Estimation method for measuring battery specific energy based on deep learning in combination with EIS data

Abstract

The invention discloses an estimation method for measuring battery specific energy based on deep learning in combination with EIS data. According to the method, EIS data of a battery is collected, feature extraction and analysis are carried out on the EIS data by using a deep learning algorithm, and a complex mapping relation between the EIS data and the specific energy of the battery is established. The estimation system is based on the method, is provided with an efficient data processing module and an accurate model prediction module, and can quickly and accurately estimate the specific energy of the battery. Compared with a traditional method, the method has the advantages that the precision and efficiency of battery specific energy estimation are remarkably improved, and more reliable technical support is provided for battery performance evaluation and related application fields.

Description

Technical Field

[0001] This invention relates to a method for estimating the specific energy of a battery, and in particular to an estimation method for measuring the specific energy of a battery based on deep learning combined with EIS data. Background Technology

[0002] In the field of battery specific energy measurement, traditional methods mainly rely on battery discharge tests to calculate energy density, which is time-consuming and causes some damage to the battery. With the development of deep learning and electrochemical impedance spectroscopy (EIS) technology, a new measurement method has emerged. EIS can obtain information about the internal electrochemical processes of a battery by measuring its impedance response at different frequencies, thus indirectly reflecting the battery's specific energy data. However, the complexity and nonlinearity of EIS data make it difficult to directly extract specific energy information from the data. Deep learning models, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have powerful feature extraction and pattern recognition capabilities, enabling them to automatically learn complex feature mapping relationships related to battery specific energy from EIS data. Combining deep learning with EIS data allows for rapid and accurate estimation of battery specific energy, providing a new technical means for battery management systems. Summary of the Invention

[0003] Objective: This invention provides a method for estimating battery specific energy based on deep learning combined with EIS data. The method first collects battery operating data and EIS data, and preprocesses these data. Then, it utilizes convolutional neural networks and recurrent neural networks from deep learning algorithms to construct a battery specific energy prediction model, and extracts battery feature information through model training. The system uses EIS data as an offline database for error evaluation to improve the accuracy of model predictions. This method enables rapid and accurate estimation of battery specific energy, providing strong support for battery performance optimization.

[0004] Technical solution: The method for estimating battery specific energy based on deep learning combined with EIS data includes the following steps: Step S1, Data Collection and Preprocessing: Collect battery EIS and specific energy data to form the first feature; Step S2: Data feature extraction. Extracted data is fused with battery-related information to form a second feature. Step S3, Model Training and Validation: Construct a deep learning model and learn the complex mapping relationship from features to specific energy. Step S4, Energy Specific Estimation and Application: Input EIS data into the trained model, output the energy specific estimate of the battery, and monitor the battery performance status in real time. Step S5: Estimation system, real-time data acquisition, and rapid estimation of specific energy.

[0005] Furthermore, Step S1: Data Collection and Preprocessing. Under different charge / discharge states of the battery, an electrochemical impedance spectroscopy (EIS) instrument is used to apply AC signals of different frequencies to the battery, measuring its impedance response and collecting EIS data. This includes EIS data under different temperatures and cycle counts to cover performance changes under various operating conditions. Simultaneously, the specific energy data of the battery under these different states is recorded, i.e., the energy stored per unit mass or unit volume. This is calculated using the battery's discharge curve, rated capacity, and mass information. The collected EIS and specific energy data are preprocessed, including data cleaning, denoising, and normalization, to obtain the first feature. Data cleaning removes outliers and noisy data; denoising uses filters to reduce noise interference; and normalization scales the data to a uniform range, such as between 0 and 1, to improve model training efficiency and stability.

[0006] Step S2, Data Feature Extraction. By using Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to capture local features and patterns in the EIS data, features reflecting the battery's specific energy state (SES) – including the real and imaginary parts of the impedance and impedance characteristics at different frequencies – are extracted. The features extracted from the EIS data are then fused with other relevant battery information (temperature, cycle count) to form a comprehensive second feature. Then, using feature selection methods such as model-based feature selection and recursive feature elimination, the subset of features most strongly correlated with the battery's specific energy is selected to reduce feature dimensionality and improve model training efficiency and estimation accuracy.

[0007] Step S3, Model Training and Validation. Based on the complexity of the problem and the characteristics of the data, a multi-layer convolutional neural network (CNN) including convolutional layers, pooling layers, and fully connected layers is constructed to extract deep features from the EIS data. Then, specific energy regression prediction is performed through fully connected layers. This deep learning model learns the complex mapping relationship from features to specific energy. During training, Adam and SGD optimization algorithms are used, and the mean squared error (MSE) regression loss function is selected. Simultaneously, regularization and dropout techniques are employed to prevent model overfitting.

[0008] Step S4, Energy Specific EIS Estimation and Application. The energy specificity (EIS) data of the battery under test is input into the trained deep learning model. Based on the extracted features and the learned mapping relationship, the model outputs an estimated energy specificity value. This estimate can be used to monitor the battery's performance status in real time, providing a reference for battery use and maintenance. The estimation results can also be analyzed to evaluate the model's estimation accuracy and stability. If a large estimation error is found, the model structure can be further optimized, model parameters adjusted, and the amount of training data increased to improve the battery's energy specificity estimation performance.

[0009] Furthermore, In step S1, the collected data includes an electrochemical impedance spectroscopy (EIS) instrument, a temperature sensor, and a current and voltage sensor, used to acquire real-time EIS data, temperature, current, voltage, and other information of the battery. The model training unit and model inference unit are used to train the deep learning model and perform real-time estimation of the specific energy. High-performance GPU computing resources are required to accelerate the model training and inference process.

[0010] In step S2, several second features are obtained by using two major neural networks, CNN and RNN, to exhibit fluctuating peaks and troughs in multiple charge and discharge cycles, and further collect relevant battery specific energy data.

[0011] In step S2, the maximum information coefficient, Pearson correlation coefficient, Kendall correlation coefficient, grey relational analysis coefficient, and Spearman correlation coefficient for each feature are calculated respectively.

[0012] For the Maximum Information Coefficient (MIC): use the formula The calculation is performed, where M(g,a.b) is the feature matrix, B(n) is the upper limit of the grid division (a×b), and a and b are the average scores of the data; For the Pearson correlation coefficient (PC), the formula is used. The calculation is performed, where cor(x, y) is the covariance. It is the product of the standard deviations of the two variables; For Kendall's correlation coefficient: use the formula... The calculation is performed, where Nc is the number of consistent pairs, Nd is the number of inconsistent pairs, and n is the total number of samples; For the grey relational analysis coefficients: the formula is used. Perform calculations, where As a reference sequence, For comparing sequences, The resolution coefficient (usually taken as 0.5); For the Spearman correlation coefficient: use the formula... Perform calculations, where Let be the rank difference of the i-th sample, and n be the total number of samples; The Mean Rule is defined as follows: A Mean Rule is a rule used for feature weight fusion. Its core is to combine feature weights from different sources through a weighted average. and Combined into a single comprehensive weight, Mean Rule = ,in, r is a random number in the range [0, 0.5], and is a random number in the range [0, 1]. It is a constant with an extremely small value. Let represent the unequal solution vectors in the population, respectively. These are the best, good, and worst solution vectors among all vectors after the g-th iteration, respectively. It is based on the weights of each vector calculated using wavelet functions, and is used to calculate the weighted average of the vectors. As a scaling factor, according to the formula Update In the scenario of battery specific energy estimation, the setting and updating mechanism of these parameters can dynamically adjust the weighting method of the vector according to the changes in input features, thereby more accurately capturing the complex relationship between input features and battery specific energy. The convergence acceleration (CA) part utilizes the optimal vector to move the current vector within the search space, thereby improving global search capability: Here, CA is the value of the objective function; random is a random value following a standard normal distribution to ensure that each vector generated in INFO has a different step size; and the function f(x) represents the fitness function of x. In this patent, the fitness function f(x) is related to the estimation accuracy of the battery specific energy. Through a convergence acceleration module, the algorithm can find the parameter combination that improves the estimation accuracy of the battery specific energy more quickly. Based on the formula, two new vectors are generated. and Its generation formula is ,in, The vector scaling factor is determined by the formula. Update ; These are the unequal solution vectors in the population after the g-th iteration. Dynamic adjustment of the vector scaling helps balance the algorithm's exploration and utilization capabilities during the search process, thus more effectively searching for the optimal solution related to battery specific energy estimation in the solution space.

[0013] In step S1, the equivalent circuit model includes a series-connected ohmic resistor R0, a solid electrolyte layer module, a charge transfer module, and a diffused CPE element CPEW. The solid electrolyte layer module includes a parallel-connected double-layer resistor Ri and a double-layer CPE element CPEi. The charge transfer module includes a parallel-connected charge transfer resistor Rct and a charge transfer CPE element CPEct.

[0014] In the equivalent circuit model , , , R0 is the ohmic resistance, Ri is the double-layer resistance, Qi is the double-layer capacitance, Ni is the order of the double-layer impedance, Rct is the charge transfer resistance, Qct is the charge transfer capacitance, Nct is the order of the charge transfer impedance, RW is the diffusion impedance, QW is the diffusion layer capacitance, and NW is the order of the diffusion layer impedance.

[0015] In step S4, the vector weighted average algorithm optimizes the hyperparameters of the echo state network through a rule update stage, a vector combination stage, and a local search stage. In step S4, the echo state network is an artificial recurrent network with a three-layer structure: an input layer, a reservoir, and an output layer.

[0016] Compared with the prior art, the present invention has the following advantages: The method and system for estimating battery specific energy using deep learning combined with EIS data offers significant advantages. First, this method fully utilizes the detailed information on the internal electrochemical processes of the battery provided by EIS data. This information more accurately reflects the battery's health status and performance changes, acquiring a feature set composed of multiple primary and secondary features. Second, deep learning models, especially CNNs and RNNs, can effectively handle the complex nonlinear relationships in EIS data, automatically extracting features related to specific energy, thereby improving the accuracy and reliability of the estimation. Furthermore, the system further enhances the accuracy of model predictions by using EIS data as an offline database for error evaluation. Finally, this method features rapid response, enabling specific energy estimation in a short time, making it suitable for real-time monitoring and battery management systems. Attached Figure Description

[0017] Figure 1 Flowchart for battery reinforcement learning; Figure 2 Create a flowchart for the battery; Figure 3 Here is the flowchart of the battery controller algorithm; Figure 4 Battery specific energy acquisition flowchart; Figure 5A diagram for determining the charging conditions for battery specific energy. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0019] Step S1: Data Collection and Preprocessing. Under different charge and discharge states of the battery, an electrochemical impedance spectroscopy (EIS) instrument is used to apply AC signals of different frequencies to the battery, measuring its impedance response and collecting EIS data. This includes EIS data under different temperatures and cycle counts to cover performance changes under various operating conditions. Simultaneously, the specific energy data of the battery under these different states is recorded, i.e., the energy stored per unit mass or unit volume. This is calculated using the battery's discharge curve, rated capacity, and mass information. The collected EIS and specific energy data are preprocessed, including data cleaning, denoising, and normalization, to obtain the first feature. Data cleaning removes outliers and noisy data; denoising uses filters to reduce noise interference; and normalization scales the data to a uniform range, such as between 0 and 1, to improve model training efficiency and stability.

[0020] Data Collection: EIS data should cover different temperature conditions (setting multiple temperature points, such as -20℃, 0℃, 25℃, 40℃, 60℃, etc.) and different cycle counts (starting from the initial state of the battery, EIS data is collected after a certain number of cycles, such as 50, 100, 150, etc.) to ensure coverage of battery performance changes under various operating conditions. Temperature sensors are used to collect the battery's temperature information in real time during EIS data measurement; current and voltage sensors are used to collect the battery's current and voltage data in real time. Based on the battery's discharge curve (which can be recorded using high-precision data acquisition equipment to show the changes in voltage and current over time during battery discharge), the battery's rated capacity (the capacity value calibrated at the time of manufacture), and mass information (which can be obtained through measurement using a high-precision balance), the specific energy data of the battery under the above different states is calculated and recorded using the corresponding calculation formula (specific energy = battery discharge energy / battery mass, where battery discharge energy can be obtained by integrating the discharge curve). This represents the energy value stored by the battery per unit mass or unit volume.

[0021] Data preprocessing Data cleaning: Set data range thresholds. For impedance values, set reasonable upper and lower limits based on battery type and normal operating range. Data points exceeding this range are considered outliers and deleted. The 3σ criterion (for normally distributed data, data points deviating from the mean by more than three standard deviations are considered outliers) is used to identify and remove noisy data.

[0022] Noise reduction: Select appropriate filters, such as low-pass filters, band-pass filters, or wavelet filters. If the noise is mainly concentrated in the high-frequency band, a low-pass filter can be used, with an appropriate cutoff frequency set to filter out high-frequency noise interference and retain the effective signal in the EIS data.

[0023] Normalization: The min-max normalization method is used to scale the EIS data and specific energy data to between 0 and 1. For a data point X, its normalized value is calculated as Xnorm = Xmax - Xmin, where Xmin and Xmax are the minimum and maximum values ​​in the data set, respectively. Through the above preprocessing steps, the first feature is obtained.

[0024] Equivalent circuit model: An equivalent circuit model is adopted, consisting of a series-connected ohmic resistor R0, a solid electrolyte layer module, a charge transfer module, and a diffusion CPE element CPEW. The solid electrolyte layer module is composed of a parallel-connected double-layer resistor Ri and a double-layer CPE element CPEi, and the charge transfer module is composed of a parallel-connected charge transfer resistor Rct and a charge transfer CPE element CPEct. In this equivalent circuit model, R0 is the ohmic resistor, Ri is the double-layer resistor, Qi is the double-layer capacitor, Ni is the order of the double-layer impedance, Rct is the charge transfer resistor, Qct is the charge transfer capacitor, Nct is the order of the charge transfer impedance, Rw is the diffusion impedance, Qw is the diffusion layer capacitor, and Nw is the order of the diffusion layer impedance.

[0025] Step S2, Data Feature Extraction. By using Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to capture local features and patterns in the EIS data, features reflecting the battery's specific energy state (SES) – including the real and imaginary parts of the impedance and impedance characteristics at different frequencies – are extracted. The features extracted from the EIS data are then fused with other relevant battery information (temperature, cycle count) to form a comprehensive second feature. Then, using feature selection methods such as model-based feature selection and recursive feature elimination, the subset of features most strongly correlated with the battery's specific energy is selected to reduce feature dimensionality and improve model training efficiency and estimation accuracy.

[0026] Feature Extraction: For CNNs, a structure containing multiple convolutional and pooling layers can be designed. For example, a convolutional layer with a kernel size of (3×3) can be used first, followed by a max pooling layer, repeating this structure 2-3 times to capture local features in the EIS data. For RNNs, variants such as Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) can be selected to process time-series features in the EIS data. The preprocessed EIS data is input into the constructed CNN and RNN models. Through model training, the real and imaginary parts of the impedance reflecting the battery's specific energy state, as well as the impedance features at different frequencies, are extracted. During multiple charge-discharge cycles, the peaks and troughs of the output results of the two neural networks are observed, and relevant battery specific energy data are further collected.

[0027] Feature fusion: Representing features as vectors and then concatenating them along their dimensions.

[0028] Feature selection: Calculate the maximum information coefficient, Pearson correlation coefficient, Kendall correlation coefficient, grey relational analysis coefficient, and Spearman correlation coefficient for each feature. Evaluate the importance of each feature in predicting battery specific energy using model-based feature selection methods, such as logistic regression or decision tree models; or employ recursive feature elimination methods, starting with all features and removing those with the least impact on model performance each time, until a preset number of features is reached or model performance no longer improves. Select the subset of features most strongly correlated with battery specific energy to reduce feature dimensionality and improve the training efficiency and estimation accuracy of subsequent models.

[0029] For the Maximum Information Coefficient (MIC): use the formula The calculation is performed, where M(g,a.b) is the feature matrix, B(n) is the upper limit of the grid division (a×b), and a and b are the average scores of the data; For the Pearson correlation coefficient (PC), the formula is used. The calculation is performed, where cor(x, y) is the covariance. It is the product of the standard deviations of the two variables; For Kendall's correlation coefficient: use the formula... The calculation is performed, where Nc is the number of consistent pairs, Nd is the number of inconsistent pairs, and n is the total number of samples; For the grey relational analysis coefficients: the formula is used. Perform calculations, where As a reference sequence, For comparing sequences, The resolution coefficient (usually taken as 0.5); For the Spearman correlation coefficient: use the formula... Perform calculations, where Let be the rank difference of the i-th sample, and n be the total number of samples; The Mean Rule is defined as follows: A Mean Rule is a rule used for feature weight fusion. Its core is to combine feature weights from different sources through a weighted average. and Combined into a single comprehensive weight, Mean Rule = ,in, r is a random number in the range [0, 0.5], and is a random number in the range [0, 1]. It is a constant with an extremely small value. Let represent the unequal solution vectors in the population, respectively. These are the best, good, and worst solution vectors among all vectors after the g-th iteration, respectively. It is based on the weights of each vector calculated using wavelet functions, and is used to calculate the weighted average of the vectors. As a scaling factor, according to the formula Update In the scenario of battery specific energy estimation, the setting and updating mechanism of these parameters can dynamically adjust the weighting method of the vector according to the changes in input features, thereby more accurately capturing the complex relationship between input features and battery specific energy; The convergence acceleration (CA) part utilizes the optimal vector to move the current vector within the search space, thereby improving global search capability: Here, CA is the value of the objective function; random is a random value following a standard normal distribution to ensure that each vector generated in INFO has a different step size; and the function f(x) represents the fitness function of x. In this patent, the fitness function f(x) is related to the estimation accuracy of the battery specific energy. Through the convergence acceleration module, the algorithm can find the parameter combination that improves the estimation accuracy of the battery specific energy more quickly.

[0030] Based on the formula, two new vectors are generated. and Its generation formula is ,in, The vector scaling factor is determined by the formula. Update ; These are the unequal solution vectors in the population after the g-th iteration. Dynamic adjustment of the vector scaling helps balance the algorithm's exploration and utilization capabilities during the search process, thus more effectively searching for the optimal solution related to battery specific energy estimation in the solution space.

[0031] (3) Vector combination stage: The vector combination stage is used to improve local search capabilities in order to generate a new, better vector. The specific operation involves combining the two vectors obtained in the previous stage with the vectors under the given conditions. The vector below The algorithm performs merging. In this way, it can explore the solution space more meticulously within a local range, uncovering potential patterns related to battery specific energy, thereby improving the accuracy of battery specific energy estimation.

[0032] (4) Local search phase Effective local search capabilities can prevent INFO from getting trapped in local optima. By using local optimum locations and performing local search operations based on mean rules, the problem of the algorithm being affected by local optima can be effectively mitigated, thereby achieving global optima.

[0033] According to the formula, around the local optimum position Generate a new vector ,in It is by and The randomly generated new solution is derived from the formula. The computation increases the randomness of the algorithm, allowing it to search the solution space more effectively. These represent unequal solution vectors in the population; and They are two random numbers, defined as follows: As shown, and Both represent random numbers within the range (0, 1). During the local search process, this method of randomly combining new solutions helps the algorithm escape the trap of local optima and search the solution space more comprehensively to find a more accurate combination of parameters to estimate the battery's specific energy.

[0034] The mean rule in the INFO algorithm enhances search capability by combining the weighted average of two sets of vectors (one set of random vectors and one set of vectors containing local optimum, good, and worst solutions). The update rule operator combines the mean rule with the convergence acceleration module to update vector positions, thereby optimizing the algorithm's search capability and convergence speed. The scaling parameter plays a role in balancing search and development capabilities. The proposed vector combination operator effectively combines the global search and local development stages, improving search capability and avoiding getting trapped in local optima. Simultaneously, adaptive parameters smoothly transition from the search to the development stage, and finally, local search operations further accelerate the development and convergence process. Overall, these optimization strategies enable the algorithm to more accurately establish the relationship between input features and battery specific energy, improving the estimation accuracy of battery specific energy.

[0035] The echo-state network is an artificial recurrent neural network with a three-layer structure: an input layer, a reservoir, and an output layer. Let the echo-state network have K input layer neurons, N reservoir neurons, and L output layer neurons. At time n, the input variable u(n), the reservoir state x(n), and the output variable y(n) are respectively... , , At time (n+1), the input signal is u(n+1), and the internal state of the echo state network's reservoir is updated using the formula. The output equation at this time is... ,in, Let W be the connection weight matrix of the input layer, and W be the connection weight matrix of the reservoir. The weight connection matrix of the output layer. and These are the activation functions for the storage pool and the output layer, respectively. In this patent, the echo state network, based on the input feature data related to the battery's specific energy, outputs an estimated value of the battery's specific energy through its internal neuron computation and information transmission.

[0036] The estimation results of this invention (INFO-ESN: Vector Weighted Average Algorithm - Echo Neural Network Estimation) were compared with those of the comparative methods in a simulation. The simulation used 25C05 battery data. The horizontal axis in the graph represents the number of charge-discharge cycles, and the vertical axis represents the battery specific energy. The comparative methods included LSTM (Long Short-Term Memory) and ESN (Echo State Network). It is clear from the graph that INFO-ESN outperforms the comparative methods in estimating battery specific energy, indicating that the method proposed in this invention, based on deep learning combined with EIS data, has higher accuracy and reliability in measuring battery specific energy. The above are merely preferred embodiments of this invention and do not constitute any limitation on the invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in this invention without departing from the scope of the invention's technical solutions are considered to be within the protection scope of this invention.

Claims

1. An estimation method of measuring battery specific energy based on deep learning combined with EIS data, characterized by, Comprise the following steps: Step S1, data collection and pretreatment, collect battery EIS and specific energy data, form the first feature; Step S2, data feature extraction, extract data and battery related information fusion, form the second feature; Step S3, model training and verification, build deep learning model, learn the complex mapping relationship from feature to specific energy; Step S4, specific energy estimation and application, EIS data input into the trained model, output the specific energy estimation value of the battery, real-time monitoring of battery performance state; Step S5, estimation system, real-time data acquisition, rapid estimation of specific energy. 2.The estimating method of battery specific energy based on deep learning combined with EIS data according to claim 1, characterized in that, The data collection of step S1: under different charge and discharge states of the battery, use electrochemical impedance spectrometer to apply different frequency AC signal to the battery, measure its impedance response, collect the EIS data of the battery, including the EIS data of the battery under different temperature, different cycle times, to cover the performance change of the battery under various working conditions, at the same time, record the specific energy data of the battery under the above different states, that is, the energy value stored by the battery per unit mass or unit volume, which is calculated by the discharge curve of the battery, the rated capacity of the battery and the mass information. 3.The estimating method of battery specific energy based on deep learning combined with EIS data according to claim 2, characterized in that, The pretreatment of step S1: pretreatment of the collected EIS data and specific energy data, including data cleaning, denoising, normalization and the like, so as to obtain the first feature. 4.The estimating method and estimating system for measuring battery specific energy based on deep learning combining EIS data according to claim 2, characterized in that, In the pretreatment, equivalent circuit model is adopted, including series ohmic resistance R0, solid electrolyte layer module, charge transfer module and diffusion CPE element CPEW, the solid electrolyte layer module includes parallel double layer resistance Ri and double layer CPE element CPEi, and the charge transfer module includes parallel charge transfer resistance Rct and charge transfer CPE element CPEct. 5.The estimating method and estimating system for measuring battery specific energy based on deep learning combined with EIS data according to claim 4, characterized in that: In the equivalent circuit model, , , , R0 is ohmic resistance, Ri is double layer resistance, Qi is double layer capacitance, Ni is the order of double layer impedance, Rct is charge transfer resistance, Qct is charge transfer capacitance, Nct is the order of charge transfer impedance, RW is diffusion impedance, QW is diffusion layer capacitance and NW is the order of diffusion layer impedance. 6.The method of claim 1, wherein the method is based on deep learning combined with EIS data. The data feature extraction of step S2: by using convolutional neural network CNN and recurrent neural network RNN, local features and patterns in EIS data are captured, features of real part and imaginary part of impedance reflecting battery specific energy state are extracted, and impedance features under different frequencies are extracted, the features extracted from EIS data are fused with other related information of the battery to form a comprehensive second feature. 7.The estimating method of battery specific energy based on deep learning combined with EIS data according to claim 2, characterized in that, In step S2, the maximum information coefficient, Pearson correlation coefficient, Kendall correlation coefficient, grey correlation analysis coefficient and Spearman correlation coefficient of each feature are calculated respectively, For the maximum information coefficient (MIC): the formula was used, where M(g, a, b) is the characteristic matrix, B(n) is the upper limit value of the grid division (a x b), and a, b are the mean division values of the data. For the Pearson correlation coefficient (PC), the formula was used, where cor(x, y) is the covariance, is the product of the standard deviations of the two variables; For Kendall's correlation coefficient: the formula was used to calculate the coefficient, where Nc is the number of concordant pairs, Nd is the number of discordant pairs, and n is the total number of samples. For the grey correlation analysis coefficient: the formula is used to calculate, where is the reference sequence, is the comparison sequence, is the resolution coefficient (usually taken as 0.5); For the Spearman's correlation coefficient: the formula is used to calculate where is the difference in rank for the ith sample and n is the total number of samples. The definition of MeanRule is as follows: MeanRule is a rule for feature weight fusion, which is to combine feature weights from different sources into a comprehensive weight by weighted average and ) MeanRule = , where , r is a random number in the range of [0, 0.5], is a random number in the range of [0, 1], is a very small constant value, respectively represent the unequal solution vectors in the group, respectively are the best, better and worst solution vectors after the gth iteration, is the weight of each vector calculated based on the wavelet function, which is used to calculate the weighted average value of the vector, is the proportion factor, which is updated according to the formula , In the battery specific energy estimation scenario, the setting and updating mechanism of these parameters can dynamically adjust the weighting method of the vector according to the change of the input feature, so as to more accurately capture the complex relationship between the input feature and the battery specific energy. The convergence acceleration part (CA) moves the current vector in the search space with the best vector to improve the global search ability: where CA is the value of the objective function; randn is a random value subject to a standard normal distribution to ensure that each vector generated in INFO has a different step size, and the function f(x) represents the fitness function of x. In this patent, the fitness function f(x) is related to the estimation accuracy of the battery specific energy, and through the convergence acceleration module, the algorithm can find the parameter combination that improves the estimation accuracy of the battery specific energy faster. According to the formula, two new vectors are generated and The generation formula is wherein, is the vector scaling rate, which is updated according to the formula ; is the unequal solution vector in the group after the gth iteration. The dynamic adjustment of the vector scaling rate helps to balance the exploration and utilization ability of the algorithm in the search process, so as to more effectively search for the optimal solution related to the battery specific energy estimation in the solution space.​ 8.The method of claim 1, wherein the method is based on deep learning combined with EIS data. The model training and verification of the step S3: a multi-layer convolutional neural network including a convolutional layer, a pooling layer and a fully connected layer is constructed to extract deep features of the EIS data, and then the regression prediction of the specific energy is performed through the fully connected layer; in the training process, Adam and SGD optimization algorithms are adopted, a regression loss function of mean square error (MSE) is selected, and regularization and dropout techniques are adopted to prevent model overfitting. 9.The estimating method of battery specific energy based on deep learning combined with EIS data measurement of claim 8, wherein, The specific energy estimation and application of the step S4: the specific energy EIS data of the battery to be tested are input into the trained deep learning model, and according to the extracted features and the learned mapping relationship, the specific energy estimation value of the battery is output. 10.The estimating method of battery specific energy based on deep learning combined with EIS data according to claim 4, characterized in that, Its features are that: In the step S4, the vector weighted average algorithm optimizes the hyperparameters of the echo state network through an updating rule stage, a vector combination stage and a local search stage, and the echo state network is an artificial cycle with three layers of an input layer, a reserve pool and an output layer.

Images