A lithium battery soh and rul collaborative prediction method based on nonlinear enhanced LSTM

Abstract

The present application relates to a kind of lithium battery SOH and RUL collaborative prediction method based on nonlinear enhancement LSTM, belong to battery life prediction technical field, solve the problem of large amount of electrochemical data demand, single and insufficient feature extraction, waste of computing resources and deep, weak nonlinear feature extraction capability and low SOH and RUL prediction accuracy in prior art.The present application is based on data-driven, through multi-channel convolutional neural network multi-dimensional feature extraction, and based on nonlinear enhancement LSTM model carries out battery sequence data feature extraction, in-depth mining nonlinear feature of battery parameter, SOH evaluation result is used as the input of RUL prediction, using neural network method carries out RUL prediction, the prediction accuracy of SOH and RUL of lithium battery is high;Data feature is fully extracted, and the amount of electrochemical data demand is reduced;Using the same pre-training nonlinear enhancement LSTM model shares input, and the computing resource is greatly saved.

Description

Technical Field

[0001] This invention relates to the field of battery life prediction technology, specifically to a method for co-predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM. Background Technology

[0002] In recent years, the development of new energy vehicles has been booming. As a key component of new energy vehicles, the reliability and safety of lithium batteries significantly impact the lifespan of these vehicles and the safety of passengers. To ensure the normal operation of vehicles, it is necessary to monitor the safety status of lithium batteries. In the event of a lithium battery malfunction, a timely warning should be issued to the user, reminding them to repair or replace the battery to ensure driving safety. Previously, battery manufacturers assessed the health status of battery packs by providing parameters such as voltage and current to ensure safety. However, due to the diversity of battery types and the influence of actual driving conditions, the accuracy of these assessments needs improvement.

[0003] Currently, methods for estimating SOH (State of Health) and RUL (Remaining Useful Life) in the battery field can be structurally divided into two categories: one is to predict SOH and RUL separately; the other is to predict SOH and RUL jointly. SOH separate prediction methods have relatively high estimation errors. Direct evaluation methods in separate prediction rely heavily on electrochemical knowledge; furthermore, data-driven SOH separate prediction methods extract SOH features from only a single dimension, resulting in limited and insufficient feature extraction, failing to fully capture battery parameter characteristics. A common problem with RUL separate prediction methods is that the battery's SOH value affects the RUL estimate, an effect not considered in these methods. Joint SOH and RUL prediction can be divided into two main approaches. One is to share input, using two independent models for SOH and RUL prediction; the other is to predict SOH and RUL using the same model. The first method first extracts features from the charging curve as model input, then uses independent SOH estimation and RUL prediction models to predict SOH and RUL. The second approach first extracts relevant features from existing datasets, performs SOH estimation, and uses the SOH estimation results as input for RUL prediction. SOH and RUL are intrinsically related, and predicting either SOH or RUL individually has limitations. Joint prediction of SOH and RUL not only uncovers their inherent connection but also saves computational resources by using the same model and sharing input. However, current joint prediction methods for SOH and RUL mostly extract features using fixed convolutional kernel sizes, limiting the extraction of multi-dimensional features of battery parameters. Furthermore, they use traditional sequence models for feature extraction, which has limited nonlinear feature extraction capabilities. Existing methods that use SOH estimation results as input for RUL prediction obtain the model's input features through correlation analysis followed by multi-health feature fusion. This method of obtaining model input features struggles to capture deep-seated features of battery parameters; moreover, existing joint estimation methods mostly rely on traditional sequence models for feature extraction, limiting their ability to extract nonlinear features.

[0004] In summary, existing technologies suffer from problems such as large demand for electrochemical data, limited and insufficient feature extraction, waste of computational resources, weak ability to extract deep and nonlinear features, and low accuracy in predicting SOH and RUL of lithium batteries. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method for synergistic prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM, which solves the problems of large electrochemical data requirements, single and insufficient feature extraction, waste of computing resources, and weak deep and nonlinear feature extraction capabilities, resulting in low prediction accuracy of SOH and RUL of lithium batteries in the prior art.

[0006] This invention provides a method for co-predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM (Long Short Term Memory), comprising the following steps:

[0007] Step S1. Obtain battery data and select battery characteristic parameters from it;

[0008] Step S2. Based on the battery feature parameters, obtain the feature vector for estimating SOH, and input it into a multi-channel convolutional neural network to extract features and obtain the final feature vector for estimating SOH;

[0009] Step S3. Input the final feature vector used to obtain the estimated SOH into the pre-trained nonlinear augmented LSTM model, and process it to obtain the deep features used to obtain the estimated SOH;

[0010] Step S4. Based on the battery data, obtain the initial capacity; based on the deep features used to obtain the estimated SOH, calculate the measured capacity, and then combine it with the initial capacity to calculate the estimated SOH;

[0011] Step S5. Obtain the historical SOH from the battery data, and combine it with the estimated SOH to obtain the feature vector for obtaining RUL. Input the feature vector into a multi-channel convolutional neural network to extract features and obtain the final feature vector for obtaining RUL.

[0012] Step S6. Input the final feature vector used to obtain RUL into the pre-trained nonlinear augmented LSTM model, and process it to obtain the deep features used to obtain RUL;

[0013] Step S7. Based on the battery data, obtain the battery cycle period at the start of the prediction; based on the deep features used to obtain RUL, calculate the battery cycle period at the end of its lifespan, and then combine it with the battery cycle period at the start of the prediction to calculate RUL.

[0014] Furthermore, step S1 includes the following specific steps:

[0015] Step S101. Obtain battery data, including battery voltage characteristics, current characteristics, temperature characteristics, time characteristics, and capacity characteristics;

[0016] Step S102. Use Pearson correlation coefficient and Spearman correlation coefficient to perform correlation calculation on battery data. Based on the correlation calculation results, select data from voltage characteristics, current characteristics, temperature characteristics and time characteristics, and use them together with capacity characteristics as battery characteristic parameters.

[0017] Furthermore, step S2 includes the following specific steps:

[0018] Step S201. Using the battery characteristic parameters as elements of the matrix, construct the battery parameter characteristic matrix, and obtain the eigenvector I used to estimate SOH from it;

[0019] Step S202. Input the feature vector I used to obtain the estimated SOH into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, where i = 1, 2, 3; after convolution processing, obtain the feature map F of the i-th channel used to obtain the estimated SOH. i n i = 1, 2, 3;

[0020] Step S203. For each channel used to obtain the estimated SOH, perform characteristic plot F. i n For i = 1, 2, 3, an attention mechanism is used to process the data to obtain the feature map F used to estimate the SOH. n ;

[0021] Step S204. For the characteristic map F used to obtain the estimated SOH n Max pooling is performed to obtain the final eigenvector I used to estimate SOH. final And output it.

[0022] Furthermore, the pre-trained nonlinear augmented LSTM model in steps S3 and S6 refers to a nonlinear augmented LSTM model pre-trained for SOH and RUL, including a forget gate, input gate, output gate, hidden layer, memory unit, and final output; wherein,

[0023] The calculation method for the forget gate is f. t =σ(tanh(c) t-1 )*P f +W xf I t +W hf h t-1 +b f ); where σ(·) is the sigmoid function, tanh(·) is the tanh function, t represents time, and c t-1 P represents the cell unit at the previous time step. f W represents the weight of the peephole in the forgetting gate.xf I represents the input vector weights of the forget gate. t W represents the input vector. hf h represents the hidden layer weights at the previous time step before the forget gate. t-1 b represents the output of the hidden layer at the previous time step. f Indicates the forget gate bias, f t Indicates the output of the forget gate;

[0024] The calculation method for the input gate is as follows: Among them, P i Indicates the input gate peephole weight, i t Indicates the input gate output;

[0025] The output gate is calculated as o t =σ(tanh(c) t )*P o +W xo I t +W ho h t-1 +b o ); where P o W represents the output gate peephole weight. xo W represents the weights of the input vector to the output gate. ho b represents the weight of the hidden layer above the output gate. o Indicates the output gate bias, o t Indicates the output of the output gate;

[0026] The hidden layer is calculated as follows: Among them, c t This represents the cell unit at time t;

[0027] The calculation method for memory units is as follows: in, W xx W represents the weights of the cell unit input vector. hc b represents the hidden layer weights of the cell unit at the previous time step. c Indicates cell unit bias;

[0028] The final output is calculated as y = W. full h t Among them, W full y represents the hidden layer weights at time t, and y represents the deep features obtained in step S3 for estimating SOH or the deep features obtained in step S6 for obtaining RUL, which are the final outputs of the pre-trained nonlinear augmented LSTM model.

[0029] Furthermore, step S5 includes the following specific steps:

[0030] Step S501. Obtain the historical SOH from the battery data, and then combine it with the estimated SOH to construct a feature matrix for obtaining RUL, and obtain the feature vector Y for obtaining RUL from it;

[0031] Step S502. Input the feature vector Y used to obtain RUL into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, i = 1, 2, 3; after convolution processing, obtain the feature map of the i-th channel used to obtain RUL.

[0032] Step S503. Develop feature maps for each channel used to obtain RUL. An attention mechanism is used to process the data, resulting in the feature map F' used to obtain the RUL. n ;

[0033] Step S504. For the feature map F' used to obtain RUL n Perform max pooling to obtain the final feature vector Y used to obtain RUL. final And output it.

[0034] Furthermore, the pre-trained nonlinear augmented LSTM model includes a model training phase for SOH and a model training phase for RUL; wherein the model training phase for SOH includes the following steps:

[0035] Obtain the battery dataset and construct the historical SOH dataset based on the battery dataset;

[0036] Perform steps S1 and S2 on the battery dataset to obtain the final feature vector dataset used to estimate SOH;

[0037] The dataset used to obtain the final feature vector of the estimated SOH is divided into SOH training set I. train SOH Validation Set I valid and SOH test set I test ;

[0038] SOH training set I train The model is trained using a nonlinear augmented LSTM model and the SOH validation set I is employed. valid The performance of the nonlinear augmented LSTM model is evaluated by calculating the model loss to determine the model hyperparameters for SOH, thus obtaining the nonlinear augmented LSTM model trained for SOH. The model loss is calculated using MSE (mean squared error) and RMSE (root mean squared error).

[0039] SOH test set Itest Input a nonlinear augmented LSTM model trained for SOH, perform nonlinear augmented LSTM model prediction, calculate the deep feature dataset used to estimate SOH, and complete the model training for SOH.

[0040] Furthermore, the model training phase for RUL includes the following steps:

[0041] Based on the deep feature dataset used to obtain the estimated SOH, the SOH is calculated and processed to obtain the estimated SOH dataset;

[0042] Perform step S5 on the historical SOH dataset and the estimated SOH dataset to obtain the final feature vector dataset used to obtain RUL;

[0043] The input dataset used to obtain the final feature vector of RUL is divided into the RUL training set Y. train RUL verification set Y valid and RUL test set Y test ;

[0044] RUL training set Y train The model is trained using a pre-trained nonlinear augmented LSTM model targeting SOH, and the RUL validation set Y is used. valid The model performance was evaluated by calculating the model loss to determine the model hyperparameters for RUL, resulting in a nonlinear augmented LSTM model trained for SOH and RUL; the model loss was calculated using MSE and RMSE.

[0045] Input RUL test set Y test We perform nonlinear enhanced LSTM model prediction, calculate the deep feature dataset used to obtain RUL, and complete the model training for RUL.

[0046] Furthermore, the model hyperparameters include the weights, biases, number of neural network layers, number of iterations, and learning rate in the nonlinear augmented LSTM model.

[0047] Furthermore, the SOH is calculated as follows:

[0048]

[0049] Among them, C measure The measured capacity is obtained from the capacity characteristics of battery data when calculating historical SOH; when calculating estimated SOH, it is calculated based on the deeper characteristics of the estimated SOH. C initial This represents the initial capacity, obtained from the capacity characteristics of the battery data.

[0050] Furthermore, the RUL is calculated as follows:

[0051] RUL=c EOL -c start ;

[0052] Among them, c EOL This represents the number of battery cycles at the end of its lifespan, calculated based on deep-level characteristics of the RUL (Round-Under-Limit). start This indicates the battery cycle time at the start of the prediction, obtained from the temporal characteristics of the battery data.

[0053] Compared with the prior art, the present invention has at least the following beneficial effects:

[0054] (1) The lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM disclosed in this embodiment of the invention is based on data-driven and uses a multi-channel convolutional neural network to extract multi-dimensional features, which can extract multi-dimensional features of battery parameters and has high prediction accuracy for lithium battery SOH and RUL.

[0055] (2) The embodiment of the present invention discloses a method for predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM. It proposes a novel nonlinear enhanced long short-term memory network for extracting features of battery sequence data, deeply explores the nonlinear features of battery parameters, and further improves the prediction accuracy of SOH and RUL of lithium batteries. The data features are fully extracted, and the amount of electrochemical data required is reduced.

[0056] (3) The embodiment of the present invention discloses a lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM, which uses the SOH evaluation result as the input for RUL prediction and uses a neural network method to predict RUL, thereby further improving the prediction accuracy of RUL of lithium battery.

[0057] (4) The lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM disclosed in the embodiments of the present invention adopts the same model and shares the input. That is, when predicting in the usage stage, a nonlinear enhanced LSTM model pre-trained for SOH and RUL is used to extract the deep features used to obtain the estimation of SOH and the deep features used to obtain RUL in turn, which greatly saves computing resources. Attached Figure Description

[0058] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0059] Figure 1 This is a flowchart illustrating the overall steps of the lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM disclosed in this embodiment of the invention.

[0060] Figure 2 This is a flowchart of the algorithm for the collaborative prediction method of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM disclosed in an embodiment of the present invention.

[0061] Figure 3 This is an architecture diagram of the nonlinear enhanced LSTM model disclosed in an embodiment of the present invention;

[0062] Figure 4 This is a flowchart of the model training process for the lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM disclosed in an embodiment of the present invention. Detailed Implementation

[0063] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can 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.

[0064] Example 1

[0065] A specific embodiment of the present invention, such as Figure 1 As shown, a collaborative prediction method for SOH and RUL of lithium batteries based on nonlinear enhanced LSTM is disclosed, including the following steps:

[0066] Step S1. Obtain battery data and select battery characteristic parameters from it.

[0067] Step S2. Based on the battery feature parameters, obtain the feature vector used to estimate SOH, and input it into a multi-channel convolutional neural network to extract features and obtain the final feature vector used to estimate SOH.

[0068] Step S3. Input the final feature vector used to obtain the estimated SOH into the pre-trained nonlinear augmented LSTM model, and after processing, obtain the deep features used to obtain the estimated SOH.

[0069] Step S4. Based on battery data, obtain the initial capacity; based on the deep features used to obtain the estimated SOH, calculate the measured capacity, and then combine it with the initial capacity to calculate the estimated SOH.

[0070] Step S5. Obtain historical SOH from battery data, and combine it with estimated SOH to obtain the feature vector used to obtain RUL. Input the vector into a multi-channel convolutional neural network to extract features and obtain the final feature vector used to obtain RUL.

[0071] Step S6. Input the final feature vector used to obtain RUL into the pre-trained nonlinear augmented LSTM model, and process it to obtain the deep features used to obtain RUL.

[0072] Step S7. Based on battery data, obtain the battery cycle period at the start of the prediction; based on the deep features used to obtain RUL, calculate the battery cycle period at the end of the lifespan, and then combine it with the battery cycle period at the start of the prediction to calculate RUL.

[0073] Compared with existing technologies, the lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM disclosed in this invention is data-driven and uses a multi-channel convolutional neural network for multi-dimensional feature extraction, enabling the extraction of multi-dimensional features of battery parameters and achieving high prediction accuracy for SOH and RUL of lithium batteries. It proposes a novel nonlinear enhanced long short-term memory network for feature extraction of battery sequence data, deeply mining the nonlinear features of battery parameters and further improving the prediction accuracy for SOH and RUL of lithium batteries. It fully extracts data features, reducing the amount of electrochemical data required. Using the SOH evaluation result as input for RUL prediction, a neural network method is employed for RUL prediction, further improving the prediction accuracy for RUL of lithium batteries. It uses the same model with shared input; that is, during the prediction phase, a pre-trained nonlinear enhanced LSTM model is used to extract deep-level features for estimating SOH and for obtaining RUL, significantly saving computational resources.

[0074] Example 2

[0075] The algorithm flow of the lithium battery SOH and RUL co-prediction method based on nonlinear enhanced LSTM in this embodiment is as follows: Figure 2 As shown.

[0076] Based on Example 1, step S1 can be further refined into the following sub-steps:

[0077] Step S101. Obtain battery data, including battery health characteristics and capacity characteristics; wherein,

[0078] Health characteristics refer to voltage characteristics, current characteristics, temperature characteristics, and time characteristics; time characteristics include the battery's cycle life at the start of prediction, and capacity characteristics include the battery's initial capacity and the measured capacity before the start of prediction.

[0079] Step S102. Use Pearson correlation coefficient and Spearman correlation coefficient to calculate the correlation of battery data. Based on the correlation calculation results, select data that are highly correlated with capacity characteristics from voltage characteristics, current characteristics, temperature characteristics and time characteristics, and use them together with capacity characteristics as battery characteristic parameters.

[0080] Specifically, the Pearson correlation coefficient is calculated as follows:

[0081]

[0082] Where X represents voltage characteristics, current characteristics, temperature characteristics, and time characteristics, and Y represents capacity samples.

[0083] The Pearson correlation coefficient indicates the degree of correlation between X and Y. The closer it is to 1, the stronger the correlation. Battery feature parameters with higher correlation are selected as elements for constructing the battery parameter feature matrix.

[0084] The Spearman correlation coefficient is calculated as follows:

[0085]

[0086] Where, x i Representing voltage characteristics, current characteristics, temperature characteristics, and time characteristics, y i Represents capacity characteristics; These represent the mean values ​​of voltage characteristics, current characteristics, temperature characteristics, and time characteristics. This represents the mean of the capacity characteristics.

[0087] Based on the correlation calculation results of battery data using Pearson and Spearman correlation coefficients, battery characteristic parameters are selected from the battery data; among them,

[0088] Battery characteristic parameters refer to the voltage, current, temperature, time, and capacity characteristics of a lithium battery during the discharge process.

[0089] It is worth noting that, on the one hand, theoretically, features with a correlation calculation result of 0.8 or higher will be selected; on the other hand, all health features are correlated with capacity features, but the importance of each health feature is different. According to the experimental results, the best prediction effect is achieved by using all correlated health features to construct the battery parameter feature matrix.

[0090] Step S2 can be further broken down into the following sub-steps:

[0091] Step S201. Using the battery characteristic parameters as elements of the matrix, construct the battery parameter characteristic matrix, and obtain the eigenvector I used to estimate SOH from it.

[0092] Specifically, the battery parameter feature matrix is ​​represented as:

[0093]

[0094] in, This represents the characteristic of the i-th battery cell in the first cycle at time t; This represents the characteristic of the i-th battery cell at time t during the m-th cycle; w represents the sliding window length; c i1 c represents the capacity of the i-th battery cell in the first cycle; im This represents the capacity of the i-th battery cell in the m-th cycle.

[0095] Based on the battery parameter feature matrix, the eigenvector I used to estimate SOH is obtained; where I satisfies W I C represents the data width used to obtain the eigenvector I for estimating SOH. I This represents the channel used to obtain the eigenvector I for estimating SOH.

[0096] Step S202. Input the feature vector I used to obtain the estimated SOH into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, where i = 1, 2, 3; after convolution processing, obtain the feature map F of the i-th channel used to obtain the estimated SOH. i n , i = 1, 2, 3.

[0097] Specifically, a multi-channel convolutional neural network has three channels, each containing a one-dimensional convolutional layer and a convolutional kernel with different settings. The i-th channel is utilized as follows: convolution kernel 3. Perform convolution on the feature vector I used to obtain the estimated SOH to obtain the feature map F of the i-th channel used to obtain the estimated SOH. i n :

[0098]

[0099] Where σ(·) represents the sigmoid activation function.

[0100] Step S203. For each channel used to obtain the estimated SOH, perform characteristic plot F. i n For i = 1, 2, 3, an attention mechanism is used to process the data to obtain the feature map F used to estimate the SOH. n .

[0101] Specifically, the feature map F used to obtain the estimated SOH is processed in the following manner. i n For i = 1, 2, 3, an attention mechanism is used to process the data to obtain the desired result. The feature map F used to obtain the estimated SOHn i = 1, 2, 3, W Fn The characteristic map F used to obtain the estimated SOH is shown. n Data width, C Fn The characteristic map F used to obtain the estimated SOH is shown. n The passage:

[0102]

[0103] Step S204. The characteristic map F used to obtain the estimated SOH is processed in the following manner. n Max pooling is performed to obtain the final eigenvector I used to estimate SOH. final And output:

[0104] I final =max{F n}

[0105] The SOH calculation method for the historical SOH dataset used in steps S4 and S5, as well as in the pre-training of the nonlinear augmented LSTM model, is as follows:

[0106]

[0107] Among them, C measure The measured capacity is obtained from the capacity characteristics of battery data when calculating historical SOH; when calculating estimated SOH, it is calculated based on the deeper characteristics of the estimated SOH. C initial This represents the initial capacity, obtained from the capacity characteristics of the battery data.

[0108] Step S5 can be further broken down into the following sub-steps:

[0109] Step S501. Obtain historical SOH from battery data, and then combine it with estimated SOH to construct a feature matrix for obtaining RUL, and obtain the feature vector Y for obtaining RUL from it.

[0110] Specifically, the feature matrix used to obtain RUL is represented as:

[0111]

[0112] in, This represents the SOH value of the i-th cell at the w-th prediction step. The table shows the SOH value of the i-th cell at the (t+1)-th prediction step. The estimated SOH and historical SOH have a temporal relationship. The SOH value at time t is the estimated value, and the SOH values ​​before time t are historical values. There is no limit to the ratio of estimated SOH to historical SOH in terms of data. The SOH value at the closest time is the estimated value.

[0113] Based on the feature matrix used to obtain the RUL, the feature vector Y used to obtain the RUL is obtained; where Y satisfies W Y C represents the data width used to obtain the feature vector Y of RUL. Y This represents the channel used to obtain the feature vector Y of RUL.

[0114] Step S502. Input the feature vector Y used to obtain RUL into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, i = 1, 2, 3; after convolution processing, obtain the feature map of the i-th channel used to obtain RUL.

[0115] Specifically, the convolution kernel of the i-th channel is used in the following manner. Convolution is performed on the feature vector Y used to obtain the RUL to obtain the feature map of the i-th channel used to obtain the RUL.

[0116]

[0117] Step S503. Develop feature maps for each channel used to obtain RUL. An attention mechanism is used to process the data, resulting in the feature map F' used to obtain the RUL. n .

[0118] Specifically, the feature map used to obtain RUL is processed in the following manner. The attention mechanism is used to process the data and achieve the desired result. The feature map F' used to obtain RUL n i = 1, 2, 3, W F’n This represents the feature map F' used to obtain the RUL. n Data width, C F’n This represents the feature map F' used to obtain the RUL. n The passage:

[0119]

[0120] Step S504. Process the feature map F' used to obtain RUL in the following manner. n Perform max pooling to obtain the final feature vector Y used to obtain RUL. final And output:

[0121] Y final =max{F n '}.

[0122] The calculation method for RUL in step S7 is as follows:

[0123] RUL=c EOL -c start ;

[0124] Among them, c EOL This represents the number of battery cycles at the end of its lifespan, calculated based on deep-level characteristics of the RUL (Round-Under-Limit). start This indicates the battery cycle time at the start of the prediction, obtained from the temporal characteristics of the battery data.

[0125] The pre-trained nonlinear augmented LSTM model in steps S3 and S6 refers to the nonlinear augmented LSTM model pre-trained for SOH and RUL, such as... Figure 3 As shown, it includes a forget gate, an input gate, an output gate, a hidden layer, memory units, and a final output; among which,

[0126] The calculation method for the forget gate is f. t =σ(tanh(c) t-1 )*P f +W xf I t +W hf h t-1 +b f ); where σ(·) is the sigmoid function, tanh(·) is the tanh function, t represents time, and c t-1 P represents the cell unit at the previous time step. f W represents the weight of the peephole in the forgetting gate. xf I represents the input vector weights of the forget gate. t W represents the input vector. hf h represents the hidden layer weights at the previous time step before the forget gate. t-1 b represents the output of the hidden layer at the previous time step. f Indicates the forget gate bias, f t This indicates the output of the forget gate.

[0127] The calculation method for the input gate is as follows: Among them, P i Indicates the input gate peephole weight, i t This indicates the input gate output.

[0128] The calculation method for the output gate is as follows: Among them, P o W represents the output gate peephole weight. xo W represents the weights of the input vector to the output gate. ho b represents the weight of the hidden layer above the output gate. o Indicates the output gate bias, o t This indicates the output of the output gate.

[0129] The hidden layer is calculated as follows: Among them, c t This represents the cell unit at time t.

[0130] The calculation method for memory units is as follows: in, W xx W represents the weights of the cell unit input vector. hc b represents the hidden layer weights of the cell unit at the previous time step. c This indicates cell unit bias.

[0131] The final output is calculated as y = W. full h t Among them, W full y represents the hidden layer weights at time t, and y represents the deep features obtained in step S3 for estimating SOH or the deep features obtained in step S6 for obtaining RUL, which are the final outputs of the pre-trained nonlinear augmented LSTM model.

[0132] It is worth noting that the pre-training of the nonlinear augmented LSTM model includes, sequentially, a model training phase for SOH and a model training phase for RUL, such as... Figure 4 As shown; where,

[0133] The model training phase for SOH includes the following steps:

[0134] Obtain the battery dataset and construct the historical SOH dataset based on the battery dataset.

[0135] Perform steps S1 and S2 on the battery dataset to obtain the final feature vector dataset used to estimate SOH.

[0136] The dataset used to obtain the final feature vector of the estimated SOH is divided into SOH training set I. train SOH Validation Set I valid and SOH test set I test .

[0137] SOH training set I train The model is trained using a nonlinear augmented LSTM model and the SOH validation set I is employed. valid To evaluate the performance of the nonlinear augmented LSTM model, the model hyperparameters for SOH are determined by calculating the model loss, resulting in a nonlinear augmented LSTM model trained for SOH. The model loss is calculated using MSE and RMSE. The model hyperparameters refer to the weights, biases, number of neural network layers, number of iterations, and learning rate in the nonlinear augmented LSTM model.

[0138] SOH test set I testInput a nonlinear augmented LSTM model trained for SOH, perform nonlinear augmented LSTM model prediction, calculate the deep feature dataset used to estimate SOH, and complete the model training for SOH.

[0139] The model training phase for RUL includes the following steps:

[0140] Based on the deep feature dataset used to obtain the estimated SOH, the SOH calculation is performed on it to obtain the estimated SOH dataset.

[0141] Perform step S5 on the historical SOH dataset and the estimated SOH dataset to obtain the final feature vector dataset used to obtain RUL.

[0142] The input dataset used to obtain the final feature vector of RUL is divided into the RUL training set Y. train RUL verification set Y valid and RUL test set Y test .

[0143] RUL training set Y train The model is trained using a pre-trained nonlinear augmented LSTM model targeting SOH, and the RUL validation set Y is used. valid The model performance is evaluated by calculating the model loss to determine the model hyperparameters for RUL, resulting in a nonlinear augmented LSTM model trained for SOH and RUL; the model loss is calculated using MSE and RMSE.

[0144] Input RUL test set Y test We perform nonlinear enhanced LSTM model prediction, calculate the deep feature dataset used to obtain RUL, and complete the model training for RUL.

[0145] Compared with Example 1, this example further discloses specific methods for calculating Pearson and Spearman correlation coefficients, methods for processing feature vectors using multi-channel convolutional neural networks, the training process of the nonlinear enhanced LSTM model and the working mode of each layer, and optimizes the model hyperparameters during the training process to obtain a pre-trained nonlinear enhanced LSTM model for SOH and RUL. The accuracy of the model is verified using test sample data, proving the feasibility and accuracy of this method for predicting the health status and remaining service life of lithium batteries.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for co-predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM, characterized in that, Includes the following steps: Step S1. Obtain battery data and select battery characteristic parameters from it; Step S2. Based on the battery feature parameters, obtain the feature vector for estimating SOH, and input it into a multi-channel convolutional neural network to extract features and obtain the final feature vector for estimating SOH; Step S3. Input the final feature vector used to obtain the estimated SOH into the pre-trained nonlinear augmented LSTM model, and process it to obtain the deep features used to obtain the estimated SOH; Step S4. Based on the battery data, obtain the initial capacity; Based on the aforementioned in-depth features used to obtain the estimated SOH, the measured capacity is calculated, and then combined with the initial capacity, the estimated SOH is calculated. Step S5. Obtain the historical SOH from the battery data, and combine it with the estimated SOH to obtain the feature vector for obtaining RUL. Input the feature vector into a multi-channel convolutional neural network to extract features and obtain the final feature vector for obtaining RUL. Step S6. Input the final feature vector used to obtain RUL into the pre-trained nonlinear augmented LSTM model, and process it to obtain the deep features used to obtain RUL; Step S7. Based on the battery data, obtain the battery cycle number at the start of the prediction; Based on the deep features used to obtain RUL, the battery cycle count at the end of its lifespan is calculated, and then combined with the battery cycle count at the start of the prediction, the RUL is calculated. The pre-trained nonlinear augmented LSTM model in steps S3 and S6 refers to a pre-trained nonlinear augmented LSTM model for SOH and RUL, including a forget gate, input gate, output gate, hidden layer, memory unit, and final output; wherein, The calculation method for the forget gate is f. t =σ(tanh(c) t-1 )*P f +W xf I t +W hf h t-1 +b f ); where σ(·) is the sigmoid function, tanh(·) is the tanh function, t represents time, and c t-1 P represents the cell unit at the previous time step. f W represents the weight of the peephole in the forgetting gate. xf I represents the input vector weights of the forget gate. t W represents the input vector. hf h represents the hidden layer weights at the previous time step before the forget gate. t-1 b represents the output of the hidden layer at the previous time step. f Indicates the forget gate bias, f t Indicates the output of the forget gate; The calculation method for the input gate is as follows: Among them, P i Indicates the input gate peephole weight, i t Indicates the input gate output; The output gate is calculated as o t =σ(tanh(c) t )*P o +W xo I t +W ho h t-1 +b o ); where P o W represents the output gate peephole weight. xo W represents the weights of the input vector to the output gate. ho b represents the weight of the hidden layer above the output gate. o Indicates the output gate bias, o t Indicates the output of the output gate; The hidden layer is calculated as follows: Among them, c t This represents the cell unit at time t; The calculation method for memory units is as follows: in, W xx W represents the weights of the cell unit input vector. hc b represents the hidden layer weights of the cell unit at the previous time step. c Indicates cell unit bias; The final output is calculated as y = W. full h t Among them, W full y represents the hidden layer weights at time t, and y represents the deep features obtained in step S3 for estimating SOH or the deep features obtained in step S6 for obtaining RUL, which are the final outputs of the pre-trained nonlinear augmented LSTM model.

2. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 1, characterized in that, Step S1 includes the following specific steps: Step S101. Obtain battery data, including battery voltage characteristics, current characteristics, temperature characteristics, time characteristics, and capacity characteristics; Step S102. Use Pearson correlation coefficient and Spearman correlation coefficient to perform correlation calculation on battery data. Based on the correlation calculation results, select data from voltage characteristics, current characteristics, temperature characteristics and time characteristics, and use them together with capacity characteristics as battery characteristic parameters.

3. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 2, characterized in that, Step S2 includes the following specific steps: Step S201. Using the battery characteristic parameters as elements of the matrix, construct the battery parameter characteristic matrix, and obtain the eigenvector I used to estimate SOH from it; Step S202. Input the feature vector I used to obtain the estimated SOH into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, where i = 1, 2, 3; after convolution processing, obtain the feature map F of the i-th channel used to obtain the estimated SOH. i n i = 1, 2, 3; Step S203. For each channel used to obtain the estimated SOH, perform characteristic plot F. i n For i = 1, 2, 3, an attention mechanism is used to process the data to obtain the feature map F used to estimate the SOH. n ; Step S204. For the characteristic map F used to obtain the estimated SOH n Max pooling is performed to obtain the final eigenvector I used to estimate SOH. final And output it.

4. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 3, characterized in that, Step S5 includes the following specific steps: Step S501. Obtain the historical SOH from the battery data, and then combine it with the estimated SOH to construct a feature matrix for obtaining RUL, and obtain the feature vector Y for obtaining RUL from it; Step S502. Input the feature vector Y used to obtain RUL into the one-dimensional convolutional layer of the i-th channel of the multi-channel convolutional neural network, i = 1, 2, 3; after convolution processing, obtain the feature map of the i-th channel used to obtain RUL. Step S503. Develop feature maps for each channel used to obtain RUL. An attention mechanism is used to process the data, resulting in the feature map F' used to obtain the RUL. n ; Step S504. For the feature map F' used to obtain RUL n Perform max pooling to obtain the final feature vector Y used to obtain RUL. final And output it.

5. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 4, characterized in that, The pre-trained nonlinear augmented LSTM model includes a model training phase for SOH and a model training phase for RUL; wherein the model training phase for SOH includes the following steps: Obtain the battery dataset and construct the historical SOH dataset based on the battery dataset; Perform steps S1 and S2 on the battery dataset to obtain the final feature vector dataset used to estimate SOH; The dataset used to obtain the final feature vector of the estimated SOH is divided into SOH training set I. train SOH Validation Set I valid and SOH test set I test ; SOH training set I train The model is trained using a nonlinear augmented LSTM model and the SOH validation set I is employed. valid The performance of the nonlinear augmented LSTM model is evaluated by calculating the model loss to determine the model hyperparameters for SOH, thus obtaining the nonlinear augmented LSTM model trained for SOH; the model loss is calculated using MSE and RMSE. SOH test set I test Input a nonlinear augmented LSTM model trained for SOH, perform nonlinear augmented LSTM model prediction, calculate the deep feature dataset used to estimate SOH, and complete the model training for SOH.

6. The method for co-predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 5, characterized in that, The model training phase for RUL includes the following steps: Based on the deep feature dataset used to obtain the estimated SOH, the SOH is calculated and processed to obtain the estimated SOH dataset; Perform step S5 on the historical SOH dataset and the estimated SOH dataset to obtain the final feature vector dataset used to obtain RUL; The input dataset used to obtain the final feature vector of RUL is divided into the RUL training set Y. train RUL verification set Y valid and RUL test set Y test ; RUL training set Y train The model is trained using a pre-trained nonlinear augmented LSTM model targeting SOH, and the RUL validation set Y is used. valid The model performance was evaluated by calculating the model loss to determine the model hyperparameters for RUL, resulting in a nonlinear augmented LSTM model trained for SOH and RUL; the model loss was calculated using MSE and RMSE. Input RUL test set Y test We perform nonlinear enhanced LSTM model prediction, calculate the deep feature dataset used to obtain RUL, and complete the model training for RUL.

7. The method for co-predicting SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 6, characterized in that, The model hyperparameters include the weights, biases, number of neural network layers, number of iterations, and learning rate in the nonlinear augmented LSTM model.

8. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 7, characterized in that, The SOH is calculated as follows: Among them, C measure The measured capacity is obtained from the capacity characteristics of battery data when calculating historical SOH; when calculating estimated SOH, it is calculated based on the deeper characteristics of the estimated SOH. C initial This represents the initial capacity, obtained from the capacity characteristics of the battery data.

9. The method for co-prediction of SOH and RUL of lithium batteries based on nonlinear enhanced LSTM according to claim 8, characterized in that, The RUL is calculated as follows: RUL=c EOL -c start ; Among them, c EOL This represents the number of battery cycles at the end of its lifespan, calculated based on deep-level characteristics of the RUL (Round-Under-Limit). start This indicates the battery cycle time at the start of the prediction, obtained from the temporal characteristics of the battery data.

Images