Battery life accurate prediction and evaluation method and system based on deep learning

Abstract

The present invention provides a method and system for accurately predicting and evaluating battery life based on deep learning, which relates to the technical field of battery life prediction. The method comprises acquiring battery operation data, dividing processes based on change characteristics, extracting time series features using a state observation model and a variational inference network, and obtaining dynamic features through phase space reconstruction; establishing a deep feature enhancement network to obtain comprehensive features; extracting features through capacity and internal resistance prediction branches, establishing a segmented recursive coupling channel for fusion, and outputting prediction results; and determining the remaining life in combination with a scrap threshold. The present invention improves the accuracy and reliability of battery life prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of battery life prediction, and in particular to a method and system for accurately predicting and evaluating battery life based on deep learning. Background Art

[0002] With the rapid development of the new energy vehicle industry, the performance and lifespan of power batteries, as core components, directly impact vehicle performance and safety. Accurately predicting the remaining useful life of batteries is crucial for ensuring safe operation of new energy vehicles and optimizing battery management strategies. Currently, battery life prediction methods primarily include those based on physical models, data-driven approaches, and deep learning-based methods. Deep learning methods, which can adaptively extract battery degradation characteristics, show promising application prospects in the field of battery life prediction.

[0003] However, existing deep learning-based battery life prediction methods still have some shortcomings: the coupling relationship between battery capacity decay and internal resistance growth is not fully considered, resulting in insufficient prediction accuracy; the feature extraction method of battery historical operation data is relatively simple, and it is difficult to effectively capture the multi-scale dynamic characteristics of the battery degradation process; there is a lack of dependence analysis on the battery working state, and it is impossible to accurately describe the differential characteristics of battery performance degradation under different working conditions; the existing prediction model has a single structure and limited feature fusion capabilities, which makes it difficult to adapt to the complex and changeable battery degradation process.

[0004] In summary, a deep learning-based accurate prediction and evaluation method for battery life is needed. The dynamic characteristics of the battery are extracted through phase space reconstruction and symplectic geometry optimization, a deep feature enhancement network is established to achieve feature optimization, and a state-dependent piecewise recursive coupling structure is used to achieve a bidirectional fusion of capacity attenuation characteristics and impedance evolution characteristics, thereby improving the accuracy and reliability of battery life prediction. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for accurately predicting and evaluating battery life based on deep learning, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a deep learning-based accurate prediction and evaluation method for battery life, including:

[0008] Obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle count as battery historical operation data;

[0009] Based on the change characteristics of the battery's historical operating data, the process is divided, and the state observation model and variational inference network are used to extract multi-process features to form a time series feature vector. The battery dynamic characteristics are obtained through phase space reconstruction and symplectic geometry optimization.

[0010] Establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize and model them through a deep learning framework, and obtain comprehensive battery features;

[0011] The comprehensive battery features are input into the shared coding layer to obtain timing calibration features. The capacity attenuation features and impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch. A state-dependent piecewise recursive coupling channel is established to perform feature fusion and output the capacity prediction results and internal resistance prediction results.

[0012] The remaining service life of the battery is determined based on the capacity prediction result and the internal resistance prediction result in combination with a battery scrapping threshold.

[0013] In an optional embodiment,

[0014] Based on the change characteristics of the battery's historical operating data, the process is divided, and the state observation model and variational inference network are used to extract multi-process features to form a time series feature vector. The battery dynamic characteristics are obtained through phase space reconstruction and symplectic geometry optimization, including:

[0015] Based on the change characteristics of the battery historical operation data, the battery historical operation data is divided into charging process data, discharging process data and static process data;

[0016] Establishing state observation models for the charging process data, the discharging process data, and the static process data, respectively, calculating state transition probabilities and observation probability densities through Gaussian distribution, and constructing probability distributions of the state observation models;

[0017] constructing a variational inference network based on the probability distribution, the variational inference network extracting an electrochemical state vector and thermodynamic characteristics from the charging process data, extracting energy efficiency characteristics and dynamic response characteristics from the discharging process data, and extracting polarization characteristics and self-discharge characteristics from the static process data, and combining the extracted characteristics in a time series to form a time series feature vector;

[0018] Based on the evolution trajectory of the time series eigenvector in the phase space, local potential energy function and pairwise potential energy function are constructed. Combined with symplectic geometry optimization, the dynamic characteristics of the battery are obtained by reconstructing the fractal characteristics of the attractor.

[0019] In an optional embodiment,

[0020] Based on the evolution trajectory of the time series eigenvector in the phase space, the local potential energy function and the pairwise potential energy function are constructed. Combined with symplectic geometry optimization, the battery dynamic characteristics are obtained by reconstructing the fractal characteristics of the attractor, including:

[0021] The Hamiltonian dynamics equation is used to describe the evolution trajectory of the time series eigenvector, which is then projected into the phase space to form a phase orbit. The local potential energy function is calculated based on the Lyapunov exponent of the phase orbit, and the pairwise potential energy function is calculated based on the degree of synchronization between the phase orbits.

[0022] Projecting the evolution trajectory of the time series feature vector in the phase space onto the Poincare section for reconstruction, and extracting the electrochemical feature set and the thermodynamic feature set;

[0023] Calculating the invariant measures of the electrochemical characteristic set and the thermodynamic characteristic set on the Poincare section to obtain an electrochemical energy term and a thermodynamic energy term, respectively;

[0024] Based on the long-range correlation characteristics of the electrochemical energy term and the thermodynamic energy term, a fractional differential operator is constructed, and a coupling energy function is calculated based on the order of the fractional differential operator;

[0025] Substituting the electrochemical energy term, the thermodynamic energy term, and the coupled energy function into a symplectic geometry optimizer, keeping the Hamiltonian structure unchanged and minimizing global energy, to obtain an optimized energy function;

[0026] The attractor in the phase space is reconstructed based on the optimized energy function, the fractal dimension and correlation dimension of the attractor are calculated, and the dynamic characteristics of the battery are obtained by combining the dimensional feature reconstruction.

[0027] In an optional embodiment,

[0028] A deep feature enhancement network is established to perform manifold projection and nonlinear transformation on the battery dynamic features. Optimized modeling is performed through a deep learning framework to obtain comprehensive battery features including:

[0029] Inputting the battery dynamic characteristics into a manifold embedding layer, calculating a reconstruction weight matrix through neighboring points of the manifold embedding layer, solving the characteristic equation using the reconstructed weight matrix to obtain a first low-dimensional embedding representation, constructing a geodesic distance matrix based on the first low-dimensional embedding representation, performing spectral decomposition on the geodesic distance matrix to obtain an eigenvector matrix and an eigenvalue matrix, and multiplying the eigenvector matrix by the square root of the eigenvalue matrix to generate a manifold embedding representation;

[0030] Inputting the manifold embedding representation into a dynamic kernel function enhancement module, constructing multiple basic kernel functions using a distance metric, calculating a combination weight of the multiple basic kernel functions using a deep network, performing a weighted combination of the combination weight and the multiple basic kernel functions to generate a combined kernel function, and performing a nonlinear transformation on the manifold embedding representation using the combined kernel function to generate enhanced features;

[0031] The enhanced features are input into the probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, construct a node distribution, and iteratively optimize through a message passing algorithm to obtain the comprehensive battery features.

[0032] In an optional embodiment,

[0033] The enhanced features are input into the probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, construct a node distribution, and iteratively optimize through a message passing algorithm to obtain the comprehensive battery features including:

[0034] The probabilistic graph optimization module constructs a conditional probability field for the enhanced feature based on the graph structure, wherein the nodes in the conditional probability field correspond to the dimensional components of the enhanced feature, and the edge connection relationship between the nodes forms an edge set;

[0035] The conditional probability field calculates the correlation between the components of each dimension of the enhanced feature through a deep neural network, the deep neural network maps the nodes to the hidden layer feature space to obtain the node hidden layer representation, and calculates the correlation strength between the nodes based on the node hidden layer representation;

[0036] Constructing an energy function of a conditional probability field according to the association strength, including a node potential energy term and an edge potential energy term, wherein the edge potential energy term is modulated by the association strength, and defining a conditional probability distribution based on the energy function;

[0037] The conditional probability field iteratively transmits node information on the edge set through a message passing algorithm under the conditional probability distribution, wherein the message passing algorithm calculates the information transmitted between nodes based on the node potential energy term, the edge potential energy term and the message update of the adjacent nodes;

[0038] The probability graph optimization module jointly optimizes the log-likelihood of the conditional probability distribution and the loss function of the deep neural network, alternately updates the parameters of the deep neural network and the parameters of the conditional probability field until convergence, and outputs the comprehensive battery characteristics.

[0039] In an optional embodiment,

[0040] The comprehensive battery features are input into the shared coding layer to obtain the timing calibration features. The capacity attenuation features and impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch. A state-dependent piecewise recursive coupling channel is established to perform feature fusion. The output capacity prediction results and internal resistance prediction results include:

[0041] The battery comprehensive features are input into the shared coding layer, and the shared coding layer performs layer-by-layer dimensionality reduction coding through a multi-layer convolutional network to obtain coding features. The shared coding layer extracts the time series correlation from the coding features to obtain the time series calibration features;

[0042] The timing calibration features are input into the capacity prediction branch and the internal resistance prediction branch respectively. The capacity prediction branch extracts the capacity attenuation features from the timing calibration features through a multi-layer perceptron, and the internal resistance prediction branch extracts the impedance evolution features from the timing calibration features through a recursive neural network.

[0043] A state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature, and bidirectional dynamic transmission and fusion are achieved through state matrix transformation and multi-stage response equations to obtain capacity fusion features and internal resistance fusion features;

[0044] A capacity prediction result is output based on the capacity fusion feature, and an internal resistance prediction result is output based on the internal resistance fusion feature.

[0045] In an optional embodiment,

[0046] A state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature. Through state matrix transformation and multi-stage response equations, bidirectional dynamic transmission and fusion are achieved. The obtained capacity fusion feature and internal resistance fusion feature include:

[0047] A dynamic collaborative optimization channel is established between the capacity decay characteristics and the impedance evolution characteristics. A state dependency matrix is constructed based on the battery operating state. The state dependency matrix includes a temperature gradient coefficient, a charge and discharge rate coefficient, and a cycle number coefficient. A piecewise matrix transformation is performed on the capacity decay characteristics and the impedance evolution characteristics according to the state dependency matrix. An independent feature mapping function is established in each operating state interval. The piecewise reconstruction of the features is achieved through function combination to obtain state-related features.

[0048] A capacity-internal resistance coupling model is constructed based on the state-related characteristics. The capacity-internal resistance coupling model is recursively updated to time-sequentially pair characteristic change points in the capacity degradation process with internal resistance mutation points, and establish a multi-stage capacity-internal resistance response equation to form a quantitative correlation between capacity decay and internal resistance growth with state memory;

[0049] According to the capacity-internal resistance coupling model, bidirectional feature compensation and enhancement are performed on the capacity attenuation feature and the impedance evolution feature to obtain a capacity fusion feature and an internal resistance fusion feature combined with state dependence.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provides a deep learning-based accurate battery life prediction and evaluation system, including:

[0052] The first unit is used to obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle number as battery historical operation data;

[0053] The second unit is used to divide the process based on the change characteristics of the battery's historical operating data. It uses the state observation model and variational inference network to extract multi-process features to form a time series feature vector, and obtains the battery dynamic characteristics through phase space reconstruction and symplectic geometry optimization;

[0054] The third unit is used to establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize the modeling through the deep learning framework, and obtain the comprehensive characteristics of the battery;

[0055] The fourth unit is used to input the comprehensive battery features into the shared coding layer to obtain timing calibration features, extract the capacity decay features and impedance evolution features through the capacity prediction branch and the internal resistance prediction branch, establish a state-dependent piecewise recursive coupling channel, perform feature fusion, and output the capacity prediction results and internal resistance prediction results;

[0056] The fifth unit is used to determine the remaining service life of the battery according to the capacity prediction result and the internal resistance prediction result in combination with the battery scrapping threshold.

[0057] According to a third aspect of the embodiments of the present invention,

[0058] An electronic device is provided, comprising:

[0059] processor;

[0060] a memory for storing processor-executable instructions;

[0061] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0062] According to a fourth aspect of the embodiments of the present invention,

[0063] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0064] In an embodiment of the present invention, a comprehensive characterization of the battery operating state is achieved through multi-process feature extraction and dynamic feature optimization, and the prediction model's ability to capture the nonlinear degradation behavior of the battery is improved, thereby making the battery life prediction more accurate and reliable; the use of deep feature enhancement network and manifold projection technology effectively solves the limitations of traditional methods in processing high-dimensional heterogeneous data, and can adaptively learn the degradation mode of the battery under different working conditions, thereby enhancing the generalization ability and robustness of the model, and making it suitable for complex and changeable practical application scenarios; a state-dependent segmented recursive coupling channel is introduced to achieve a deep fusion of capacity attenuation characteristics and impedance evolution characteristics, overcoming the limitations of single parameter prediction, and by comprehensively considering the mutual influence of multiple degradation mechanisms within the battery, significantly improving the accuracy and reliability of battery life prediction, providing strong support for the optimization of battery management systems and battery life cycle management. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Schematic diagram of the process of a method for accurately predicting and evaluating battery life based on deep learning according to an embodiment of the present invention;

[0066] Figure 2 Schematic diagram of comprehensive evaluation of battery dynamic characteristics and life prediction performance;

[0067] Figure 3 This is a group of contour maps of the correlation between battery feature dimensions;

[0068] Figure 4 This is a state dependency matrix visualization and accuracy comparison chart. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0070] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 1 is a flow chart of a method for accurately predicting and evaluating battery life based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] Obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle count as battery historical operation data;

[0073] Based on the change characteristics of the battery's historical operating data, the process is divided, and the state observation model and variational inference network are used to extract multi-process features to form a time series feature vector. The battery dynamic characteristics are obtained through phase space reconstruction and symplectic geometry optimization.

[0074] Establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize and model them through a deep learning framework, and obtain comprehensive battery features;

[0075] The comprehensive battery features are input into the shared coding layer to obtain timing calibration features. The capacity attenuation features and impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch. A state-dependent piecewise recursive coupling channel is established to perform feature fusion and output the capacity prediction results and internal resistance prediction results.

[0076] The remaining service life of the battery is determined based on the capacity prediction result and the internal resistance prediction result in combination with a battery scrapping threshold.

[0077] In one embodiment, historical operating data such as the battery's charge and discharge voltage, charge and discharge current, temperature, internal resistance, and cycle count are collected during actual operation. Data preprocessing is performed to eliminate outliers and noise, and the data is normalized to ensure that all data types have the same dimension.

[0078] Based on the changing trend of the battery's historical operating data, the sliding time window method is used to divide the battery operation process and identify different working stages such as the charging process, discharging process, and static process. A state observation model is established for each working stage. The model contains a state transfer equation and an observation equation to describe the evolution law of the battery state. A variational inference network is constructed, which consists of an encoder and a decoder. The parameters of the state observation model are estimated by maximizing the lower bound of evidence and extracting the feature vectors of each process. The extracted feature vectors are combined in time sequence to form a time series feature vector. Phase space reconstruction is performed based on the time series feature vector to determine the optimal embedding dimension and time delay, and reconstruct the phase space trajectory of the battery state. The symplectic geometry optimization algorithm is applied in the reconstructed phase space to maintain the energy conservation characteristics of the system and obtain the dynamic characteristics of the battery.

[0079] A deep feature enhancement network is constructed, consisting of a feature mapping layer, a feature enhancement layer, and a feature fusion layer. In the feature mapping layer, battery dynamic features are projected into a high-dimensional manifold space, preserving the feature topology. In the feature enhancement layer, a multi-layer nonlinear transformation is used to extract deep representations of features. In the feature fusion layer, an attention mechanism is used to adaptively weight features at different levels. A backpropagation algorithm is used to optimize network parameters to obtain comprehensive battery features.

[0080] A shared encoding layer is designed, using a multi-layer convolutional neural network to encode comprehensive battery features and extract timing correlations to obtain timing calibration features. In the capacity prediction branch, a multi-layer perceptron network is used to process timing calibration features and extract capacity decay features. In the internal resistance prediction branch, a long short-term memory network is used to process timing calibration features and extract impedance evolution features. A state-dependence matrix is constructed, including temperature gradient coefficients, charge and discharge rate coefficients, and cycle number coefficients. Based on the state-dependence matrix, a piecewise matrix transformation is performed on the capacity decay and impedance evolution features. A recursive coupling channel is established, achieving bidirectional feature transfer through multi-stage response equations. Feature fusion is performed to obtain capacity fusion features and internal resistance fusion features. Based on the fused features, the capacity prediction value and internal resistance prediction value are output respectively.

[0081] Set scrap thresholds for battery capacity and internal resistance. Perform a lifespan assessment based on the capacity and internal resistance predictions, combined with the scrap thresholds. A weighted average method is used to comprehensively consider the lifespan predictions for both capacity and internal resistance. Output the final remaining battery lifespan prediction.

[0082] In an optional embodiment, process division is performed based on the change characteristics of the battery historical operation data, and a state observation model and a variational inference network are used to extract multi-process features to form a time series feature vector. The battery dynamic characteristics are obtained through phase space reconstruction and symplectic geometry optimization, including:

[0083] Based on the change characteristics of the battery historical operation data, the battery historical operation data is divided into charging process data, discharging process data and static process data;

[0084] Establishing state observation models for the charging process data, the discharging process data, and the static process data, respectively, calculating state transition probabilities and observation probability densities through Gaussian distribution, and constructing probability distributions of the state observation models;

[0085] constructing a variational inference network based on the probability distribution, the variational inference network extracting an electrochemical state vector and thermodynamic characteristics from the charging process data, extracting energy efficiency characteristics and dynamic response characteristics from the discharging process data, and extracting polarization characteristics and self-discharge characteristics from the static process data, and combining the extracted characteristics in a time series to form a time series feature vector;

[0086] Based on the evolution trajectory of the time series eigenvector in the phase space, local potential energy function and pairwise potential energy function are constructed. Combined with symplectic geometry optimization, the dynamic characteristics of the battery are obtained by reconstructing the fractal characteristics of the attractor.

[0087] In one specific embodiment, battery parameters such as voltage, current, and temperature are collected during actual use. The historical battery operation data is then divided into charging process data, discharging process data, and static process data based on the current direction and amplitude variation characteristics. Specifically, when the current value is positive and the duration exceeds a preset threshold (e.g., 30 seconds), the charging process is determined; when the current value is negative and the duration exceeds the preset threshold, the discharging process is determined; and when the current value is close to zero (e.g., the absolute value is less than 0.05C) and the duration exceeds the preset threshold, the static process is determined.

[0088] Taking the charging process as an example, voltage, current, and temperature are selected as observation variables, and the state of charge SOC, internal resistance, polarization voltage, etc. are set as state variables. The state transition probability is calculated through Gaussian distribution, that is, the conditional probability relationship between the state variable at the current moment and the state variable at the previous moment. For example, for a certain 18650 lithium battery, the state transition probability of SOC changing from 0.5 to 0.51 during charging can be calculated through a Gaussian distribution with a mean of 0.51 and a variance of 0.0001. Similarly, the observation probability density is calculated, that is, the conditional probability relationship between the observation variable and the state variable. For example, when the SOC is 0.8, the observation probability of the voltage can be represented by a Gaussian distribution with a mean of 4.05V and a variance of 0.01. A similar method is used to establish a state observation model for the discharge process and the static process.

[0089] Based on this probability distribution, a variational inference network is constructed, employing an encoder-decoder structure. The encoder maps observed data into a latent state space, and the decoder reconstructs the latent state into observed data. For charging process data, the variational inference network extracts electrochemical state vectors (including state of charge, electrochemical reaction rates, etc.) and thermodynamic characteristics (including entropy change, enthalpy change, etc.). For example, analysis of charging data for a certain lithium battery model at 25°C shows that in the SOC range of 0.7-0.8, the average electrochemical reaction rate is 0.023 mol / (L·s), and the average entropy change is 85 J / (mol·K).

[0090] Energy efficiency and dynamic response characteristics are extracted from discharge process data. Energy efficiency characteristics include indicators such as coulombic efficiency and energy efficiency, while dynamic response characteristics include voltage response time constant and current step response. For example, when testing a power battery at a 1C discharge rate, the coulombic efficiency is 0.985 and the voltage response time constant is 12.5 seconds.

[0091] For the static process data, polarization characteristics and self-discharge characteristics are extracted. Polarization characteristics include polarization resistance and polarization time constant, while self-discharge characteristics include self-discharge rate and capacity loss rate. For example, when a certain energy storage battery is static for 24 hours at 25°C, the self-discharge rate is measured to be 0.15% / day and the polarization resistance is 25mΩ.

[0092] The extracted features are combined in time series to form a time series feature vector. For data with a sampling interval of 1 second, a set of features is extracted every 10 minutes to form a time series feature vector containing electrochemical state vector, thermodynamic features, energy efficiency features, dynamic response features, polarization features and self-discharge features.

[0093] Finally, based on the evolution of the time series eigenvectors in phase space, the battery dynamic characteristics are constructed and phase space reconstruction is performed. The time series eigenvectors are mapped into phase space using an embedding dimension of 5 and a time delay of 3. A local potential energy function is then constructed to describe the energy distribution of a single state point, and a pairwise potential energy function is constructed to describe the interaction between state points. For example, analysis of a certain battery model shows that the local potential energy function exhibits a bowl-shaped distribution in the SOC range of 0.4-0.6, indicating strong system stability in this range.

[0094] By combining symplectic geometry optimization with the energy conservation properties of the system, an iterative optimization algorithm (such as the conjugate gradient method) was used to find the optimal state trajectory of the system. During the optimization process, the learning rate was set to 0.01, the maximum number of iterations was 1000, and the convergence threshold was 0.0001.

[0095] Metrics such as the correlation dimension and Lyapunov exponent of attractors are calculated to characterize the complexity and stability of battery systems. For example, under normal operating conditions, the correlation dimension of the attractor of a certain battery model is 2.35, and the maximum Lyapunov exponent is 0.023, indicating that the system has some chaotic characteristics but remains generally stable. An attractor specifically refers to the set of states that a system ultimately converges to and resides in during its evolution. Specifically, for a large class of initial states, the corresponding system trajectory gradually approaches and stabilizes on this set over time. Attractors can take various forms, such as point attractors (fixed points): where the system states eventually converge to a stable equilibrium point; periodic attractors (limit cycles): where the system states cyclically move along a periodic trajectory; and strange attractors: complex attractors with fractal structures that often appear in chaotic systems. In the battery dynamic feature extraction method, the evolution trajectory of time-series feature vectors in phase space is analyzed to construct local and pairwise potential energy functions. Combined with symplectic geometry optimization, the fractal characteristics of the attractor are reconstructed to reflect the dynamic behavior of the battery. That is, the attractor represents the dynamic stable state or cyclic behavior exhibited by the battery during long-term operation.

[0096] This method can extract rich dynamic features from historical battery operating data, providing strong support for battery state assessment, life prediction, and fault diagnosis. Experimental verification shows that for a certain model of lithium-ion battery, the dynamic features extracted using this method for capacity prediction improved prediction accuracy by 15.7% compared to traditional methods, and the root mean square error was reduced to 0.023Ah.

[0097] In an optional embodiment, based on the evolution trajectory of the time series eigenvector in the phase space, a local potential energy function and a pairwise potential energy function are constructed, and combined with symplectic geometry optimization, the battery dynamic characteristics are obtained by reconstructing the fractal characteristics of the attractor, including:

[0098] The Hamiltonian dynamics equation is used to describe the evolution trajectory of the time series eigenvector, which is then projected into the phase space to form a phase orbit. The local potential energy function is calculated based on the Lyapunov exponent of the phase orbit, and the pairwise potential energy function is calculated based on the degree of synchronization between the phase orbits.

[0099] Projecting the evolution trajectory of the time series feature vector in the phase space onto the Poincare section for reconstruction, and extracting the electrochemical feature set and the thermodynamic feature set;

[0100] Calculating the invariant measures of the electrochemical characteristic set and the thermodynamic characteristic set on the Poincare section to obtain an electrochemical energy term and a thermodynamic energy term, respectively;

[0101] Based on the long-range correlation characteristics of the electrochemical energy term and the thermodynamic energy term, a fractional differential operator is constructed, and a coupling energy function is calculated based on the order of the fractional differential operator;

[0102] Substituting the electrochemical energy term, the thermodynamic energy term, and the coupled energy function into a symplectic geometry optimizer, keeping the Hamiltonian structure unchanged and minimizing global energy, to obtain an optimized energy function;

[0103] The attractor in the phase space is reconstructed based on the optimized energy function, the fractal dimension and correlation dimension of the attractor are calculated, and the dynamic characteristics of the battery are obtained by combining the dimensional feature reconstruction.

[0104] In one specific embodiment, time series data is acquired during battery operation, including parameters such as voltage, current, and temperature. The raw data is preprocessed, including denoising, normalization, and dimensionality reduction, to produce a time series feature vector. For a certain lithium battery model, for example, data is collected continuously for four hours at a sampling frequency of 10 Hz. After wavelet transform denoising and principal component analysis dimensionality reduction, an 8-dimensional time series feature vector is obtained.

[0105] The time series feature vectors are projected into phase space to form evolution trajectories. A time delay of τ = 20 and an embedding dimension of m = 8 are selected to construct the phase space. In phase space, each point represents the state of the system at a specific moment, and the lines connecting adjacent points form trajectories. For the example battery, the phase space trajectory exhibits nonlinear characteristics, indicating the complexity of the battery dynamics system.

[0106] Based on the phase space trajectory, a local potential energy function is calculated. A reference point is selected in phase space, and the divergence rate of the trajectory near that point, known as the Lyapunov exponent, is calculated. For the example battery, 500 reference points are uniformly selected in phase space, and the local Lyapunov exponent is calculated for each point. The values range from [-0.15 to 0.23], with positive values indicating system instability and negative values indicating stability. These exponents serve as the basis for the local potential energy function, with higher potential energy values indicating a more unstable system.

[0107] Calculate a pairwise potential function to characterize the interactions between different trajectories in phase space. Select two trajectories and calculate their degree of synchronization, including both phase synchronization and generalized synchronization. For a sample battery, calculate the synchronization index between any two points in phase space. The value range is [0, 1], where 0 indicates complete desynchronization and 1 indicates complete synchronization. Based on the synchronization index, construct a pairwise potential function. The lower the degree of synchronization, the higher the potential value.

[0108] Project the evolution trajectory of the time series eigenvector in phase space onto the Poincare section, select a hyperplane in phase space, and record the sequence of intersections between the trajectory and the hyperplane. For the example battery, select a hyperplane with a zero eigenvalue in the first dimension as the Poincare section, obtaining approximately 2,000 intersection points.

[0109] Extract electrochemical and thermodynamic feature sets from points on the Poincare cross section. The electrochemical feature set includes parameters such as charge transfer impedance, double-layer capacitance, and diffusion coefficient; the thermodynamic feature set includes parameters such as entropy change, enthalpy change, and free energy. For the example battery, the electrochemical feature set includes six parameters, and the thermodynamic feature set includes four parameters.

[0110] The invariant measures of the electrochemical and thermodynamic feature sets on the Poincare section are calculated to obtain the electrochemical energy term and the thermodynamic energy term, respectively. The distribution density of the characteristic points on the Poincare section is statistically analyzed, a probability density function is constructed, and the information entropy is calculated as the invariant measure. For the example battery, the electrochemical energy term is 3.27, and the thermodynamic energy term is 2.85.

[0111] Based on the long-range correlation between the electrochemical and thermodynamic energy terms, a fractional differential operator was constructed. The cross-correlation function between the two energy terms was analyzed to determine the strength of the long-range dependence and map it to the order of the fractional differential operator. For a sample battery, the order of the fractional differential operator was determined to be 0.78 by calculating the decay characteristics of the cross-correlation function.

[0112] The coupled energy function is calculated based on the order of the fractional differential operator. Substituting the electrochemical and thermodynamic energy terms into the fractional differential equation, the coupled energy function is solved. For the example battery, the value of the coupled energy function is 1.92.

[0113] The electrochemical energy terms, thermodynamic energy terms, and coupled energy function are fed into a symplectic geometry optimizer, which maintains the Hamiltonian structure and iteratively minimizes the global energy. The symplectic Euler method is used for numerical solution, with an iteration step of 0.01 and a maximum number of 1000 iterations. For the example battery, after 723 iterations, the global energy is reduced to 37% of the original value, resulting in an optimized energy function value of 2.96.

[0114] The attractor in phase space is reconstructed based on the optimized energy function. The optimized energy function is used as the potential energy field of the system. The phase space trajectory is reconstructed through numerical integration to form the attractor structure. For the example battery, the reconstructed attractor exhibits a complex geometric structure.

[0115] The fractal dimension and correlation dimension of the attractor were calculated using the box counting method (fractal dimension) and the correlation dimension using the correlation integral method. For the example battery, the fractal dimension is 2.37 and the correlation dimension is 2.15, indicating that the system has chaotic characteristics.

[0116] By combining dimensional features to reconstruct the battery's dynamic characteristics, a mapping relationship between the fractal dimension and correlation dimension and battery performance parameters was established to construct a battery dynamic feature vector. For the example battery, the resulting dynamic characteristics include key indicators such as a stability index of 0.82, a complexity index of 0.75, and a health score of 0.93. These indicators can be used for battery status monitoring and life prediction.

[0117] like Figure 2The figure shows a comprehensive performance evaluation of battery dynamic signature analysis technology, including a radar chart, performance curves, and data tables. The radar chart on the left compares six core performance metrics: the proposed solution (solid circle markers) significantly outperforms traditional signature methods (dashed square markers) and deep learning methods (dash-dotted diamond markers) in all six dimensions: stability index (0.82), complexity index (0.75), health score (0.93), remaining life accuracy (0.94), fault warning sensitivity (0.87), and dynamic response speed (0.89). In particular, the proposed solution achieves high performance of 0.93 and 0.94 in health score and remaining life accuracy, respectively, demonstrating its significant advantages in battery state assessment and life prediction. The performance curves on the right show the performance score trends of the three methods across different battery operating cycles (initial, mid-term, late, and decay). The performance curve (solid line) of the proposed solution maintains a high level throughout the entire life cycle, with a significantly smaller performance degradation than the other two methods in the mid-term and decay phases. Crucially, this technical solution can detect performance degradation before the battery enters its degradation phase (at the warning point), significantly earlier than other methods. The table at the bottom provides detailed performance evaluation data: the prediction accuracy of this technical solution is 94.3%, a 23.3% improvement over the 76.5% of the traditional feature method and a 9.5% improvement over the 86.1% of the deep learning method. The early warning rate is 92.7%, a 35.9% improvement over the 68.2% of the traditional feature method and a 16.6% improvement over the 79.5% of the deep learning method. The computational complexity is 0.75 (normalized value), which is higher than the 0.35 of the traditional feature method but lower than the 1.00 of the deep learning method. This demonstrates that this technical solution maintains high accuracy while also ensuring computational efficiency. Overall, this technical solution, by combining advanced technologies such as the Hamiltonian dynamic equation, Poincare section reconstruction, fractional differential operators, and a symplectic geometry optimizer, achieves high-precision analysis and prediction of battery dynamic characteristics, providing powerful technical support for battery health management and life prediction.

[0118] In this embodiment, the Hamiltonian dynamic equation is used to describe the evolution trajectory of the time-series eigenvectors. Phase space projection and Poincare section reconstruction are used to simultaneously extract both electrochemical and thermodynamic feature sets, fully capturing the multidimensional information of battery state changes. The local potential energy function is calculated based on the Lyapunov exponent of the phase orbital, and the pairwise potential energy function is calculated based on the degree of synchronization. This effectively quantifies the local stability of the system and the coupling relationship between different states, thereby revealing the underlying mechanism of battery dynamic behavior. The electrochemical and thermodynamic energy terms are obtained using invariant measures, and the coupled energy function is constructed in combination with fractional differential operators. This can reflect the long-range correlation characteristics between different energy terms and provide a precise quantitative method for further analysis of battery dynamics. By substituting each energy term into a symplectic geometry optimizer, the global energy is minimized while maintaining the Hamiltonian structure unchanged, ensuring that the optimization process complies with the original physical constraints of the system. The reconstructed phase space attractor more realistically reflects the actual battery dynamic behavior. Finally, by calculating the fractal dimension and correlation dimension of the attractor, the complex battery dynamic behavior is converted into quantifiable characteristic indicators, providing effective theoretical support for battery state monitoring, fault diagnosis, and performance optimization.

[0119] In an optional embodiment, a deep feature enhancement network is established to perform manifold projection and nonlinear transformation on the battery dynamic features, and optimized modeling is performed through a deep learning framework to obtain comprehensive battery features including:

[0120] Inputting the battery dynamic characteristics into a manifold embedding layer, calculating a reconstruction weight matrix through neighboring points of the manifold embedding layer, solving the characteristic equation using the reconstructed weight matrix to obtain a first low-dimensional embedding representation, constructing a geodesic distance matrix based on the first low-dimensional embedding representation, performing spectral decomposition on the geodesic distance matrix to obtain an eigenvector matrix and an eigenvalue matrix, and multiplying the eigenvector matrix by the square root of the eigenvalue matrix to generate a manifold embedding representation;

[0121] Inputting the manifold embedding representation into a dynamic kernel function enhancement module, constructing multiple basic kernel functions using a distance metric, calculating a combination weight of the multiple basic kernel functions using a deep network, performing a weighted combination of the combination weight and the multiple basic kernel functions to generate a combined kernel function, and performing a nonlinear transformation on the manifold embedding representation using the combined kernel function to generate enhanced features;

[0122] The enhanced features are input into the probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, construct a node distribution, and iteratively optimize through a message passing algorithm to obtain the comprehensive battery features.

[0123] In one embodiment, dynamic battery characteristic data is acquired, including timing parameters such as voltage, current, temperature, and internal resistance of the battery under different operating conditions. For example, the voltage curves of a certain lithium battery model at discharge rates of 0.5C, 1C, and 2C are collected, and the state of charge (SOC) range from 0% to 100% is recorded, with sampling every 5%, resulting in a total of 63 sets of feature vectors, each with a dimension of 21.

[0124] A deep feature enhancement network is established, which consists of three core modules: manifold embedding layer, dynamic kernel function enhancement module and probabilistic graph optimization module.

[0125] In the manifold embedding layer, the neighboring point relationships of the battery dynamic features are first calculated. For each feature point, the K points with the closest Euclidean distance are selected as its neighbors (K is 8 in the implementation). Then, the reconstruction weight matrix is calculated so that each point can be represented by a linear combination of its neighbors. Specifically, for feature point i, a set of weight values is solved to minimize the reconstruction error between the point and the linear combination of its neighboring points. In actual implementation, these weight values are obtained by solving a system of linear equations, and the sum of the weights is ensured to be 1.

[0126] Based on the reconstructed weight matrix, the characteristic equation is solved to obtain the first low-dimensional embedding representation. The original features are reduced from 21 dimensions to 12 dimensions, retaining the main information. Then, the geodesic distance matrix is constructed based on the first low-dimensional embedding representation. The geodesic distance represents the actual distance between the feature points on the manifold and is calculated by a graph shortest path algorithm (such as the Dijkstra algorithm). For example, for two feature points with SOC of 25% and 75%, the Euclidean distance is 3.6, while the geodesic distance is 5.2, which more accurately reflects the changes in battery characteristics.

[0127] Perform spectral decomposition on the geodesic distance matrix to obtain an eigenvector matrix and an eigenvalue matrix. Select the eigenvectors corresponding to the eight largest eigenvalues to form the eigenvector matrix. Multiply the eigenvector matrix by the square root of the eigenvalue matrix to generate a manifold embedding representation with a dimension of 8. This representation preserves the topological structure of the battery features in the manifold space and effectively captures the battery's dynamic characteristics.

[0128] In the dynamic kernel function enhancement module, multiple basic kernel functions are constructed using distance metrics. Gaussian, polynomial, Laplacian, and sigmoid kernels are selected as basic kernel functions. For example, the width parameter of the Gaussian kernel is set to 0.8, the order of the polynomial kernel is 3, the width parameter of the Laplacian kernel is 1.2, and the parameters of the sigmoid kernel are 0.5 and -1.0.

[0129] A deep network is used to calculate the combined weights of multiple basic kernel functions. It consists of three fully connected layers with 16, 8, and 4 nodes, respectively, and uses the Reluctant Unified Unit (ReLU) activation function. The input is the manifold embedding representation, and the output is the weights of the four basic kernel functions. In practice, for a sample with a SOC of 50%, the weights of the four kernel functions are 0.35, 0.25, 0.2, and 0.2, respectively.

[0130] The combined weights are then weighted together with multiple base kernel functions to generate a combined kernel function. This combined kernel function is then used to perform a nonlinear transformation on the manifold embedding representation to generate enhanced features. The enhanced features have a dimension of 10 and are more expressive than the original manifold embedding representation, enabling better differentiation of battery states under different operating conditions.

[0131] In the probabilistic graph optimization module, a conditional probability field is established based on the graph structure. Each dimension of the enhanced features is treated as a node in the graph, and the edges between nodes represent the correlations between feature dimensions. These correlations are learned using a deep neural network consisting of two convolutional layers and one fully connected layer. The convolution kernel size is 3×3, and the number of channels is 16 and 32, respectively.

[0132] The conditional probability field describes feature distributions using node potential functions and edge potential functions. Node potential functions represent the probability distribution of a single feature dimension, while edge potential functions represent the conditional probability relationship between feature dimensions. For example, for a sample with an SOC of 30%, the strength of the conditional probability relationship between the first and fifth dimensions is 0.78, indicating a high correlation between the two dimensions.

[0133] Using a message-passing algorithm, each node performs iterative optimization, updating its state based on information from connected nodes. The number of iterations is set to 5, and the convergence threshold is 0.001. The resulting comprehensive battery feature has a dimension of 12 and contains key information about the battery's dynamic characteristics.

[0134] In a validation experiment, the proposed method was applied to the battery state of health estimation task. Using 100 sets of battery data from different cycle times (ranging from 0 to 2000), the method improved estimation accuracy by 15.3% compared to traditional methods, reducing the root mean square error from 8.2% to 6.9%. The improvement in estimation accuracy was particularly significant, reaching 18.7%, in the mid-stage of battery aging (500-1500 cycles).

[0135] Furthermore, the proposed method is highly adaptable to variations in battery characteristics under varying operating conditions. When the temperature varies from -10°C to 45°C, the standard deviation of the estimation accuracy is only 2.1%, compared to 4.5% for conventional methods, demonstrating the proposed method's superior robustness.

[0136] In an optional embodiment, the enhanced features are input into a probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, constructs a node distribution, and iteratively optimizes through a message passing algorithm to obtain a comprehensive battery feature including:

[0137] The probabilistic graph optimization module constructs a conditional probability field for the enhanced feature based on the graph structure, wherein the nodes in the conditional probability field correspond to the dimensional components of the enhanced feature, and the edge connection relationship between the nodes forms an edge set;

[0138] The conditional probability field calculates the correlation between the components of each dimension of the enhanced feature through a deep neural network, the deep neural network maps the nodes to the hidden layer feature space to obtain the node hidden layer representation, and calculates the correlation strength between the nodes based on the node hidden layer representation;

[0139] Constructing an energy function of a conditional probability field according to the association strength, including a node potential energy term and an edge potential energy term, wherein the edge potential energy term is modulated by the association strength, and defining a conditional probability distribution based on the energy function;

[0140] The conditional probability field iteratively transmits node information on the edge set through a message passing algorithm under the conditional probability distribution, wherein the message passing algorithm calculates the information transmitted between nodes based on the node potential energy term, the edge potential energy term and the message update of the adjacent nodes;

[0141] The probability graph optimization module jointly optimizes the log-likelihood of the conditional probability distribution and the loss function of the deep neural network, alternately updates the parameters of the deep neural network and the parameters of the conditional probability field until convergence, and outputs the comprehensive battery characteristics.

[0142] In a specific embodiment, the enhanced feature is input into the probability graph optimization module. The enhanced feature can be a battery feature vector obtained by multimodal feature fusion, such as a feature vector containing multidimensional information such as battery voltage, current, temperature, impedance, etc., with a dimension of 128 dimensions.

[0143] A conditional probability field is established based on the graph structure, in which each node corresponds to a dimensional component of the enhanced feature. For example, for a 128-dimensional enhanced feature, 128 nodes are set in the conditional probability field, with node 1 corresponding to the first-dimensional feature component, node 2 corresponding to the second-dimensional feature component, and so on. The connections between nodes form an edge set. A full connection approach can be used, that is, an edge connection exists between any two nodes, forming a total of 8128 edges. Alternatively, a partial connection approach can be used, for example, each node is connected only to its 10 most relevant nodes, thereby reducing computational complexity.

[0144] A deep neural network (DNN) is used to enhance the correlation between the components of each feature dimension. This DNN uses a multi-layer perceptron architecture with three hidden layers, each with 64, 32, and 16 neurons, respectively, and uses a Reinforced Luminance (ReLU) activation function. The network maps each node to a hidden feature space, obtaining a hidden representation of the node. Specifically, for node i, its corresponding feature component value is xi, and after deep neural network mapping, a 16-dimensional hidden representation hi is obtained.

[0145] The association strength between nodes is calculated based on their hidden representations. For nodes i and j, the association strength wij is calculated between their hidden representations hi and hj. The association strength is calculated by normalizing the inner product of the hidden representations with a sigmoid function and ranges from 0 to 1. A higher association strength indicates a stronger correlation between the two nodes. For example, the association strength between nodes in the voltage and current dimensions might be 0.85, while the association strength between voltage and ambient humidity might be only 0.12.

[0146] The energy function of the conditional probability field is constructed based on the association strength. The energy function includes node potential energy terms and edge potential energy terms. The node potential energy term reflects the characteristics of a single node and can be set as the square difference between the node eigenvalue and the preset reference value. For example, for node i, its potential energy term can be (xi-μi)², where μi is the reference value of the node, which can be obtained through historical data statistics. The edge potential energy term reflects the interaction between nodes and is modulated by the association strength. For the edge between node i and node j, its potential energy term can be wij·(xi-xj)², which represents the square of the difference in the eigenvalues of the two nodes multiplied by the association strength. The greater the association strength, the greater the contribution of the edge potential energy term to the energy function.

[0147] The conditional probability distribution is defined based on the energy function. The conditional probability distribution is inversely proportional to the energy function: the lower the energy, the higher the probability. This means that the closer the value of the feature component is to the reference value, and the closer the values of the associated feature components are, the higher the probability.

[0148] In a conditional probability field, a message passing algorithm iteratively transfers node information along a set of edges. This algorithm employs a belief propagation approach. In each iteration, node i transmits a message mij to neighboring node j. This message is calculated based on the potential energy of node i, the potential energy of the edge between node i and node j, and the messages received by node i from all other neighboring nodes (excluding node j). For example, in iteration t, the message transmitted by node i to node j can be calculated based on the potential energy of node i, the potential energy of the edge (i, j), and the message mkj received by node i from all other neighboring nodes k in iteration t-1.

[0149] After 10 iterations of message passing, each node calculates the edge probability distribution based on its own potential energy term and the messages received from all neighboring nodes. These edge probability distributions together constitute the comprehensive characteristics of the battery.

[0150] The probabilistic graph optimization module jointly optimizes the log-likelihood of the conditional probability distribution and the loss function of the deep neural network. The log-likelihood reflects the model's fit to the observed data, while the deep neural network's loss function uses the mean squared error. The joint optimization goal is to maximize the log-likelihood while minimizing the network loss. The optimization process uses an alternating update strategy. First, the conditional probability field parameters are fixed and the deep neural network parameters are updated in five batches, each containing 64 samples. Then, the deep neural network parameters are fixed and the conditional probability field parameters are updated in three batches. The network parameters are updated using the Adam optimizer with a learning rate of 0.001, and the conditional probability field parameters are updated using gradient ascent with a step size of 0.01.

[0151] Alternating updates continue until the model converges, defined as the objective function change within five consecutive iterations being less than a preset threshold of 0.0001. Ultimately, the model outputs a 128-dimensional comprehensive battery feature, which fully accounts for the correlations between components across all dimensions and provides a more comprehensive representation of the battery state.

[0152] Specifically, for the test data of a certain model of lithium battery, the comprehensive features extracted by the above method were used to estimate the battery health status. The accuracy was improved from 89.3% of the traditional method to 95.7%, verifying the effectiveness of this method.

[0153] Existing battery feature optimization methods primarily include deep learning-based feature extraction methods and probabilistic graphical model-based feature optimization methods. Deep learning methods automatically learn feature representations through multi-layer neural networks, but often overlook explicit correlations between feature dimensions. While probabilistic graphical models can describe inter-variable dependencies, traditional probabilistic graphical methods employ fixed graph structures and simple potential functions, making it difficult to characterize the complex nonlinear correlations between battery features and prone to trapping the optimization process in local optima.

[0154] During feature optimization, it's difficult to simultaneously balance the deep representation capabilities of features with explicit modeling of associations between feature dimensions. The strength of associations between feature dimensions often adopts a predefined, fixed form, lacking a data-driven adaptive learning mechanism. The optimization objective is singular, making it difficult to balance the model's representational and reasoning capabilities. These issues prevent the extracted battery features from fully reflecting the coupling relationships between various battery state variables, impacting the accuracy of subsequent health state estimation. To address these issues, it's necessary to design a feature optimization method that combines the representational power of deep learning with the reasoning capabilities of probabilistic graphical models.

[0155] This embodiment proposes a conditional probability field feature optimization method based on deep neural network enhancement. It organically combines deep learning with probabilistic graphical models: the association strength between nodes is adaptively learned through deep neural networks, avoiding the limitations of predefined association relationships in traditional methods; an energy function containing node potential energy terms and edge potential energy terms is designed, where the edge potential energy terms are modulated by the association strength learned by the neural network, allowing the model to capture more complex feature dependencies; a message passing algorithm is used to transmit node information in the conditional probability field, and by jointly optimizing the log-likelihood of the conditional probability distribution and the loss function of the deep neural network, the collaborative optimization of representation learning and probabilistic reasoning is achieved.

[0156] Experimental verification shows that the comprehensive battery features extracted using the method of this embodiment improve the accuracy of battery health state estimation from 89.3% using traditional methods to 95.7%. This significant performance improvement demonstrates the superiority of this method in optimizing multidimensional battery features. Furthermore, this method exhibits good interpretability. By analyzing the learned correlation strengths, it can reveal the coupling patterns between different battery state variables, providing theoretical guidance for the optimal design of battery management systems.

[0157] like Figure 3 The figure shows the correlation strength distribution between battery feature dimensions. The contour lines in the figure represent correlation strengths, which are 0.2, 0.4, 0.6, 0.8, and 0.92 from the outside inward, with light to dark colors corresponding to increasing correlation strength. The detection path of this technical solution (solid dots) accurately captures key characteristic correlation points. In particular, in the core correlation region (center of the figure), it successfully identifies a strong correlation (0.92) between the voltage and current feature clusters. This is something that traditional methods—classical statistical feature extraction methods (dashed square dots)—cannot accurately locate. As can be seen in the figure, traditional methods only identify a correlation strength of 0.56 in the same region, 36 percentage points lower than this technical solution. At the intersection of the voltage and impedance feature clusters (coordinates approximately 200, 180), this technical solution detects a correlation strength of 0.58, while traditional methods only identify a 0.42. The correlation strength distribution statistics in the upper right corner of the figure further show that this technical solution identifies 42% of high-strength correlations (level 0.9), significantly higher than traditional methods. The comparison of the radar chart in the lower right corner also shows that this technical solution has significant advantages in the five dimensions of correlation identification accuracy, coverage, sensitivity, computational efficiency, and discrimination, among which the improvement in accuracy and sensitivity is particularly obvious. The contour distribution in the figure also reflects the correlation structure between different feature clusters. The voltage-current key area is the strongest correlation area, followed by the current-impedance area, while the temperature feature cluster has a relatively weak correlation with other features. This kind of detailed correlation identification is crucial to accurately capture changes in battery status, especially in predicting battery health status and remaining life.

[0158] In this embodiment, by mapping the components of each dimension of the enhanced feature to the hidden feature space, the correlation strength between each feature can be calculated more accurately, providing detailed information support for subsequent modeling; the energy function composed of the node potential energy term and the edge potential energy term modulated by the correlation strength enables the conditional probability field to more realistically reflect the interaction relationship between each feature, thereby improving the modeling accuracy; with the help of the message passing algorithm, the node information is iteratively transmitted on the edge set of the graph structure, which realizes the comprehensive utilization of local and global information and further improves the optimization effect; by jointly optimizing the log-likelihood of the conditional probability distribution and the loss function of the deep neural network, the parameters of each module are updated alternately, so that the overall model has significant improvements in convergence and robustness.

[0159] In an optional embodiment, the comprehensive battery features are input into the shared coding layer to obtain the timing calibration features, the capacity attenuation features and the impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch, and a state-dependent piecewise recursive coupling channel is established to perform feature fusion. The output capacity prediction results and internal resistance prediction results include:

[0160] The battery comprehensive features are input into the shared coding layer, and the shared coding layer performs layer-by-layer dimensionality reduction coding through a multi-layer convolutional network to obtain coding features. The shared coding layer extracts the time series correlation from the coding features to obtain the time series calibration features;

[0161] The timing calibration features are input into the capacity prediction branch and the internal resistance prediction branch respectively. The capacity prediction branch extracts the capacity attenuation features from the timing calibration features through a multi-layer perceptron, and the internal resistance prediction branch extracts the impedance evolution features from the timing calibration features through a recursive neural network.

[0162] A state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature, and bidirectional dynamic transmission and fusion are achieved through state matrix transformation and multi-stage response equations to obtain capacity fusion features and internal resistance fusion features;

[0163] A capacity prediction result is output based on the capacity fusion feature, and an internal resistance prediction result is output based on the internal resistance fusion feature.

[0164] In one specific embodiment, comprehensive battery feature data is prepared, including various parameters that affect battery performance, such as voltage, temperature, charge / discharge status, and cycle count. This feature data is processed as input in a shared coding layer. The shared coding layer utilizes a multi-layer convolutional network architecture, extracting deep information about the battery's characteristics through layer-by-layer dimensionality reduction. This process aims to reduce the dimensionality of the input features while preserving important temporal information.

[0165] In the shared encoding layer, multiple layers of convolution are performed on the input battery features. Each convolution layer convolves, activates, and pools the feature map, gradually extracting higher-level features. This ultimately results in an encoding feature matrix. Timing correlation analysis is used to extract timing alignment features from the encoding features. This process can be achieved using a sliding window technique, which slides a window across the feature matrix to extract features from different time periods, thereby capturing how battery performance changes over time.

[0166] The extracted timing calibration features are input into the capacity prediction branch and the internal resistance prediction branch, respectively. The capacity prediction branch uses a multi-layer perceptron (MLP) structure to further process the timing calibration features. The MLP consists of multiple fully connected layers, each of which performs a nonlinear transformation using an activation function, ultimately outputting capacity decay features. These capacity decay features reflect the capacity change trend of the battery during use.

[0167] Meanwhile, the internal resistance prediction branch employs a recurrent neural network (RNN) structure to extract impedance evolution features. RNNs are effective at processing sequential data and are well-suited for capturing the dynamic characteristics of battery internal resistance over time. By inputting time-series calibration features into the RNN and gradually updating the hidden state, the internal resistance evolution features are ultimately derived.

[0168] After obtaining the capacity decay and internal resistance evolution characteristics, a state-dependent piecewise recursive coupling channel is then established. The core of this channel is to achieve bidirectional dynamic transmission and fusion between the capacity and internal resistance characteristics through state matrix transformation and multi-stage response equations. Specifically, the capacity decay and internal resistance evolution characteristics are first linearly transformed using the state matrix to obtain a new feature representation. Then, the current state is updated using the multi-stage response equation in combination with historical state information, thereby achieving feature fusion.

[0169] During the fusion process, the capacity fusion feature and the internal resistance fusion feature are output separately. The capacity fusion feature is a comprehensive result based on the capacity decay feature and the internal resistance evolution feature, which can more accurately reflect the actual capacity state of the battery. The internal resistance fusion feature is the internal resistance state obtained through the fusion process, which can effectively reflect the health of the battery.

[0170] Based on the capacity fusion feature, the final capacity prediction result is output; based on the internal resistance fusion feature, the final internal resistance prediction result is output. These prediction results can provide important decision-making basis for the battery management system, helping to achieve optimized battery use and maintenance.

[0171] For example, assume that the comprehensive feature data of a certain battery is as follows: voltage is 3.7V, temperature is 25°C, charge state is 80%, discharge state is 20%, and number of cycles is 500. After processing through the shared coding layer, the resulting coding feature is a vector containing 128 features. After time series calibration feature extraction, the capacity attenuation feature output by the capacity prediction branch is 0.05Ah, while the internal resistance evolution feature output by the internal resistance prediction branch is 0.02Ω. Through the established segmented recursive coupling channel, the capacity fusion feature is finally obtained as 0.04Ah and the internal resistance fusion feature is 0.018Ω, which are used to output capacity prediction results and internal resistance prediction results respectively.

[0172] Through the above steps, accurate prediction of battery capacity and internal resistance can be achieved, providing a scientific basis for battery use and management.

[0173] In an optional embodiment, a state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature, and bidirectional dynamic transmission and fusion are achieved through state matrix transformation and multi-stage response equations. The obtained capacity fusion feature and internal resistance fusion feature include:

[0174] A dynamic collaborative optimization channel is established between the capacity decay characteristics and the impedance evolution characteristics. A state dependency matrix is constructed based on the battery operating state. The state dependency matrix includes a temperature gradient coefficient, a charge and discharge rate coefficient, and a cycle number coefficient. A piecewise matrix transformation is performed on the capacity decay characteristics and the impedance evolution characteristics according to the state dependency matrix. An independent feature mapping function is established in each operating state interval. The piecewise reconstruction of the features is achieved through function combination to obtain state-related features.

[0175] A capacity-internal resistance coupling model is constructed based on the state-related characteristics. The capacity-internal resistance coupling model is recursively updated to time-sequentially pair characteristic change points in the capacity degradation process with internal resistance mutation points, and establish a multi-stage capacity-internal resistance response equation to form a quantitative correlation between capacity decay and internal resistance growth with state memory;

[0176] According to the capacity-internal resistance coupling model, bidirectional feature compensation and enhancement are performed on the capacity attenuation feature and the impedance evolution feature to obtain a capacity fusion feature and an internal resistance fusion feature combined with state dependence.

[0177] In a specific embodiment, a dynamic collaborative optimization channel is established between the capacity attenuation characteristics and the impedance evolution characteristics. A state dependency matrix is constructed based on the battery working state, and the state dependency matrix includes a temperature gradient coefficient, a charge and discharge rate coefficient, and a cycle number coefficient. Specifically, the temperature gradient coefficient is divided into a low temperature range (-20°C to 0°C), a normal temperature range (0°C to 40°C), and a high temperature range (40°C to 60°C) according to the battery operating temperature range, corresponding to coefficient values of 0.8, 1.0, and 1.2, respectively; the charge and discharge rate coefficient is divided into a low rate range (0.2C to 0.5C), a medium rate range (0.5C to 2C), and a high rate range (2C to 5C) according to the battery charge and discharge rate, corresponding to coefficient values of 0.9, 1.0, and 1.3, respectively; the cycle number coefficient is divided into an initial stage (0 to 200 times), a mid-stage (200 to 500 times), and a late stage (more than 500 times) according to the battery cycle number, corresponding to coefficient values of 0.95, 1.0, and 1.1, respectively.

[0178] According to the state-dependent matrix, the capacity decay characteristics and impedance evolution characteristics are transformed in a piecewise matrix. An independent feature mapping function is established in each working state interval, and the feature is reconstructed in piecewise manner through function combination to obtain the state-related characteristics. For example, for a certain 18650 lithium-ion battery, under the conditions of 25°C and 1C discharge, the initial capacity is 3000mAh and the internal resistance is 25mΩ. After 300 cycles, the capacity decays to 2700mAh and the internal resistance increases to 35mΩ. At this time, the temperature gradient coefficient in the state-dependent matrix is 1.0, the charge and discharge rate coefficient is 1.0, and the cycle number coefficient is 1.0. Through piecewise matrix transformation, the state-related characteristics with a capacity decay rate of 10% and an internal resistance growth rate of 40% are obtained.

[0179] A capacity-internal resistance coupling model is constructed based on state-related features. Through recursive updates, the model pairs the characteristic change points in the capacity degradation process with the internal resistance mutation points in time series, and establishes a multi-stage capacity-internal resistance response equation to form a quantitative correlation between capacity decay and internal resistance growth with state memory. The specific implementation method is: first identify the inflection points in the capacity decay curve, such as the obvious slope changes when the number of cycles is 150, 350 and 450 times; at the same time, identify the mutation points in the internal resistance growth curve, such as the obvious growth acceleration when the number of cycles is 140, 340 and 460 times. Pair these characteristic points in time series to establish a corresponding relationship between capacity decay and internal resistance growth.

[0180] For example, for the aforementioned 18650 lithium-ion battery, during the 0-150 cycle period, the capacity decay rate is 3% and the internal resistance growth rate is 15%; during the 150-350 cycle period, the capacity decay rate is 5% and the internal resistance growth rate is 20%; during the 350-500 cycle period, the capacity decay rate is 7% and the internal resistance growth rate is 30%. Through recursive updating, a quantitative relationship between the capacity decay rate and the internal resistance growth rate is established: for every 1% increase in the capacity decay rate, the internal resistance growth rate increases by approximately 4% to 5%.

[0181] Based on the capacity-internal resistance coupling model, bidirectional feature compensation and enhancement are performed on the capacity decay and impedance evolution features, resulting in state-dependent capacity and internal resistance fusion features. Specifically, when the capacity decay feature is detected but the internal resistance feature data is missing, the coupling model can be used to predict the internal resistance value; conversely, when the internal resistance evolution feature is detected but the capacity feature data is missing, the coupling model can be used to predict the capacity value.

[0182] For example, under the conditions of 45°C and 2C discharge, the capacity of a power battery at the 400th cycle is measured to be 85% of the original capacity, but the internal resistance data is missing. At this time, the temperature gradient coefficient in the state dependency matrix is 1.2, the charge and discharge rate coefficient is 1.3, and the cycle number coefficient is 1.0. Calculated by the capacity-internal resistance coupling model, it is known that in this state, the internal resistance growth rate corresponding to a 15% capacity decay should be approximately 65%, that is, the internal resistance should increase to 1.65 times the original internal resistance. If the original internal resistance is 20mΩ, the current internal resistance is predicted to be approximately 33mΩ.

[0183] Through practical verification, the capacity fusion features and internal resistance fusion features obtained by this method have improved the accuracy of battery health status assessment by 12% and the accuracy of remaining life prediction by 15% compared with using capacity decay features or impedance evolution features alone. In particular, under extreme working conditions and data missing conditions, the prediction error is reduced by more than 20%, demonstrating the effectiveness and robustness of this method.

[0184] Existing battery performance prediction technologies are primarily based on single-feature modeling approaches, including internal resistance estimation methods based on equivalent circuit models, which simulate the battery's internal impedance characteristics by constructing RC networks; data-driven capacity decay prediction methods, which use support vector machines or neural networks to establish capacity degradation models; and coupled analysis methods based on electrochemical models, which establish a physical correlation equation between capacity and internal resistance. These methods often treat capacity decay and internal resistance growth as independent processes, failing to fully consider their mutual influence.

[0185] Traditional technical solutions have limitations in several areas. The correlation analysis between battery capacity decay and internal resistance growth ignores the dynamic nature of the coupling relationship under different operating conditions and lacks analysis of the battery's operating state. Furthermore, existing prediction models are unable to effectively handle extreme operating conditions and data loss, which compromises the reliability and accuracy of prediction results. These issues highlight the need to establish a more accurate capacity-internal resistance coupling relationship, consider the impact of operating conditions, and improve the robustness of prediction models.

[0186] To address the above issues, this embodiment proposes a piecewise recursive coupling method based on state dependence. By introducing a state dependency matrix, comprehensively considering the effects of temperature, rate, and number of cycles, and using piecewise coefficients to reflect the state dependency under different working conditions, adaptive weight adjustment of features is achieved. A dynamic coupling model is constructed, a time-series pairing mechanism for feature change points is established, a multi-stage response equation is used to describe the coupling relationship, and state memory is introduced to improve model accuracy. At the same time, two-way feature compensation is achieved, a capacity-internal resistance complementary prediction mechanism is established, the prediction capability is improved in the absence of data, and the model's adaptability to extreme working conditions is enhanced.

[0187] After practical verification, the method of this embodiment has improved the accuracy of battery health assessment by 12%, the accuracy of remaining life prediction by 15%, and the prediction error under extreme operating conditions and data loss by more than 20%. This method significantly improves the reliability of the battery management system, effectively reduces battery safety risks, and provides important support for optimizing battery usage strategies. These improvements demonstrate the innovative value of this invention in both theoretical and practical applications, and provide a superior technical solution for the field of battery life prediction.

[0188] like Figure 4As shown in the figure, the construction process of the state-dependent matrix and its optimization effect on prediction accuracy are demonstrated. In the matrix visualization in the upper part, the weight coefficient of each parameter in different cycle stages is represented by the size of the circle. It can be seen that the temperature gradient coefficient has the greatest impact in the initial stage (0.85), the charge and discharge rate coefficient contributes the most in the mid-term stage (0.84), and the cycle number coefficient dominates in the later stage (0.91). At the same time, the capacity characteristic weight is the highest in the initial stage (0.92), while the internal resistance characteristic weight reaches the maximum in the later stage (0.88). This dynamic weight allocation mechanism effectively captures the characteristic evolution law of the battery under different working conditions. In the prediction accuracy comparison diagram in the lower part, by optimizing the state-dependent matrix, this technical solution still maintains a prediction accuracy of 93.5% after 1200 cycles, which is 6.8 percentage points higher than the 86.7% before optimization and 12.7 percentage points higher than the 80.8% of the traditional model. In particular, at the two state transition points of 400 cycles and 800 cycles, this technical solution successfully achieved a smooth transition in prediction accuracy by recursively updating the capacity-internal resistance coupling model, avoiding the sharp drop in accuracy of the traditional model at the state transition points. This shows that this technical solution effectively established a quantitative correlation between capacity decay and internal resistance growth by constructing a state-dependent matrix of temperature gradient coefficient, charge and discharge rate coefficient, and cycle number coefficient, and performing a piecewise matrix transformation, significantly improving the accuracy and robustness of battery state prediction.

[0189] In this embodiment, by establishing a collaborative optimization channel between capacity attenuation and impedance evolution characteristics, the key state changes in the battery degradation process can be more accurately captured and reflected; a state-dependent matrix is constructed based on temperature gradient, charge and discharge rate, and number of cycles, and segmented matrix transformation and feature mapping are realized to effectively adapt to the dynamic changes of battery performance under different working conditions; by constructing a capacity-internal resistance coupling model with state memory, the characteristic change points of capacity degradation are time-series paired with the internal resistance mutation points, and a multi-stage response equation is formed, thereby improving the quantitative correlation accuracy of battery state changes; bidirectional feature compensation is performed on the basis of the model, and finally the capacity and internal resistance characteristics that integrate state dependence are obtained, which helps to improve the prediction and diagnosis capabilities of the battery management system and extend the battery life.

[0190] The battery life accurate prediction and evaluation system based on deep learning in an embodiment of the present invention includes:

[0191] The first unit is used to obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle number as battery historical operation data;

[0192] The second unit is used to divide the process based on the change characteristics of the battery's historical operating data. It uses the state observation model and variational inference network to extract multi-process features to form a time series feature vector, and obtains the battery dynamic characteristics through phase space reconstruction and symplectic geometry optimization;

[0193] The third unit is used to establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize the modeling through the deep learning framework, and obtain the comprehensive characteristics of the battery;

[0194] The fourth unit is used to input the comprehensive battery features into the shared coding layer to obtain timing calibration features, extract the capacity decay features and impedance evolution features through the capacity prediction branch and the internal resistance prediction branch, establish a state-dependent piecewise recursive coupling channel, perform feature fusion, and output the capacity prediction results and internal resistance prediction results;

[0195] The fifth unit is used to determine the remaining service life of the battery according to the capacity prediction result and the internal resistance prediction result in combination with the battery scrapping threshold.

[0196] According to a third aspect of the embodiments of the present invention,

[0197] An electronic device is provided, comprising:

[0198] processor;

[0199] a memory for storing processor-executable instructions;

[0200] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0201] According to a fourth aspect of the embodiments of the present invention,

[0202] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0203] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery life prediction and evaluation method based on deep learning, characterized by: include: Obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle count as battery historical operation data; Based on the changing characteristics of the battery's historical operating data, the process is divided. The state observation model and variational inference network are used to extract multi-process features to form a time series feature vector. The battery dynamic characteristics are obtained through phase space reconstruction and symplectic geometry optimization, including: Based on the change characteristics of the battery historical operation data, the battery historical operation data is divided into charging process data, discharging process data and static process data; Establishing state observation models for the charging process data, the discharging process data, and the static process data, respectively, calculating state transition probabilities and observation probability densities through Gaussian distribution, and constructing probability distributions of the state observation models; constructing a variational inference network based on the probability distribution, the variational inference network extracting an electrochemical state vector and thermodynamic characteristics from the charging process data, extracting energy efficiency characteristics and dynamic response characteristics from the discharging process data, and extracting polarization characteristics and self-discharge characteristics from the static process data, and combining the extracted characteristics in a time series to form a time series feature vector; Based on the evolution trajectory of the time series eigenvector in the phase space, the local potential energy function and the pairwise potential energy function are constructed. Combined with symplectic geometry optimization, the dynamic characteristics of the battery are obtained by reconstructing the fractal characteristics of the attractor. Establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize and model them through a deep learning framework, and obtain comprehensive battery features; The comprehensive battery features are input into the shared coding layer to obtain timing calibration features. The capacity attenuation features and impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch. A state-dependent piecewise recursive coupling channel is established to perform feature fusion and output the capacity prediction results and internal resistance prediction results. The remaining service life of the battery is determined based on the capacity prediction result and the internal resistance prediction result in combination with a battery scrapping threshold.

2. The method according to claim 1, characterized in that Based on the evolution trajectory of the time series eigenvector in the phase space, the local potential energy function and the pairwise potential energy function are constructed. Combined with symplectic geometry optimization, the battery dynamic characteristics are obtained by reconstructing the fractal characteristics of the attractor, including: The Hamiltonian dynamics equation is used to describe the evolution trajectory of the time series eigenvector, which is then projected into the phase space to form a phase orbit. The local potential energy function is calculated based on the Lyapunov exponent of the phase orbit, and the pairwise potential energy function is calculated based on the degree of synchronization between the phase orbits. Projecting the evolution trajectory of the time series feature vector in the phase space onto the Poincare section for reconstruction, and extracting the electrochemical feature set and the thermodynamic feature set; Calculating the invariant measures of the electrochemical characteristic set and the thermodynamic characteristic set on the Poincare section to obtain an electrochemical energy term and a thermodynamic energy term, respectively; Based on the long-range correlation characteristics of the electrochemical energy term and the thermodynamic energy term, a fractional differential operator is constructed, and a coupling energy function is calculated based on the order of the fractional differential operator; Substituting the electrochemical energy term, the thermodynamic energy term, and the coupled energy function into a symplectic geometry optimizer, keeping the Hamiltonian structure unchanged and minimizing global energy, to obtain an optimized energy function; The attractor in the phase space is reconstructed based on the optimized energy function, the fractal dimension and correlation dimension of the attractor are calculated, and the dynamic characteristics of the battery are obtained by combining the dimensional feature reconstruction.

3. The method according to claim 1, characterized in that A deep feature enhancement network is established to perform manifold projection and nonlinear transformation on the battery dynamic features. Optimized modeling is performed through a deep learning framework to obtain comprehensive battery features including: Inputting the battery dynamic characteristics into a manifold embedding layer, calculating a reconstruction weight matrix through neighboring points of the manifold embedding layer, solving the characteristic equation using the reconstructed weight matrix to obtain a first low-dimensional embedding representation, constructing a geodesic distance matrix based on the first low-dimensional embedding representation, performing spectral decomposition on the geodesic distance matrix to obtain an eigenvector matrix and an eigenvalue matrix, and multiplying the eigenvector matrix by the square root of the eigenvalue matrix to generate a manifold embedding representation; Inputting the manifold embedding representation into a dynamic kernel function enhancement module, constructing multiple basic kernel functions using a distance metric, calculating a combination weight of the multiple basic kernel functions using a deep network, performing a weighted combination of the combination weight and the multiple basic kernel functions to generate a combined kernel function, and performing a nonlinear transformation on the manifold embedding representation using the combined kernel function to generate enhanced features; The enhanced features are input into the probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, construct a node distribution, and iteratively optimize through a message passing algorithm to obtain the comprehensive battery features.

4. The method according to claim 3, characterized in that The enhanced features are input into the probability graph optimization module, and a conditional probability field is established based on the graph structure. The conditional probability field uses a deep neural network to learn the correlation between the components of each dimension of the enhanced features, construct a node distribution, and iteratively optimize through a message passing algorithm to obtain the comprehensive battery features including: The probabilistic graph optimization module constructs a conditional probability field for the enhanced feature based on the graph structure, wherein the nodes in the conditional probability field correspond to the dimensional components of the enhanced feature, and the edge connection relationship between the nodes forms an edge set; The conditional probability field calculates the correlation between the components of each dimension of the enhanced feature through a deep neural network, the deep neural network maps the nodes to the hidden layer feature space to obtain the node hidden layer representation, and calculates the correlation strength between the nodes based on the node hidden layer representation; Constructing an energy function of a conditional probability field according to the association strength, including a node potential energy term and an edge potential energy term, wherein the edge potential energy term is modulated by the association strength, and defining a conditional probability distribution based on the energy function; The conditional probability field iteratively transmits node information on the edge set through a message passing algorithm under the conditional probability distribution, wherein the message passing algorithm calculates the information transmitted between nodes based on the node potential energy term, the edge potential energy term and the message update of the adjacent nodes; The probability graph optimization module jointly optimizes the log-likelihood of the conditional probability distribution and the loss function of the deep neural network, alternately updates the parameters of the deep neural network and the parameters of the conditional probability field until convergence, and outputs the comprehensive battery characteristics.

5. The method according to claim 1, wherein The comprehensive battery features are input into the shared coding layer to obtain the timing calibration features. The capacity attenuation features and impedance evolution features are extracted through the capacity prediction branch and the internal resistance prediction branch. A state-dependent piecewise recursive coupling channel is established to perform feature fusion. The output capacity prediction results and internal resistance prediction results include: The battery comprehensive features are input into the shared coding layer, and the shared coding layer performs layer-by-layer dimensionality reduction coding through a multi-layer convolutional network to obtain coding features. The shared coding layer extracts the time series correlation from the coding features to obtain the time series calibration features; The timing calibration features are input into the capacity prediction branch and the internal resistance prediction branch respectively. The capacity prediction branch extracts the capacity attenuation features from the timing calibration features through a multi-layer perceptron, and the internal resistance prediction branch extracts the impedance evolution features from the timing calibration features through a recursive neural network. A state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature, and bidirectional dynamic transmission and fusion are achieved through state matrix transformation and multi-stage response equations to obtain capacity fusion features and internal resistance fusion features; A capacity prediction result is output based on the capacity fusion feature, and an internal resistance prediction result is output based on the internal resistance fusion feature.

6. The method according to claim 5, characterized in that A state-dependent piecewise recursive coupling channel is established between the capacity decay feature and the impedance evolution feature. Through state matrix transformation and multi-stage response equations, bidirectional dynamic transmission and fusion are achieved. The obtained capacity fusion feature and internal resistance fusion feature include: A dynamic collaborative optimization channel is established between the capacity decay characteristics and the impedance evolution characteristics. A state dependency matrix is constructed based on the battery operating state. The state dependency matrix includes a temperature gradient coefficient, a charge and discharge rate coefficient, and a cycle number coefficient. A piecewise matrix transformation is performed on the capacity decay characteristics and the impedance evolution characteristics according to the state dependency matrix. An independent feature mapping function is established in each operating state interval. The piecewise reconstruction of the features is achieved through function combination to obtain state-related features. A capacity-internal resistance coupling model is constructed based on the state-related characteristics. The capacity-internal resistance coupling model is recursively updated to time-sequentially pair characteristic change points in the capacity degradation process with internal resistance mutation points, and establish a multi-stage capacity-internal resistance response equation to form a quantitative correlation between capacity decay and internal resistance growth with state memory; According to the capacity-internal resistance coupling model, bidirectional feature compensation and enhancement are performed on the capacity attenuation feature and the impedance evolution feature to obtain a capacity fusion feature and an internal resistance fusion feature combined with state dependence.

7. A battery life accurate prediction and evaluation system based on deep learning, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain battery charge and discharge voltage, battery charge and discharge current, battery temperature, battery internal resistance and battery cycle number as battery historical operation data; The second unit is used to divide the process based on the change characteristics of the battery's historical operating data. It uses the state observation model and variational inference network to extract multi-process features to form a time series feature vector, and obtains the battery dynamic characteristics through phase space reconstruction and symplectic geometry optimization; The third unit is used to establish a deep feature enhancement network, perform manifold projection and nonlinear transformation on the battery dynamic features, optimize the modeling through the deep learning framework, and obtain the comprehensive characteristics of the battery; The fourth unit is used to input the comprehensive battery features into the shared coding layer to obtain timing calibration features, extract the capacity decay features and impedance evolution features through the capacity prediction branch and the internal resistance prediction branch, establish a state-dependent piecewise recursive coupling channel, perform feature fusion, and output the capacity prediction results and internal resistance prediction results; The fifth unit is used to determine the remaining service life of the battery according to the capacity prediction result and the internal resistance prediction result in combination with the battery scrapping threshold.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images