Lithium battery health diagnosis method and system based on scalable recurrent neural network

Abstract

The present invention discloses a lithium battery health diagnosis method and system based on an extensible recurrent neural network, comprising the following steps: performing wavelet decomposition on the voltage segment data near the switching segment of the fast charging stage to obtain the aging characteristics of the battery to be selected; performing correlation analysis on the aging characteristics of the battery to be selected to obtain an aging characteristic sequence; combining a long short-term memory neural network with a Gaussian process regression layer to form an extensible deep recurrent neural network model; inputting the aging characteristic sequence into the trained extensible deep recurrent neural network model to obtain the health state estimation and remaining service life estimation results of the lithium battery. The present invention includes two parts: the construction of a neural network structure and the implementation of a data-driven method, which can effectively extract features from battery data under fast charging conditions, and obtain high-precision health state estimation and remaining service life estimation results.

Description

Technical Field

[0001] The present invention belongs to the field of battery technology, and in particular relates to a lithium battery health diagnosis method and system based on an expandable recurrent neural network. Background Art

[0002] The sustainable development of transportation electrification requires the integration of highly energy-efficient and zero-pollution electric vehicles. A major problem with electric vehicles is that their performance depends largely on the lifetime, safety, and cost of the energy storage system. Lithium-ion batteries have become the most popular energy storage devices due to their advantages such as high power density, long life, and low self-discharge rate. However, battery performance will degrade during repeated charge and discharge operations due to the loss of active lithium content caused by irreversible side reactions. The degradation of battery performance can be evaluated by state of health and remaining service life, which are the most critical indicators in battery system prognosis and health management. However, complex degradation mechanisms, unpredictable dynamic stress factors, and complex path dependencies of degradation phenomena pose significant challenges to accurate estimation and prediction. Therefore, the battery management system must accurately perform prognosis and health management functions and use measurable variables such as voltage, current, and temperature to avoid safety risks and extend battery life.

[0003] At present, the methods for achieving accurate health status estimation and remaining service life prediction can basically be divided into three categories, namely empirical methods, model-based methods and data-driven methods.

[0004] Empirical methods: These methods aim to establish an open-loop empirical model that describes the relationship between external stress factors (voltage, current, temperature, depth of discharge, etc.) and internal aging behavior (capacity reduction and resistance increase) and use complex experimental battery tests. These models can be further divided into two subcategories, namely chemical degradation models and mechanical degradation models. Chemical models describe battery aging behavior based on the electrochemical mechanism of side reactions during battery use. After establishing a well-calibrated model, the battery health state of the battery can be directly calculated. Obviously, these models can provide a physical explanation for battery degradation and are easy to implement. However, significant health state estimation errors are observed when the operating conditions are different from the test matrix.

[0005] Model-based approaches: Battery capacity and internal resistance are two important parameters in battery management. Therefore, model-based approaches first consider the battery state of health as an unknown parameter. Then, an adaptive online estimator is designed based on an equivalent circuit model or an electrochemical model. The advantages of these approaches lie in their closed-loop design and high robustness to unknown external disturbances. However, their performance is highly dependent on the accuracy, robustness, and reliability of the battery model. Equivalent circuit models and electrochemical models are two of the most popular battery models. Each model has its own advantages and challenges. For example, equivalent circuit models are simple and easy to implement, and this computationally efficient feature makes them suitable for online battery state of health estimation, but achieving high accuracy remains a challenge. Electrochemical models consist of a set of partial differential equations that can provide internal information about battery degradation, but suffer from high computational complexity. Despite progress in developing more accurate and faster models, identifiability issues, model mismatch, and unknown disturbances pose challenges to these model-based schemes.

[0006] Data-driven methods: To eliminate the dependency of the state of health estimation algorithm on the battery model, a large number of data-driven methods have been developed, taking advantage of the model independence and high flexibility of machine learning algorithms. These data-driven algorithms establish a nonlinear mapping model between battery operating data and health state. A three-stage machine learning for battery state of health estimation is: feature engineering for health indicator selection, offline training using experimental data, and online estimation using the trained model. Data-driven methods can be divided into several categories according to the type of health indicator or machine learning algorithm. Existing health factors for battery state of health estimation can be further divided into three subcategories: measurable variables, signal processing-based, and statistical indicators. The health factors based on measurable variables directly use the data collected from the battery management system sensors, directly taking the battery's historical voltage, voltage, and current as input. The health factors based on signal processing are extracted by applying signal processing algorithms on the battery operating data. The health factors based on statistical indicators are extracted from the statistical characteristics of the battery data. Despite the progress in developing robust and highly relevant health factors, the availability of health factors remains a major challenge for battery state of health estimation. Most existing studies focus on extracting health factors during constant current and constant voltage charging of batteries, as constant current and constant voltage charging protocols for batteries are widely used. However, with the rapid development of fast charging strategies, charging protocols are becoming more and more complex. In this case, the random and highly dynamic fast charging profiles may make the existing methods inoperable. Summary of the invention

[0007] In order to solve the problem of predicting the health status and remaining service life of lithium batteries with uncertain indicators, the present invention proposes a lithium battery health diagnosis method and system based on a scalable recurrent neural network, which is a scalable deep recurrent neural network and can adapt to a variety of fast charging conditions. It includes two parts: neural network structure construction and data-driven method implementation. It can effectively extract features of battery data under fast charging conditions and obtain high-precision health status estimation and remaining service life estimation results.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] The lithium battery health diagnosis method based on scalable recurrent neural network includes the following steps:

[0010] Perform wavelet decomposition on the voltage segment data near the switching section in the fast charging stage to obtain the battery aging characteristics to be selected;

[0011] Performing correlation analysis on the aging characteristics of the battery to be selected to obtain an aging characteristic sequence;

[0012] Combining long short-term memory neural networks with Gaussian process regression layers to form a scalable deep recurrent neural network model;

[0013] The aging feature sequence is input into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery.

[0014] Preferably, the wavelet decomposition is to perform multiple discrete wavelet transforms on a discrete sequence, each discrete wavelet transform decomposes the data segment into an approximate component and a detail component, and the obtained approximate component is subjected to discrete wavelet transform multiple times. The method of performing wavelet decomposition on the voltage segment data near the switching segment in the fast charging stage includes:

[0015]

[0016] Where x(t) is the discrete sequence to be decomposed, j and k are scaling parameters and translation parameters, ψ(·) is the wavelet function, the superscript * is the conjugate function symbol, and t is the time series number of the data.

[0017] Preferably, the method for performing correlation analysis on the battery aging characteristics to be selected includes:

[0018] For multiple battery aging features to be selected, the aging feature sequence is obtained by selecting them through the Pearson correlation coefficient;

[0019] Among them, the expression of Pearson correlation coefficient is:

[0020] In the formula, x is the health status feature sequence to be selected, y is the corresponding health status value in the data, the superscript dash represents the mean of the sequence, and x i is the i-th element of the health status sequence to be selected, y i is the i-th element of the corresponding health status value sequence.

[0021] Preferably, the kernel structure of the Gaussian process regression layer uses sparseness to achieve the scalability of the Gaussian process regression algorithm. The specific process is:

[0022] The sparse method used by the kernel structure of the Gaussian process regression layer is to introduce induction points to obtain the regression subset approximate kernel function:

[0023]

[0024] Where u j 、u a 、u b is the element in the introduced induction point, where u j for and wait for and x i The induction point for kernel function calculation, u a 、u b For x i The two nearest induction points have u a <x i b , W i For x i to u a 、u b Linear interpolation weights as a function of the relative distance;

[0025] The introduced induction points are evenly spaced, and the kernel function is further simplified by linear interpolation or cubic interpolation:

[0026] K U，U is the covariance matrix of the induced point sequence generated by the kernel function, where W is the network weight;

[0027] The scalability of Gaussian process regression is achieved by using induced points with equal spacing to make the kernel matrix satisfy the Toeplitz structure or Kronecker structure.

[0028] Preferably, the method for training a scalable deep recurrent neural network model includes: performing gradient descent using a semi-stochastic asynchronous gradient descent algorithm and using a negative log marginal likelihood function as a loss function.

[0029] Preferably, the method of performing gradient descent using a semi-stochastic asynchronous gradient descent algorithm includes:

[0030] ​S1: Collect labeled training data, i.e., battery aging features X and corresponding health status or remaining service life;

[0031] S2: Initialize Gaussian process layer parameters θ and neural network layer weight parameters W;

[0032] S3: Calculate the kernel matrix K based on the input battery aging feature X and Gaussian process layer parameters θ;

[0033] S4: Gradient descent optimization of the Gaussian process layer parameters θ using the full batch training data;

[0034] S5: Perform mini-batch stochastic gradient descent on the neural network layer weight parameter W;

[0035] S6: Determine whether all small batches have been optimized. If so, execute S7. Otherwise, return to S5.

[0036] S7: Determine whether the algorithm converges. If so, the training of the scalable deep recurrent network model is completed and this process ends. If not, return to S4.

[0037] Preferably, the method of inputting the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery includes:

[0038] The extracted aging feature sequence is input into the trained scalable deep recurrent neural network model and then passed through the long short-term memory recurrent neural network structure. The result is then used as the input of a sparse Gaussian process regression layer. The output of the sparse Gaussian process regression layer is the battery health status or remaining service life to be obtained.

[0039] The present invention also provides a lithium battery health diagnosis system based on an expandable recurrent neural network, comprising: a feature acquisition module to be selected, a feature sequence acquisition module, a model construction module and an estimation module;

[0040] The module for obtaining the features to be selected is used to perform wavelet decomposition on the voltage segment data near the switching segment in the fast charging phase to obtain the battery aging features to be selected;

[0041] The feature sequence acquisition module is used to perform correlation analysis on the aging features of the battery to be selected to obtain an aging feature sequence;

[0042] The model building module is used to combine the long short-term memory neural network and the Gaussian process regression layer to form a scalable deep recurrent neural network model;

[0043] The estimation module is used to input the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention can effectively extract features of battery data under fast charging conditions and obtain high-precision health status estimation and remaining service life estimation results.

[0046] The present invention can process long sequence data represented by voltage segments and obtain the uncertainty quantitative interval of the output result by combining long short-term memory neural network and sparse Gaussian process regression. Compared with traditional Gaussian process regression, the sparse Gaussian process regression using kernel structure can greatly reduce the limitation of input data length and significantly reduce the time of training and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0048] Figure 1 It is a schematic flow chart of a lithium battery health diagnosis method based on an expandable recurrent neural network according to an embodiment of the present invention;

[0049] Figure 2 It is a flowchart of a neural network semi-stochastic asynchronous gradient descent process for realizing battery health status diagnosis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Embodiment 1

[0053] like Figure 1 As shown, the present invention provides a lithium battery health diagnosis method based on an expandable recurrent neural network, comprising the following steps:

[0054] Perform wavelet decomposition on the voltage segment data near the switching section in the fast charging stage to obtain the battery aging characteristics to be selected;

[0055] Perform correlation analysis on the selected battery aging characteristics to obtain an aging characteristic sequence;

[0056] Combining long short-term memory neural networks with Gaussian process regression layers to form a scalable deep recurrent neural network model;

[0057] The aging feature sequence is input into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery.

[0058] In this embodiment, the battery health state is defined as the percentage of the current capacity to the nominal capacity, and the remaining service life is defined as the number of remaining cycles from the current cycle period to when the battery health state degrades to 80%.

[0059] For batteries that accept fast charging, there are multiple stages of constant voltage charging process. Using the voltage and current data of the battery fast charging process, the point where the current is continuously approximately 0 is taken as the voltage switching interval. The voltage data of the last voltage switching interval is selected, and the voltage segment of this charging cycle is decomposed by wavelet.

[0060] Wavelet decomposition is to perform multiple discrete wavelet transforms on a discrete sequence. Each discrete wavelet transform decomposes the data fragment into an approximate component and a detail component, and performs discrete wavelet transforms on the obtained approximate components multiple times. The method of performing wavelet decomposition on the voltage fragment data near the switching segment in the fast charging stage includes:

[0061]

[0062] Where x(t) is the discrete sequence to be decomposed, j and k are scaling parameters and translation parameters, ψ(·) is the wavelet function, the superscript * is the conjugate function symbol, and t is the time series number of the data.

[0063] The wavelet function used in the present invention is the Haar function:

[0064]

[0065] The length of the segment selected in this embodiment is 64, which can be decomposed into 7 segments of data by wavelet. The variance and average value of these 7 segments are calculated respectively to obtain 14 health status features to be selected.

[0066] In this embodiment, the method for performing correlation analysis on the selected battery aging characteristics includes:

[0067] For multiple battery aging features to be selected, the Pearson correlation coefficient is used to select and obtain an aging feature sequence. The larger the absolute value of the Pearson correlation is, the more likely it is to be used as the final selected aging feature.

[0068] Among them, the expression of Pearson correlation coefficient is:

[0069] In the formula, x is the health status feature sequence to be selected, y is the corresponding health status value in the data, the superscript dash represents the mean of the sequence, and x i is the i-th element of the health status sequence to be selected, y i is the i-th element of the corresponding health status value sequence.

[0070] Based on the battery fast charging data, the mean of the last layer of approximate components of the wavelet decomposition data is finally selected as the input feature.

[0071] In this embodiment, there are corresponding aging features for different aging cycles of the battery, and a sequence of aging features is used as the input of the long short-term memory neural network.

[0072] In the neural network construction part, a long short-term memory network structure is used to process the sequence aging characteristics, and the input results are input into the sparse Gaussian process regression layer.

[0073] As an extension of the recurrent neural network, the long short-term memory network introduces the gating structure of input gate, forget gate, and output gate and the memory unit structure on the basis of the recurrent neural network. It can capture the long-term dependencies between sequences and avoid the gradient vanishing problem caused by too long sequences.

[0074] The long short-term neural network can be described by the following formula:

[0075] i t =σ(W xi x t +W hi h t-1 +b i )

[0076]

[0077] f t =σ(W xf x t +W hf h t-1 +b f )

[0078]

[0079] o t =σ(W xo x t +W ho h t-1 +b o )

[0080]

[0081] y t =W y h t

[0082] in represents point-by-point multiplication between vectors, σ(·) and tanh(·) represent the sigmoid activation function and the hyperbolic tangent activation function respectively, i t 、f t , o t They correspond to the results of the input gate, forget gate, and output gate respectively. The variable range of these gate control structures is in [0, 1]. t , is the value of the memory cell and its intermediate state. ·，· , b. represents the weights and bias parameters of each layer of the neural network

[0083] The sparse Gaussian process regression is implemented by utilizing the kernel structure.

[0084] Gaussian Process Regression:

[0085] in

[0086]

[0087] K is the sum of the training input X and the prediction input X * Determine the kernel matrix, For the effect on X * The mean function on , y is the output of the training set, and I is the identity matrix represents a normal distribution, With Cov[f * ] distribution represents the predicted noise-free output f * The expectation and covariance matrices of .

[0088] For the input of Gaussian process regression, the loss function used is the negative log marginal likelihood function NLML:

[0089]

[0090] where σ is the noise parameter of Gaussian process regression, represents the transpose of the training output matrix, K y represents the covariance matrix of the training output y, is the negative log marginal likelihood loss function with respect to the covariance matrix K.

[0091] In order to achieve the gradient descent of the loss function, the partial derivative of the loss function with respect to the parameter to be optimized is calculated:

[0092] tr(·) is the operation of finding the matrix trace.

[0093] Unlike the usual Gaussian process regression and recurrent neural network, Gaussian process regression is connected in series with the neural network, so the partial derivatives of some weight parameters of the neural network are no longer obvious.

[0094]

[0095] That is, the partial derivative of K with respect to the network weight W in this formula is no longer obvious.

[0096] Apply the chain rule:

[0097] Where H is the output of the neural network layer.

[0098] By calculating the matrix derivative, we get:

[0099]

[0100] where h i 、h j where W represents two elements of the output sequence of the recurrent neural network layer, m is the network weight parameter of the mth layer.

[0101] Through the above formula, it is possible to calculate the gradient of a certain weight alone, that is, to achieve factor decomposition.

[0102] Specifically, the kernel structure of the Gaussian process regression layer uses sparseness to achieve the scalability of the Gaussian process regression algorithm as follows:

[0103] The sparse method used by the kernel structure of the Gaussian process regression layer is to use the method of introducing induced points to approximate the kernel function by regression subset:

[0104]

[0105] Where U represents the introduced virtual induction point set, the induction points can be selected arbitrarily, and here U is selected to be evenly spaced.

[0106] Due to the equidistant nature of the induced point set, the kernel function is further approximated:

[0107]

[0108] Where u j 、u a 、u b is the element in the introduced induction point, where u j for and wait for and x iThe induction point for kernel function calculation, u a 、u b For x i The two nearest induction points have u a <x i b , w i For x i to u a 、u b Linear interpolation weights as a function of the relative distance;

[0109] here is the constructed new approximate kernel function, converted into matrix form:

[0110] K X，U ≈WK U，U

[0111] Substitute the previous formula and simplify it to get:

[0112]

[0113] In the formula, K U，U is the covariance matrix of the induced point sequence generated by the kernel function, where W is the network weight;

[0114] If the original kernel function is a static kernel function, that is, the value of the kernel function is only related to the input distance (this satisfies most kernel functions), then for one-dimensional input, it satisfies the Toeplitz structure; for multi-bit input, it satisfies the Kronecker structure. Using these structures can reduce the computational complexity to Achieved scalability of Gaussian process regression.

[0115] In this embodiment, the method for training a scalable deep recurrent neural network model includes: performing gradient descent using a semi-stochastic asynchronous gradient descent algorithm and using a negative log marginal likelihood function as a loss function.

[0116] For the problem of estimating the battery health status, the input of the neural network is the aging feature sequence extracted from the voltage and current data, and the output is the estimated health status value. For the problem of the remaining battery life, the input of the neural network is a sequence of battery health status values, and the output is the predicted remaining service life. Specifically, Figure 2 As shown, the method of gradient descent using the semi-stochastic asynchronous gradient descent algorithm includes:

[0117] S1: Collect labeled training data, i.e., battery aging features X and corresponding health status y or remaining service life;

[0118] S2: Initialize Gaussian process layer parameters θ and neural network layer weight parameters W; ​

[0119] S3: Calculate the kernel matrix K based on the input battery aging feature X and Gaussian process layer parameters θ;

[0120] S4: Gradient descent optimization of the Gaussian process layer parameters θ using the full batch training data;

[0121] S5: Perform mini-batch stochastic gradient descent on the neural network layer weight parameter W;

[0122] S6: Determine whether all small batches have been optimized. If so, execute S7. Otherwise, return to S5.

[0123] S7: Determine whether the algorithm converges. If so, the training of the scalable deep recurrent network model is completed and this process ends. If not, return to S4.

[0124] In this embodiment, the method of inputting the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery includes:

[0125] It consists of two parts: neural network structure construction and data-driven method implementation.

[0126] Among them, the neural network structure is constructed by connecting the recurrent neural network structure of the long short-term memory network and the sparse Gaussian process regression layer in series. After the extracted features are input into the neural network structure, they first pass through the long short-term memory recurrent neural network structure, and then the obtained results are used as the input of a sparse Gaussian process regression layer, and the output of the sparse Gaussian process regression layer is the battery health status or remaining service life to be obtained.

[0127] The three stages of implementing the data-driven approach are aging feature extraction, offline training of the network structure, and online prediction using the trained network structure.

[0128] The long short-term neural network is a cyclic structure with memory units. Compared with general neural networks, it can better avoid the gradient vanishing problem caused by long sequence information and reduce the possibility of training failure.

[0129] Sparse Gaussian process regression uses a kernel structure approximation method to reduce the training cost of Gaussian process regression. The computational complexity has, to a certain extent, solved the scalability problem of Gaussian process regression.

[0130] The aging characteristics of the present invention are obtained by performing wavelet decomposition on the voltage segment data near the switching section in the fast charging stage.

[0131] The training method is to perform semi-stochastic asynchronous gradient descent on the LSTM network and the sparse Gaussian process layer. Semi-stochastic means that the LSTM network is trained with random mini-batch gradient descent, and the Gaussian process regression layer is trained with full batch gradient descent. Asynchronous means that the LSTM network is trained with gradient descent multiple times before the Gaussian process regression layer is trained with gradient descent.

[0132] The loss function corresponding to gradient descent is obtained by the difference between the actual health state or remaining service life and the health state or remaining service life output by the network.

[0133] Embodiment 2

[0134] The present invention also provides a lithium battery health diagnosis system based on an expandable recurrent neural network, comprising: a feature acquisition module to be selected, a feature sequence acquisition module, a model construction module and an estimation module;

[0135] The module for obtaining the features to be selected is used to perform wavelet decomposition on the voltage segment data near the switching section in the fast charging stage to obtain the battery aging features to be selected;

[0136] The feature sequence acquisition module is used to perform correlation analysis on the battery aging features to be selected to obtain an aging feature sequence;

[0137] The model building module is used to combine the long short-term memory neural network with the Gaussian process regression layer to form a scalable deep recurrent neural network model;

[0138] The estimation module is used to input the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery.

[0139] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A lithium battery health diagnosis method based on a scalable recurrent neural network, characterized in that: The following steps are involved: Perform wavelet decomposition on the voltage segment data near the switching section in the fast charging stage to obtain the battery aging characteristics to be selected; Performing correlation analysis on the aging characteristics of the battery to be selected to obtain an aging characteristic sequence; Combining long short-term memory neural networks with Gaussian process regression layers to form a scalable deep recurrent neural network model; Inputting the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery; Wavelet decomposition is to perform multiple discrete wavelet transforms on a discrete sequence. Each discrete wavelet transform decomposes the data fragment into an approximate component and a detail component, and performs discrete wavelet transforms on the obtained approximate components multiple times. The method of performing wavelet decomposition on the voltage fragment data near the switching segment in the fast charging stage includes: Where x(t) is the discrete sequence to be decomposed, j and k are scaling parameters and translation parameters, ψ(·) is the wavelet function, the superscript * is the conjugate function symbol, and t is the time series number of the data; The method for performing correlation analysis on the battery aging characteristics to be selected includes: For multiple battery aging features to be selected, the aging feature sequence is obtained by selecting them through the Pearson correlation coefficient; Among them, the expression of Pearson correlation coefficient is: In the formula, x is the health status feature sequence to be selected, y is the corresponding health status value in the data, the superscript dash represents the mean of the sequence, and x i is the i-th element of the health status sequence to be selected, y i is the i-th element of the corresponding health status value sequence; The kernel structure of the Gaussian process regression layer uses sparseness to achieve the scalability of the Gaussian process regression algorithm. The specific process is: The sparse method used by the kernel structure of the Gaussian process regression layer is to introduce induction points to obtain the regression subset approximate kernel function: Where u j 、u a 、u b is the element in the introduced induction point, where u j is with x i The induction point for kernel function calculation, u a 、u b For x i The two nearest induction points have u a <x i b , w i For x i to u a 、u b Linear interpolation weights as a function of relative distance;​ The introduced induction points are evenly spaced, and the kernel function is further simplified by linear interpolation or cubic interpolation: K U，U is the covariance matrix of the induced point sequence generated by the kernel function, where W is the network weight; The scalability of Gaussian process regression is achieved by using the equal spacing of induced points to make the kernel matrix satisfy the Toeplitz structure or Kronecker structure; The method for training a scalable deep recurrent neural network model includes: performing gradient descent using a semi-stochastic asynchronous gradient descent algorithm and using a negative log marginal likelihood function as a loss function; Methods for gradient descent using the semi-stochastic asynchronous gradient descent algorithm include: S1: Collect labeled training data, i.e., battery aging features X and corresponding health status or remaining service life; S2: Initialize Gaussian process layer parameters θ and neural network layer weight parameters W; S3: Calculate the kernel matrix K based on the input battery aging feature X and Gaussian process layer parameters θ; S4: Gradient descent optimization of the Gaussian process layer parameters θ using the full batch training data; S5: Perform mini-batch stochastic gradient descent on the neural network layer weight parameter W; S6: Determine whether all small batches have been optimized. If so, execute S7. Otherwise, return to S5. S7: Determine whether the algorithm converges. If so, the training of the scalable deep recurrent neural network model is completed and the process ends. If not, return to S4. For batteries that accept fast charging, their charging voltage has a multi-stage constant-voltage charging process. Using the voltage and current data of the battery fast charging process, the point where the current is continuously approximately 0 is used as the voltage switching interval; the voltage data of the last voltage switching interval is selected, and the voltage segment of this charging cycle is decomposed by wavelet.

2. The lithium battery health diagnosis method based on scalable recurrent neural network according to claim 1 is characterized in that: The method of inputting the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery includes: The extracted aging feature sequence is input into the trained scalable deep recurrent neural network model and then passed through the long short-term memory recurrent neural network structure. The result is then used as the input of a sparse Gaussian process regression layer. The output of the sparse Gaussian process regression layer is the battery health status or remaining service life to be obtained.

3. A lithium battery health diagnosis system based on a scalable recurrent neural network, characterized in that: The method for diagnosing lithium battery health based on an expandable recurrent neural network according to any one of claims 1 to 2 comprises: a module for obtaining features to be selected, a module for obtaining feature sequences, a model building module and an estimation module; The module for obtaining the features to be selected is used to perform wavelet decomposition on the voltage segment data near the switching segment in the fast charging phase to obtain the battery aging features to be selected; The feature sequence acquisition module is used to perform correlation analysis on the aging features of the battery to be selected to obtain an aging feature sequence; The model building module is used to combine the long short-term memory neural network and the Gaussian process regression layer to form a scalable deep recurrent neural network model; The estimation module is used to input the aging feature sequence into the trained scalable deep recurrent neural network model to obtain the health status estimation and remaining service life estimation results of the lithium battery.

Images