Battery pack performance evaluation method and system based on big data

Abstract

The present invention provides a battery pack performance evaluation method and system based on big data, which relates to the technical field of battery performance evaluation, and comprises: obtaining charging and discharging data of a battery to be evaluated under different working conditions, performing denoising processing on the charging and discharging data, performing adaptive normalization processing to obtain standard charging and discharging data, performing equalization processing and respectively extracting time domain features and frequency domain features, and combining them to obtain a high-dimensional feature vector; performing a dimensionality reduction operation on the high-dimensional feature vector to obtain a low-dimensional embedded representation, generating an initial data set, dividing it into a training data set and a test data set, constructing an initial prediction model and performing hyperparameter optimization to obtain a first prediction model; obtaining real-time charging and discharging data and constructing a real-time feature vector, adding the real-time feature vector to the first prediction model to generate real-time performance parameters, filtering the real-time performance parameters, constructing a health status prediction model for prediction, and obtaining a comprehensive performance evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical field of battery performance evaluation, and in particular to a battery pack performance evaluation method and system based on big data. Background Art

[0002] As an important energy storage device, batteries are widely used in new energy vehicles, energy storage power stations, backup power supplies and other fields. With the continuous expansion of battery application, the evaluation and prediction of their performance has become increasingly important. Accurately evaluating the health status, remaining capacity and life of batteries is of great significance for ensuring the reliable operation of the system, optimizing energy management strategies and reducing operation and maintenance costs.

[0003] Traditional battery performance evaluation methods mainly include rest voltage method, internal resistance method and capacity method, which usually rely on specific test conditions and environments and are difficult to adapt to the complex working conditions in actual battery applications. These methods are often based on limited sample data and empirical models and are difficult to make full use of massive operating monitoring data, resulting in limited accuracy and applicability of evaluation results.

[0004] With the rapid development of sensor technology, communication technology and data storage technology, the massive data generated during battery operation has provided new opportunities for performance evaluation. By collecting key parameters such as battery voltage, current, temperature, etc., and combining big data analysis technologies such as machine learning and data mining, the performance degradation law of batteries can be mined from massive historical data, and a data-driven performance evaluation model can be constructed. However, the existing battery pack performance evaluation technology based on big data is difficult to ensure consistency, and data noise and outliers will interfere with the evaluation results. At the same time, the performance degradation mechanism of batteries is complex and has many influencing factors. It is still challenging to extract key features from massive data and build an effective evaluation model.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0006] The embodiments of the present invention provide a battery pack performance evaluation method and system based on big data, which can at least solve some problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a battery pack performance evaluation method based on big data, comprising:

[0008] Acquire the charging and discharging data of the battery to be evaluated under different working conditions, perform denoising on the charging and discharging data by combining wavelet transform, perform adaptive normalization processing by combining gradient descent algorithm to obtain standard charging and discharging data, perform equalization processing by synthetic minority class oversampling technology and extract time domain features and frequency domain features respectively, and combine them to obtain a high-dimensional feature vector;

[0009] Performing a dimensionality reduction operation on the high-dimensional feature vector by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generating an initial data set, dividing the initial data set into a training data set and a test data set, constructing an initial prediction model based on a bidirectional gated recurrent unit network and using the data in the training data set as input, and optimizing the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model;

[0010] Real-time charge and discharge data of the battery to be evaluated is obtained and a real-time feature vector is constructed. The real-time feature vector is added to the first prediction model to generate real-time performance parameters. The real-time performance parameters are filtered by an online learning algorithm. A health status prediction model is constructed in combination with a Bayesian learning algorithm. The filtered real-time performance parameters are used as input for prediction to obtain a comprehensive performance evaluation result.

[0011] In an optional embodiment,

[0012] The charging and discharging data of the battery to be evaluated under different working conditions are obtained, and the charging and discharging data are denoised by wavelet transform, and the standard charging and discharging data are obtained by adaptive normalization processing combined with the gradient descent algorithm. The synthetic minority class oversampling technology is used to perform equalization processing and extract time domain features and frequency domain features respectively, and the high-dimensional feature vectors are obtained by combination, including:

[0013] Acquire charge and discharge data of the battery to be evaluated under different working conditions, wherein the different working conditions include different ambient temperatures, different charge and discharge rates, and different usage durations, collect charge and discharge voltage and charge and discharge current data under different working conditions, and combine them to obtain the charge and discharge data;

[0014] Select a wavelet basis function and determine the corresponding decomposition scale, perform wavelet decomposition on the charge and discharge data through a wavelet transform operation to obtain a low-frequency approximate coefficient and a high-frequency detail coefficient corresponding to each decomposition scale, determine the minimum value of an unbiased estimate of the high-frequency detail coefficient for each decomposition scale through a threshold method, perform soft threshold processing on the high-frequency detail coefficient using the minimum value of the unbiased estimate as a threshold, remove high-frequency noise, perform wavelet reconstruction on the high-frequency detail coefficient and the low-frequency approximate coefficient after soft threshold processing, map them in combination with a normalization function and take minimizing the mean square error between the normalized data and the original data as the goal, construct a first objective function, and solve the first objective function to obtain the standard charge and discharge data;

[0015] According to the standard charge and discharge data, the number of data samples of each operating condition category is counted and minority class samples and majority class samples are determined, the nearest neighbor samples corresponding to the minority class samples are selected and combined with the interpolation algorithm to synthesize new minority class samples, and the synthesis is repeated until the number of minority class samples is balanced with the number of majority class samples to obtain balanced charge and discharge data;

[0016] The balanced charge and discharge data are subjected to feature extraction in time series to obtain corresponding time domain features, the balanced charge and discharge data are subjected to fast Fourier transform to obtain a spectrum diagram and feature extraction to obtain frequency domain features, and the frequency domain features and the time domain features are combined to obtain a high-dimensional feature vector.

[0017] In an optional embodiment,

[0018] The high-frequency detail coefficients after soft threshold processing and the low-frequency approximate coefficients are reconstructed by wavelet as shown in the following formula:

[0019]

[0020] Among them, x(t) represents the reconstructed signal, c j，k represents the low-frequency approximation coefficient, w c represents the adaptive low-frequency reconstruction threshold, φ jk (t) represents the low-frequency approximate wavelet basis function, j represents the decomposition level, J represents the maximum number of levels, k represents the translation parameter, and d j，k represents the high frequency detail coefficient, represents the adaptive high-frequency reconstruction threshold of the jth layer, ψ j，k (t) represents the high-frequency detail wavelet basis function.

[0021] In an optional embodiment,

[0022] The high-dimensional feature vector is reduced in dimension by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, an initial data set is generated, the initial data set is divided into a training data set and a test data set, an initial prediction model is constructed based on a bidirectional gated recurrent unit network and the data in the training data set is used as input, and the initial prediction model is optimized for hyperparameters in combination with an automatic machine learning algorithm to obtain a first prediction model including:

[0023] Determine a high-dimensional feature space based on the high-dimensional feature vector, calculate the Euclidean distance from the current data point to other data points for each data point in the high-dimensional feature space, select multiple data points closest to the current data point as the neighborhood of the current data point, construct an undirected weighted graph corresponding to the neighborhood, and determine the weight of the connecting edge between every two data points by using a Gaussian kernel function;

[0024] For the undirected weighted graph, the Laplace matrix is ​​calculated in combination with the connection edge weights and the Laplace matrix is ​​eigen-decomposed to obtain eigenvalues ​​and eigenvectors, the eigenvectors are arranged according to the size of the eigenvalues, the eigenvectors corresponding to the first five smallest non-zero eigenvalues ​​are retained and constitute a low-dimensional embedded representation after dimensionality reduction, and the initial data set is generated;

[0025] The initial data set is divided into a training data set and a test data set, a forward gated recurrent unit and a backward gated recurrent unit are stacked in sequence to form a deep bidirectional gated recurrent unit network, and a fully connected layer is used as an output layer to obtain an initial prediction model;

[0026] The training data in the training data set is added as input data to the initial prediction model, the forward gated recurrent unit processes the training data according to the time sequence corresponding to the training data to obtain a forward hidden state, the backward gated recurrent unit processes the training data according to the reverse time sequence corresponding to the training data to obtain a backward hidden state, the forward hidden state and the backward hidden state are concatenated to obtain an output state, the output state is mapped to a predicted performance indicator through the fully connected layer, the predicted performance indicator is obtained and compared with the true label of the current training data, the prediction loss function is constructed in combination with the mean square error and the absolute percentage error, and the prediction loss value is calculated;

[0027] Based on the predicted loss value, the value range of the hyperparameters in the initial prediction model is defined as a search space, wherein each hyperparameter corresponds to a dimension. The Gaussian process is used as a probability model to model the relationship between the objective function and the hyperparameters. The probability model is initialized and a hyperparameter combination is randomly selected for evaluation to obtain the objective function value. The objective function value is used as historical data to update the probability model. The update is repeated until a preset maximum number of iterations is reached to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is added to the initial prediction model to obtain the first prediction model.

[0028] In an optional embodiment,

[0029] The prediction loss function is constructed by combining the mean square error and the absolute percentage error. The prediction loss value is calculated as shown in the following formula:

[0030]

[0031] Among them, L(θ) represents the prediction loss value under the given parameter θ, α represents the mean square error weight coefficient, n represents the number of samples, and y i represents the true value of the i-th sample, represents the predicted value of the i-th sample, β represents the absolute percentage error weight coefficient, and λ represents the regularization parameter.

[0032] In an optional embodiment,

[0033] The real-time charge and discharge data of the battery to be evaluated is obtained and a real-time feature vector is constructed. The real-time feature vector is added to the first prediction model to generate real-time performance parameters. The real-time performance parameters are filtered by an online learning algorithm. A health status prediction model is constructed in combination with a Bayesian learning algorithm. The filtered real-time performance parameters are used as input for prediction. The comprehensive performance evaluation results obtained include:

[0034] Acquire real-time charge and discharge data of the battery to be evaluated and preprocess the real-time charge and discharge data, perform time synchronization and alignment on data with different sampling frequencies, extract charge and discharge capacity, internal resistance and peak power from the aligned real-time charge and discharge data as key features, construct a key feature space, determine a key feature subset in combination with a principal component analysis algorithm, use a feature vector in the key feature subset as a real-time feature vector and add it to the first prediction model to obtain real-time performance parameters;

[0035] The first prediction model sets the initial value and covariance matrix of the state variable according to the input real-time feature vector in combination with the Kalman filter algorithm, predicts the state variable and covariance matrix at the current moment according to the state estimation value and the state transition model at the previous moment, calculates the Kalman gain and updates the covariance matrix in combination with the observation value and the observation model at the current moment and the preset damping factor, and repeats the update until convergence to obtain the filtered real-time performance parameters;

[0036] According to historical data and expert experience, the prior probability distribution of the health status is estimated in combination with the Gaussian distribution. According to the relationship between the filtered real-time performance parameters and the health status, a likelihood function is constructed. According to the Bayesian theorem, the posterior probability distribution of the health status is calculated. Based on the likelihood function and the prior probability distribution and the posterior probability distribution, a health status prediction model is constructed. The filtered real-time performance parameters are added to the health status prediction model to obtain a comprehensive performance evaluation result.

[0037] In an optional embodiment,

[0038] Combining the current observation value and observation model, combined with the preset damping factor, the Kalman gain is calculated as shown in the following formula:

[0039]

[0040] Among them, K t represents the Kalman gain matrix, η represents the damping factor, represents the prediction error covariance matrix, H t represents the observation matrix, T represents the transpose, R t represents the observation noise covariance matrix, ()-1 Represents the inverse of a matrix.

[0041] A second aspect of an embodiment of the present invention provides a battery pack performance evaluation system based on big data, comprising:

[0042] The first unit is used to obtain the charging and discharging data of the battery to be evaluated under different working conditions, perform denoising on the charging and discharging data in combination with wavelet transform, perform adaptive normalization processing in combination with gradient descent algorithm to obtain standard charging and discharging data, perform equalization processing through synthetic minority class oversampling technology and respectively extract time domain features and frequency domain features, and combine them to obtain a high-dimensional feature vector;

[0043] The second unit is used to perform a dimensionality reduction operation on the high-dimensional feature vector through a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generate an initial data set, divide the initial data set into a training data set and a test data set, build an initial prediction model based on a bidirectional gated recurrent unit network and use the data in the training data set as input, and optimize the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model;

[0044] The third unit is used to obtain real-time charging and discharging data of the battery to be evaluated and construct a real-time feature vector, add the real-time feature vector to the first prediction model to generate real-time performance parameters, filter the real-time performance parameters through an online learning algorithm, and build a health status prediction model in combination with a Bayesian learning algorithm, and use the filtered real-time performance parameters as input for prediction to obtain a comprehensive performance evaluation result.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] According to a fourth aspect of the embodiments of the present invention,

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] In the present invention, the charge and discharge data are denoised by wavelet transform, which effectively removes high-frequency noise and interference in the data, improves the signal-to-noise ratio and quality of the data, fully mines the effective information contained in the charge and discharge data by extracting time domain features and frequency domain features, obtains the real-time charge and discharge data of the battery and constructs a real-time feature vector, which is added to the prediction model, thereby realizing the real-time evaluation and prediction of the battery performance, adopts an online learning algorithm to filter the real-time performance parameters, dynamically updates the model parameters, adapts to the changing trend of the battery performance, and improves the real-time and accuracy of the evaluation results, and obtains the comprehensive performance evaluation results of the battery by comprehensively considering the real-time performance parameters and the health status prediction results, which provides a quantitative basis for the battery status monitoring, performance optimization and maintenance decision-making, and in summary, the present invention not only improves the accuracy and reliability of the battery performance evaluation, but also provides data support for the health management, fault diagnosis and life prediction of the battery, and has significant technical advantages and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of a battery pack performance evaluation method based on big data according to an embodiment of the present invention;

[0054] Figure 2 It is a schematic diagram of the structure of a battery pack performance evaluation system based on big data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0057] Figure 1 FIG. 1 is a flow chart of a battery pack performance evaluation method based on big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S1. Obtain the charge and discharge data of the battery to be evaluated under different working conditions, perform denoising on the charge and discharge data by combining wavelet transform, perform adaptive normalization processing by combining gradient descent algorithm to obtain standard charge and discharge data, perform equalization processing by synthetic minority class oversampling technology and extract time domain features and frequency domain features respectively, and combine to obtain a high-dimensional feature vector;

[0059] The wavelet transform is a signal processing technique that decomposes a signal into parts of different scales (or frequencies) in order to analyze different features of the signal. The gradient descent is an optimization algorithm used to minimize (or maximize) a function. The synthetic minority class oversampling is a technique for processing unbalanced data sets, particularly suitable for classification problems. The high-dimensional feature vector refers to a vector containing a large number of features, which usually appears when processing complex data (such as images, text or genetic data).

[0060] In an optional embodiment,

[0061] The charging and discharging data of the battery to be evaluated under different working conditions are obtained, and the charging and discharging data are denoised by wavelet transform, and the standard charging and discharging data are obtained by adaptive normalization processing combined with the gradient descent algorithm. The synthetic minority class oversampling technology is used to perform equalization processing and extract time domain features and frequency domain features respectively, and the high-dimensional feature vectors are obtained by combination, including:

[0062] Acquire charge and discharge data of the battery to be evaluated under different working conditions, wherein the different working conditions include different ambient temperatures, different charge and discharge rates, and different usage durations, collect charge and discharge voltage and charge and discharge current data under different working conditions, and combine them to obtain the charge and discharge data;

[0063] Select a wavelet basis function and determine the corresponding decomposition scale, perform wavelet decomposition on the charge and discharge data through a wavelet transform operation to obtain a low-frequency approximate coefficient and a high-frequency detail coefficient corresponding to each decomposition scale, determine the minimum value of an unbiased estimate of the high-frequency detail coefficient for each decomposition scale through a threshold method, perform soft threshold processing on the high-frequency detail coefficient using the minimum value of the unbiased estimate as a threshold, remove high-frequency noise, perform wavelet reconstruction on the high-frequency detail coefficient and the low-frequency approximate coefficient after soft threshold processing, map them in combination with a normalization function and take minimizing the mean square error between the normalized data and the original data as the goal, construct a first objective function, and solve the first objective function to obtain the standard charge and discharge data;

[0064] According to the standard charge and discharge data, the number of data samples of each operating condition category is counted and minority class samples and majority class samples are determined, the nearest neighbor samples corresponding to the minority class samples are selected and combined with the interpolation algorithm to synthesize new minority class samples, and the synthesis is repeated until the number of minority class samples is balanced with the number of majority class samples to obtain balanced charge and discharge data;

[0065] The balanced charge and discharge data are subjected to feature extraction in time series to obtain corresponding time domain features, the balanced charge and discharge data are subjected to fast Fourier transform to obtain a spectrum diagram and feature extraction to obtain frequency domain features, and the frequency domain features and the time domain features are combined to obtain a high-dimensional feature vector.

[0066] The charge and discharge rate is an indicator that measures the charging or discharging rate of a battery within a certain period of time. The low-frequency approximation coefficient is the low-frequency component of the signal obtained by wavelet transform, which represents the main characteristics or outline of the signal and usually retains most of the energy and main information of the signal. The high-frequency detail coefficient is the high-frequency component of the signal obtained by wavelet transform, which represents the details and changing parts of the signal. The unbiased estimation is an important concept in statistics, which means that the expected value of the estimated value is equal to the true value of the estimated parameter. The soft threshold processing is a signal denoising method, which is often used in wavelet transform. The wavelet reconstruction is the process of reconstructing the original signal from the coefficients obtained from wavelet decomposition. The original signal can be reconstructed by inverse transforming the low-frequency approximation coefficient and the high-frequency detail coefficient. The minority class sample refers to the sample belonging to the class with a smaller number of samples in the unbalanced data set. The majority class sample refers to the sample belonging to the class with a larger number of samples in the unbalanced data set. The nearest neighbor sample refers to other samples closest to the given sample in the feature space.

[0067] Select different ambient temperatures, charge and discharge rates, and usage durations as different operating conditions. Under each operating condition, perform charge and discharge tests on the battery to be evaluated, collect voltage and current data during the charge and discharge process, and combine the collected voltage and current data to obtain charge and discharge data under each operating condition.

[0068] Select wavelet basis function, determine the scale of wavelet decomposition, usually select 3 to 5 levels of decomposition, perform wavelet transform on the charge and discharge data under each working condition, obtain low-frequency approximate coefficients and high-frequency detail coefficients under each decomposition scale, determine the minimum value of unbiased estimation as the threshold value for the high-frequency detail coefficient under each decomposition scale by threshold method, use soft threshold processing method to denoise the high-frequency detail coefficient, remove high-frequency noise, perform wavelet reconstruction on the denoised high-frequency detail coefficient and the low-frequency approximate coefficient of the corresponding scale, and obtain denoised charge and discharge data;

[0069] Normalize the charge and discharge data after wavelet reconstruction, map the data to the interval [0, 1], select the normalization function and construct the objective function, take the minimization of the mean square error between the normalized data and the original data as the optimization goal, use the gradient descent algorithm to solve the constructed objective function, obtain the optimal normalization function parameters, and use the optimal normalization function to normalize the charge and discharge data under all working conditions to obtain standardized charge and discharge data;

[0070] Count the number of data samples under each operating condition category (such as temperature, rate, number of cycles), determine the minority class samples (the category with a smaller number) and the majority class samples (the category with a larger number), calculate the Euclidean distance between each minority class sample and all samples, select the k samples with the closest distance as its nearest neighbor samples, and use an interpolation algorithm (such as the SMOTE algorithm) to randomly interpolate and generate new synthetic samples in the feature space for each minority class sample and its nearest neighbor sample, repeat the generation until the number of minority class samples is balanced with the number of majority class samples, combine the original majority class samples with the synthesized minority class samples to obtain balanced charge and discharge data;

[0071] The balanced charge and discharge data is subjected to time domain feature extraction, a spectrum is obtained by fast Fourier transform and frequency domain feature extraction is performed, and the extracted time domain features and frequency domain features are combined to obtain the high-dimensional feature vector.

[0072] In this embodiment, by considering factors such as different ambient temperatures, charge and discharge rates, and usage time, the charge and discharge data of the battery under various working conditions are obtained, which fully reflects the actual usage of the battery and provides sufficient data support for subsequent health status assessment and life prediction. Through wavelet decomposition and denoising, high-frequency noise and interference in the charge and discharge data are effectively removed, the signal-to-noise ratio and quality of the data are improved, and a good foundation is laid for subsequent analysis. By performing unbalanced sample processing on the standard charge and discharge data, the problem of unbalanced sample quantity under different working conditions is solved, the negative impact of data imbalance on subsequent modeling and prediction is avoided, and the generalization performance of the model is improved. The time domain and frequency domain features of the balanced charge and discharge data are extracted, and the charge and discharge characteristics of the battery are characterized from multiple angles, which provides more information and judgment basis for subsequent health status assessment and life prediction. In summary, this embodiment constructs a high-quality, balanced, and rich battery charge and discharge data feature set through systematic data processing and feature extraction, which lays a solid data foundation for subsequent health status assessment, fault diagnosis, and life prediction.

[0073] In an optional embodiment,

[0074] The high-frequency detail coefficients after soft threshold processing and the low-frequency approximate coefficients are reconstructed by wavelet as shown in the following formula:

[0075]

[0076] Among them, x(t) represents the reconstructed signal, c j，k represents the low-frequency approximation coefficient, w c represents the adaptive low-frequency reconstruction threshold, φ j，k (t) represents the low-frequency approximate wavelet basis function, j represents the decomposition level, J represents the maximum number of levels, k represents the translation parameter, and d j，k represents the high frequency detail coefficient, represents the adaptive high-frequency reconstruction threshold of the jth layer, ψ j，k (t) represents the high-frequency detail wavelet basis function.

[0077] In this embodiment, by using an adaptive threshold, the threshold size can be automatically adjusted according to the characteristics of the signal and the noise level, so as to suppress noise more flexibly and effectively. By performing threshold processing and reconstruction on coefficients of different scales, the noise can be effectively removed while retaining the main features and detail information of the signal. By weightedly combining low-frequency approximate coefficients and high-frequency detail coefficients, the denoised signal is reconstructed, which can better retain the original form and detail features of the signal. By reasonably selecting the wavelet basis function and the number of decomposition layers, the real information of the signal can be retained to the maximum extent while denoising, thereby improving the fidelity of the reconstructed signal. In summary, this embodiment can effectively improve the signal-to-noise ratio of the signal and retain the real features of the signal, thus laying a good foundation for subsequent feature extraction, health status assessment and life prediction.

[0078] S2. Performing a dimensionality reduction operation on the high-dimensional feature vector by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generating an initial data set, dividing the initial data set into a training data set and a test data set, constructing an initial prediction model based on a bidirectional gated recurrent unit network and taking the data in the training data set as input, and optimizing the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model;

[0079] The manifold learning is a type of nonlinear dimensionality reduction technology, which aims to find low-dimensional manifold structures in high-dimensional data, that is, the data actually exists in a space with lower dimensionality than its apparent dimension. The low-dimensional embedding representation refers to mapping high-dimensional data into a low-dimensional space through dimensionality reduction technology for easy visualization, storage and processing. The bidirectional gated recurrent unit network is an improved recurrent neural network.

[0080] In an optional embodiment,

[0081] The high-dimensional feature vector is reduced in dimension by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, an initial data set is generated, the initial data set is divided into a training data set and a test data set, an initial prediction model is constructed based on a bidirectional gated recurrent unit network and the data in the training data set is used as input, and the initial prediction model is optimized for hyperparameters in combination with an automatic machine learning algorithm to obtain a first prediction model including:

[0082] Determine a high-dimensional feature space based on the high-dimensional feature vector, calculate the Euclidean distance from the current data point to other data points for each data point in the high-dimensional feature space, select multiple data points closest to the current data point as the neighborhood of the current data point, construct an undirected weighted graph corresponding to the neighborhood, and determine the weight of the connecting edge between every two data points by using a Gaussian kernel function;

[0083] For the undirected weighted graph, the Laplace matrix is ​​calculated in combination with the connection edge weights and the Laplace matrix is ​​eigen-decomposed to obtain eigenvalues ​​and eigenvectors, the eigenvectors are arranged according to the size of the eigenvalues, the eigenvectors corresponding to the first five smallest non-zero eigenvalues ​​are retained and constitute a low-dimensional embedded representation after dimensionality reduction, and the initial data set is generated;

[0084] The initial data set is divided into a training data set and a test data set, a forward gated recurrent unit and a backward gated recurrent unit are stacked in sequence to form a deep bidirectional gated recurrent unit network, and a fully connected layer is used as an output layer to obtain an initial prediction model;

[0085] The training data in the training data set is added as input data to the initial prediction model, the forward gated recurrent unit processes the training data according to the time sequence corresponding to the training data to obtain a forward hidden state, the backward gated recurrent unit processes the training data according to the reverse time sequence corresponding to the training data to obtain a backward hidden state, the forward hidden state and the backward hidden state are concatenated to obtain an output state, the output state is mapped to a predicted performance indicator through the fully connected layer, the predicted performance indicator is obtained and compared with the true label of the current training data, the prediction loss function is constructed in combination with the mean square error and the absolute percentage error, and the prediction loss value is calculated;

[0086] Based on the predicted loss value, the value range of the hyperparameters in the initial prediction model is defined as a search space, wherein each hyperparameter corresponds to a dimension. The Gaussian process is used as a probability model to model the relationship between the objective function and the hyperparameters. The probability model is initialized and a hyperparameter combination is randomly selected for evaluation to obtain the objective function value. The objective function value is used as historical data to update the probability model. The update is repeated until a preset maximum number of iterations is reached to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is added to the initial prediction model to obtain the first prediction model.

[0087] The high-dimensional feature space refers to a space in which the feature dimension of the data is very high. Each data point is represented by a high-dimensional feature vector in this space. The neighborhood refers to a set of adjacent points around a point in the feature space. The undirected weighted graph is a graph structure in which the edges have no direction but each edge has a weight. The Gaussian kernel function is a commonly used kernel function, which is widely used in support vector machines and kernel methods. By mapping data from a low-dimensional space to a high-dimensional space, nonlinear problems are transformed into linear problems in high-dimensional space. The absolute percentage error is a metric for evaluating the performance of a prediction model, which represents the average percentage error between the predicted value and the actual value. The Gaussian process is a non-parametric Bayesian method for regression and classification tasks.

[0088] The extracted high-dimensional feature vector is used as the data point in the high-dimensional feature space. For each data point, the Euclidean distance between it and other data points is calculated. The k data points closest to the current data point are selected as its neighborhood. Each data point is used as a node. There are connecting edges between the data points in the neighborhood. An undirected weighted graph is constructed. The Gaussian kernel function is used to calculate the weight of the connecting edge between each two data points, where the weight reflects the similarity between the data points.

[0089] According to the edge weights of the undirected weighted graph, the Laplace matrix is ​​calculated, and the Laplace matrix is ​​eigen-decomposed to obtain the eigenvalues ​​and eigenvectors. The eigenvectors are sorted according to the size of the corresponding eigenvalues, and the eigenvectors corresponding to the first five smallest non-zero eigenvalues ​​are selected to form a low-dimensional embedded representation after dimensionality reduction, and the low-dimensional embedded representation is used as the initial data set;

[0090] The initial data set is divided into a training data set and a test data set, and a deep bidirectional gated recurrent unit network is constructed, including a forward gated recurrent unit and a backward gated recurrent unit. The forward gated recurrent unit and the backward gated recurrent unit are connected in a stacked manner, and a fully connected layer is added on the top of the network as an output layer to obtain an initial prediction model.

[0091] The training data in the training data set is used as input and added to the initial prediction model, wherein the forward gated recurrent unit processes the training data in the time sequence to obtain the forward hidden state, and the backward gated recurrent unit processes the training data in the reverse time sequence to obtain the backward hidden state, the forward hidden state and the backward hidden state are concatenated to obtain the output state, the output state is mapped to the predicted performance index through the fully connected layer to obtain the predicted performance index, the predicted performance index is compared with the true label of the training data, a prediction loss function is constructed, and the prediction loss value is calculated by combining the mean square error and the absolute percentage error;

[0092] According to the predicted loss value, the hyperparameter value range of the initial prediction model is defined as the search space. Each hyperparameter corresponds to a dimension in the search space. The Gaussian process is selected as the probability model to model the relationship between the objective function and the hyperparameters. The probability model is initialized, and a random hyperparameter combination is selected for evaluation to obtain the objective function value. The objective function value is used as historical data to update the probability model, and the evaluation and update are repeated until the preset maximum number of iterations is reached to obtain the optimal hyperparameter combination.

[0093] The optimal hyperparameter combination obtained by Bayesian optimization is added to the initial prediction model, and the initial prediction model is retrained using the optimized hyperparameters to obtain the first prediction model.

[0094] In this embodiment, by calculating the Euclidean distance between high-dimensional feature vectors and selecting the nearest neighbor to construct an undirected weighted graph, the local structure and similarity relationship between data points are effectively captured, laying the foundation for subsequent dimensionality reduction and prediction. The low-dimensional embedded representation after dimensionality reduction retains the essential structure of the original high-dimensional feature space, while reducing data redundancy and noise, improving the training efficiency and generalization ability of subsequent prediction models, and obtaining a more comprehensive and rich feature representation through the concatenation of forward and backward hidden states, thereby improving the expressive power of the prediction model. Bayesian optimization can find the global optimal or approximately optimal hyperparameter combination within a limited number of evaluations, greatly reducing the workload of manual parameter adjustment and improving the performance of the prediction model. In summary, this embodiment constructs an efficient and accurate battery health status assessment and life prediction model, providing reliable support for the decision-making of the battery management system.

[0095] In an optional embodiment,

[0096] The prediction loss function is constructed by combining the mean square error and the absolute percentage error. The prediction loss value is calculated as shown in the following formula:

[0097]

[0098] Among them, L(θ) represents the prediction loss value under the given parameter θ, α represents the mean square error weight coefficient, n represents the number of samples, and y i represents the true value of the i-th sample, represents the predicted value of the i-th sample, β represents the absolute percentage error weight coefficient, and λ represents the regularization parameter.

[0099] In this embodiment, by minimizing the mean square error, the prediction model is prompted to continuously adjust parameters during the training process to fit the real data distribution and improve the accuracy of the prediction. By minimizing the absolute percentage error, the prediction model pays more attention to the reduction of relative error during the training process, thereby improving the reliability and interpretability of the prediction results. By adjusting the size of the weight coefficient, the degree of emphasis of the prediction model on different error metrics can be flexibly controlled. By controlling the regularization parameter, a trade-off can be made between fitting performance and model simplicity to prevent overfitting and improve the generalization ability of the model. In summary, this embodiment can guide the model to pay attention to both absolute error and relative error during the training process, balance fitting performance and model complexity, improve the accuracy, reliability and generalization ability of the prediction results, and provide reliable decision support for the battery management system.

[0100] S3. Obtain real-time charge and discharge data of the battery to be evaluated and construct a real-time feature vector, add the real-time feature vector to the first prediction model to generate real-time performance parameters, filter the real-time performance parameters through an online learning algorithm, and build a health status prediction model in combination with the Bayesian learning algorithm. Use the filtered real-time performance parameters as input for prediction to obtain a comprehensive performance evaluation result.

[0101] The real-time feature vector refers to a feature vector extracted and updated in real time from a data stream. The performance parameter is a standard and indicator for evaluating the performance of a model or system. The online learning algorithm is a type of machine learning algorithm that can gradually update the model when the data stream arrives, rather than training on the entire data set at one time. The Bayesian learning algorithm is based on the Bayesian theorem and performs learning and prediction by updating the posterior probability distribution.

[0102] In an optional embodiment,

[0103] The real-time charge and discharge data of the battery to be evaluated is obtained and a real-time feature vector is constructed. The real-time feature vector is added to the first prediction model to generate real-time performance parameters. The real-time performance parameters are filtered by an online learning algorithm. A health status prediction model is constructed in combination with a Bayesian learning algorithm. The filtered real-time performance parameters are used as input for prediction. The comprehensive performance evaluation results obtained include:

[0104] Acquire real-time charge and discharge data of the battery to be evaluated and preprocess the real-time charge and discharge data, perform time synchronization and alignment on data with different sampling frequencies, extract charge and discharge capacity, internal resistance and peak power from the aligned real-time charge and discharge data as key features, construct a key feature space, determine a key feature subset in combination with a principal component analysis algorithm, use a feature vector in the key feature subset as a real-time feature vector and add it to the first prediction model to obtain real-time performance parameters;

[0105] The first prediction model sets the initial value and covariance matrix of the state variable according to the input real-time feature vector in combination with the Kalman filter algorithm, predicts the state variable and covariance matrix at the current moment according to the state estimation value and the state transition model at the previous moment, calculates the Kalman gain and updates the covariance matrix in combination with the observation value and the observation model at the current moment and the preset damping factor, and repeats the update until convergence to obtain the filtered real-time performance parameters;

[0106] According to historical data and expert experience, the prior probability distribution of the health status is estimated in combination with the Gaussian distribution. According to the relationship between the filtered real-time performance parameters and the health status, a likelihood function is constructed. According to the Bayesian theorem, the posterior probability distribution of the health status is calculated. Based on the likelihood function and the prior probability distribution and the posterior probability distribution, a health status prediction model is constructed. The filtered real-time performance parameters are added to the health status prediction model to obtain a comprehensive performance evaluation result.

[0107] The internal resistance refers to the resistance to the flow of current inside a battery or electrical device. The key feature space refers to the space composed of features that can best describe and distinguish data categories selected from a high-dimensional data set in data analysis and machine learning. The key feature subset is a subset selected from the original feature set, which contains the features that have the greatest impact on model performance. The state variable is a set of variables used to describe the state of the system, and is widely used in fields such as power systems, control systems, and random processes. The damping factor refers to a parameter in the Kalman filter algorithm, which is used to control the update speed of the system state estimation and the degree of trust in the observed value. The prior probability distribution refers to the estimate of the probability distribution of parameters or events based on prior knowledge or assumptions before the data is observed. The posterior probability distribution refers to the probability distribution of parameters or events obtained by combining the prior distribution and the data likelihood after the data is observed.

[0108] Collect the charging and discharging data of the battery to be evaluated in real time, including parameters such as voltage, current, and temperature. Select a suitable time interval (such as 1 second) to resample all parameters to ensure that they are aligned at the same time point. Extract key features that can reflect the performance of the battery from the aligned real-time charging and discharging data. Combine the extracted key features into a feature vector to construct a key feature space. Each feature vector represents the performance status of the battery at a certain moment.

[0109] Combined with the principal component analysis algorithm, the key feature space is reduced in dimension, the key feature subset is determined, the feature combination that best represents the battery performance is selected, the feature vector in the key feature subset is used as the real-time feature vector, and the real-time feature vector is added to the first prediction model for predicting the real-time performance parameters;

[0110] Input the real-time feature vector into the first prediction model, combine with the Kalman filter algorithm, set the initial value and covariance matrix of the state variable, predict the state variable and covariance matrix at the current moment according to the state estimation value and the state transition model at the previous moment, calculate the Kalman gain by combining the observation value and the observation model at the current moment, introduce the preset damping factor, adjust the size of the Kalman gain, update the state variable and covariance matrix, repeat the prediction and update until the filtering process converges, and obtain the real-time performance parameters after filtering, including indicators such as battery capacity and internal resistance;

[0111] According to historical data and expert experience, the prior probability distribution of health status is estimated, and it is assumed to obey Gaussian distribution. According to the relationship between the filtered real-time performance parameters and health status, the likelihood function is constructed. According to Bayes' theorem, the posterior probability distribution of health status is calculated by combining the prior probability distribution and the likelihood function. Based on the prior probability distribution, the likelihood function and the posterior probability distribution, a health status prediction model is constructed.

[0112] The filtered real-time performance parameters are added to the health status prediction model. The health status prediction model predicts the health status of the battery based on the input real-time performance parameters, and gives a comprehensive performance evaluation result of the battery by combining the health status prediction result and the real-time performance parameters.

[0113] In this embodiment, key features are extracted and real-time feature vectors are constructed, which can dynamically reflect the state of the battery and provide a data basis for subsequent performance evaluation and prediction. Dimensionality reduction algorithms such as principal component analysis are used to extract the most representative key feature subset from the high-dimensional feature space, which effectively reduces the data dimension and reduces the computational complexity. Through feature optimization, redundant and highly correlated features are removed, and the generalization ability and prediction accuracy of the model are improved. By estimating the prior probability distribution of the health state, combining the likelihood function and Bayesian inference, the posterior probability distribution of the health state is obtained, the health state of the battery is classified and predicted, and a quantitative health state evaluation result is provided. In summary, this embodiment can monitor the performance of the battery in real time, predict its health state and remaining life, and provide support for maintenance decisions.

[0114] In an optional embodiment,

[0115] Combining the current observation value and observation model, combined with the preset damping factor, the Kalman gain is calculated as shown in the following formula:

[0116]

[0117] Among them, K t represents the Kalman gain matrix, η represents the damping factor, represents the prediction error covariance matrix, H t represents the observation matrix, T represents the transpose, R t represents the observation noise covariance matrix, () -1 Represents the inverse of a matrix.

[0118] In this embodiment, by reasonably setting the damping factor, the best balance can be found between prediction and observation, and the accuracy of state estimation can be improved. The introduction of the damping factor can suppress the impact of observation noise on state estimation to a certain extent, which helps to improve the robustness of the filtering algorithm so that it can still maintain good performance in the face of uncertainty and noise. In summary, this embodiment helps to obtain accurate, reliable and real-time state estimation, and provides strong technical support for subsequent calculations and performance evaluation.

[0119] Figure 2 FIG. 1 is a schematic diagram of a battery pack performance evaluation system based on big data according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0120] The first unit is used to obtain the charging and discharging data of the battery to be evaluated under different working conditions, perform denoising on the charging and discharging data in combination with wavelet transform, perform adaptive normalization processing in combination with gradient descent algorithm to obtain standard charging and discharging data, perform equalization processing through synthetic minority class oversampling technology and respectively extract time domain features and frequency domain features, and combine them to obtain a high-dimensional feature vector;

[0121] The second unit is used to perform a dimensionality reduction operation on the high-dimensional feature vector through a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generate an initial data set, divide the initial data set into a training data set and a test data set, build an initial prediction model based on a bidirectional gated recurrent unit network and use the data in the training data set as input, and optimize the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model;

[0122] The third unit is used to obtain real-time charging and discharging data of the battery to be evaluated and construct a real-time feature vector, add the real-time feature vector to the first prediction model to generate real-time performance parameters, filter the real-time performance parameters through an online learning algorithm, and build a health status prediction model in combination with a Bayesian learning algorithm, and use the filtered real-time performance parameters as input for prediction to obtain a comprehensive performance evaluation result.

[0123] According to a third aspect of the embodiments of the present invention,

[0124] An electronic device is provided, comprising:

[0125] processor;

[0126] a memory for storing processor-executable instructions;

[0127] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0128] According to a fourth aspect of the embodiments of the present invention,

[0129] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0130] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery pack performance evaluation method based on big data, characterized in that: include: Acquire the charging and discharging data of the battery to be evaluated under different working conditions, perform denoising on the charging and discharging data by combining wavelet transform, perform adaptive normalization processing by combining gradient descent algorithm to obtain standard charging and discharging data, perform equalization processing by synthetic minority class oversampling technology and extract time domain features and frequency domain features respectively, and combine them to obtain a high-dimensional feature vector; Performing a dimensionality reduction operation on the high-dimensional feature vector by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generating an initial data set, dividing the initial data set into a training data set and a test data set, constructing an initial prediction model based on a bidirectional gated recurrent unit network and using the data in the training data set as input, and optimizing the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model; Real-time charge and discharge data of the battery to be evaluated is obtained and a real-time feature vector is constructed. The real-time feature vector is added to the first prediction model to generate real-time performance parameters. The real-time performance parameters are filtered by an online learning algorithm. A health status prediction model is constructed in combination with a Bayesian learning algorithm. The filtered real-time performance parameters are used as input for prediction to obtain a comprehensive performance evaluation result.

2. The method according to claim 1, characterized in that The charging and discharging data of the battery to be evaluated under different working conditions are obtained, and the charging and discharging data are denoised by wavelet transform, and the standard charging and discharging data are obtained by adaptive normalization processing combined with the gradient descent algorithm. The synthetic minority class oversampling technology is used to perform equalization processing and extract time domain features and frequency domain features respectively, and the high-dimensional feature vectors are obtained by combination, including: Acquire charge and discharge data of the battery to be evaluated under different working conditions, wherein the different working conditions include different ambient temperatures, different charge and discharge rates, and different usage durations, collect charge and discharge voltage and charge and discharge current data under different working conditions, and combine them to obtain the charge and discharge data; Select a wavelet basis function and determine the corresponding decomposition scale, perform wavelet decomposition on the charge and discharge data through a wavelet transform operation to obtain a low-frequency approximate coefficient and a high-frequency detail coefficient corresponding to each decomposition scale, determine the minimum value of an unbiased estimate of the high-frequency detail coefficient for each decomposition scale through a threshold method, perform soft threshold processing on the high-frequency detail coefficient using the minimum value of the unbiased estimate as a threshold, remove high-frequency noise, perform wavelet reconstruction on the high-frequency detail coefficient and the low-frequency approximate coefficient after soft threshold processing, map them in combination with a normalization function and take minimizing the mean square error between the normalized data and the original data as the goal, construct a first objective function, and solve the first objective function to obtain the standard charge and discharge data; According to the standard charge and discharge data, the number of data samples of each operating condition category is counted and minority class samples and majority class samples are determined, the nearest neighbor samples corresponding to the minority class samples are selected and combined with the interpolation algorithm to synthesize new minority class samples, and the synthesis is repeated until the number of minority class samples is balanced with the number of majority class samples to obtain balanced charge and discharge data; The balanced charge and discharge data are subjected to feature extraction in time series to obtain corresponding time domain features, the balanced charge and discharge data are subjected to fast Fourier transform to obtain a spectrum diagram and feature extraction to obtain frequency domain features, and the frequency domain features and the time domain features are combined to obtain a high-dimensional feature vector.

3. The method according to claim 2, characterized in that The high-frequency detail coefficients after soft threshold processing and the low-frequency approximate coefficients are reconstructed by wavelet as shown in the following formula: Among them, x(t) represents the reconstructed signal, c j，k represents the low-frequency approximation coefficient, w c represents the adaptive low-frequency reconstruction threshold, φ j，k (t) represents the low-frequency approximate wavelet basis function, j represents the decomposition level, J represents the maximum number of levels, k represents the translation parameter, and d j，k represents the high frequency detail coefficient, represents the adaptive high-frequency reconstruction threshold of the jth layer, ψ j，k (t) represents the high-frequency detail wavelet basis function.

4. The method according to claim 1, characterized in that The high-dimensional feature vector is reduced in dimension by a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, an initial data set is generated, the initial data set is divided into a training data set and a test data set, an initial prediction model is constructed based on a bidirectional gated recurrent unit network and the data in the training data set is used as input, and the initial prediction model is optimized for hyperparameters in combination with an automatic machine learning algorithm to obtain a first prediction model including: Determine a high-dimensional feature space based on the high-dimensional feature vector, calculate the Euclidean distance from the current data point to other data points for each data point in the high-dimensional feature space, select multiple data points closest to the current data point as the neighborhood of the current data point, construct an undirected weighted graph corresponding to the neighborhood, and determine the weight of the connecting edge between every two data points by using a Gaussian kernel function; For the undirected weighted graph, the Laplace matrix is ​​calculated in combination with the connection edge weights and the Laplace matrix is ​​eigen-decomposed to obtain eigenvalues ​​and eigenvectors, the eigenvectors are arranged according to the size of the eigenvalues, the eigenvectors corresponding to the first five smallest non-zero eigenvalues ​​are retained and constitute a low-dimensional embedded representation after dimensionality reduction, and the initial data set is generated; The initial data set is divided into a training data set and a test data set, a forward gated recurrent unit and a backward gated recurrent unit are stacked in sequence to form a deep bidirectional gated recurrent unit network, and a fully connected layer is used as an output layer to obtain an initial prediction model; The training data in the training data set is added as input data to the initial prediction model, the forward gated recurrent unit processes the training data according to the time sequence corresponding to the training data to obtain a forward hidden state, the backward gated recurrent unit processes the training data according to the reverse time sequence corresponding to the training data to obtain a backward hidden state, the forward hidden state and the backward hidden state are concatenated to obtain an output state, the output state is mapped to a predicted performance indicator through the fully connected layer, the predicted performance indicator is obtained and compared with the true label of the current training data, the prediction loss function is constructed in combination with the mean square error and the absolute percentage error, and the prediction loss value is calculated; Based on the predicted loss value, the value range of the hyperparameters in the initial prediction model is defined as a search space, wherein each hyperparameter corresponds to a dimension. The Gaussian process is used as a probability model to model the relationship between the objective function and the hyperparameters. The probability model is initialized and a hyperparameter combination is randomly selected for evaluation to obtain the objective function value. The objective function value is used as historical data to update the probability model. The update is repeated until a preset maximum number of iterations is reached to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is added to the initial prediction model to obtain the first prediction model.

5. The method according to claim 4, characterized in that The prediction loss function is constructed by combining the mean square error and the absolute percentage error. The prediction loss value is calculated as shown in the following formula: Among them, L(θ) represents the prediction loss value under the given parameter θ, α represents the mean square error weight coefficient, n represents the number of samples, and y i represents the true value of the i-th sample, represents the predicted value of the i-th sample, β represents the absolute percentage error weight coefficient, and λ represents the regularization parameter.

6. The method according to claim 1, characterized in that The real-time charge and discharge data of the battery to be evaluated is obtained and a real-time feature vector is constructed. The real-time feature vector is added to the first prediction model to generate real-time performance parameters. The real-time performance parameters are filtered by an online learning algorithm. A health status prediction model is constructed in combination with a Bayesian learning algorithm. The filtered real-time performance parameters are used as input for prediction. The comprehensive performance evaluation results obtained include: Acquire real-time charge and discharge data of the battery to be evaluated and preprocess the real-time charge and discharge data, perform time synchronization and alignment on data with different sampling frequencies, extract charge and discharge capacity, internal resistance and peak power from the aligned real-time charge and discharge data as key features, construct a key feature space, determine a key feature subset in combination with a principal component analysis algorithm, use a feature vector in the key feature subset as a real-time feature vector and add it to the first prediction model to obtain real-time performance parameters; The first prediction model sets the initial value and covariance matrix of the state variable according to the input real-time feature vector in combination with the Kalman filter algorithm, predicts the state variable and covariance matrix at the current moment according to the state estimation value and the state transition model at the previous moment, calculates the Kalman gain and updates the covariance matrix in combination with the observation value and the observation model at the current moment and the preset damping factor, and repeats the update until convergence to obtain the filtered real-time performance parameters; According to historical data and expert experience, the prior probability distribution of the health status is estimated in combination with the Gaussian distribution. According to the relationship between the filtered real-time performance parameters and the health status, a likelihood function is constructed. According to the Bayesian theorem, the posterior probability distribution of the health status is calculated. Based on the likelihood function and the prior probability distribution and the posterior probability distribution, a health status prediction model is constructed. The filtered real-time performance parameters are added to the health status prediction model to obtain a comprehensive performance evaluation result.

7. The method according to claim 6, characterized in that Combining the current observation value and observation model, combined with the preset damping factor, the Kalman gain is calculated as shown in the following formula: Among them, K t represents the Kalman gain matrix, η represents the damping factor, represents the prediction error covariance matrix, H t represents the observation matrix, T represents the transpose, R t represents the observation noise covariance matrix, () -1 Represents the inverse of a matrix.

8. A battery pack performance evaluation system based on big data, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the charging and discharging data of the battery to be evaluated under different working conditions, perform denoising on the charging and discharging data in combination with wavelet transform, perform adaptive normalization processing in combination with gradient descent algorithm to obtain standard charging and discharging data, perform equalization processing through synthetic minority class oversampling technology and respectively extract time domain features and frequency domain features, and combine them to obtain a high-dimensional feature vector; The second unit is used to perform a dimensionality reduction operation on the high-dimensional feature vector through a nonlinear dimensionality reduction algorithm based on manifold learning to obtain a low-dimensional embedded representation, generate an initial data set, divide the initial data set into a training data set and a test data set, build an initial prediction model based on a bidirectional gated recurrent unit network and use the data in the training data set as input, and optimize the hyperparameters of the initial prediction model in combination with an automatic machine learning algorithm to obtain a first prediction model; The third unit is used to obtain real-time charging and discharging data of the battery to be evaluated and construct a real-time feature vector, add the real-time feature vector to the first prediction model to generate real-time performance parameters, filter the real-time performance parameters through an online learning algorithm, and build a health status prediction model in combination with a Bayesian learning algorithm, and use the filtered real-time performance parameters as input for prediction to obtain a comprehensive performance evaluation result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images