A lithium battery soh estimation method and system based on multi-source data fusion

Abstract

This invention discloses a lithium battery SOH estimation method based on multi-source data fusion, comprising the following steps: S1, constructing a vehicle-side local sensor and a cloud-based global memory network, processing them, and outputting an estimated SOH value; S2, filtering future degradation trajectories; S3, generating an aging state encoding vector and a battery internal aging artifact map, mapping them, and obtaining a confidence interval; S4, outputting a consensus-reached SOH; S5, obtaining a personalized SOH estimate; S6, inputting the SOH estimate, the consensus-reached SOH, and the personalized SOH estimate as three channels into an electrochemical-data dual-driven game theory fusion module, fusing them according to the future degradation trajectory and the confidence interval, and outputting the fused SOH estimation result. This invention uses meta-learning to customize a model for each battery, and through continuous learning, the overall system performance continuously improves over time.

Description

Technical Field

[0001] This invention relates to the field of battery processing technology, specifically to a lithium battery SOH estimation method and system based on multi-source data fusion. Background Technology

[0002] Accurate estimation of the state of health (SOH) of lithium batteries is crucial for battery safety management and lifespan prediction. Existing technologies suffer from the following main shortcomings:

[0003] Poor data adaptability: Most methods rely on complete charge and discharge cycle data, which makes it difficult to cope with random and fragmented charging behaviors common in real-world applications (such as daily short-term charging), leading to failure in real-world scenarios.

[0004] The estimation dimension is singular: most of them only provide a single-point estimate of the current SOH, lack the ability to predict the future aging trajectory of the battery, and cannot meet the needs of early warning and life planning.

[0005] Mechanism and data are disconnected: Purely data-driven methods lack physical interpretability and may give results that violate electrochemical laws; while pure mechanistic models are too complex and difficult to parameterize precisely.

[0006] Rigid system collaboration: Existing vehicle-cloud collaboration solutions are mostly simple data upload-model distribution modes, which fail to achieve deep and adaptive collaboration between the powerful computing power of the cloud and the real-time perception of the vehicle. Summary of the Invention

[0007] To address the above problems, this invention proposes a lithium battery SOH estimation method and system based on multi-source data fusion.

[0008] The technical solution of this invention is: a lithium battery SOH estimation method based on multi-source data fusion, comprising the following steps:

[0009] S1. Construct vehicle-side local sensors and cloud-based global memory networks, process them, and output the estimated SOH value;

[0010] S2. Extract the current state snapshot and historical path fingerprint, process them, and filter future decay trajectories;

[0011] S3. Based on the virtual stress waveform, generate the aging state encoding vector and the battery internal aging artifact map, perform mapping, and obtain the confidence interval;

[0012] S4. Conduct game bidding and consensus decision-making to obtain the game optimization objective and output the consensus-reached SOH;

[0013] S5. Construct personalized calculation units to obtain personalized SOH estimates;

[0014] S6. Input the SOH estimate, the consensus SOH, and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, and output the fused SOH estimate result.

[0015] Furthermore, S1 includes the following sub-steps:

[0016] S11. Construct a vehicle-side local perceptron and extract memory fragment feature vectors;

[0017] S12. Construct a cloud-based global memory network and determine memory retrieval weights;

[0018] S13. Determine the virtual complete features based on the memory fragment feature vector and memory retrieval weight;

[0019] S14. Feed the virtual complete features into the prediction network and output the SOH estimate.

[0020] Furthermore, in S11, the memory fragment feature vector The expression is:

[0021] ;

[0022] in, The average voltage. For voltage variance, The average temperature. For the set of real numbers, The dimension of the feature vector;

[0023] In S12, the memory read weight w i The expression is:

[0024] ;

[0025] Where β is the attention scaling factor, sim(⋅) is the similarity function, exp(⋅) is the exponential function, g(⋅) is the query vector mapping function, k is the query vector, and M... i For the memory matrix of the first Okay, M j For the memory matrix of the th Okay, SOC0 is the starting SOC for charging, SOC end The state of charge (SOC) is the end of charging; T is the ambient temperature.

[0026] In S13, the virtual complete features The expression is:

[0027] ;

[0028] Where W1 is the memory projection weight matrix, M i For the memory matrix of the th Line, W2 is the fragment projection weight matrix, and b is the bias vector;

[0029] In S14, the estimated value of SOH The expression is:

[0030] ;

[0031] MLP stands for Multilayer Perceptron.

[0032] Furthermore, S2 includes the following sub-steps:

[0033] S21. Based on the SOH estimation value, extract state features and path features, which are used as the current state snapshot and historical path fingerprint, respectively.

[0034] S22. Using historical path fingerprints as conditions, generate several future SOH decay trajectories;

[0035] S23. Match the current state snapshot with the current point of all future SOH decay trajectories, calculate the matching probability, and filter future decay trajectories.

[0036] Furthermore, in S22, the expression for generating several future SOH decay trajectories is as follows:

[0037] ;

[0038] Where z is a latent variable. (⋅) represents a multivariate normal distribution, μ(⋅) is the mean function, and σ 2 (⋅) is the variance function, x path For path features, Let Dec(⋅) be the k-th prediction at time t, where Dec(⋅) is the decoder function and T is the total prediction step size of the future decay trajectory.

[0039] In S23, the matching probability The expression is:

[0040] ;

[0041] Where, x state As a characteristic of the current state, g k Let τ be the prototype of the k-th target, and τ be the temperature parameter.

[0042] Furthermore, S3 includes the following sub-steps:

[0043] S31. Collect voltage, current and temperature time-series data to generate virtual stress waveforms;

[0044] S32. Compress the virtual stress waveform with SOC and cumulative throughput into an aging state encoding vector;

[0045] S33. Construct a spatiotemporal hierarchical mapping network, use the spatial decoder of the spatiotemporal hierarchical mapping network to receive the aging state encoding vector, and use the deconvolutional neural network of the spatiotemporal hierarchical mapping network to reconstruct it to generate an aging artifact map inside the battery.

[0046] S34. Input the battery internal aging artifact image into the convolutional network, perform mapping, and obtain the confidence interval.

[0047] Furthermore, in S31, the expression for the virtual stress waveform s(t) is:

[0048] ;

[0049] in, Let I(t) be the stress mapping function, I(t) be the current at time t, T(t) be the temperature at time t, and Θ be the temperature at time t. base Basic model parameters;

[0050] In S32, the expression for the aging state encoding vector z is:

[0051] ;

[0052] Among them, Q acc For cumulative scalars, For encoder functions, For the set of real numbers, The dimension of the encoded vector;

[0053] In S33, the expression for the battery internal aging artifact map M is:

[0054] ;

[0055] Where DeconvNet(⋅) is a deconvolutional network, Let H be a real number matrix space, where H is the image height and W is the image width.

[0056] In S34, the expression for mapping is:

[0057] ;

[0058] in, For spatial SOH estimates, CNN health (⋅) represents a healthy state convolutional network, σ spatial Given spatial uncertainty, softplus(⋅) is the softplus activation function, CNN unc (⋅) represents an uncertain convolutional network.

[0059] Furthermore, S4 includes the following sub-steps:

[0060] S41. Construct a mathematical driving model on the vehicle side and an electrochemical mechanism model in the cloud.

[0061] S42. Based on the mathematical driving model and the electrochemical mechanism model, we conduct game bidding and consensus decision-making to obtain the game optimization objective;

[0062] S43. Based on the game optimization objective, output the consensus reached at SOH.

[0063] Furthermore, in S41, the electrochemical mechanism model The expression is:

[0064] ;

[0065] in, Here, u is the aging parameter, V(t) is the input variable, and T(t) is the terminal voltage and temperature.

[0066] In S42, the expression for the game optimization objective is:

[0067] ;

[0068] Where, q ∗ Let SOH be the consensus value, and Θ be the electrochemical model parameter. For vehicle-side model weights, Estimating the SOH data. The loss represents data consistency, and λ is the regularization coefficient. For electrochemical model simulation output, u obs For observation of operating conditions, V obs For the observed voltage, T obs To observe the temperature, For physical simulation loss, For mechanism SOH, Consensus-based loss mitigation;

[0069] S5 includes the following sub-steps:

[0070] S51. Construct a feature encoder to obtain the local representation vector;

[0071] S52. In the cloud, personalized computing units are generated using meta-learning objectives;

[0072] S53. Input the local representation vector into the personalized calculation unit to obtain the personalized SOH estimate;

[0073] In S51, the expression for the local representation vector h is:

[0074] ;

[0075] in, For the set of real numbers, Enc shared (⋅) represents the shared encoder, data represents multi-source time-series data, and d h As a representation dimension;

[0076] In S52, the expression for the meta-learning objective is:

[0077] ;

[0078] Where Φ is a primitive parameter, For the i-th meta-training task, For the meta-loss function, For initial personalized parameters, For the adapted personalized parameters, α is the inner learning rate, ∇ θp For the initial parameter θ p gradient, For training set loss, MLP personal (⋅) represents a personalized multilayer perceptron;

[0079] In S53, personalized SOH estimation The expression is:

[0080] ;

[0081] Where, θ p For personalized parameters.

[0082] Based on the above methods, this invention also proposes a lithium battery SOH estimation system based on multi-source data fusion, comprising:

[0083] Vehicle-side adaptive perception module: used to collect multi-source data;

[0084] Vehicle-side SOH estimation module: used to construct the vehicle-side local sensor and the cloud-based global memory network, process them, and output the SOH estimation value;

[0085] Cloud-based multimodal intelligent central module: used to embed a spatiotemporal decoder, trajectory inferencer, memory synthesizer, game fusion decision-maker, and output the consensus-reached SOH;

[0086] Cloud-based meta-learning and model management module: used to build personalized computing units to obtain personalized SOH estimates;

[0087] Update data sending module: It is used to input the SOH estimate, the consensus SOH and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, output the fused SOH estimate result, and send it to the vehicle system to realize vehicle-cloud collaborative update.

[0088] The beneficial effects of this invention are:

[0089] (1) Multimodal parallel processing: This invention is no longer a single model, but starts different professional computing channels for different data conditions, which greatly improves the data adaptability and robustness of the system.

[0090] (2) Game theory fusion decision: This invention creatively introduces the idea of ​​game theory, solves the conflict problem of multi-source estimation results, and ensures that the final result is optimal in both data and physical aspects.

[0091] (3) Dynamic task distribution: The system can dynamically allocate computing tasks between the vehicle end (lightweight, real-time) and the cloud (high precision, prediction) according to data quality and application scenario, so as to achieve optimal resource allocation.

[0092] (4) Personalization and self-evolution: This invention customizes a model for each battery through meta-learning and continuously improves the overall performance of the system over time through continuous learning. Attached Figure Description

[0093] Figure 1 The flowchart shows a lithium battery SOH estimation method based on multi-source data fusion.

[0094] Figure 2 This is a graph showing the SOH estimation results of lithium batteries based on multi-source data fusion and cloud collaboration;

[0095] Figure 3 A graph showing the validity of SOH estimation results after different processing channels and game fusion. Detailed Implementation

[0096] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0097] like Figure 1 As shown, this invention provides a lithium battery SOH estimation method based on multi-source data fusion, comprising the following steps:

[0098] S1. Construct vehicle-side local sensors and cloud-based global memory networks, process them, and output the estimated SOH value;

[0099] S2. Extract the current state snapshot and historical path fingerprint, process them, and filter future decay trajectories;

[0100] S3. Based on the virtual stress waveform, generate the aging state encoding vector and the battery internal aging artifact map, perform mapping, and obtain the confidence interval;

[0101] S4. Conduct game bidding and consensus decision-making to obtain the game optimization objective and output the consensus-reached SOH;

[0102] S5. Construct personalized calculation units to obtain personalized SOH estimates;

[0103] S6. Input the SOH estimate, the consensus SOH, and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, and output the fused SOH estimate result.

[0104] This invention proposes a deep architecture that fuses a cloud-based multimodal intelligent hub with vehicle-side adaptive perception nodes. The vehicle extracts multiple features (such as complete features, memory fragments, state snapshots, and path fingerprints) based on the charging scenario and uploads them to the cloud. The cloud processes different features through three parallel channels (spatiotemporal decoder, trajectory inferencer, and memory synthesizer), and uses an electrochemical-data dual-driven game theory fusion module to verify and optimize the results of each channel, outputting a consensus SOH value and its future prediction that combines accuracy and physical consistency. The cloud generates personalized vehicle-side models for each battery based on meta-learning and distributes them periodically. The system continuously learns and evolves its estimation capabilities. Specifically:

[0105] Vehicle-side multi-mode perception and lightweight feature extraction: The vehicle-side system has multiple operating modes. During full charging, traditional health features (such as capacity increment curve features) are extracted; during fragmented charging, a local sensor mode is activated to extract memory fragment features from a fixed reference SOC window. Simultaneously, a lightweight state-trajectory sensor continuously runs on the vehicle, extracting state snapshots and path fingerprints representing the current health status and recent aging trends from time-series data.

[0106] Cloud-based multimodal data fusion and intelligent computing: After receiving various features uploaded from the vehicle, the cloud-based central processing unit does not process them individually, but instead initiates a multimodal parallel processing channel.

[0107] Channel 1 (Spatiotemporal Decoder): For complete charging data, a spatiotemporal hierarchical mapping network is used to attempt to reconstruct the aging artifact map inside the battery and obtain the SOH estimate based on the physical spatial distribution.

[0108] Channel 2 (Trajectory Inferencer): Based on state snapshots and path fingerprints, it runs a state-trajectory co-inference network to predict the future decay trajectory and confidence interval of SOH.

[0109] Channel 3 (Memory Synthesizer): For memory fragments, it runs a global memory network to synthesize virtual complete features and estimate SOH by retrieving and fusing historical fragments.

[0110] The results from the three channels will be input into an electrochemical-data dual-driven game theory fusion module. This module treats the results from each channel as participants, performs physical validity checks on the results based on a simplified cloud-based electrochemical model, and finally outputs a consensus SOH value and its future prediction through a game theory optimization algorithm, which combines data accuracy and physical consistency.

[0111] Personalized Model Deployment and System Evolution: Utilizing a distributed representation learning and meta-learning framework, the cloud generates a unique and minimalist on-board SOH estimation model based on historical data for each battery. The cloud periodically deploys these personalized parameters, along with the latest game-theoretic fusion results, to the vehicle. When there is no network or low-power requirement, the vehicle can use this lightweight personalized model for rapid estimation. The cloud continuously receives global data and uses new data to continuously optimize the various sub-networks and meta-models within its multimodal hub, enabling the entire system's estimation capabilities to continuously evolve.

[0112] In this embodiment of the invention, S1 includes the following sub-steps:

[0113] S11. Construct a vehicle-side local perceptron and extract memory fragment feature vectors;

[0114] S12. Construct a cloud-based global memory network and determine memory retrieval weights;

[0115] S13. Determine the virtual complete features based on the memory fragment feature vector and memory retrieval weight;

[0116] S14. Feed the virtual complete features into the prediction network and output the SOH estimate.

[0117] Addressing the reality that users rarely complete a full charge-discharge cycle, this invention simulates the human ability to piece together the whole picture from fragmented memories. This invention designs an architecture combining a cloud-based global memory network and vehicle-side local sensors to directly estimate State of Harmony (SOH) from scattered, irregular charging segments.

[0118] In this embodiment of the invention, in S11, during each charge, regardless of the starting and ending SOC, a voltage rise segment is captured. For a fixed, narrow reference SOC window (e.g., always from SOC 45% to 55%), several simple features within this segment, such as average voltage, voltage variance, and temperature coefficient, are extracted; these are referred to as memory fragments. The memory fragments, along with the start and end SOCs of the current segment and the ambient temperature, are packaged and uploaded.

[0119] Memory fragment feature vector The expression is:

[0120] ;

[0121] in, The average voltage. For voltage variance, The average temperature. For the set of real numbers, The dimension of the feature vector;

[0122] In S12, each row of the memory matrix stores the ideal feature vector of a memory fragment. When a fragment packet uploaded from the vehicle is received, the network dynamically reads and weights multiple relevant memory fragments from the memory matrix based on the start and end SOC and ambient temperature, forming a virtual complete charging feature. The real memory fragments uploaded this time are selectively written into the memory matrix to achieve memory update and reinforcement.

[0123] The cloud acts as a constantly growing and evolving global memory, while the vehicle acts as a sensory probe that triggers memory retrieval. The vehicle does not need to perform complex calculations; it only uploads fragments. The cloud, through its powerful associative and synthetic capabilities, reconstructs the health status from these fragments. This solution is particularly suitable for the everyday plug-in charging scenarios of electric vehicles.

[0124] Memory read weight w i The expression is:

[0125] ;

[0126] Where β is the attention scaling factor, sim(⋅) is the similarity function, exp(⋅) is the exponential function, g(⋅) is the query vector mapping function, k is the query vector, and M... i For the memory matrix of the th Okay, M j For the memory matrix of the th Okay, SOC0 is the starting SOC for charging, SOC end The state of charge (SOC) is the end of charging; T is the ambient temperature.

[0127] In S13, the virtual complete features The expression is:

[0128] ;

[0129] Where W1 is the memory projection weight matrix, M i For the memory matrix of the th Line, W2 is the fragment projection weight matrix, and b is the bias vector;

[0130] In S14, the estimated value of SOH The expression is:

[0131] ;

[0132] MLP stands for Multilayer Perceptron.

[0133] In this embodiment of the invention, S2 includes the following sub-steps:

[0134] S21. Based on the SOH estimation value, extract state features and path features, which are used as the current state snapshot and historical path fingerprint, respectively.

[0135] S22. Using historical path fingerprints as conditions, generate several future SOH decay trajectories;

[0136] S23. Match the current state snapshot with the current point of all future SOH decay trajectories, calculate the matching probability, and filter future decay trajectories.

[0137] This invention creates a dual-stream extrapolation network, the core of which lies in distinguishing and fusing the battery's current health state and historical aging path, and utilizing the powerful computing capabilities of the cloud to simulate future trajectories under different usage strategies. This method can not only estimate the current state of health (SOH) but also predict the future degradation trajectory of SOH.

[0138] Two types of features are extracted: A) State features (such as current internal resistance and current constant current charging voltage plateau slope); B) Path features (such as the slope of the trend line indicating the shift of the main peak position of the capacity increment curve in each of the past 100 cycles). Features A and B are uploaded as a snapshot of the current state and a historical path fingerprint.

[0139] The core model, the state-trajectory co-inference network, is run. It consists of two coupled sub-networks:

[0140] Trajectory generator: Based on historical path fingerprints, it generates multiple possible future SOH decay trajectories (a probability distribution).

[0141] State verifier: Matches the current state snapshot with the current point of all possible trajectories generated by the trajectory generator, calculates the matching probability, and selects the most likely future trajectories.

[0142] The final output is not a single SOH value, but rather a most probable SOH range and its evolution over several future cycles. The system can also make reverse recommendations based on a cloud-stored strategy-trajectory library, such as how the trajectory would improve if a gentle charging strategy were adopted in the future.

[0143] The vehicle-side system provides high-quality state snapshots and path fingerprints. The cloud-based system handles energy-intensive trajectory extrapolation and strategy simulation. Based on received future trajectory predictions, the vehicle-side system can adjust its charging strategy (e.g., requesting a reduction in charging power), forming a closed loop of estimation, early warning, and control.

[0144] In this embodiment of the invention, in S22, the expression for generating several future SOH decay trajectories is as follows:

[0145] ;

[0146] Where z is a latent variable. (⋅) represents a multivariate normal distribution, μ(⋅) is the mean function, and σ 2 (⋅) is the variance function, x path For path features, Let Dec(⋅) be the k-th prediction at time t, where Dec(⋅) is the decoder function and T is the total prediction step size of the future decay trajectory.

[0147] In S23, the matching probability The expression is:

[0148] ;

[0149] Where, x state As a characteristic of the current state, g k Let τ be the prototype of the k-th target, and τ be the temperature parameter.

[0150] In this embodiment of the invention, S3 includes the following sub-steps:

[0151] S31. Collect voltage, current and temperature time-series data to generate virtual stress waveforms;

[0152] S32. Compress the virtual stress waveform with SOC and cumulative throughput into an aging state encoding vector;

[0153] S33. Construct a spatiotemporal hierarchical mapping network, use the spatial decoder of the spatiotemporal hierarchical mapping network to receive the aging state encoding vector, and use the deconvolutional neural network of the spatiotemporal hierarchical mapping network to reconstruct it to generate an aging artifact map inside the battery.

[0154] S34. Input the battery internal aging artifact image into the convolutional network, perform mapping, and obtain the confidence interval.

[0155] Abandoning the traditional approach of treating the battery as a whole for black-box estimation, battery aging is a physical and chemical process that evolves over time and is non-uniform in space. The core of this solution is to construct a spatiotemporally decoupled encoder-vehicle-cloud collaborative decoder, which aims to decode the (virtual) two-dimensional aging distribution inside the battery from one-dimensional time-series data from the vehicle, and then map it to the overall SOH (State of Health).

[0156] In this embodiment of the invention, in step S31, time-series data of voltage, current, and temperature are collected, but traditional health characteristics are not directly extracted. An ultra-lightweight local stress waveform generator is run. This model, based on real-time current and temperature, combined with a fundamental electrochemical response spectrum from the cloud specific to this battery model, generates a short sequence of virtual stress waveforms representing the stress experienced by each micro-region within the battery during this charge-discharge process. This waveform is physically driven, not data-driven. This virtual stress waveform, along with scalars such as SOC and cumulative throughput, is compressed into an aging state encoding vector and uploaded to the cloud.

[0157] The expression for the virtual stress waveform s(t) is:

[0158] ;

[0159] in, Let I(t) be the stress mapping function, I(t) be the current at time t, T(t) be the temperature at time t, and Θ be the temperature at time t. base Basic model parameters;

[0160] In S32, a large number of battery aging state encoding vectors and their final true aging distribution are stored.

[0161] The core model, a spatiotemporal hierarchical mapping network, is run. This network consists of two layers. The first layer (spatial decoder) receives the aging state encoding vector and reconstructs a low-resolution, two-dimensional aging artifact map of the battery's interior through a deconvolutional neural network. Pixel values ​​in different regions of this map represent different degrees of aging (such as lithium loss and SEI growth). The second layer (health mapper) inputs the aging artifact map into a convolutional network, directly mapping it to a high-precision overall SOH and its confidence interval.

[0162] The cloud periodically trains the spatial decoder and health mapper with new data, and sends the trained, simplified direct mapping from the encoded vector to the SOH to the vehicle. The vehicle uses this simplified model when there is no network, and uploads the encoded vector when there is a network to obtain more accurate cloud results with spatial distribution information.

[0163] The expression for the aging state encoding vector z is:

[0164] ;

[0165] Among them, Q acc For cumulative scalars, For encoder functions, For the set of real numbers, The dimension of the encoded vector;

[0166] In S33, the expression for the battery internal aging artifact map M is:

[0167] ;

[0168] Where DeconvNet(⋅) is a deconvolutional network, Let H be a real number matrix space, where H is the image height and W is the image width.

[0169] In S34, the expression for mapping is:

[0170] ;

[0171] in, For spatial SOH estimates, CNN health (⋅) represents a healthy state convolutional network, σ spatial Given spatial uncertainty, softplus(⋅) is the softplus activation function, CNN unc (⋅) represents an uncertain convolutional network.

[0172] In this embodiment of the invention, S4 includes the following sub-steps:

[0173] S41. Construct a mathematical driving model on the vehicle side and an electrochemical mechanism model in the cloud.

[0174] S42. Based on the mathematical driving model and the electrochemical mechanism model, we conduct game bidding and consensus decision-making to obtain the game optimization objective;

[0175] S43. Based on the game optimization objective, output the consensus reached at SOH.

[0176] The aging process inside the battery (such as lithium loss and SEI growth) is viewed as a game. This invention constructs a framework that coordinates a cloud-based electrochemical game simulator and an on-vehicle data referee, allowing the mechanistic model and the data model to compete and correct each other, ultimately reaching a consensus on the true state of energy loss (SOH).

[0177] On the vehicle side: Maintain a lightweight data-driven model (such as Extreme Learning Machine ELM), whose input is measurable data from the vehicle side.

[0178] In the cloud: maintain a parameterized electrochemical mechanism model (such as a simplified version of the P2D model), where key aging parameters (such as the volume fraction of positive and negative electrode active materials and the lithium-ion diffusion coefficient) are the game participants.

[0179] The vehicle uploads the current operating condition data. Based on the current parameters, the cloud simulator extrapolates the voltage, temperature, and other data that should be exhibited under this operating condition (simulated data). The vehicle's data model then outputs a SOH (State of Health) estimate (data estimate) based on the actual data.

[0180] The collaborative module compares simulated data with real data, and the differences between the mechanistic SOH (calculated from simulator parameters) and the data SOH. Through a game theory-based optimization algorithm, it adjusts the aging parameters of the cloud simulator to make the simulated data approximate real data. Simultaneously, the adjusted, more physically realistic mechanistic SOH is used to constrain and correct the output of the vehicle-side data model.

[0181] The iterative update rule expression is:

[0182] ;

[0183] Where η is the learning rate, ∇ Θ Let θ be the gradient operator, and PhysLoss(⋅) be the physical loss function. This is the observation dataset.

[0184] Output the consensus-reached State of Interest (SOH). Simultaneously, update the corrected vehicle-side data model parameters to the vehicle, giving it a higher starting point for the next estimation.

[0185] Therefore, this is not a simple weighted average, but a dynamic, theoretically based game process. The mechanistic model ensures that the estimation results conform to physical laws, while the data model compensates for the errors caused by the simplification of the mechanistic model. Together, they approximate reality in the game.

[0186] In this embodiment of the invention, in S41, the electrochemical mechanism model... The expression is:

[0187] ;

[0188] in, Here, u is the aging parameter, V(t) is the input variable, and T(t) is the terminal voltage and temperature.

[0189] In S42, the expression for the game optimization objective is:

[0190] ;

[0191] Where, q ∗ Let SOH be the consensus value, and Θ be the electrochemical model parameter. For vehicle-side model weights, Estimating the SOH data. The loss represents data consistency, and λ is the regularization coefficient. For electrochemical model simulation output, u obs For observation of operating conditions, V obs For the observed voltage, T obs To observe the temperature, For physical simulation loss, For mechanism SOH, Consensus-based loss mitigation;

[0192] S5 includes the following sub-steps:

[0193] S51. Construct a feature encoder to obtain the local representation vector;

[0194] S52. In the cloud, personalized computing units are generated using meta-learning objectives;

[0195] S53. Input the local representation vector into the personalized calculation unit to obtain the personalized SOH estimate;

[0196] It is assumed that the aging process of each battery is unique, but the aging of all batteries shares an implicit common characterization space. The goal of this invention is to learn this aging characterization space in the cloud and customize a minimal SOH calculation function for each battery on the vehicle side.

[0197] On the vehicle side: A fixed, generic feature encoder (a small neural network) runs, encoding multi-source time-series data into a low-dimensional local representation vector. This encoder is delivered from the cloud and shared by all batteries. A personalized, minimalist SOH calculation unit runs (possibly a linear regression or a miniature network with only a few layers of neurons). The parameters of this unit are battery-specific and customized and delivered from the cloud. During estimation, the data is processed by the encoder to obtain a local representation, which is then input into the personalized calculation unit to obtain the SOH.

[0198] Cloud-based: At its core is a meta-learning framework. Its goal is to quickly generate an optimal set of parameters for a personalized computing unit based on a small number of local representation vectors of a battery and its corresponding historical SOH ground truth values. To achieve this, the cloud trains a meta-model using massive amounts of battery data. This meta-model learns the meta-knowledge of how to customize the SOH estimation model based on the representation of a battery.

[0199] When a new vehicle is initialized, a universal encoder and an initial personalized computing unit based on the average parameters of similar batteries are downloaded from the cloud. After the vehicle has been running for a period of time, the accumulated local representation vectors and historical estimates are uploaded. Based on this data, the cloud-based metamodel quickly generates a new set of personalized computing unit parameters that are more closely aligned with the battery and sends them to the vehicle for updates. Subsequently, the vehicle achieves high-precision personalized estimation with extremely low computational cost, while the cloud only needs to periodically fine-tune the metamodel and universal encoder with new data.

[0200] In S51, the expression for the local representation vector h is:

[0201] ;

[0202] in, For the set of real numbers, Enc shared (⋅) represents the shared encoder, data represents multi-source time-series data, and d h As a representation dimension;

[0203] In S52, the expression for the meta-learning objective is:

[0204] ;

[0205] Where Φ is a primitive parameter, For the i-th meta-training task, For the meta-loss function, For initial personalized parameters, For the adapted personalized parameters, α is the inner learning rate, ∇ θp For the initial parameter θ p gradient, For training set loss, MLP personal (⋅) represents a personalized multilayer perceptron;

[0206] In S53, personalized SOH estimation The expression is:

[0207] ;

[0208] Where, θ p For personalized parameters.

[0209] Based on the above methods, this invention also proposes a lithium battery SOH estimation system based on multi-source data fusion, comprising:

[0210] Vehicle-side adaptive perception module: used to collect multi-source data;

[0211] Vehicle-side SOH estimation module: used to construct the vehicle-side local sensor and the cloud-based global memory network, process them, and output the SOH estimation value;

[0212] Cloud-based multimodal intelligent central module: used to embed a spatiotemporal decoder, trajectory inferencer, memory synthesizer, game fusion decision-maker, and output the consensus-reached SOH;

[0213] Cloud-based meta-learning and model management module: used to build personalized computing units to obtain personalized SOH estimates;

[0214] Update data sending module: It is used to input the SOH estimate, the consensus SOH and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, output the fused SOH estimate result, and send it to the vehicle system to realize vehicle-cloud collaborative update.

[0215] The system of the present invention also includes:

[0216] Vehicle communication module: responsible for data uploading and command model downlink.

[0217] Data acquisition module: Used to receive vehicle-side battery data uploaded by electric vehicles.

[0218] Data preprocessing module: This module performs time alignment, anomaly removal, normalization, and feature standardization on the received multi-source data before sending it to the data classification module.

[0219] Data classification and scheduling module: It is used to classify input data according to data type and perception mode, and distribute different types of data to the corresponding multimodal intelligent computing channels.

[0220] In this embodiment of the invention, the specific performance testing method is as follows:

[0221] (1) Select 50 power lithium batteries (ternary lithium batteries, nominal capacity 20Ah, nominal voltage 3.7V, rated charging voltage 4.2V, discharge termination voltage 2.5V) of the same model and batch to ensure initial consistency (initial capacity deviation ≤ ±1%, initial internal resistance deviation ≤ ±5%).

[0222] (2) Sample aging treatment:

[0223] Gradient aging treatment was performed on 50 battery samples, simulating the natural aging process of batteries through constant current-constant voltage (CC-CV) cyclic charge-discharge. Specific parameters:

[0224] Charging: Charge at a constant current of 0.5C to 4.2V, then switch to constant voltage charging until the current is ≤0.05C;

[0225] Discharge: 0.5C constant current discharge to 2.5V;

[0226] Aging gradient: After every 100 cycles, the current capacity is tested, and finally 50 samples with SOH values ​​distributed between 50% and 100% are obtained (sample number 1-50, the SOH value increases with the number and shows a gradient distribution, number 1 corresponds to SOH≈50%, number 50 corresponds to SOH≈100%).

[0227] (3) Specific test steps:

[0228] Step 1: True SOH value calibration (all samples)

[0229] For 50 aged battery samples, the true SOH value was obtained using the full charge-discharge capacity calibration method. The steps are as follows:

[0230] The battery sample was placed in a constant temperature environment of 25°C for 2 hours;

[0231] Use a reference SOH calibrator to charge at a constant current of 0.5C to 4.2V, then charge at a constant voltage until the current is ≤0.05C, and let stand for 1 hour;

[0232] Discharge at a constant current of 0.5C to 2.5V and record the total discharge capacity (i.e., the current actual capacity C_actual).

[0233] Calculate the true SOH value: SOH_true = (C_actual / C_rated) × 100%, where C_rated = 20Ah (rated capacity);

[0234] Repeat the above steps 3 times, and take the average value of SOH_true as the final true SOH value of the sample, and record it in the sample ledger (sample 1 - SOH_true ≈ 50%, sample 50 - SOH_true ≈ 100%).

[0235] Step 2: Test Scenario Simulation and Data Acquisition

[0236] For each sample, two real-world application scenarios are simulated, multi-source data is collected, and the system of this invention is triggered to run:

[0237] 1) Complete charging scenario:

[0238] The battery was discharged at a constant current of 0.5C to 2.5V and left to stand for 1 hour.

[0239] Charge at a constant current of 0.5C to 4.2V, then charge at a constant voltage until the current is ≤0.05C;

[0240] The vehicle-side adaptive perception module extracts traditional health features (capacity increment curve features) and status snapshots + path fingerprints, and uploads them to the cloud through the vehicle-cloud communication module.

[0241] 2) Fragmented charging scenarios:

[0242] Discharge the battery to three initial SOC points of 20%, 40%, and 60%, and let it stand for 1 hour at each point.

[0243] Each initial SOC point is charged at a constant current of 0.5C for 15 minutes (simulating short-term charging), then charging is stopped and the device is left to stand for 30 minutes.

[0244] The vehicle-side adaptive perception module activates the local sensor mode, extracts memory fragment features from a fixed reference SOC window (20%-35%, 40%-55%, 60%-75%), and uploads them synchronously to the cloud.

[0245] Step 3: System Estimation and Data Recording

[0246] After receiving the feature data uploaded from the vehicle, the cloud-based multimodal intelligent central module automatically initiates three-channel parallel processing:

[0247] Channel 1 (Spatiotemporal Decoder): Processes complete charging data and outputs the estimated SOH value SOH_1;

[0248] Channel 2 (Trajectory Inferencer): Processes state snapshots and path fingerprints, outputs SOH estimate SOH_2 and future decay trajectory;

[0249] Channel 3 (Memory Synthesizer): Processes memory fragments and outputs the estimated SOH value SOH_3;

[0250] Game-theoretic fusion decision: Input SOH_1, SOH_2, and SOH_3 into the electrochemical-data dual-driven game-theoretic fusion module, perform physical rationality verification based on a simplified cloud electrochemical model (considering key electrochemical parameters such as lithium-ion diffusion coefficient and polarization resistance), and output the consensus SOH value SOH_consensus through the Nash equilibrium optimization algorithm.

[0251] Record the SOH_1, SOH_2, SOH_3, and SOH_consensus of each sample, as well as the SOH_true calibrated in step 1.

[0252] Step 4: Test repeatability verification

[0253] For each sample, repeat steps 2-3 a total of 5 times, and take the average of SOH_1, SOH_2, SOH_3, and SOH_consensus as the final estimation result for that sample to ensure the repeatability of the test results (coefficient of variation ≤ ±1%).

[0254] Figure 2This is a scatter plot showing the fit between the estimated and actual SOH values ​​of lithium batteries based on multi-source data fusion and cloud collaboration. The horizontal axis represents the actual SOH value of the lithium batteries (unit: %), and the vertical axis represents the predicted SOH value of the method of this invention (unit: %). The data points represent the correspondence between the actual and predicted SOH values ​​of 50 battery samples (a total of 50 data points). This plot visually demonstrates the prediction accuracy of the lithium battery SOH estimation method based on multi-source data fusion and cloud collaboration. The coefficient of determination R² = 0.98, the absolute error between the predicted and actual SOH values ​​for all samples is ≤ ±2%, the mean absolute error (MAE) is 0.8%, and the root mean square error (RMSE) is 1.2%, indicating that the predicted values ​​and actual values ​​have a very high degree of fit, verifying the high accuracy characteristics of the estimation method of this invention.

[0255] Figure 3 This is a line graph comparing the SOH estimation results of different processing channels and after game-theoretic fusion. The horizontal axis represents the battery sample number (1-50, covering lithium batteries with different aging levels), and the vertical axis represents the estimated SOH value (range 0.6-1.0). The graph contains five curves, corresponding to the SOH estimated value SOH_1 of channel 1 (spatiotemporal decoder), the SOH estimated value SOH_2 of channel 2 (trajectory inferencer), the SOH estimated value SOH_3 of channel 3 (memory synthesizer), the consensus SOH value SOH_consensus after game-theoretic fusion, and the actual SOH value SOH_true (baseline) of the lithium battery. This graph clearly presents the estimation performance of each multimodal parallel processing channel, and the high consistency between the consensus SOH value and the actual value after processing by the electrochemical-data dual-driven game-theoretic fusion module, demonstrating the superiority of the game-theoretic fusion decision-making mechanism in improving the reliability and accuracy of the estimation results.

[0256] In summary, the cloud-based multimodal intelligent hub + vehicle-side adaptive perception node fusion architecture proposed in this invention has significant advantages in achieving comprehensiveness, foresight, high accuracy and high reliability in lithium battery SOH estimation, as well as system self-adaptation and self-evolution. It is suitable for complex dynamic scenarios with high reliability and long-term prediction requirements for battery health management.

[0257] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A lithium battery SOH estimation method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Construct vehicle-side local sensors and cloud-based global memory networks, process them, and output the estimated SOH value; S2. Extract the current state snapshot and historical path fingerprint, process them, and filter future decay trajectories; S3. Based on the virtual stress waveform, generate the aging state encoding vector and the battery internal aging artifact map, perform mapping, and obtain the confidence interval; S4. Conduct game bidding and consensus decision-making to obtain the game optimization objective and output the consensus-reached SOH; S5. Construct personalized calculation units to obtain personalized SOH estimates; S6. Input the SOH estimate, the consensus SOH, and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, and output the fused SOH estimate result.

2. The lithium battery SOH estimation method based on multi-source data fusion according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Construct a vehicle-side local perceptron and extract memory fragment feature vectors; S12. Construct a cloud-based global memory network and determine memory retrieval weights; S13. Determine the virtual complete features based on the memory fragment feature vector and memory retrieval weight; S14. Feed the virtual complete features into the prediction network and output the SOH estimate.

3. The lithium battery SOH estimation method based on multi-source data fusion according to claim 2, characterized in that, In S11, the memory fragment feature vector The expression is: ； in, Average voltage For voltage variance, The average temperature. For the set of real numbers, The dimension of the feature vector; In S12, the memory read weight w i The expression is: ； Where β is the attention scaling factor, sim(⋅) is the similarity function, exp(⋅) is the exponential function, g(⋅) is the query vector mapping function, k is the query vector, and M... i For the memory matrix of the first Okay, M j For the memory matrix of the first Okay, SOC0 is the starting SOC for charging. end The state of charge (SOC) is the end of charging; T is the ambient temperature. In S13, the virtual complete features The expression is: ； Where W1 is the memory projection weight matrix, M i For the memory matrix of the first Line, W2 is the fragment projection weight matrix, and b is the bias vector; In S14, the estimated value of SOH The expression is: ； MLP stands for Multilayer Perceptron.

4. The lithium battery SOH estimation method based on multi-source data fusion according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Based on the SOH estimation value, extract state features and path features, which are used as the current state snapshot and historical path fingerprint, respectively. S22. Using historical path fingerprints as conditions, generate several future SOH decay trajectories; S23. Match the current state snapshot with the current point of all future SOH decay trajectories, calculate the matching probability, and filter future decay trajectories.

5. The lithium battery SOH estimation method based on multi-source data fusion according to claim 4, characterized in that, In step S22, the expression for generating several future SOH decay trajectories is as follows: ； Where z is a latent variable. (⋅) represents a multivariate normal distribution, μ(⋅) is the mean function, and σ 2 (⋅) is the variance function, x path For path features, Let Dec(⋅) be the k-th prediction at time t, where Dec(⋅) is the decoder function and T is the total prediction step size of the future decay trajectory. In S23, the matching probability The expression is: ； Where, x state As a characteristic of the current state, g k Let τ be the prototype of the k-th target, and τ be the temperature parameter.

6. The lithium battery SOH estimation method based on multi-source data fusion according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Collect voltage, current and temperature time-series data to generate virtual stress waveforms; S32. Compress the virtual stress waveform with SOC and cumulative throughput into an aging state encoding vector; S33. Construct a spatiotemporal hierarchical mapping network, use the spatial decoder of the spatiotemporal hierarchical mapping network to receive the aging state encoding vector, and use the deconvolutional neural network of the spatiotemporal hierarchical mapping network to reconstruct it to generate an aging artifact map inside the battery. S34. Input the battery internal aging artifact image into the convolutional network, perform mapping, and obtain the confidence interval.

7. The lithium battery SOH estimation method based on multi-source data fusion according to claim 6, characterized in that, In S31, the expression for the virtual stress waveform s(t) is: ； in, Let I(t) be the stress mapping function, I(t) be the current at time t, T(t) be the temperature at time t, and Θ be the temperature at time t. base Basic model parameters; In step S32, the expression for the aging state encoding vector z is: ； Among them, Q acc For cumulative scalars, For encoder functions, For the set of real numbers, The dimension of the encoded vector; In S33, the expression for the battery internal aging artifact map M is: ； Where DeconvNet(⋅) is a deconvolutional network, Let H be a real number matrix space, where H is the image height and W is the image width. In S34, the expression for mapping is: ； in, For spatial SOH estimates, CNN health (⋅) represents a healthy state convolutional network, σ spatial Given spatial uncertainty, softplus(⋅) is the softplus activation function, CNN unc (⋅) represents an uncertain convolutional network.

8. The lithium battery SOH estimation method based on multi-source data fusion according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Construct a mathematical driving model on the vehicle side and an electrochemical mechanism model in the cloud. S42. Based on the mathematical driving model and the electrochemical mechanism model, we conduct game bidding and consensus decision-making to obtain the game optimization objective; S43. Based on the game optimization objective, output the consensus reached at SOH.

9. The lithium battery SOH estimation method based on multi-source data fusion according to claim 8, characterized in that, In S41, the electrochemical mechanism model The expression is: ； in, Here, u is the aging parameter, V(t) is the input variable, and T(t) is the terminal voltage and temperature. In S42, the expression for the game optimization objective is: ； Where, q ∗ Let SOH be the consensus value, and Θ be the electrochemical model parameter. For vehicle-side model weights, Estimating the SOH data. The loss represents data consistency, and λ is the regularization coefficient. For electrochemical model simulation output, u obs For observation of operating conditions, V obs For the observed terminal voltage, T obs To observe the temperature, For physical simulation loss, For mechanism SOH, Consensus-based loss mitigation; S5 includes the following sub-steps: S51. Construct a feature encoder to obtain the local representation vector; S52. In the cloud, personalized computing units are generated using meta-learning objectives; S53. Input the local representation vector into the personalized calculation unit to obtain the personalized SOH estimate; In S51, the expression for the local representation vector h is: ； in, For the set of real numbers, Enc shared (⋅) represents the shared encoder, data represents multi-source time-series data, and d h As a representation dimension; In S52, the expression for the meta-learning objective is: ； Where Φ is a primitive parameter, For the i-th meta-training task, For the meta-loss function, For initial personalized parameters, For the adapted personalized parameters, α is the inner learning rate, ∇ θp For the initial parameter θ p gradient, For training set loss, MLP personal (⋅) represents a personalized multilayer perceptron; In S53, personalized SOH estimation The expression is: ； Where, θ p For personalized parameters.

10. A lithium battery SOH estimation system based on multi-source data fusion, characterized in that, include: Vehicle-side adaptive perception module: used to collect multi-source data; Vehicle-side SOH estimation module: used to construct the vehicle-side local sensor and the cloud-based global memory network, process them, and output the SOH estimation value; Cloud-based multimodal intelligent central module: used to embed a spatiotemporal decoder, trajectory inferencer, memory synthesizer, game fusion decision-maker, and output the consensus-reached SOH; Cloud-based meta-learning and model management module: used to build personalized computing units to obtain personalized SOH estimates; Update data sending module: It is used to input the SOH estimate, the consensus SOH and the personalized SOH estimate as three channels into the electrochemical-data dual-drive game fusion module, and fuse them according to the future decay trajectory and confidence interval, output the fused SOH estimate result, and send it to the vehicle system to realize vehicle-cloud collaborative update.

Images