Automobile power battery health monitoring method and system based on multi-modal enhancement

Abstract

The application belongs to the technical field of power battery safety monitoring, and particularly relates to a kind of automobile power battery health monitoring method and system based on multi-modal enhancement. The original sensing data of the vehicle is collected and preprocessed;The original sensing data after preprocessing is subjected to feature construction, obtaining multi-dimensional features including time series main features, statistical features and signal energy and spectral entropy;According to the vehicle working condition feature vector, the current working condition of the vehicle is judged, the corresponding SOH prediction sub-model is allocated according to the working condition, then the corresponding features are selected from the multi-dimensional features and input into the SOH prediction main model and the SOH prediction sub-model respectively, to obtain the main model SOH prediction value and the sub-model SOH prediction value;The main model SOH prediction value and the sub-model SOH prediction value are weighted and fused to obtain the final SOH prediction value. The SOH prediction model is selected according to the vehicle working condition, and the accuracy of the SOH prediction value is significantly improved through weighted fusion.

Description

Technical Field

[0001] This invention belongs to the field of power battery safety monitoring technology, specifically relating to a method and system for monitoring the health of automotive power batteries based on multimodal enhancement. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the safety accidents and health degradation of power batteries have become increasingly prominent, becoming a key bottleneck restricting the high-quality development of the industry. Traditional monitoring technologies, which rely on the collection of single physical quantities and fixed threshold judgments, are unable to process multi-dimensional dynamic correlation data of battery systems, cannot capture weak early signs of thermal runaway, and have an early warning time of less than 3 minutes, which is insufficient to meet the requirement of "thermal runaway early warning time ≥ 5 minutes" in GB / T380-322021 "Safety Requirements for Power Batteries for Electric Vehicles".

[0003] From a health management perspective, the health status of the power battery directly affects the vehicle's driving range, charging efficiency, and lifespan. Under dynamic conditions such as rapid acceleration and low-temperature starts, battery capacity degradation exhibits non-linear characteristics. Traditional estimation methods based on static parameters have an error rate as high as 15%-20%, leading to real pain points for users such as "falsely advertised range" and "breakdowns on the road." Over 60% of new energy vehicle users list "accurate range" as a primary consideration when purchasing a vehicle, and insufficient accuracy in battery health monitoring has become a significant factor hindering consumer confidence.

[0004] Furthermore, the raw monitoring data suffers from issues such as data silos and extreme imbalance between normal and faulty data samples, further impacting the accuracy of fault diagnosis and health assessment. Existing technologies suffer from limitations such as heavy communication burden, slow convergence speed, and poor interpretability, failing to meet the industry's demands for high-precision, robust, and fast-response power battery health monitoring.

[0005] Therefore, designing an intelligent monitoring technology that integrates multimodal features, adapts to complex scenarios, and has the ability to converge within a preset time is of great significance for promoting the high-quality development of the new energy vehicle industry. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for monitoring the health of automotive power batteries based on multimodal enhancement. The method selects a specific SOH prediction model according to the vehicle's operating conditions and significantly improves the accuracy of the SOH prediction value through weighted fusion.

[0007] To achieve the above objectives, the present invention provides a method for health monitoring of automotive power batteries based on multimodal enhancement, comprising the following steps:

[0008] S1. Collect the vehicle's raw sensor data and preprocess it;

[0009] S2. Feature construction is performed on the preprocessed raw sensor data to obtain multi-dimensional features including time-series main features, statistical features, and signal energy and spectral entropy.

[0010] S3. Based on the vehicle operating condition feature vector, determine the current operating condition of the vehicle, assign the corresponding SOH prediction sub-model according to the operating condition, and then select the corresponding features from the multi-dimensional features and input them into the SOH prediction main model and the SOH prediction sub-model respectively to obtain the SOH prediction value of the main model and the SOH prediction value of the sub-model.

[0011] S4. The SOH prediction values ​​of the main model and the SOH prediction values ​​of the sub-model are weighted and fused to obtain the final SOH prediction value.

[0012] Furthermore, in step S2, the multidimensional features also include multimodal sample pairs composed of images and text, and the acquisition methods for the multimodal sample pairs composed of images and text include:

[0013] The temporal main features are used to generate a visualization image. Then, text descriptions are extracted based on the trend features presented in the visualization image. The visualization image and text descriptions are processed by a visual encoder and a language encoder, respectively, to obtain image features and text features. Then, the image features and text features are semantically aligned by a linear mapping to obtain a multimodal sample pair composed of images and text.

[0014] Furthermore, the main SOH prediction model consists of a CNN convolutional neural network, a bidirectional LSTM, and an attention mechanism, while the sub-SOH prediction models include a lightweight LSTM model, an XGBoost model, and a GCNN model.

[0015] The bidirectional LSTM consists of two layers, each with 64 hidden units; the lightweight LSTM model adopts a single-layer LSTM structure with 32 hidden units.

[0016] Furthermore, the inputs to the SOH prediction master model include the temporal main features, statistical features, signal energy, spectral entropy, and the multimodal sample pairs composed of images and text;

[0017] The input to the lightweight LSTM model is the temporal main feature; the input to the XGBoost model is the statistical feature, signal energy, and spectral entropy; and the input to the GCNN model is the temporal feature of a local time window extracted from the temporal main feature.

[0018] Furthermore, the SOH prediction sub-model allocates data using a scene discriminant function:

[0019]

[0020] Where c represents the working condition type and s represents the working condition feature vector;

[0021] Once the operating condition type is determined, the SOH prediction sub-model corresponding to each operating condition type is directly allocated according to the pre-stored SOH sub-model in the system.

[0022] Furthermore, in step S4, the weighted fusion obtains the weights of the main model's SOH prediction value and the sub-model's SOH prediction value under the current operating condition by querying a weight allocation table or by using a lightweight function; the lightweight function is shown in the following formula:

[0023]

[0024]

[0025] Where ε is a coefficient greater than 0. The weights of the SOH predictions of the main model. The weights of the SOH prediction values ​​in the sub-model are... The reliability coefficient of the SOH prediction value of the master model under operating condition c. The reliability coefficient of the SOH prediction value of the sub-model under operating condition c. To predict the RMSE of the main model on the validation set for the SOH, The RMSE is the error of the SOH prediction sub-model on the validation set.

[0026] Furthermore, in step S1, the raw sensing data includes basic parameters, time information, status information, and individual cell data; the basic parameters include total voltage, total current, state of charge, highest individual cell voltage, lowest individual cell voltage, highest temperature, and lowest temperature; the status information includes charging status and total mileage; the individual cell data includes several columns of individual cell voltage and several columns of temperature probe data.

[0027] In step S2, the main time-series features include the basic parameters, time information, and state information; the statistical features include voltage range, temperature range, and total voltage change rate; the signal energy E and spectral entropy H are calculated by extracting frequency domain features through Fourier transform.

[0028] In step S3, the vehicle operating condition feature vector includes ambient temperature, current current and its fluctuation range, SOC change rate, charging mode, number of cycles, cumulative mileage, and data missing flag.

[0029] Furthermore, the formulas for calculating the signal energy E and the spectral entropy H are as follows:

[0030]

[0031]

[0032]

[0033] in, It represents each frequency component of the signal in the frequency domain; N is the signal length, and k is the index of the frequency component. It is the normalized power spectrum.

[0034] Furthermore, in step S1, the preprocessing includes: performing validity checks on the time information; when the time difference is zero, it is necessary to determine whether it is a duplicate record or a sampling anomaly; if it is a duplicate record, the redundant rows are deleted; if it is a sampling anomaly, the timestamp sequence is recalibrated.

[0035] Missing data and NaN values ​​are repaired using interpolation or nearest-neighbor filling.

[0036] The present invention also provides a multimodal enhancement-based automotive power battery health monitoring system, comprising:

[0037] The data acquisition module is used to collect and preprocess the vehicle's raw sensor data.

[0038] The feature construction module is used to construct features from the preprocessed raw sensor data to obtain multi-dimensional features including time-series main features, statistical features, and signal energy and spectral entropy.

[0039] The model allocation module is used to determine the current operating condition of a vehicle based on its operating condition feature vector, and to allocate the corresponding SOH prediction sub-model according to the operating condition.

[0040] The SOH prediction module is used to select corresponding features from the multidimensional features and input them into the SOH prediction main model and the SOH prediction sub-model respectively to obtain the SOH prediction value of the main model and the SOH prediction value of the sub-model.

[0041] The weighted fusion module is used to weight and fuse the SOH prediction values ​​of the main model and the SOH prediction values ​​of the sub-models to obtain the final SOH prediction value.

[0042] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0043] 1. The multimodal enhanced automotive power battery health monitoring method provided by this invention constructs a multi-algorithm collaborative prediction system consisting of a "main model + supplementary model pool + scenario adaptive scheduling + weighted fusion" based on unified data preprocessing and feature construction. Finally, it outputs a single battery health prediction value through weighted fusion, significantly improving prediction accuracy. This invention achieves high-precision SOH detection under different operating conditions by unifying input features, clearly defining the roles of the main model and supplementary models, implementing adaptive scheduling based on operating conditions, and using reliability-based weighted fusion. It balances global robustness with scenario-specific optimization capabilities, significantly improving the engineering usability of power battery health assessment in actual vehicle operating environments.

[0044] 2. This invention uses CNN-LSTM-Attention as the main model under complex operating conditions. An additive attention mechanism is introduced at the LSTM output, assigning learnable importance weights to features at each time step. After softmax normalization, a time weight distribution is formed, and the weighted summation is performed over the time dimension. This automatically focuses on critical moments most sensitive to SOH decay, such as the constant voltage termination phase, the voltage plateau contraction range, and the temperature rise abrupt change range. Three supplementary models (lightweight LSTM, XGBoost, and GCNN) are specifically designed for their respective advantageous operating conditions. Lightweight LSTM, while significantly reducing the computational burden on the vehicle side, can still accurately depict long-term trends such as slow capacity decay and steady internal resistance increase. XGBoost identifies the energy concentration and random changes of temperature and voltage signals in the frequency domain based on frequency domain features such as spectral energy and spectral entropy, thereby capturing early decay patterns. GCNN enhances sensitivity to peak features in fast charging environments by lowering the gating threshold, effectively compensating for the main model's insufficient ability to capture local drastic fluctuations under extreme conditions.

[0045] 3. This invention achieves high-dimensional representation and semantic enhancement of battery operating status through the synergistic fusion of visual, linguistic, and structured numerical modalities, improving the stability and robustness of the power battery health prediction model in complex scenarios such as high-dimensionality, non-stationarity, and class imbalance. The multimodal large model feature enhancement algorithm achieves a scene recognition rate of up to 95.5%, effectively alleviating the sample imbalance problem through semantic enhancement and providing a high-quality feature foundation for health assessment. The accuracy of health assessment is improved by 24.2%, with RMSE and MAE as low as 0.0109 and 0.0087 respectively. The error range is controlled within 2.1% under different scenarios, far below the national standard requirement of 7%, effectively solving the user pain point of "falsely advertised range".

[0046] 4. In terms of engineering adaptability, it does not require the collection of bus voltage signals or global data, but only local sampling information and neighbor transmission information, or only local signals. It is flexible to constant power load changes, plug-and-play capability and communication link failures, reducing communication burden by more than 90% and significantly reducing deployment difficulty. In terms of model performance, the distributed algorithm can converge within a preset time, and the convergence time is independent of network size and initial state. At the same time, through attention weight tracing and feature importance quantification, it outputs interpretable decision basis, which is convenient for engineering application and provides efficient, accurate and reliable technical support for power battery management in multiple fields.

[0047] 5. This invention is applicable to real-time monitoring, fault prediction, and health status tracking of power batteries for new energy vehicles. It can be widely applied in the integration of on-board battery management systems in new energy vehicle manufacturing, factory quality inspection during power battery production, and data support for maintenance and repair in the new energy vehicle aftermarket. It provides an integrated solution for power battery safety and health management across multiple fields, encompassing "data acquisition, feature processing, model reasoning, and decision output." Attached Figure Description

[0048] Figure 1 The flowchart of the vehicle power battery health monitoring method based on multimodal enhancement provided by the present invention is shown.

[0049] Figure 2 This invention provides an overall technical roadmap for the multimodal enhancement-based automotive power battery health monitoring method.

[0050] Figure 3 This is a flowchart of the SOH prediction algorithm.

[0051] Figure 4 This is a graph showing the root mean square error (RMSE) analysis results obtained when comparing the performance of different battery health prediction models under the scenario of partial data loss.

[0052] Figure 5 For the present invention in conjunction with Figure 4 A schematic diagram illustrating the results of battery health prediction under the same operating conditions. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0054] This invention aims to address the core problems existing in current health monitoring technologies for new energy vehicle power batteries: low accuracy in health assessment, difficulty in adapting to complex dynamic conditions such as high and low temperatures and high-rate charging, insufficient ability to model the nonlinear degradation law of batteries, and inability to meet users' needs for accurate range; furthermore, imbalanced data samples and insufficient multimodal information fusion result in poor model generalization ability and low anomaly localization accuracy, making it difficult to cope with monitoring needs in different scenarios; simultaneously, existing technologies rely on global data acquisition or complex communication networks, resulting in heavy communication burdens and poor model interpretability, which is not conducive to engineering applications; finally, the model convergence speed is slow, and the convergence time of traditional distributed algorithms is too long, making it impossible to complete state assessment and early warning response within a preset time, affecting the real-time monitoring effect and making it difficult to meet the timeliness requirements in practical applications.

[0055] Please see Figure 1-3 This invention provides a method for monitoring the health of automotive power batteries based on multimodal enhancement, comprising the following steps:

[0056] S1. Collect the vehicle's raw sensor data and preprocess it;

[0057] S2. Feature construction is performed on the preprocessed raw sensor data to obtain multi-dimensional features including time-series main features, statistical features, and signal energy and spectral entropy.

[0058] S3. Based on the vehicle operating condition feature vector, determine the current operating condition of the vehicle, assign the corresponding SOH prediction sub-model according to the operating condition, and then select the corresponding features from the multi-dimensional features and input them into the SOH prediction main model and the SOH prediction sub-model respectively to obtain the SOH prediction value of the main model and the SOH prediction value of the sub-model.

[0059] S4. The SOH prediction values ​​of the main model and the SOH prediction values ​​of the sub-model are weighted and fused to obtain the final SOH prediction value.

[0060] Figure 1 This diagram demonstrates the complete processing flow of the power battery health prediction task, starting from the underlying data acquisition, through anomaly detection, linear interpolation, normalization, and feature construction, to multi-algorithm model training, scenario adaptation, dynamic scheduling, and finally, health output—a complete chain path. The diagram clearly presents the input data types, feature generation mechanisms, and interaction boundaries of each module, providing an overall schematic of the SOH health prediction subsystem.

[0061] Figure 2This invention presents an integrated multi-layered architecture, encompassing data preprocessing, multimodal feature enhancement, DAVAE fault warning model, SOH health prediction model, and vehicle-cloud collaborative deployment. The technology roadmap is presented in a left-to-right link format: input layer → feature layer → model layer → decision layer → output layer, reflecting the invention's integrated intelligent battery management system that connects "early warning – diagnosis – health prediction".

[0062] Figure 3 The data flow and internal structure of the SOH health prediction model are demonstrated, including the data preprocessing module, the unified feature input format (128×11), the input positions, functional modules, and output health values ​​(SOH) of the four core algorithms (CNN-LSTM-Attention, LSTM, XGBoost, and GCNN). The interaction between the multi-algorithm pool and the scene-adaptive scheduling module is also shown, reflecting the "scene-driven model collaboration" concept of this invention.

[0063] The following is a detailed explanation of steps S1 and S2:

[0064] The first step in SOH (Health Status) detection is to systematically preprocess the raw data collected in real time from vehicles, ensuring its continuity, completeness, and interpretability suitable for subsequent modeling. Data preprocessing includes not only cleaning and repairing the raw data but also extracting representative health-related features as the basis for model training and prediction.

[0065] First, the raw sensor data from the vehicle battery is stored in CSV files, with each file corresponding to the raw sampling data of one vehicle. Each file contains multiple feature columns:

[0066] (1) Basic parameters: total voltage (SUM_VOLTAGE), total current (SUM_CURRENT), state of charge (SOC), maximum cell voltage (MAX_CELL_VOLT), minimum cell voltage (MIN_CELL_VOLT), maximum temperature (MAX_TEMP), minimum temperature (MIN_TEMP), where the unit of total voltage is V, the unit of cell voltage is mV, the unit of total current is A, and the unit of temperature is ℃;

[0067] (2) Time information: TIME (Unix timestamp);

[0068] (3) Status information: Charging status (CHARGE_STATUS), total mileage (SUM_MILE_AGE);

[0069] (4) Individual data: Includes 95 / 96 individual voltage columns (U_xx) and 34 / 32 temperature probe columns (T_xx).

[0070] First, a systematic anomaly detection and repair process is performed on the collected raw data (such as timestamps, total voltage, total current, state of charge, cell voltage, temperature probes, etc.). For invalid timestamp correction, when duplicate records or time differences are zero, the time series needs to be realigned according to a set sampling frequency (e.g., 10 seconds) to avoid the impact of data duplication or loss. If missing data (such as NaN values) exists, linear interpolation is used to fill in the missing points, thereby ensuring the continuity of key variables such as battery voltage, current, and temperature. The specific interpolation method is as follows:

[0071]

[0072] Among them, V interp V(t) represents the parameter at the current time t, V(t1) represents the parameter at the previous time t1, and V(t2) represents the parameter at the previous time t2.

[0073] Meanwhile, outlier detection follows the 3σ criterion:

[0074]

[0075] in, It is the sample mean. It is the sample standard deviation. This rule is used to eliminate abnormal voltage, current, or temperature points that are significantly outside the statistical range, in order to ensure the stability and continuity of the input signal.

[0076] Based on the cleaned raw signals, further health-related features are constructed. Feature engineering is a crucial step in SOH detection, helping the model extract effective signals related to battery health degradation. First, a series of statistical features are constructed, which help reveal the battery's degradation trend.

[0077] Voltage Range Defined as the difference between the maximum and minimum voltage of a single cell:

[0078]

[0079] Temperature range , defined as the difference between the maximum and minimum battery temperature:

[0080]

[0081] Among them, voltage range reflects the consistency between individual cells, while temperature range characterizes the uniformity of thermal distribution in the battery. As the battery ages, uneven temperature distribution may worsen, which is closely related to battery performance degradation. Furthermore, to enhance the sensitivity to changes in time-series characteristics, the total voltage change rate is calculated.

[0082]

[0083] The results can reveal transient fluctuations and nonlinear response behavior under charging and discharging conditions.

[0084] Voltage change rate This represents the rate of change of voltage per unit time.

[0085]

[0086] This feature helps to reveal the exacerbation of voltage polarization in aging batteries during charging and discharging, and the increased polarization is usually associated with a decrease in battery health.

[0087] Building upon statistical characteristics, it is also necessary to extract frequency domain features from battery data. These features can better reveal battery health degradation patterns, especially under dynamic operating conditions. Fast Fourier Transform (FFT) is used to perform frequency domain analysis on time-series signals such as voltage and temperature, extracting the following key features:

[0088] Frequency energy: By calculating the energy distribution of voltage or temperature signals in the frequency domain using FFT, it is possible to reveal the energy concentration of the battery at different frequencies. The concentration of energy in the frequency domain typically changes as battery health deteriorates.

[0089] Spectral Entropy: Spectral entropy measures the randomness of signal energy distribution and reflects the stability of battery status. When a battery is healthy, the spectrum is usually relatively concentrated, and the entropy value is low; however, as the battery ages, the randomness of the signal increases, and the entropy value rises. Changes in spectral entropy H are important for identifying early battery degradation. The calculation formula is as follows:

[0090]

[0091] in, It represents each frequency component of the signal in the frequency domain, and N is the length of the signal.

[0092] Spectral energy: Spectral energy E is the sum of the energy of a signal in the frequency domain, and its calculation formula is:

[0093]

[0094] To further analyze the frequency domain structure of the time-series signal, the Fast Fourier Transform (FFT) is used to extract frequency domain features:

[0095]

[0096] As the state of health (SOH) declines, the battery's temperature signal will exhibit typical changes such as concentrated spectral quantity and decreased entropy value. These characteristics can be used to predict changes in battery health in advance.

[0097] By extracting these statistical and frequency domain features, a complete feature set of battery health status can be constructed, providing reliable input data for subsequent health prediction models. Each step of data cleaning, anomaly repair, and feature construction lays a solid foundation for accurate estimation of battery health.

[0098] Specifically, in step S2, the multidimensional features also include multimodal sample pairs composed of images and text, and the methods for obtaining the multimodal sample pairs composed of images and text include:

[0099] The temporal main features are used to generate a visualization image. Then, text descriptions are extracted based on the trend features presented in the visualization image. The visualization image and text descriptions are processed by a visual encoder and a language encoder, respectively, to obtain image features and text features. Then, the image features and text features are semantically aligned by a linear mapping to obtain a multimodal sample pair composed of images and text.

[0100] Specifically, multimodal sample pairs are constructed using a multimodal feature enhancement model. The construction process is as follows:

[0101] To enhance the stability and robustness of power battery health prediction models in complex scenarios such as high-dimensionality, non-stationarity, and class imbalance, this invention constructs a general multimodal feature enhancement framework. Through the synergistic fusion of visual, linguistic, and structured numerical modalities, it achieves high-dimensional representation and semantic enhancement of battery operating states. The operating behavior of power batteries exhibits strong temporal fluctuations and condition-dependent characteristics; relying solely on raw numerical features often fails to fully characterize their potential structural patterns and degradation mechanisms. Therefore, after completing the data preprocessing and feature construction in step one, this invention transforms the processed time-series signals into various visual image formats, such as scatter plots, trend curves, and heatmaps, to reveal the battery's structural changes, abnormal fluctuation patterns, and spatial distribution characteristics at different time scales. Simultaneously, based on the trend features presented in the images, corresponding text descriptions are generated by expert experience, enabling the system to obtain additional domain semantic priors. This constructs a "image + text" multimodal sample pair, providing a composite input for the subsequent SOH model that simultaneously includes structural, semantic, and statistical information.

[0102] In the feature encoding stage, this invention employs a general visual encoder and a language encoder to process the image and text modalities respectively, obtaining their semantic embedding representations. For each multimodal sample pair... Visual encoder Output image features Language encoder Output text features The expression is as follows:

[0103]

[0104] in, For image feature vectors, This represents the text feature vector. The encoder described above can be a pre-trained visual Transformer, a convolutional neural network, or the text encoding end of a large language model, and can obtain strong general semantic capabilities through transfer learning.

[0105] To achieve cross-modal semantic alignment, enabling image and text features to be measured and matched in the same space, this invention further introduces a linear mapping to project the two high-dimensional modalities onto a shared semantic space. Let the dimension of the shared semantic space be dsd_sds, and define a learnable linear projection matrix. , and bias terms , The projection process is as follows:

[0106]

[0107] in, , representing the alignment representations of the image modality and the text modality in the shared semantic space, respectively. , It is a learnable linear projection matrix. This invention enables... (The sentence is incomplete and requires more context to translate accurately.) and Maintaining semantic consistency means that for semantically matched image-text pairs, their vector distances in the shared space should be as close as possible, thereby achieving cross-modal fusion of image structural information and text semantic information. This shared projection space provides a foundation for image-text semantic alignment and also provides more discriminative and stable high-order semantic feature inputs for subsequent SOH main models (such as CNN-LSTM-Attention).

[0108] Regarding parameter optimization strategies, to significantly reduce the parameter size and improve training stability during fine-tuning while maintaining the basic semantic capabilities of pre-training, this invention introduces a low-rank adaptation (LoRA) structure in the multimodal feature enhancement module, restricting parameter updates to a low-rank subspace. Specifically, for the projection matrix in the aforementioned linear mapping layer... and In this invention, the original weights are not directly updated during the fine-tuning stage; instead, low-rank residual terms are introduced only. This applies to the visual projection matrix. The update format is as follows:

[0109]

[0110] For the text projection matrix The update format is as follows:

[0111]

[0112] in and Let be a low-rank matrix to be learned, satisfying:

[0113]

[0114] and This is a preset low-rank constraint. (Passed via:)

[0115]

[0116] This low-rank decomposition method, in this invention, preserves the original projection matrix. While keeping the model frozen, training is performed only on its low-rank residuals, so that parameter updates are concentrated in a lower-dimensional subspace, which greatly reduces the number of new parameters and training computation overhead, while preserving the model's ability to adapt to new domain data.

[0117] In the multimodal feature fusion stage, this invention can concatenate or weightedly fuse numerical features with multimodal semantic features. For example, the numerical feature vector obtained in step one... Image features in shared space Text features By splicing the components together, an enhanced composite feature is formed:

[0118]

[0119]

[0120] Represents the average state of charge, and... This serves as a higher-order input to the subsequent SOH master prediction model (CNN-LSTM-Attention). In this way, numerical time-series features are responsible for accurately characterizing the changes in the physical quantities of the power battery, image features capture the curve shape and spatial distribution structure, text features inject expert knowledge and semantic priors, and multimodal features achieve complementary enhancement in a unified embedding space.

[0121] In summary, through the multimodal feature enhancement mechanism in step S2, the SOH detection model is no longer limited to prediction based on a single numerical feature. Instead, it significantly improves its expressive and interpretive capabilities for battery degradation signals through collaborative modeling of image, text, and numerical values. The cross-modal alignment technique between image and text ensures the synergistic effect of different types of features in a shared semantic space, while the LoRA low-rank adaptation strategy guarantees efficient domain fine-tuning and model updates even under resource-constrained conditions in both in-vehicle and cloud environments. Ultimately, with this multimodal enhancement scheme, the accuracy and robustness of SOH detection under complex operating conditions and imperfect data are significantly improved, providing a richer and more stable high-order feature foundation for the subsequent multi-algorithm collaborative prediction model in step S3.

[0122] Step S3 specifically includes: constructing the SOH health prediction model.

[0123] The degradation process of battery state of health (SOH) exhibits significant nonlinearity and strong scene dependence. Its health evolution pattern varies considerably under different temperatures, current rates, SOC evolution paths, and cycle mileages. To improve the accuracy and robustness of SOH prediction, this invention constructs a multi-algorithm collaborative prediction system based on unified data preprocessing and feature construction. This system consists of a "main model + supplementary model pool + scene-adaptive scheduling + weighted fusion." CNN-LSTM-Attention serves as the main model under complex operating conditions, while three supplementary models (lightweight LSTM, XGBoost, and GCNN) perform specialized modeling for their respective advantageous operating conditions. Finally, a single battery health estimate is output through weighted fusion.

[0124] First, at the input end, this invention specifies a unified time-series feature format for all participating models. Assuming a single prediction uses a time window of length T, and each time step contains d-dimensional variables (total voltage, current, SOC, extreme temperature, individual consistency index, etc.), then the time-series principal feature matrix is ​​denoted as:

[0125]

[0126] The statistical and frequency domain features defined in step S2 are concatenated to obtain a statistical feature vector for use in some models:

[0127]

[0128] For scenarios involving multimodal enhancement, this invention can also denote the semantic embedding output of the image-text multimodal large model as:

[0129]

[0130] Image-text multimodal sample pairs and numerical features The data is then assembled and fed into the main model, thereby supporting collaborative modeling of multiple models within a unified feature space.

[0131] (1) Core model: CNN-LSTM-Attention

[0132] Based on the unified input described above, this invention first constructs a main model, CNN-LSTM-Attention, to provide baseline SOH predictions under complex conditions. The input X to the main model... main It can be represented as:

[0133]

[0134] The mapping relationship is as follows:

[0135]

[0136] in, This represents the normalized SOH prediction value. Structurally, the convolutional module consists of two one-dimensional convolutional layers. The first layer uses 32 convolutional kernels of size 3, and the second layer uses 64 convolutional kernels of size 5. A sliding convolution operation is performed to extract local instantaneous patterns such as voltage fluctuations and temperature gradients, capturing early degradation signals such as sudden voltage range changes, short-term current surges, and local temperature rises. The convolution output is stabilized by BatchNormalization and Dropout before being input into a bidirectional LSTM. The bidirectional LSTM consists of two layers, each with 64 hidden units, which model the time series forward and backward respectively. The outputs are concatenated in the time dimension to form a 128-dimensional feature vector, thereby representing the long-term decay trend across cycles.

[0137] This invention also sets the time step features in the non-charging phase to zero to suppress overfitting of the model to operating conditions with weak correlation to SOH, making feature learning closer to the actual health evolution mechanism. At the LSTM output, an additive attention mechanism is introduced to assign learnable importance weights to each time step feature. After softmax normalization, a time weight distribution is formed, and the time dimension is weighted and summed to automatically focus on the critical moments most sensitive to SOH decay, such as the constant voltage closing stage, the voltage plateau contraction interval, and the temperature rise abrupt change section.

[0138] Regarding the loss function, this invention employs the Huber loss with parameter δ to balance robustness to outliers with fine fitting of small errors. Let the single-sample error e be:

[0139]

[0140] in, If the true state of harmonics (SOH) is labeled, then the Huber loss is defined as:

[0141]

[0142] In this invention, the following is taken: .

[0143] The optimization process uses the AdamW optimizer (learning rate 0.001), combined with cosine annealing learning rate scheduling and early stopping strategy to ensure that the main model achieves stable convergence and good generalization in the later stages of training, and outputs a stable benchmark under unified input features.

[0144] (2) Supplement the model pool (for automatic selection in different scenarios)

[0145] For different typical working conditions, this invention constructs three types of supplementary sub-models in addition to the main model, forming a model pool that can be automatically called according to the scenario.

[0146] (2.1) Lightweight LSTM:

[0147] The lightweight LSTM model is designed for slow charging scenarios with normal temperature, low current, and stable charge / discharge curves. Its input directly uses the same time-series feature matrix as the main model.

[0148]

[0149] The output is the SOH prediction value of a lightweight LSTM:

[0150]

[0151] The model adopts a single-layer LSTM structure with 32 hidden units and adaptively adjusts the proportion of historical information retention through the forget gate mechanism. While significantly reducing the computational burden on the vehicle side, it can still accurately depict long-term trends such as slow capacity decay and steady increase in internal resistance.

[0152] (2.2) XGBoost:

[0153] The XGBoost model is designed for scenarios with sparse data, locally missing data, or where only statistical features are available. It does not directly process the complete time series matrix, but instead uses statistical and frequency domain feature vectors as input Z. XGB ,Right now:

[0154]

[0155] The corresponding mapping is:

[0156]

[0157] This model, using approximately 500 trees and a learning rate of 0.01, combined with feature importance evaluation, maintains good fitting performance by leveraging the natural tolerance of tree models for discrete features and missing values, even with insufficient initial iterations or high missing values. The spectral energy is as follows:

[0158]

[0159] Frequency domain features such as spectral entropy H(X) help the model identify the energy concentration and random changes of temperature and voltage signals in the frequency domain, thereby capturing early decay patterns.

[0160] (2.3) GCNN:

[0161] For GCNN models designed for highly fluctuating operating conditions such as fast charging, incomplete charging, and frequent start-stop cycles, this invention uses a local time window as the input X. local That is, extracting a local segment of length L from the original time series:

[0162]

[0163] And define the mapping:

[0164]

[0165] GCNN employs a three-layer gated convolutional structure, using one-dimensional convolutional kernels of different scales to capture short-term, intense fluctuations such as voltage spikes, transient temperature surges, and current surges. In fast-charging environments, the sensitivity to spike features is enhanced by lowering the gate threshold, thereby effectively compensating for the main model's insufficient ability to capture local, severe fluctuations under extreme conditions.

[0166] (3) Scene adaptive scheduling mechanism

[0167] To automatically select the most suitable supplementary model during the inference phase, this invention constructs a scenario adaptive scheduling mechanism based on operating condition features. A 12-dimensional operating condition feature vector is defined as follows:

[0168]

[0169] This includes information such as ambient temperature, current current and its fluctuation range, SOC change rate, charging mode (fast charging / slow charging), number of cycles, cumulative mileage, and data missing flags. A scene discrimination function is used:

[0170]

[0171] The current operating conditions are divided into several typical types, for example:

[0172]

[0173] in, For stable slow charging at room temperature, For work conditions with sparse or partially missing data, This is suitable for fast charging or high-fluctuation operating conditions. During inference, when c=normal, the system selects a lightweight LSTM as the current supplementary model, letting:

[0174]

[0175] When c=sparse, select the XGBoost model and let:

[0176]

[0177] When c=fast, choose the GCNN model, let:

[0178]

[0179] Regardless of the scenario, the main model CNN-LSTM-Attention always outputs... Only one supplementary model in the model pool is activated for output. This is to facilitate subsequent weighted fusion of the two models.

[0180] In real-world operating conditions, multiple characteristics may simultaneously occur within a single time window, such as data loss, high current during fast charging, and rapid changes in SOC. To ensure the determinism of the scheduling logic, this invention sets the following priority and decision rules for the supplementary model pool:

[0181] First, determine whether to enable the XGBoost model based on data integrity. If, after timestamp correction, interpolation completion, and 3σ cleaning, more than a preset proportion (20%) of sampling points are still missing or removed, it is considered that the time series information is insufficient to support the deep time series model. In this case, XGBoost is selected as the only supplementary model, and the weights of other supplementary models in this window are reset to 0.

[0182] Provided that the data integrity requirements are met, it is then determined whether the condition is fast charging or a condition with strong fluctuations. If the current window is found to meet the conditions such as fast charging, large current amplitude, high SOC change rate, and obvious voltage / temperature spikes, then GCNN is selected as a supplementary model to enhance the characterization of local severe fluctuations.

[0183] When there is neither severe data gaps nor the conditions for fast charging / strong fluctuations are met, a lightweight LSTM is selected as a supplementary model. In this case, the overall operating conditions are assumed to be stable, and the LSTM is responsible for modeling the long-term degradation trend across cycles.

[0184] Therefore, even if multiple features appear simultaneously within the same window, this invention still activates only one supplementary model according to the priority of "data integrity first, then fluctuation intensity, and finally stability as a backup". This model is then used for dual-model weighted fusion with the main model, and the outputs of the other supplementary models do not participate in the weighted calculation within this time window.

[0185] like Figure 4 The error magnitudes of models such as GCNN, LSTM, CNN-LSTM-Attention, and XGBoost on the same dataset are presented in bar chart form to visually demonstrate the differences in fitting accuracy among the models in this scenario, providing a basis for subsequent multi-model collaboration and weight allocation.

[0186] (4) Multi-model fusion stage

[0187] In the multi-model fusion stage, this invention constructs a reliability weighting mechanism based on the historical errors of each model under different scenarios, synthesizing the outputs of the main model and the supplementary model into a single SOH prediction result. In the offline validation stage, for each scenario c, the RMSE error between the main model and the corresponding supplementary model on the validation set is calculated and denoted as:

[0188]

[0189] To predict the RMSE of the main model on the validation set for the SOH, The RMSE is the error of the SOH prediction sub-model on the validation set.

[0190] In order to convert error into reliability, this invention defines the reliability coefficient under scenario c as follows:

[0191]

[0192] Where ε>0 represents a small positive number to prevent division by zero. The reliability coefficient of the SOH prediction value of the master model under operating condition c. The reliability coefficients for the SOH predictions of the sub-model under operating condition c are then normalized to the fusion weights using a lightweight function.

[0193]

[0194] The weights of the SOH predictions of the main model. Assuming the weights of the SOH predictions from the sub-model, the final SOH predictions clearly satisfy:

[0195]

[0196] During engineering deployment, to ensure that the main model always maintains a dominant position in any scenario, this invention can also impose constraints:

[0197]

[0198] During online inference, within each prediction window, the main model is based on the input. Output The scheduling mechanism is based on Calculate scene labels And select the corresponding supplementary model from the model pool to output. The weights for the current scenario can be obtained by looking up a table or using lightweight functions. , Finally, the predicted SOH value is calculated using the following formula:

[0199]

[0200] in, This invention provides the final battery health estimate under the current operating conditions using a multi-algorithm collaborative system. By unifying input features, clearly defining the roles of the main model and supplementary models, adaptive scheduling based on operating conditions, and reliability-based weight fusion, this invention achieves high-precision SOH detection under different operating conditions, balancing global robustness with scenario-specific local optimization capabilities, and significantly improving the engineering usability of power battery health assessment in actual vehicle operating environments.

[0201] For specific working conditions, CNN-LSTM-Attention and XGBoost are used for prediction and weighted fusion. Figure 5 The paper presents a scatter plot comparison of predicted SOH and actual SOH, the distribution of residuals as a function of health status, and a comparison of accuracy by health level, demonstrating that the method of the present invention has good prediction accuracy, error distribution characteristics, and health level determination capability in this scenario.

[0202] This invention achieves a scene recognition rate of up to 95.5% through a multimodal large model feature enhancement algorithm, effectively alleviates the sample imbalance problem through semantic enhancement, and provides a high-quality feature foundation for health assessment; the average accuracy of health assessment is improved by 24.2%, with RMSE and MAE as low as 0.0109 and 0.0087 respectively, satisfying the formula:

[0203]

[0204] The error range is controlled within 2.1% under different scenarios, which is far lower than the national standard requirement of 7%.

[0205] Finally, the vehicle-cloud collaborative deployment is carried out, as detailed below:

[0206] To achieve high-speed response, high reliability, and online learning capabilities, this invention employs a collaborative deployment system between the vehicle-mounted and cloud-based systems. This system fully leverages the respective strengths of both the vehicle-mounted and cloud-based systems, ensuring efficient operation of battery health assessment and fault warning under various working conditions. In this system, the vehicle-mounted system is primarily responsible for data acquisition, data preprocessing, lightweight inference, rapid scene recognition, and some health prediction tasks, while the cloud-based system is responsible for executing computationally intensive multimodal augmentation models and managing a global knowledge base across vehicles. This approach reduces the computational burden on the cloud while ensuring prediction accuracy, and guarantees real-time assessment of battery health status within milliseconds.

[0207] The vehicle-mounted terminal plays a crucial role in this invention, primarily undertaking the following tasks:

[0208] Data Acquisition and Preprocessing: The vehicle-mounted device acquires various battery data in real time from the Battery Management System (BMS), including voltage, current, temperature, and SOC (State of Charge), and performs preliminary processing. Data preprocessing includes missing value imputation, noise removal, and outlier detection to ensure the quality and reliability of the input data.

[0209] Lightweight Inference and Health Assessment: The on-board unit uses an optimized LSTM model for real-time inference and health assessment of the battery's state of health. Under stable operating conditions at room temperature, the LSTM model can quickly assess the long-term degradation trend of the battery (such as capacity decay, impedance increase, etc.). The model's input includes parameters such as battery temperature and SOC, and its output is the battery's health assessment result.

[0210] Rapid Scene Recognition: The vehicle-mounted system identifies the battery's current operating condition based on real-time data, such as fast charging or low temperature, and automatically selects the appropriate prediction model according to different scenarios. For example, under special conditions such as fast charging or low temperature, the vehicle-mounted system can quickly schedule the GCNN model to predict local fluctuations, ensuring an accurate assessment of the battery's health status.

[0211] Cloud-based functions and tasks: The cloud is responsible for handling computationally intensive tasks, mainly including:

[0212] Multimodal Enhancement Model: The CNN-LSTM-Attention core model is executed in the cloud, performing complex multimodal enhancement inference through deep learning. It combines visual data, language descriptions, and sensor data to further improve the accuracy of battery health prediction. The cloud possesses powerful computing resources, capable of processing large-scale data and executing deep learning models, adapting to the characteristics of data under various operating conditions.

[0213] Global Knowledge Base Management: The cloud manages a global knowledge base across vehicles, aggregating data from various in-vehicle devices to analyze and optimize health assessment models. The cloud regularly updates and optimizes these models, then distributes them to the in-vehicle devices in an encrypted manner, ensuring continuous system optimization.

[0214] Periodic Model Updates and Optimization: The cloud periodically retrains and optimizes the model, especially when new battery types or usage scenarios emerge. The cloud updates the health assessment model and distributes the updated model to the vehicle. Through periodic updates and fine-tuning, the model is ensured to adapt to new data distributions and changes in battery health.

[0215] Information exchange between vehicle and cloud

[0216] The vehicle-mounted system and the cloud exchange information using a low-frequency synchronous method. This avoids the bandwidth pressure of high-frequency communication while ensuring the real-time nature of health assessments and fault warnings. The vehicle-mounted system periodically uploads key battery health assessment results to the cloud for verification. The cloud then provides optimized health assessment models based on these results and sends them back to the vehicle-mounted system in encrypted form. This collaborative model ensures the system's efficiency and security.

[0217] In summary, this invention constructs an integrated intelligent technology system covering multimodal feature enhancement, unified feature representation, and SOH health prediction.

[0218] First, there is the multimodal large-model feature enhancement algorithm. This algorithm enhances the semantic level of battery degradation features by aggregating visual and linguistic information from the data. Specifically, the original multidimensional time-series data (voltage U, current I, SOC, temperature T, cell consistency indicators, etc.) is transformed into visual images such as scatter plots, time-series trend charts, and heatmaps. Based on battery aging mechanisms and expert experience, corresponding descriptive text is generated, thus constructing image-text multimodal sample pairs. Subsequently, image features and semantic features are extracted using a visual encoder and a linguistic encoder, respectively. A learnable linear projection matrix is ​​used to map the bimodal embeddings to a shared semantic space, achieving semantic alignment of cross-modal health features. Simultaneously, a parameter optimization strategy of "freezing the backbone + fine-tuning the higher layers" is employed, combined with LoRA technology, to achieve domain-specific adaptation to battery scenarios while maintaining general representation capabilities. This reduces training costs and significantly improves the separability and interpretability of SOH degradation features.

[0219] Secondly, there is the SOH (State of Health) prediction model. This model is designed around the core principles of "temporal correlation—multi-scale perception—attenuation pattern capture," constructing a multi-algorithm collaborative architecture comprising a main model and a comparison model. The core predictor employs a CNN-LSTM-Attention structure, extracting local fluctuation patterns through convolutional modules, capturing long-term degradation trends across cycles through recurrent networks, and focusing on the most sensitive critical stages of battery degradation, such as voltage change rate at the end of constant voltage and temperature range evolution, using attention weights. Simultaneously, it incorporates a multimodal large-model feature enhancement algorithm, fusing high-dimensional semantic features with original numerical features to improve the model's ability to represent SOH evolution behavior under complex operating conditions. Furthermore, to adapt to differences in scenarios such as different temperatures, charging rates, and charging completeness, the model further constructs lightweight LSTM, GCNN, and XGBoost as supplementary predictors, fitting the degradation curve through a unified feature space. During the training phase, each model is uniformly optimized based on health regression error, robustness index, and generalization ability. During the inference phase, a dynamic scheduling strategy is used to select the optimal predictor according to the working conditions, thereby achieving accurate estimation of SOH, stable output, and consistent performance across working conditions.

[0220] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the health of automotive power batteries based on multimodal enhancement, characterized in that, Includes the following steps: S1. Collect the vehicle's raw sensor data and preprocess it; S2. Feature construction is performed on the preprocessed raw sensor data to obtain multi-dimensional features including time-series main features, statistical features, and signal energy and spectral entropy. S3. Based on the vehicle operating condition feature vector, determine the current operating condition of the vehicle, assign the corresponding SOH prediction sub-model according to the operating condition, and then select the corresponding features from the multi-dimensional features and input them into the SOH prediction main model and the SOH prediction sub-model respectively to obtain the SOH prediction value of the main model and the SOH prediction value of the sub-model. S4. The SOH prediction values ​​of the main model and the SOH prediction values ​​of the sub-models are weighted and fused to obtain the final SOH prediction value; In step S2, the multidimensional features also include multimodal sample pairs composed of images and text, and the methods for obtaining the multimodal sample pairs composed of images and text include: The temporal main features are used to generate a visualization image. Then, text descriptions are extracted based on the trend features presented by the visualization image. The visualization image and text descriptions are processed by a visual encoder and a language encoder, respectively, to obtain image features and text features. Then, the image features and text features are semantically aligned by a linear mapping to obtain a multimodal sample pair composed of images and text. The main SOH prediction model consists of a CNN convolutional neural network, a bidirectional LSTM, and an attention mechanism. The SOH prediction sub-models include a lightweight LSTM model, an XGBoost model, and a GCNN model. The bidirectional LSTM consists of two layers, each with 64 hidden units; the lightweight LSTM model uses a single-layer LSTM structure with 32 hidden units. The input to the SOH prediction master model includes the temporal main features, statistical features, signal energy, spectral entropy, and the multimodal sample pairs composed of images and text; The input to the lightweight LSTM model is the temporal main feature; the input to the XGBoost model is the statistical feature, signal energy, and spectral entropy; and the input to the GCNN model is the temporal feature of a local time window extracted from the temporal main feature.

2. The method for monitoring the health of automotive power batteries based on multimodal enhancement according to claim 1, characterized in that, The SOH prediction sub-model is assigned using a scene discrimination function: Where c represents the working condition type and s represents the working condition feature vector; Once the operating condition type is determined, the SOH prediction sub-model corresponding to each operating condition type is directly allocated according to the pre-stored SOH sub-model in the system.

3. The method for monitoring the health of automotive power batteries based on multimodal enhancement according to claim 2, characterized in that, In step S4, the weighted fusion obtains the weights of the main model's SOH prediction value and the sub-model's SOH prediction value under the current operating condition by querying a weight allocation table or by using a lightweight function; the lightweight function is shown in the following formula: Where ε is a coefficient greater than 0. The weights of the SOH predictions of the main model. The weights of the SOH prediction values ​​in the sub-model are... The reliability coefficient of the SOH prediction value of the master model under operating condition c. The reliability coefficient of the SOH prediction value of the sub-model under operating condition c. To predict the RMSE of the main model on the validation set for the SOH, The RMSE is the error of the SOH prediction sub-model on the validation set.

4. The method for monitoring the health of automotive power batteries based on multimodal enhancement according to claim 1, characterized in that, In step S1, the raw sensing data includes basic parameters, time information, status information, and individual cell data; the basic parameters include total voltage, total current, state of charge, highest individual cell voltage, lowest individual cell voltage, highest temperature, and lowest temperature; the status information includes charging status and total mileage; the individual cell data includes several columns of individual cell voltage and several columns of temperature probe data. In step S2, the main time-series features include the basic parameters, time information, and state information; the statistical features include voltage range, temperature range, and total voltage change rate; the signal energy E and spectral entropy H are calculated by extracting frequency domain features through Fourier transform. In step S3, the vehicle operating condition feature vector includes ambient temperature, current current and its fluctuation range, SOC change rate, charging mode, number of cycles, cumulative mileage, and data missing flag.

5. The method for monitoring the health of automotive power batteries based on multimodal enhancement according to claim 4, characterized in that, The formulas for calculating the signal energy E and the spectral entropy H are as follows: in, It represents each frequency component of the signal in the frequency domain; N is the signal length, and k is the index of the frequency component. It is the normalized power spectrum.

6. The method for monitoring the health of automotive power batteries based on multimodal enhancement according to claim 4, characterized in that, In step S1, the preprocessing includes: performing validity checks on the time information; when the time difference is zero, it is necessary to determine whether it is a duplicate record or a sampling anomaly; if it is a duplicate record, the extra rows are deleted; if it is a sampling anomaly, the timestamp sequence is recalibrated. Missing data and NaN values ​​are repaired using interpolation or nearest-neighbor filling.

7. A health monitoring system for automotive power batteries based on multimodal enhancement, characterized in that, include: The data acquisition module is used to collect and preprocess the vehicle's raw sensor data. The feature construction module is used to construct features from the preprocessed raw sensor data to obtain multi-dimensional features including time-series main features, statistical features, and signal energy and spectral entropy. The multidimensional features also include multimodal sample pairs composed of images and text, and the methods for obtaining the multimodal sample pairs composed of images and text include: The temporal main features are used to generate a visualization image. Then, text descriptions are extracted based on the trend features presented by the visualization image. The visualization image and text descriptions are processed by a visual encoder and a language encoder, respectively, to obtain image features and text features. Then, the image features and text features are semantically aligned by a linear mapping to obtain a multimodal sample pair composed of images and text. The model allocation module is used to determine the current operating condition of a vehicle based on its operating condition feature vector, and to allocate the corresponding SOH prediction sub-model according to the operating condition. The SOH prediction module is used to select corresponding features from the multidimensional features and input them into the SOH prediction main model and the SOH prediction sub-model respectively to obtain the SOH prediction value of the main model and the SOH prediction value of the sub-model. The SOH prediction main model is composed of a CNN convolutional neural network, a bidirectional LSTM and an attention mechanism, and the SOH prediction sub-model includes a lightweight LSTM model, an XGBoost model and a GCNN model. The bidirectional LSTM consists of two layers, each with 64 hidden units; the lightweight LSTM model uses a single-layer LSTM structure with 32 hidden units. The input to the SOH prediction master model includes the temporal main features, statistical features, signal energy, spectral entropy, and the multimodal sample pairs composed of images and text; The input to the lightweight LSTM model is the temporal main feature; the input to the XGBoost model is the statistical feature, signal energy, and spectral entropy; and the input to the GCNN model is the temporal feature of a local time window extracted from the temporal main feature. The weighted fusion module is used to weight and fuse the SOH prediction values ​​of the main model and the SOH prediction values ​​of the sub-models to obtain the final SOH prediction value.

Images