Lithium battery health state estimation method fusing VMD and BiLSTM

Abstract

The invention discloses a lithium battery health state estimation method fusing VMD and BiLSTM. The method comprises the following steps: firstly, taking a minimum envelope entropy as a fitness function, introducing a vector weighted average optimization algorithm to adaptively optimize core parameters of variational mode decomposition, and realizing optimal decomposition of capacity data and quantitative discrimination of decomposition component regularity; then, components which are obtained through decomposition and contain original capacity recession and regeneration information are respectively predicted through a bidirectional long short-term memory (BiLSTM) network; and finally, reconstructing the prediction result to obtain an estimation result of the health state. According to the method, the estimation precision of the health state of the lithium ion battery can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technology for evaluating the health state of lithium batteries, and particularly to a method for estimating the SOH (State of Health) of lithium batteries by integrating VMD (Variational Mode Decomposition) and BiLSTM (Bidirectional Long Short-Term Memory Network), which is applicable to the health state monitoring and prediction in battery management systems. Background Art

[0002] Due to their high energy density, long cycle life and environmental friendliness, lithium batteries are widely used in the fields of consumer electronics, automobiles and national defense. However, during the use of lithium batteries, capacity attenuation and regeneration phenomena will occur, increasing the risk of equipment failure. Therefore, accurately evaluating the health state of lithium batteries has become an important topic for ensuring the reliability of battery management systems.

[0003] Existing methods for estimating the health state of lithium batteries include model-based methods and data-driven methods. The data-driven methods use historical data to establish a mathematical model to extract battery aging characteristics, and among them, the deep learning methods perform excellently in non-linear feature extraction. However, the existing deep learning methods have insufficient accuracy in dealing with the capacity regeneration phenomenon of lithium batteries, and at the same time, there are problems with insufficient decomposition adaptability in variational mode decomposition, making it difficult to effectively extract health state features.

[0004] In summary, lithium batteries are widely used due to their high energy density and long cycle life, but their capacity attenuation and regeneration phenomena increase the risk of failure. Among the existing methods, although the data-driven models can extract non-linear features, there are problems such as insufficient adaptability of decomposition parameters and poor adaptability to the capacity regeneration phenomenon. For this reason, the present invention proposes a method combining optimized variational mode decomposition and bidirectional long short-term memory network to improve the accuracy of health state estimation. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology in the estimation of the health state of lithium batteries, the present invention proposes a method for estimating the health state of lithium batteries based on vector weighted average optimized variational mode decomposition and bidirectional long short-term memory network. The main technical problems to be solved are as follows: how to adaptively optimize the decomposition parameters of variational mode decomposition to improve the ability to extract multi-scale capacity features of lithium batteries; how to combine the bidirectional long short-term memory network to model and predict the capacity data of lithium batteries to improve the adaptability and estimation accuracy to the capacity regeneration phenomenon.

[0006] Technical Solution: A method for estimating the health state of lithium batteries by integrating VMD and BiLSTM networks, characterized in that:

[0007] Step 1) Obtain the lithium battery capacity attenuation data, usually the capacity records of multiple charge and discharge cycles. Normalize the capacity data so that its range is unified to [0,1] to adapt to subsequent decomposition and modeling steps.

[0008] Step 2) Optimize variational mode decomposition (VMD) using the vector weighted average optimization algorithm. Take the capacity data as the input signal and perform multi-scale decomposition on it using the VMD method. Introduce the vector weighted average optimization algorithm to optimize the decomposition parameters of VMD, including the number of decomposition modes and the penalty factor. Use the minimum envelope entropy as the fitness function and find the best decomposition effect by iteratively updating the algorithm parameters.

[0009] Step 3) Bidirectional long short-term memory (BiLSTM) network modeling and prediction. Normalize each intrinsic mode function (IMF) component and then input it into the BiLSTM network. The BiLSTM network consists of a forward and a backward layer of long short-term memory networks, which capture the past and future information in the lithium battery capacity sequence respectively. Use the vector weighted average optimization algorithm to further optimize the hyperparameters of the network, including the number of hidden layer units and the learning rate, to improve the modeling accuracy.

[0010] Step 4) Result reconstruction and health state estimation. Denormalize the prediction results of the BiLSTM network for each IMF component. Synthesize the prediction results of each component into the final lithium battery health state estimation value by superposition reconstruction.

[0011] Furthermore, in the above step 1), four 18650 lithium batteries with a nominal capacity of 2 Ah, namely B0005, B0006, B0007, and B0018 in the NASA lithium battery dataset, are used as the research objects. The four batteries are subjected to charge and discharge experiments at 24°C. Specifically, in the charging mode, a constant current of 1.5 A is used for charging. When the battery reaches the cut-off voltage of 4.2 V, it is switched to constant voltage charging and the charging stops when the current drops to 20 mA. During the discharging process, a constant current of 2 A is used for discharging until the voltage drops to the cut-off voltage. Batteries B0005 - B0007 have 168 charge and discharge cycles, and B0018 has 132 cycles. Each cycle corresponds to a capacity value.

[0012] Furthermore, in the above step 2), use the weighted mean idea to process the entity structure, and update the position of the vector by combining three core stages: the update rule, vector combination, and local search. In a D-dimensional space, a vector population composed of L vectors is X=(x1,x2…x L ), and the position of the l-th vector in the search space is x l =(x l1 ,x l2 ,…x lD ). The update rule is as follows:

[0013] (1) Update the position of the vector based on the Mean Rule (MR) to obtain vectors and

[0014]

[0015] Among them and are the vector positions updated by the mean rule for the vector population after the m-th iteration.

[0016] (2) Combine the two vectors calculated in the previous stage with the vector to generate a new vector Improve the local search ability:

[0017]

[0018] Among them is the new vector obtained by combining the vectors in the m-th generation, μ = 0.05×randn, randn is a random number subject to the standard normal distribution, and rand1, rand2 are random values within [0,1].

[0019] (3) In the local search stage, prevent the algorithm from falling into a local optimal solution. When rand < 0.5, generate a new vector around the optimal solution of all vectors in the m-th generation population :

[0020]

[0021] where x rnd randomly combines the components of the optimal, worst, and random solutions in the vector population, increasing the randomness of the algorithm to better search in the solution space; MR is the new vector created by the vector weighted mean; a1≠1 is an integer randomly selected within [1,L]; v1, v2 are two random numbers.

[0022] In the variational mode decomposition process, the decomposition number k and the penalty factor α have a significant impact on the decomposition result. The vector weighted average optimization algorithm is used to optimize the parameters k and α. The minimum envelope entropy is selected as the fitness function. After the battery capacity signal is decomposed by variational mode decomposition, if the obtained intrinsic mode function components contain a lot of noise and the periodic characteristics are not obvious, it indicates that the component sparsity is weak and the envelope entropy is large. On the contrary, if the signal shows strong sparse characteristics, it means that the envelope entropy value is small.

[0023] The envelope entropy reflects the randomness and irregularity degree of the signal amplitude envelope, and the specific definition is:

[0024]

[0025] In the formula, E P (k) is the envelope entropy of the k-th modal component, and P j is the normalized envelope, which is specifically expressed as:

[0026]

[0027] where a j is the envelope signal obtained by performing a Hilbert transform on the modal component u k , and N is the length of the battery capacity sequence.

[0028] Furthermore, in step 3), the long short-term memory network, as a variant of the recurrent neural network, improves the limitation of the recurrent neural network for long-term prediction information and is applicable to sequence analysis and prediction. However, the classical long short-term memory network does not consider the influence of future information on current information. The calculation process of the long short-term memory network is as follows:

[0029]

[0030] where f t , i t , C t , o t and h t represent the forget gate, input gate, memory cell, output gate, and output value, respectively; σ(·) and tanh(·) represent the Sigmoid activation function and tanh activation function, respectively; w f , w i , w o and w c are weight parameters; b f , b i , b o and b c are bias parameters.

[0031] The bidirectional long short-term memory network uses the hidden layer nodes in the forward long short-term memory network and the backward long short-term memory network to capture past and future information, which can improve the model's ability to represent the dynamic characteristics of the battery. The calculation process of the bidirectional long short-term memory network is as follows:

[0032]

[0033] where x t is the input at time t, and h t is the output corresponding to time t. The bidirectional long short-term memory network consists of a forward and a backward layer of long short-term memory networks, which capture past and future information in the lithium battery capacity sequence respectively. The vector weighted average optimization algorithm is used to further optimize the hyperparameters of the network to improve the estimation accuracy.

[0034] Further, in step 4), the intrinsic mode function components obtained by decomposing the original battery capacity are sorted in ascending order of frequency. The first several components are high-frequency and low-amplitude components, which reflect the noise and capacity recovery phenomena during the degradation process of the lithium battery. The last component is a low-frequency and high-amplitude component, which reflects the overall degradation trend of the lithium battery capacity. It is necessary to consider the correlation between each component and the health state when predicting the health state of the lithium battery. The prediction results of the network for each component are de-normalized, and the prediction results of each intrinsic mode function component are synthesized into the final lithium battery health state estimation result by the method of superposition and reconstruction.

[0035] Beneficial effects: The present invention improves the adaptability and accuracy of variational mode decomposition, and can better extract the multi-scale features of the lithium battery capacity. The bidirectional long short-term memory network combined with vector weighted average optimization improves the adaptability to the capacity regeneration phenomenon of the lithium battery and reduces the health state estimation error. Description of the Drawings

[0036] Figure 1 It is a flow chart of variational mode decomposition optimized by vector weighted average;

[0037] Figure 2 It is a structural diagram of a long short-term memory network and a bidirectional long short-term memory network; (a) is a structural diagram of a bidirectional long short-term memory network, and (b) is a structural diagram of a long short-term memory network;

[0038] Figure 3 It is an overall framework diagram of the model;

[0039] Figure 4 It is a schematic diagram of the decomposition result of lithium battery capacity data;

[0040] Figure 5 It is a comparison diagram of lithium battery health state estimation results; (a) is a comparison diagram of the estimation result of battery No. B0005, (b) is a comparison diagram of the estimation result of battery No. B0006, (c) is a comparison diagram of the estimation result of battery No. B0007, and (d) is a comparison diagram of the estimation result of battery No. B0018. Detailed Embodiments

[0041] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Step 1): Data collection and preprocessing;

[0043] Select the capacity decay data of batteries numbered B0005, B0006, B0007, and B0018 from the NASA lithium battery dataset. This dataset contains the records of the capacity changes of the batteries in multiple charge-discharge cycles, and the data includes the battery capacity values for each charge-discharge cycle. Check whether there are missing values or outliers in the data. If there are outliers, ensure the integrity of the data by removing the outliers or performing interpolation. After that, normalize the capacity data so that the numerical range of the capacity data is unified to [0, 1] to avoid the influence of different data dimensions on the model. Divide the processed data into a training set and a test set in chronological order. During the model training process, use the training set for training, and the test set is used for model validation and evaluation.

[0044] Step 2): Optimize the variational mode decomposition by the vector weighted average optimization algorithm;

[0045] Input the normalized capacity data into the variational mode decomposition. The goal of the variational mode decomposition is to decompose the lithium battery capacity data into several intrinsic mode function components. Each component represents a frequency component in the capacity data. The variational mode decomposition finds the optimal center frequency and bandwidth of each mode component through an iterative process to ensure that each component can better reflect the characteristics of the battery capacity decline. For an input signal g, the corresponding constrained variational model is:

[0046]

[0047] where u1,…,u k are k components; ω1,…,ω k are the center frequencies corresponding to each component; * is the convolution operation; is the time derivative of the function; δ(t) is the unit impulse function. Introduce the Lagrange multiplier λ and the penalty factor α to transform the constrained variational problem into an unconstrained form:

[0048]

[0049] The alternating direction multiplier method is used to update u k ,ω k , until the convergence condition is met. The capacity degradation data of the lithium battery contains random fluctuations caused by factors such as battery capacity regeneration. Through the variational mode decomposition, the influence of the capacity regeneration phenomenon on the estimation accuracy can be avoided, and the fluctuation information of the capacity regeneration is not completely discarded, and the characteristics of the battery capacity degradation at different scales can be effectively extracted.

[0050] Introduce the vector weighted average optimization algorithm to optimize the decomposition key parameters k and α, thereby improving the accuracy and self-adaptability of the variational mode decomposition results. Specifically, in a D-dimensional (D is 2) space, a vector population composed of L vectors is X = (x1, x2…xL ), the position in the search space of the l-th vector is x l =(x l1 , x l2 , … x lD ). The steps are as follows:

[0051] (1) Update the position of the vector based on the Mean Rule (MR) to obtain the vectors and

[0052]

[0053] where and are the vector positions updated by the mean rule for the vector population after the m-th iteration.

[0054] (2) Combine the two vectors calculated in the previous stage with the vector to generate a new vector to improve the local search ability:

[0055]

[0056] where is the new vector obtained by combining the vectors in the m-th generation, μ = 0.05×randn, randn is a random number obeying the standard normal distribution, and rand1, rand2 are random values within [0, 1].[[]]

[0057] (3) In the local search stage, prevent the algorithm from falling into a local optimal solution. When rand < 0.5, generate a new vector around the optimal solution of all vectors in the m-th generation population:

[0058]

[0059] where x rnd randomly combines the components of the optimal, worst, and random solutions in the vector population, increasing the randomness of the algorithm to better search in the solution space; MR is the new vector created by the weighted mean of the vectors; a1 ≠ 1 is an integer randomly selected within [1, L]; v1, v2 are two random numbers.

[0060] Select the minimum envelope entropy as the fitness function. The envelope entropy reflects the randomness and irregularity degree of the signal amplitude envelope, and the specific definition is:

[0061]

[0062] In the formula, E P (k) is the envelope entropy of the k-th modal component, Pj is the normalized envelope, specifically expressed as:

[0063]

[0064] where a j is the envelope signal obtained after the modal component u k undergoes the Hilbert transform, and N is the length of the battery capacity sequence. The vector weighted average optimization algorithm adaptively optimizes two important parameters of variational mode decomposition by minimizing the envelope entropy as the fitness function. The minimum envelope entropy reflects the randomness of the signal envelope. The smaller the entropy value, the smoother the signal decomposition. Otherwise, it may contain more noise.

[0065] Step 3): Use a bidirectional long short-term memory network for modeling and prediction;

[0066] Input the decomposed components into the bidirectional long short-term memory network for modeling. The bidirectional long short-term memory network has the ability of bidirectional learning and can learn information from both the past and the future simultaneously, which is very effective for processing time series data (such as battery capacity degradation). For each component, first perform normalization processing to ensure that the data meets the input requirements of the bidirectional long short-term memory network. Then, split each component into multiple time steps (for example, take 7 time points as an input sample) for training.

[0067] The calculation process of the bidirectional long short-term memory network is as follows:

[0068]

[0069] where x t is the input at time t, and h t is the output corresponding to time t.

[0070] The input of each layer is the input sequence at the current moment and the state information at the previous moment, and the output is the predicted value and state information at the current moment. Use the vector weighted average optimization algorithm to optimize the hyperparameters of the network, including the number of hidden layer units, learning rate, batch size, etc. The vector weighted average optimization algorithm optimizes the values of each hyperparameter by searching different combinations in the search space to achieve the best training effect. Through the backpropagation algorithm and the gradient descent method, train the weights of the network so that it can accurately predict the degradation trend of each intrinsic mode function component.

[0071] Step 4): Reconstruct the prediction results and estimate the health state;

[0072] The predicted results of the network are denormalized to the actual battery capacity range. This is achieved by reversing the normalization ratio of the predicted values with respect to the data. For the predicted results of each intrinsic mode function component, they are superimposed with the predicted results of other intrinsic mode function components to reconstruct the final battery capacity sequence. This reconstructed sequence is the final result of the state of health estimation. Through the reconstructed capacity sequence, the state of health of the battery at different cycles can be calculated. The state of health value can reflect the remaining health degree of the battery at present.

[0073] Step 5): Experiment and result analysis;

[0074] By comparing the predicted results with the true capacity values, evaluation metrics such as the root mean square error and the mean absolute error are calculated to evaluate the prediction accuracy of the model. In the embodiment, the state of health estimation error obtained by using the method of the present invention is reduced, and the proposed method is compared with existing baseline models such as ABMS-CEEMDAN-LSTM and PSO-BP.

[0075] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A lithium battery health state estimation method integrating VMD and BiLSTM, characterized in that: The following steps are involved: Collect lithium battery aging data sets, normalize and preprocess the original capacity data, and divide them into training sets, validation sets, and test sets in proportion; Using the vector weighted average optimization algorithm and taking the minimum envelope entropy as the fitness function, the decomposition parameters of the variational mode decomposition, including the decomposition mode number k and the penalty factor α, are optimized. The capacity data is decomposed at multiple scales to obtain several intrinsic mode function components. Each intrinsic mode function component is normalized and input into a bidirectional long short-term memory network. The network hyperparameters are optimized by a vector weighted average optimization algorithm, and the network is trained to predict the components. The prediction results of the bidirectional long short-term memory network are denormalized and superimposed and reconstructed to obtain the estimated value of the health status of the lithium battery.

2. The method according to claim 1, characterized in that The vector weighted average optimization algorithm comprises the following steps: Update the position of the vector population based on the mean rule to generate a new vector; Combined with the local search strategy, new vectors are randomly generated to avoid falling into the local optimal solution; Taking the minimum envelope entropy as the fitness function, the decomposition parameters of variational mode decomposition are iteratively optimized.

3. The method according to claim 1, characterized in that: The bidirectional long short-term memory network consists of a forward long short-term memory network and a backward long short-term memory network, which respectively capture the past and future information of the capacity sequence. The network hyperparameters include the number of hidden layer units and the initial learning rate, which are optimized by a vector weighted average optimization algorithm.

4. The method according to claim 1, characterized in that The envelope entropy calculation method of the variational mode decomposition is: Among them, Pj is the normalized envelope signal, which is obtained by Hilbert transform of the modal component, and N is the capacity sequence length.

5. The method according to claim 1, characterized in that The data set is the capacity decay data of batteries B0005, B0006, B0007, and B0018 in the NASA lithium battery data set, and the charge and discharge conditions are constant current charge and discharge to the cut-off voltage.

Images