Residual error enhanced lithium ion battery health state large model evaluation method and system

Abstract

The invention provides a residual error enhanced lithium ion battery health state large model evaluation method and system, and relates to the technical field of lithium ion battery state evaluation. The method comprises the following steps: acquiring original monitoring data of a full life cycle charging and discharging process of the lithium ion battery, and performing denoising and standardization preprocessing to obtain a high-quality time sequence data set; constructing a hierarchical time sequence token coding mechanism, and converting the multi-dimensional battery cycle characteristic sequence into a semantic token sequence which can be processed by a large model; constructing a residual enhanced feature mapping network; constructing a time sequence reasoning trunk based on a large language model architecture, and capturing a long-range degradation dependency relationship by using rotation position coding and a multi-head self-attention mechanism; and outputting a health state evaluation value of the lithium ion battery through the feedforward neural network and the regression head. According to the method, the defects of a traditional deep learning method in the aspect of capturing long-range dependence and generalization ability are overcome, end-to-end accurate SOH estimation is achieved, and the lithium ion battery health state evaluation value can be directly and accurately output.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery state assessment technology, specifically to a residual-enhanced lithium-ion battery health state large-scale model assessment method and system. Background Technology

[0002] As the core energy unit of high-end equipment such as new energy vehicles, large-scale energy storage power stations, and aerospace equipment, the accurate assessment of the State of Health (SOH) of lithium-ion batteries directly determines the reliability and safety of equipment operation and is also a core prerequisite for achieving intelligent operation and maintenance and optimizing the entire life cycle management. However, lithium-ion batteries are complex electrochemical systems. During repeated charge-discharge cycles, a series of irreversible microscopic degradations occur internally, such as lithium dendrite growth, electrolyte decomposition, and loss of active materials. Macroscopically, the degradation process presents three major challenges: first, strong nonlinearity and time-varying nature, with huge differences in capacity decay rates at different stages, accompanied by abrupt phenomena such as capacity drops; second, multi-factor coupling, with the degradation rate dynamically affected by factors such as temperature, rate, and depth of discharge; and third, individual variability, where even batteries from the same batch may exhibit vastly different lifespans.

[0003] Existing SOH estimation methods are mainly divided into two categories. The first category is model-based physical methods, such as equivalent circuit models or pseudo-two-dimensional models. These methods have clear theories, but high model complexity, require accurate identification of a large number of electrochemical parameters that are difficult to measure online, and have a large computational load and poor environmental adaptability. The second category is data-driven methods, such as support vector regression, CNN, or LSTM. However, traditional deep learning models still have fundamental defects: First, fragmented architecture: feature engineering, temporal modeling, regression prediction, and other stages are stacked with different modules, resulting in a redundant architecture that is difficult to optimize and fails to achieve end-to-end holistic learning. Second, insufficient capture of long-range dependencies: battery degradation is a long process spanning hundreds or even thousands of cycles. Traditional RNN-type models are limited by the vanishing or exploding gradient problem, making it difficult to effectively model the decay trajectory over such a long period of time, resulting in insufficient memory. Finally, weak generalization ability: models trained on specific datasets often experience a sharp drop in performance when applied to new battery models or new operating conditions, lacking the robustness required for industrial scenarios.

[0004] With the rise of large-scale models (LMs) like the Transformer, their powerful contextual learning and cross-task transfer capabilities offer new insights into solving the aforementioned problems. However, industrial data consists of low-sampling-rate, multi-variable continuous time-series signals, which differ fundamentally from the discrete text semantics that large-scale models excel at processing—a "modal gap." Furthermore, early battery degradation features are weak and easily dissipated during transmission through deep networks. Therefore, directly applying existing large-scale model architectures will not be successful; there is an urgent need to design encoding and feature enhancement mechanisms adapted to the characteristics of battery data. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a cross-modal adaptive residual-enhanced large-scale model evaluation method and system for lithium-ion battery health status. By designing a hierarchical temporal token encoding and residual-enhanced feature mapping network, the long-term inference capability of the large model is transferred to the field of battery health management, thus solving the shortcomings of existing methods in long-range dependency capture and generalization capabilities.

[0006] Specifically, on the one hand, the present invention provides a large-scale model evaluation method for the health status of residual-enhanced lithium-ion batteries, which includes the following steps: S1. Obtain voltage, current and temperature monitoring data of lithium-ion batteries in different cycle periods, and clean and remove outliers from the monitoring data to obtain a preprocessed battery cycle dataset. S2. Based on the preprocessed battery cycling dataset, extract multidimensional statistical and physical features that reflect battery aging characteristics, and then normalize them to construct a normalized feature sequence. S3. Construct a hierarchical temporal token encoding mechanism to convert the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, forming a token sequence arranged in chronological order. S4. Construct a residual enhancement feature mapping network. Input the token sequence into the residual enhancement feature mapping network, and after high-dimensional linear projection, nonlinear manifold transformation, and residual connection processing, obtain the enhanced high-dimensional semantic feature vector. The specific steps are as follows: S41, Using a linear projection layer to project the first The original input vector for each cycle period The transformation vector is obtained by mapping to a high-dimensional latent space. Then, layer normalization and ReLU activation are performed to obtain intermediate variables. : ; ; In the formula: For a learnable projection weight matrix, For bias vectors, and These are the statistical mean and standard deviation of the feature dimension, respectively. Here is the numerical stability constant. and These are the affine transformation parameters; S42. The original information stream before linear projection is fused with the high-level semantic features after nonlinear transformation through residual connection to obtain the final feature output. : ; In the formula: For mapping functions; S5. Construct a SOH estimation model based on a large language model architecture. The SOH estimation model includes a stacked multi-head self-attention module, a rotation position encoding module, and a feedforward neural network module. S6. Input the enhanced high-dimensional semantic feature vector into the SOH estimation model, and inject relative position information into the feature vector through the rotation position encoding module; S7. Calculate the global dependencies within the sequence using a multi-head self-attention mechanism, and perform deep feature extraction and fusion through a feedforward neural network; S8. Based on the output of the SOH estimation model, the health status assessment value of the lithium-ion battery is calculated through the regression prediction head.

[0007] Preferably, in step S42, when the input and output dimensions are the same, ;otherwise , This is a dimension matching matrix.

[0008] Preferably, the specific process of cleaning and outlier removal of monitoring data in step S1 includes: Calculate the original sequence mean with standard deviation : ; ; in, For the first Data from each sampling point Set the total number of samples; set a threshold. , retain satisfaction The samples were discarded, and the remaining abnormal samples were removed. threshold The value of is 3; .

[0009] Preferably, step S8 specifically includes: S81. Construct a modular regression prediction head: The regression prediction head adopts a fully connected network structure. The first layer receives the deep fusion feature vector output in step S7 and maps the high-dimensional feature vector to the intermediate layer. The intermediate layer uses the ReLU activation function to obtain the non-linear mapping relationship between the features and the SOH value. A Dropout regularization layer is introduced to prevent overfitting. Each layer is configured with a layer normalization operation. S82. Output layer design: The output layer of the regression prediction head uses a linear activation function to output the preliminary normalized prediction value after mapping. S83. Inverse Normalization Operation: The actual SOH value is restored using the inverse normalization formula, which is: in, The normalized predicted value output by the regression head. This represents the maximum value of SOH. This is the minimum value of SOH; The prediction result is after inverse normalization; S84, Physical Constraint Calibration: Perform boundary constraint calibration on the inverse normalized prediction results. Then it is truncated to 1.0, if Then truncate to 0.0 and output the final value. The value is used as a prediction result, and the health status of the lithium-ion battery is assessed based on the output value.

[0010] Preferably, in step S1, piecewise linear interpolation is used to fill in the outliers or missing values ​​that have been removed. The piecewise linear interpolation expression is as follows: ; in, This indicates the preceding valid data point at the start of the missing interval. This indicates the next valid data point after the end of the missing interval. The segment length of consecutive missing points. This represents the number of intervals within the missing interval for the current interpolation point.

[0011] Preferably, in step S6, the rotation position encoding module uses a block-based diagonal rotation matrix. For query vector and key vector Encode, rotation matrix The expression is: ; in, For sequence position index, For feature dimension, For the rotating fundamental frequency parameter, This represents a matrix concatenation operation; Encoded query vector With key vector The relative positional relationships satisfy: This allows for the explicit preservation of relative positional information during attention calculations.

[0012] Preferably, the calculation expression for the multi-head self-attention mechanism in step S7 is: ; in, , , ; Given the input feature sequence, These are weight matrices for the query, key, and value, respectively. Let T be the feature dimension and T be the transpose matrix.

[0013] Preferably, the calculation expression for the feedforward neural network in step S7 is: ; in, Here is the weight matrix of the feedforward network. This is the bias vector.

[0014] Preferably, in step S2, the extracted multidimensional statistical features include the mean, standard deviation, kurtosis, skewness, and entropy of the voltage and current sequences; the extracted physical features include constant current charging time, constant voltage charging time, constant current charging charge, constant voltage charging charge, voltage curve slope, current curve slope, time difference, and charge difference.

[0015] On the other hand, the present invention provides a residual-enhanced lithium-ion battery health state large model evaluation system, which includes: The data acquisition module is used to acquire voltage, current and temperature monitoring data of lithium-ion batteries at different cycle periods, and to clean the monitoring data using the mean-variance method and piecewise linear interpolation method to obtain a preprocessed battery cycle dataset. The feature extraction module is used to extract multi-dimensional features, including mean, standard deviation, kurtosis, skewness, charging time and capacity, based on the preprocessed battery cycle dataset, and perform Min-Max normalization to construct a normalized feature sequence. The temporal token encoding module is used to construct a hierarchical temporal token encoding mechanism, which converts the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, forming a token sequence arranged in chronological order. The residual feature mapping module is used to construct a residual enhanced feature mapping network, perform linear projection and nonlinear transformation on the token sequence, and output an enhanced high-dimensional semantic feature vector through residual connections; the residual connection expression is as follows: ; The model building module is used to build an SOH estimation model based on a large language model architecture. The model includes a rotation position encoding module, a multi-head self-attention module, and a feedforward neural network module. The position encoding module is used to inject relative position information into features using a rotated position encoding matrix; The attention inference module is used to inject relative position information into the feature vector using rotational position encoding, and then uses the formula... Calculate the global dependencies of sequence features; The SOH prediction module is used to calculate the health status assessment value of the lithium-ion battery through a regression layer based on the output of the feedforward neural network.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention proposes a new method for estimating the state of health (SOH) of a battery driven by a basic large model. By leveraging the excellent long-term sequence reasoning ability and cross-scene context awareness mechanism learned by the basic large model during the pre-training process of massive data, it effectively overcomes the technical limitations of traditional deep learning methods such as recurrent neural networks and shallow Transformers, which are difficult to accurately capture the long-term dynamic dependencies of the battery throughout its entire life cycle due to limitations in model capacity and pre-training data. This method can achieve high-precision, full-cycle estimation of the health status of lithium-ion batteries.

[0017] (2) The residual-enhanced lithium-ion battery health status large model evaluation method proposed in this invention designs a hierarchical time-series token encoding mechanism and a residual-enhanced feature mapping network, which effectively solves the modal gap between continuous numerical battery data and the semantic space of the large model, and enhances the transmission efficiency of weak degradation signals through deep residual connections, realizing accurate alignment from the original sensor data to the high-dimensional semantic space, and ensuring the accuracy of the output results.

[0018] (3) The residual-enhanced lithium-ion battery health status large model evaluation method proposed in this invention introduces rotation position coding, which explicitly preserves the relative position relationship in attention calculation. Compared with absolute position coding, rotation position coding is more suitable for processing variable-length sequences and has better numerical stability, avoiding gradient problems in long sequence scenarios.

[0019] (4) The comprehensive verification of the method of the present invention on four public datasets, HUST, MIT, XJTU and TJU, shows that compared with mainstream benchmark models such as BiGRU and Transformer, it achieves the best or near-best levels in terms of root mean square error (RMSE) and mean absolute error (MAE), demonstrating the excellent prediction accuracy, stability and generalization ability of the method of the present invention. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall process framework of the residual-enhanced lithium-ion battery health status large-scale model evaluation method of the present invention. Figure 2 This is a schematic diagram of the battery health status prediction model based on a large model according to the present invention. Figure 3 This is a comparison chart showing the change in prediction error of the method of this invention with other models on the HUST dataset as a function of the number of iterations; Figure 4 This is a histogram of the error distribution of the method of the present invention on the HUST dataset; Figure 5 This is a block diagram of the residual-enhanced lithium-ion battery health status large model evaluation system of the present invention. Detailed Implementation

[0021] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating the overall process framework of the cross-modal adaptive residual-enhanced lithium-ion battery health state assessment method of this invention. Figure 1 As shown, this invention provides a large-scale model evaluation method for the state of health of a residual-enhanced lithium-ion battery with cross-modal adaptation, which specifically includes the following steps: S1. Obtain voltage, current and temperature monitoring data of lithium-ion batteries at different cycle periods, and clean and remove outliers from the monitoring data to obtain a preprocessed battery cycle dataset.

[0023] Specifically, in practical applications, data cleaning is necessary due to measurement errors and interference signals in the actual acquisition environment. This applies to the original sequence. First, calculate its mean. with standard deviation : ; ; To balance robustness to extreme anomalies and long-tailed distributions, a threshold is used. Outlier removal is performed, combined with interquartile range (ICM) anomaly detection. The index set is retained. Data points within, that is, satisfying The samples are processed, and the remaining abnormal samples that do not meet the threshold are deleted.

[0024] in, .

[0025] To avoid introducing non-physical oscillations, piecewise linear interpolation is used to fill in outliers or missing values ​​that have been removed. For multiple consecutive missing values, the interpolation formula is as follows: ; in, This indicates the preceding valid data point at the start of the missing interval. This indicates the next valid data point after the end of the missing interval. The segment length of consecutive missing points. This represents the number of intervals within the missing interval for the current interpolation point. The above interpolation process reduces the impact of noise and improves data reliability.

[0026] S2. Based on the preprocessed battery cycling dataset, extract multidimensional statistical and physical features that reflect battery aging characteristics, and normalize the extracted features to construct a normalized feature sequence.

[0027] Specifically, to fully extract implicit information related to battery aging from the original voltage and current curves, a series of health indicators with both statistical and physical significance were extracted. The extracted features encompass statistical quantities describing the data distribution pattern and physical quantities reflecting changes in the electrochemical process, including: multidimensional statistical features such as the mean, standard deviation, kurtosis, skewness, and entropy of the voltage series; and the mean, standard deviation, kurtosis, skewness, and entropy of the current series. Physical features include: constant current charging time, constant voltage charging time, constant current charging charge, constant voltage charging charge, voltage curve slope, current curve slope, time difference, and charge difference.

[0028] The extracted features are highly correlated with internal battery degradation mechanisms such as lithium dendrite growth and loss of active materials. A Min-Max normalization method was then used to unify the extracted features to the same scale, ensuring that the model's convergence speed is not affected by differences in feature scale during training, thus effectively capturing the core information of battery state changes. Finally, the normalized features were sorted to obtain a normalized feature sequence.

[0029] S3. Construct a hierarchical temporal token encoding mechanism, which converts the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, and sorts all the semantic tokens by time to form a token sequence arranged in chronological order.

[0030] Basic large-scale models excel at handling sequential information composed of discrete tokens. This invention constructs a hierarchical token encoding mechanism: the normalized multidimensional feature vector extracted from each charge-discharge cycle of the battery is treated as an independent token, used to compactly represent the overall health status of that cycle. Under this representation, all cycles experienced by the battery from its initial use to the current moment are organized into a token sequence in chronological order. After this processing, the original continuous-time-based SOH estimation task is transformed into a sequence-to-sequence (Seq2Seq) modeling problem. This serialized representation not only preserves the temporal structure of the battery aging process but also enables the model to understand the interactions between cycles at the token level, thereby capturing short-term changes and long-term trends in degradation behavior.

[0031] S4. Construct a residual enhancement feature mapping network. Input the token sequence into the residual enhancement feature mapping network, and obtain the enhanced high-dimensional semantic feature vector through high-dimensional linear projection, nonlinear manifold transformation and residual connection processing.

[0032] The state-of-the-art (SOH) degradation process of batteries is essentially a strongly nonlinear, multi-scale electrochemical evolution process. Early aging signals often manifest as weak perturbations in the original data space, which are easily dissipated during transmission in deep networks. This invention constructs a hierarchical mapping architecture that includes high-dimensional linear projection, nonlinear manifold transformation, and residual feature enhancement.

[0033] First, the linear projection layer is used to project the first... The original input vector for each cycle period The model is mapped to a high-dimensional latent feature space. Subsequently, layer normalization is introduced to eliminate covariate bias, and ReLU activation is used to induce sparsity to capture complex nonlinear degradation patterns. The specific transformation process is as follows: ; ; In the formula: For a learnable projection weight matrix, For bias vectors, and These are the statistical mean and standard deviation of the feature dimension, respectively. Here is the numerical stability constant. and These are the affine transformation parameters.

[0034] Considering the long-term temporal characteristics of battery lifecycle data, as network depth increases, weak early degradation features are prone to gradient vanishing. This invention designs a residual enhancement path, fusing the original information stream before linear projection with high-level semantic features after nonlinear transformation to obtain the final feature output after residual enhancement. As an enhanced high-dimensional semantic feature vector The calculation formula is: ; in, This is an identity mapping or dimension-matching projection. It occurs when the input and output dimensions are the same. When the input dimension and the output dimension are inconsistent, This design establishes fast gradient connections, allowing error signals to be directly transmitted to shallower layers during backpropagation, effectively preserving the original physical characteristics that reflect the battery's health state. Furthermore, a Dropout regularization mechanism is introduced at the residual block output. This further enhances the algorithm's generalization robustness when faced with small sample battery data.

[0035] S5. Construct a SOH estimation model based on a large language model architecture. The SOH estimation model includes a stacked multi-head self-attention module, a rotation position encoding module, and a feedforward neural network module. A schematic diagram of the model is shown below. Figure 2 As shown.

[0036] The SOH estimation model of this invention is based on the DeepSeek-LLM-7B-base pre-trained model to design the time-series inference backbone, and uses its powerful general inference capabilities to solve the problem of battery aging prediction.

[0037] S6. Input the enhanced high-dimensional semantic feature vector into the SOH estimation model, and inject relative position information into the feature vector through the rotation position encoding module.

[0038] The absolute position encoding in the standard Transformer model is difficult to generalize effectively to sequences of different lengths and cannot explicitly model time intervals. This invention introduces Rotational Position Encoding (RoPE), which transforms absolute position encoding into a geometric rotation operation in the complex field, thereby explicitly preserving relative positional relationships during attention computation.

[0039] Given sequence position index With feature dimension index Rotational Position Encoding (RoPE) constructs a block-based diagonal rotation matrix. Implement position encoding: ; In the formula: These are dimension-dependent rotational fundamental frequency parameters. For the query vector... and key vector Position-aware projection is equivalent to a complex field rotation operation: , .

[0040] During the attention score calculation phase, this mechanism ensures that the attention weights depend only on the relative position offset. It satisfies translation invariance: This allows for the explicit preservation of relative positional information during attention computation. This relative positional modeling method ensures numerical stability and avoids the problems of gradient explosion and gradient vanishing in long sequence scenarios.

[0041] S7. Calculate the global dependencies within the sequence using a multi-head self-attention mechanism, and perform deep feature extraction and fusion through a feedforward neural network.

[0042] Multi-head self-attention mechanisms enable the model to automatically capture dependencies between data. First, a query matrix is ​​constructed. Key matrix and value matrix : ; Subsequently, attention calculation is performed. The calculation expression for the multi-head self-attention mechanism is as follows: ; in, , , ; Given the input feature sequence, These are weight matrices for the query, key, and value, respectively. Let T be the feature dimension and T be the transpose matrix.

[0043] The above calculations do not depend on the distance of tokens in the sequence, allowing the model to directly establish dependencies between any two cycles, thereby constructing a global degradation view spanning the entire battery lifecycle. The multi-head mechanism allows the model to capture dependencies at different scales in parallel.

[0044] After undergoing self-attention operation, the data is fed into a feedforward neural network (FFN) to further extract deep features. The calculation expression is as follows: ; Feedforward neural networks (FFNs) effectively enhance the nonlinear expressive power of models, ensuring more thorough information transfer and fusion between layers. Residual connections are introduced in both the self-attention module and the feedforward network, and layer normalization is employed to guarantee the overall network training stability.

[0045] S8. Based on the output of the SOH estimation model, the health status assessment value of the lithium-ion battery is calculated through a regression prediction head. Finally, a linear regression layer is used to map the high-dimensional semantic features back to the SOH numerical space, completing the end-to-end SOH prediction.

[0046] In specific applications, step S8 includes the following sub-steps: S81. Constructing a modular regression prediction head: The regression prediction head adopts a fully connected network structure. The first layer receives the deep fusion feature vector output by S7. The high-dimensional feature vector is mapped to the intermediate feature space through dimension adaptive mapping. The intermediate layer uses the ReLU activation function to capture the non-linear mapping relationship between features and SOH values. At the same time, a Dropout regularization layer is introduced to prevent overfitting. Each layer is configured with a layer normalization operation to maintain the stability of data distribution and ensure efficient convergence during the training process.

[0047] S82. Output layer design: The output layer of the regression prediction head adopts a linear activation function to avoid range compression of the SOH prediction value. The output layer directly outputs the preliminary prediction result after mapping. The weights of the output layer are initialized with a normal distribution, and the bias term is initialized to 0 to ensure that the initial prediction bias is minimized.

[0048] S83. Inverse Normalization Operation: Since the SOH labels were synchronously normalized using Min-Max during the training phase, the actual SOH value is recovered here as the predicted value after inverse normalization using the inverse normalization formula. The inverse normalization formula is as follows: in, The normalized predicted value output by the regression head. The maximum and minimum values ​​of SOH represent the physical reasonable range boundaries of SOH.

[0049] S84, Physical Constraint Calibration: Perform boundary constraint calibration on the inverse normalized prediction results. Then it is truncated to 1.0, if Then truncate it to 0.0, so that the output falls within the interval. Internally, and output the final result. The value is used as the prediction result, and the health status is evaluated based on the output value. The above physical constraints ensure that the output result conforms to the physical definition of the health status of lithium-ion batteries and avoids meaningless outliers.

[0050] In specific embodiments, the regression prediction head can be modularly adapted and optimized: the regression prediction head supports dynamic adjustment of structural parameters according to the deployment scenario: when facing edge computing devices such as battery management systems (BMS), a 2-layer lightweight fully connected network is adopted with a 32-dimensional intermediate layer to reduce inference latency; when facing cloud servers, it can be expanded to a 3-layer fully connected network with a 128-dimensional intermediate layer to further improve prediction accuracy and achieve a flexible balance between accuracy and real-time performance.

[0051] Secondly, this invention also provides a large-scale model assessment system for the state of health of lithium-ion batteries with cross-modal adaptation and residual enhancement, such as... Figure 5 As shown, it includes: The data acquisition module 1 is used to acquire voltage, current and temperature monitoring data of lithium-ion batteries at different cycle periods, and to clean the monitoring data using the mean-variance method and piecewise linear interpolation method to obtain a preprocessed battery cycle dataset.

[0052] Feature extraction module 2 is used to extract multi-dimensional features, including mean, standard deviation, kurtosis, skewness, charging time and capacity, based on the preprocessed battery cycle dataset, and perform Min-Max normalization processing to construct a normalized feature sequence.

[0053] The temporal token encoding module 3 is used to construct a hierarchical temporal token encoding mechanism, which converts the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, forming a token sequence arranged in chronological order.

[0054] Residual feature mapping module 4 is used to construct a residual enhanced feature mapping network, performing linear projection and nonlinear transformation on the token sequence, and outputting an enhanced high-dimensional semantic feature vector through residual connections; the residual connection expression is as follows: .

[0055] Model building module 5 is used to build an SOH estimation model based on a large language model architecture. The model includes a rotation position encoding module, a multi-head self-attention module, and a feedforward neural network module.

[0056] Position encoding module 6 is used to inject relative position information into features using a rotated position encoding matrix.

[0057] Attention reasoning module 7 is used to inject relative position information into the feature vector using rotational position encoding, and uses the formula Calculate the global dependencies of sequence features.

[0058] SOH prediction module 8 is used to calculate the health status assessment value of lithium-ion battery through a regression layer based on the output of the feedforward neural network.

[0059] To verify the effectiveness of this invention, the proposed SOH-LM model was trained and tested end-to-end against mainstream baseline models such as CNN, ResNet, LSTM, Transformer, BiGRU, and P-CLCL in a multi-source battery lifetime prediction experimental environment constructed using four battery datasets. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used as the benchmarks. 2 As performance evaluation metrics, RMSE highlights the predictive stability of the model under extreme conditions, MAE intuitively reflects the average deviation of the prediction results, and R... 2The closer the value is to 1, the stronger the model's ability to fit the SOH decay process. Each model was tested three times, and the average result and standard deviation were taken to fully verify its effectiveness. The experimental results are shown in the table below.

[0060] The following conclusions can be clearly drawn from the experimental results: The SOH-LM model proposed in this invention achieves high RMSE, MAE, and R-values ​​across all four datasets. 2 SOH-LM achieved best performance across all three core metrics: On the HUST dataset, SOH-LM achieved an RMSE of 0.63%±0.04% and a MAE of 0.43%±0.03%, representing reductions of approximately 35% and 34% respectively compared to the best baseline model, CNN; on the MIT dataset, its RMSE was 0.26%±0.01% and MAE was 0.14%±0.01%, outperforming the best performance of baseline models such as ResNet; on the TJU dataset, despite the small performance differences among models due to periodic fluctuations in the data, SOH-LM maintained a slight advantage with an RMSE of 0.45%±0.01% and a MAE of 0.22%±0.01%; on the XJTU dataset, SOH-LM achieved an RMSE as low as 0.13%±0.06% and a MAE of 0.09%±0.04%, significantly outperforming the baseline model's highest RMSE of nearly 1.0%.

[0061] R of the SOH-LM model of this invention 2 The values ​​reached 99.46%±0.07% (HUST), 99.58%±0.04% (MIT), 98.85%±0.03% (TJU), and 99.83%±0.17% on the four datasets, respectively, all of which were higher than all baseline methods. This indicates that its fitting of the battery SOH degradation process is close to the "no residual" level, and it has extremely strong feature capture and fitting capabilities.

[0062] This invention demonstrates excellent adaptability and robustness for different datasets: on the MIT dataset, which has low noise and easily captures cyclic features, all models perform similarly, but SOH-LM maintains its advantage due to its nonlinear feature deep mapping and long-range dependency modeling capabilities; on the HUST and XJTU datasets, which have complex degradation patterns, SOH-LM performs better in R... 2 While maintaining an accuracy of over 99%, it also maintains an extremely low standard deviation, enabling precise capture of short-term minor fluctuations and long-term decline trends. On the TJU dataset, which exhibits significant cyclical fluctuations, its comprehensive modeling capability of features at both the inner and outer loop levels is more complete, ensuring the stability of its predictive performance.

[0063] 4) Based on the test results of the four datasets, the SOH-LM model of this invention performs more robustly in the multi-source battery lifetime prediction task, consistently achieving optimal or near-optimal prediction accuracy. Its powerful feature extraction ability, generalization ability, and adaptability to complex operating conditions are significantly better than existing mainstream baseline models. Among them, the error distribution histogram ( Figure 4 The results show that the prediction error of SOH-LM exhibits a significant central tendency and light-tailed nature, with the vast majority of errors concentrated within ±0.01, without any obvious systematic overestimation or underestimation, which is better than the comparison model. The error variation curve with the number of cycles (…) Figure 3 The results show that the error of SOH-LM remains stable throughout the cycle without cumulative drift, verifying the effectiveness of multi-head self-attention and RoPE in long-term feature modeling.

[0064] This invention presents a cross-modal adaptive residual-enhanced large-scale model assessment method for the state of health (SOH) of lithium-ion batteries. It eliminates modal differences through hierarchical encoding, preserves weak features through residual enhancement, and captures long-range dependencies using a large-scale model architecture, thereby achieving high-precision and strong-generalization SOH estimation.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A large-scale model evaluation method for the state of health of residual-enhanced lithium-ion batteries, characterized in that: It includes the following steps: S1. Obtain voltage, current and temperature monitoring data of lithium-ion batteries in different cycle periods, and clean and remove outliers from the monitoring data to obtain a preprocessed battery cycle dataset. S2. Based on the preprocessed battery cycling dataset, extract multidimensional statistical and physical features that reflect battery aging characteristics, and then normalize them to construct a normalized feature sequence. S3. Construct a hierarchical temporal token encoding mechanism to convert the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, forming a token sequence arranged in chronological order. S4. Construct a residual enhancement feature mapping network. Input the token sequence into the residual enhancement feature mapping network, and after high-dimensional linear projection, nonlinear manifold transformation, and residual connection processing, obtain the enhanced high-dimensional semantic feature vector. The specific steps are as follows: S41, Using a linear projection layer to project the first The original input vector for each cycle period The transformation vector is obtained by mapping to a high-dimensional latent space. Then, layer normalization and ReLU activation are performed to obtain intermediate variables. : ； ； In the formula: For a learnable projection weight matrix, For bias vectors, and These are the statistical mean and standard deviation of the feature dimension, respectively. Here is the numerical stability constant. and These are the affine transformation parameters; S42. The original information stream before linear projection is fused with the high-level semantic features after nonlinear transformation through residual connection to obtain the final feature output. : ； In the formula: For mapping functions; S5. Construct a SOH estimation model based on a large language model architecture. The SOH estimation model includes a stacked multi-head self-attention module, a rotation position encoding module, and a feedforward neural network module. S6. Input the enhanced high-dimensional semantic feature vector into the SOH estimation model, and inject relative position information into the feature vector through the rotation position encoding module; S7. Calculate the global dependencies within the sequence using a multi-head self-attention mechanism, and perform deep feature extraction and fusion through a feedforward neural network; S8. Based on the output of the SOH estimation model, the health status assessment value of the lithium-ion battery is calculated through the regression prediction head.

2. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: In step S42, when the input and output dimensions are the same, ;otherwise , This is a dimension matching matrix.

3. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: The specific process of cleaning and outlier removal of monitoring data in step S1 includes: Calculate the original sequence mean with standard deviation : ； ； in, For the first Data from each sampling point Set the total number of samples; set a threshold. , retain satisfaction Remove the remaining abnormal samples from the sample; threshold The value of is 3; 。 4. The residual-enhanced lithium-ion battery health state assessment method according to claim 3, characterized in that: Step S8 is as follows: S81. Construct a modular regression prediction head: The regression prediction head adopts a fully connected network structure. The first layer receives the deep fusion feature vector output in step S7 and maps the high-dimensional feature vector to the intermediate layer. The intermediate layer uses the ReLU activation function to obtain the non-linear mapping relationship between the features and the SOH value. A Dropout regularization layer is introduced to prevent overfitting. Each layer is configured with a layer normalization operation. S82. Output layer design: The output layer of the regression prediction head uses a linear activation function to output the preliminary normalized prediction value after mapping. S83. Inverse Normalization Operation: The actual SOH value is restored using the inverse normalization formula, which is: in, The normalized predicted value output by the regression head. This represents the maximum value of SOH. This is the minimum value of SOH; The prediction result is after inverse normalization; S84, Physical Constraint Calibration: Perform boundary constraint calibration on the inverse normalized prediction results. Then it is truncated to 1.0, if Then truncate to 0.0 and output the final value. The value is used as a prediction result, and the health status of the lithium-ion battery is assessed based on the output value.

5. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: In step S1, for the removed outliers or missing values, piecewise linear interpolation is used to fill in the gaps. The piecewise linear interpolation expression is as follows: ； in, This indicates the preceding valid data point at the start of the missing interval. This indicates the next valid data point after the end of the missing interval. The segment length of consecutive missing points. This represents the number of intervals within the missing interval for the current interpolation point.

6. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: In step S6, the rotation position encoding module uses a block-based diagonal rotation matrix. For query vector and key vector Encode, rotation matrix The expression is: ； in, For sequence position index, For feature dimension, For the rotating fundamental frequency parameter, This represents a matrix concatenation operation; Encoded query vector With key vector The relative positional relationships satisfy: This allows for the explicit preservation of relative positional information during attention calculations.

7. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: The calculation expression for the multi-head self-attention mechanism in step S7 is as follows: ； in, , , ; Given the input feature sequence, These are weight matrices for the query, key, and value, respectively. Let T be the feature dimension and T be the transpose matrix.

8. The residual-enhanced lithium-ion battery health state assessment method according to claim 7, characterized in that: The calculation expression for the feedforward neural network in step S7 is: ； in, These are the weight matrices of the feedforward network. These are the bias vectors.

9. The residual-enhanced lithium-ion battery health state assessment method according to claim 1, characterized in that: In step S2, the extracted multidimensional statistical features include the mean, standard deviation, kurtosis, skewness, and entropy of the voltage and current sequences; the extracted physical features include constant current charging time, constant voltage charging time, constant current charging charge, constant voltage charging charge, voltage curve slope, current curve slope, time difference, and charge difference.

10. A residual-enhanced lithium-ion battery health state large model evaluation system for the residual-enhanced lithium-ion battery health state large model evaluation method according to claim 1, characterized in that: It includes a data acquisition module, which is used to acquire voltage, current and temperature monitoring data of lithium-ion batteries at different cycle periods, and clean the monitoring data using the mean-variance method and piecewise linear interpolation method to obtain a preprocessed battery cycle dataset. The feature extraction module is used to extract multi-dimensional features, including mean, standard deviation, kurtosis, skewness, charging time and capacity, based on the preprocessed battery cycle dataset, and perform Min-Max normalization to construct a normalized feature sequence. The temporal token encoding module is used to construct a hierarchical temporal token encoding mechanism, which converts the multidimensional feature vector of each cycle in the normalized feature sequence into a semantic token, forming a token sequence arranged in chronological order. The residual feature mapping module is used to construct a residual enhanced feature mapping network, perform linear projection and nonlinear transformation on the token sequence, and output an enhanced high-dimensional semantic feature vector through residual connections; the residual connection expression is as follows: ; The model building module is used to build an SOH estimation model based on a large language model architecture. The model includes a rotation position encoding module, a multi-head self-attention module, and a feedforward neural network module. The position encoding module is used to inject relative position information into features using a rotated position encoding matrix; The attention inference module is used to inject relative position information into the feature vector using rotational position encoding, and then uses the formula... Calculate the global dependencies of sequence features; The SOH prediction module is used to calculate the health status assessment value of the lithium-ion battery through a regression layer based on the output of the feedforward neural network.

Images