Lithium battery life prediction method based on SVMD-Transform-LSTM fusion algorithm

Abstract

The invention discloses a lithium battery life prediction method based on an SVMD-Transform-LSTM fusion algorithm, and the method comprises the steps: decomposing the original capacity data of a target lithium battery through an SVMD method, and obtaining an original residual signal and an original mode component; inputting the original modal component into a trained LSTM model, and outputting a modal component predicted value of the target lithium battery in a preset round of circulation; inputting the original residual error signal into the trained Transform network model, and outputting a residual error signal prediction value of the target lithium battery in a preset round of circulation; and integrating the modal component prediction value and the residual signal prediction value to obtain a final life prediction result of the target lithium battery. According to the method, the random volatility of the capacity regeneration signal can be effectively captured through the SVMD-Transform-LSTM model, so that the prediction of the service life of the battery is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery life prediction, and more specifically to a lithium battery life prediction method based on an SVMD-Transformer-LSTM fusion algorithm. Background Art

[0002] Lithium batteries have been widely used in many fields such as electric vehicles, aerospace, and energy storage systems due to their advantages such as light weight, high energy density, and long cycle life. However, as the number of charge and discharge cycles of lithium batteries increases, the performance of lithium batteries gradually decreases due to the influence of multiple internal and external factors. When the capacity of a lithium battery decreases to the end of its life (End of Life, EOL), the lithium battery function fails and no longer meets the power consumption requirements of electrical equipment. The battery must be replaced or repaired, otherwise it may cause a catastrophic accident. The remaining useful life (RUL) of a battery refers to the number of remaining charge and discharge cycles between the predicted starting point of the battery life and the end of its life. Accurately predicting the RUL of a battery is crucial for battery status monitoring, operation and maintenance, and battery replacement, which can improve the stability and utilization efficiency of lithium batteries.

[0003] At present, there are more and more studies on the prediction methods of the remaining service life of lithium batteries, and the mainstream prediction methods are lithium battery model-based prediction methods and data-driven prediction methods.

[0004] The prediction method based on the lithium battery model is to establish a mathematical model for battery life prediction by analyzing the internal mechanism of the battery, the battery degradation mechanism and the external environmental factors. The existing technology uses electrochemical models and equivalent circuit models, and combines particle filtering and its derivative algorithms to estimate the model parameters to achieve battery life prediction. However, the prediction method based on the lithium battery model has limitations in use: (1) Lithium batteries are affected by external environmental factors such as temperature and humidity, and their battery degradation model is complex and has too many parameters; (2) Particle filtering and its derivative algorithms are overly dependent on accurate battery degradation models and particle filtering has obvious particle degradation problems, which affects the prediction accuracy.

[0005] The data-driven method uses machine learning algorithms to analyze future battery life trends using historical degradation data of lithium batteries. This method does not require the exploration of the internal mechanism of the battery, nor does it require the establishment of a large number of mathematical models, but can also achieve high prediction accuracy. In the prior art, a single machine learning algorithm is used for battery RUL prediction, such as support vector machines, Gaussian regression processes, and correlation vector machines. This improves the prediction accuracy to a certain extent, but fails to effectively handle the long-term dependencies of the sequence. The above data-driven algorithms have achieved great results in the field of battery life prediction, but there are still some areas that need to be improved: (1) The battery life prediction model is single, and the prediction generalization and accuracy are poor; (2) A large number of studies directly process the raw battery data and do not consider the capacity regeneration phenomenon separately; (3) Battery capacity data has strong long-term memory and random volatility. How to accurately map this relationship and achieve accurate long-term online prediction of future capacity is still a challenging problem.

[0006] Therefore, how to avoid the low accuracy of battery life prediction and difficulty in capturing capacity regeneration in the prediction results of a single machine learning prediction model during the performance degradation of lithium batteries due to the nonlinear, random and non-stationary time series characteristics of the battery's original capacity data is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] In view of the above problems, the present invention provides a lithium battery life prediction method based on the SVMD-Transformer-LSTM fusion algorithm to at least solve some of the technical problems mentioned in the above background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm includes the following steps:

[0010] S1. Decompose the original capacity data of the target lithium battery by the SVMD method to obtain the original residual signal and the original modal component;

[0011] S2, inputting the original modal component into the trained LSTM model, and outputting the modal component prediction value of the target lithium battery in a preset round cycle;

[0012] S3, inputting the original residual signal into the trained Transformer network model, and outputting the residual signal prediction value of the target lithium battery in a preset round cycle;

[0013] S4. Integrate the modal component prediction value and the residual signal prediction value to obtain a final life prediction result of the target lithium battery in a preset number of cycles.

[0014] Furthermore, in the step S1, it also includes: using a filter to perform a minimization standard constraint on the original residual signal and the original modal component, and the constraint condition is expressed as:

[0015]

[0016] in:

[0017]

[0018] Among them, u n (t) represents the nth original modal component; J 1 represents an objective function that minimizes the bandwidth of the original modal component in the frequency domain; J 2 represents the minimum error objective function of the extracted original modal components at each frequency of the original capacity data; J 3 represents the objective function of ensuring that the frequency characteristics of each original modal component conform to the expected frequency range; represents the derivative with respect to time t; δ(t) represents the unit impulse function, which is used to select characteristic time points in signal processing; j represents the imaginary unit in the complex number; ω n represents the nth original modal component u n (t) represents the center frequency; f(t) represents the original capacity data; * represents the convolution operation; t represents the time t; f r (t) represents the residual signal, which is the original capacity data f(t) minus u n The residual component of (t); β n (t) represents the weight function, and the center frequency ω n Correlation, used to control the residual signal f r Extraction of each component in (t); i (t) is the impulse response of the filter; f u (t) represents the remaining part of the original capacity signal after removing all modal components from 1 to n, which is the original residual signal of the global degradation of the target lithium battery; It means to sum the first n-1 modal functions, i is the index variable, and its value ranges from 1 to n-1.

[0019] Furthermore, in the step S1, it also includes:

[0020] An orthogonality index is introduced to analyze the correlation between the original capacity data, the original residual signal and the original modal component to obtain an orthogonality value;

[0021] The decomposition result is judged according to the orthogonality value; if the orthogonality value tends to 0, it is judged that the interaction between the original capacity data and the original residual signal and the original modal component is lower than a preset value, and the decomposition quality is high.

[0022] Furthermore, the original modal component includes an original capacity regeneration signal and an original fluctuation signal.

[0023] Furthermore, the training steps of the LSTM model include:

[0024] The original capacity signals of a large number of lithium battery samples are obtained and recorded as historical original capacity signals; the historical original capacity signals are decomposed by using the SVMD method to obtain corresponding historical modal components;

[0025] Dividing the historical modal components into a training set and a test set according to a preset ratio;

[0026] Inputting the historical modal components in the training set into the LSTM model, and calculating the modal component prediction values ​​of the lithium battery samples in the cycle rounds through forward propagation;

[0027] Compare the predicted value of the modal component with the actual value of the modal component of the lithium battery sample in the cycle round, and calculate the LSTM model loss function;

[0028] Optimizing the LSTM model based on the LSTM model loss function;

[0029] The trained LSTM model is verified by the historical modal components in the test set, and the hyperparameters in the LSTM model are adjusted according to the verification results to achieve secondary optimization of the LSTM model.

[0030] Furthermore, the mean square error or mean absolute error is used as the LSTM model loss function.

[0031] Furthermore, the training steps of the Transformer network model include:

[0032] Obtaining the original capacity signals of a large number of lithium battery samples and recording them as historical original capacity signals; using the SVMD method to decompose the historical original capacity signals to obtain corresponding historical residual signals;

[0033] Dividing the historical residual signal into a training set and a test set according to a preset ratio;

[0034] Inputting the historical residual signal in the training set into the Transformer network model, and calculating the residual signal prediction value of the lithium battery sample in the cycle round through forward propagation;

[0035] Compare the predicted value of the residual signal with the actual value of the residual signal of the lithium battery sample in the cycle round, calculate the loss function of the Transformer network model; optimize the Transformer network model based on the loss function of the Transformer network model;

[0036] The trained Transformer network model is verified by the historical residual signal in the test set, and the hyperparameters in the Transformer network model are adjusted according to the verification result to achieve secondary optimization of the Transformer network model.

[0037] Furthermore, the mean square error or mean absolute error is used as the loss function of the Transformer network model.

[0038] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a lithium battery life prediction method based on the SVMD-Transformer-LSTM fusion algorithm, which has the following beneficial effects:

[0039] The present invention can decompose the original battery capacity data with time series characteristics such as nonlinearity and non-stationarity into a series of time series signals with periodicity and stability through the SVMD sequence decomposition technology, thereby ensuring the effectiveness and stability of subsequent modeling.

[0040] The present invention can effectively fit the global degradation trend of the residual signal through the Transformer network model, and the LSTM network model can effectively capture the periodic changes and serial correlation of the capacity regeneration signal and the fluctuation signal. Different network models are used to model signals with different timing characteristics, which improves the accuracy of the model while avoiding the limitations of a single algorithm.

[0041] The method provided by the present invention can integrate the prediction results of the global degraded residual signal by Transformer and the prediction results of the capacity regeneration signal and other fluctuation signals by LSTM to obtain the life prediction results of the lithium battery. To a certain extent, it can solve the long-term dependency of the sequence and effectively capture the random volatility of the capacity regeneration signal, so as to more accurately predict the capacity degradation trend of the battery. At the same time, since the battery capacity is an important indicator reflecting the RUL, the prediction method proposed by the present invention can also realize the prediction of the RUL of the battery.

[0042] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0043] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0045] Figure 1 A schematic flow chart of a lithium battery life prediction method based on the SVMD-Transformer-LSTM fusion algorithm provided in an embodiment of the present invention.

[0046] Figure 2 A schematic diagram of a NASA lithium battery capacity decay curve provided in an embodiment of the present invention.

[0047] Figure 3 A schematic diagram of the SVMD decomposition results of B05 battery capacity data provided by an embodiment of the present invention.

[0048] Figure 4 A schematic diagram of the prediction results of the B05 battery residual signal provided in an embodiment of the present invention.

[0049] Figure 5 A schematic diagram of the IMF prediction results of a B05 battery provided in an embodiment of the present invention.

[0050] Figure 6 A schematic diagram of the integration of B05 prediction results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] See also Figure 1 As shown, the embodiment of the present invention discloses a lithium battery life prediction method based on the SVMD-Transformer-LSTM fusion algorithm, comprising the following steps:

[0053] S1. Decompose the original capacity data of the target lithium battery by the SVMD method to obtain the original residual signal and the original modal component;

[0054] S2, input the original modal component into the LSTM model, and output the modal component prediction value of the target lithium battery in the preset cycle;

[0055] S3, inputting the original residual signal into the Transformer network model, and outputting the residual signal prediction value of the target lithium battery in the preset round cycle;

[0056] S4. Integrate the modal component prediction value and the residual signal prediction value to obtain the final life prediction result of the target lithium battery in a preset number of cycles.

[0057] Next, each of the above steps will be described in detail.

[0058] In the above step S1, as the cycle period increases, the overall capacity of the lithium battery shows a downward trend, but due to the capacity regeneration phenomenon in some cycles, the curve rebounds slightly, so the lithium battery capacity data has time series characteristics such as nonlinearity and non-stationarity, and the prediction model trained using the original capacity data cannot fully utilize its time series characteristics. Therefore, in the embodiment of the present invention, a sub-modal decomposition algorithm (SVMD algorithm) is used to decompose the original capacity data of the target lithium battery to obtain the globally degraded original residual signal and the original modal component; the original modal component includes the original capacity regeneration signal and other original fluctuation signals, and the decomposed signal is more periodic and stable than the original capacity data, so the modeling can be achieved more effectively.

[0059] SVMD is an efficient and fast adaptive method. SVMD does not need to set the exact number of signals when processing the sequence. Instead, it obtains an approximation of the center frequency of the modal signal and continuously extracts the modal signal until all modal signals are extracted or the reconstruction error drops below the specified threshold. Specifically, the SVMD method is used to perform multimodal decomposition on the original capacity data of the target lithium battery. The decomposition process is as follows:

[0060] (1) Input the original capacity data of the target lithium battery into the constructed SVMD decomposition algorithm model, and decompose the original capacity data f(t) into two types of signals, including the nth original modal component u n (t) and the original residual signal f r (t); where the original residual signal f r (t) is the original capacity data f(t) minus u n The input signal of (t) includes the obtained capacity decomposition signal; the process is expressed as:

[0061] f(t)=u n (t)+f r (t)

[0062] (2) To ensure the original modal component u n The bandwidth of (t) can be tightly concentrated at the center frequency ω in the frequency domain.n Therefore, it is necessary to n (t) The bandwidth in the frequency domain is minimized, and the objective function J 1 It is expressed as:

[0063]

[0064] Among them, J 1 An objective function that represents the minimization of the bandwidth of the original modal component in the frequency domain; represents the partial derivative with respect to time t; δ(t) represents the unit impulse function, which is used to control the local characteristics and time resolution of the modal component; j represents the imaginary unit in the complex number; ω n is the center frequency of the nth original modal component, * represents the convolution operation; t represents the time t;

[0065] (3) At the same time, in order to avoid the influence of noise on the original modal extraction results, resulting in a large error in the extracted original modal component at a certain frequency component, it is necessary to ensure that the original modal component u n (t) has the minimum error at a certain frequency under the original capacity data, the objective function J 2 It is expressed as:

[0066]

[0067] Among them, J 2 represents the minimum error objective function of the extracted original modal components at each frequency of the original capacity data; β n (t) is the impulse response of the filter; f r (t) represents the original residual signal, which is the original capacity data f(t) minus u n The remaining amount of (t);

[0068] (4) Through the above two constraints function J 1 and J 2 , the nth original modal component u of the battery capacity signal is obtained n (t);

[0069] (5) To avoid obtaining the nth original modal component u n (t), there is an obvious mode superposition phenomenon with other original mode components; it is necessary to ensure that the nth original mode component u n The center frequency of (t) should not overlap with the center frequency of the previous original modal component as much as possible, that is, to ensure that the nth original modal component u n The center frequency of (t) is consistent with the expected frequency range, and the objective function J 3 It is expressed as:

[0070]

[0071] Among them, J 3 It represents the objective function that ensures that the frequency characteristics of each original modal component meet the expected frequency range. Its main function is to constrain the spectral characteristics of each original modal component, ensure that the frequency characteristics of each original modal component meet the expected frequency range, and no modal superposition phenomenon occurs; i is an index variable, with a value of 1 to n-1; β i (t) is the impulse response of the filter.

[0072] (6) To ensure the original modal component u n (t) and the unprocessed part can be reconstructed into the original capacity data, and a constraint is imposed on the original capacity data, which is expressed as:

[0073]

[0074] Among them, f u (t) represents the remaining part of the original capacity signal after removing all modal components from 1 to n, which is the original residual signal of the global degradation of the target lithium battery; It means summing the first n-1 modal functions, i is the index variable, and its value ranges from 1 to n-1.

[0075] (7) Assuming that SVMD is to obtain the nth modal component, the first n-1 modal components are known and can be used as constraints. If n is 1, then n-1 is 0. When n-1 modes are known, the problem of extracting the nth mode can be expressed as a constrained minimization problem, where J 1 , J 2 and J 3 The combination of is constrained, expressed as:

[0076]

[0077] It should be noted that the principle of the modal decomposition algorithm is: (1) first obtain an original modal component, assuming it is the first original modal component, and the rest are collectively referred to as the residual signal f r (t) (but at this time the residual signal has not been completely decomposed and the modal components have not been completely extracted). At this time, it is necessary to ensure that the extracted modal components have concentrated frequencies and minimize errors, so the constraint J needs to be satisfied. 1 and J 2 (2) Then we should start from f r (t) Continue to extract the second modal component. The extracted modal component does not overlap with the center frequency of the first modal component (if it overlaps, it proves that the parameter setting is unreasonable and that all modal components have been completely extracted). At this time, the third constraint condition J is required. 3, ensuring that all extracted modal components do not overlap. So n is used here to represent the modal component currently being extracted.

[0078] After decomposing the original capacity data of the target lithium battery into the original residual signal and the original modal component, the orthogonality index (Index of Orthogonality, IO) is introduced to analyze the correlation between the original capacity data and the original residual signal and the original modal component to obtain the orthogonality value; the decomposition result is judged according to the orthogonality value; generally speaking, the orthogonality value tends to 0, which means that the better the orthogonality between the signals, the smaller the modal superposition phenomenon, and the lower the correlation; the value tends to 1, which means that the correlation between the signals is higher; in the embodiment of the present invention, it is assumed that if the orthogonality value is low and tends to 0, the interaction between the modes is small and the decomposition quality is high; the orthogonality value is expressed as:

[0079]

[0080] Among them, IO represents the orthogonality value, which is used to measure the orthogonality between different decomposed modes. The higher the orthogonality, the smaller the interaction between the modes, indicating a good decomposition effect; m represents m modal functions; i is the target index, indicating the i-th modal function; j is the target index, indicating the j-th modal function; u i (t) represents the i-th mode function; u j (t) represents the jth mode function different from the ith one; x(t) represents the original signal; u i (t)u j (t) / x 2 (t) indicates that the cross term between the ith mode and the jth mode is normalized to the energy of the original signal, so that the calculation of the orthogonality value is related to the energy of the original signal.

[0081] In the above step S2, in order to effectively capture the periodic changes and serial correlation of the capacity regeneration signal and fluctuation signal of the target lithium battery in the preset cycle, the LSTM model is used for prediction. Therefore, the original capacity regeneration signal and other original fluctuation signals obtained by SVMD decomposition are trained separately using the LSTM model. The following are the basic steps of the LSTM network model training process:

[0082] (1) Data preparation:

[0083] Obtaining original capacity signals of a large number of lithium batteries (hereinafter referred to as lithium battery samples) as historical original capacity signals; decomposing the historical original capacity signals by using the SVMD method to obtain corresponding historical modal components; dividing the historical modal components into a training set and a test set according to a preset ratio, for example, using 65% as a training set and 35% as a test set;

[0084] (2) LSTM model construction:

[0085] The LSTM model includes an input layer, an LSTM unit (hidden layer) and an output layer; the LSTM unit is mainly composed of a forget gate, an input gate and an output gate; these gates control the input, retention and output of information, helping the network to effectively remember and forget information; the forget gate detects the data that needs to be cleared in the data through the Sigmoid layer and removes redundant data. Its mathematical expression is as follows:

[0086] f t =σ(W f ·[o t-1 ,i t ])+b f

[0087] The new cell state requires the Sigmoid layer of the input gate to update the cell value, and then use the tanh layer to select new candidate values.

[0088] The final output unit state result is determined by the LSTM neural network, and its expression is:

[0089]

[0090] In the above formula, f t W represents the forget gate unit; f represents the weight matrix; σ represents the Sigmoid activation function, the output value is between 0 and 1, which is used to determine how much previous memory to retain; · represents the dot product operation is the multiplication of the matrix and the vector, which is used to map the concatenated input and hidden state to the output of the forget gate; b f Represents the bias term of the forget gate, which is a vector added to the linear transformation result of the forget gate; i t represents the input gating unit; G t represents the tanh layer of the input gate; y t Represents the output of the forget gate; o t represents the output gating unit; b y represents the bias vector; Represents the Sigmoid function;

[0091] (3) Loss function selection:

[0092] Battery life prediction is a regression task, and the mean squared error (MSE) or mean absolute error (MAE) is usually used as the LSTM model loss function to achieve higher robustness.

[0093] (4) Model training: Input the historical modal components in the training set into the LSTM model, and calculate the predicted value of the modal components of the lithium battery sample in the cycle through forward propagation; compare the predicted value of the modal components with the actual value of the modal components of the lithium battery sample in the cycle, and calculate the LSTM model loss function; optimize the LSTM model based on the LSTM model loss function;

[0094] (5) The prediction ability of the LSTM model is verified by using the historical modal components in the test set, and the difference between the prediction results and the actual capacity is compared to verify the accuracy of the model; and the hyperparameters (such as learning rate, regularization parameters, etc.) in the LSTM model are adjusted according to the verification results to achieve secondary optimization of the LSTM model;

[0095] In the above step S3, in order to solve the long-term dependency problem of the sequence, the Transformer network model is used to predict the residual signal of the global degradation of the target lithium battery in the preset round cycle. The original residual signal obtained by SVMD decomposition is input into the Transformer network model for training. The basic steps of the training process of the Transformer network model include:

[0096] (1) Data preparation:

[0097] Obtaining original capacity signals of a large number of lithium batteries (hereinafter referred to as lithium battery samples) as historical original capacity signals; decomposing the historical original capacity signals by using the SVMD method to obtain corresponding historical residual signals; dividing the historical residual signals into a training set and a test set according to a preset ratio, for example, using 65% as a training set and 35% as a test set;

[0098] (2) Transformer model construction:

[0099] The Transformer network framework includes an encoder and a decoder. The Transformer-based encoder can extract battery degradation features in the battery capacity sequence. In order to make full use of the position information of the sequence, a position encoder (PE) is introduced to inject the position information.

[0100] In Transformer, each encoder is divided into two complete sub-layers: a multi-head attention mechanism and a feed-forward fully connected sub-layer; where:

[0101] Multi-head attention mechanism: The multi-head attention mechanism is an improvement based on the attention mechanism, and is better at learning the correlation within the sequence. The multi-head attention mechanism calculates the feature matrix attention value by scaling the dot product attention, and uses its own vector Q (query), K (key), and V (value) as variables to participate in the calculation, expressed as:

[0102]

[0103] Among them, QK T Express concern; represents normalization processing, d k Indicates that the scaling factor is the dimension of the key vector, which is used to scale the effect of the dot product;

[0104] Feedforward fully connected sublayer: In the encoder structure, Transformer adds a feedforward fully connected sublayer, which is applied to each position, performing two linear transformations and one nonlinear transformation to obtain a higher-dimensional mapping output;

[0105] The Transformer-based decoder is the part used to generate the target sequence, and in this invention, it is used for the battery life prediction task.

[0106] (3) Loss function selection: Select mean square error (MSE), mean absolute error (MAE), etc. as the loss function of the Transformer network model;

[0107] (4) Training process: inputting the historical residual signal in the training set into the Transformer network model for training; outputting the predicted value of the residual signal of the lithium battery sample in the cycle round; comparing the predicted value of the residual signal with the actual value of the residual signal of the lithium battery sample in the cycle round, and calculating the loss function of the Transformer network model; optimizing the Transformer network model based on the Transformer network model loss function; calculating the gradient and updating the model weights through backpropagation;

[0108] (5) The trained Transformer network model is verified by the historical residual signal in the test set, and the hyperparameters (such as learning rate, number of Transformer layers, etc.) in the Transformer network model are adjusted according to the verification results to optimize the generalization ability and prediction accuracy of the model, thereby achieving secondary optimization of the Transformer network model.

[0109] In the above step S4, the modal component prediction value and the residual signal prediction value obtained above are superimposed and integrated, that is, the two are added together to obtain the life prediction result of the lithium battery.

[0110] In order to verify the effectiveness of the lithium battery life prediction method based on the SVMD-Transformer-LSTM fusion algorithm provided by the embodiment of the present invention, a public data set provided by NASA is selected, and a group of battery models are 18650 lithium batteries (B05, B06, B07, B18) for verification. Table 1 below shows the experimental parameter information of the lithium battery.

[0111] Table 1 Experimental parameter information

[0112] Battery Cycle times Failure threshold / Ah B05 168 1.4 B06 168 1.4 B07 168 1.5 B18 132 1.4

[0113] The battery failure threshold is set to 70% to 80% of the rated capacity, so the failure threshold of B05, B06 and B18 batteries is set to 70%, i.e. 1.4Ah. However, considering that the battery capacity of B07 battery did not drop to 1.4Ah during the entire cycle charge and discharge experiment, the failure threshold of B07 is set to 75%, i.e. 1.5Ah. The lithium battery capacity decay curves of the B05, B06, B07 and B18 battery data sets are shown in Figure 2. Figure 2 shown.

[0114] from Figure 2 It can be seen that as the cycle period increases, the overall capacity of the battery shows a downward trend, but due to the capacity regeneration phenomenon in some cycles, the curve rebounds slightly. Therefore, the lithium battery capacity data has time series characteristics such as nonlinearity and non-stationarity. Using the original capacity data to train the prediction model cannot fully utilize its time series characteristics. Therefore, in order to improve the accuracy, the present invention uses SVMD to decompose the original capacity data. The decomposed signal is more periodic and stable than the original sequence. Taking the original capacity data of B05 as an example, SVMD decomposition is performed on it. The decomposition result is as follows: Figure 3 shown.

[0115] The orthogonality index (IO) is introduced to analyze the correlation between the original capacity sequence and each decomposed signal. In the sequence decomposition technology, the value tends to 0, which means that the orthogonality between the signals is better, the modal superposition phenomenon is smaller, and the correlation is low; the value tends to 1, which means that the correlation between the signals is higher. The orthogonality index of each sequence is shown in Table 2, where each IO value represents the IO value of the signal in the row and the signal in the column.

[0116] Table 2 Orthogonality index of each sequence

[0117] Signal B05 residual signal IMF1 IMF2 IMF3 B05 1 0.99996 0.00612 0.00348 0.00253 residual signal 0.99996 1 0.00174 0.00007 0.00001 IMF1 0.00612 0.00174 1 0.00095 0.00042 IMF2 0.00348 0.00007 0.00095 1 0.01651 IMF3 0.00253 0.00001 0.00042 0.01651 1

[0118] It can be seen from Table 2 that the IO value between the residual signal after decomposition of the original capacity data and the original capacity data is 0.99996, which is close to 1, indicating that the correlation between the residual signal and the original capacity is extremely high, and the residual signal can represent the global degradation trend of the lithium battery; the IO value of the modal signal IMF1~IMF3 and the original capacity tends to 0, indicating that the modal signal IMF1~IMF3 has a low correlation with the original capacity, and the modal signal IMF1~IMF3 represents the noise in the original capacity sequence; at the same time, a signal is extracted from the abnormal local extreme point caused by the capacity regeneration phenomenon, and IMF1 is obtained after SVMD decomposition, which fully represents the characteristics of the capacity regeneration phenomenon.

[0119] In summary, the original capacity sequence of B05 is decomposed into 4 signals by SVMD, among which the residual signal represents the trend signal of long-term degradation of lithium battery capacity, IMF1 represents the fluctuation signal of capacity regeneration, and IMF2 and IMF3 represent other noise fluctuation information in the degradation process of lithium battery. The 4 signals decomposed by SVMD are more time-series, stable and periodic as a whole, which is convenient for subsequent modeling and prediction.

[0120] The residual signal and IMF signal obtained by SVMD decomposition are input into the deep learning model for training. Taking B05 as an example, there are 168 groups of battery capacity data. The first 109 groups of data are selected as the model training set, and the last 59 groups are selected as the model test set, so the starting point of battery life prediction is the 110th charge and discharge cycle. The residual signal obtained by SVMD decomposition is input into the Transformer network model for training. Table 3 shows the parameter settings of the Transformer network model.

[0121] Table 3 Transformer network model parameter settings

[0122]

[0123] Then, the residual signal of the test set is input into the trained Transformer network model, and the prediction results are as follows: Figure 4 As shown. Figure 4 It can be seen that the Transformer network model has a higher prediction accuracy in the early stage of prediction, which proves that the Transformer network model has a certain accuracy in solving the long-term dependency of the sequence.

[0124] The three IMF signals obtained by SVMD decomposition are trained separately using the LSTM network model. Table 4 shows the parameter settings of the LSTM network model.

[0125] Table 4 LSTM network model parameter settings

[0126]

[0127] Similarly, the test set of three IMF signals is input into the trained LSTM network model, and the prediction results are as follows: Figure 5 As shown. Figure 5 It can be seen that the LSTM network model can accurately fit the fluctuation trend of the IMF signal, proving that the trained LSTM network model has a certain degree of accuracy in processing sequence volatility, capturing the periodic changes of high-frequency signals and long-term prediction.

[0128] Finally, the prediction results of the global degradation residual signal by Transformer and the prediction results of the capacity regeneration signal and other fluctuation signals by LSTM are integrated to obtain the life prediction results of the lithium battery. Figure 6 As shown, the lithium battery life prediction method proposed in the present invention solves the long-term dependency of the sequence to a certain extent, and effectively captures the random volatility of the capacity regeneration signal, which can more accurately predict the capacity degradation trend of the battery. At the same time, since the battery capacity is an important indicator reflecting the RUL, the prediction method proposed in the present invention can also predict the RUL of the battery.

[0129] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. The above description of the disclosed embodiments enables professionals and technicians in the field to implement or use the present invention. Various modifications to these embodiments will be obvious to professionals and technicians in the field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm, It is characterized in that The steps include: S1. Decompose the original capacity data of the target lithium battery by the SVMD method to obtain the original residual signal and the original modal component; S2, inputting the original modal component into the trained LSTM model, and outputting the modal component prediction value of the target lithium battery in a preset round cycle; S3, inputting the original residual signal into the trained Transformer network model, and outputting the residual signal prediction value of the target lithium battery in a preset round cycle; S4. Integrate the modal component prediction value and the residual signal prediction value to obtain a final life prediction result of the target lithium battery in a preset number of cycles.

2. According to claim 1, a lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm, It is characterized in that In the step S1, it also includes: using a filter to perform a minimization standard constraint on the original residual signal and the original modal component, and the constraint condition is expressed as: in: Among them, u n (t) represents the nth original modal component; J 1 represents an objective function that minimizes the bandwidth of the original modal component in the frequency domain; J 2 represents the minimum error objective function of the extracted original modal components at each frequency of the original capacity data; J 3 represents the objective function of ensuring that the frequency characteristics of each original modal component conform to the expected frequency range; represents the derivative with respect to time t; δ(t) represents the unit impulse function, which is used to select characteristic time points in signal processing; j represents the imaginary unit in the complex number; ω n represents the nth original modal component u n (t) represents the center frequency; f(t) represents the original capacity data; * represents the convolution operation; t represents the time t; f r (t) represents the residual signal, which is the original capacity data f(t) minus u n The residual component of (t); β n (t) represents the weight function, and the center frequency ω n Correlation, used to control the residual signal f r Extraction of each component in (t); β i (t) is the impulse response of the filter; f u (t) represents the remaining part of the original capacity signal after removing all modal components from 1 to n, which is the original residual signal of the global degradation of the target lithium battery; It means summing the first n-1 modal functions, i is the index variable, and its value ranges from 1 to n-1.

3. According to the lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm according to claim 1, It is characterized in that In the step S1, it also includes: An orthogonality index is introduced to analyze the correlation between the original capacity data, the original residual signal and the original modal component to obtain an orthogonality value; The decomposition result is judged according to the orthogonality value; if the orthogonality value tends to 0, it is judged that the interaction between the original capacity data and the original residual signal and the original modal component is lower than a preset value, and the decomposition quality is high.

4. According to claim 1, a lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm, It is characterized in that The original modal component includes an original capacity regeneration signal and an original fluctuation signal.

5. According to claim 1, a lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm, It is characterized in that The training steps of the LSTM model include: The original capacity signals of a large number of lithium battery samples are obtained and recorded as historical original capacity signals; the historical original capacity signals are decomposed by using the SVMD method to obtain corresponding historical modal components; Dividing the historical modal components into a training set and a test set according to a preset ratio; Inputting the historical modal components in the training set into the LSTM model, and calculating the modal component prediction values ​​of the lithium battery samples in the cycle rounds through forward propagation; Compare the predicted value of the modal component with the actual value of the modal component of the lithium battery sample in the cycle round, and calculate the LSTM model loss function; Optimizing the LSTM model based on the LSTM model loss function; The trained LSTM model is verified by the historical modal components in the test set, and the hyperparameters in the LSTM model are adjusted according to the verification results to achieve secondary optimization of the LSTM model.

6. A lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm according to claim 5, It is characterized in that The mean square error or mean absolute error is used as the LSTM model loss function.

7. A lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm according to claim 1, It is characterized in that The training steps of the Transformer network model include: Obtaining the original capacity signals of a large number of lithium battery samples and recording them as historical original capacity signals; using the SVMD method to decompose the historical original capacity signals to obtain corresponding historical residual signals; Dividing the historical residual signal into a training set and a test set according to a preset ratio; Inputting the historical residual signal in the training set into the Transformer network model, and calculating the residual signal prediction value of the lithium battery sample in the cycle round through forward propagation; Compare the predicted value of the residual signal with the actual value of the residual signal of the lithium battery sample in the cycle round, calculate the loss function of the Transformer network model; optimize the Transformer network model based on the loss function of the Transformer network model; The trained Transformer network model is verified by the historical residual signal in the test set, and the hyperparameters in the Transformer network model are adjusted according to the verification result to achieve secondary optimization of the Transformer network model.

8. A lithium battery life prediction method based on SVMD-Transformer-LSTM fusion algorithm according to claim 7, It is characterized in that The mean square error or mean absolute error is used as the loss function of the Transformer network model.

Images