Lithium battery SOH estimation method based on Box-Cox transform exponential model and Conv-LSTM

Abstract

The present invention discloses a lithium battery SOH estimation method based on a Box‑Cox transform exponential model and Conv‑LSTM, and belongs to the field of automation technology. In order to overcome the problems of the huge amount of data required for model establishment and insufficient SOH estimation accuracy in the prior art, a hybrid neural network lithium-ion battery SOH modeling method based on deep learning is proposed. The technical solution of the present invention consists of five aspects: data acquisition, construction of an overall trend prediction model, establishment of a residual data set, construction of a nonlinear feature prediction model, and performance evaluation. The method is based on a Box‑Cox transform exponential model, a residual structure, and a convolutional long short-term memory network, and utilizes the charging fragments in the charging process of a lithium-ion battery to realize the SOH estimation within the service life of the lithium-ion battery. Compared with the traditional deep learning SOH estimation method, the present invention decomposes the SOH features into overall trends and nonlinear features, and then estimates them separately, which can effectively improve the SOH estimation accuracy performed by the model.

Description

Technical Field

[0001] The present invention belongs to the field of automation technology, and in particular relates to a lithium battery SOH estimation method based on a Box-Cox transformation exponential model and Conv-LSTM. Background Art

[0002] Against the backdrop of the Chinese government's "carbon peak and carbon neutrality" goals, new energy vehicles have gradually become the focus of the industry. With the strong support of national policies, the development of pure electric vehicles has been particularly rapid. However, as the core component of pure electric vehicles, batteries still have many technical problems that need to be solved. For example, there are still technical bottlenecks in battery thermal management and accurate estimation of battery state of health (SOH). In recent years, the number of electric vehicles has increased year by year, and a large number of lithium batteries will be retired in the near future. They need to be properly handled. In order to respond to the relevant national policies on the cascade utilization of lithium batteries and enable lithium batteries to continue to play a role in other places after being retired from electric vehicles, it is necessary to achieve high-precision estimation of battery SOH values. Therefore, the task of breaking through the technical barriers of high-precision SOH estimation is urgent.

[0003] Currently, the commonly used methods for estimating lithium battery SOH can be roughly divided into the following two categories: 1. Physical model-based methods; 2. Data-driven methods. Method 1 requires detailed research on the internal physical mechanisms of the battery, and the differences in internal mechanisms between different battery types require model remodeling. Therefore, a data-driven approach was chosen this time. However, in previous studies, the estimation model needed to simultaneously estimate both linear and nonlinear features in the SOH curve, which placed high demands on the estimation model's generalization capabilities. Summary of the Invention

[0004] In order to overcome the high generalization capability requirement of an estimation model in the prior art that needs to simultaneously understand the overall trend and nonlinear characteristics of the SOH transformation process, a lithium battery SOH estimation method based on the Box-Cox transformation exponential model and the neural network model is proposed. The technical solution of the present invention is to collect data, construct an SOH overall trend model, construct an SOH nonlinear residual data set, establish a nonlinear neural network estimation model and conduct online training, and finally integrate the linear estimation module and the nonlinear estimation module to establish the final lithium-ion battery SOH estimation model. This method can decompose complex features into overall trends and nonlinear characteristics, and then estimate them separately, which can effectively improve the SOH estimation accuracy of the model.

[0005] The technical solutions specifically adopted in the present invention are as follows:

[0006] A lithium battery health status estimation method based on the Box-Cox transform exponential model and Conv-LSTM, the steps of which are as follows:

[0007] S1: Collect charging segments of lithium-ion batteries during charge and discharge cycles under different working conditions as sample data, and record the battery health status (SOH) values ​​corresponding to the sample data during the battery charge and discharge cycles as labels, thereby forming a series of first training samples consisting of sample data and labels;

[0008] S2: Perform Box-Cox transformation on the labels of all the first training samples collected in S1, and use the sample data in each first training sample as the independent variable and the transformed SOH value as the dependent variable to obtain the fitting equation by exponential fitting. Then, cascade the Box-Cox inverse transformation operation after the fitting equation, and the obtained Box-Cox transformation exponential model is used as the linear part estimation model;

[0009] S3: For each first training sample collected in S1, the sample data in the first training sample is input into the Box-Cox transformation exponential model to obtain the overall trend of the SOH, and then the residual of the label in the first training sample and the obtained overall trend of the SOH is calculated. The label in the first training sample is replaced by the obtained residual value to obtain a second training sample;

[0010] S4: using all the second training samples to train a neural network model, fine-tuning the weights of the neural network so that it can predict the residual value based on the sample data, and obtaining a nonlinear part estimation model after the training is completed;

[0011] S5: Combine the linear part estimation model and the nonlinear part estimation model to form a complete lithium-ion battery SOH estimation model; use the charging segment of the lithium-ion battery during the charge and discharge cycle under the predicted working conditions as the input of the lithium-ion battery SOH estimation model, and use the linear part estimation model to predict the overall trend part of the lithium-ion battery SOH based on the input, and use the nonlinear part estimation model to predict the residual part of the lithium-ion battery SOH based on the input, and add the overall trend part and the residual part to obtain the estimated value of the lithium-ion battery SOH.

[0012] Preferably, in step S1, the sample data of the lithium-ion battery during the t-th cycle is the charging segment characteristic data x of the lithium battery charging from the power I1 to the power I2. t , the label Y corresponding to the sample data t The SOH value is defined as the ratio of the current capacity of the battery to the nominal capacity.

[0013] Preferably, in step S2, the fitting equation is fitted using the least squares method.

[0014] Preferably, in step S3, for each first training sample (x t ,Y t), the charging segment feature data x t Input the Box-Cox transformation index model, and the model output value is the overall trend of SOH Then Y t and Perform residual calculation and use the difference between the two as the new label Thus, the first training sample is transformed into the second training sample (x t ,N t ).

[0015] As an advantage, in step S4, the Conv-LSTM model is used in the neural network model, and the input feature data x t After the Conv-LSTM network, the output y is obtained t , this output y t With the original input x t After the residual connection, the features are extracted again through one-dimensional convolution, and finally the residual part of the final output of the lithium-ion battery SOH is obtained through the fully connected layer.

[0016] Preferably, in step S5, the final lithium-ion battery SOH estimation model calculates the charging segment x of the lithium-ion battery charge and discharge cycle under the predicted working condition when performing SOH estimation. t At the same time, the two combined models in the lithium-ion battery SOH estimation model are input. On the one hand, the Box-Cox transformation index model is used to perform a Box-Cox inverse transformation operation on the output value of the fitting equation to obtain the overall trend of SOH. On the other hand, the SOH residual is estimated by the neural network model Finally, the estimated SOH value of lithium-ion battery is obtained

[0017] Preferably, the charging segment during the charge-discharge cycle of the lithium-ion battery is the voltage, current and capacity during the charging process of the lithium-ion battery.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The method of the present invention is based on the Box-Cox transform exponential model, residual structure and convolutional long short-term memory network, and uses the voltage, current and capacity of the lithium-ion battery during the charging process to realize the SOH estimation of the lithium-ion battery during the service life. Compared with the traditional deep learning SOH estimation method, the present invention decomposes the SOH characteristics into overall trends and nonlinear characteristics, and then estimates them separately, which can effectively improve the SOH estimation accuracy performed by the model. The technical solution of the present invention consists of five aspects: data acquisition, construction of an overall trend prediction model, establishment of a residual data set, construction of a nonlinear feature prediction model, and performance evaluation. It establishes a lithium-ion battery SOH estimation method based on the Box-Cox transform exponential model and convolutional long short-term memory network. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the flow chart for estimating SOH described in this embodiment;

[0021] Figure 2 This is a schematic diagram of the exponential model fitting after Box-Cox transformation described in this embodiment;

[0022] Figure 3 This is the neural network estimation flow chart described in this embodiment;

[0023] Figure 4 This is a schematic diagram of the nonlinear part of the neural network fitting described in this embodiment;

[0024] Figure 5 This is the overall estimated effect diagram described in this embodiment. DETAILED DESCRIPTION

[0025] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. It should be noted that the technical features of the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0027] As a preferred implementation of the present invention, a method for estimating the health status of a lithium battery based on a Box-Cox transform exponential model and Conv-LSTM is provided, and the steps are as follows:

[0028] S1: Collect charging segments of lithium-ion batteries during charge and discharge cycles under different working conditions as sample data, and record the battery health status (SOH) values ​​corresponding to the sample data during the battery charge and discharge cycles as labels, thereby forming a series of first training samples consisting of sample data and labels.

[0029] In step S1 of the present invention, the sample data of the lithium-ion battery during the t-th cycle is the charging segment characteristic data x of the lithium battery charging from the power I1 to the power I2. t , the label Y corresponding to the sample data t The SOH value is defined as the ratio of the current capacity of the battery to the nominal capacity.

[0030] According to the definition of the SOH value above, it is necessary to collect the nominal capacity and full charge capacity of the battery and calculate the SOH value of the battery system. The SOH value is defined from the perspective of capacity, and the definition formula is as follows:

[0031]

[0032] Where C0 represents the nominal capacity of the battery, C t Indicates the current capacity of the battery. The calculated SOH value of the tth battery charge and discharge cycle is recorded as Y t . Use this SOH value as the label of the training set, and use this label Y t With input x t One-to-one correspondence, in order to meet the form requirements of the training samples, the training samples at this time (x t ,Y t ) is recorded as the first training sample.

[0033] It should be noted that the specific power values ​​of the above power I1 and power I2 need to be reasonably selected according to the power range of the actual working conditions of the lithium-ion battery, as long as all samples and subsequent predictions are kept consistent. t The specific indicators need to be selected according to actual conditions. In subsequent embodiments of the present invention, the charging segment during the charge and discharge cycle of the lithium-ion battery is the voltage, current and capacity during the charging process of the lithium-ion battery.

[0034] To ensure consistency across samples, lithium-ion batteries must be collected at a preset, uniform temperature. Characteristic data can be collected during constant-current / constant-voltage charging. The collected raw data is also ideally preprocessed to remove incomplete battery data from the battery cycle and construct the data into a sample matrix for model input.

[0035] S2: Perform Box-Cox transformation on the labels of all the first training samples collected in S1, and use the sample data in each first training sample as the independent variable and the transformed SOH value as the dependent variable. Use exponential fitting to obtain the fitting equation, and then cascade the Box-Cox inverse transformation operation after the fitting equation to obtain the Box-Cox transformation exponential model as the linear part estimation model.

[0036] Step S2 above primarily constructs a Box-Cox transformation exponential model based on the historical SOH values ​​of lithium-ion batteries. Ideally, a battery's SOH and cycle count are linearly related. However, due to the complexity of the battery aging process, this correlation is unsatisfactory. Empirical prediction methods lack the ability to withstand data fluctuations, which can easily lead to divergent predictions. Therefore, the Box-Cox transformation is employed, using a parameterized generalized power transformation method to mitigate data irregularities such as non-additivity, non-normality, and heteroscedasticity in statistical modeling.

[0037] The Box-Cox transformation belongs to the prior art and is used to reduce the correlation between unobservable errors and predictive variables. In the present invention, the calculation formula of the collected lithium-ion battery SOH using the Box-Cox transformation is as follows:

[0038]

[0039] Where y t and are SOH before and after transformation respectively; c is the guarantee (y t +c) is a positive constant; λ is the parameter to be identified. Methods for determining λ mainly include maximum likelihood estimation and Bayesian methods. In this paper, maximum likelihood estimation is used to identify λ. That is, λ is searched traversally until the joint distribution likelihood function of the data reaches its maximum value.

[0040] In step S2 of the present invention, the fitting equation is fitted using the least squares method, that is, the x corresponding to all first training samples t and By performing exponential fitting using the least squares method, an exponential fitting equation can be obtained. The dependent variable of the exponential fitting equation is Therefore, a Box-Cox inverse transformation is cascaded after the exponential fitting equation, and the curve after exponential fitting can be subjected to Box-Cox inverse transformation to obtain the required y t The corresponding Box-Cox inverse transformation formula is as follows:

[0041]

[0042] S3: For each first training sample collected in S1, the sample data in the first training sample is input into the Box-Cox transformation exponential model, and the exponential fitting equation is used to calculate The overall trend of SOH is obtained by inverse Box-Cox transformation, and the residual of the label in the first training sample and the obtained overall trend of SOH is calculated. The residual value is used to replace the label in the first training sample to obtain the second training sample.

[0043] In step S3 of the present invention, for each first training sample (x t ,Y t ), the charging segment feature data x t Input the Box-Cox transformation index model, and the model output value is the overall trend of SOH Then Y t and Perform residual calculation and use the difference between the two as the new label Thus, the first training sample is transformed into the second training sample (x t ,N t ). All the second training samples constitute the residual data set, which enables the neural network module to predict the nonlinear part in the SOH time series.

[0044] S4: Using all the second training samples to train the neural network model, fine-tuning the weights of the neural network so that it can predict the residual value based on the sample data, and obtaining a nonlinear part estimation model after the training is completed.

[0045] In step S4 of the present invention, the Conv-LSTM model is used in the neural network model, and the neural network estimation process is as follows: Figure 3 As shown. Input feature data x t After the Conv-LSTM network, the output y is obtained t , this output y t With the original input x t After the residual connection, the features are extracted again through one-dimensional convolution, and finally the residual part of the final output of the lithium-ion battery SOH is obtained through the fully connected layer.

[0046] The Conv-LSTM model is an existing technology. The specific operations in Conv-LSTM are briefly introduced as follows:

[0047] 1) Forget gate f tThe sigmoid activation function generates a value ranging from 0 to 1, which determines the amount of memory cells from the previous time step that enter the current time step. When it is close to 0, most of the value of the memory cell in the previous time step will be erased; when it is close to 1, the opposite will happen. The formula is as follows:

[0048]

[0049] 2) Input gate i t The sigmoid activation function generates a value ranging from 0 to 1 to determine how many candidate memory units of the current time step can be added to the storage unit. The formula is as follows:

[0050]

[0051] 3) Candidate memory units It is calculated by the tanh activation function, and the formula is as follows:

[0052]

[0053] 4) Update Gate Then manage the memory unit of the previous time step Memory cells up to the current time step The update process is mainly controlled by the previous forget gate and input gate. The forget gate determines the last memory unit. How much is effective at the current time step? The memory gate determines the candidate memory unit of the current time step. The formula for how much it will take effect is as follows:

[0054]

[0055] 5) Input gate o t Determine the specific output of the network and make adjustments to the hidden state based on this Update to transfer the information of the current time step to the next time step. The formula is as follows:

[0056]

[0057]

[0058] In the above formula, σ represents the sigmoid activation function, w represents the parameters of the weight matrix, Represents the current input, Represents the previous hidden layer output.

[0059] The Conv-LSTM model is designed to complete the information transfer between cycles while encoding the time information within the cycle. Based on the Conv-LSTM model, the battery characteristic data x is input.t It actually contains two types of time information, one is x1, x2, x3..., x n The time information between the cycles is x t The characteristic curve itself changes with time during the charging cycle, and it also carries time information. Adding a one-dimensional convolution operator to the traditional LSTM network to form Conv-LSTM allows the neural network to extract the time information within the cycle while extracting the timing information between cycles. After the training is completed, the estimated result of the network is recorded as

[0060] S5: Combine the linear part estimation model and the nonlinear part estimation model to form a complete lithium-ion battery SOH estimation model; use the charging segment of the lithium-ion battery during the charge and discharge cycle under the predicted working conditions as the input of the lithium-ion battery SOH estimation model, and use the linear part estimation model to predict the overall trend part of the lithium-ion battery SOH based on the input, and use the nonlinear part estimation model to predict the residual part of the lithium-ion battery SOH based on the input, and add the overall trend part and the residual part to obtain the estimated value of the lithium-ion battery SOH.

[0061] In step S5 of the present invention, the final lithium-ion battery SOH estimation model calculates the charging segment x of the lithium-ion battery charge and discharge cycle under the predicted working condition when performing SOH estimation. t At the same time, the two combined models in the lithium-ion battery SOH estimation model are input. On the one hand, the Box-Cox transformation index model is used to perform a Box-Cox inverse transformation operation on the output value of the fitting equation to obtain the overall trend of SOH. On the other hand, the SOH residual is estimated by the neural network model Finally, the estimated SOH value of lithium-ion battery is obtained

[0062] Based on the above method, the technical effects thereof will be further demonstrated below in combination with specific embodiments. The definitions of some parameters are as described above and will not be repeated here.

[0063] Example

[0064] like Figure 1 As shown, in this embodiment, the lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM includes the following steps:

[0065] Step (1): Data collection, collecting various data characteristics of the lithium-ion battery during constant current-constant voltage charging at a preset temperature. The specific process is as follows:

[0066] 1) Collect input features of lithium-ion batteries. The collected battery data is preprocessed by removing incomplete battery data from the battery cycle and concatenating the data into a sample matrix for model input. The input for this example is a 4000×3 sample matrix, where 4000 and 3 represent the number of data samples and input variables within a charging cycle, respectively. The specific form of the sample matrix is ​​as follows:

[0067]

[0068] t represents the number of charge and discharge cycles of the current lithium-ion battery, V i , I i 、C i They represent the voltage, current, and capacity collected during the current lithium-ion charge and discharge cycle.

[0069] 2) Collect the SOH value of lithium ions as the label of the training sample data. The SOH value collected in this embodiment is defined from the perspective of capacity. It is necessary to calculate the SOH value of the battery system by collecting the nominal capacity and full charge capacity of the battery. The calculation formula is as follows:

[0070]

[0071] Where C0 represents the rated capacity of the battery, C t Indicates the current capacity of the battery. And the calculated SOH value of the tth battery charge and discharge cycle is recorded as Y t . Use this SOH value as the label of the training set, and correspond this label to the input one by one, and then get a series of first training samples (x t ,Y t ).

[0072] Step (2): Construct a Box-Cox transformation exponential model for estimating the overall trend of lithium-ion battery SOH. The Box-Cox transformation is used to reduce the correlation between unobservable errors and predictive variables. The specific operation is as follows:

[0073] 1) The collected lithium-ion battery SOH is transformed using the Box-Cox transformation. The Box-Cox transformation calculation formula is as follows:

[0074]

[0075] Where y t and are the variables before and after the transformation respectively; c is to ensure (y t +c) is a positive constant; λ is the parameter to be identified. Methods for determining λ include maximum likelihood estimation and Bayesian methods. This paper uses maximum likelihood estimation to identify λ. Specifically, it searches for λ until the joint distribution likelihood function of the data reaches its maximum value.

[0076] 2) Perform an exponential fit on the Box-Cox transformed data using the least squares method, and then connect the inverse Box-Cox transformation to the exponential fitting equation to form a Box-Cox transformation exponential model. The corresponding inverse Box-Cox transformation formula is as follows:

[0077]

[0078] In this embodiment, the SOH index fitting after Box-Cox transformation is compared with the direct index fitting. Figure 2 shown.

[0079] Step (3): Combine the Box-Cox exponential model to construct the residual data set. For each first training sample (x t ,Y t ), change x t Input the Box-Cox transformation exponential model in step (2), and the output is the predicted value The actual SOH value Y of the lithium-ion battery recorded by the training sample t Subtract the predicted value predicted by the Box-Cox exponential model in step (2) The difference between the two is used as the label N trained by the neural network t ,Right now Thus, we get a series of second training samples (x t ,N t ). The residual dataset is used to train the neural network module so that it can predict the nonlinear part in the SOH time series.

[0080] Step (4): Construct a lithium-ion battery SOH estimation model based on a convolutional long short-term memory network. The neural network estimation process in the present invention is as follows: Figure 3 As shown, its core is the Conv-LSTM network. Input x t The output is obtained through the Conv-LSTM network, and this output will be combined with the input x t The concatenation is then fed into the network to form a residual map. After that, the residual connection result is subjected to a one-dimensional convolution to extract features again, and finally the output is integrated through a fully connected layer to obtain the residual part of the final output lithium-ion battery SOH.

[0081] Like LSTM, Conv-LSTM's core concepts are memory cells and "gate" structures. Specifically, the three gates are the forget gate, the input gate, and the output gate. The forget gate determines how many memory cells from the previous time step can enter the current time step, the input gate determines how many candidate memory cells from the current time step can be added to the memory cells, and the output gate is responsible for the final output of the current time step and the update of the hidden state. Conv-LSTM adds a convolution operator to the traditional LSTM to encode temporal information within the loop. The specific operation is as follows:

[0082] 1) Forget gate f t The sigmoid activation function generates a value ranging from 0 to 1, which determines the amount of memory cells from the previous time step that enter the current time step. When it is close to 0, most of the value of the memory cell in the previous time step will be erased; when it is close to 1, the opposite will happen. The formula is as follows:

[0083]

[0084] 2) Input gate i t The sigmoid activation function generates a value ranging from 0 to 1 to determine how many candidate memory units of the current time step can be added to the storage unit. The formula is as follows:

[0085]

[0086] 3) Candidate memory units It is calculated by the tanh activation function, and the formula is as follows:

[0087]

[0088] 4) Update Gate Then manage the memory unit of the previous time step Memory cells up to the current time step The update process is mainly controlled by the previous forget gate and input gate. The forget gate determines the last memory unit. How much is effective at the current time step? The memory gate determines the candidate memory unit of the current time step. The formula for how much it will take effect is as follows:

[0089]

[0090] 5) Input gate o t Determine the specific output of the network and make adjustments to the hidden state based on this Update to transfer the information of the current time step to the next time step. The formula is as follows:

[0091]

[0092]

[0093] In the above formula, σ represents the sigmoid activation function, w represents the parameters of the weight matrix, Represents the current input, Represents the previous hidden layer output.

[0094] The residual data set obtained in step (3) is used to train the above neural network model. The model parameters are adjusted through the test set. In this embodiment, the loss function used is RMSE, the trainer used is Adam trainer, and the initial learning rate is set to 1×10 -5 The total number of training rounds is 2000, and the learning rate is reduced by half after 500 rounds. After the model training is completed, the test set data is used to evaluate the neural network model obtained by training. The convolutional long short-term memory network module is mainly responsible for predicting the nonlinear part of the lithium-ion battery. The network prediction result is Neural network output The effect is as Figure 4 shown.

[0095] Step (5): Integrate the linear estimation module and the nonlinear estimation module to construct a lithium-ion battery SOH estimation model based on the Box-Cox exponential model and Conv-LSTM, that is, in The Box-Cox transformation index model is used to perform a Box-Cox inverse transformation on the output value of the fitting equation to obtain the overall trend of SOH. The Conv-LSTM module is responsible for predicting the nonlinear part of the output, namely the SOH residual part.

[0096] To verify the prediction accuracy of the above model, the input matrix of the test set is fed into the overall estimation module in sequence to obtain the SOH estimation result. The result is compared with the true value using RMSE and MAE, which are defined as follows:

[0097]

[0098]

[0099] Where n is the number of cycles of the burn-in dataset, is the SOH value predicted by the neural network based on the input data, y i(X) is the real SOH value. The real-time SOH estimation is achieved based on the learned lithium-ion battery SOH estimation model, and the result is compared with the real value to test the model estimation performance. The lithium-ion battery SOH estimation model of the present invention is compared with the traditional convolutional neural network model for estimation effect. The overall prediction value and error are shown in Figure 2. Figure 5 shown.

[0100] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A lithium battery SOH estimation method based on Box-Cox transform exponential model and Conv-LSTM, characterized in that: Here are the steps: S1: Collect charging segments of lithium-ion batteries during charge and discharge cycles under different working conditions as sample data, and record the battery health status (SOH) values ​​corresponding to the sample data during the battery charge and discharge cycles as labels, thereby forming a series of first training samples consisting of sample data and labels; S2: Perform Box-Cox transformation on the labels of all the first training samples collected in S1, and use the sample data in each first training sample as the independent variable and the transformed SOH value as the dependent variable to obtain the fitting equation by exponential fitting. Then, cascade the Box-Cox inverse transformation operation after the fitting equation, and the obtained Box-Cox transformation exponential model is used as the linear part estimation model; S3: For each first training sample collected in S1, the sample data in the first training sample is input into the Box-Cox transformation exponential model to obtain the overall trend of the SOH, and then the residual of the label in the first training sample and the obtained overall trend of the SOH is calculated. The label in the first training sample is replaced by the obtained residual value to obtain a second training sample; S4: using all the second training samples to train a neural network model, fine-tuning the weights of the neural network so that it can predict the residual value based on the sample data, and obtaining a nonlinear part estimation model after the training is completed; S5: Combine the linear part estimation model and the nonlinear part estimation model to form a complete lithium-ion battery SOH estimation model; use the charging segment of the lithium-ion battery during the charge and discharge cycle under the predicted working conditions as the input of the lithium-ion battery SOH estimation model, and use the linear part estimation model to predict the overall trend part of the lithium-ion battery SOH based on the input, and use the nonlinear part estimation model to predict the residual part of the lithium-ion battery SOH based on the input, and add the overall trend part and the residual part to obtain the estimated value of the lithium-ion battery SOH.

2. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 1 is characterized in that: In step S1, the sample data of the lithium-ion battery during the t-th cycle is the charging segment characteristic data x of the lithium battery charging from the power I1 to the power I2. t , the label Y corresponding to the sample data t The SOH value is defined as the ratio of the current capacity of the battery to the nominal capacity.

3. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 1 is characterized in that: In step S2, the fitting equation is fitted using the least squares method.

4. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 2, characterized in that: In step S3, for each first training sample (x t ,Y t ), the charging segment feature data x t Input the Box-Cox transformation index model, and the model output value is the overall trend of SOH Then Y t and Perform residual calculation and use the difference between the two as the new label Thus, the first training sample is transformed into the second training sample (x t ,N t ).

5. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 4 is characterized in that: In step S4, the Conv-LSTM model is used in the neural network model, and the input feature data x t After the Conv-LSTM network, the output y is obtained t , this output y t With the original input x t After the residual connection, the features are extracted again through one-dimensional convolution, and finally the residual part of the final output of the lithium-ion battery SOH is obtained through the fully connected layer.

6. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 1, characterized in that: In step S5, the final lithium-ion battery SOH estimation model calculates the charging segment x of the lithium-ion battery charge and discharge cycle under the predicted working condition when performing SOH estimation. t At the same time, the two combined models in the lithium-ion battery SOH estimation model are input. On the one hand, the Box-Cox transformation index model is used to perform a Box-Cox inverse transformation operation on the output value of the fitting equation to obtain the overall trend of SOH. On the other hand, the SOH residual is estimated by the neural network model Finally, the estimated SOH value of lithium-ion battery is obtained 7. The lithium battery SOH estimation method based on the Box-Cox transform exponential model and Conv-LSTM according to claim 1, characterized in that: The charging segment during the charge-discharge cycle of the lithium-ion battery is the voltage, current and capacity during the charging process of the lithium-ion battery.

Images