Lithium battery health state prediction method based on CNN-Transform-LSTM

Abstract

The invention discloses a lithium battery health state prediction method based on CNN-Transform-LSTM, and belongs to the technical field of lithium battery state monitoring and health management. Meanwhile, constant-current charging time and average charging voltage and average charging current in the whole process are extracted from a charging curve to serve as health features, and high-relevance features are screened through relevance analysis. Firstly, local features in data are extracted by using a convolutional neural network, and redundant information is reduced; then, using Transform to capture global dependence, and identifying a long-term health trend; and finally, capturing long-term dependence and trend in time through a long-short-term memory network, and ensuring smooth and stable prediction. According to the method, the later aging process can be predicted according to the first 70% of data of the whole life cycle of the lithium battery, and compared with an existing advanced model, the prediction precision and generalization ability are improved.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery state monitoring and health management technology, specifically involving a lithium battery health state prediction method based on CNN-Transformer-LSTM. Background Technology

[0002] As a core component of modern energy storage, the performance degradation of lithium batteries directly impacts the safe operation and lifespan of systems such as electric vehicles, smart grids, and portable electronic devices. Battery state of health (SOH), a key indicator for quantifying the degree of battery performance degradation, is crucial for accurate prediction and is an essential prerequisite for implementing proactive battery safety management, optimizing charge and discharge strategies, extending battery life, and realizing the reuse of batteries. Therefore, developing high-precision and robust SOH prediction technology has been a focus of attention for both academia and industry.

[0003] Existing SOH prediction methods can be broadly categorized into three types: those based on physical models, those based on traditional machine learning models, and those based on deep learning models. Physical model-based methods attempt to characterize the complex internal processes of batteries through electrochemical mechanisms (such as equivalent circuit models and degradation mechanism models). However, these models typically rely on precise identification of physical parameters, resulting in complex modeling processes. Furthermore, they struggle to adequately adapt to the challenges posed by individual battery differences, complex and variable operating conditions, and multi-factor coupled degradation mechanisms in real-world applications, limiting their universality and accuracy in practical engineering applications. Traditional machine learning-based methods (such as support vector machines and random forests) rely on manually designing and extracting features reflecting battery aging (such as capacity increment curve features and constant current charging time features). While these methods reduce model complexity to some extent, their predictive performance is highly dependent on the quality and experience of feature engineering. The feature extraction process is cumbersome and highly subjective, making it difficult to automate and efficiently mine and utilize the deep, multi-scale spatiotemporal correlation features containing degradation information from battery operating data, especially high-dimensional, non-stationary time-series data such as voltage, current, and temperature.

[0004] Because battery degradation patterns are influenced by multiple factors, exhibiting nonlinear and dynamic characteristics, the complexity of SOH prediction for lithium-ion batteries is increased. Although data-driven methods learn the degradation patterns of lithium-ion batteries from historical data, single models often cannot efficiently complete the task of predicting SOH. Existing hybrid models, while having achieved advantages in processing specific types of battery data, still suffer from insufficient feature utilization and weak noise robustness.

[0005] Therefore, how to collaboratively integrate multiple features and achieve multi-level modeling within a unified framework remains an important research question. Summary of the Invention

[0006] The purpose of this invention is to provide a lithium battery state of health prediction method based on CNN-Transformer-LSTM, which improves the prediction accuracy of lithium battery SOH through multi-level feature fusion and staged temporal modeling. This provides a reliable guarantee for the safe operation of systems such as electric vehicles, smart grids, and portable electronic devices, and can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a lithium battery health status prediction method based on CNN-Transformer-LSTM, the method comprising the following steps:

[0008] S1. Obtain lithium battery charge / discharge cycle degradation data, including voltage, current, time, and capacity.

[0009] S2. Process the collected data, replace abnormal data with the data from the previous moment, and calculate SOH.

[0010] S3. Capacity increment curves are extracted using capacity increment analysis. Simultaneously, the constant current charging time Tcc, the average charging voltage Tv,avg, and the average charging current Tc,avg are extracted from the charging curves as health characteristics. Pearson correlation analysis is then used to select the components with high capacity correlation as the final health characteristics.

[0011] S4. Construct the CNN-Transformer-LSTM model, which includes the encoder part of CNN and Transformer and the Long Short-Term Memory network (LSTM);

[0012] S5. After the model is built, select lithium battery data for processing, take the first 70% as the training set and the last 30% as the test set, and send them to the CNN-Transformer-LSTM model for training.

[0013] S6. After the model training is completed, it is sent to the test set, and the SOH prediction results are evaluated using evaluation metrics.

[0014] S7. Preprocess other datasets of the same battery type as those used during model training, extract features according to experimental steps S2-S3, and put the final health features into the trained model to verify the model's generalization ability and effectiveness in practical applications.

[0015] Preferably, the SOH of the lithium battery is calculated based on the battery capacity, as described in step S2, using the following formula:

[0016]

[0017] Where C0 represents the rated capacity of the battery; C k This represents the capacity after k charge-discharge cycles.

[0018] Preferably, the capacity increment analysis method described in step S3 is a key method for studying battery aging and degradation mechanisms. Its core is to differentiate the rate of change of battery capacity relative to open-circuit voltage. This method can convert voltage fluctuations within the battery into peaks and troughs on the IC curve. However, the IC curve obtained by traditional methods is easily affected by noise, resulting in instability. This characteristic increases the difficulty of data processing and adds uncertainty to the analysis of battery health status. Therefore, this method introduces an improved IC curve calculation method. Compared with the traditional IC curve calculation method, this method introduces incremental accumulation and normalization calculation at a fixed reference voltage point in data processing, which is equivalent to performing a certain degree of data smoothing. This method can effectively reduce the influence of noise, making the IC curve more stable, thereby improving the reliability of the IC curve for predicting health status. At the same time, since only the IC curve corresponding to a specific voltage point needs to be calculated, the computational load is reduced.

[0019] Preferably, the Pearson correlation analysis method described in step S3 is a commonly used correlation analysis method with the advantage that the calculation results are not affected by the dimensions or magnitude of the variables. The calculation formula is as follows:

[0020]

[0021] Where: P represents the correlation coefficient, n is the number of health characteristic elements, and x i y i These represent the health characteristics indexed by i and the actual battery capacity, respectively. These are the average values ​​of health characteristics and actual capacity, respectively.

[0022] Preferably, the Transformer part described in step S4 consists of an encoder and a decoder. It efficiently captures global dependencies in the data through a multi-head attention mechanism, making it suitable for handling long-term trends in battery health. The decoder part typically requires an existing target value as input to guide the generation of predicted values; however, it is difficult to input the actual values ​​in real time during online prediction. Therefore, this invention only uses the encoder part of the Transformer.

[0023] Preferably, the CNN-Transformer-LSTM model described in step S4 is a hybrid model combining a convolutional neural network (CNN), a Transformer, and a long short-term memory network (LSTM). This model achieves multi-level feature fusion and dynamic temporal modeling through a phased collaborative framework, improving the accuracy of SOH prediction. The model first extracts local features from the lithium battery input data through a CNN layer, capturing local fluctuations in data changes through convolutional operations. Then, the extracted features are fed into the Transformer encoder for global dependency modeling. A multi-head attention mechanism, as the core of the encoder, linearly transforms the input, generating three matrices: query, key, and value. These three matrices are then used for independent attention processing, capturing the dependencies between different positions within the sequence through multiple attention heads. The outputs of these multiple attention heads are then merged to obtain the final multi-head self-attention output. Next, an LSTM layer further processes the long-term dependencies of the data, smoothing the prediction results and improving the ability to predict future states. Finally, a fully connected layer maps the output of the long short-term memory network (LSTM) to the SOH prediction value. This model fully leverages the complementary properties of Convolutional Neural Networks (CNNs) and Transformers, combining them to focus on both local features and global dependencies. Then, a Long Short-Term Memory (LSTM) network is used to process temporal relationships, further improving the overall performance of the model.

[0024] Preferably, in step S5, the training model is configured such that the input and output channels of the convolutional layers are 9 and 64 respectively, and the kernel size is 3. The Transformer layer uses a 64-dimensional embedding vector and employs two self-attention heads and a 128-dimensional feedforward network. A single LSTM layer with 128 hidden units is used. The fully connected layer maps the output of the LSTM to the SOH prediction result. During model training, the Adam optimizer is used with a learning rate of 0.0015, training for 100 epochs, and a batch size of 32.

[0025] Preferably, the evaluation index described in step S6. This method selects mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) as evaluation indexes, with the specific formulas as follows:

[0026]

[0027] Among them, y kThis is the actual value of the battery's SOH; is the predicted SOH value for the battery; n represents the predicted length.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] This invention presents a novel CNN-Transformer-LSTM hybrid model for predicting state of health (SOH) in lithium batteries, proposing a three-stage hybrid structure: First, a convolutional neural network (CNN) is used to extract local features from the data to reduce redundant information; then, a Transformer is used to capture global dependencies and identify long-term health trends; finally, a Long Short-Term Memory (LSTM) network is used to capture long-term dependencies and trends over time, ensuring smooth and stable predictions.

[0030] Multiple feature fusion strategies are employed: through correlation analysis, highly correlated features are selected from capacity increment curves, constant current charging time, average voltage, and average current for fusion.

[0031] The lithium battery health state prediction method based on the CNN-Transformer-LSTM model exhibits excellent generalization ability, effectively fitting SOH decay data and significantly improving SOH prediction accuracy. This provides a reliable guarantee for the safe operation of systems such as electric vehicles, smart grids, and portable electronic devices. Attached Figure Description

[0032] To fully reflect the technical features of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below.

[0033] Figure 1 The overall flowchart of the lithium battery health status prediction method based on the CNN-Transformer-LSTM model provided for embodiments of the present invention is shown below.

[0034] Figure 2 A model structure diagram of a lithium battery health status prediction method based on a CNN-Transformer-LSTM model provided for embodiments of the present invention.

[0035] Figure 3 The diagram illustrates the convolution kernel operation process of the lithium battery health status prediction method based on the CNN-Transformer-LSTM model provided in the embodiments of the present invention.

[0036] Figure 4 The diagram shows the multi-head self-attention mechanism of the lithium battery health state prediction method based on the CNN-Transformer-LSTM model provided in the embodiments of the present invention.

[0037] Figure 5The LSTM structure diagram of the lithium battery health state prediction method based on the CNN-Transformer-LSTM model provided for embodiments of the present invention is shown. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] As an embodiment of the present invention, a method for predicting the health status of lithium batteries based on a CNN-Transformer-LSTM model is provided. Figure 1 The provided flowchart outlines a lithium battery health state prediction method based on the CNN-Transformer-LSTM model, which mainly includes the following steps:

[0040] S1. Obtain lithium battery charge-discharge cycle degradation data, including at least voltage, current, time, and capacity;

[0041] S2. Process the collected data, replace abnormal data with the data from the previous moment, and calculate SOH.

[0042] S3. Extract the capacity increment curve using the capacity increment analysis method, and simultaneously extract the constant current charging time T from the charging curve. cc The average charging voltage T throughout the entire process v,avg and the average charging current T throughout the process c,avg As a health characteristic, Pearson correlation analysis was used to screen for components with high correlation to volume;

[0043] S4. Construct the CNN-Transformer-LSTM model, which includes a convolutional neural network (CNN), the encoder part of the Transformer, and a long short-term memory network (LSTM).

[0044] S5. After the model is built, select lithium battery data for processing, take the first 70% as the training set and the last 30% as the test set, and send them to the CNN-Transformer-LSTM model for training.

[0045] S6. After the model training is completed, it is sent to the test set, and the SOH prediction results are evaluated using evaluation metrics.

[0046] S7. Preprocess other datasets of the same battery type as those used during model training, extract features according to experimental steps S2-S3, and put the final health features into the trained model to verify the model's generalization ability and effectiveness in practical applications.

[0047] Step S2 includes the following specific steps:

[0048] Step S2.1: Preprocess the dataset, including handling outliers;

[0049] Step S2.2: Calculate the State of Harm (SOH) of the lithium battery. SOH is an important indicator for assessing the degree of battery aging or degradation. The SOH of a lithium battery is calculated based on its capacity, using the following formula:

[0050]

[0051] Where C0 represents the rated capacity of the battery; C k This represents the capacity after k charge-discharge cycles.

[0052] Preferably, step S3 includes the following specific steps:

[0053] Step S3.1: Extract capacity increment curves (IC1~IC14) using the capacity increment analysis method. Capacity increment analysis is a key method for studying battery aging and degradation mechanisms. Its core is to differentiate the rate of change of battery capacity relative to open-circuit voltage. This method can convert voltage fluctuations within the battery into peaks and troughs on the IC curve. However, IC curves obtained by traditional methods are easily affected by noise, resulting in instability. This characteristic increases the difficulty of data processing and adds uncertainty to the analysis of battery health status. Therefore, this method introduces an improved IC curve calculation method. Compared with the traditional IC curve calculation method, this method introduces incremental accumulation and normalization calculations at a fixed reference voltage point in data processing, which is equivalent to a certain degree of data smoothing. This method can effectively reduce the influence of noise, making the IC curve more stable, thereby improving the reliability of the IC curve for predicting health status. At the same time, since only the IC curve corresponding to the characteristic voltage point needs to be calculated, the computational load is reduced.

[0054] During the constant current charging phase, time-series data T = [t1, t2, ..., t] of the battery are collected. l-1 ,t l Voltage sequence data V = [V1, V2, ..., V l-1 V l And the corresponding current sequence data I=[I1,I2,…,I l-1 ,I lBased on the current sequence and time interval, the instantaneous capacity increment dQ at each sampling point is calculated according to the following rules. i : Calculate the capacity change per unit time using these data, dQ = [dQ1, dQ2, ..., dQ] l-1 ,dQ l ]

[0055]

[0056] This generates the instantaneous capacity increment sequence dQ = [dQ1, dQ2, ..., dQ2]. l-1 ,dQ l ].

[0057] Select the boundary value V of the battery charging voltage range s (starting voltage) and V e (Termination voltage), constructing an equally spaced reference voltage sequence with a fixed voltage interval ΔV.

[0058]

[0059] Where m represents the index of the reference voltage point; V e This indicates the maximum reference voltage value.

[0060] Initialize a capacity increment accumulator array of the same length as the reference voltage sequence, for each reference voltage point. Allocate a capacity increment IC with an initial value of 0. m Iterate through the voltage V at each sampling point. i ,implement:

[0061] (1) In the reference voltage sequence V ref Mid-positioning and V i The reference voltage point with the smallest distance

[0062] (2) V i The corresponding instantaneous capacity increment dQ i The calculated capacity increment accumulator accumulated to this reference voltage point:

[0063] IC m =IC m +dQ i (4)

[0064] After completing the full charge cycle data processing, for each reference voltage point The accumulated values ​​are normalized by voltage intervals to generate the final capacity increment curve:

[0065]

[0066] Step S3.2: Extract the constant current charging time T from the charging curvecc The average charging voltage T throughout the entire process v,avg and the average charging current T throughout the process c,avg Here, the subscript cc represents the charging current; the subscripts c,avg represent the average charging current; and the subscripts v,avg represent the average charging voltage. These features, combined with capacity increment features, can more comprehensively capture battery health degradation characteristics, helping to improve the accuracy of SOH prediction.

[0067] Step S3.3: Use Pearson correlation analysis to select the components with high correlation to volume as the final health characteristics. Pearson correlation analysis is a commonly used correlation analysis method with the advantage that the calculation results are not affected by the dimensions or magnitude of the variables. The calculation formula is as follows:

[0068]

[0069] Where: P represents the correlation coefficient, n is the number of health characteristic elements, and x i y i These represent the health characteristics indexed by i and the actual battery capacity, respectively. These are the average values ​​of health characteristics and actual capacity, respectively.

[0070] Preferably, step S4 includes the following specific steps:

[0071] Step S4.1: The CNN-Transformer-LSTM model is a hybrid model combining a Convolutional Neural Network (CNN), a Transformer, and a Long Short-Term Memory (LSTM) network. The model structure diagram is shown below. Figure 2 As shown in the figure. This model improves the accuracy of SOH prediction by achieving multi-level feature fusion and dynamic temporal modeling through a phased collaborative framework. The model first extracts local features from the lithium battery input data through a CNN layer, and then captures local fluctuations in data changes through convolution operations. Its convolution kernel operation is as follows: Figure 3 As shown. The extracted features are then fed into the Transformer encoder for global dependency modeling. The multi-head attention mechanism, as the core of the encoder, linearly transforms the input, generating three matrices: Query, Key, and Value. These three matrices are then used for independent attention processing, capturing the dependencies between different positions within the sequence through multiple attention heads. The outputs of these multiple attention heads are then merged to obtain the final multi-head self-attention output. The multi-head attention mechanism is as follows: Figure 4As shown in the diagram. Secondly, a Long Short-Term Memory (LSTM) layer is used to further process the long-term dependencies of the data, smoothing the prediction results and improving the ability to predict future states. Finally, a fully connected layer maps the output of the LSTM to the SOH prediction value. This model fully utilizes the complementary properties of the Convolutional Neural Network (CNN) and Transformer architectures, combining them to focus on both local features and capture global dependencies. Then, the LSTM network is used to process temporal relationships, further improving the overall performance of the model.

[0072] Figure 5 The structure of LSTM is shown. Long Short-Term Memory (LSTM) networks, as a variant of RNNs, are better suited for handling long-term sequence problems and include forget gates, input gates, and output gates.

[0073] The forget gate receives the output sequence of the previous unit and the input sequence of the current unit, and uses the sigmoid function to determine the degree to which the previous layer's output has been forgotten. The calculation formula is as follows:

[0074] f(t)=σ(W f ·[h(t-1),x(t)]+b f (7)

[0075]

[0076] Where σ represents the sigmoid activation function; W f and b f Let represent the weight matrix and bias parameters of the forget gate, respectively; h(t-1) represents the hidden state at the previous time step; x(t) represents the current input; W C and b C These represent the weight matrix and bias parameters of the memory cell state, respectively.

[0077] The input gate stores the new information into the cell state, calculated using the following formula:

[0078] i(t)=σ(W i ·[h(t-1),x(t)]+b i )

[0079] Among them, W i b represents the weights of the input gate; i is the bias of the input gate; h(t-1) is the output at time t-1; x(t) is the input at time t.

[0080] The formula for calculating the output gate is as follows:

[0081] σ(t)=σ(W σ ·[h(t-1),x(t)]+bσ (9)

[0082] Among them, W σ and b σ These represent the weights and biases of the output gate, respectively.

[0083] Finally, the hidden state and memory unit are updated using the results of the forget gate, input gate, and output gate.

[0084]

[0085] h(t)=σ(t)×tanh(C(t)) (11)

[0086] Where C(t) represents the current memory cell; h(t) represents the hidden state at the current moment.

[0087] Preferably, step S5 includes the following specific steps:

[0088] Step S5.1: Set up the training model. During training, set the input and output channels of the convolutional layers to 9 and 64, respectively, with a kernel size of 3. The Transformer layer uses a 64-dimensional embedding vector and employs two self-attention heads and a 128-dimensional feedforward network. A single LSTM layer with 128 hidden units is used. The fully connected layer maps the output of the Long Short-Term Memory (LSTM) network to the SOH prediction result. The model is trained using the Adam optimizer with a learning rate of 0.0015 for 100 epochs and a batch size of 32.

[0089] Step S5.2: Process the selected lithium battery data, i.e., take the first 70% as the training set and the last 30% as the test set, and feed it into the CNN-Transformer-LSTM model for training. To verify the effectiveness of CNN-Transformer-LSTM, this invention compares the method of this invention with the Long Short-Term Memory Network (LSTM) model, the Transformer model, and the Transformer-LSTM model.

[0090] Step S6 includes the following specific steps:

[0091] Step S6.1: After the model training is completed, it is sent to the test set, and the SOH prediction results are evaluated using evaluation metrics. Mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as evaluation metrics, with the specific formulas as follows:

[0092]

[0093] Among them, y k This is the actual value of the battery's SOH; is the predicted SOH value for the battery; n represents the predicted length.

[0094] Preferably, step S7 includes the following specific steps:

[0095] Step S7.1: Process the dataset of other lithium batteries of the same model as those used in training the model according to steps S2 to S3, put the final health features into the trained model, and verify the generalization ability and effectiveness of the model in practical applications based on the prediction results.

[0096] In summary, the lithium battery health state prediction method based on the CNN-Transformer-LSTM model of this invention extracts the capacity increment curve using capacity increment analysis, and simultaneously extracts constant current charging time, average charging voltage throughout the process, and average charging current as health features from the charging curve, and filters highly correlated features through correlation analysis. First, a convolutional neural network (CNN) is used to extract local features from the data to reduce redundant information; then, a Transformer is used to capture global dependencies and identify long-term health trends; finally, a Long Short-Term Memory (LSTM) network is used to capture long-term temporal dependencies and trends, ensuring smooth and stable predictions. Experimental results show that the model achieves good results on the dataset, improving prediction accuracy and generalization ability compared to existing state-of-the-art models.

[0097] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0098] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the health status of lithium batteries based on CNN-Transformer-LSTM, characterized in that, The method includes the following steps: S1. Obtain lithium battery charge-discharge cycle degradation data, including at least voltage, current, time, and capacity data; S2. Process the collected data, replace abnormal data with the data from the previous moment, and calculate SOH. S3. Extract the capacity increment curve using the capacity increment analysis method, and simultaneously extract the constant current charging time T from the charging curve. cc The average charging voltage T throughout the entire process v,avg and the average charging current T throughout the process c,avg As a health trait, Pearson correlation analysis was used to select the components with high correlation to volume as the final health trait. S4. Construct a CNN-Transformer-LSTM model, which includes a convolutional neural network, a Transformer encoder, and a long short-term memory network. S5. After the model is built, select lithium battery data for processing, take the first 70% as the training set and the last 30% as the test set, and send them into the CNN-Transformer-LSTM model for training. S6. After the model training is completed, it is sent to the test set, and the SOH prediction results are evaluated using evaluation metrics. S7. Preprocess other datasets of the same battery type as those used during model training, extract features according to steps S2 to S3, and put the final health features into the trained model to verify the model's generalization ability and effectiveness in practical applications.

2. The lithium battery health state prediction method based on CNN-Transformer-LSTM according to claim 1, characterized in that, The SOH mentioned in step S2 is calculated based on the battery capacity using the following formula: Where C0 represents the rated capacity of the battery; C k This represents the capacity after k charge-discharge cycles.

3. The lithium battery health status prediction method based on CNN-Transformer-LSTM according to claim 1, characterized in that, The capacity increment analysis method described in step S3 involves differentiating the rate of change of battery capacity relative to open-circuit voltage, converting voltage fluctuations inside the battery into peaks and troughs on the IC curve. Simultaneously, in the data processing of the IC curve, incremental accumulation and normalization calculations at a fixed reference voltage point are introduced to effectively reduce the impact of noise, making the IC curve more stable and thus improving the reliability of the IC curve for predicting health status. The specific process for calculating the incremental accumulation and normalization at the fixed reference voltage point includes: During the constant current charging phase, time-series data of the battery, T = [t1, t2, ..., t], are collected. l-1 ,t l Voltage sequence data V = [V1, V2, ..., V l-1 V l ], and the corresponding current sequence data I=[I1,I2,…,I l-1 ,I l Based on the current sequence and time interval, the instantaneous capacity increment dQ at each sampling point is calculated according to the following rules. i ; Calculate the capacity change per unit time using these data, dQ = [dQ1, dQ2, ..., dQ] l-1 ,dQ l ]: This generates the instantaneous capacity increment sequence dQ = [dQ1, dQ2, ..., dQ2]. l-1 ,dQ l ]; Select the boundary value V of the battery charging voltage range s and V e A series of equally spaced reference voltages are constructed with a fixed voltage interval ΔV: Where m represents the index of the reference voltage point; V e Indicates the maximum reference voltage value; Initialize a capacity increment accumulator array of the same length as the reference voltage sequence, for each reference voltage point. Allocate a capacity increment IC with an initial value of 0. m ; Traverse the voltage V at each sampling point i ,implement: (1) In the reference voltage sequence V ref Mid-positioning and V i The reference voltage point with the smallest distance (2) V i The corresponding instantaneous capacity increment dQ i The calculated capacity increment accumulator accumulated to this reference voltage point: IC m =IC m +dQ i After completing the full charge cycle data processing, for each reference voltage point The accumulated values ​​are normalized by voltage intervals to generate the final capacity increment curve:

4. The lithium battery health status prediction method based on CNN-Transformer-LSTM according to claim 1, characterized in that, The Pearson correlation analysis method described in step S3 has the advantage that the calculation results are not affected by the dimensions or magnitude of the variables; the calculation formula is as follows: Where: P represents the correlation coefficient, n is the number of health characteristic elements, and x i y i These represent the health characteristics indexed by i and the actual battery capacity, respectively. These are the average values ​​of health characteristics and actual capacity, respectively.

5. The lithium battery health status prediction method based on the CNN-Transformer-LSTM model according to claim 1, characterized in that, The CNN-Transformer-LSTM model described in step S4 is a hybrid model that combines convolutional neural networks, Transformers, and long short-term memory networks. This model achieves multi-level feature fusion and dynamic temporal modeling through a phased collaborative framework, thereby improving the accuracy of SOH prediction. The specific working process of the CNN-Transformer-LSTM model includes: First, local features in the lithium battery input data are extracted through convolutional neural network layers, and local fluctuations in data changes are captured through convolution operations. Then, the extracted features are fed into the Transformer encoder for global dependency modeling. The multi-head attention mechanism, as the core part of the encoder, performs linear transformation on the input to generate three matrices: query, key, and value. At the same time, these three matrices are used to perform independent attention processing. Multiple attention heads capture the dependencies between various positions within the sequence. Then, the outputs of multiple attention heads are merged to obtain the final multi-head self-attention output. Secondly, the long short-term memory network layer further processes the long-term dependencies of the data, smooths the prediction results, and improves the ability to predict future states. Finally, the output of the Long Short-Term Memory network is mapped to the SOH prediction value through a fully connected layer.

6. The method for predicting the health status of lithium batteries based on a CNN-Transformer-LSTM model according to claim 1, characterized in that, The training model described in step S5 has the following parameters: during training, the number of input and output channels of the convolutional layer is 9 and 64, respectively, and the kernel size is 3; the Transformer layer uses a 64-dimensional embedding vector and employs two self-attention heads and a 128-dimensional feedforward network; a single-layer Long Short-Term Memory network with 128 hidden units is used; the fully connected layer maps the output of the Long Short-Term Memory network to the SOH prediction result; the model uses the Adam optimizer with a learning rate of 0.0015, trains for 100 epochs, and has a batch size of 32.

7. The lithium battery health status prediction method based on the CNN-Transformer-LSTM model according to claim 1, characterized in that, The evaluation indicators mentioned in step S6 are selected as mean square error, root mean square error, and mean absolute percentage error, and the specific formulas are as follows: Among them, y k This is the actual value of the battery's SOH; is the predicted SOH value for the battery; n represents the predicted length.

Images