Lithium ion battery prediction and health management method based on deep neural network

Abstract

The invention discloses a lithium ion battery prediction and health management method based on a deep neural network. The method comprises the following steps: step 1, constructing a model based on the deep neural network; 2, learning a model based on a deep neural network: inputting a learning sample into the full-connection deep neural network, and outputting a predicted value of the lithium ion health state; calculating partial differential of the U (t, x) to each characteristic parameter; inputting at least one of partial differential, learning samples and U (t, x) corresponding to secondary characteristic parameters in the partial differential set into a first module to obtain output, inputting the output into a second module, and performing nonlinear mapping through a plurality of learnable parameters to obtain output characteristics; obtaining a learned model based on the loss of the parameter optimization module; and step 3, inputting the characteristic parameters of the to-be-tested battery into the learned model to obtain a health state prediction value of the to-be-tested battery. According to the method, the fuzzy system is linked with the deep neural network, and the accuracy of lithium ion health state estimation is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium-ion battery fault prediction and health management, and specifically relates to a lithium-ion battery prediction and health management method based on a deep neural network. Background Art

[0002] Lithium-ion batteries are lower carbon and more environmentally friendly than fossil energy sources, and have significant advantages over traditional nickel-cadmium batteries, such as high energy density, no heavy metal pollution, and long cycle life, so lithium-ion batteries have been widely used in many fields. Lithium-ion batteries mainly rely on the movement of lithium ions between the positive and negative electrodes to work. During the charging and discharging process, lithium ions are intercalated and deintercalated back and forth between the two electrodes. In addition to this main reaction, the charging and discharging process includes a series of complex electrochemical reactions, as well as a large number of side reactions, such as electrolyte decomposition and metallic lithium deposition. When lithium-ion batteries are being discharged, charged, or even stationary, side reactions that consume lithium ions or electrons may cause changes in battery capacity. This change is irreversible and may cause the storage performance to decline, that is, aging.

[0003] According to the classification of model mechanisms, the aging assessment methods of lithium-ion batteries mainly include two categories: model-based methods and data-driven methods. Model-based methods mainly include electrochemical models. The model uses a series of electrochemical equations to describe the movement of lithium ions and electrons inside the battery based on the principles of physical and chemical reactions, and contains multiple partial differential equations. Its advantages are high accuracy and the ability to explore the microscopic reactions inside the battery, but some parameters cannot be measured after the battery is packaged, and too many parameters need to be solved to solve partial differential equations, so parameter measurement is difficult and the amount of calculation is too large. The data-driven model does not consider the internal chemical reactions and parameters of lithium-ion batteries, but only fits the input-output mapping through the external data of charge and discharge, and mainly uses machine learning algorithms. With the rapid development of deep learning, such as convolutional deep learning and long short-term memory, it has been used for lithium-ion battery aging analysis and evaluation, but it requires a large number of data samples, cannot couple the electrochemical reactions inside lithium-ion batteries, and does not analyze the internal reaction mechanism of the battery, so model learning mainly relies on external data of charge and discharge. In real applications, the charging time is shortened, and high-rate current fast charging is often used. In practical applications, fast charging methods are flexible and varied. If a health state estimation model is established and trained for each fast charging method, a large number of data samples are required and it is time-consuming and laborious. A more efficient way is to establish and train a model and method that can be used for a variety of different fast charging methods, which requires the model to have good generalization ability. If the electrochemical model and deep learning can be integrated into one, the advantages can complement each other, which can improve the accuracy of model prediction and the generalization ability of the model. The existing physical information neural network incorporates the physical laws described by differential equations into its loss function, and through the network architecture, the learning process obtains solutions that are more in line with the relevant physical laws. These physical laws can cover the electrochemical reactions of lithium-ion batteries, so the physical information neural network is considered to be a fusion of electrochemical models and deep learning. The existing physical information neural network is used to estimate the health state of lithium-ion batteries. Usually, a deep neural network (DNN) is used to represent the potential solution, and another DNN is used to represent nonlinear dynamics. However, the lack of transparency and interpretability of DNN leads to uncertainty in existing physical information neural network models and methods, which has a negative impact on prediction accuracy and insufficient generalization ability. Summary of the invention

[0004] The purpose of the present invention is to provide a lithium-ion battery prediction and health management method based on deep neural network, which adopts fuzzy system and deep neural network to link and integrate into the health status of lithium-ion battery to fit the nonlinearity of lithium-ion battery system more carefully and efficiently, and improve the accuracy of lithium-ion health status estimation.

[0005] The present invention adopts the following technical solution: a lithium-ion battery prediction and health management method based on a deep neural network, comprising the following steps:

[0006] Step 1: Build a model based on deep neural network:

[0007] The model based on the deep neural network includes a fully connected deep neural network and a global normalized deep neural fuzzy system, and the global normalized deep neural fuzzy system includes a first module and a second module connected to each other;

[0008] The output of the fully connected deep neural network is connected to the first module and the parameter optimization module respectively; the output of the second module is connected to the parameter optimization module;

[0009] Step 2: Learn the deep neural network based model in step 1:

[0010] Extract the characteristic parameter set of the battery charging and discharging process as a learning sample;

[0011] The learning samples are input into a fully connected deep neural network. The output of the fully connected deep neural network is the predicted value of the health status of lithium ions, which is represented by U(t,x), where t is the main feature parameter and x is the secondary feature parameter.

[0012] Calculate the partial differential of U(t,x) with respect to each characteristic parameter and obtain the partial differential set;

[0013] Taking at least one of the partial differential corresponding to the secondary feature parameter in the partial differential set, the learning sample and U(t,x) as a feature set, inputting the feature set into the first module to obtain an expression of the feature set, inputting the expression of the feature set into the second module, performing nonlinear mapping through multiple learnable parameters, and obtaining an output feature of the global normalized deep neural fuzzy system;

[0014] Based on the loss between the predicted value and the measured value of the lithium ion health status obtained by the parameter optimization module, the loss of the expression characteristics of the feature set, and the loss of the partial differential of the main feature parameters, a deep neural network-based model after learning is obtained;

[0015] Step 3: Input the characteristic parameters of the battery to be tested into the deep neural network-based model learned in step 2 to obtain a predicted value of the health status of the battery to be tested.

[0016] Furthermore, the second module is a deep neural network, which includes N G Small modules BG are connected in sequence and stacked in a chain. A fully connected layer is connected after the last small module BG. The deep neural network is as follows:

[0017]

[0018] Among them, BG 1 , BG 2 ,…,BG NGrepresents the input module from the input end to the output module at the output end; f 1 (·) represents a fully connected layer, f 2 (·) represents the activation function; is the output of the first module.

[0019] Further, when t selects the internal impedance of the battery, x includes: the average temperature of the battery, the charging time, the variance of the incremental capacity analysis during the charging or discharging process, the skewness of the incremental capacity analysis during the charging or discharging process, and the minimum value of the incremental capacity analysis during the charging or discharging process;

[0020] When t is selected as the minimum value of the incremental capacity analysis during the battery charging or discharging process, x includes: battery internal impedance, average temperature, charging time, variance of the incremental capacity analysis during the charging or discharging process, and skewness of the incremental capacity analysis during the charging or discharging process.

[0021] Furthermore, the parameter optimization module is a loss function L, as follows:

[0022]

[0023] Where: L U =|U(t,x)-y| 2 represents the error between the predicted value U(t,x) of the lithium ion health state by the fully connected deep neural network and the measured value y of the lithium ion health state; L F =|F| 2 Represents the square error of function F, where function F is defined as F:=U t -G;L Ft =|F t | 2 Indicates F t The square error of t It means the partial derivative of F with respect to parameter t.

[0024] Further, the first module includes a batch normalization layer, a first layer, a second layer, a third layer and a fourth layer connected in sequence; wherein:

[0025] In the first layer, each neuron represents a membership function, which indicates the input feature z d Satisfy fuzzy set Z R,D r=1,...,R,d=1,...,D, and output; using Gaussian membership function, as follows:

[0026]

[0027] Where: m r,d and σ r,dis the Gaussian mean parameter and Gaussian standard deviation parameter, indicating the mean and standard deviation of the Gaussian membership function, m r,d The value range is real number, σ r,d The value range is positive real number; input feature z d The partial differentials {t,x,U,U) corresponding to the secondary features in the learning sample input features, the fully connected deep neural network, and the partial differential set are selected. x} after normalization;

[0028] In the second layer, each neuron multiplies the input and outputs the product as the activation strength of the corresponding fuzzy rule, as follows:

[0029]

[0030] In the third layer, the normalized activation strength calculated for each neuron is:

[0031]

[0032] Where: R is the number of fuzzy rules.

[0033] In the fourth layer, each neuron corresponds to a fuzzy rule: where g r (z) = [b r,0 ,b r,1 …,b r,D ]×[1,z 1 ,…,z D ] T ,r=1,...,R,d=0,1,...,Dwhere b r,d Represents a fuzzy parameter, whose value range is a real number.

[0034] Furthermore, the fully connected deep neural network includes N U Small modules BL connected in sequence, and N U The small modules BL are stacked sequentially in a chain. At the back end of the chain, a fully connected layer is connected after the last small module BL, as follows:

[0035] U(t,x)=f 1 (BL NU (BL NU-1 (...BL 2 (BL 1 (t,x))))) (6);

[0036] Among them, BL 1 BL 2 ,…,BL NU represents the input module from the input end to the output module at the output end; f 1is a fully connected layer.

[0037] The present invention also discloses a model of a lithium-ion battery prediction and health management method based on a deep neural network of Shanshu, including a fully connected deep neural network and a global normalized deep neural fuzzy system, wherein the global normalized deep neural fuzzy system includes a first module and a second module connected to each other;

[0038] The output of the fully connected deep neural network is connected to the first module and the parameter optimization module respectively; the output of the second module is connected to the parameter optimization module.

[0039] The beneficial effects of the present invention are as follows: 1. A fully connected deep neural network is used to represent the nonlinear mapping between the input and output of the health state characteristic of the lithium-ion battery, and its partial differential and input characteristics are calculated as the input of the global normalized deep neural fuzzy system, and the output of the fully connected deep neural network and the output of the global normalized deep neural fuzzy system together constitute a physical information neural network architecture, so that the electrochemical reaction equation described by the differential equation is incorporated into this loss function to make the network architecture approximate the electrochemical model, and then the training data is used to optimize the model so that the learning process obtains a solution that is more in line with the electrochemical reaction of the lithium-ion battery, which has the advantages of a data-driven model, and the electrochemical model and the data-driven model are integrated into one, and the advantages complement each other, which can improve the accuracy of the model prediction and the generalization ability of the model. 2. The global normalized deep neural fuzzy system uses a fuzzy system to link with a deep neural network. Both the fuzzy system and the deep neural network have universal approximation theorems and can fit complex nonlinear systems. The two universal approximators are integrated into one to fit complex nonlinear lithium-ion battery systems. The integration into the health state of the lithium-ion battery can fit the nonlinearity of the lithium-ion battery system more carefully and efficiently, and improve the accuracy of the lithium-ion health state estimation. 3. The first module uses fuzzy rules and fuzzy set membership to improve the ability to handle complex and uncertain systems such as lithium-ion batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a diagram of the lithium-ion battery prediction and health management model based on deep neural network;

[0041] Figure 2 Charging conditions and battery life graphs for the fast charging dataset;

[0042] Figure 3 Schematic diagram of the extracted input features. DETAILED DESCRIPTION

[0043] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] The present invention discloses a lithium-ion battery prediction and health management method based on a deep neural network, such as Figure 1As shown in Figure 2, the main body of the model architecture is a physical information neural network, which includes two connected modules: a fully connected deep neural network U(t,x) and a global normalized deep neural fuzzy system G(t,x,U,U x ,); the outputs of the fully connected deep neural network U(t,x) are connected to the global normalized deep neural fuzzy system G(t,x,U,U x ) input and the input of the loss function L, the global normalized deep neural fuzzy system G(t,x,U,U x )’s output is connected to the input of the loss function L;

[0045] Where: t is a one-dimensional real number; x = [x 1 ,...,x n ] T is an n-dimensional real number, t and x are both input features, their values ​​are real numbers, and n is a positive integer.

[0046] Physical information neural network includes partial differentials, and the subscripts are used to indicate partial differentials or partial derivatives of t or x, U t and U x represents the first-order partial differential, such as Figure 1 As shown, the partial differential solution uses automatic differentiation.

[0047] The above-mentioned lithium-ion battery prediction and health management method based on deep neural network includes the following steps:

[0048] Step 1: Build a model based on deep neural network:

[0049] The deep neural network-based model includes a fully connected deep neural network and a global normalized deep neural fuzzy system, and the global normalized deep neural fuzzy system includes a first module and a second module connected to each other;

[0050] The output of the fully connected deep neural network is connected to the first module and the parameter optimization module respectively; the output of the second module is connected to the parameter optimization module.

[0051] The fully connected deep neural network includes N U Small modules BL connected in sequence, and N U The small modules BL are stacked sequentially in a chain. At the back end of the chain, a fully connected layer is connected after the last small module BL, as follows:

[0052] U(t,x)=f 1 (BL NU (BL NU-1 (...BL 2 (BL 1 (t,x))))) (6);

[0053] Among them, BL1 BL 2 ,…,BL NU represents the input module from the input end to the output module at the output end; f 1 is a fully connected layer.

[0054] Each fully connected layer has parameter weight coefficients and biases, which can be learned through the back propagation algorithm of the entire network. The number of parameters is related to the number of neurons. 2 Including but not limited to activation functions such as Sigmoid function, Tanh function, ReLU function, Softmax function. Each neuron in the batch normalization layer can be represented as:

[0055]

[0056] Where: x bn is the input of the batch normalization layer, β and γ are the offset and scale parameters, respectively, which can be learned through the back-propagation algorithm, μ B and are the mean and variance of the mini-batch dataset, respectively, and ε is a very small value added to prevent the denominator from being zero.

[0057] The learnable parameter set of this fully connected deep neural network includes all fully connected layers f 1 The parameter weight coefficient and bias of the normalization layer and the parameters β and γ constitute the parameter set Ω U , the value range is real numbers.

[0058] The above-mentioned lithium-ion battery prediction and health management method based on deep neural network can flexibly adjust the input, which can be {t, x, U, U x} any combination of these four variables, t is the main characteristic of the battery, and x is the secondary characteristic of the battery. In the embodiment of the present invention, these four variables [t, x, U, U x ] T All are input into the global normalized deep neural fuzzy system. The global normalized deep neural fuzzy system is first normalized, z = BN ([t, x, U, U x ] T ), where z is a D-dimensional column vector and D is a positive integer; z can be used d Represents the d-th dimension element of z, and then enters the first module and the second module of the global normalized deep neural fuzzy system in turn.

[0059] The first module includes a batch normalization layer, a first layer, a second layer, a third layer and a fourth layer connected in sequence; wherein:

[0060] In the first layer, each neuron represents a membership function, which indicates the input feature z d Satisfy fuzzy set ZR,D r=1,...,R,d=1,...,D, and output; using Gaussian membership function, as follows:

[0061]

[0062] Where: m r,d and σ r,d is the Gaussian mean parameter and Gaussian standard deviation parameter, indicating the mean and standard deviation of the Gaussian membership function, m r,d The value range is real number, σ r,d The value range is positive real number; input feature z d The partial differentials {t,x,U,U) corresponding to the secondary features in the learning sample input features, the fully connected deep neural network and the partial differential set are selected. x}one of the;

[0063] In the second layer, each neuron multiplies the input and outputs the product as the activation strength of the corresponding fuzzy rule, as follows:

[0064]

[0065] In the third layer, the normalized activation strength calculated for each neuron is:

[0066]

[0067] Where: R is the number of fuzzy rules.

[0068] In the fourth layer, each neuron corresponds to a fuzzy rule: where g r (z) = [b r,0 ,b r,1 …,b r,D ]×[1,z 1 ,…,z D ] T ,r=1,...,R,d=0,1,...,Dwhere b r,d Represents a fuzzy parameter, whose value range is a real number.

[0069] The second module is a deep neural network, and the deep neural network includes N G Small modules BG are connected in sequence and stacked in a chain. A fully connected layer is connected after the last small module BG. The deep neural network is as follows:

[0070]

[0071] Among them, BG 1 , BG 2 ,…,BG NGrepresents the input module from the input end to the output module at the output end; f 1 (·) represents a fully connected layer, f 2 (·) represents the activation function; is the output of the first module.

[0072] The fourth layer output of the global normalized deep neural fuzzy system is R neurons, each neuron is The output of the fourth layer is the input of the fifth layer. After entering the fifth layer of the deep neural network, it passes through the first small module BG 1 After obtaining Each fully connected layer f 1 (·) includes learnable parameter weight coefficients and biases, and each normalization layer includes learnable parameters β and γ. And this N G The small modules BG are stacked sequentially in a chain and finally connected to a fully connected layer to obtain G(z) as shown in formula (4).

[0073] The fifth layer of the traditional fuzzy system will r=1,...,R, add and sum, excluding learnable parameters; compared with the traditional fuzzy system, the fifth layer of the deep neural network includes multiple learnable parameters, such as the weight coefficient and bias of each fully connected layer and the parameters β and γ of each normalized layer. These learnable parameters are represented by the parameter set Ω G It means that it can improve the learning ability of the fifth level; f 2 (·) indicates that the activation function is a nonlinear mapping, which can enhance the nonlinear fitting ability of the global normalized deep neural fuzzy system after learning from the data.

[0074] Because both fuzzy systems and deep neural networks have universal approximation theorems and can fit complex nonlinear systems, the global normalized deep neural fuzzy system and deep neural network, two universal approximators, are integrated into one to fit G in the lithium-ion battery prediction and health management model based on deep neural networks. This can fit the nonlinearity of the lithium-ion battery system more carefully and efficiently, and improve the accuracy of lithium-ion health status estimation. As shown in formula (4), the global normalized deep neural fuzzy system uses a fuzzy system to link with a deep neural network. In addition, the global normalized deep neural fuzzy system can use fuzzy rules and fuzzy set membership to improve the ability to handle complex and uncertain systems such as lithium-ion batteries.

[0075] The parameters of the lithium-ion battery prediction and health management model based on deep neural network include the following:

[0076] Parameter set Ω in a fully connected deep neural network U ;

[0077] Globally normalized deep neural fuzzy system, including Gaussian mean parameter m r,d , Gaussian standard deviation parameter σ r,d and fuzzy parameter b r,d , and the parameter set Ω in the deep neural network G .

[0078] The parameter optimization module is the loss function L, as follows:

[0079]

[0080] Where: L U =|U(t,x)-y| 2 represents the error between the predicted value U(t,x) of the lithium ion health state by the fully connected deep neural network and the measured value y of the lithium ion health state; L F =|F| 2 Represents the square error of function F, where function F is defined as F:U t -G;L Ft =|F t | 2 Indicates F t The square error of t It means the partial derivative of F with respect to parameter t.

[0081] Step 2: Learn the deep neural network based model in step 1:

[0082] Extract the characteristic parameter set of the battery charging and discharging process as a learning sample;

[0083] Inputting the learning sample into the fully connected deep neural network, the output of the fully connected deep neural network is the predicted value of the health state of the lithium ion, represented by U(t,x), where t is the main characteristic parameter and x is the secondary characteristic parameter;

[0084] Calculate the partial differential of U(t,x) with respect to each characteristic parameter to obtain a partial differential set;

[0085] Taking at least one of the partial differentials corresponding to the secondary feature parameters in the partial differential set, the learning samples and U(t,x) as a feature set, inputting the feature set into the first module to obtain an expression of the feature set, inputting the expression of the feature set into the second module to perform nonlinear mapping through multiple learnable parameters, and obtaining output features of the global normalized deep neural fuzzy system;

[0086] Based on the loss between the predicted value and the measured value of the lithium ion health state obtained by the parameter optimization module, the loss of the expression characteristics of the feature set, and the loss of the partial differential of the main feature parameters, a deep neural network-based model after learning is obtained;

[0087] Calculate the partial differential of the fully connected deep neural network for each feature in the feature set, and obtain the partial differential set including: U t is the first-order partial differential of the fully connected deep neural network with respect to parameter t, U x is the first-order partial differential of the fully connected deep neural network with respect to parameter x; U xx is the second-order partial differential of the fully connected deep neural network with respect to the parameter x.

[0088] Step 3: Input the characteristic parameters of the battery to be tested into the deep neural network-based model learned in step 2 to obtain a health status prediction value of the battery to be tested.

[0089] When t is selected as the internal impedance of the battery, x includes: the average temperature of the battery, the charging time, the variance of the incremental capacity analysis during the charging or discharging process, the skewness of the incremental capacity analysis during the charging or discharging process, and the minimum value of the incremental capacity analysis during the charging or discharging process;

[0090] When t selects the minimum value of the incremental capacity analysis during the battery charging or discharging process, the x includes: battery internal impedance, average temperature, charging time, variance of the incremental capacity analysis during the charging or discharging process, and skewness of the incremental capacity analysis during the charging or discharging process.

[0091] The method of the present invention mainly includes two links: learning and prediction. After the model is built, it is necessary to learn and optimize the parameters, and then save the learned parameters, with the goal of accurate prediction. The learning step mainly includes two parts: forward propagation to calculate the loss function and back propagation to optimize the parameters. These two parts are iteratively cycled, and finally the optimized parameters meet the conditions, and the optimized parameters and model are saved. The test sample is input into the optimized model, and the forward propagation is calculated to predict the health status of lithium ions.

[0092] In the first step of forward propagation, the input (x, t) is first calculated using equation (4) to obtain the fully connected deep neural network U, and then the partial differential is solved by automatic differentiation to obtain Ux.

[0093] The second step of forward propagation is [t, x, U, U x ] T The input vector that constitutes the global normalized deep neural fuzzy system G is calculated by equations (1)-(4).

[0094] In the third step of forward propagation, the output of the globally normalized deep neural-fuzzy system G is used to calculate the loss function L of the fully connected deep neural network U as shown in formula (5).

[0095] The present invention uses the back propagation method to learn and optimize these adjustable parameters from the data: the parameter set Ω in the fully connected deep neural network U; In the global normalized deep neural fuzzy system, including the Gaussian mean parameter m r,d , Gaussian standard deviation parameter σ r,d and fuzzy parameter b r,d , and the parameter set Ω in the deep neural network G The back propagation method is a neural network optimization method, which is currently widely used in neural network learning. As follows:

[0096] First, calculate the forward propagation from the input to the loss function L, as described above, forward propagation from the first step to the third step; the loss function of a sample is used to calculate the error as shown in formula (5); if it is a batch or small batch of sample learning, the error is the sum of the loss function formula (5) of these samples, that is, the loss function sum; use the optimization algorithm extended by gradient descent to adjust the parameters, the calculation process is to advance from the loss function L to the input, that is, the error is back propagated, and the goal is to make the error smaller.

[0097] Specifically as follows: Optimize the lithium-ion battery prediction and health management model based on deep neural network in step 1:

[0098] First, extract the characteristics of the battery charging and discharging process as training data;

[0099] Input the training data into the lithium-ion battery prediction and health management model based on the deep neural network, and calculate the loss function L according to the first step to the third step of the forward propagation;

[0100] The error between the predicted value U(t,x) of the lithium ion health state of the fully connected deep neural network obtained based on the loss function L and the measured value y of the lithium ion health state, the square error of the function F and F t The square error is calculated; and the optimized lithium-ion battery prediction and health management model based on deep neural network is obtained through back propagation algorithm;

[0101] Iterate the above steps until the stopping criterion is met, such as the change in error between two consecutive iterations is very small or the number of iterations is sufficient.

[0102] In the optimized deep neural network-based lithium-ion battery prediction and health management model, the fully connected deep neural network U is used to predict the health status of lithium ions. Specifically, the characteristics of the test battery charge and discharge process are extracted as test data, and the test data is input into the fully connected deep neural network U(t,x) as shown in formula (6).

[0103] The fully connected deep neural network U is obtained by automatic differentiation to solve the partial differential x , and then together with the input (t, x) constitutes the input vector [t, x, U, U] of the global normalized deep neural fuzzy system G. x ] T; The output of the global normalized deep neural network fuzzy system G is linked to the fully connected deep neural network U through the loss function L as shown in formula (5), so that the input and output of the fully connected deep neural network and the global normalized deep neural network fuzzy system are seamlessly linked. The physical information neural network architecture is similar to the electrochemical model, so that the model has information about the electrochemical reaction of the lithium-ion battery, and then learns from the data to optimize the model parameters, integrating the electrochemical model and the data-driven model into one, complementing each other.

[0104] To verify the method of the present invention, the following experiment was conducted using a fast-charged lithium-ion battery data set, which includes 124 lithium-ion phosphate / graphite batteries, each with a nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V. The battery was charged under different fast-charging conditions and then discharged under the same conditions, repeatedly cycling until the battery capacity was less than 80% of the nominal capacity. Figure 2 As shown in the figure, there are three main conditions for fast charging: the initial charging current, the subsequent charging current, and the charging state when the current is converted. Compared with the traditional constant current and constant voltage charging, the fast charging conditions are complex and flexible, making the estimation of the health state of lithium ions more difficult. External data such as TU current and temperature are collected during the charging and discharging process. In this experiment, more than 90,000 effective cycles were collected, and about 90,000 samples were extracted from them.

[0105] Features are extracted from the above-mentioned effective cycles, and the extracted features have information related to the health status of lithium ions. First, internal impedance, charging time and average charging temperature are extracted as features. Then, considering that incremental capacity analysis is a widely used electrochemical method, the rate of change of capacity with voltage incremental capacity analysis is used to identify and quantify the electrochemical characteristics of the battery by using the voltage change during the charging or discharging stage. Therefore, the present invention also extracts features: the minimum value, variance and skewness of the incremental capacity analysis during the discharge process. These six features: internal impedance, charging time and average charging temperature, as well as the minimum value, variance and skewness of the incremental capacity analysis during the discharge process, are first preprocessed and scaled to the range of 0 to 1 using a linear scale, such as Figure 3 shown.

[0106] All lithium-ion batteries are cycled under multiple fast charging conditions. The charging conditions affect the health status of lithium ions. Therefore, the extracted features also include: number of cycles, initial charging current, subsequent charging current, and charging status during current conversion. These four features must also be preprocessed and scaled to the range of 0 to 1 using a commonly used linear scale. In addition, zero-mean normalization is used to convert the 10 features extracted from each effective cycle of each battery into a range of 0 to 1 with zero mean and unit variance. The 10 features extracted and preprocessed above are the input variables (t, x) of the present invention.

[0107] There are many definitions of lithium ion health status. The present invention adopts the commonly used definition as follows: the capacity C of the ith cycle i Ratio C to the nominal capacity nom ,Right now:

[0108]

[0109] Where: y represents the measured value of lithium ion health status, such as Figure 1 shown.

[0110] Deep neural networks rely on big data, so the present invention tries to increase the training and validation data sets as much as possible. The training and validation data sets include 115 batteries from No. 1 to No. 115, and the training and validation include more than 80,000 samples. In the training and validation data sets, 90% of the samples are randomly selected as training data sets for learning in each experiment, and the remaining samples are used for validation. The training validation set and the test data set are mutually exclusive, so 6 test batteries are selected from batteries No. 116 to No. 124, namely, No. 116, No. 117, No. 118, No. 119, No. 121 and No. 124 batteries constitute the test data set. Four groups of test experiments were performed on the six batteries respectively. In order to reduce randomness, each experiment was repeated 3 times and the average was taken. The experiments were all conducted on workstations equipped with NVIDIA GPU: GeForce RTX3080 and Intel CPU: i9.

[0111] The present invention adopts the back propagation method to learn and optimize parameters from the training data set, wherein the optimization settings are mainly as follows:

[0112] The optimization algorithm extended by gradient descent uses an adaptive moment estimation method to calculate the adaptive learning rate of each parameter, which not only includes the exponential decay average of the previous squared gradient, but also maintains the exponential decay average of the previous gradient. The initial learning rate is set to 0.01, and the learning rate is reduced by 0.1 times every 100 rounds of learning. Using small batch learning, the batch size is set to 4096; the parameter initialization uses the Xavier Normal algorithm, and the corresponding activation function uses Tanh.

[0113] In the present invention, the input feature t is in the function F:=U t -G and loss function L are more important than x as shown in formula (5), so t is the main feature and x is the secondary feature. Therefore, the selection of feature t has an impact on the accuracy of the model and prediction results. First, one feature is selected as t from the 10 extracted features, and the other 9 features are x. It is known that incremental capacity analysis is an electrochemical method for identifying the electrochemical properties of batteries, and according to the electrochemical model, internal impedance is related to the health status of lithium ions, so the experiment uses two schemes to select feature t: the minimum value of incremental capacity analysis or internal impedance. The specific features set by the two schemes are as follows:

[0114] When t is set as the internal impedance of the battery, x includes: charging time, average temperature, minimum value, variance and skewness of incremental capacity analysis during discharge, as well as number of cycles, initial charging current, subsequent charging current and charging state during conversion current;

[0115] When t is set to the minimum value of the incremental capacity analysis during battery discharge, x includes: battery internal impedance, average temperature, charging time, variance and skewness of the incremental capacity analysis during discharge, as well as number of cycles, initial charging current, subsequent charging current and charging state during conversion current.

[0116] Four sets of experiments were performed on the six test batteries in the two t selection schemes. Each set of experiments used two choices of t for learning and testing, and each selection scheme used two sets of hyperparameter settings, so a total of four experiments were performed in each group. The experimental labels are LiB1, LiB2, LiB3, and LiB4, as shown in Table 1. The detailed information of the hyperparameters of each set of experiments. Figure 1 As shown, N U N represents the number of small modules BL connected sequentially in the fully connected deep neural network in U; G Represents the number of sequentially connected small modules BG in the global normalized deep neural fuzzy system.

[0117] Table 1 Detailed information of experimental hyperparameters

[0118]

[0119] To test the prediction effect of the method of the present invention, it is compared with three configurations in the existing physical information neural network method: PINN-DeepHPM (AdpBal), PINN-Verhulst (Sum) and PINN-Verhulst (AdpBal). The examples A of these three existing methods are learned from the data of batteries No. 91 and No. 100, and then tested on battery No. 124. The present invention is not only tested on battery No. 124, but also on another 5 test batteries: No. 116, No. 117, No. 118, No. 119 and No. 121 to verify the prediction effect.

[0120] The above variables (x, t) and y represent the input characteristics of a sample and the corresponding measured value of the lithium ion health status. However, the present invention uses a large number of samples, so the variables are added with a subscript to indicate the sample number. For example, y m represents the measured value of the mth lithium ion health status, and the corresponding mth input feature is (x m ,t m ), the predicted value of the lithium ion health status of the present invention is U(x m ,t m ) is simplified to U mThe present invention uses two commonly used error measurement experimental results: mean absolute error and root mean square error. The mean absolute error MAE formula is:

[0121]

[0122] The root mean square error RMSE formula is:

[0123]

[0124] Where: M represents the total number of samples, and its value range is a positive integer. The smaller the error value, the better the prediction accuracy.

[0125] Table 2 Average absolute error of six test batteries

[0126]

[0127] Table 3 Root mean square error of six test batteries

[0128]

[0129] Tables 2 and 3 list the mean absolute error and root mean square error of six test batteries under four experimental schemes and configurations, as well as three existing physical information neural network methods, where the black bold font indicates the minimum error obtained in the test on each battery. The lithium-ion battery prediction and health management method based on deep neural network proposed in the present invention predicts the health status of lithium ions on six test batteries after learning four optimization models from the training data under four different experimental schemes and configurations. On battery No. 116, LiB2 proposed in the present invention achieves the best prediction performance with a mean absolute error of 0.00335 and a root mean square error of 0.00442. On battery No. 117, LiB1 proposed in the present invention achieves a mean absolute error of 0.00511 and a root mean square error of 0.00593, both of which are better than other methods. On battery No. 118, LiB1 proposed in the present invention achieves the best prediction performance with a mean absolute error of 0.00406 and a root mean square error of 0.00535. On battery No. 119, LiB1 proposed by the present invention achieves the best prediction performance with a mean absolute error of 0.00504 and a root mean square error of 0.00645. On battery No. 121, LiB4 proposed by the present invention achieves the best prediction performance with a mean absolute error of 0.00510 and a root mean square error of 0.00557. On battery No. 124, LiB1 proposed by the present invention achieves a mean absolute error of 0.00324, which is better than other methods. The existing physical information neural network method PINN-DeepHPM (AdpBal) achieves a root mean square error of 0.00448, which is better than other methods.

[0130] It can be seen that on the five test batteries No. 116, No. 117, No. 118, No. 119 and No. 121, the optimization model and method with the smallest prediction error are all the lithium-ion battery prediction and health management method based on deep neural network proposed in the present invention, and the four scheme configurations of the lithium-ion battery prediction and health management method based on deep neural network proposed in the present invention obtain average absolute error and root mean square error that are better than the three existing physical information neural network methods, so the accuracy of lithium ion health status predicted by the model and method proposed in the present invention is improved. Only in the test of battery No. 124, the average absolute error of LiB1 and LiB2 predicted by the present invention is better than the three existing physical information neural network methods; however, the root mean square error of the predictions of the two existing physical information neural network methods PINN-DeepHPM (AdpBal) and PINN-Verhulst (AdpBal) is smaller than that of the method of the present invention. Because the three existing physical information neural network methods are learned from the data of batteries No. 91 and No. 100, and because the three batteries No. 91, No. 100 and No. 124 use the same fast charging conditions, good prediction results are achieved on the No. 124 test battery; however, the other five test batteries use different fast charging conditions, so the test results of the three existing physical information neural network methods have degraded performance.

[0131] Fast charging conditions are flexible and changeable. If a health state estimation model is established and trained for each fast charging, a large number of data samples are required and it is time-consuming and laborious. A more efficient way is to train a model that can be used for a variety of different fast charging conditions, which requires the model to have good generalization ability. In order to measure the generalization ability of the method and model proposed in the present invention, the prediction experimental results are further analyzed, and the mean and variance of the average absolute error and root mean square error obtained on 6 test batteries for the four schemes and configurations proposed in the present invention and the three existing physical information neural network methods are calculated. As shown in Table 4, mean(RMSE) represents the mean of the root mean square error calculation obtained on 6 test batteries, var(RMSE) represents the variance of the root mean square error calculation obtained on 6 test batteries, mean(MAE) represents the mean of the average absolute error calculation obtained on 6 test batteries, and var(MAE) represents the variance of the average absolute error calculation obtained on 6 test batteries. The smaller the mean and variance of the two errors, the smaller the test error and the smaller the fluctuation, that is, the better the generalization performance.

[0132] Table 4 Mean and variance of the error on six test batteries

[0133]

[0134] As can be seen from Table 4, LiB1 proposed in the present invention obtains the smallest mean and variance of the root mean square error, as well as the smallest mean of the mean absolute error; LiB3 proposed in the present invention obtains the optimal variance of the mean absolute error. The mean and variance of the mean absolute error and root mean square error obtained by the four schemes and configurations proposed in the present invention on 6 test batteries are smaller than those of the three existing physical information neural network methods. It can be seen that the model and method proposed in the present invention have better generalization performance than the three existing physical information neural network methods under four different schemes and configurations.

[0135] Based on the experimental results and analysis, the lithium-ion battery prediction and health management method based on deep neural network proposed in the present invention improves the accuracy of lithium-ion health status prediction and has better generalization performance.

Claims

1. A lithium-ion battery prediction and health management method based on a deep neural network, characterized in that: The steps include: Step 1: Build a model based on deep neural network: The deep neural network-based model includes a fully connected deep neural network and a global normalized deep neural fuzzy system, and the global normalized deep neural fuzzy system includes a first module and a second module connected to each other; The outputs of the fully connected deep neural network are respectively connected to the first module and the parameter optimization module; the output of the second module is connected to the parameter optimization module; Step 2: Learn the deep neural network based model in step 1: Extract the characteristic parameter set of the battery charging and discharging process as a learning sample; Inputting the learning sample into the fully connected deep neural network, the output of the fully connected deep neural network is the predicted value of the health state of the lithium ion, represented by U(t,x), where t is the main characteristic parameter and x is the secondary characteristic parameter; Calculate the partial differential of U(t,x) with respect to each characteristic parameter to obtain a partial differential set; Taking at least one of the partial differential corresponding to the secondary feature parameter in the partial differential set, the learning sample and U(t,x) as a feature set, inputting the feature set into the first module to obtain an expression of the feature set, inputting the expression of the feature set into the second module, performing nonlinear mapping through multiple learnable parameters, and obtaining output features of the global normalized deep neural fuzzy system; Based on the loss between the predicted value and the measured value of the lithium ion health state obtained by the parameter optimization module, the loss of the expression characteristics of the feature set, and the loss of the partial differential of the main feature parameters, a deep neural network-based model after learning is obtained; Step 3: Input the characteristic parameters of the battery to be tested into the deep neural network-based model learned in step 2 to obtain a predicted value of the health status of the battery to be tested.

2. The lithium-ion battery prediction and health management method based on deep neural network according to claim 1, characterized in that: The second module is a deep neural network, and the deep neural network includes N G Small modules BG are connected in sequence and stacked in a chain. A fully connected layer is connected after the last small module BG. The deep neural network is as follows: Among them, BG1, BG2, …, BG NG They represent the input small module starting from the input end to the output small module at the output end; f1(·) represents the fully connected layer, and f2(·) represents the activation function; is the output of the first module.

3. The lithium-ion battery prediction and health management method based on deep neural network according to claim 2, characterized in that: When t is selected as the internal impedance of the battery, x includes: the average temperature of the battery, the charging time, the variance of the incremental capacity analysis during the charging or discharging process, the skewness of the incremental capacity analysis during the charging or discharging process, and the minimum value of the incremental capacity analysis during the charging or discharging process; When t selects the minimum value of the incremental capacity analysis during the battery charging or discharging process, the x includes: battery internal impedance, average temperature, charging time, variance of the incremental capacity analysis during the charging or discharging process, and skewness of the incremental capacity analysis during the charging or discharging process.

4. The lithium-ion battery prediction and health management method based on deep neural network according to claim 3, characterized in that: The parameter optimization module is the loss function L, as follows: Where: L U =|U(t,x)-y| 2 represents the error between the predicted value U(t,x) of the lithium ion health state by the fully connected deep neural network and the measured value y of the lithium ion health state; L F =|F| 2 Represents the square error of function F, where function F is defined as F: =U t -G;L Ft =|F t | 2 Indicates F t The square error of t It means the partial derivative of F with respect to parameter t.

5. The lithium-ion battery prediction and health management method based on deep neural network according to claim 4, characterized in that: The first module includes a batch normalization layer, a first layer, a second layer, a third layer and a fourth layer connected in sequence; wherein: In the first layer, each neuron represents a membership function, which indicates the input feature z d Satisfy fuzzy set Z R,D r=1,...,R,d=1,...,D, and output; using Gaussian membership function, as follows: Where: m r,d and σ r,d is the Gaussian mean parameter and Gaussian standard deviation parameter, indicating the mean and standard deviation of the Gaussian membership function, m r,d The value range is real number, σ r,d The value range is positive real number; input feature z d The partial differentials {t,x,U,U) corresponding to the secondary features in the learning sample input features, the fully connected deep neural network and the partial differential set are selected. x } after normalization; In the second layer, each neuron multiplies the input and outputs the product as the activation strength of the corresponding fuzzy rule, as follows: In the third layer, the normalized activation strength calculated for each neuron is: Where: R is the number of fuzzy rules. In the fourth layer, each neuron corresponds to a fuzzy rule: where g r (z) = [b r，0 , b r，1 …, b r，D ]×[1,z1,…,z D ] T ,r=1,...,R,d=0,1,...,Dwhere b r,d Represents a fuzzy parameter, whose value range is a real number.

6. The lithium-ion battery prediction and health management method based on deep neural network according to claim 5, characterized in that: The fully connected deep neural network includes N U Small modules BL connected in sequence, and N U The small modules BL are stacked sequentially in a chain. At the back end of the chain, a fully connected layer is connected after the last small module BL, as follows: U(t,x)=f1(BL NU (BL NU-1 (...BL2(BL1(t,x) ) ) ) ) (6); Among them, BL1, BL2, ..., BL NU They represent the input small module starting from the input end to the output small module at the output end; f1 is the fully connected layer.

7. The model in the lithium-ion battery prediction and health management method based on deep neural network according to any one of claims 1 to 6, characterized in that: It includes a fully connected deep neural network and a global normalized deep neural fuzzy system, wherein the global normalized deep neural fuzzy system includes a first module and a second module connected to each other; The output of the fully connected deep neural network is connected to the first module and the parameter optimization module respectively; the output of the second module is connected to the parameter optimization module.

Images