network

Abstract

The present application relates to a kind of intelligent network connection new energy vehicle power battery SOH prediction method based on bidirectional constraint Koopman network, belong to battery health management field.The method includes: S1: the signal data of long time series operation of lithium battery is collected;S2: extract lithium battery multiscale health characteristics and carry out correlation screening;S3: FES module is constructed;S4: based on deep Koopman self-encoder, bidirectional linear dynamics evolution model is constructed in latent space, and forward and backward evolution matrix and bidirectional consistency constraint mechanism are introduced;S5: after model training, network weight and evolution matrix are fixed, SOH is predicted by bidirectional constraint Koopman network model.S6: based on linear operator spectrum theory, eigenvalue decomposition is carried out to evolution matrix, and the physical stability of SOH prediction is evaluated.The present application solves the error accumulation and gradient explosion problem in long time series prediction, realizes the physical interpretability of model.

Description

Technical Field

[0001] This invention belongs to the field of battery health management, specifically the field of SOH estimation for lithium-ion batteries, and more specifically, it relates to a method for predicting the SOH of power batteries for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network. Background Technology

[0002] Lithium-ion batteries, due to their high power and energy density, long cycle life, and low self-discharge rate, are now widely used in electric vehicles, mobile phones, aircraft, and satellites. The market demands increasingly higher performance from battery management systems (BMS). Accurately estimating the State of Charge (SOC) and State of Health (SOH) of lithium-ion batteries is a core function of BMS. SOC reflects the current state of battery charge, expressed as the ratio of current remaining capacity to current maximum usable capacity, and plays a crucial role in the charging and discharging process. SOH reflects the battery's aging state over its entire lifespan, expressed as the ratio of maximum usable capacity to the battery's initial capacity.

[0003] Currently, many scholars have conducted extensive research on SOH estimation for lithium batteries, and the main technical routes can be summarized into the following three categories: (1) Physical measurement method: SOH assessment is achieved by directly obtaining key parameters such as battery internal resistance, impedance and capacity. It mainly includes internal resistance method, impedance method and ampere-hour integration method.

[0004] The internal resistance method mainly includes DC pulse signal testing, pulse power testing, and hybrid pulse power characteristic testing. Its basic principle is to apply a short-time current pulse to the battery and calculate the internal resistance through the voltage response. This method is intuitive, but it is easily affected by factors such as SOC, temperature, polarization effect, and electrode material corrosion, and cannot accurately separate the ohmic internal resistance and polarization resistance.

[0005] Impedance spectroscopy is used to obtain parameters such as ohmic internal resistance, charge transfer resistance, and double-layer capacitance over a wide frequency range through electrochemical impedance spectroscopy analysis. It offers advantages such as wide applicability, high measurement accuracy, and non-destructive testing. Its basic principle involves applying a small-amplitude sinusoidal AC perturbation signal to the battery within a fixed frequency range to obtain the impedance spectrum curve. However, in practical applications, temperature fluctuations and changes in charge / discharge states can affect the particle conduction process inside the battery, easily leading to larger errors in the impedance measurement results.

[0006] The ampere-hour integration method obtains the capacity parameter by calculating the integral of the charging and discharging current over time. Since the ampere-hour integration method is essentially an open-loop operation, small current measurement deviations accumulate during the integration process, which can easily lead to significant discrepancies between the estimated and actual capacity values.

[0007] (2) Mechanism model method: By analyzing the dynamic external characteristics or internal electrochemical reaction mechanism of the battery, a capacity loss model is established, which can be divided into two categories: electrochemical model and equivalent circuit model.

[0008] Electrochemical models quantitatively describe the internal mechanisms of Li+ concentration distribution and phase transitions in active materials through partial differential equations, accurately depicting the aging state of batteries. Currently, the more mature electrochemical models mainly include the single-particle (SP) model and the pseudo-two-dimensional (P2D) model. These models involve multi-scale coupled nonlinear partial differential equation systems; even with model order reduction or simplification techniques, their computational complexity remains high, making them difficult to meet the needs of real-time simulation and engineering applications.

[0009] The equivalent circuit model constructs a mathematical description of the battery's external characteristic response through series and parallel combinations of lumped-parameter components such as resistors and capacitors. The accuracy of parameter identification in the equivalent circuit model is highly dependent on prior assumptions about the model structure, a characteristic particularly pronounced during long-term battery aging. When the prior model structure fails to accurately reflect the actual battery degradation mechanism, it leads to distorted parameter representation, thus limiting the long-term estimation accuracy of the equivalent circuit model.

[0010] (3) Data-driven approach: This approach does not rely on prior knowledge or modeling of complex coupled aging mechanisms. It establishes a nonlinear mapping relationship between input and output through historical monitoring data. Machine learning methods include Gaussian process regression (GPR), support vector regression (SVR), and correlation vector machine (RVM). However, machine learning models lack physical interpretability. Summary of the Invention

[0011] In view of this, the purpose of this invention is to provide a State of Health (SOH) prediction method for power batteries of intelligent connected new energy vehicles based on a bidirectional constrained Koopman network, which solves the problems of long-term prediction divergence in existing traditional time series models, eigenvalue divergence caused by unidirectional prediction, and lack of physical interpretability in deep learning models. This method solves the problems of error accumulation and gradient explosion in long-term prediction through bidirectional consistency constraints, and endows deep learning with a certain degree of physical interpretability by utilizing spectral radius analysis.

[0012] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the state of harm (SOH) of power batteries in intelligent connected new energy vehicles based on bidirectional constrained Koopman networks, specifically including the following steps: S1: Collect signal data from the long-term operation of the lithium battery; S2: Extract multi-scale health characteristics of lithium batteries and perform correlation screening; S3: Construct the Feature Extraction and Decoupling (FES) module; S4: Constructing a bidirectional constrained Koopman network model: Based on a deep Koopman autoencoder, a fully parameterized bidirectional linear dynamic evolution model is constructed in the latent space, introducing forward and backward evolution matrices and a bidirectional consistency constraint mechanism. S5: After the model is trained and the network weights and evolution matrix are fixed, the SOH (health status) prediction is performed using a bidirectional constrained Koopman network model. S6: Eigenvalue decomposition of the evolution matrix is ​​performed based on linear operator spectral theory to evaluate the physical stability of the SOH prediction.

[0013] Furthermore, step S1 specifically includes: performing cyclic aging on the battery using a strategy of full charging and full discharging, acquiring data for each charge-discharge cycle, capacity aging data, and current during the charge-discharge phase. ,Voltage ,capacity and temperature .

[0014] The charging phase initially uses a constant current. When the battery voltage reaches its upper limit, it switches to a constant voltage charging mode until the charging current decreases to a set value. The discharging phase uses a constant current, with the battery's discharge cutoff voltage set. Each complete charge-discharge cycle is counted as one cycle. According to the standard, when the battery capacity decreases to 70% of its rated capacity, the battery is considered to have reached the end of its service life, and the cycle experiment is terminated.

[0015] Furthermore, step S2 specifically includes the following steps: S21: Based on the data in step S1, extract features from the collected battery signals. The aging of lithium-ion batteries is a complex dynamic process spanning multiple physical dimensions. To construct a high-quality observation space for the subsequent Deep Koopman network, a multi-scale feature extraction method integrating macroscopic external characteristics and microscopic electrochemical mechanisms is proposed. This method extracts multi-scale features from the collected battery signals, including directly measurable features, capacity increment features, and similarity features. Specifically, it extracts 11 key health features that comprehensively characterize the battery's SOH degradation trajectory. These health features include: isobaric rise time, constant current charging time, constant voltage charging time, isobaric drop time, peak value of the incremental capacity (IC) curve, voltage corresponding to the peak value of the IC curve, maximum temperature during discharge, area enclosed by current and time during charging, area enclosed by current and time during constant current charging, area enclosed by current and time during constant voltage charging, and similarity features.

[0016] The IC curve reflects the rate of change of charge and discharge capacity of a lithium-ion battery under different terminal voltages. Its basic idea is to use a fixed voltage interval... replace Calculate each Battery within voltage variation range Transformation ,replace And then use replace That is, the voltage range is Divide into, among which , may include: (1) In constant current charging mode, by differentiating the charging capacity from its terminal voltage, the slowly changing voltage plateau can be transformed into a distinct peak on the IC curve. The relationship between charging capacity and voltage can be described by a formula. By converting the charging current into the derivative of capacity with respect to voltage, the continuous QV equation can be obtained: (2) (3) in, I Let be the current of the constant current discharge. From equations (2) and (3), we can obtain: (4) Discretizing the continuous QV relationship yields: (5) (6) in, , These represent differential calculations for voltage and capacity, respectively. For the terminal voltage from Increase to Changes in capacity over time i For time steps.

[0017] Raw voltage, current, and temperature data often contain sensor noise, and direct use can cause severe fluctuations in differential curves (such as IC curves), masking the true electrochemical characteristics. Outlier removal and interpolation: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The criteria remove outliers from the sampling and use linear interpolation to fill in missing data, ensuring the continuity of the time series. The Kalman filter algorithm preprocesses the original IC curve to obtain the filtered IC curve, which effectively removes high-frequency noise while preserving the peak position and height information reflecting the phase transition process in the IC curve to the greatest extent.

[0018] The Kalman filter (KF) algorithm is used to preprocess the original IC curve, and its state and measurement equations are as follows: (7) in, x k For capacity increment data, y k Indicates to x k Noise pollution measurement, w k Indicates process noise. v k This represents the measurement noise. The KF calculation process can be expressed as: (8) in, Q and R Let be the covariance matrix of process noise and measurement noise. K For Kalman gain, P is x k The covariance matrix, For posterior state estimation, To estimate the covariance matrix a priori, To estimate the covariance matrix posteriorly, For Kalman gain, This is a priori state estimation.

[0019] Although voltage and current curves under different cycle periods exhibit similar changing patterns, subtle differences often contain rich information about battery degradation. Dynamic Time Warping (DTW) and Discrete Fréchet (DFD) algorithms are used to calculate the similarity between voltage and current curves under different cycles, and this similarity is used as a health characteristic to quantify battery degradation.

[0020] S22: Ten-dimensional features were obtained using a multi-scale feature extraction method. To verify the strong correlation between the multi-scale features and the battery SOH and to eliminate redundant interference, the Pearson correlation coefficient was used to calculate the relationship between the features and the capacity, and the three features with the largest absolute values ​​were retained.

[0021] For the i a set of features Battery capacity set The mean of these two time series is: (9) The covariance between the two time series is: (10) X and Y The Pearson correlation coefficient between them is: (11) in, For set X variance For set Y variance For set X With sets Y The mean of the product of corresponding elements. Pearson coefficient. The value range is [-1, 1]. The closer the absolute value is to 1, the stronger the correlation between the two sequences. The three features with the largest absolute values ​​are retained.

[0022] Furthermore, in step S3, the Feature Extraction and Decoupling (FES) module includes an independent feature extraction branch and a cross-feature interaction branch; The independent feature extraction branch treats each feature channel as an independent stream and applies a separate 1D convolutional layer to each feature, where the number of groups equals the feature dimension. The construction of cross-feature interaction branches involves pairing the original health features in pairs and combining them into three pairs, applying a 1D convolutional layer to the pairwise combinations, and applying a 3x3D convolutional layer to the three-pair combinations.

[0023] Further, in step S4, a bidirectional constrained Koopman network model is constructed, specifically including: constructing a deep Koopman autoencoder comprising a deep neural network (DNN) encoder and a symmetric decoder; the deep neural network (DNN) encoder, acting as an observation function, is responsible for mapping the nonlinear observation state to a high-dimensional, linear Koopman invariant subspace to obtain the latent state; the decoder is responsible for inversely mapping the latent state back to the original physical space, and retaining key physical information about battery aging by reconstructing error-supervised latent features. Within the potential space, a fully parameterized bidirectional linear dynamic evolution model is constructed, specifically including: defining a learnable forward evolution matrix A to simulate the forward extrapolation process of battery aging state over time; defining a learnable backward backtracking matrix B to utilize the invertibility of the dynamic system to backtrack historical states; and introducing bidirectional consistency constraints in network training to ensure the physical stability of the evolution operators, forcing the forward and backward matrices to be inverses of each other. A SOH regression predictor based on the Koopman latent space is constructed. After obtaining the dynamic law of the Koopman latent space state through an autoencoder and a bidirectional linear dynamic evolution model, a mapping mechanism from the latent linear space to the specific battery health state (SOH) is established.

[0024] Furthermore, in step S4, the SOH regression predictor employs a nonlinear regressor with the Swish activation function.

[0025] Furthermore, in step S4, the total loss function for training the bidirectional constrained Koopman network model is:

[0026] in, For autoencoder reconstruction loss, ensure that the latent state contains all the information needed to recover the original physical features; To predict losses forward, For backward prediction of loss, This is a loss due to bidirectional consistency. For the SOH regression loss, the forward evolution matrix A and the backward regression matrix B are forced to be inverses of each other, and the eigenvalues ​​are strictly constrained within the edge band of the unit circle in the complex plane. For the spectral norm constraint of the matrix; , , , , , These are the weights corresponding to the loss and constraints, respectively. Autoencoder reconstruction loss: ,in, The encoder inputs the raw input data at time... τ The value, The latent representation reconstructed by the decoder at time... τ The value, This represents the mean square error. Ensure that the latent states extracted by the encoder contain enough information to be recovered from the original state without loss by the decoder, and prevent information loss.

[0027] Forward prediction loss: ,in, For the future The true value, The future moments predicted by the model The value, The total number of samples in the batch. To predict the step size; Backward prediction loss: ; Bidirectional consistency loss: ,in, and These are the backward backtracking matrices. B upper part m Row and forward evolution matrix A left side m List, and They are A upper partm lines and B left side m List; Spectral norm constraints of a matrix: ,in, For matrix A spectral radius, It is a very small constant; To ensure the long-term stability of the system state, the spectral radius of the Koopman operator matrix A is constrained to be less than 1.

[0028] The online mapping process may compromise the consistency of the original data. To address the vanishing gradient problem encountered during long-term prediction training of neural networks, a consistency loss function is employed. To penalize the difference between forward and backward dynamics, this prompts the model to learn dynamic patterns of meaningful temporal relationships and dependencies in the data. It does not enforce the invertibility of the entire latent space, but locks in the dominant subspace that plays a decisive role in SOH prediction, enforces the strict invertibility and stability of the dynamics of this subspace, and places the most important dynamic features in the first m dimensions.

[0029] SOH regression loss: ,in, For a moment τ The true battery health status, For model prediction time τ Battery health status.

[0030] The error between the predicted SOH output by the regressor and the true SOH label is calculated to ensure that the Koopman latent space can be accurately mapped to the final SOH value.

[0031] The Adam optimizer is used to synchronously update the encoder network weights, decoder network weights, evolution matrix, and regression network weights based on the total loss function calculated above, until the model converges.

[0032] Furthermore, step S6 specifically includes: performing eigenvalue decomposition on the trained forward evolution matrix A to extract intrinsic modes and eigenvalues ​​that characterize the inherent laws of battery aging; establishing an asymptotic stability evaluation mechanism based on spectral radius: verifying the dissipative structure of the battery system when the eigenvalues ​​are distributed inside the unit circle of the complex plane; quantifying the dominant degradation trend of irreversible battery capacity decay by identifying the dominant eigenvalues ​​near the unit circle; eigenvalues ​​far from the unit circle correspond to the rapid and transient dynamics during battery charging and discharging, and contribute less to long-term SOH prediction.

[0033] The beneficial effects of this invention are as follows: This scheme constructs a bidirectional constrained Koopman network model, realizes the linearization mapping of nonlinear data through an autoencoder, and introduces bidirectional evolution and consistency constraints in the latent space, effectively avoiding the gradient explosion and error accumulation problems commonly found in long-term multi-step prediction. Simultaneously, by combining linear operator spectral theory to extract eigenmodes and analytical spectral radii, it endows the deep learning black-box model with interpretability at the physical and dynamic levels, providing technical support for the safety management of the entire battery lifecycle.

[0034] This scheme utilizes a multi-scale feature extraction method to extract multi-dimensional health features and employs the Pearson correlation coefficient to calculate the relationship between features and capacity, eliminating redundant interference. Furthermore, a Feature Extraction and Decoupling (FES) module is constructed, dynamically capturing the evolution trend of each health feature (HF) and the nonlinear coupling relationships between them through independent feature extraction branches and cross-feature interaction branches. Before entering the Koopman encoder, the original signal is preprocessed through the convolutional layers of the FES module, enabling the extraction of higher-order health features that are more robust to noise and more discriminative, thus improving the model's convergence speed and prediction accuracy.

[0035] This scheme evaluates the physical stability of SOH predictions by performing eigenvalue decomposition on the evolution matrix based on linear operator spectral theory.

[0036] This scheme solves the problems of error accumulation and gradient explosion in long-term time series prediction by using bidirectional consistency constraints, and gives deep learning a certain degree of physical interpretability by using spectral radius analysis. It can be applied to the SOH prediction of power batteries for intelligent connected new energy vehicles.

[0037] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the lithium battery SOH prediction method based on a bidirectional constrained Koopman network according to the present invention. Figure 2 Here is a flowchart of the multi-scale feature extraction method; Figure 3 Here is a structural diagram of the Feature Extraction and Decoupling Module (FES); Figure 4A framework for SOH estimation using bidirectional constrained Koopman networks; Figure 5 Comparison of SOH predictions for bidirectional constrained Koopman networks; Figure 6 Here is the structure diagram of the forward matrix operator A; Figure 7 This is a plot of the poles of the forward matrix operator A. Detailed Implementation

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0040] Please see Figures 1 to 7 This invention provides a method for predicting the state of harm (SOH) of power batteries in intelligent connected new energy vehicles based on a bidirectional constrained Koopman network. The specific implementation process is as follows: S1: Collect long-term operational signal data of the power battery in new energy vehicles. The battery is subjected to cyclic aging using a full charge and full discharge strategy, acquiring data for each charge-discharge cycle, capacity aging data, and current. ,Voltage ,capacity ,temperature .

[0041] S2: The aging of lithium-ion batteries is a complex dynamic process spanning multiple physical dimensions. To construct a high-quality observation space for the subsequent Deep Koopman network, a multi-scale feature extraction method integrating macroscopic external characteristics and microscopic electrochemical mechanisms is proposed. This method extracts multi-scale features from the acquired battery signals, including directly measurable features, capacity increment features, and similarity features. It extracts 10 key health factors that comprehensively characterize the battery's SOH degradation trajectory, and proposes the following physical feature vectors: isobaric rise time, constant current charging time, constant voltage charging time, isobaric drop time, IC curve peak value, voltage corresponding to the IC curve peak value, maximum temperature during discharge, area enclosed by current and time during charging, area enclosed by current and time during constant current charging, area enclosed by current and time during constant voltage charging, and similarity features. The raw voltage, current, and temperature data often contain sensor noise; direct use of this data can cause severe fluctuations in differential curves (such as the IC curve), masking the true electrochemical characteristics. Outlier removal and interpolation are employed. The criteria for eliminating outliers and using linear interpolation to fill in missing data ensures the continuity of the time series. A Kalman filter algorithm preprocesses the original IC curve to obtain a filtered IC curve, effectively removing high-frequency noise while preserving the peak positions and height information reflecting the phase transition process to the greatest extent. Although voltage and current curves under different cycle periods exhibit similar change patterns, subtle differences often contain rich information about battery degradation. Dynamic Time Warping (DTW) and Discrete Fréchet (DFD) algorithms are used to calculate the similarity between voltage and current curves under different cycles, and this is used as a health feature to quantify battery degradation. To verify the strong correlation between the above multi-scale features and battery SOH, and to eliminate redundant interference, the Pearson correlation coefficient is used to calculate the relationship between features and capacity, retaining the three health features (HF) with the largest absolute values.

[0042] S3: Construct the Feature Extraction and Decoupling (FES) module. To eliminate the dimensionality influence between different features, a min-max normalization method is used for scaling, mapping the input features to the [0,1] interval. Independent feature extraction branches are constructed to capture the unique temporal degradation trajectory of each HF. Cross-feature interaction branches are constructed, pairing the original health factors in pairs and combining them into three pairs, applying different 1D and 3x3D convolutional layers to each pair. Z ind 、 Z cro1 、 Z cro2 It is then passed as input to subsequent Koopman networks.

[0043] Step S3 specifically includes the following steps: S31: To eliminate the dimensionality influence between different features, a min-max normalization method is used for scaling, thus reducing the input features... x Mapped to the interval [0,1].

[0044]

[0045] S32: Construct independent feature extraction branches to capture the unique temporal degradation trajectory of each HF (Health Factor). Treat each feature channel as an independent stream and apply a separate one-dimensional convolutional layer to each feature, where the number of groups equals the feature dimension, ensuring that the learned patterns are specific to each individual HF, and outputting an independent feature sequence Z. ind .

[0046] A cross-feature interaction branch is constructed, pairing the original health factors together and applying a different 1D convolutional layer to each pair. This structure forces the model to learn the joint patterns that emerge when two HFs co-evolve, outputting a cross-feature sequence Z. cro1 。 The three original health factors are combined, and a 3x3D convolutional layer is applied to output the cross-feature sequence Z. cro2 The FES module outputs three distinct feature sequences, encapsulating the decoupled independent dynamics and cross-coupling effects, respectively. These three representations are then passed to the input of the subsequent bidirectional constrained Koopman network model. Preprocessing the original signal through convolutional layers before it enters the Koopman encoder extracts higher-order health features that are more robust to noise and more discriminative, thus improving the model's convergence speed and prediction accuracy.

[0047]

[0048]

[0049]

[0050] S4: Constructing a Bidirectional Constrained Koopman Network Model: Based on a deep Koopman autoencoder, a fully parameterized bidirectional linear dynamic evolution model is constructed within the latent space, introducing forward and backward evolution matrices and a bidirectional consistency constraint mechanism. A deep Koopman autoencoder is constructed, comprising a deep neural network (DNN) encoder and a symmetric decoder. The encoder is responsible for losslessly mapping the nonlinear observation state to a high-dimensional, linear Koopman invariant subspace; the decoder is responsible for inversely mapping the latent state back to the original physical space. Within the latent space, a fully parameterized bidirectional linear dynamic evolution model is constructed: a learnable forward evolution matrix A is defined to simulate the forward evolution of the battery aging state over time, and a learnable backward backtracking matrix B is defined to utilize the invertibility of the dynamic system for historical state backtracking. To ensure the physical stability of the evolution operator, a bidirectional consistency constraint is introduced, forcing the forward and backward matrices to be inverses of each other.

[0051] Step S4 specifically includes: To address the highly nonlinear and long-term time-dependent characteristics of the SOH degradation process in lithium batteries, a deep Koopman autoencoder is constructed as the core architecture, which uses a deep neural network to map the nonlinear observation space to the linear dynamic space.

[0052] A deep neural network (DNN)-based encoder is constructed as the observation function. This module is responsible for mapping the nonlinear observation state to a high-dimensional, linear Koopman invariant subspace to obtain the latent state. Within this space, the complex electrochemical degradation trajectory of the battery is straightened out into an easily predictable linear manifold. Simultaneously, a symmetric decoder is configured to inversely map the latent state back to the original physical space, preserving key physical information about battery aging by reconstructing error-supervised latent features.

[0053] A fully parameterized bidirectional linear evolution is constructed, establishing a fully parameterized linear evolution layer within the latent space. A learnable weight matrix is ​​defined. Simulate the forward extrapolation process of battery aging over time. ,in, For a moment t The latent state vector, This refers to the nth power operation on matrix A. A learnable weight matrix is ​​defined. Using the reversibility of dynamic systems to backtrack to historical states ,in, This involves n-fold power operations on matrix B. This design transforms complex nonlinear time series forecasting into efficient matrix exponentiation. This significantly reduces the cumulative error of multi-step prediction.

[0054] A bidirectional consistency and spectral constraint mechanism is introduced to ensure the physical stability of the evolution operator. This mechanism mandates that the forward and backward matrices be inverses of each other during network training. This not only regulates the distribution of the eigenvalue spectrum, preventing gradient explosion or vanishing during long-term iterations, but also ensures the mathematical invertibility and closed-loop consistency of the dynamic model, thereby achieving robust prediction of the entire battery lifecycle.

[0055] A SOH regression predictor based on the Koopman latent space is constructed, and the dynamic states of the Koopman latent space are obtained through an autoencoder and a bidirectional evolution module. Next, a mapping mechanism needs to be established from the latent linear space to the specific state of battery health (SOH). Considering that the Koopman encoder has already mapped the highly nonlinear raw data to a structurally regular and smooth linear manifold, the regression network should not be too complex to avoid introducing unnecessary overfitting risks. A nonlinear regressor with the Swish activation function is used to model the mapping function between the high-level features learned from the Koopman feature function space and the corresponding SOH.

[0056] To achieve end-to-end model training, a multi-task joint loss function was designed, which includes reconstruction, prediction, linear and bidirectional constraints.

[0057] The total loss function is defined as follows:

[0058] Among them, the autoencoder reconstruction loss is: ,in, The encoder inputs the raw input data at time... τ The value, The latent representation reconstructed by the decoder at time... τ The value, Mean square error; Ensure that the latent states extracted by the encoder contain enough information to be recovered from the original state without loss by the decoder, and prevent information loss.

[0059] Forward prediction loss: ,in, For the future The true value, The future moments predicted by the model The value, The total number of samples in the batch. To predict the step size; Backward prediction loss: ; Bidirectional consistency loss: ,in, and These are the backward backtracking matrices. B upper part m Row and forward evolution matrix A left side m List, and They are A upper part m lines and B left side m List; Spectral norm constraints of a matrix: ,in, For matrix A spectral radius, It is a very small constant; To ensure the long-term stability of the system state, the spectral radius of the Koopman operator matrix A is constrained to be less than 1.

[0060] The online mapping process may compromise the consistency of the original data. To address the vanishing gradient problem encountered during long-term prediction training of neural networks, a consistency loss function is employed. To penalize the difference between forward and backward dynamics, this prompts the model to learn dynamic patterns of meaningful temporal relationships and dependencies in the data. It does not enforce the invertibility of the entire latent space, but locks in the dominant subspace that plays a decisive role in SOH prediction, enforces the strict invertibility and stability of the dynamics of this subspace, and places the most important dynamic features in the first m dimensions.

[0061] SOH regression loss: ,in, For a moment τ The true battery health status, For model prediction time τ Battery health status.

[0062] The error between the predicted SOH output by the regressor and the true SOH label is calculated to ensure that the Koopman latent space can be accurately mapped to the final SOH value.

[0063] The Adam optimizer is used to synchronously update the encoder network weights, decoder network weights, evolution matrix, and regression network weights based on the total loss function calculated above, until the model converges.

[0064] S5: After training, the weights and evolution matrix of the bidirectional constrained Koopman network model are fixed, and SOH (health status) prediction is performed through the bidirectional constrained Koopman network model.

[0065] Step S5 specifically includes the following steps: S51: Collect the battery external characteristic signal from the most recent battery cycle and extract multi-scale health features using the same preprocessing method as the training set. Input the extracted multi-scale health features into the trained encoder. At this point, the highly nonlinear raw aging data is mapped almost losslessly to a well-structured, smooth linear manifold.

[0066] S52: Predict the SOH for the next n cycles, then in the latent space, use the trained forward evolution matrix. A Perform efficient matrix exponentiation on the current potential state (i.e., multiply the state by the power of ... A n This allows us to deduce the potential dynamic state at future moments.

[0067] S53: After obtaining the Koopman latent space state of the dynamic law through the autoencoder and bidirectional evolution module, the state is input into the SOH regression network, and finally the specific battery health state (SOH) prediction value is output.

[0068] S6: Eigenvalue decomposition of the forward evolution matrix A based on linear operator spectral theory is performed to evaluate the physical stability of the SOH prediction. After model training, the forward evolution matrix A is extracted and eigenvalue decomposition is performed to extract the eigenvalue diagonal matrix and the corresponding eigenvectors (i.e., Koopman modes). Spectral radius analysis is introduced to verify whether the model captures the physical essence of the battery dissipative system: a complex plane unit circle is drawn. When the eigenvalues ​​are distributed inside the unit circle, the dissipative structure of the battery system is verified. The Koopman modes corresponding to the dominant eigenvalues ​​near the unit circle represent the dominant degradation trend of slow and irreversible capacity decay. Eigenvalues ​​far from the unit circle correspond to the battery's fast and transient dynamics (such as polarization effects or noise), contributing less to long-term SOH prediction. Prediction failure is warned by checking whether the spectral radius deviates from the unit circle.

[0069] Step S6 specifically includes: after the model training is completed, extracting the forward evolution matrix. According to Koopman theory, the dynamic behavior of a linear system is entirely determined by its eigenpairs. For the matrix... A Perform eigenvalue decomposition ,in It is an eigenvalue diagonal matrix. For the corresponding feature vector (i.e., the Koopman mode), the latent state at any time step It can be represented as a linear combination of Koopman modes. This formula shows that the evolutionary trajectory of SOH is essentially caused by... K It is composed of the superposition of several independent modes; among which For the first jProjection coefficients of each Koopman mode.

[0070] Eigenvalues The modulus length determines the decay or growth rate of the mode over time, and the argument angle... This determines the oscillation frequency of the mode. Based on the asymptotic stability assessment using spectral radius, lithium batteries, as dissipative systems, do not exhibit infinite divergence of physical quantities. To mathematically verify whether the model captures this physical essence, spectral radius analysis is introduced. The unit circle in the complex plane is drawn. The eigenvalues ​​λ are distributed inside the unit circle. This verified the dissipation structure of the battery system. Near the unit circle... There exists a dominant eigenvalue. The Koopman mode corresponding to this eigenvalue represents the dominant degradation trend of slow, irreversible capacity decline in the battery. (Far from the unit circle) The eigenvalues ​​correspond to rapid, transient dynamics during battery charging and discharging, such as polarization effects or measurement noise. These modes contribute little to long-term SOH prediction. By plotting the complex plane eigenvalue distribution (Unit Circle Plot), the "black box" weights of deep learning are transformed into a visible "spectrum." If the predicted SOH curve shows abnormal fluctuations, prediction failure can be detected by checking whether the spectral radius deviates from the unit circle.

[0071] Figure 5 The battery SOH prediction results are presented. (By...) Figure 5 As can be seen, the model's predicted values ​​(red dashed line) are consistent with the actual SOH (blue solid line) in an overall trend, accurately tracking the battery aging and degradation process, thus verifying the effectiveness of the proposed method.

[0072] Figure 6 This is a weighted heatmap of the Koopman forward evolution matrix A. Figure 6 As can be seen, matrix A exhibits a clear diagonal dominance, with the main diagonal elements significantly larger than the off-diagonal elements, indicating that the dynamic behavior within the latent space is dominated by independent evolution of each dimension, with weak coupling between dimensions. This structure verifies that the proposed method can effectively learn low-dimensional linearized dynamic representations, consistent with the expectations of Koopman theory.

[0073] Figure 7 This is a pole distribution plot for the forward matrix operator A. Figure 7 As can be seen, all eigenvalues ​​lie within the unit circle, satisfying the stability constraints; and the poles are mainly distributed near the positive real axis, with a small imaginary part, indicating that the system is dominated by slow, monotonically decaying, with weak oscillatory characteristics. This distribution characteristic is consistent with the physical law of gradual degradation during lithium battery aging, verifying the rationality of the proposed Koopman operator in characterizing battery dynamics.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the state of harm (SOH) of power batteries in intelligent connected new energy vehicles based on a bidirectional constrained Koopman network, characterized in that, The method specifically includes the following steps: S1: Collect signal data from the long-term operation of the lithium battery; S2: Extract multi-scale health characteristics of lithium batteries and perform correlation screening; S3: Construct a feature extraction and decoupling module; S4: Constructing a bidirectional constrained Koopman network model: Based on a deep Koopman autoencoder, a fully parameterized bidirectional linear dynamic evolution model is constructed in the latent space, introducing forward and backward evolution matrices and a bidirectional consistency constraint mechanism. S5: After the model is trained and the network weights and evolution matrix are fixed, SOH prediction is performed using a bidirectional constrained Koopman network model.

2. The method for predicting the State of Harm (SOH) of power batteries for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to claim 1, characterized in that, Step S1 specifically includes: using a strategy of full charging and full discharging to perform cyclic aging of the battery, obtaining data for each charge-discharge cycle, capacity aging data, and current, voltage, capacity, and temperature during the charge-discharge phase.

3. The lithium battery SOH prediction method based on a bidirectional constrained Koopman network according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21: For the collected battery signals, multi-scale features, including direct measurement features, capacity increment features, and similarity features, are extracted using a multi-scale feature extraction method. That is, health features that can characterize the battery's SOH degradation trajectory are extracted. The health features include: isobaric rise time, constant current charging time, constant voltage charging time, isobaric drop time, peak value of the capacity increment curve, voltage corresponding to the peak value of the capacity increment curve, maximum temperature during discharge, area enclosed by current and time during charging, area enclosed by current and time during constant current charging, area enclosed by current and time during constant voltage charging, and similarity features. S22: The relationship between features and capacity is calculated using the Pearson correlation coefficient, and the three features with the largest absolute values ​​are retained.

4. The method for predicting the State of Harm (SOH) of power batteries for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to claim 1, characterized in that, In step S3, the feature extraction and decoupling module includes an independent feature extraction branch and a cross-feature interaction branch; The independent feature extraction branch treats each feature channel as an independent stream and applies a separate 1D convolutional layer to each feature, where the number of groups equals the feature dimension. The construction of cross-feature interaction branches involves pairing the original health features in pairs and combining them into three pairs. A 1D convolutional layer is applied to the pairwise combinations, and a 3x3D convolutional layer is applied to the three-pair combinations.

5. The method for predicting the State of Harm (SOH) of power batteries for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to claim 1, characterized in that, In step S4, a bidirectional constrained Koopman network model is constructed, specifically including: constructing a deep Koopman autoencoder comprising a deep neural network encoder and a symmetric decoder; the deep neural network encoder, as the observation function, is responsible for mapping the nonlinear observation state to a high-dimensional linear Koopman invariant subspace to obtain the latent state; the decoder is responsible for inversely mapping the latent state back to the original physical space. Within the potential space, a bidirectional linear dynamic evolution model is constructed, specifically including: defining a learnable forward evolution matrix to simulate the forward extrapolation process of battery aging state over time; defining a learnable backward backtracking matrix to utilize the invertibility of the dynamic system to backtrack historical states; and introducing bidirectional consistency constraints in network training to force the forward and backward matrices to be inverses of each other. A SOH regression predictor based on the Koopman latent space is constructed. After obtaining the Koopman latent space state of dynamic laws through an autoencoder and a bidirectional linear dynamic evolution model, a mapping mechanism from the latent linear space to the SOH of a specific battery is established.

6. The method for predicting the State of Harm (SOH) of a power battery for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to claim 5, characterized in that, In step S4, the SOH regression predictor uses a nonlinear regressor with the Swish activation function.

7. The method for predicting the State of Harm (SOH) of a power battery for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to claim 5, characterized in that, In step S4, the total loss function for training the bidirectional constrained Koopman network model is: in, For autoencoder reconstruction loss, ensure that the latent state contains all the information needed to recover the original physical features; To predict losses forward, For backward prediction of loss, This is a loss due to bidirectional consistency. For the SOH regression loss, the forward evolution matrix A and the backward regression matrix B are forced to be inverses of each other, and the eigenvalues ​​are strictly constrained within the edge band of the unit circle in the complex plane. For the spectral norm constraint of the matrix; , , , , , These are the weights corresponding to the loss and constraints, respectively. Autoencoder reconstruction loss: ,in, The encoder inputs the raw input data at time... τ The value, The latent representation reconstructed by the decoder at time... τ The value; Mean squared error measures the difference between the reconstructed value and the true value. Forward prediction loss: ,in, For the future The true value, The future moments predicted by the model The value, The total number of samples in the batch. To predict the step size; Backward prediction loss: ; Bidirectional consistency loss: ,in, and These are the backward backtracking matrices. B upper part m Row and forward evolution matrix A left side m List, and They are A upper part m lines and B left side m List; Spectral norm constraints of a matrix: ,in, For matrix A spectral radius, It is a constant; SOH regression loss: ,in, For a moment τ The true battery health status, For model prediction time τ Battery health status.

8. The method for predicting the state of harm (SOH) of a power battery for intelligent connected new energy vehicles based on a bidirectional constrained Koopman network according to any one of claims 1 to 7, characterized in that, The method also includes step S6: performing eigenvalue decomposition on the evolution matrix based on linear operator spectral theory to evaluate the physical stability of the SOH prediction; Specifically, this includes: performing eigenvalue decomposition on the trained forward evolution matrix A to extract intrinsic modes and eigenvalues ​​that characterize the inherent laws of battery aging; Establish an asymptotic stability assessment mechanism based on spectral radius: when the eigenvalues ​​are distributed inside the unit circle of the complex plane, verify the dissipative structure of the battery system; by identifying the dominant eigenvalues ​​near the unit circle, quantify the dominant degradation trend of irreversible battery capacity decay; eigenvalues ​​far from the unit circle correspond to the rapid and transient dynamics during battery charging and discharging, and contribute little to the long-term SOH prediction.

Images