Flow battery state of health estimation method and system based on physical information and cross-modal collaboration

Abstract

The application provides a liquid flow battery SOH estimation method and system based on physical information and cross-modal collaboration, relates to the technical field of all-vanadium liquid flow batteries, and comprises the following steps: acquiring optical spectrum signals and electrical signals in the operation process of a liquid flow battery; performing inversion on the optical spectrum signals according to a physical mechanism constraint to obtain intrinsic characteristics; wherein the intrinsic characteristics are estimated ion component concentrations of electrolyte of the battery; decomposing the electrical signals according to the intrinsic characteristics through electrochemical principles to obtain extrinsic characteristics; and performing cross-modal residual error fusion and mapping output on the intrinsic characteristics and the extrinsic characteristics to obtain an SOH estimation value of the liquid flow battery. Through the technical solution, the accuracy, real-time performance and robustness of SOH estimation of the liquid flow battery can be improved.

Description

Technical Field

[0001] This invention relates to the field of vanadium redox flow battery technology, and more specifically, to a method and system for estimating the state of harm (SOH) of a flow battery based on physical information and cross-modal collaboration. Background Technology

[0002] Vanadium redox flow batteries (VRFBs), as an electrochemical energy storage technology, possess significant advantages such as independent power and capacity design, high system safety, long cycle life, and recyclable electrolyte. These characteristics make them valuable for large-scale, long-term energy storage applications. In the operation and management of VRFBs, accurate assessment of their State of Health (SOH) plays a crucial supporting role in optimizing power system dispatch, developing predictive maintenance strategies, and extending the system's overall lifespan.

[0003] Currently, existing SOH estimation schemes can be mainly divided into the following two categories: 1. Physical Model-Based Approaches: These methods primarily use equivalent circuit models (ECMs) or electrochemical mechanism models (such as P2D models) to describe the physicochemical processes inside the battery. The basic logic is to use observer algorithms (such as extended Kalman filtering) to identify model parameters (such as internal resistance and diffusion coefficient) online, and then calculate the state of charge (SOH) based on the mapping relationship between these parameters and capacity.

[0004] 1.1. For example, CN110261779A discloses a method for estimating the state of harmonics (SOH) of ternary lithium batteries based on an improved Thevenin model. This method updates the model parameters in real time using recursive least squares. 1.2. CN121348112A discloses an improved equivalent circuit model, a multi-heat source thermal network model, and a degradation evolution model that integrates solid-phase and liquid-phase diffusion mechanisms for state estimation. It identifies model parameters through offline experiments and an online dual-filter framework, and estimates the battery's state of charge (SOC), internal temperature (SOT), and state of health (SOH) in real time. 2. Data-driven approach: This type of approach does not rely on complex internal mechanisms. It collects electrical signals (voltage, current, temperature) during the battery charging and discharging process, extracts feature parameters (such as incremental capacity IC, voltage relaxation curve, etc.), and uses machine learning models (such as Long Short-Term Memory Network LSTM, Gaussian Process Regression GPR, etc.) to establish a nonlinear mapping from features to SOH.

[0005] 2.1. CN120334752A trains multiple machine learning models by capturing some features near the peak values ​​of each IC to cope with changes in charge and discharge depth. However, the applicable scenarios are limited.

[0006] 2.2. Existing research uses joint features (PDF, IC, DV) combined with Bayesian neural networks and considers the inconsistency of individual cells in secondary batteries used in energy storage power stations. It introduces a distributed calibration step to optimize the prediction results, but it needs to cover at least 20% of the SOC range.

[0007] 2.3. CN121522490A discloses a battery SOH calculation method based on incremental capacity and differential voltage curves, which still requires an electrical signal of 100s-600s and cannot effectively separate internal and external states.

[0008] Although the above methods have made significant progress in fields such as lithium-ion batteries, their direct application to all-vanadium redox flow batteries has the following limitations: 1. Operating Condition Coupling and External Interference: Unlike lithium batteries, flow batteries dynamically adjust pump flow rates during actual operation to maintain optimal energy conversion efficiency. This causes complex concentration polarization fluctuations. This "voltage noise" generated by external operating conditions often masks subtle aging characteristics, leading to a significant decrease in the accuracy of traditional voltage-based methods.

[0009] 2. Stringent Sampling Conditions: Most existing solutions assume that the battery can undergo a complete and stable constant-current charge-discharge cycle. However, industrial-grade flow battery systems have huge capacities, and in actual operation, they are often subjected to fragmented shallow charge and discharge states. Within an extremely narrow voltage window, traditional electrical characteristics often cannot be fully revealed.

[0010] 3. Inefficient use of multimodal signals: Current technologies mainly rely on a single electrical signal mode. Although the optical properties (spectrum) of electrolytes can intuitively reflect the internal chemical state, due to their strong nonlinearity and the difficulty in obtaining concentration labels online in real time, there is still a lack of an efficient fusion framework that can physically decouple the internal optical factors from the external electrical factors for health diagnosis. Summary of the Invention

[0011] In view of this, the present invention proposes a method and system for estimating the state of harmonics (SOH) of flow batteries based on physical information and cross-modal collaboration, in order to solve the problems existing in the prior art.

[0012] To achieve the above objectives, this invention proposes a method and system for estimating the state of harmonics (SOH) of flow batteries based on physical information and cross-modal collaboration, comprising: Acquire optical and electrical signals during the operation of the flow battery; The optical spectral signal is inverted based on physical mechanism constraints to obtain intrinsic characteristics; wherein the intrinsic characteristics are the estimated ionic component concentrations of the battery electrolyte. Based on the intrinsic characteristics, the electrical signal is decomposed using electrochemical principles to obtain the extrinsic characteristics; The SOH estimate of the flow battery is obtained by performing cross-modal residual fusion and mapping output on the intrinsic and extrinsic features.

[0013] Optionally, after acquiring the optical spectral signal and the electrical signal, the process may also include: The optical spectral signal and the electrical signal are preprocessed, wherein the preprocessing includes normalization and consistency check. The optical spectral signals are ultraviolet-visible spectral signals for the positive and negative electrodes of the flow battery, and the bands of the spectral signals are characteristic bands related to the ion concentration of the flow battery; the electrical signals include terminal voltage signals, current signals, incremental capacity curve data, and differential voltage curve data.

[0014] Optionally, the process of inverting the optical spectral signal includes: A constrained bond-value pair grid is constructed based on Beer-Lambert's law. The bond-value pair grid is used to store relational pairs with different spectral signals as bonds and corresponding ion component concentrations as values. During the construction of the bond-value pair grid, the mass conservation relationship of different ion component concentrations and the key to chemical equilibrium are used as constraints. The optical spectral signal is used as a query to perform matching and inversion in a key-value pair grid to obtain intrinsic features.

[0015] Optionally, the matching inversion process includes: The optical spectral signal is matched with the spectral signals corresponding to different bonds to obtain the candidate spectral similarity. The candidate spectral similarity is transformed using a fitting function to obtain the attention weight; the concentrations of the ion components corresponding to the attention weights are weighted and calculated to obtain the estimated ion component concentrations.

[0016] Optionally, the process of decomposing the electrical signal includes: The equilibrium potential is obtained by calculating the estimated ion component concentration using the Nernst equation; wherein the Nernst coefficient in the Nernst equation is a learnable parameter. The ohmic polarization component is obtained by calculating the current signal in the electrical signal through linear fitting; the linear operator in the linear fitting method is related to the battery internal resistance parameters and is constrained to be positive, and the battery internal resistance parameters are learnable parameters. The basic polarization intensity is calculated based on the estimated ion component concentration and current signal; the flow rate suppression factor is calculated based on the attenuation coefficient and the electrolyte flow rate of the flow cell; the concentration polarization component is calculated based on the basic polarization intensity and the flow rate suppression factor; wherein the attenuation coefficient is used to characterize the nonlinear attenuation effect of flow rate on polarization, and the attenuation coefficient is a learnable parameter.

[0017] Optionally, the process of performing cross-modal residual fusion on the intrinsic and extrinsic features includes: The intrinsic and extrinsic features are mapped to obtain optical embedding features and electrochemical embedding features; the optical embedding features and electrochemical embedding features are then spliced ​​together to obtain joint features; The optical embedding features and electrochemical embedding features are processed with the joint features through linear layer and layer normalization to obtain fused features; the fused features are then mapped to obtain the SOH estimate of the flow battery.

[0018] Optionally, learnable parameters are set in the process of optical spectral signal inversion, electrical signal decomposition, cross-modal residual fusion, and mapping output. The learnable parameters are optimized according to a comprehensive objective function. The comprehensive objective function includes a main task loss and an auxiliary task loss to characterize the deviation of the SOH estimate. The auxiliary task loss includes spectral reconstruction constraints, ion dynamics consistency constraints, and voltage decomposition constraints.

[0019] On the other hand, the present invention also provides a flow battery SOH estimation system based on physical information and cross-modal collaboration for performing the above-described method.

[0020] On the other hand, the present invention also provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a flow battery SOH estimation method based on physical information and cross-modal collaboration.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the shortcomings of existing flow battery SOH estimation methods, such as low accuracy, poor robustness, and reliance on offline calibration, through core technical features such as multimodal fusion, physical information guidance, and decoupling of internal and external factors. The introduction of physical mechanisms enhances the model's generalization ability under various operating conditions of shallow charging and discharging, achieving accurate, real-time, and robust SOH estimation. This provides data support for the refined operation and maintenance of battery systems, effectively guides maintenance decisions, and reduces operation and maintenance costs. Attached Figure Description

[0022] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the flow battery SOH estimation method based on physical information and cross-modal collaboration in an embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] This invention proposes a flow battery SOH estimation system based on physical information and cross-modal collaboration. The overall technical solution of this invention adopts a modular architecture, with each module working collaboratively to form a complete data processing and estimation chain. The functional boundaries and innovative features of each module are clearly defined. Specifically, it includes the following four modules, and the functions, technical features, and logical relationships of each module are as follows: 1. Multimodal operation data preprocessing module: This module, serving as the front-end module of the entire method, is responsible for integrating multi-sensor data and providing standardized input data for the entire estimation framework. Its core function is to process multi-source observation signals of flow batteries under complex operating conditions.

[0025] The front-end module is connected to a spectral sensor and an electrical sensor. The spectral sensor is used to acquire the ultraviolet-visible (UV-Vis) spectral signal of the electrolyte in the vanadium redox flow battery. The electrical sensor is used to acquire the terminal voltage signal, current signal, and electrolyte flow rate signal of the vanadium redox flow battery.

[0026] This module aligns the data based on time and removes outliers caused by equipment failures. It then extracts the differential voltage and differential capacitance characteristics of the electrical signals as supplementary signals, laying the foundation for subsequent cross-modal fusion.

[0027] 2. Physical Information-Guided Spectral Attention (PISA) Intrinsic Factor Extraction Module: This module constructs a physical benchmark anchor for the internal chemical state of the battery for the entire estimation framework. Its core function is to guide the model using the spectral formation mechanism (Beer-Lambert's law) to achieve self-supervised differentiable inversion of ion component concentration from the spectral signal. It does not require expensive concentration labels and solves the technical problem that traditional spectral analysis relies on offline calibration and cannot obtain the internal chemical state in real time, providing a reliable internal state benchmark for subsequent decoupling from external factors.

[0028] The key technical features of this module are: constructing a bond-value pair grid under physical constraints, embedding Beer-Lambert's law and the chemical equilibrium relationship of the flow battery electrolyte into the model structure to ensure the physical consistency of concentration inversion; proposing the spectral similarity kernel (SSK) as a spectral matching degree calculation function to enhance the model's robustness to measurement noise and baseline drift; and achieving a fully differentiable design, which can be jointly optimized with subsequent modules through backpropagation, eliminating the need for real ion concentration calibration tests and reducing deployment costs.

[0029] Alternatives to this module: Spectral Similarity Kernel (SSK) can use similar correlation measures such as cosine similarity and root mean square error to replace Pearson correlation coefficient, adapting to different deployment scenarios.

[0030] 3. Electrochemically Driven Voltage Decomposition (EDDL) Exogenous Decoupling Module: This module effectively separates the influence of external operating conditions from internal aging drift in electrical signals. Its core function is to rely on electrochemical a priori mechanisms and combine the ion concentration benchmark provided by the PISA module to explicitly decompose the complex terminal voltage signal into components with clear physical meaning. This achieves effective decoupling of voltage signal interference from external operating conditions (flow rate, current), solves the technical problem that traditional methods cannot distinguish whether voltage fluctuations originate from internal aging or external polarization, and highlights the inherent aging characteristics of the battery.

[0031] The key technical features of this module are as follows: It constructs an equilibrium potential calculation branch based on the Nernst equation, directly linking ion concentration with the equilibrium potential to ensure that the reference voltage is anchored to the internal chemical state of the battery in real time; it adopts a condition-aware polarization component decomposition strategy to model ohmic polarization and concentration polarization separately, embedding the influence mechanism of external operating conditions on polarization to ensure the physical rationality of voltage decomposition; and it provides physical constraints for subsequent joint training through a voltage decomposition verification mechanism, thereby improving the reliability of the model.

[0032] Alternatives for this module: The physical equations used in this module are not limited to the Nernst equation, the Ohm polarization equation, and the concentration polarization equation. Other equivalent physical equations that conform to the electrochemical mechanism of flow batteries can be used for alternative modeling.

[0033] 4. Cross-modal residual fusion and physical consistency mapping module This module achieves deep fusion of internal and external features and accurate mapping of SOH. Its core function is to deeply interact and fuse the internal features (ion concentration) extracted by the PISA module with the external features (voltage components and electrical signal features) decomposed by the EDDL module. Through multi-task joint training with physical consistency constraints, the fused features are mapped to the final SOH estimate, solving the technical problems of insufficient single-modal feature information and insufficient fusion of multi-modal features, and ensuring the accuracy and robustness of SOH estimation.

[0034] The key technical features of this module are: adopting a cross-modal residual fusion mechanism to achieve mutual correction and deep interaction of features from two modes, avoiding feature redundancy and information loss; introducing a physical consistency multi-task joint training strategy to transform physical constraints such as spectral reconstruction, ion dynamics consistency, and voltage decomposition into auxiliary losses, guiding model parameters to converge to values ​​that conform to physical laws, and avoiding error accumulation in intermediate feature extraction stages; and achieving accurate mapping of fused features to SOH, ensuring estimation accuracy even under extreme conditions.

[0035] Alternatives for this module: The auxiliary loss weights in the comprehensive objective function can be adjusted according to actual needs to adapt to application scenarios with different signal quality; the fully connected layer of the feature mapping can be replaced by lightweight neural networks such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) to enhance the ability to capture temporal features and adapt to scenarios where temporal features are more prominent.

[0036] On the other hand, the present invention also provides a physical information-guided cross-modal collaborative method for estimating the state of health (SOH) of an all-vanadium redox flow battery, comprising the following steps: Step 1. Collect multimodal data during the operation of the vanadium redox flow battery, including optical spectral signals and electrical signals; Step 2. Based on physical mechanism constraints, the optical spectral signal is inverted to extract intrinsic features reflecting the internal chemical state of the battery, wherein the intrinsic features are the concentrations of ionic components in the battery electrolyte; Step 3. Based on the concentration of the ion components and the electrochemical principle, the terminal voltage in the electrical signal is decomposed into multiple voltage components with physical meaning to achieve decoupling between external operating condition interference and internal battery aging drift, and to obtain the decoupled external characteristics. Step 4. The intrinsic features and the decoupled extrinsic features are fused through cross-modal residual processing to output the SOH estimate of the all-vanadium redox flow battery.

[0037] Furthermore, the inversion processing of the optical spectral signal based on physical mechanism constraints in step 2 specifically involves: constructing a physical constraint bond-value pair grid based on Beer-Lambert's law, and considering the mass conservation relationship and chemical equilibrium relationship of ions in the electrolyte of the vanadium redox flow battery during the construction process, and inverting the optical spectral signal based on the physical constraint bond-value pair grid.

[0038] Furthermore, when inverting the optical spectral signal based on the physical constraint key-value pair grid, a differentiable inversion is performed using spectral similarity matching. The matching index of the spectral similarity kernel is Pearson correlation coefficient, cosine similarity, or root mean square error similarity.

[0039] Furthermore, the physically significant voltage components mentioned in step 3 specifically include components such as equilibrium potential (OCV), ohmic polarization, and concentration polarization.

[0040] Furthermore, the equilibrium potential is calculated based on the Nernst equation, and the Nernst coefficient, battery internal resistance related parameters, and flow rate suppression factor in concentration polarization calculation are all learnable parameters of the model.

[0041] Furthermore, the calculation logic for the concentration polarization component incorporates the nonlinear coupling relationship between current and electrolyte flow rate. Calculate using the following formula: ;in, λ represents the basic polarization intensity under concentration-current coupling, λ is a learnable parameter, and Qp is the electrolyte flow rate.

[0042] Furthermore, it also includes a model training step: a multi-task joint training method with physical consistency constraints is used to perform parameter optimization training on the SOH-estimated model.

[0043] Furthermore, the objective function of the multi-task joint training with physical consistency constraints includes the SOH prediction main task loss and at least three auxiliary loss terms; the auxiliary loss terms are specifically spectral reconstruction loss, ion dynamics consistency loss, and voltage decomposition conservation loss; wherein, the ion dynamics consistency loss is the correlation loss of the rate of change of ion concentration with respect to the charge, and the voltage decomposition conservation loss is the correlation loss of the measured terminal voltage with the sum of the decomposed voltage components.

[0044] The above technical solution is described in detail: This invention addresses the technical problems of existing flow battery SOH estimation methods, which rely on a single electrical signal, struggle to handle random shallow charge / discharge cycles and dynamic flow rate adjustments, and fail to effectively separate internal aging from external interference. It proposes a physically-guided, integrated internal and external factor-based SOH estimation framework for flow batteries. This scheme deeply couples optical and electrical sensing modes at the physical mechanism level, relying on prior physical knowledge to effectively decouple internal chemical states from external operating condition interference, thereby achieving robust and accurate estimation of battery health status.

[0045] like Figure 1As shown, to achieve the above technical objectives, this invention provides a method for estimating the state of health (SOH) of a vanadium redox flow battery based on physical information and cross-modal collaboration. It provides a method for online health state estimation of a vanadium redox flow battery by combining optical spectroscopy and electrical signals and utilizing a network guided by physical information, specifically including the following steps: Step 1: Preprocessing of multimodal operation data The core of this step is to achieve synchronous alignment of multimodal signals, data standardization, and extraction of IC and DV features, providing high-quality input for subsequent modules. The detailed steps are as follows: 1. Simultaneous Acquisition and Alignment of Dual-Mode Signals: The system simultaneously acquires UV-Vis spectral signals (one signal each for the positive and negative electrodes), terminal voltage signals, and current signals during the operation of the flow battery. The sampling frequency is uniformly set to 1Hz, and the dual-mode signals are aligned using timestamp matching. The data acquisition time is T seconds (T can be 1, 10, 60, or 120 seconds to adapt to different prediction scenarios). The voltage, current, and spectral signals within the window are averaged to smooth sensor noise and avoid interference from noise in subsequent feature extraction. For the spectral signals, only the characteristic bands related to ion concentration (400~1000nm) are retained to avoid interference from irrelevant bands in subsequent feature extraction.

[0046] 2. IC (Incremental Capacity) / DV (Differential Voltage) Feature Processing: Based on the segmented voltage and current data, calculate the capacity change ΔQ and voltage change ΔV between sample points. Obtain the incremental capacity IC curve by calculating ΔQ / ΔV. Based on the inverse operation of the IC curve, calculate ΔV / ΔQ to obtain the differential voltage DV curve.

[0047] 3. Feature and Data Standardization: The extracted IC, DV features, and spectral signals are normalized, mapping all feature values ​​to... The interval is used to eliminate dimensional differences; at the same time, the consistency of the aligned dual-mode signal is checked to remove abnormal data points and ensure the reliability of the input data.

[0048] Step 2: Extraction of intrinsic factors for spectral attention (PISA) guided by physical information The core of this step is to achieve self-supervised differentiable inversion of ion component concentrations, and to construct an internal chemical state benchmark based on physical constraints. The specific steps are as follows: 1. Physically Constrained Bond-Value Pair Mesh Construction: Based on Beer-Lambert's law and the chemical equilibrium relationship of the flow battery electrolyte, constraint equations are established for the absorbance and ion concentration of the positive and negative electrode electrolytes. Considering the coexistence relationships of ions and complexes of various valence states in the positive and negative electrode electrolytes, the mass conservation and chemical equilibrium equations are solved. A set of concentration vectors that strictly satisfy the physical constraints is pre-calculated, generating a bond-value pair mesh containing G candidate states. Each candidate state is represented by a set of concentration vectors (denoted as "values"). Simultaneously, learnable parameters are introduced to generate the corresponding candidate spectral matrix (denoted as "bonds"). Beer-Lambert's law describes the physical law that the degree of light absorption by a substance is proportional to the concentration of the absorbing substance.

[0049] Taking the positive electrode as an example, the chemical constraint equations for spectrum formation and electrolyte composition are as follows:

[0050] in, This is the absorption spectrum of the electrolyte; The molar absorbance coefficients of each component; for , and its complexes Concentration of each component; This represents the total vanadium concentration. This is the equilibrium constant for the complexation reaction, for the negative electrode. V 2+ and V 3+ Similarly.

[0051] 2. Differentiable Inversion Based on Spectral Similarity Kernel (SSK): The measured spectrum is used as the input for matching (denoted as the "query"), and similarity matching is performed with the candidate spectral matrix. Based on a spectral similarity kernel function (SSK), the similarity between the measured spectrum and each candidate spectrum is calculated, and the square of the similarity is taken as the final matching index. The proposed formula for calculating spectral similarity is as follows:

[0052] in, Indicates the measured spectrum and the first The similarity of the candidate spectra; The measured spectrum; The first of the candidate spectral matrices OK; and These are their respective mean values.

[0053] 3. Attention Weight Calculation and Concentration Inversion: Using a Softmax function with a temperature coefficient τ=0.1, the similarity index is converted into attention weights. The concentration vectors in the key-value pair grid are then weighted and summed to obtain the estimated ionic component concentrations. This serves as a baseline for the battery's internal chemical state and is output to the next step. The calculation formula is as follows:

[0054] in, Attention weights; Temperature coefficient; Let j represent the total number of candidate states, j represent the candidate spectral label, and G represent the total number of candidate spectra.

[0055]

[0056] in, This represents the ion concentration vector that best matches the measured spectrum. This is a pre-calculated concentration vector in the key-value pair grid.

[0057] Step 3: Electrochemically Driven Voltage Decomposition (EDDL) Exogenous Decoupling The core of this step is to utilize the ion concentration benchmark output by the PISA module to decompose the physical components of the terminal voltage and eliminate external operating condition interference. The specific steps are as follows: 1. Equilibrium Potential (OCV) Calculation: Calculate the estimated ion concentration output from step 2. As input, an equilibrium potential calculation model is constructed using the Nernst equation. A learnable Nernst scaling factor α is introduced to replace the thermodynamic coefficient RT / F, and the equilibrium potential Vocv, directly related to the internal chemical state of the battery, is calculated as the benchmark for voltage decomposition. The calculation formula is as follows:

[0058] in, To balance the potential (open circuit voltage); The formal potential is approximately 1.4V; The concentration of vanadium ions obtained from the inversion; For learnable Nernst scale coefficients (alternate thermodynamic coefficients) ).

[0059] 2. Ohmic polarization component decomposition: A constrained linear operator is used for modeling. The current signal acquired in step 1 is used as input, and the Softplus operator is used to ensure that the equivalent internal resistance is non-negative, thus calculating the ohmic polarization. This characterizes the interference of external current conditions on the voltage. The calculation formula is as follows:

[0060] in, For Ohmic polarization; For current; To ensure that the equivalent internal resistance is always physically positive; These are parameters related to the battery's internal resistance.

[0061] 3. Concentration polarization component decomposition: Modeling is performed using the "product of basic polarization and velocity inhibition factor", based on the concentration estimate from step 2. Using the current and flow velocity signals from step 1, calculate the fundamental polarization intensity under the coupling effect of concentration and current, and combine it with the flow velocity suppression factor ( (λ is initially set to 0.05), thus obtaining concentration polarization. This characterizes the interference of external flow velocity and current conditions on voltage. The calculation formula is as follows:

[0062] in, For concentration polarization; The basic polarization intensity under concentration-current coupling; Electrolyte flow rate; A learnable parameter to characterize the nonlinear attenuation effect of flow velocity on polarization.

[0063] Step 4: Cross-modal residual fusion and physical consistency mapping The core of this step is to achieve deep fusion of internal and external features. Through physical constraint training, the fused features are mapped to SOH estimates. The specific steps are as follows: 1. Feature Definition and Mapping: Define intrinsic feature Fc (positive and negative electrode electrolyte component characteristics output by the PISA module) and extrinsic feature Fe (voltage component decomposed by the EDDL module, IC / DV features extracted in step 1, and control signals, including battery operating current and pump flow rate); through two independent linear layers, map Fc and Fe to a shared feature space (uniformly 64 dimensions) to form an optical embedding. and electrochemical intercalation .

[0064] 2. Cross-modal residual fusion: A fusion mechanism based on residual connections is adopted to splice optical embedding and electrochemical embedding to form joint features. Through linear projection and layer normalization operations, mutual correction and deep interaction of the two modal features are achieved, avoiding feature redundancy and information loss, and obtaining the fused feature vector. , The calculation formula is as follows:

[0065] in, and These are optical embedding and electrochemical embedding, respectively. This refers to the combined feature of the two components; and It is a learnable linear projection matrix.

[0066] 3. SOH Mapping and Output: The fused feature vector is fed into the fully connected layer. Through the jointly trained model, the final SOH estimate is output. This value has effectively removed external operating condition interference and can accurately reflect the actual health status of the battery.

[0067] 4. Physically Consistent Multi-Task Joint Training: A comprehensive objective function L_total is defined, comprising the main task loss (root mean square error between SOH predictions and true labels) and auxiliary task losses (spectral reconstruction constraints, ion dynamics consistency constraints, and voltage decomposition constraints). These are weighted and integrated into a unified optimization objective, and gradient optimization is used to converge the model parameters to values ​​consistent with physical laws. The calculation formula is as follows:

[0068] in, The mean square error of the SOH prediction value represents the main task loss. These are auxiliary loss terms for spectral reconstruction, ion dynamics consistency, and voltage decomposition, respectively. These are the corresponding balancing weights. In this example, the loss weights are set to... =0.1、 =0.1、 =0.01 Specifically: Spectral reconstruction constraints:

[0069] in For the measured spectrum, To reconstruct the spectrum.

[0070] Ion kinetic consistency constraints:

[0071] in This represents the rate of change of ions with respect to the amount of charge introduced. Let T be the mean of the rate of change, and t be the time period.

[0072] Voltage decomposition constraints:

[0073] in This is the measured terminal voltage. These represent the equilibrium potential, ohmic polarization, and concentration polarization components predicted by the model, respectively.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not 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 modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A flow battery SOH estimation method based on physical information and cross-modal collaboration, characterized in that, include: Acquire optical and electrical signals during the operation of the flow battery; The optical spectral signal is inverted based on physical mechanism constraints to obtain intrinsic characteristics; The intrinsic characteristic mentioned above is the estimated ionic component concentration of the battery electrolyte; Based on the intrinsic characteristics, the electrical signal is decomposed using electrochemical principles to obtain extrinsic characteristics; the extrinsic characteristics include equilibrium potential, ohmic polarization component, and concentration polarization component; the extrinsic characteristics also include IC / DV characteristics and control signals, the control signals including battery operating current and pump flow rate; the IC characteristic is incremental capacity characteristic, and the DV characteristic is differential voltage characteristic; Cross-modal residual fusion and mapping output are performed on the intrinsic and extrinsic features to obtain the SOH estimate of the flow battery; The process of cross-modal residual fusion of the intrinsic and extrinsic features includes: The intrinsic and extrinsic features are mapped to obtain optical embedding features and electrochemical embedding features; the optical embedding features and electrochemical embedding features are then spliced ​​together to obtain joint features; The optical embedding features and electrochemical embedding features are processed with the joint features through linear layer and layer normalization to obtain fused features; the fused features are then mapped to obtain the SOH estimate of the flow battery.

2. The method according to claim 1, characterized in that, After acquiring optical spectral signals and electrical signals, the following steps are also included: The optical spectral signal and the electrical signal are preprocessed, wherein the preprocessing includes normalization and consistency check. The optical spectral signals are ultraviolet-visible spectral signals for the positive and negative electrodes of the flow battery, and the bands of the spectral signals are characteristic bands related to the ion concentration of the flow battery; the electrical signals include terminal voltage signals, current signals, incremental capacity curve data, and differential voltage curve data.

3. The method according to claim 1, characterized in that, The process of inverting the optical spectral signal includes: A bond-value pair grid is constructed based on Beer-Lambert's law, wherein the bond-value pair network is used to store relational pairs with different spectral signals as bonds and corresponding ion component concentrations as values; wherein the mass conservation relationship of different ion component concentrations and the key to chemical equilibrium are used as constraints during the construction of the bond-value pair grid. The optical spectral signal is used as a query to perform matching and inversion in a key-value pair grid to obtain intrinsic features.

4. The method according to claim 3, characterized in that, The matching inversion process includes: The optical spectral signal is matched with the spectral signals corresponding to different bonds to obtain the candidate spectral similarity. The candidate spectral similarity is transformed using a fitting function to obtain the attention weight; the concentrations of the ion components corresponding to the attention weights are weighted and calculated to obtain the estimated ion component concentrations.

5. The method according to claim 1, characterized in that, The process of decomposing the electrical signal includes: The equilibrium potential is obtained by calculating the estimated ion component concentration using the Nernst equation; wherein the Nernst coefficient in the Nernst equation is a learnable parameter. The ohmic polarization component is obtained by calculating the current signal in the electrical signal through linear fitting; the linear operator in the linear fitting method is related to the battery internal resistance parameters and is constrained to be positive, and the battery internal resistance parameters are learnable parameters. The basic polarization intensity is calculated based on the estimated ion component concentration and current signal; the flow rate suppression factor is calculated based on the attenuation coefficient and the electrolyte flow rate of the flow cell; the concentration polarization component is calculated based on the basic polarization intensity and the flow rate suppression factor; wherein the attenuation coefficient is used to characterize the nonlinear attenuation effect of flow rate on polarization, and the attenuation coefficient is a learnable parameter.

6. The method according to claim 1, characterized in that, Learnable parameters are set in the process of optical spectral signal inversion, electrical signal decomposition, cross-modal residual fusion and mapping output, and the learnable parameters are optimized according to the comprehensive objective function; The comprehensive objective function includes a main task loss and an auxiliary task loss to characterize the deviation of the SOH estimate. The auxiliary task loss includes spectral reconstruction constraints, ion dynamics consistency constraints, and voltage decomposition constraints.

7. A flow battery SOH estimation system based on physical information and cross-modal collaboration, characterized in that, Used to perform the method described in any one of claims 1-6.

8. A computer device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

Images