Multi-modal power battery health prediction method for complex real vehicle scenarios

Abstract

This invention discloses a multimodal power battery health prediction method for complex real-vehicle scenarios, comprising: acquiring electric vehicle battery data from laboratory and real-world environments to construct a baseline SOH sequence, and synthesizing a noisy SOH sequence based on real-world vehicle noise characteristics; extracting segments from the noisy and smooth sequences respectively, and converting them into paired grayscale image training samples; constructing a deep generative prediction model including visual branches, numerical branches, and adaptive fusion modules, training the model using sample pairs, outputting a future SOH trajectory image, and reconstructing it into a numerical prediction sequence through pixel-to-numerical mapping; and finally adapting the laboratory model to real-vehicle scenarios through transfer learning. This invention exhibits good generalization performance for different driving conditions, different vehicle types, and various battery material systems.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery prediction and health management technology, specifically to a multimodal power battery health prediction method for complex real-vehicle scenarios. Background Technology

[0002] State of Health (SOH) of lithium-ion batteries is a core indicator for assessing their remaining lifespan and performance degradation. Existing methods for predicting SOH of lithium batteries can be mainly divided into mechanistic model-based methods and data-driven methods. Mechanism-based methods rely on accurate modeling of the electrochemical reaction processes inside the battery, but their model parameters vary significantly with temperature, charge / discharge rate, and aging stage, leading to complex online parameter identification and difficulty in adapting to the irregular charging behavior and data gaps commonly found in real electric vehicles. While data-driven methods can learn the nonlinear relationship of battery degradation from historical data, they still face the following prominent problems in real-world vehicle scenarios: First, the SOH (State of Harm) sequences collected in real vehicles are noisy and significantly non-stationary, making it easy for traditional time-series regression models to misjudge short-term noise fluctuations as long-term degradation trends. Second, existing methods mostly use numerical sequences as the sole modeling object, failing to explicitly mine and utilize the overall morphology, structural inflection points, and geometric consistency information inherent in the SOH degradation trajectory itself. Third, for multi-source measurement information such as voltage, current, and temperature, existing methods mostly adopt simple splicing or fixed-weight fusion methods, lacking the ability to dynamically adjust the contribution of each information source according to actual noise levels and operating conditions. Finally, most existing methods are developed based on strictly controlled laboratory battery data, whose data distribution differs significantly from data in real-world complex vehicle environments. Directly deploying laboratory-trained models to real-world vehicle scenarios often leads to a severe decrease in prediction accuracy. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a multimodal power battery health prediction method for complex real-vehicle scenarios, which solves the problems of high noise, irregular trajectory, insufficient feature utilization, and poor generalization across battery systems / operating conditions in existing technologies under real electric vehicle operating conditions.

[0004] Technical solution: The multimodal power battery health prediction method for complex real-vehicle scenarios described in this invention includes:

[0005] S1 acquires battery cycling data under controlled laboratory conditions and constructs the corresponding first SOH sequence of battery health status; collects operating data of electric vehicles and on-board BMS (Battery Management System) data under complex real-world conditions and constructs the second SOH sequence.

[0006] S2, based on the noise statistical distribution characteristics of the SOH sequence of real electric vehicles, superimposes synthetic noise on the first SOH sequence to generate a noisy third SOH sequence that simulates the noise characteristics of real vehicles;

[0007] For laboratory scenarios, S3 extracts noise segments from historical periods from the third SOH sequence and extracts smooth segments containing historical periods and future prediction periods from the first SOH sequence; for real electric vehicle scenarios, it extracts noise segments from historical periods from the second SOH sequence, performs trend extraction on the second SOH sequence to obtain the fourth SOH sequence, and extracts smooth segments containing historical periods and future prediction periods from the fourth SOH sequence.

[0008] S4 converts the noise segments and smooth segments in the laboratory scene into corresponding first grayscale images and second grayscale images, respectively, to form image training sample pairs for the controlled laboratory scene; and converts the noise segments and smooth segments in the real electric vehicle scene into corresponding third grayscale images and fourth grayscale images, respectively, to form image training sample pairs for the real electric vehicle scene.

[0009] S5 constructs a computer vision-driven deep generative prediction model, including a visual branch, a numerical branch, and an adaptive fusion module. The visual branch is used to extract and reconstruct the spatial structure features of smooth degradation trajectories from the input noisy SOH grayscale image. The numerical branch is used to extract temporal degradation features from the battery multi-source feature sequence corresponding to historical time periods. The adaptive fusion module is used to dynamically generate fusion weights and perform cross-modal feature fusion and noise suppression based on the output features of the visual branch and the numerical branch.

[0010] S6 uses image training sample pairs and corresponding multi-source feature sequences to train the deep generation prediction model, and the model outputs a smooth SOH trajectory prediction image for the future prediction period.

[0011] S7 converts the SOH trajectory prediction image into a numerical prediction sequence for the future prediction period through pixel-to-numerical mapping and curve reconstruction.

[0012] S8 transfers the lab-trained model to real-world vehicle scenarios using transfer learning techniques.

[0013] Furthermore, in step S2, the distribution parameters of the synthesized noise are determined based on the statistical properties of the residual sequence between the real vehicle SOH sequence and its inherent smoothing trend component, so that the injected noise statistically approximates the real vehicle noise.

[0014] Furthermore, in step S4, converting the SOH sequence fragment into a grayscale image specifically involves mapping the cycle number to the horizontal coordinate of the image, mapping the SOH value to the vertical coordinate of the image, and drawing a trajectory curve at a uniform pixel scale to generate a grayscale image of a fixed size.

[0015] Furthermore, in step S5, the vision branch includes an encoder and a decoder. The encoder is used to perform multi-scale downsampling on the input SOH grayscale image to extract features, and the decoder is used to upsample the features to reconstruct the SOH trajectory image. A cross-layer skip connection is set between the encoder and the decoder. The encoder includes a multi-level module consisting of convolutional layers, activation functions, and downsampling layers. The decoder includes a multi-level upsampling module consisting of transposed convolutional layers and convolutional layers.

[0016] Furthermore, in step S5, the numerical branch uses a bidirectional recurrent neural network to perform temporal encoding on the multi-source feature sequence to extract forward and backward long-term dependencies, and maps the extracted temporal features into feature maps aligned with the deep feature space size of the visual branch; the multi-source feature sequence is extracted from the sequence of one or more parameters among voltage, current, temperature, state of charge (SOC), charging duration, and discharging duration during battery operation, and is aligned with the historical window on the cyclic index or timestamp.

[0017] Furthermore, in step S5, the adaptive fusion module includes a weight generation unit and a fusion unit; the weight generation unit generates dynamic weights for feature fusion based on the features output by the visual branch and the numerical branch; the fusion unit performs a weighted combination of the visual features and temporal features according to the dynamic weights to obtain fused features; the adaptive fusion module further includes an attention enhancement unit, which performs channel attention operations and spatial attention operations on the fused features in sequence to enhance the feature responses related to the degradation trajectory and suppress noise.

[0018] Furthermore, in step S6, the training includes two stages: source domain pre-training and target domain transfer fine-tuning. First, pre-training is performed on samples constructed from laboratory data, and then transfer fine-tuning is performed using a small number of samples constructed from real vehicle data. During the transfer fine-tuning process, all or part of the parameters of the visual branch encoder are frozen, and all or part of the parameters of the numerical branch and the visual branch decoder are updated.

[0019] Further, in step S7, the pixel-value mapping is as follows: the SOH trajectory prediction image is scanned horizontally column by column to determine the pixel position representing the SOH trajectory in each column; according to the preset image coordinates and SOH value calibration relationship, the pixel position is converted into the corresponding SOH value to form a discrete SOH point set; wherein, determining the pixel position representing the SOH trajectory involves determining candidate pixels in each column of pixels based on grayscale thresholds or connectivity rules, and selecting pixels that satisfy the preset trajectory continuity constraint as target pixels; the discrete SOH point set is subjected to curve fitting or interpolation to reconstruct a continuous SOH curve, and the continuous curve is sampled at equal intervals to obtain the SOH value prediction sequence for the future prediction period.

[0020] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0021] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the program is executed by the processor, it implements any of the methods described herein.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) By performing trend-residual separation on the real vehicle SOH observation signal and constraining the noise injection intensity with residual statistical characteristics, the training disturbance is closer to the vehicle measurement fluctuation, thereby reducing the model's sensitivity to random fluctuations and improving the stability of the prediction curve; (2) The SOH sequence is converted into a grayscale image according to a unified coordinate rule, so that the model can learn the consistency of degradation morphology, inflection point and low-frequency trend information from the spatial structure level, rather than just performing numerical fitting, thereby improving the ability to identify and predict long-term degradation trajectories; (3) The visual branch extracts the spatial degradation mode of the SOH image, and the numerical branch extracts the temporal degradation mode composed of multi-source measurement signals such as voltage, current, temperature, SOC, and charging time, and dynamically adjusts the contribution of the two types of modes through the cross-branch adaptive weight mechanism, so that the model still maintains good generalization performance under different vehicles, different working conditions and different noise levels; (4) The prediction image is converted into a discrete SOH point set through pixel-numerical mapping, and the continuous SOH curve is obtained by curve fitting / interpolation and then the future q is sampled and output at equal intervals. The periodic SOH sequence facilitates direct integration with BMS health management and lifespan decision-making. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention;

[0024] Figure 2This is a schematic diagram of the noise statistical analysis and denoising trend construction process of the real vehicle-mounted SOH sequence of the present invention;

[0025] Figure 3 This is a schematic diagram of the structure of the deep generative prediction model CPDG of the present invention;

[0026] Figure 4 This is a schematic diagram of the pixel-to-value mapping and SOH numerical sequence reconstruction process of the model output image of the present invention;

[0027] Figure 5 This invention relates to an ablation study of training strategies, multimodal fusion, and time window settings. (a) shows a comparison of SOH prediction performance under three training strategies: transfer learning (with fine-tuning), transfer learning (without fine-tuning), and training based solely on real vehicle data. (b) shows the prediction results of different model structures: visual branch only, numerical branch only, and the proposed multimodal fusion model. (c) shows the impact of historical window length k and prediction step size q on prediction accuracy (averaged across all datasets).

[0028] Figure 6 This paper compares the absolute prediction error distribution of each model on three real electric vehicle datasets of the present invention; where (a) represents real vehicle dataset 2, (b) represents real vehicle dataset 3, and (c) represents real vehicle dataset 4. The violin plot shows the empirical distribution of the absolute error of the proposed method and 8 baseline models, and the corresponding box statistics (median and interquartile range, etc.) are superimposed. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, this invention provides a multimodal power battery health prediction method for complex real-vehicle scenarios. This method uses laboratory aging cycle data as source domain data and real vehicle battery system data as target domain data. Through the processing link of "on-board noise statistical modeling - noisy image synthesis and slicing - multimodal generative model learning - pixel mapping and sequence reconstruction - migration fine-tuning deployment", it achieves robust prediction of SOH for the next q cycles and generalization across vehicles and operating conditions.

[0031] For ease of implementation, the historical window length is denoted as k, the prediction step size as q, and the sliding window step size as s. Through a series of experimental verifications, the optimal step size can be selected. , To balance prediction accuracy and deployment cost.

[0032] S101 (Benchmark Sequence): Obtain laboratory battery aging cycle data to construct a smooth SOH degradation benchmark sequence. The laboratory data includes recorded information such as capacity, temperature, and time during battery cycling. SOH can be characterized by capacity and forms a degradation trajectory that evolves with the number of cycles. Let the original SOH sequence obtained under laboratory conditions be represented as: in, This represents the SOH value corresponding to the t-th cycle. To obtain a smooth degradation baseline sequence that can be used for supervised learning, outlier removal, missing value interpolation, and smoothing are performed sequentially on the original SOH sequence to obtain a smoothed SOH degradation baseline sequence: The smoothed SOH degradation baseline sequence is used to characterize typical degradation trends under controlled conditions and serves as a baseline input for subsequent noise injection and reference image generation.

[0033] S102 (Noise Statistics): Based on the noise statistical characteristics of the actual vehicle-mounted SOH observation sequence, determine the noise distribution type and parameter range. Let the vehicle-mounted SOH observation sequence be: The onboard SOH observation sequence is subjected to trend-noise separation processing, preferably using empirical mode decomposition (EMD) to extract low-frequency trend components, resulting in a smoothed trend sequence. And define the noise residual sequence as: Further statistical modeling of the noisy residual sequence is performed, and its mean and variance are expressed as follows: Based on this, the noise distribution type and parameter range are determined, which are used to constrain the perturbation injection process on the subsequent source domain reference sequence, so that the synthesized noise can statistically approximate the real vehicle-mounted SOH fluctuation characteristics.

[0034] S103 (Perturbation Injection): A perturbation is superimposed on the smoothed SOH degradation reference sequence according to the statistical characteristics of vehicle noise to generate a noisy SOH sequence. Specifically, based on the noise distribution and parameter range determined in step S102, a perturbation is injected into the smoothed SOH reference sequence. Random perturbation injection was performed on the sequence to obtain a noisy SOH sequence: in, This represents a noise distribution model consistent with the statistical characteristics of vehicle noise.

[0035] S104 (Window Slicing): A sliding window is used to slice the noisy SOH sequence and the smoothed SOH baseline sequence separately. A historical segment of length k is extracted from the noisy SOH sequence, and a segment of length k is extracted from the smoothed baseline sequence that is aligned with the starting point of the historical segment. The reference segment includes historical intervals and future prediction intervals. Preferably, the sliding window step size s is used to control the sample overlap rate and the number of samples to achieve a balance between training overhead and sample diversity.

[0036] S105 (calibration and mapping) aligns and scales historical and reference segments, converting them into noisy SOH images and smoothed reference SOH images, respectively. Specifically, a sequence-image mapping operator is defined. Mapping the SOH numerical sequence into a two-dimensional trajectory image: In this model, the horizontal axis pixels represent the cyclic index, and the vertical axis pixels represent the SOH amplitude. After uniform calibration and scaling, the images are converted into grayscale images, thus forming training sample pairs consisting of a noisy input image and a smooth reference image.

[0037] S106 (Feature Sequence): Obtain the vehicle-mounted multi-source measurement signals corresponding to historical segments, and construct a numerical feature sequence of length k and dimension d: The numerical features are preferably extracted from signals such as voltage, current, temperature, SOC, and charging time trajectory through engineering methods, and aligned with the historical window on the cyclic index or timestamp.

[0038] S107 (Transfer Training): Construct a computer vision-driven deep generative prediction model (CPDG) and perform transfer training, consisting of "source domain pre-training – vehicle-mounted fine-tuning". The CPDG employs a dual-channel structure to process different modal inputs: one is a visual data channel, and the other is a numerical data channel. A cross-branch fusion mechanism is used to jointly model the features of both modalities. The overall model can be represented by the following multimodal mapping relationship:

[0039]

[0040] in, This represents a noisy SOH image input. This represents a multi-source measurement feature sequence aligned with a historical window. For model parameters, The image shows the predicted SOH trajectory.

[0041] (11) Visual data channel modeling

[0042] The visual data channel is used to extract spatial structural features of degraded trajectories from noisy SOH images. Preferably, the visual channel uses a U-shaped encoder-decoder structure as its backbone, and its encoding end extracts multi-scale visual features through a multi-level downsampling convolution module.

[0043]

[0044] in, , Indicates the first Layer-based encoding features. The decoder recovers spatial resolution step-by-step through multi-level upsampling convolutional modules and introduces corresponding-scale encoded features via skip connections to achieve high-resolution reconstruction of the SOH degenerate morphology.

[0045]

[0046] (12) Numerical data channel modeling

[0047] The numerical data channel is used to extract time-series features related to SOH degradation from a multi-source measurement feature sequence. Let the numerical feature sequence be:

[0048]

[0049] Preferably, the numerical channel uses a bidirectional recurrent neural network to encode the feature sequence, and its forward and backward hidden states are represented as follows:

[0050]

[0051]

[0052] The two are then concatenated to obtain the temporal feature representation:

[0053]

[0054] The temporal features are further mapped to a spatially alignable feature map through an upsampling alignment module:

[0055]

[0056] in, This represents the feature projection and upsampling mapping operator.

[0057] (13) Cross-branch adaptive weighted fusion

[0058] To achieve joint modeling of features from different modalities, a cross-branch adaptive weighted attention mechanism is introduced. Fusion weights are generated based on global descriptions of visual and numerical features.

[0059]

[0060] in, This represents the global average pooling operator. This represents the Sigmoid activation function. The two modal features are weighted and combined according to the fusion weights to obtain the fused features:

[0061]

[0062] (14) Enhanced attention to passageways and spaces

[0063] To further highlight the effective response associated with the SOH degradation trajectory and suppress noise interference, a channel and spatial saliency enhancement mechanism is applied before the fused features enter the decoding stage. Preferably, the attention enhancement can adopt a CBAM structure, where the channel attention and spatial attention are expressed as follows:

[0064]

[0065]

[0066] in, and These represent the channel attention mapping and spatial attention mapping operators, respectively. This indicates element-wise multiplication.

[0067] (15) Decoding, reconstruction and transfer training strategies

[0068] The attention-enhanced fused features are input into the visual decoding module to generate a predicted SOH trajectory image.

[0069]

[0070] Regarding the training strategy, it is preferable to first pre-train using source domain lab sample pairs, by minimizing the loss function between the predicted image and the smoothed reference image:

[0071]

[0072] The model learns a generative mapping from "noisy SOH image to smooth SOH trajectory image". Subsequently, transfer fine-tuning is performed using sample pairs constructed from a small amount of real-world vehicle data to adapt to the data distribution differences in the target vehicle domain. Preferably, during the vehicle fine-tuning stage, all or some parameters of the visual encoder are frozen, and only all or some parameters of the numerical channel, cross-modal fusion module, and visual decoder are updated. This ensures rapid adaptation to vehicle domain differences while maintaining the prior stability of the degenerate morphology.

[0073] S108 (Sequence Output): The trained CPDG is used to generate a smooth SOH trajectory image for the next q periods, and the predicted SOH sequence for the next q periods is output through pixel-to-value mapping and curve reconstruction. In one embodiment, the pixel-to-value mapping process includes: determining the trajectory pixel positions column by column along the horizontal direction of the image (e.g., using the position of the minimum pixel value in each column as the trajectory point), converting the pixel coordinates into corresponding SOH values ​​according to the image calibration relationship, thereby obtaining a discrete SOH point set based on the image; subsequently, curve fitting (e.g., spline fitting) is performed on the discrete point set and equidistant sampling is performed to obtain an SOH sequence containing historical reconstruction segments and future prediction segments, where the future prediction segments are the predicted SOH results for the next q periods.

[0074] In the target domain vehicle dataset, the input image can be generated from the vehicle's noisy SOH curve, while the output supervision image can be obtained from the EMD smoothing trend transformation of the vehicle SOH sequence, thus completing the transfer fine-tuning supervised training under a small number of annotations.

[0075] To address the issues of noise fluctuations and a lack of high-quality smooth labels in real-world vehicle-mounted SOH sequences, this embodiment performs two processes on the target domain vehicle-mounted data: "noise statistical parameter extraction" and "denoising trend label construction." The former is used to constrain... Figure 1 The noise perturbation injection in step S103 ensures that the source domain synthesized noise and the vehicle noise are statistically consistent; the latter is used to construct... Figure 1 The transfer fine-tuning supervision signal in step S107 enables the model to complete domain adaptation training under conditions of limited on-vehicle data. Specifically, it includes the following steps.

[0076] S201 (Sequence Acquisition): Acquire the onboard SOH observation sequence of the target vehicle. The onboard SOH observation sequence is the SOH time series acquired or estimated by the battery management system during actual vehicle operation, and its form is as follows:

[0077]

[0078] in, This represents the onboard SOH observation value corresponding to the t-th cycle or time index. The sequence typically includes real-world characteristics such as measurement errors, operating condition disturbances, and uneven sampling. Preferably, to ensure the stability of subsequent trend extraction and statistical estimation, the onboard SOH observation sequence undergoes basic normalization processing, including outlier removal, cycle index or timestamp unification, and one or more of the following: missing segment interpolation completion or marker masking, to obtain a normalized observation sequence. .

[0079] S202 (Trend-Residual Separation): Trend-residual separation is performed on the vehicle-mounted SOH observation sequence obtained in step S201 to obtain a smoothed trend sequence and a residual noise sequence. In one embodiment, EMD is used to decompose the vehicle-mounted SOH observation sequence, representing it as the sum of several intrinsic mode functions (IMFs) and a residual term:

[0080]

[0081] Among them, the low-frequency components and residual terms constitute a smooth degradation trend sequence:

[0082]

[0083] And define the residual noise sequence as:

[0084]

[0085] Thus, a smooth trend sequence is obtained. With residual noise sequence .

[0086] S203 (Residual Statistical Characterization): The residual noise sequence obtained in step S202 is statistically characterized to extract statistics describing the intensity and distribution of vehicle-mounted noise. Preferably, the central level and discrete level of the residual noise sequence are calculated, and expressed as follows:

[0087]

[0088] Furthermore, statistical measures such as skewness, kurtosis, and extreme value range can be calculated to characterize the distribution symmetry, tail thickness, and fluctuation boundaries of the residual noise. More preferably, the residual statistics can be estimated piecewise according to different vehicles, operating conditions, or aging stages to obtain a set of noise parameters.

[0089]

[0090] To more closely resemble the characteristics of real vehicle noise.

[0091] S204 (Noise Parameter Determination and Output): Based on the statistics obtained in step S203, the noise distribution type and parameter range are determined and output as noise injection constraints. Specifically, a matching noise distribution model is selected based on the statistical characteristics of the residual noise. And give the parameter range:

[0092]

[0093] The parameter range is used to limit Figure 1In step S103, the amplitude and shape of the perturbation injection are implemented on the source domain smoothed SOH degradation benchmark sequence to ensure that the injected perturbation is statistically consistent with the real vehicle noise. Preferably, when there is data from multiple vehicles, noise parameter ranges can be output separately for different vehicles; when the number of vehicle samples is limited, a unified parameter range can be output using global statistical results for rapid deployment.

[0094] S205 (Smoothing Trend as Reference Label for Vehicle Fine-tuning): The smoothed trend sequence obtained in step S202 is used as the reference label for the transfer fine-tuning stage, and paired training samples in the vehicle domain are generated. Specifically, the smoothed trend sequence is coordinate-calibrated and scale-unified according to the same image-based rules as the source domain, and converted into a target SOH image as the supervision label for transfer fine-tuning; simultaneously, the vehicle SOH sequence is converted into an input SOH image according to the same rules, thus forming a "input image - target image" sample pair in the vehicle domain, used for... Figure 1 Step S107 involves transfer fine-tuning training. More preferably, the vehicle-mounted multi-source measurement feature sequence can be aligned with the aforementioned vehicle-mounted domain SOH image samples by cyclic index or timestamp, and simultaneously used as numerical channel input during fine-tuning training to enhance the model's adaptability and generalization ability to real operating conditions and irregular charging behavior.

[0095] Reference Figure 3 As shown, this embodiment constructs a computer vision-driven depth generative prediction model (CPDG) to achieve denoising reconstruction and multi-step prediction of the SOH trajectory under real vehicle noise interference and operating condition fluctuations. The CPDG adopts a dual-input structure of visual and numerical modalities, and achieves information complementarity and noise suppression through cross-modal adaptive fusion, layer-by-layer injection, and attention enhancement mechanisms. Its composition and working process are as follows.

[0096] Input Interface

[0097] Input 1: Let the noisy SOH image be represented as: The noisy SOH image is composed of... Figure 1 Step S105 generates information to provide the spatial morphology, boundary structure, and local undulation characteristics of the SOH degradation trajectory within the historical window.

[0098] Input 2: Let the multi-source feature sequence be represented as: The multi-source feature sequence is composed of Figure 1 Step S106 generates a time-driven information and operating condition-related information that is strictly aligned with the historical window to provide information on degradation evolution.

[0099] S301 (Visual Encoding) performs visual encoding on input 1 and outputs a multi-scale visual feature representation. Preferably, the visual encoding module uses a U-shaped encoder structure as the backbone and extracts multi-scale spatial features from shallow to deep through a multi-level downsampling convolution module. Its encoding process can be represented as follows:

[0100]

[0101] in, , Indicates the first Layer-encoded features. More preferably, a skip connection is set between the encoding and decoding ends to pass shallow detail features to the reconstruction path, thereby enhancing the continuity of trajectory boundaries and the ability to preserve shape.

[0102] S302 (numerical encoding) performs temporal encoding on input 2 and outputs a temporal feature representation. Preferably, the numerical encoding module uses a bidirectional loop structure to encode the multi-source feature sequence, and its forward and backward hidden states are represented as follows:

[0103]

[0104]

[0105] The two are then concatenated to obtain the temporal feature representation:

[0106]

[0107] This yields a set of high-level temporal features associated with SOH degradation. In other implementations, gated loop structures, temporal convolutions, or other equivalent temporal coding structures can be used instead, as long as they can output stable temporal representations.

[0108] S303 (Feature Mapping): This step performs mapping and alignment processing on the temporal features output from step S302 to generate mapped features that can be fused with visual features. The feature mapping module achieves cross-modal dimensional consistency through feature projection and upsampling alignment operators. Its mapping process can be represented as follows:

[0109]

[0110] in, This indicates that a linear mapping combined with upsampling, transposed convolutional mapping, or equivalent mapping operators is used to expand temporal features into feature maps that can be aligned with visual features in terms of spatial scale and channel dimension.

[0111] S304 (Adaptive Fusion and Layer-by-Layer Injection) adaptively fuses the multi-scale visual features output from step S301 with the mapping features output from step S303 to generate fused features, which are then introduced into the visual branch decoding stage via layer-by-layer injection. The adaptive fusion module includes a weight generation unit and a fusion unit, wherein the fusion weights are generated based on the global descriptions of the two modalities.

[0112]

[0113] in, This represents the global average pooling operator. This represents the Sigmoid activation function. The fused features are then obtained by weighting and combining the values ​​according to the fusion weights.

[0114]

[0115] In the layer-by-layer injection implementation, the fused features are injected at one or more upsampling reconstruction layers at the decoding end according to the hierarchical structure of visual decoding. The injection method can be expressed as follows:

[0116] in, This represents splicing injection, element-wise addition injection, or gated injection operators. Through this design, global degradation trend information is introduced in deeper decoding stages, while local morphological correction information is introduced in shallower decoding stages, thereby dynamically balancing the contributions of different modes under varying noise levels and operating conditions.

[0117] S305 (Attention Enhancement): Attention enhancement is performed on the decoded features formed after injection at each layer in step S304 to highlight the effective responses related to the SOH trajectory and suppress noise responses. Preferably, attention enhancement includes channel recalibration and spatial saliency enhancement, the processes of which are respectively expressed as follows:

[0118]

[0119]

[0120] in, and These represent the channel attention mapping and spatial attention mapping operators, respectively. This represents element-wise multiplication. The attention enhancement module can be implemented using SE, CBAM, or equivalent structures.

[0121] S306 (Visual Decoding and Reconstruction) performs visual decoding and upsampling reconstruction on the decoded features after layer-by-layer injection and attention enhancement, outputting a predicted SOH trajectory image. Preferably, the visual decoding module and the visual encoding module correspond at the hierarchical scale, using upsampling, transposed convolution, or other equivalent reconstruction modules to restore spatial resolution step by step; and fine-grained features at the encoding end are introduced through skip connections to ensure that the output trajectory is consistent with the historical structure in morphology and has good continuity. This can be represented as:

[0122]

[0123] in, This indicates the visual decoding module.

[0124] S307 (Output): Outputs the SOH trajectory image for the next q cycles. The output image is a smoothed prediction of the SOH degradation trajectory for the next q cycles. Subsequent pixel-to-numerical mapping and curve reconstruction processes can convert it into an SOH prediction sequence for use in applications such as battery health management, lifespan assessment, and maintenance decision-making.

[0125] To convert the future SOH trajectory image output by CPDG into a numerical SOH prediction sequence that can be directly used by the Battery Management System (BMS), this embodiment provides a pixel-to-numerical mapping and curve reconstruction method. This method assumes that "the image coordinates are consistent with the SOH calibration relationship," and extracts trajectory pixels column by column, performs inverse calibration conversion, and reconstructs a continuous curve, ultimately outputting a future q-period SOH prediction sequence. The specific steps include the following.

[0126] S401 (Trajectory Image): Obtain the predicted SOH trajectory image output by CPDG as the object to be decoded, denoted as: Among them, the horizontal pixel coordinates Used to characterize the cyclic index or equivalent time index, vertical pixel coordinates Used to characterize the location of SOH amplitude. The size, coordinate range, and calibration rules of the predicted SOH trajectory image are consistent with the sample generation rules in the training phase to ensure the consistency and repeatability of subsequent pixel-to-numerical conversions. Preferably, before entering column-by-column scanning, the predicted image can be... Perform lightweight preprocessing operators The preprocessed image is obtained:

[0127]

[0128] The preprocessing operators may include one or more of thresholding, de-isolation, or morphological operations, but do not change the main structure of the trajectory.

[0129] S402 (column-by-column scan): Scans column by column along the horizontal pixel coordinates of the predicted SOH trajectory image, obtaining a set of candidate trajectory pixels for each column x:

[0130]

[0131] in, This represents the set of trajectory discrimination conditions consisting of grayscale thresholds, connected component features, or combinations thereof. Through the above column-by-column scanning, a sequence of candidate pixels covering the entire prediction interval is obtained. .

[0132] S403 (pixel localization), in each column of candidate pixel set The position of the target pixel in the trajectory of this column is determined. This forms a sequence of trajectory pixels arranged in columns:

[0133]

[0134] Preferably, the determination of the target pixel position follows the following constraints:

[0135] 1) Response significance constraint:

[0136] 2) Trajectory continuity constraints:

[0137] in, This represents the allowed threshold for changes in the trajectory of adjacent columns. More preferably, when a single column presents multiple solutions or a risk of local breakage, a sequence optimization operator can be used across the entire column. Find a globally continuous optimal pixel path:

[0138]

[0139] in, This represents the energy function that combines the response intensity with the continuity cost.

[0140] S404 (Numerical Conversion): Based on the calibration rules described in step S401, the vertical pixel coordinates of the trajectory pixel sequence obtained in step S403 are converted back into SOH values ​​to obtain a discrete SOH point set.

[0141]

[0142] in, This represents the fixed pixel-SOH inverse calibration mapping function determined during the training phase. Preferably, when image calibration uses a linear mapping, the inverse calibration relationship can be expressed as:

[0143]

[0144] This yields a discrete set of SOH points arranged in columns. To reduce the impact of pixel quantization errors and local positioning fluctuations, a curve reconstruction operator can be applied to the discrete SOH point set. A continuous and smooth SOH prediction curve was obtained:

[0145]

[0146] in, It may include one or more of spline interpolation, local regression, or robust smoothing.

[0147] S405 (Output Sequence): The reconstructed SOH curve obtained in step S404 is sampled at equal intervals according to the cyclic index, and the predicted SOH sequence for the future q periods is output.

[0148] Please refer to the attached diagram for further explanation:

[0149] like Figures 1-2 As shown, the source domain noisy sample construction and vehicle noise consistency constraints (refer to...) Figure 1 , Figure 2 Laboratory cycle life data are typically collected under controlled conditions, resulting in smooth SOH degradation trajectories with weak random fluctuations. However, real-world vehicle-mounted SOH sequences are affected by measurement errors, operating condition fluctuations, and irregular sampling, exhibiting stronger short-term fluctuations. Directly training and deploying laboratory data in vehicle-mounted scenarios can easily lead to prediction biases caused by domain differences. Therefore, this embodiment adopts a "source domain noisy synthesis driven by vehicle noise statistics" strategy: on the one hand, based on… Figure 2 The steps shown perform trend-residual separation and residual statistical modeling on the target domain vehicle-mounted SOH sequence to obtain the statistical characteristics and parameter range of vehicle-mounted noise. On the other hand, using the above statistical results, a perturbation injection is applied to the source domain smoothed SOH degradation benchmark sequence to generate a noisy SOH sequence that is statistically consistent with vehicle-mounted noise, thus pre-covering the vehicle-mounted noise distribution and observed fluctuation patterns during the training phase. Through this strategy, while maintaining the dominant information of the "true degradation trend," the source domain samples are superimposed with random perturbations that "conform to vehicle-mounted statistical characteristics," providing a controllable and scalable data foundation for the subsequent model learning of "denoising and generating mappings."

[0150] SOH sequence imaging and vehicle-mounted multi-source feature alignment construction (see reference) Figure 1 After constructing the noisy SOH sequence, this embodiment generates paired training samples using a sliding window mechanism: the noisy sequence provides a historical segment of length k as input, and the smoothed baseline sequence provides a segment of length k aligned with the same starting point. The reference fragment is used as the supervision output (refer to) Figure 1 (S104–S105). To ensure cross-sample comparability, it is preferable to perform unified coordinate calibration and scale unification on different segments, mapping the cyclic index to horizontal pixel coordinates and the SOH amplitude to vertical pixel coordinates, and performing trajectory pixelation under a unified coordinate system, ultimately obtaining a sample pair of "noisy SOH image - smooth reference SOH image". Meanwhile, to enhance adaptability to real-world conditions, this embodiment further constructs an onboard multi-source measurement feature sequence aligned with the historical window (refer to...). Figure 1 (S106). The feature sequence is preferably extracted by engineering from signals such as voltage, current, temperature, SOC, and charge / discharge duration, and is strictly aligned with the image input by cyclic index or timestamp, so that the model can simultaneously utilize "degenerative morphology information" (image) and "operating condition driving information" (numerical sequence), thereby improving robustness in scenarios where operating condition changes and observation noise coexist.

[0151] Explanation of the cross-modal fusion mechanism and noise resistance sources of the CPDG model (see reference) Figure 3 ):like Figure 3 As shown, CPDG adopts a dual-input, dual-branch structure: the visual branch is used to extract multi-scale spatial representations from noisy SOH images and reconstruct the output trajectory image at the decoding end; the numerical branch is used to extract temporal degradation representations from multi-source measurement feature sequences and convert them into representations that can be aligned with visual features through feature mapping to support cross-modal fusion. To enable the model to maintain stable output under different noise levels and different operating conditions, this embodiment preferably introduces an "adaptive fusion + layer-by-layer injection" mechanism: the fusion module generates dynamic fusion weights based on the global statistics or key responses of the two types of modal features to reflect the relative credibility of different modalities under the current conditions; and injects the fusion features layer by layer at one or more scale levels at the decoding end, so that the deep decoding stage focuses more on stabilizing the global degradation trend, and the shallow decoding stage focuses more on correcting the local morphology and boundary continuity, thereby achieving both "trend stability" and "morphological consistency" in the generated results. Furthermore, to suppress noise response and highlight the trajectory backbone, this embodiment can configure channel recalibration and spatial saliency enhancement modules (such as SE, CBAM, or equivalent structures) during the fusion and decoding stages to enhance the effective response related to the trajectory and reduce background noise interference. Through the above structural design, CPDG can not only achieve "noisy → smooth" generation mapping, but also improve its anti-interference capability against complex on-board fluctuations through cross-modal complementarity and attention enhancement.

[0152] Training deployment methods and migration fine-tuning strategies are explained (refer to...) Figure 1 , Figure 2This embodiment employs a two-stage training strategy: the first stage is source domain pre-training, which uses paired samples constructed from laboratory data to learn the generative mapping from "noisy SOH image → smoothed SOH trajectory image," thereby obtaining general degradation morphology priors and denoising capabilities; the second stage is target domain transfer fine-tuning, which uses a small amount of real vehicle-mounted data to construct vehicle-mounted domain sample pairs (input is the vehicle-mounted observed SOH image, supervision is the smoothed label image after trend extraction, referencing...). Figure 2 The model is updated for domain adaptation (S201–S205), enabling it to quickly align with the noise distribution and operating conditions of the target vehicle domain while maintaining morphological prior stability. In a preferred embodiment, during the fine-tuning stage, all or some parameters of the visual encoder are frozen, while focusing on updating all or some parameters of the numerical branch, fusion module, and visual decoder to achieve more stable adaptation results and reduce the risk of overfitting with a smaller on-board sample size. Through the above transfer training mechanism, this method can achieve cross-vehicle and cross-operating condition generalization for real-world deployments under "limited on-board supervision".

[0153] Pixel-to-numerical decoding output and BMS availability description: (e.g.) Figure 4 As shown, this embodiment converts the SOH trajectory image output by CPDG into a numerical SOH sequence that can be directly used by BMS through pixel-to-numerical mapping and curve reconstruction. This process is based on consistent coordinate calibration rules during the training phase: the trajectory pixel positions are extracted by scanning column by column, and a discrete SOH point set is obtained by reverse calculation according to the calibration relationship. Curve reconstruction and equidistant sampling are then performed on the point set to complete the closed-loop connection from "image prediction results" to "engineering usable indicators." Therefore, BMS can directly conduct health monitoring, trend assessment, and strategy triggering applications based on future SOH prediction sequences.

[0154] Experimental verification, comparison results and explanation of beneficial effects (refer to) Figure 5 – Figure 6 The method of this invention has been validated on laboratory recurrent data and multiple real-world vehicle datasets. Overall results show that the model can stably learn denoised generation mappings and maintain good prediction consistency on source domain data; on target domain vehicle data, a small amount of transfer fine-tuning can significantly improve domain adaptation, and it maintains robust prediction performance across different vehicles and operating conditions.

[0155] Ablation verification ( Figure 5 ): Figure 5This is used to quantify the necessity and contribution of the key design of this invention. Training strategy ablation shows that "two-stage training (source domain pre-training + target domain fine-tuning)" is superior to training only in the source domain or only in the target domain; modal ablation shows that the visual branch or numerical branch alone has insufficient information, and cross-modal fusion results in smaller errors and stronger stability; parameter ablation shows that there is a trade-off between accuracy and coverage between the historical window length k and the prediction step size q, and the parameter configuration selected in this embodiment can achieve a good balance between deployment cost and prediction performance.

[0156] Comparative verification ( Figure 6 ): Figure 6 The results of this method and a multi-class comparison method are presented on representative vehicles / datasets. The predicted trajectory of this method can more closely track the smooth trend of the vehicle's SOH, especially in the case of strong noise, non-stationary degradation, or late aging stages, it still maintains good trend consistency and output smoothness, demonstrating the advantages of the multimodal generation paradigm and transfer adaptation strategy in real vehicle scenarios.

[0157] In summary, this invention, through its overall scheme of "construction of source domain noisy samples with vehicle noise statistical consistency constraints, image-numerical cross-modal fusion generation and prediction, target domain migration fine-tuning adaptation, and pixel-numerical engineering output," can achieve robust prediction of SOH under real vehicle noise interference and irregular operating conditions, and maintains good generalization ability in cross-vehicle and cross-operating condition scenarios, thus having significant engineering application value.

Claims

1. A multimodal power battery health prediction method for complex real-world vehicle scenarios, characterized in that, include: S1 acquires battery cycling data under controlled laboratory conditions and constructs the corresponding first SOH sequence of battery health status; collects operating data of electric vehicles and on-board BMS data under complex real-world conditions and constructs the second SOH sequence; S2, based on the noise statistical distribution characteristics of the SOH sequence of real electric vehicles, superimposes synthetic noise on the first SOH sequence to generate a noisy third SOH sequence that simulates the noise characteristics of real vehicles; For laboratory scenarios, S3 extracts noise segments from historical periods from the third SOH sequence and extracts smooth segments containing historical periods and future prediction periods from the first SOH sequence; for real electric vehicle scenarios, it extracts noise segments from historical periods from the second SOH sequence, performs trend extraction on the second SOH sequence to obtain the fourth SOH sequence, and extracts smooth segments containing historical periods and future prediction periods from the fourth SOH sequence. S4 converts the noisy segments and smooth segments in the laboratory scene into corresponding first grayscale images and second grayscale images, respectively, to form image training sample pairs of the controlled laboratory scene; Noisy segments and smooth segments in real electric vehicle scenes are converted into corresponding third grayscale images and fourth grayscale images, respectively, to form image training sample pairs of real electric vehicle scenes; S5 constructs a computer vision-driven deep generative prediction model, including a visual branch, a numerical branch, and an adaptive fusion module. The visual branch is used to extract and reconstruct the spatial structure features of smooth degradation trajectories from the input noisy SOH grayscale image. The numerical branch is used to extract temporal degradation features from the battery multi-source feature sequence corresponding to historical time periods. The adaptive fusion module is used to dynamically generate fusion weights and perform cross-modal feature fusion and noise suppression based on the output features of the visual and numerical branches. S6 uses image training sample pairs and corresponding multi-source feature sequences to train the deep generation prediction model, and the model outputs a smooth SOH trajectory prediction image for the future prediction period. S7 converts the SOH trajectory prediction image into a numerical prediction sequence for the future prediction period through pixel-to-numerical mapping and curve reconstruction. S8 transfers the lab-trained model to real-world vehicle scenarios using transfer learning techniques.

2. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S2, the distribution parameters of the synthesized noise are determined based on the statistical properties of the residual sequence between the real vehicle SOH sequence and its inherent smoothing trend component, so that the injected noise approximates the real vehicle noise in terms of statistical properties.

3. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S4, converting the SOH sequence fragment into a grayscale image specifically involves mapping the cycle number to the horizontal coordinate of the image, mapping the SOH value to the vertical coordinate of the image, and drawing a trajectory curve at a uniform pixel scale to generate a grayscale image of a fixed size.

4. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S5, the vision branch includes an encoder and a decoder. The encoder is used to perform multi-scale downsampling on the input SOH grayscale image to extract features, and the decoder is used to upsample the features to reconstruct the SOH trajectory image. A cross-layer skip connection is set between the encoder and the decoder. The encoder contains a multi-level module consisting of convolutional layers, activation functions and downsampling layers. The decoder contains a multi-level upsampling module consisting of transposed convolutional layers and convolutional layers.

5. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S5, the numerical branch uses a bidirectional recurrent neural network to perform temporal encoding on the multi-source feature sequence to extract forward and backward long-term dependencies, and maps the extracted temporal features into feature maps aligned with the size of the deep feature space of the visual branch. The multi-source feature sequence is extracted from the sequence of one or more parameters of voltage, current, temperature, state of charge (SOC), charging duration, and discharging duration during battery operation, and aligned with the historical window on the cyclic index or timestamp.

6. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S5, the adaptive fusion module includes a weight generation unit and a fusion unit; the weight generation unit generates dynamic weights for feature fusion based on the features output by the visual branch and the numerical branch. The fusion unit performs a weighted combination of visual features and temporal features based on dynamic weights to obtain fused features; The adaptive fusion module further includes an attention enhancement unit for sequentially performing channel attention and spatial attention operations on the fused features to enhance the feature responses associated with the degenerate trajectory and suppress noise.

7. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S6, the training includes two stages: source domain pre-training and target domain transfer fine-tuning. First, pre-training is performed on samples constructed from laboratory data, and then transfer fine-tuning is performed using a small number of samples constructed from real vehicle data. During migration fine-tuning, freeze all or some parameters of the visual branch encoder and update all or some parameters of the numerical branch and visual branch decoder.

8. The multimodal power battery health prediction method for complex real-vehicle scenarios according to claim 1, characterized in that, In step S7, the pixel-value mapping is as follows: the SOH trajectory prediction image is scanned horizontally column by column to determine the pixel position representing the SOH trajectory in each column; according to the preset image coordinates and SOH value calibration relationship, the pixel position is converted into the corresponding SOH value to form a discrete SOH point set; wherein, determining the pixel position representing the SOH trajectory involves determining candidate pixels in each column of pixels based on grayscale thresholds or connectivity rules, and selecting pixels that satisfy the preset trajectory continuity constraint as target pixels; the discrete SOH point set is subjected to curve fitting or interpolation to reconstruct a continuous SOH curve, and the continuous curve is sampled at equal intervals to obtain the SOH value prediction sequence for the future prediction period.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1-8.

Images