Battery life prediction method and device based on time sequence and multi-modal model fusion

Abstract

The invention belongs to the technical field of battery health management, and particularly relates to a time sequence and multi-modal model fusion-based battery life prediction method and device, and the method comprises the steps: collecting electrochemical performance data, text semantic data and visual image data of a to-be-detected battery, and extracting corresponding time sequence features, text features and image features; and dynamically adjusting each modal contribution weight through a learnable gating coefficient, determining a life degradation task driving correlation coefficient matrix and a structure preserving mask matrix, and obtaining each modal enhancement feature by using a cross-modal multi-head attention mechanism. And performing semantic enhancement decoupling fusion on the enhanced features to obtain multi-modal fusion features, and finally inputting the multi-modal fusion features into a regression predictor to output a prediction result of the remaining service life of the battery. According to the method, the contribution of each modal is adaptively adjusted through a dynamic weighting mechanism, cross-modal feature deep interaction is realized through a physically guided attention mechanism, redundant information is inhibited through decoupling fusion, and full fusion and utilization of multi-source heterogeneous data are realized.

Description

A method and apparatus for predicting battery lifetime based on the fusion of time series and multimodal models Technical Field

[0001] This invention belongs to the field of battery health management technology, specifically relating to a battery life prediction method and device based on the fusion of time series and multimodal models. Background Technology

[0002] Health status monitoring and remaining lifespan prediction of lithium-ion batteries are key technologies in new energy vehicles, energy storage systems, and portable electronic devices. Currently, battery remaining lifespan prediction methods are mainly divided into two categories: model-based methods and data-driven methods. Among them, data-driven methods have received widespread attention because they do not require the establishment of complex electrochemical models. These methods primarily involve collecting electrochemical time-series data such as battery charge / discharge voltage, current, capacity, and internal resistance, performing regression analysis, establishing battery degradation models, and predicting remaining lifespan. In recent years, with the development of sensor technology and artificial intelligence, some studies have begun to explore using multi-source data to assist in prediction. For example, textual data such as operation and maintenance logs and scene annotations can be used to describe the battery's usage environment and maintenance records, or visual data such as battery appearance images and thermal imaging can be used to perceive physical damage and microstructural changes in the battery.

[0003] Among existing multimodal fusion technologies, Chinese patent application CN118795336A proposes a battery state assessment method based on multimodal data fusion. This method collects electrical and non-electrical quantity monitoring data of the battery, extracts feature vectors through convolutional neural networks and bidirectional sequence models, calculates the information entropy of each feature vector and performs weighted fusion, and then performs secondary fusion of the fused features through a self-attention mechanism to finally output battery state assessment data. This method attempts to fuse multi-source data to assess battery state, which improves the comprehensiveness of the assessment to a certain extent. However, the information entropy weighting method used in this approach is a static weight allocation method, which cannot adaptively adjust the modal contribution according to the dynamic changes in the reliability of data from different modalities at different battery usage stages. When there is noise or missing data in a certain modality, the model performance will significantly decrease. In addition, the self-attention mechanism of this method only operates on the fused single feature vector, failing to achieve deep interaction between the original features of different modalities, and making it difficult to fully explore the nonlinear coupling relationship between heterogeneous data such as time series, text, and images. Finally, this method does not determine the physical laws guiding the battery degradation process, such as thermal diffusion paths, electrochemical reaction active regions, and aging expansion directions, resulting in a lack of physical interpretability in cross-modal feature associations, making it difficult to meet the requirements for high-precision and quantifiable remaining service life prediction. Summary of the Invention

[0004] The purpose of this invention is to provide a battery life prediction method and apparatus based on the fusion of time series and multimodal models, so as to solve the problem that the existing technology cannot effectively utilize and fuse multi-source heterogeneous data of batteries.

[0005] The present invention achieves the above-mentioned objectives through the following technical solutions: Firstly, the present invention proposes a battery lifetime prediction method based on the fusion of time-series and multimodal models. The method includes: acquiring multi-source heterogeneous data of the battery under test and extracting multi-source heterogeneous features from it; wherein the multi-source heterogeneous features include electrochemical time-series features, text features, and image features; generating corresponding learnable gating coefficients based on the multi-source heterogeneous features; constraining the cross-modal multi-head attention mechanism by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix based on the multi-source heterogeneous features, thereby obtaining cross-modal attention enhancement features for each modality; performing semantic enhancement decoupling fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multimodal fusion features; and inputting the multimodal fusion features into a regression predictor to obtain the remaining lifetime prediction result of the battery under test.

[0006] Furthermore, the multi-source heterogeneous data includes electrochemical performance data, textual semantic data, and visual image data; after preprocessing, the electrochemical performance data is input into a time-series encoder to extract the electrochemical time-series features characterizing the evolution of the battery's internal state; the textual semantic data is input into a pre-trained language model to extract the textual features characterizing the semantic information of the battery's use and maintenance background; and the visual image data is input into a visual encoder to extract the image features characterizing the battery's physical damage and electrochemical behavior.

[0007] Furthermore, the electrochemical performance data includes at least one of the following: charge / discharge voltage curves, current, capacity, internal resistance, open-circuit voltage, coulombic efficiency, operating status time-series data, charge rate, depth of discharge, number of cycles, resting time, and environmental parameters; the textual semantic data includes at least one of the following: operation and maintenance logs, scene annotation text, and technical documents; and the visual image data includes at least one of the following: battery appearance image, electrode microstructure image, separator state image, electrochemical visualization image, and screenshot of the battery management system.

[0008] Furthermore, generating corresponding learnable gating coefficients based on the multi-source heterogeneous features includes: for the k-th modality among the temporal modality, text modality, and image modality in the multi-source heterogeneous features, calculating the gating coefficient using the following formula. : ;in, The Sigmoid function restricts the gating coefficient to the range [0,1]. For modal embedding vectors, and Trainable parameters; gating coefficients Used to adjust the effectiveness of corresponding modalities in the fusion process, achieving interpretable weighting at the sample level; These are time-series, text, and image data, respectively.

[0009] Furthermore, based on the aforementioned multi-source heterogeneous features, the cross-modal multi-head attention mechanism is constrained by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix to obtain cross-modal attention enhancement features for each modality, including: features for modality m. and the characteristics of mode n The features of mode m are mapped to a query vector using a learnable projection matrix. Map the features of mode n to key vectors Sum value vector ;in To query the projection matrix; The key projection matrix; The projection matrix is ​​defined; the lifetime degradation task-driven correlation coefficient matrix C is determined to quantify the contribution strength of each feature in mode n to the degradation evolution of mode m at different times; the structure preservation mask matrix M is determined to limit the effective range of cross-modal attention; based on the query vector... The key vector The correlation coefficient matrix C and the structure-preserving mask matrix M are used to calculate the attention weights of mode m on mode n. As shown in the following formula: Where ⊙ denotes element-wise multiplication, and softmax denotes normalization over all features of mode n. The feature dimension of the key vector. The normalized modulation matrix is ​​determined based on the attention weights. Sum value vector The enhanced features are calculated as follows: Under a multi-head mechanism, the query vector will be... Key vector Sum value vector The system is divided into h subspaces, and each head independently computes attention and enhancement features to obtain the outputs of each head. These outputs are then concatenated and linearly mapped to form the final modality enhancement features. .

[0010] Furthermore, the process of generating the lifetime degradation task-driven correlation coefficient matrix C includes: acquiring at least one degradation index among the capacity decay rate, internal resistance growth rate, or remaining lifetime label of the battery under test; mapping the degradation index to a degradation state sequence or degradation spatial distribution, and performing scale alignment on each modal feature according to the degradation state sequence or degradation spatial distribution; applying a perturbation to each feature in the multi-source heterogeneous features, calculating the degree of influence of feature value changes on degradation rate changes, and obtaining the sensitivity of each feature; identifying lifetime-sensitive features or degradation-sensitive features from the multi-source heterogeneous features based on the sensitivity of each feature; for the identified lifetime-sensitive features or degradation-sensitive features, converting their influence on the degradation rate into lifetime response weights through local regression or piecewise fitting methods to obtain the initial correlation coefficient matrix elements. Where i represents the time step index and j represents the feature index; combining the thermal diffusion path, electrochemical reaction active region, and aging propagation direction, the elements of the initial correlation coefficient matrix are physically consistent to obtain the corrected correlation coefficient matrix elements. As shown in the following formula: The modified correlation coefficient matrix is ​​normalized row by row to obtain the correlation coefficient matrix C of the probability distribution, and updated as the training process or degradation state progresses.

[0011] Furthermore, the structure-preserving mask matrix M is constructed based on modal characteristics, specifically including: for temporal features, the mask matrix M is constructed based on a window size k determined by thermal diffusion or charge / discharge lag time, allowing attention to be calculated only within the current degradation state and the physically propagable time window, as shown in the following equation: For image or two-dimensional parametric features, the mask matrix M is constructed based on the local aging diffusion neighborhood to maintain the spatial continuity of the degradation morphology and the local topological structure.

[0012] Furthermore, the step of performing semantic enhancement decoupling and fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multimodal fusion features includes: for the fusion feature vector of the i-th time step or sample, calculating the multimodal fusion features using the following formula. : ;in, These represent the original features of the temporal, text, and image modalities, respectively. This is a cross-modal attention enhancement vector; , , These are the gating coefficients for the corresponding modes, used to regulate the contribution of the original features; These are the weighting coefficients for the cross-modal enhancement terms.

[0013] Furthermore, the step of inputting the multimodal fusion features into the regression predictor to obtain the battery remaining life prediction result includes: assuming that the enhancement features of the i-th sample in each modality are respectively... , , The enhanced features of each modality are mapped to a unified dimensional space using a learnable linear mapping matrix, and then integrated through a concatenation operation to obtain the final fused features. As shown in the following formula: ;in, , , is a learnable linear mapping matrix; ⊕ denotes a vector concatenation operation; For normalization or nonlinear transformation functions; the final fused features Input the regression predictor to generate the remaining useful life prediction for the sample, as shown in the following formula: ;in, This is the regression prediction function.

[0014] Secondly, this invention proposes a battery life prediction device based on the fusion of time-series and multimodal models, used to implement the battery life prediction method described above. The device includes: a time-series acquisition module for acquiring electrochemical time-series data of the battery under test and extracting time-series features; a text acquisition module for acquiring text data of the battery under test and extracting text features; an image acquisition module for acquiring image data of the battery under test and extracting image features; a gating module for generating corresponding learnable gating coefficients based on the time-series features, text features, and image features; an attention module for constraining the cross-modal multi-head attention mechanism by determining the life degradation task-driven correlation coefficient matrix and structure-preserving mask matrix based on the time-series features, text features, and image features, thereby obtaining cross-modal attention enhancement features for each modality; a fusion module for performing semantic enhancement decoupling fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multimodal fusion features; and a prediction module for inputting the multimodal fusion features into a regression predictor to obtain the battery remaining life prediction result.

[0015] The beneficial effects of the present invention are as follows: 1. The present invention dynamically adjusts the contribution of each mode by learning gating coefficients, and can adaptively adjust the fusion weight according to the reliability of each mode data at different battery usage stages. When there is noise or missing data in a certain mode, its influence is automatically suppressed, thereby improving the robustness of the prediction model.

[0016] 2. This invention constrains the cross-modal multi-head attention mechanism by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix, quantifies the contribution intensity of different modal features to the battery degradation process, and uses physical laws such as thermal diffusion paths and electrochemical reaction active regions as constraints to limit the attention range, ensuring that cross-modal correlations are established only within physically interpretable degradation channels, thereby enhancing the interpretability of the model.

[0017] 3. This invention achieves full fusion of multi-source heterogeneous data by semantic enhancement and decoupling fusion, which retains the original modal information while absorbing cross-modal complementary information, suppresses redundancy and noise interference, and finally outputs quantifiable battery remaining life prediction results. Attached Figure Description

[0018] Figure 1 is a flowchart of a battery life prediction method in an embodiment of the present invention; Figure 2 is a flowchart of a cross-modal multi-head attention interaction mechanism in an embodiment of the present invention; Figure 3 is a flowchart of a gating-attention joint multimodal fusion mechanism in an embodiment of the present invention; Figure 4 is another flowchart of a battery life prediction method in an embodiment of the present invention. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0020] Example 1

[0021] Referring to Figures 1 and 4, a specific embodiment of this disclosure proposes a battery lifetime prediction method based on the fusion of time-series and multimodal models. The method includes: acquiring multi-source heterogeneous data of the battery under test and extracting multi-source heterogeneous features from it; wherein the multi-source heterogeneous features include electrochemical time-series features, text features, and image features; generating corresponding learnable gating coefficients based on the multi-source heterogeneous features; obtaining cross-modal attention enhancement features for each modality by determining a cross-modal multi-head attention mechanism that drives the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix based on the multi-source heterogeneous features; performing semantic enhancement decoupling fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multimodal fusion features; and inputting the multimodal fusion features into a regression predictor to obtain the prediction result of the remaining lifetime of the battery under test.

[0022] Optionally, this method can be applied to the health status monitoring and remaining service life prediction of lithium-ion batteries in scenarios such as new energy vehicles, energy storage systems, and portable electronic devices.

[0023] In a preferred embodiment, the multi-source heterogeneous data includes electrochemical performance data, textual semantic data, and visual image data; wherein, the electrochemical performance data is used to characterize the evolution of the internal electrochemical state of the battery, the textual semantic data is used to describe the battery's usage and maintenance background, and the visual image data is used to reflect the battery's physical damage and microstructural changes.

[0024] Specifically, electrochemical performance data includes at least one of the following: charge / discharge voltage curves, current, capacity, internal resistance, open-circuit voltage, coulombic efficiency, operating state time-series data, charge rate, depth of discharge, number of cycles, resting time, and environmental parameters. Environmental parameters include ambient temperature, temperature gradient within the battery pack, and humidity, which significantly impact the battery's degradation rate. Textual semantic data includes at least one of the following: maintenance logs, scenario-annotated text, and technical documents. Maintenance logs record fault descriptions, maintenance records, and abnormal events during battery use. Scenario-annotated text describes the battery's application scenarios, operating environment, and charging strategies. Technical documents provide battery specifications and electrochemical mechanism descriptions. Visual image data includes at least one of the following: battery appearance images, electrode microstructure images, separator state images, electrochemical visualization images, and screenshots from the battery management system. Electrochemical visualization images include electrochemical impedance spectroscopy curves, cyclic voltammetry curves, and thermal imaging images. These images can reflect the battery's electrochemical behavior and thermal distribution characteristics from different dimensions.

[0025] After preprocessing, the electrochemical performance data is input into a time-series encoder to extract electrochemical time-series features characterizing the internal state evolution of the battery. The preprocessing steps include standardization, filtering, resampling, and normalization to eliminate noise interference and dimensional differences in the original data, ensuring comparability of time-series data from different sources and scales. The time-series encoder employs the TimeLLM model, which effectively captures long-term and short-term dependencies in the electrochemical time-series data, extracting high-dimensional time-series feature representations.

[0026] Textual semantic data is input into a pre-trained language model to extract textual features representing the semantic information of battery use and maintenance. The pre-trained language model adopts BERT, RoBERTa, or LLaMA series models. After being pre-trained on large-scale corpora, these models can fully understand textual semantics and transform unstructured textual data into feature vectors rich in semantic information, providing textual modal input for subsequent multimodal fusion.

[0027] Visual image data is input into a visual encoder to extract image features characterizing the physical damage and electrochemical behavior of the battery. The visual encoder employs a convolutional neural network, a visual Transformer, or a hybrid architecture of the two, and is able to extract multi-level, multi-scale visual features from images, including macroscopic damage features of the battery appearance (such as bulging, leakage, and shell damage) and microstructural features (such as electrode grain size, porosity, and separator state).

[0028] For example, time-series data encoding and characterization includes the following: First, electrochemical time-series data is collected, including charge-discharge voltage curves. Current ,capacity Internal resistance Open circuit voltage Coulomb efficiency and runtime status timing data Simultaneously record the charging rate. Depth of discharge Number of loops Settling time and environmental parameters (such as ambient temperature, internal temperature gradient, humidity, etc.).

[0029] The time-series data are standardized, filtered, resampled, and normalized to form a unified electrochemical feature vector. As shown in the following formula: Subsequently, the processed electrochemical time series characteristics are input into the time series model. (TimeLLM) is used for characterization to obtain a high-dimensional temporal feature representation: For example, text data encoding and representation includes the following: collecting and organizing multi-source text data, including: operation and maintenance logs: fault descriptions, maintenance records, and abnormal events; scenario-annotated text: application scenarios, usage environments, and charging strategies; and technical documents: battery specifications and electrochemical mechanism descriptions.

[0030] Text data is encoded using pre-trained language models (such as BERT, RoBERTa, and LLaMA series) to obtain text feature vectors: ;in It contains all text entries.

[0031] The vector output from this step retains semantic information and can be used to help the RUL / lifetime prediction model understand the battery's usage and maintenance context.

[0032] For example, image data encoding and characterization include the following: acquiring images of the battery appearance and structure, including: cell appearance: bulging, leakage, shell damage; electrode microstructure: grain size, porosity; membrane state: damage, wrinkles; electrochemical visualization images: EIS, CV curves, thermal imaging; BMS screenshots and monitoring interface status diagrams.

[0033] Extracting high-dimensional image features using convolutional neural networks (CNNs), visual Transformers, or hybrid visual encoders: This output can capture the physical damage and electrochemical behavior characteristics of the battery, providing important information for multimodal fusion.

[0034] In this embodiment, after completing the vectorization of three modalities—time series, text, and image—the present invention determines a learnable modal gating mechanism to dynamically adjust the contribution of different modalities in each sample, in order to cope with the differences in modal reliability caused by changes in battery usage stage and experimental conditions.

[0035] In a preferred embodiment, learnable gating coefficients are generated based on the multi-source heterogeneous features, including: for the k-th modality among the temporal modality, text modality, and image modality in the multi-source heterogeneous features, the gating coefficient is calculated using the following formula. : ;in, The Sigmoid function restricts the gating coefficient to the range [0,1]. For modal embedding vectors, and Trainable parameters; gating coefficients Used to adjust the effectiveness of corresponding modalities in the fusion process, achieving interpretable weighting at the sample level; These are time-series, text, and image data, respectively.

[0036] In this embodiment, the modality embedding vector This refers to the high-dimensional semantic representation obtained by encoding the original data (including time-series data, text data, and image data) of the k-th modality through the corresponding feature extraction network. The time-series modality uses a Transformer encoder to extract time-dependent features, the text modality uses a pre-trained language model for semantic vectorization, and the image modality uses a convolutional neural network or a visual Transformer to extract spatial structure features, which are then pooled or aggregated to obtain a fixed-dimensional embedding vector. The trainable parameters... and The linear mapping weights and bias terms in the gating network are generated using random initialization methods (such as Xavier or normal distribution initialization) during model construction. During the end-to-end training of the overall model, they are automatically updated and optimized through the backpropagation algorithm under the constraint of the loss function, so that the gating coefficients can adaptively adjust the contribution weight of each modality in the fusion process according to the modal characteristics of different samples.

[0037] Understandably, through gradient backpropagation, the model can automatically learn the optimal mapping between the gating coefficient and the remaining useful life prediction accuracy, thereby automatically suppressing the impact of noise or missing modal data. For example, when image modal noise is severe, the gating coefficient of the image modality... Approaching 0; when the text modality is missing or anomalous, the gating coefficient of the text modality approaches 0. Approaching 0; when abnormal spikes appear in the time series modal data, the gating coefficient of the time series mode approaches 0. It approaches 0. Simultaneously, the gating coefficient can dynamically adjust the mode weights according to the battery life cycle stage: in the early degradation stage, time-series modes dominate. The contribution of image modalities is relatively high; in the mid-to-late degradation stage, the contribution of image modalities increases. Enhancement; during the fault-related phase, the text modality dominates. Increase.

[0038] Meanwhile, the gating mechanism can dynamically adjust the mode weights according to the battery life cycle stage: Ultimately, the gating vector for each sample It can quantify the contribution of each modality to RUL prediction, providing accurate input for subsequent attention interaction and multimodal fusion, while achieving interpretability and verifiability of prediction results.

[0039] Referring to Figure 2, to effectively model the nonlinear coupling relationships between different modalities, this invention defines a cross-modal multi-head attention interaction mechanism. In a preferred embodiment, based on multi-source heterogeneous features, the cross-modal multi-head attention mechanism is constrained by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix, thereby obtaining the cross-modal attention enhancement features for each modality, including: features for modality m. and the characteristics of mode n The features of mode m are mapped to a query vector using a learnable projection matrix. Map the features of mode n to key vectors Sum value vector ;in To query the projection matrix; The key projection matrix; For the value projection matrix; specifically, It is the characteristic matrix of mode m. The length of the feature sequence for this modality (number of tokens / number of time steps / number of patches). This is the feature dimension (embedding dimension) of this modality. The length of the feature sequence of modality n (number of tokens / time steps / number of patches). Let n be the feature dimension (embedding dimension).

[0040] Determine the lifetime degradation task-driven correlation coefficient matrix C to quantify the contribution strength of each feature in mode n to the degradation evolution of mode m at different times; determine the structure preservation mask matrix M to limit the effective range of cross-modal attention; and determine the structure preservation mask matrix M based on the query vector. Key vector The correlation coefficient matrix C and the structure preservation mask matrix M are used to calculate the attention weights of mode m on mode n. Specifically, the query vector is first calculated. With key vector dot product and divide by scaling factor The original attention score matrix is ​​obtained, and then the original attention score matrix is ​​multiplied element-wise with the normalized correlation coefficient matrix C and the structure-preserving mask matrix M. Finally, the attention weights are normalized over all features of mode n using the softmax function, as shown in the following formula: Where ⊙ denotes element-wise multiplication, and softmax denotes normalization over all features of mode n. The feature dimension of the key vector. The normalized modulation matrix is ​​determined based on the attention weights. Sum value vector The enhanced features are calculated as follows: Under a multi-head mechanism, the query vector will be... Key vector Sum value vector The system is divided into h subspaces, and each head independently computes attention and enhancement features to obtain the outputs of each head. These outputs are then concatenated and linearly mapped to form the final modality enhancement features. Through this mechanism, each modality not only retains its own feature information, but also fully absorbs the complementary information of other modalities, realizing deep cross-modal interaction and feature enhancement.

[0041] Through the above process, each modality not only retains its own feature information but also fully absorbs complementary information from other modalities, achieving deep cross-modal interaction and feature enhancement. This provides richer input features for subsequent prediction, classification, or degradation mechanism analysis. The entire mechanism is naturally applicable to the joint modeling of multimodal data such as time series, images, and text, and can capture complex nonlinear coupling relationships between modalities.

[0042] In a preferred embodiment, the process of generating the lifetime degradation task-driven correlation coefficient matrix C includes: (1) obtaining at least one degradation index among the capacity decay rate, internal resistance growth rate, or remaining lifetime label of the battery under test; mapping the degradation index to a degradation state sequence or degradation spatial distribution, and based on the degradation state sequence (Temporal modality) or degenerate spatial distribution (image / 2D parametric modality) scales the features of each modality to ensure clear alignment between different modalities.

[0043] (2) Apply a perturbation to each feature in the multi-source heterogeneous features, calculate the degree of influence of the feature value change on the degradation rate change, and obtain the sensitivity of each feature. To further distinguish between lifespan-sensitive features and degradation-sensitive features, the specific method is as follows: First, take the absolute value of the sensitivity of each sample and normalize it to obtain the relative contribution weight. This is to facilitate comparison of the contribution sizes of different features: ;in, Let be the sensitivity of the j-th feature of the i-th sample to changes in the degradation rate, and d be the total number of features. After normalization, the weights of all features sum to 1.

[0044] Based on this, feature type determination is performed: Lifetime-sensitive features: These are features that significantly drive changes in the battery degradation rate. Technically, lifetime-sensitive features typically correspond to features with large and positive sensitivity values, meaning that an increase in these values ​​leads to a faster degradation rate or faster capacity decay. In the determination process, features with positive sensitivity and high weight are marked as lifetime-sensitive features. For example, statistical characteristics of a sudden increase in battery temperature during a certain cycle, or indicators of negative electrode active material degradation, may fall under the category of lifetime-sensitive features.

[0045] Degradation-sensitive features: These are features that play a central role in degradation rate prediction or remaining lifetime estimation and have the greatest impact on the stability of model output. Technically, degradation-sensitive features can be features with high absolute sensitivity values, but their directionality can be positive or negative, representing their core contribution to the degradation rate. In determining their sensitivity, features with high weight (regardless of whether they are positive or negative) are labeled as degradation-sensitive features. For example, higher-order statistical features such as the slope of capacity change or the growth rate of internal resistance, even if the change is small, can have a decisive impact on lifetime prediction results.

[0046] In this embodiment, let the degradation rate of the i-th battery sample within the current observation window be . Its relative change is denoted as ,in It can be defined as a degradation index obtained through time difference or regression fitting, such as the capacity decay slope, internal resistance growth rate, or state of health (SOH) change rate. Let the feature vector after multi-source heterogeneous feature fusion be... ,in This represents the j-th feature component in the n-th modality, and the feature can be derived from temporal statistical features, text semantic embedding components, or image structural feature components.

[0047] Among them, degradation rate The capacity is obtained by performing sliding window linear regression, exponential fitting, or difference calculation on the battery capacity versus cycle number or time curve. For example, the least squares method can be used to fit the capacity-time curve, and the slope of the fit can be used as the degradation rate. .

[0048] Eigencomponents The method of obtaining the feature vector is as follows: it is directly obtained from the j-th dimension component of the embedding vector output by the multimodal feature extraction network. Specifically, this includes: obtaining statistical or deep representations of temporal modalities through temporal coding networks; obtaining semantic vectors of text modalities through pre-trained language models; obtaining spatial feature vectors of image modalities through convolutional neural networks or visual Transformers; and forming a unified feature representation after concatenation or fusion.

[0049] (3) For life-sensitive or degradation-sensitive features identified in the sensitivity analysis, if they have time-series patterns, spatial structures, or periodic evolution characteristics, it is necessary to further model their local response behavior within a specific time window or operating condition interval, and quantify this response into specific life-response weights to obtain the initial correlation coefficient matrix elements. , where i represents the time step index and j represents the feature index.

[0050] Understandably, a local response refers to the response at the current moment. Within a certain time window nearby, a certain lifespan-sensitive characteristic Battery degradation rate caused by minute changes The change in local response. Its difference from global sensitivity lies in the fact that global sensitivity reflects the average impact of a feature throughout its entire lifespan, while local response captures the instantaneous intensity of the impact at the current moment, reflecting the nonlinearity and time-varying nature of battery aging. For example, a 5°C increase in temperature can cause completely different degrees of accelerated degradation in the early and late stages of battery life; this difference is characterized by the local response.

[0051] As an example, the calculation and weight transformation process of the local response is as follows: Select the current time... The window consists of N adjacent data points. Within the window, the feature values ​​are used to... The x-axis represents the degradation rate. Using the vertical axis as the ordinate, perform linear regression. The slope k of the fitted line is the original response value of that feature at the current time. Normalizing the absolute values ​​of the slopes of each feature yields the weights. .

[0052] For example, the 100th cycle ( The sliding window takes a total of 5 points, both before and after. Two degradation-sensitive features have been identified: temperature ( ) and internal resistance increment ( ).

[0053] 1. Local data and fitting in ~ Within the window, the data trend is as follows: Temperature ( ): Degradation rate from 44℃ to 46℃ It rose from 0.04% to 0.05%.

[0054] Fitting slope .

[0055] Internal resistance ( ): Degradation rate from 1.0 mΩ to 1.2 mΩ It surged from 0.04% to 0.08%.

[0056] Fitting slope .

[0057] 2. Calculate the elements of the weight matrix. Total response intensity: Temperature weighting: Internal resistance weight: It can be seen that, in the 100th cycle, although the temperature is rising, the local response weight of the internal resistance increment is as high as 0.976, indicating that it is the dominant factor leading to accelerated degradation at this point. Based on this, the model will assign a higher predictive weight to the internal resistance feature.

[0058] (4) Combining the thermal diffusion path, the active region of electrochemical reaction, and the direction of aging expansion, the elements of the initial correlation coefficient matrix are physically consistent to obtain the elements of the corrected correlation coefficient matrix. As shown in the following formula: In this embodiment, "feature j can affect time step i" means that the physical quantity corresponding to feature j has the ability to act on the degradation state at time step i within the battery. That is, changes in this feature can have a direct and explainable impact on the degradation rate of the i-th time step through physical propagation mechanisms such as thermal diffusion paths, electrochemical reaction activity regions, or aging propagation directions. Specifically: Thermal diffusion path constraint: If feature j originates from the statistics of local battery temperature, its range of influence is limited by the thermal diffusion rate and direction. It can only be considered to affect that time step if the feature is located within the thermal response region corresponding to the i-th time step.

[0059] Electrochemical reaction active region: If feature j reflects the state of certain electrochemically active substances (such as local capacity, internal resistance or surface reaction activity), its effectiveness is constrained by the distribution of active regions. Only when the feature is within the active region can it be considered to affect the degradation process at the i-th time step.

[0060] Aging propagation direction: Battery degradation is spatially or temporally continuous. The influence of feature j propagates along the aging propagation direction. Only when the feature is located in the aging propagation direction or on its action chain can it be considered to affect the i-th time step.

[0061] Accordingly, "if feature j cannot affect time step i" means that the change in feature j cannot affect the degradation state of the i-th time step through any of the aforementioned physical mechanisms; that is, there is no physical propagation path or causal relationship between the two that conforms to the thermal diffusion path, the distribution of the electrochemical reaction active region, or the direction of aging expansion. In this case, the correlation between the feature and time step i is considered a non-physical spurious correlation and should be eliminated during the correction process.

[0062] Specifically, the elements of the corrected correlation coefficient matrix The acquisition method is as follows: 1. For each time step i and each feature j, determine whether the feature can be physically applied to that time step (based on thermal diffusion path, electrochemical active region and aging direction).

[0063] 2. If feature j is effective, then retain the original locally calculated correlation coefficient. .

[0064] 3. If feature j is not applicable, the corresponding matrix element is set to 0 to shield the sensitivity analysis and lifetime prediction from statistical noise or non-physical correlation.

[0065] This correction method not only preserves the statistical correlation information of the correlation coefficient matrix, but also keeps it consistent with the actual physical mechanism of the battery, thus improving the model's interpretability and prediction reliability.

[0066] (5) Normalize the modified correlation coefficient matrix row by row to obtain the correlation coefficient matrix C of the probability distribution, and update it as the training process or degradation state progresses to ensure that attention continues to focus on the lifetime-sensitive features, as shown in the following formula: ;in, For matrix The i-th row vector, This indicates the degree of modified correlation between the i-th feature unit and the j-th feature unit.

[0067] The above process uses a temporal modality as an example, but the generation principle of matrix C can be extended to modalities such as images, two-dimensional parameters, or text. Different modalities only need to map the indices to the time, space, or feature dimensions, and adjust the local response modeling and masking strategies according to the modal characteristics.

[0068] In a preferred embodiment, the structure-preserving mask matrix M is constructed based on modal characteristics, specifically including: determining the structure-preserving mask matrix to preserve the causal order of temporal data and the local topological structure of spatial data. This is used to limit the effective range of cross-modal attention. For temporal features, the mask matrix M is constructed based on a window size k determined by thermal diffusion or charge / discharge lag time, allowing attention to be computed only within the current degradation state and the physically propagable time window, as shown below: ;in, For mask matrix elements; This refers to the mask matrix elements at the corresponding positions when the absolute time distance between time step i and time step j does not exceed the preset window size k. Setting it to 1 indicates that the cross-modal attention computation is allowed to establish a correlation between the two time steps. This refers to the mask matrix element at the corresponding position when the absolute time distance between time step i and time step j exceeds the preset window size k. Setting it to 0 indicates that the cross-modal attention computation is disabled; the window size k is determined by the thermal diffusion time constant or the charge / discharge lag time to ensure that the causal order in time and the physical propagation range are maintained.

[0069] For image or two-dimensional parametric features, the mask matrix M is constructed based on the local aging diffusion neighborhood to maintain the spatial continuity of the degradation morphology and the local topology. For example, in electrode microstructure images, aging often starts from grain boundaries or defects and spreads outwards. Therefore, the mask matrix should ensure that attention is established only within the local neighborhood, avoiding non-physical connections across regions.

[0070] Referring to Figure 3, in a preferred embodiment, semantic enhancement decoupling and fusion are performed based on learnable gating coefficients and cross-modal attention enhancement features to obtain multimodal fusion features. This includes: for the fusion feature vector of the i-th time step or sample, the multimodal fusion features are calculated using the following formula. : ;in, These represent the original features of the temporal, text, and image modalities, respectively. This is a cross-modal attention enhancement vector; , , These are the gating coefficients for the corresponding modes, used to regulate the contribution of the original features; These are the weighting coefficients for the cross-modal enhancement terms.

[0071] This weighting coefficient The training process can be adaptively adjusted if the attention enhancement vector contains redundant or noisy information. This will be automatically suppressed, thus ensuring the stability of the fused features and avoiding noise amplification. Through this module design, the fused features can achieve a dynamic balance between the direct path (preserving the original modal information) and the interaction path (cross-modal enhancement information), effectively preserving the most valuable cross-modal semantic information and suppressing interference from irrelevant or redundant information.

[0072] In a preferred embodiment, the multimodal fusion features are input into a regression predictor to obtain the battery remaining life prediction result, including: assuming the enhanced features of the i-th sample in each modality are as follows: , , The enhanced features of each modality are mapped to a unified dimensional space using a learnable linear mapping matrix, and then integrated through a concatenation operation to obtain the final fused features. As shown in the following formula: ;in, , , is a learnable linear mapping matrix; ⊕ denotes a vector concatenation operation; It is a normalization or nonlinear transformation function used to balance the differences in characteristic amplitudes and distributions of different modes.

[0073] Final fusion features Input the regression predictor to generate the remaining useful life (RUL) prediction for the sample. As shown in the following formula: ;in, This is the regression prediction function.

[0074] In a preferred embodiment, the regression predictor employs a multilayer perceptron structure with two hidden layers, each containing 128 and 64 neurons respectively. The ReLU activation function is used, and the output layer uses a linear activation function to directly output the predicted remaining lifespan. Model training uses mean squared error as the loss function, Adam as the optimizer, with an initial learning rate of 0.001, a batch size of 32, 100 training epochs, and an early stopping mechanism to prevent overfitting.

[0075] In a preferred approach, multi-source heterogeneous data is collected from lithium-ion battery systems operating in real-world environments, with a sampling frequency of 1 Hz. The data acquisition duration covers the entire lifecycle of the battery, from its newest state to the end of its lifespan. Electrochemical time-series data is collected in real-time through the battery management system, text data is recorded and structured through an operation and maintenance log system, and image data is collected periodically through industrial cameras and thermal imagers. During the data preprocessing stage, missing values ​​are filled using linear interpolation, outliers are detected and replaced using the 3σ principle, and multi-source data is aligned by timestamps to ensure consistency across different modalities in the time dimension.

[0076] Example 2

[0077] This disclosure presents a specific embodiment of a battery life prediction device based on the fusion of time-series and multimodal models. This device implements the steps of the battery life prediction method proposed in Embodiment 1. It can be integrated into a battery management system or deployed as a standalone edge computing device in application scenarios such as new energy vehicles and energy storage power stations, providing real-time and accurate predictions of remaining battery life for health status monitoring and maintenance decisions. The device includes: a time-series acquisition module for acquiring electrochemical time-series data of the battery under test and extracting time-series features; this module is connected to the battery management system's communication interface and acquires in real-time at least one of the following: charge / discharge voltage curves, current, capacity, internal resistance, open-circuit voltage, coulombic efficiency, operating status time-series data, charge rate, discharge depth, cycle count, rest time, and environmental parameters. The acquired raw data undergoes standardization, filtering, resampling, and normalization preprocessing before being input into a time-series encoder to obtain a high-dimensional time-series feature representation. The time-series encoder uses the TimeLLM model, which can fully capture the long-term and short-term dependencies in the electrochemical time-series data.

[0078] The text acquisition module is used to acquire text data of the battery under test and extract text features. This module connects to the operation and maintenance log database, cloud server, or local storage device via wired or wireless means to collect text data from at least one of the following: operation and maintenance logs, scene annotation text, and technical documents. After cleaning and word segmentation preprocessing, the acquired text data is input into a pre-trained language model for encoding, resulting in text feature vectors rich in semantic information. The pre-trained language model adopts BERT, RoBERTa, or LLaMA series models, which can fully understand the semantics of the text and transform unstructured text data into structured feature representations.

[0079] The image acquisition module is used to acquire image data of the battery under test and extract image features. This module connects to an industrial camera, infrared thermal imager, or battery management system screenshot interface to acquire at least one type of image data, including battery appearance images, electrode microstructure images, separator state images, electrochemical visualization images, and screenshots from the battery management system. After preprocessing such as size normalization and data augmentation, the acquired image data is input into a visual encoder to extract image features. The visual encoder employs a convolutional neural network, a visual Transformer, or a hybrid architecture of both, enabling it to extract multi-level, multi-scale visual features from the images, including macroscopic damage features and microstructural features of the battery appearance.

[0080] The gating module is used to generate corresponding learnable gating coefficients based on temporal features, text features, and image features. This module receives feature vectors output by the temporal acquisition module, text acquisition module, and image acquisition module. For the k-th modality among the temporal modality, text modality, and image modality, the corresponding gating coefficient is generated using the gating coefficient calculation formula in Example 1. The gating coefficient can dynamically adjust the contribution of the corresponding modality in the fusion process, automatically suppress the influence of noise or missing modal data, and dynamically adjust the modal weights according to the battery life cycle stage.

[0081] The attention module, based on temporal, textual, and image features, constrains the cross-modal multi-head attention mechanism by determining the lifetime degradation task-driven correlation coefficient matrix and structure-preserving mask matrix, thereby obtaining cross-modal attention enhancement features for each modality. This module receives feature vectors output from the gating module. First, it maps each modal feature to query vectors, key vectors, and value vectors using a learnable projection matrix. Then, it determines the lifetime degradation task-driven correlation coefficient matrix C and the structure-preserving mask matrix M, calculating the attention weights between modalities. Finally, it weights and sums the value vectors to obtain the enhanced features, which are then integrated into the final modal enhancement features through concatenation and linear mapping under the multi-head mechanism. The correlation coefficient matrix C is generated based on battery degradation indices, quantifying the contribution strength of each feature to the degradation process; the mask matrix M is constructed according to modal characteristics, ensuring that attention establishes associations only within physically interpretable limits.

[0082] The fusion module is used to perform semantic enhancement decoupling fusion based on learnable gating coefficients and cross-modal attention enhancement features to obtain multimodal fusion features. This module receives the gating coefficients output by the gating module and the enhancement features output by the attention module, and calculates the multimodal fusion features using the semantic enhancement decoupling fusion formula in Example 1.

[0083] The prediction module is used to input multimodal fusion features into the regression predictor to obtain the battery remaining life prediction result. This module receives the multimodal fusion features output by the fusion module, first maps each modality enhancement feature to a unified dimensional space through a learnable linear mapping matrix, and integrates them through a concatenation operation to obtain the final fusion feature. Then, the final fusion feature is input into the regression predictor to output the remaining life prediction value. The regression predictor adopts a multilayer perceptron structure, containing two hidden layers, and the output layer is a linear activation function, directly outputting the remaining life prediction value in units of loop count or calendar time.

[0084] Optionally, the above modules are integrated on the same chip, using a system-on-a-chip architecture, including an embedded processor, memory and multiple communication interfaces, which can process multi-source heterogeneous data in real time and output prediction results.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0086] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0087] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A battery lifetime prediction method based on the fusion of time series and multimodal models, characterized in that, The method includes: acquiring multi-source heterogeneous data of the battery under test and extracting multi-source heterogeneous features from it; wherein the multi-source heterogeneous features include electrochemical time-series features, text features, and image features; generating corresponding learnable gating coefficients based on the multi-source heterogeneous features; constraining the cross-modal multi-head attention mechanism by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix based on the multi-source heterogeneous features to obtain cross-modal attention enhancement features for each modality; performing semantic enhancement decoupling fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multi-modal fusion features; and inputting the multi-modal fusion features into a regression predictor to obtain the remaining lifespan prediction result of the battery under test.

2. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 1, characterized in that, The multi-source heterogeneous data includes electrochemical performance data, textual semantic data, and visual image data. After preprocessing, the electrochemical performance data is input into a time-series encoder to extract the electrochemical time-series features characterizing the evolution of the battery's internal state. The textual semantic data is input into a pre-trained language model to extract the textual features characterizing the semantic information of the battery's use and maintenance background. The visual image data is input into a visual encoder to extract the image features characterizing the battery's physical damage and electrochemical behavior.

3. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 2, characterized in that, The electrochemical performance data includes at least one of the following: charge / discharge voltage curves, current, capacity, internal resistance, open-circuit voltage, coulombic efficiency, operating status time-series data, charge rate, discharge depth, cycle count, resting time, and environmental parameters; the textual semantic data includes at least one of the following: operation and maintenance logs, scene annotation text, and technical documents; the visual image data includes at least one of the following: battery appearance image, electrode microstructure image, separator state image, electrochemical visualization image, and screenshot of the battery management system.

4. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 1, characterized in that, The step of generating corresponding learnable gating coefficients based on the multi-source heterogeneous features includes: for the k-th modality among the temporal modality, text modality, and image modality in the multi-source heterogeneous features, calculating the gating coefficient using the following formula. ： ;in, The Sigmoid function restricts the gating coefficient to the range [0,1]. For modal embedding vectors, and For trainable parameters; gating coefficients Used to adjust the effectiveness of corresponding modalities in the fusion process, achieving interpretable weighting at the sample level; These are time-series, text, and image data, respectively.

5. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 1, characterized in that, Based on the multi-source heterogeneous features, the cross-modal multi-head attention mechanism is constrained by determining the lifetime degradation task-driven correlation coefficient matrix and the structure-preserving mask matrix to obtain cross-modal attention enhancement features for each modality; including: features for modality m. and the characteristics of mode n The features of mode m are mapped to a query vector using a learnable projection matrix. Map the features of mode n to key vectors Sum value vector ;in To query the projection matrix; The key projection matrix; The projection matrix is ​​defined; the lifetime degradation task-driven correlation coefficient matrix C is determined to quantify the contribution strength of each feature in mode n to the degradation evolution of mode m at different times; the structure preservation mask matrix M is determined to limit the effective range of cross-modal attention; based on the query vector... The key vector The correlation coefficient matrix C and the structure-preserving mask matrix M are used to calculate the attention weights of mode m on mode n. As shown in the following formula: Where ⊙ denotes element-wise multiplication, and softmax denotes normalization over all features of mode n. The feature dimension of the key vector. The normalized modulation matrix is ​​determined based on the attention weights. Sum value vector The enhanced features are calculated as follows: Under a multi-head mechanism, the query vector will be... Key vector Sum value vector The system is divided into h subspaces, and each head independently computes attention and enhancement features to obtain the outputs of each head. These outputs are then concatenated and linearly mapped to form the final modality enhancement features. 。 6. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 5, characterized in that, The process of generating the lifetime degradation task-driven correlation coefficient matrix C includes: acquiring at least one degradation index among the capacity decay rate, internal resistance growth rate, or remaining lifetime label of the battery under test; mapping the degradation index to a degradation state sequence or degradation spatial distribution, and performing scale alignment on each modal feature according to the degradation state sequence or degradation spatial distribution; applying a perturbation to each feature in the multi-source heterogeneous features, calculating the degree of influence of feature value changes on degradation rate changes, and obtaining the sensitivity of each feature; identifying lifetime-sensitive features or degradation-sensitive features from the multi-source heterogeneous features based on the sensitivity of each feature; for the identified lifetime-sensitive features or degradation-sensitive features, converting their influence on the degradation rate into lifetime response weights through local regression or piecewise fitting methods to obtain the initial correlation coefficient matrix elements. Where i represents the time step index and j represents the feature index; combining the thermal diffusion path, electrochemical reaction active region, and aging propagation direction, the elements of the initial correlation coefficient matrix are physically consistent to obtain the corrected correlation coefficient matrix elements. As shown in the following formula: The modified correlation coefficient matrix is ​​normalized row by row to obtain the correlation coefficient matrix C of the probability distribution, and updated as the training process or degradation state progresses.

7. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 6, characterized in that, The structure-preserving mask matrix M is constructed based on modal characteristics, specifically including: for temporal features, the mask matrix M is constructed based on a window size k determined by thermal diffusion or charge / discharge lag time, allowing attention to be calculated only within the current degradation state and the physically propagable time window, as shown in the following equation: For image or two-dimensional parametric features, the mask matrix M is constructed based on the local aging diffusion neighborhood to maintain the spatial continuity of the degradation morphology and the local topological structure.

8. The battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 1, characterized in that, The step of performing semantic enhancement decoupling and fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multimodal fusion features includes: for the fusion feature vector of the i-th time step or sample, calculating the multimodal fusion features using the following formula. ： ;in, These represent the original features of the temporal, text, and image modalities, respectively. This is a cross-modal attention enhancement vector; , , These are the gating coefficients for the corresponding modes, used to regulate the contribution of the original features; These are the weighting coefficients for the cross-modal enhancement terms.

9. A battery lifetime prediction method based on the fusion of time series and multimodal models according to claim 5, characterized in that, The step of inputting the multimodal fusion features into the regression predictor to obtain the battery remaining life prediction result includes: assuming the enhanced features of the i-th sample in each modality are as follows: , , The enhanced features of each modality are mapped to a unified dimensional space using a learnable linear mapping matrix, and then integrated through a concatenation operation to obtain the final fused features. As shown in the following formula: ;in, ， ， is a learnable linear mapping matrix; ⊕ denotes a vector concatenation operation; For normalization or nonlinear transformation functions; the final fused features Input the regression predictor to generate the remaining useful life prediction for the sample, as shown in the following formula: ;in, This is the regression prediction function.

10. A battery life prediction device based on the fusion of time series and multimodal models, used to implement the battery life prediction method as described in any one of claims 1-9, characterized in that, The device includes: a time-series acquisition module for acquiring electrochemical time-series data of the battery under test and extracting time-series features; a text acquisition module for acquiring text data of the battery under test and extracting text features; an image acquisition module for acquiring image data of the battery under test and extracting image features; a gating module for generating corresponding learnable gating coefficients based on the time-series features, text features, and image features; an attention module for constraining the cross-modal multi-head attention mechanism based on the multi-source heterogeneous features by determining the lifetime degradation task-driven correlation coefficient matrix and structure-preserving mask matrix, thereby obtaining cross-modal attention enhancement features for each modality; a fusion module for performing semantic enhancement decoupling fusion based on the learnable gating coefficients and the cross-modal attention enhancement features to obtain multi-modal fusion features; and a prediction module for inputting the multi-modal fusion features into a regression predictor to obtain the battery's remaining lifespan prediction result.

Images