A Method for Predicting the Remaining Useful Life of Lithium-Ion Batteries Based on SAE-CEEMDAN-LSTM

Abstract

The present invention discloses a method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM. It belongs to the technical field of lithium-ion battery capacity detection. The specific steps are as follows: The discharge power P of the lithium-ion battery, the constant current charging time T c and the voltage V at the battery terminal during the constant current charging stage are used as the HI for predicting the remaining service life of the lithium-ion battery. An SAE is used to construct a fused HI. This method generates a high-order abstract complex function through self-learning and adaptively transforms the complex multi-dimensional HI into a fused HI that can centrally express the characteristics of the remaining capacity of the battery. CEEMDAN is used to perform scale decomposition on the fused HI to obtain multiple groups of components, and through correlation analysis, several groups of components with strong correlation are selected to achieve good generalization for different data. The trained LSTM model is used to predict the RUL of the lithium-ion battery for the selected groups of components with strong correlation, and finally the prediction results of several groups are accumulated to achieve accurate prediction of the RUL of the lithium-ion battery.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery capacity detection, and in particular to a method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM. Background Art

[0002] Lithium batteries have characteristics such as high energy density, low self-discharge rate, long cycle life, wide operating temperature range, and no pollution, and are widely used in intelligent manufacturing fields such as computing engineering, logistics, aerospace, electric vehicles, and electronic devices. The degradation of lithium batteries will cause battery failure, resulting in shortened battery life and even serious accidents. Therefore, a high accuracy rate in predicting the remaining useful life of lithium-ion batteries is one of the important measures for protecting lithium batteries and has important significance. RUL (Remaining Useful Life) generally refers to the remaining effective working time, and is usually defined as: under real-time working conditions, the number of cycle periods experienced when the real-time capacity of the battery decays to 70%-80% of the rated capacity. It is generally considered that when the battery capacity drops to 70%-80% of the rated capacity, the battery reaches the end of its service life. Therefore, an accurate and convenient RUL prediction method is beneficial to making a correct judgment on the degradation degree of lithium batteries and timely preventing the harm caused by the life attenuation of lithium batteries to actual production and life.

[0003] Currently, through a large number of studies, it has been found that parameters such as the discharge power of lithium-ion batteries, the constant current charging time, the voltage at the battery terminal during the constant current charging stage, and the average temperature during discharge have a large correlation with the battery capacity and can be used as HI (Health Indicator) for RUL prediction. HI solves the difficulty of obtaining the capacity online, but there are still the following deficiencies: 1. For some lithium-ion battery HIs, due to their low correlation with the battery capacity, it is easy to result in uneven expression capabilities for battery capacity attenuation. 2. A single HI cannot effectively reflect the local trend during capacity attenuation and lacks the ability to analyze the change characteristics of HI at multiple scales.

[0004] In summary, in the field of lithium-ion battery RUL prediction, existing methods have problems such as low prediction accuracy, poor generalization, and weak robustness for lithium-ion battery RUL prediction due to the lack of multi-scale analysis of HI and the characteristics of non-linearity, multi-modal, and multi-noise. Summary of the Invention

[0005] In view of the problems existing in the above technical background, the present invention provides a method for predicting the remaining useful life (RUL) of lithium-ion batteries based on SAE-CEEMDAN-LSTM. After being extracted and fused by SAE (Stacked Auto Encoder), there is a very high correlation between the fused HI and the capacity. Therefore, the fused HI can represent the RUL of the lithium-ion battery, and then indirectly predict the RUL, overcoming the problem of low quality of the existing HI. At the same time, CEEMDAN (Complete EEMD with Adaptive Noise) and correlation analysis are used to perform multi-scale analysis on HI, effectively overcoming the problems of incomplete EEMD decomposition and large reconstruction error, and at the same time ensuring the strong correlation between HI and capacity.

[0006] The technical solution adopted by the present invention includes the following steps:

[0007] Step 1: Since the discharge power of the lithium-ion battery, the constant-current charging time, and the voltage at the battery terminal during the constant-current charging stage have a strong correlation with the battery capacity, they can be used as HI for predicting the RUL of the lithium-ion battery. Use SAE to construct the fused HI. This method generates a high-order abstract complex function through self-learning and adaptively transforms the complex multi-dimensional HI into a fused HI that can centrally express the characteristics of the remaining battery capacity.

[0008] Step 2: Use CEEMDAN to perform multi-scale decomposition on the fused HI to obtain multiple groups of components, and through correlation analysis, screen out several groups of components with strong correlation, with the goal of having good generalization for different data.

[0009] Step 3: Use the trained long short-term memory network model to predict the RUL of the lithium-ion battery for several groups of components with strong correlation that are screened out, and finally add up the prediction results of several groups to achieve accurate prediction of the RUL of the lithium-ion battery.

[0010] Step 1 is specifically implemented according to the following steps:

[0011] Step 1.1: The experimental data comes from the public dataset provided by NASA's PCoE. Select a set of charge and discharge data of 18650-type lithium-ion batteries (B5, B6, B7) collected in the same experimental environment, and the rated capacity of the battery is 2 A·h.

[0012] Obtain the discharge power of the lithium-ion battery, the constant-current charging time, and the voltage at the battery terminal during the constant-current charging stage through the public dataset, and calculate the Spearman correlation coefficient between these three lithium-ion battery HIs and the capacity.

[0013] By calculation, it is found that the correlation coefficients between the battery discharge power, the constant-current charging time, and the voltage and capacity at the battery terminal during the constant-current charging stage are relatively high, which can be used as HI to indirectly reflect the RUL of lithium-ion batteries.

[0014] Step 1.2: Use the battery discharge power, the constant-current charging time, and the voltage at the battery terminal during the constant-current charging stage as HI.

[0015] SAE is a deep learning network composed of a certain number of autoencoders.

[0016] Step 1.3: The fusion process using SAE is as follows:

[0017] Apply the indirect HI as the input layer data of the SAE network structure, train the first AE, and then initialize the weights and biases of the AE.

[0018] Parse the indirect HI through the encoder of the first AE and use it as the output of the hidden layer. The specific mathematical process is as follows: H = f(w (1) x + b (1) ) (1)

[0019] In formula (1), w 1 is the weight from the input layer to the hidden layer neurons in the currently trained AE structure; x = [x1,..., x k corresponds to the indirect HI and serves as the input data of the AE; b 1 is the bias from the hidden layer to the output layer neurons in the AE structure, and H is the output of the hidden layer.

[0020] Use H = f(w (1) x + b (1) ) as the input of the decoder to perform data reconstruction to obtain the reconstructed data

[0021] In formula (2), is the output data of the AE; w (2) is the weight from the hidden layer to the output layer neurons in the AE structure, and b (2) is the bias from the hidden layer to the output layer neurons in the AE structure.

[0022] The original input data and the reconstructed data will form a reconstruction error. The expression of the reconstruction error is as follows:

[0023] Aim at minimizing the reconstruction error and apply the error backpropagation method to continuously adjust the weights and biases of the current AE to complete the training of the first layer of AE.

[0024] Only retain the AE encoding part completed in this training. Use the data H obtained from the hidden layer as the input layer data of the next AE and continue to train the next AE in the same way until the training of the next AE is finally completed.

[0025] Repeat step 1.3 until all AEs in the SAE are trained.

[0026] In the present invention, the SAE consists of two AEs. The output of the hidden layer of the first AE is used as the result after the first layer of fusion. Immediately, the result after the first layer of fusion is used as the input data of the input layer of the second AE, and the number of hidden layer nodes in the second layer is set to 1. Finally, the output is a sequence, that is, the corresponding fused HI.

[0027] Extract a fused HI through the SAE, and verify that the correlation coefficient between the fused HI and the capacity is better than that of the single HI.

[0028] Step 2.1: The HI after being fused by the SAE is H(t). Now, decompose H(t) into k IMFs (Intrinsic Mode Functions), and each IMF is represented by IMF k For presentation.

[0029] Define the operator E k (*) as the k-th modal component generated by the EMD decomposition of H(t).

[0030] Step 2.2: Add Gaussian white noise n k (t) that satisfies the standard normal distribution to H(t). Its expression is as follows: H k (t) = H(t) + δ i n k (t) (3)

[0031] In formula (3), t is the battery cycle period, H k (t) is the original signal with white noise added for the k-th time, and δ i is the signal-to-noise ratio between the i-th noise and the original signal.

[0032] Step 2.3: Use the EMD algorithm to decompose the H k (t) signal N times repeatedly. Calculate the first modal component IMF1(t) through the mean value, and then obtain the first residual signal R1(t). Its expression is as follows: R1(t) = C(t) - IMF1(t)

[0033] Step 2.4: Obtain the second modal component and the residual. For the signal R1(t) + δ1E1(n kPerform N repeated decompositions on (t) and calculate the mean value to obtain the expressions of the second mode component IMF2(t) and the second residual signal R2(t) as follows: R2(t) = R1(t) - IMF2(t)

[0034] Step 2.5: For the calculation methods of subsequent mode components and residual signals, by analogy with Step 2.4, the (k + 1)-th mode component IMF k+1 (t) and the (k + 1)-th residual signal R k+1 (t) can be obtained.

[0035] Repeat Step 2.5 until the obtained residual signal can no longer be decomposed. The final signal is decomposed into

[0036] Perform correlation analysis on the obtained several groups of mode components, and set the correlation threshold CT = max(IMF k ) / K, which is used to select the components with a correlation coefficient greater than CT to obtain several groups of IMF(t) with relatively large correlations.

[0037] Step 3.1: Divide the several groups of IMF(t) and a group of R(t) signals after decomposition into a training set and a test set in a certain proportion, and substitute them into the trained LSTM model respectively.

[0038] Step 3.2: Accumulate the prediction results of the LSTM model to obtain the final prediction result of the remaining life of the lithium-ion battery.

[0039] The advantages and beneficial effects of the present invention are as follows:

[0040] 1. It overcomes the problem of low quality of the existing indirect HI, and the fused HI can well represent the battery capacity for indirect prediction.

[0041] 2. Use CEEMDAN to perform multi-scale analysis on the fused HI, and solve the problems of incomplete EEMD decomposition and large reconstruction error.

[0042] 3. Through screening by the correlation threshold, combine the mode components IMF(t) with strong correlations and the residual signal R(t). While retaining local features, it avoids the problem of large cumulative errors caused by a large number of groups. Description of the Drawings

[0043] Figure 1 It is a schematic flowchart of a method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM disclosed by the present invention.

[0044] Figure 2 It is a schematic structural diagram of the encoder AE.

[0045] Figure 3It is a schematic diagram of the stacked autoencoder (SAE) structure.

[0046] Figure 4 It is a schematic diagram of the LSTM network structure.

[0047] Figure 5 It is the constant current charging time T c The curve diagram showing the change with the battery cycle number.

[0048] Figure 6 It is the curve diagram of HI after SAE fusion and the battery capacity changing with the cycle number.

[0049] Figure 7 It is the RUL prediction diagram of the lithium-ion battery based on SAE-CEEMDAN-LSTM (taking B5 as an example) Specific implementation manners

[0050] In order to more clearly express the implementation purpose, implementation scheme and advantages of the present invention, the present invention will be further explained below in conjunction with the drawings. It should be noted that the described implementation process is a part of the overall implementation process of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0051] See Figure 1 , Figure 1 The method for predicting the remaining service life (RUL) of a lithium-ion battery based on SAE-CEEMDAN-LSTM in the embodiment of the present invention includes the following steps:

[0052] Step 1. Since the discharge power P of the lithium-ion battery, the constant current charging time T c and the voltage V at the battery terminal during the constant current charging stage have a significant linear correlation with the battery capacity, they can be used as indirect HIs for predicting the RUL of the lithium-ion battery. An SAE is used to construct a fused HI. This method generates a high-order abstract complex function through self-learning and adaptively converts the complex multi-dimensional HI into a fused HI that can centrally express the characteristics of the remaining battery capacity.

[0053] Step 1 is specifically implemented according to the following steps:

[0054] Step 1.1. The experimental data is from the public dataset provided by NASA's PCoE. Select a set of charge and discharge data of 18650-type lithium-ion batteries (B5, B6, B7, B18) collected under the same experimental environment, and the rated capacity of the battery is 2 A·h.

[0055] The discharge power P of the lithium-ion battery and the constant current charging time T are obtained through the public dataset cAnd the voltage V at the battery terminal during the constant current charging stage, and calculate the Spearman correlation coefficient between these three lithium-ion batteries HI and the capacity. As shown in Table 1:

[0056] Table 1 Analysis table of the correlation between health factors and capacity (taking B5 battery as an example)

[0057]

[0058] It is calculated that there is a strong correlation between the battery discharge power, the constant current charging time, and the voltage at the battery terminal during the constant current charging stage and the capacity, which can be used as HI to indirectly reflect the RUL of lithium-ion batteries. Taking the constant current charging time as an example, Figure 5 Give the curve of constant current charging changing with the battery cycle.

[0059] Step 1.2: Take the battery discharge power, the constant current charging time, and the voltage at the battery terminal during the constant current charging stage as HI, and then extract the fused HI through SAE.

[0060] SAE is a deep learning network composed of a large number of autoencoders. Figure 2 The structural schematic diagram of the encoder AE is given, Figure 3 The structural schematic diagram of the stacked encoder SAE is given.

[0061] Step 1.3: The fusion process using SAE is as follows:

[0062] Apply the indirect HI as the input layer data of the SAE network structure, train the first AE, and then initialize the weights and biases of the AE.

[0063] Parse the indirect HI through the encoder of the first AE and use it as the output of the hidden layer. The specific mathematical process is as follows: H = f(w (1) x + b (1) )

[0064] Where w 1 is the weight from the input layer to the hidden layer neurons in the currently trained AE structure; x = [x1,..., x k corresponds to the indirect HI and serves as the input data of the AE; b 1 is the bias from the hidden layer to the output layer neurons in the AE structure, and H is the output of the hidden layer.

[0065] Take H = f(w (1) x + b (1) ) as the input of the decoder to perform data reconstruction to obtain the reconstructed data

[0066] Where is the output data of the AE; w (2)is the weight from the hidden layer to the output layer neurons in the AE structure, b (2) is the bias from the hidden layer to the output layer neurons in the AE structure.

[0067] The original input data and the reconstructed data will form a reconstruction error, and the expression of the reconstruction error is as follows:

[0068] Aiming at the minimum reconstruction error, the error backpropagation method is applied to continuously adjust the weights and biases of the current AE to complete the training of the first-layer AE.

[0069] Only keep the encoding part of the AE completed in this training, and use the data H obtained from the hidden layer as the input layer data of the next AE to continue training the next AE in the same way, and finally complete the training of the next AE.

[0070] Repeat step 1.3 until all AEs in the SAE are trained.

[0071] In the present invention, the SAE consists of two AEs. The output of the hidden layer of the first AE is used as the result of the first-layer fusion, and then the result of the first-layer fusion is used as the input data of the input layer of the second AE, and the number of hidden layer nodes in the second layer is set to 1, and the final output is a sequence, that is, the corresponding fused HI.

[0072] Extract a fused HI through the SAE, and verify that the correlation coefficient between the fused HI and the capacity is better than that of the single HI.

[0073] Table 2 Correlation analysis between the fused HI and the battery capacity

[0074]

[0075] As Figure 6 shown, the curve graph of the fused HI and the battery capacity changing with the cycle is given. It can be visually found from the graph that the HI fused by the SAE can indirectly represent the battery capacity.

[0076] Step 2: Use EEMD to perform scale decomposition on the fused HI to obtain multiple groups of components, and through correlation analysis, screen out several groups of components with strong correlation to achieve good generalization for different data.

[0077] Step 2.1: Denote the HI fused by the SAE as H(t). Now, decompose H(t) as the original signal into k IMFs, and each IMF is represented by IMF k for representation.

[0078] Define the operator E k (*) as the k-th modal component generated by the EMD decomposition of H(t).

[0079] Step 2.2: Add Gaussian white noise \(n(t)\) that satisfies the standard normal distribution to \(H(t)\). Its expression is as follows: \(H(t)=\hat{H}(t)+\delta n(t)\) k (t), where the expression is as follows: \(H\) k (t)=\(H(t)+\delta\) i n k (t)

[0080] In the formula, \(t\) is the battery cycle period, and \(H\) k (t) is the original signal with white noise added for the \(k\)th time, and \(\delta\) i is the signal-to-noise ratio between the \(i\)th noise and the original signal.

[0081] Step 2.3: Use the EMD algorithm to decompose the \(H(t)\) signal \(N\) times repeatedly. Calculate the first mode component IMF1(t) through the mean value, and then obtain the first residue signal R1(t). Its expression is as follows: R1(t) = C(t) - IMF1(t) k (t), and its expression is as follows; R1(t)=C(t)-IMF1(t)

[0082] Step 2.4: Obtain the second mode component and residue. Decompose the signal \(R1(t)+\delta1E1(n\) k (t)) \(N\) times repeatedly and calculate the mean value. The expressions for the second mode component IMF2(t) and the second residue signal R2(t) obtained are as follows:

[0083] Step 2.5: For the calculation methods of subsequent mode components and residue signals, by analogy with Step 2.4, the \((k + 1)\)th mode component IMF k+1 (t) and the \((k + 1)\)th residue signal R k+1 (t) can be obtained.

[0084] Repeat Step 2.5 until the obtained residue signal can no longer be decomposed. The final signal is decomposed into

[0085] Perform correlation analysis on the obtained several groups of mode components. Set the correlation threshold CT = max(IMF k ) / K, which is used to select the components with a correlation coefficient greater than CT, and obtain several groups of IMF(t) with relatively large correlations.

[0086] Step 3: Use the trained LSTM model to predict the RUL of lithium-ion batteries for the several groups of components with strong correlations that are screened out. Finally, add the several groups of prediction results to achieve accurate prediction of the RUL of lithium-ion batteries.

[0087] Figure 4 , and the schematic diagram of the LSTM network structure is given

[0088] Step 3.1: Divide the decomposed several groups of IMF(t) and a group of R(t) signals into a training set and a test set in a certain proportion, and substitute them into the trained LSTM model respectively.

[0089] Step 3.2: Accumulate the prediction results of the LSTM model to obtain the final prediction result of the lithium-ion battery life.

[0090] Figure 7 The RUL prediction diagram of the lithium-ion battery based on SAE-CEEMDAN-LSTM is given (taking B5 as an example).

Claims

1. A method for predicting the remaining useful life of a lithium-ion battery based on SAE-CEEMDAN-LSTM, characterized in that, It includes the following steps: Step 1. Take the discharge power P of the lithium-ion battery, the constant-current charging time T c and the voltage V at the battery terminal during the constant-current charging stage as the HI for predicting the RUL of the lithium-ion battery, and use SAE to construct the fused HI. This method generates a complex function with high-order abstraction through self-learning and adaptively converts the complex multi-dimensional HI into a fused HI that can centrally express the characteristics of the remaining battery capacity; Step 2: Use CEEMDAN to perform scale decomposition on the fused HI to obtain multiple groups of components, and through correlation analysis, screen out several groups of components with strong correlation, aiming to achieve good generalization for different data; Step 3: Use the trained LSTM model to predict the remaining useful life (RUL) of the lithium-ion battery for the several groups of components with strong correlation screened out, and finally accumulate the several groups of prediction results to achieve accurate prediction of the RUL of the lithium-ion battery.

2. The method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM according to claim 1, wherein The specific content of Step 1 includes: Step 1.1: The experimental data comes from the public dataset provided by the NASA PCoE. Select a set of charge and discharge data of 18650-type lithium-ion batteries collected in the same experimental environment, and the rated capacity of the battery is 2 A·h; Obtain the discharge power P and constant-current charging time T of the lithium-ion battery through the publicly available dataset c and the voltage V at the battery terminal during the constant-current charging stage, and calculate the Spearman correlation coefficient between these three lithium-ion battery HIs and the capacity; through calculation, obtain the battery discharge power P and constant-current charging time T c and those with a relatively high correlation coefficient between the voltage V at the battery terminal during the constant-current charging stage and the capacity are used as HIs to indirectly reflect the RUL of the lithium-ion battery Step 1.2, take the battery discharge power P, the constant current charging time T c and the V voltage at the battery terminal during the constant current charging stage as HI; Step 1.

3. The fusion process using SAE is as follows: Taking HI as the input layer data of the SAE network structure, train the first AE, and then initialize the weights and biases of the AE; Parse HI indirectly through the encoder of the first AE and use it as the output of the hidden layer. The specific mathematical process is as follows: H = f(w (1) x + b (1) ) where w (1) is the weight from the input layer to the hidden layer neurons in the currently trained AE structure; x = [x1,..., x k corresponds to HI before fusion and serves as the input data of the AE; b (1) is the bias from the hidden layer to the output layer neurons in the AE structure, and H is the output of the hidden layer; Take H = f(w (1) x + b (1) ) as the input of the decoder, and perform data reconstruction to obtain the reconstructed data where is the output data of the AE; w (2) is the weight from the hidden layer to the output layer neurons in the AE structure, and b (2) is the bias from the hidden layer to the output layer neurons in the AE structure; the original input data and the reconstructed data will form a reconstruction error, and the reconstruction error expression is as follows: Taking the minimum reconstruction error as the goal, apply the error backpropagation method to continuously adjust the weights and biases of the current AE to complete the training of the first layer of AE; only retain the encoding part of the AE completed in this training, and use the data H obtained from the hidden layer as the input layer data of the next AE to continue training the next AE in the same way, and finally complete the training of the next AE; repeat step 1.3 until all AEs in the SAE are trained; The SAE consists of two AEs. The output of the hidden layer of the first AE is used as the result of the first-layer fusion. Then, the result of the first-layer fusion is used as the input data of the input layer of the second AE, and the number of hidden layer nodes in the second layer is set to 1. Finally, the output is a sequence, that is, the corresponding fused HI; A fused HI is extracted through the SAE, and it is verified that the correlation coefficient between the fused HI and the capacity is better than that of the single HI.

3. The method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM according to claim 1, wherein, The specific content of Step 2 includes: Step 2.1: Denote the HI after SAE fusion as H(t). Now, decompose H(t) as the original signal into k IMFs, and each IMF is represented by IMF k ; Define the operator E k (*) as the k-th modal component generated by the EMD decomposition of H(t); Step 2.2: Add Gaussian white noise \(n(t)\) that satisfies the standard normal distribution to \(H(t)\), and its expression is as follows: \(H^{(i)}(t)=H(t)+\delta n(t)\), where \(t\) is the battery cycle period, \(H^{(i)}(t)\) is the original signal with white noise added for the \(i\)-th time, and \(\delta\) is the signal-to-noise ratio between the \(i\)-th noise and the original signal. k (t), and its expression is as follows: \(H\) k (t) = H(t) + \(\delta\) i n k (t), where \(t\) is the battery cycle period, \(H\) k (t) is the original signal with white noise added for the \(k\)-th time, and \(\delta\) i is the signal-to-noise ratio between the \(i\)-th noise and the original signal; Step 2.

3. Perform N - times repeated decomposition on the H k (t) signal using the EMD algorithm. Obtain the first mode component IMF1(t) through mean - value calculation, and then obtain the first residue signal R1(t), whose expression is as follows: R1(t)=H(t) - IMF1(t); Step 2.4, obtain the second mode component and the residue. Repeat the decomposition of the signal R1(t)+δ1E1(n k (t)) N times and calculate the mean value. The expressions for the second mode component IMF2(t) and the second residue signal R2(t) are as follows: R2(t) = R1(t) - IMF2(t); Step 2.5: For the calculation method of subsequent modal components and residue signals, by analogy with Step 2.4, the (k + 1)-th modal component IMF k+1 (t) and the (k + 1)-th residue signal R k+1 (t) are obtained;​​ Repeat step 2.5 until the resulting residual signal can no longer be decomposed, and the final signal is decomposed into Perform a correlation analysis on several groups of modal components obtained, and set the correlation threshold CT = max(IMF k ) / K, which is used to select the components with a correlation coefficient greater than CT, and obtain several groups of IMF(t) with relatively large correlations.

4. The method for predicting the remaining life of a lithium-ion battery based on SAE-CEEMDAN-LSTM according to claim 1, wherein The specific content of Step 3 includes: Step 3.1: Divide the decomposed several groups of IMF(t) and a group of R(t) signals into a training set and a test set in a certain proportion, and substitute them into the trained LSTM model respectively; Step 3.2: Accumulate the prediction results of the LSTM model to obtain the final prediction result of the lithium-ion battery life.

Images