Adaptive Incremental Battery RUL Prediction Method Combining Multi-Source Degradation Features

Abstract

The present invention discloses an adaptive incremental battery RUL prediction method combining multi-source degradation features, belonging to the field of battery RUL prediction technology. In this method, differential thermovoltammetry is used to extract multi-source degradation features from surface temperature and terminal voltage data, and HI preprocessing with adaptive noise is obtained, effectively eliminating the capacity regeneration phenomenon of the fused data and the influence of noise on the prediction process. Then, the meta-heuristic algorithm sand cat swarm optimization is used to optimize the regularization parameters in the random configuration network, and the parameter adaptive incremental neural network SC-SCN is obtained. This network better meets the real-time data requirements of the battery RUL prediction workflow industry. Adding an adaptive optimization algorithm can effectively improve the flexibility and prediction efficiency of the prediction model, and effectively solve the technical problems existing in the existing deep learning prediction methods, such as the single health index cannot stably and effectively characterize battery degradation and the network structure needs to be determined in advance.

Description

Technical Field

[0001] The present invention belongs to the technical field of life prediction and incremental learning, and relates to an adaptive incremental battery RUL prediction method combining multi-source degradation features. Background Art

[0002] Lithium-ion batteries (LIBs) are widely used in aerospace, electric vehicles, mobile electronic devices and other fields due to their wide operating temperature range and high energy density. Establishing an accurate and effective remaining useful life prediction model has always been the core issue of the battery management system (BMS) of LIBs. The prediction of the battery RUL in seconds can improve the use safety and extend the battery life.

[0003] There are mainly three methods for battery RUL prediction: model-driven, data-driven and hybrid methods. The model-based prediction method uses the physical model or electrochemical equation of the battery to explore its degradation mechanism from the physical structure of the battery, and has the advantage of stable performance. However, the disadvantage is that it requires accurate battery electrochemical and physical equations and is not suitable for off-line testing. The data-driven RUL prediction relies on deep learning algorithms to establish a neural network model to explore the correlation between battery life degradation and related data. The data-driven prediction method utilizes the precise analysis of statistical methods and the powerful learning ability of neural networks to achieve efficient and fast model approximation without knowing the battery operation model. The most significant advantage of the data-driven RUL prediction algorithm is that it does not need to know the exact model of the battery and only requires a certain amount of degradation data to complete, so it has become the mainstream method for battery life prediction.

[0004] Data-driven battery RUL prediction methods mainly include artificial neural networks, support vector machines, support vector regression, particle filtering, etc. Among them, artificial neural networks, due to their powerful nonlinear processing capabilities, adaptability and self-learning capabilities, are designed to simulate the operation of the human brain nervous system and shine in the problem of battery life prediction. In the prior art, some people use hybrid gated recurrent units (CNNs) to learn the degradation characteristics and time dependence in the charging curve, and use new voltage and current data to achieve battery SOH estimation. Some people have also disclosed an effective long-cycle battery health management method, first comparing the performance difference FFNN and NARXRNN 30-cycle predictions, and then selecting a more accurate long-cycle battery prediction, and experimentally proved that this method has the advantages of small error and low complexity error and low complexity. Others use the elastic mean square back propagation method to adaptively optimize long-term and short-term data, and obtain a small batch data training prediction method. In addition, there are literature records: various hybrid model-driven methods are compared, and the results show that the combined LSTM+GPR model has the best prediction accuracy and can achieve accurate results for single-step and multi-step capacity prediction. Deep learning requires a pre-determined network structure, which is hardly applicable to process industries, especially for battery life prediction with real-time requirements.

[0005] The above neural network-based method requires the hidden layer structure of the prediction model to be determined in advance. The degradation data of the battery continues to grow during real-time operation, and the fixed hidden layer nodes and layers cannot meet similar real-time industrial requirements. If too many nodes are selected, there will be less data in the early stage of prediction, and the model will not fit well; if the number of fixed nodes is too small, the increase in input data in the later stage of prediction will lead to insufficient overfitting of the model, which will seriously affect the accuracy of battery RUL prediction and threaten the safety of battery use.

[0006] To address the above problems, some people adopt a constructed incremental method, the Stochastic Configuration Network (SCN), which can effectively solve the above problems. Its characteristics are as follows: starting from a small network, a set of weights is randomly generated, and a supervision mechanism is used to verify that it meets the universal approximation condition. The best set of weights is selected from them, and hidden layer nodes are gradually generated; the least squares method is used to determine the output weights. The SCN has the advantages of fast convergence speed, low network cost, and strong generalization ability; it can be used not only for classification research but also for regression prediction. In the SCN, [a certain parameter] is an important regularization hyperparameter, which determines the selection of weights and biases of hidden layer nodes. It needs to be manually adjusted according to the volume and type of input data. The emergence of meta-heuristic algorithms provides a solution for the adaptive selection of neural network hyperparameters. To address the defects of the SCN in large-scale data regression, it is documented in the literature that: first, a logical mapping and mutation operator are introduced into the sparse search algorithm, and then combined with the SCN to obtain an effective regression CSSA-SCN. Different from the constructed structure, it is also documented in other literature that: ISSA-FSCN proposed for the ISSA. Under the optimization conditions of an improved sparse search algorithm, the number of hidden layer nodes is corrected to achieve hyperparameter adaption. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an adaptive incremental battery RUL prediction method combining multi-source degradation features, aiming to solve the problem of predicting the remaining useful service life of lithium-ion batteries.

[0008] The technical solution adopted by the present invention is as follows: an adaptive incremental battery RUL prediction method combining multi-source degradation features, the method comprising the following steps:

[0009] Step (1), using differential thermovoltammetry to extract multi-source degradation features from surface temperature and terminal voltage data, obtaining HI preprocessing with adaptive noise;

[0010] Step (2), using the HI data matrix as the input of the SC-SCN, and optimizing the regularization parameter in the random configuration network by the meta-heuristic algorithm sand cat swarm optimization, obtaining the parameter adaptive incremental neural network SC-SCN;

[0011] Step (3), using the parameter adaptive incremental neural network SC-SCN obtained in step (2) to predict the adaptive incremental battery RUL.

[0012] The detailed method in step (1) is as follows: by calculating the entropy increase ratio of the battery surface temperature and terminal voltage during the constant current charge and discharge process of the battery, DTV is obtained, and its discrete expression is as shown in formula (1):

[0013]

[0014] Wherein, T, t, V, and k respectively represent temperature, time, voltage, and time step, and V (t) represents the terminal voltage at time t, and T v(t) is the surface temperature.

[0015] Furthermore, the detailed steps in the above step (2) are as follows:

[0016] Step 1: Parameter initialization:

[0017] Initialize the overall size of Sand Cat (pop), iteration (iter), upper and lower bounds of search hyperparameters (ub and lb), dimension (D), maximum number of hidden nodes (L max ), candidate nodes (T max ), fault tolerance (ε), and initialize the parameter value r within the given range [lb, ub] as Sand r ;

[0018] Step 2: Construct and train SCN:

[0019] Use L max , T max , ε, and Sand r to construct and train SCN, and train the dataset;

[0020] Step 3: Calculate the fitness of the sand cat:

[0021] Use the root mean square error (RMSE) as the fitness function of SC - SCN, and RMSE can be calculated as:

[0022]

[0023] where, y i is the true value of the i - th data point, is the output of SCN for the i - th data point, and N represents the total number of samples;

[0024] Step 4: Update the parameter r:

[0025] Step 5: Obtain the final parameter r:

[0026] Repeat steps 2 to 5 until SCN meets the set conditions, so as to obtain the optimal value of r and the minimum fitness;

[0027] Step 6: Obtain the output of SCN:

[0028] After calculating the optimal value of r, train SCN to obtain the battery RUL prediction value.

[0029] Furthermore, the dataset training method in the above step 2 is:

[0030] Assume that L - 1 SCNs with hidden nodes have been constructed, and the output results are as follows:

[0031]

[0032] where g represents an activation function β represents the weight of the output, and the input of the model training dataset is X = {x1, x2, …, x N}, x i = [x i,1 , …, x i,d T ; w j and b j are the random weights and biases of the j - th node:

[0033] w j = λ × (2 × rand(n, T max )) - 1) (3)

[0034] b j = λ × (2 × rand(1, T max )) - 1) (4)

[0035] T max is the maximum value of the candidate nodes, and the SCN is updated:

[0036] Y L = Y L-1 +β L g L (5)

[0037] Calculate the error of the current network e L-1 :

[0038] e L-1 = Y - Y L-1 (6)

[0039] Before reaching the error range, adding a hidden node in the SCN introduces an inequality constraint, as shown in Equation (7), which adopts a self - monitoring mechanism to allocate hidden - layer parameters:

[0040]

[0041] where h L (X) = [g L (x1), g L (x2), …, g L (x N )] T , h L (X) represents the output of the L - th hidden node after input activation; r ∈ (0, 1) is a random regularization parameter; μ L ​Is a given sequence that satisfies It is a series of non - negative real numbers, ξ L The maximum value in Equation (7) is used to calculate the parameters of node L, which is related to the candidate nodes;

[0042] β is the output weight of the hidden layer, calculated based on the least - squares method:

[0043]

[0044] In Equation (8), ||·|| F Represents the Frobenius norm, H L =[h1,h2,…,h L is an inverse matrix, representing the output of the hidden layer;

[0045] There are two ways to end the network training, namely setting the error limit e or the maximum number of hidden nodes L max ; Adopt the method of early stopping. If one of the termination conditions in Equation (9) is met, finally determine the network structure and parameters of the SCN, and obtain the relevant output according to these parameters and the test data set. Otherwise, continue to add nodes.

[0046]

[0047] Furthermore, the method of updating the parameter r in Step 4 above is: update r through Equations (12) and (16):

[0048]

[0049]

[0050] In Equation (12), Represents the sensitivity range of each sand cat, S M =2 represents the maximum value of the sand cat's hunting range (which can be appropriately adjusted under different problems), iter c Represents the current iteration, iter Max Is the maximum number of iterations; Represents the current position of the cat's capture, Represents the optimal position of the cat's capture;

[0051] In Equation (16), Is a random position of the cat's capture to ensure that the sand cat randomly explores the prey;

[0052] Is the optimal solution of the cat's capture, Is the current position of the cat's capture, Indicates the variation range of R in the linear decline during the iterative process

[0053] Furthermore, the set condition in Step 5 above is Equation (9).

[0054] Advantages of the present invention: Compared with the prior art, the effects of the present invention are as follows:

[0055] 1) The present invention uses a health factor that integrates multi-source degradation features, combined with an adaptive random configuration algorithm to accurately predict the remaining useful life (RUL) of lithium-ion batteries; uses differential thermovoltammetry to extract multi-source degradation features from surface temperature and terminal voltage data, and obtains HI preprocessing with adaptive noise, effectively eliminating the capacity regeneration phenomenon of the fused data and the influence of noise on the prediction process.

[0056] 2) The present invention uses the meta-heuristic algorithm sand cat swarm optimization to optimize the regularization parameters in the random configuration network, and obtains the parameter adaptive incremental neural network SC-SCN. The experimental results show that SC-SCN applies the fused multi-source degradation characteristics and is an effective multi-step advance prediction algorithm for battery RUL. This constructive incremental learning network can better meet the real-time data requirements of the battery RUL prediction workflow industry. Adding an adaptive optimization algorithm can effectively improve the flexibility and prediction efficiency of the prediction model. Constructing an incremental neural network can effectively meet the industrial requirements of the battery RUL prediction process. SC-SCN can effectively achieve multi-step advance prediction of battery RUL, and can obtain good prediction performance even with only a small amount of training data, and the prediction error is significantly smaller than that of the comparison network. Brief Description of the Drawings

[0057] Figure 1 It is the differential process diagram of the DTV algorithm;

[0058] Figure 2 It is the flowchart of the CEEMDAN algorithm;

[0059] Figure 3 It is the incremental construction process diagram of SCN;

[0060] Figure 4 It is the schematic diagram of the population initialization of SCSO;

[0061] Figure 5 It is the optimization process diagram of SCSO;

[0062] Figure 6 It is the flowchart of the SC-SCN algorithm;

[0063] Figure 7 It is the battery life decay process diagram of the NASA battery dataset;

[0064] Figure 8 It is a voltage degradation comparison chart of B#5 battery;

[0065] Figure 9 It is a temperature degradation comparison chart of B#5 battery;

[0066] Figure 10 It is a health factor chart;

[0067] Figure 11 It is a similarity comparison chart between the health factor obtained by the DTV algorithm with different step sizes and the relationship between battery life degradation;

[0068] Figure 12 It is a comparison chart of the health factor after being processed by EMD and CEEMDAN;

[0069] Figure 13 It is a multi-step ahead prediction result and error chart of #B5;

[0070] Figure 14 It is a multi-step ahead prediction result and error chart of #B6;

[0071] Figure 15 It is a multi-step ahead prediction result and error chart of #B7. Detailed implementation manner

[0072] The present invention will be further introduced below in conjunction with specific embodiments.

[0073] Embodiment 1: As Figures 1-15 shown, an adaptive incremental battery RUL prediction method combining multi-source degradation characteristics, the method includes the following steps:

[0074] Step (1), in the existing algorithms based on deep learning, the insufficient degradation characteristics of the model due to a single health metric is a serious problem. In order to fully integrate the degradation characteristics from multiple data sources to characterize the attenuation of RUL, this application adopts the differential thermovoltammetry (DTV) that combines the battery terminal voltage and surface temperature, and uses the differential thermovoltammetry to extract the multi-source degradation characteristics in the surface temperature and terminal voltage data, and obtains the HI preprocessing with adaptive noise;

[0075] The detailed method is: by calculating the entropy increase ratio of the battery surface temperature and terminal voltage during the constant current charge and discharge process of the battery, the DTV is obtained, and the calculation process is as Figure 1 shown, which can fully reflect the thermodynamic change process during the battery decline process, and the discrete expression of DTV is as formula (1):

[0076]

[0077] where T, t, V, and k represent temperature, time, voltage, and time step, respectively, and V (t) represents the terminal voltage at time t, and T v(t) is the surface temperature.

[0078] It can be seen from Equation (1) that the DTV effectively integrates the battery aging information contained in the multi-sensor data features. In the differential calculation of multi-sensor data, the value of t directly affects the eigenvalue of the DTV function; if the value of t is too small, the meaning of entropy increase is lost, and if the value of t is too large, many repeated feature points are included, resulting in overfitting. In subsequent experiments, univariate experiments will be conducted, and the maximum correlation between the calculated HI and battery life will be calculated to determine the optimal value.

[0079] Step (2), using the meta-heuristic algorithm sand cat swarm optimization to optimize the regularization parameter in the random configuration network to obtain the parameter adaptive incremental neural network SC-SCN;

[0080] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) solves the "mode mixing" and "end point problems" of EMD and EEMD by adding Gaussian white noise. Different from EEMD, in EEMD, after each order component of the IMF component is further averaged and iterated in the residual, white noise is added to the IMF. Compared with EMD and EEMD, it can obtain better computational efficiency and modal decomposition results. The decomposed IMFs have better integrity and less interference from low-frequency components.

[0081] The calculation process of CEEMDAN is as Figure 2 shown. Taking a non-linear signal as the input x and setting the average value of a single IMF as t; adding Gaussian white noise (n1……nt) to the decomposed signal to obtain a new signal x, and performing EMD decomposition; the overall average of the generated t modal components gives the component IMF1 and the residual r1. Then, positive and negative paired Gaussian white noise is added to r1 to obtain a new signal. The new signal is subjected to EMD decomposition to obtain the first-order modal component E k (*)(E k (*) is the k-th IMF component generated by the EMD algorithm), thus obtaining the second characteristic mode component, and repeating the above steps until the obtained residual signal is a monotonic function.

[0082] Sand Cat Swarm Optimization (SCSO) is a meta-heuristic algorithm that simulates the wild survival habits of sand cats using low-frequency noise detection capabilities. SCSO consists of three main steps: population initialization, prey search, and prey attack. SCSO also proposes a mechanism for balancing exploration and exploitation. SCSO has the advantages of excellent performance for single-objective optimization problems, high complexity, few parameters and operators, and easy implementation. This algorithm assumes that sand cats are a social animal.

[0083] According to the foregoing description of SCN, it is necessary to randomly select the input node weights and hidden node biases. These parameters directly determine the model performance of SCN, and SCN is highly correlated with the hyperparameter r. However, the setting of the parameter r is related to the input data, and manual adjustment of the parameter is too time-consuming and laborious for real-time applications.

[0084] To improve the prediction accuracy and efficiency, the meta-heuristic algorithm SCSO is introduced to dynamically seek the optimization of the regularization parameter r, and the parameter r adaptive prediction model SC-SCN is obtained. SCSO has obvious advantages in single-objective optimization. It can combine low complexity and high accuracy when searching for the regularization parameter of thiocyanate.

[0085] The implementation of SC-SCN is divided into two steps: the establishment of SCN and the optimization of parameters. As described above, DTV is used to fuse the battery surface temperature and terminal voltage, and then CEEMDAN is used for feature extraction and denoising preprocessing of HI. The HI data matrix obtained after preprocessing is used as the input of SC-SCN for model training. As Figure 6 shown, the detailed steps of SC-SCN are as follows:

[0086] Step 1: Parameter initialization:

[0087] Initialize the overall size of Sand Cat (pop), iteration (iter), the upper and lower bounds of the search hyperparameters (ub and lb), dimension (D), the maximum number of hidden nodes (L max ), candidate nodes (T max ), fault tolerance (ε), and initialize the parameter value r within the given range [lb, ub] as Sand r ;

[0088] As Figure 4 shown, in a D-dimensional optimization instance, the position of the sand cat is a D-dimensional matrix representing the solution of the instance, and the size of the sand cat population (N pop ×N d ), (pop = 1,..., n), (x1, x2,..., xd) are all floating-point numbers. The adaptive value of the sand cat is determined by a custom adaptive function. After each iteration, the sand cat will output the corresponding fitness value. By selecting the sand cat with the best cost, the hunting plan (direction, distance) is closest to the prey. The remaining sand cats will search for the best sand cat in the next iteration in terms of direction.

[0089] Step 2: Construct and train the SCN:

[0090] Use L max , T max , ε and Sand r to construct and train the SCN and train the data set;

[0091] The data set training method in Step 2 is as follows:

[0092] Applying a prediction model based on deep learning to the process industry is challenging, especially for those industries with real-time requirements, to cope with flexible and variable battery real-time applications. Battery RUL prediction has very strict requirements for real-time performance. The prediction method based on deep learning cannot correctly determine the appropriate structure of the neural network (the number of hidden nodes), which makes it impossible for the deep learning model to learn and generalize good performance simultaneously. A network structure that can vary with the amount of data is needed to adapt to the continuously updated degradation data during the battery's service life. Such a network is called incremental learning.

[0093] Incremental learning has strong network flexibility and can effectively handle the increase in degradation data in battery RUL prediction.

[0094] Generally, the construction structure is trained using a two-step training paradigm. First, a random assignment algorithm is used to set the input weights and the biases of the hidden nodes. Second, the least squares method is used to calculate the output weights, and the least squares method is used to calculate the output weights, and its improved version is often adopted.

[0095] The SCN constructs a network structure by introducing a random method based on a supervision mechanism. This random method is feasible and effective in large-scale data calculations. The SCN has the advantages of less manual intervention in network scale setting, adaptation to the range of random parameters, and fast learning. Its greatest contribution is the ability to assume random parameters using inequality constraints and adaptively select the range of random parameters.

[0096] Assume that L - 1 SCNs with hidden nodes have been constructed, and the output results are as follows:

[0097]

[0098] where g represents an activation function β represents the output weight, and the input of the model training data set is X = {x1, x2, …, x N}, x i = [x i,1 , …, x i,d T ; w j and b j are the random weight and bias of the j-th node:

[0099] w j = λ × (2 × rand(n, T max )) - 1) (3)

[0100] b j = λ × (2 × rand(1, T max )) - 1) (4)

[0101] T max is the maximum value of the candidate nodes, SCN update:

[0102] Y L = Y L-1 + β L g L (5)

[0103] Calculate the error of the current network e L-1 :

[0104] e L-1 = Y - Y L-1 (6)

[0105] Before reaching the error range, adding a hidden node in the SCN introduces an inequality constraint, as shown in Equation (7), which adopts a self-monitoring mechanism to allocate hidden layer parameters:

[0106]

[0107] where h L (X) = [g L (x1), g L (x2), …, g L (x N )] T , h L (X) represents the output of the L-th hidden node after input activation; r ∈ (0, 1) is the random regularization parameter; μ L is a given sequence that satisfies It is a series of non-negative real numbers, ξ L The maximum value in Equation (7) is used to calculate the parameters of node L, which is related to the candidate nodes;

[0108] β is the output weight of the hidden layer, calculated based on the least squares method:

[0109] ​

[0110] In Equation (8), ||·|| F represents the Frobenius norm, and H L = [h1, h2, …, h L is an inverse matrix, representing the output of the hidden layer;

[0111] There are two ways to end the network training, that is, setting the error limit e or the maximum number of hidden nodes L max ; By adopting the method of early stopping, if one of the termination conditions in Equation (9) is satisfied, the network structure and parameters of the SCN are finally determined, and the relevant output is obtained according to these parameters and the test data set. Otherwise, please continue to add nodes. The training process of the SCN is as Figure 3 shown, which intuitively illustrates the construction method of the SCN.

[0112]

[0113] Step 3: Calculate the fitness of the sand cat:

[0114] Use the root mean square error (RMSE) as the fitness function of the SC - SCN. The RMSE can be calculated as:

[0115]

[0116] where, y i is the true value of the i-th data point, is the output of the SCN for the i-th data point, and N represents the total number of samples;

[0117] Step 4: Update the parameter r:

[0118] The sand cat can hear low frequencies below 2 kHz and can search for various prey on and under the desert ground. By setting the value of the variable (Equation (10)) linearly decreases from 2 to 0. The gradual linear change ensures that the sand cat population does not lose or miss prey; S M = 2 represents the maximum value of the sand cat's hunting range (which can be appropriately adjusted under different problems), iter c represents the current iteration, and iter Max is the maximum number of iterations; represents the current position of the cat's capture, represents the optimal position of the cat's capture.

[0119]

[0120] The method for updating the parameter r is: update r through Equations (12) and (16):

[0121]

[0122] The search space of the prey is obtained by random initialization from the defined boundaries. After obtaining the initial random positions, the optimal value of the current population is calculated. Each sand cat will update its state according to its current position, the best sand cat position, and its detection sensitivity (Equation (12)). This optimal search not only ensures that each updated position is close to the optimum, but also the randomness of the calculation guarantees the low cost and complexity of the algorithm;

[0123]

[0124] After finding the prey, the sand cat group enters the attack phase. Assume that the attack limit of each sand cat is circular, which can be obtained by defining a random angle θ, and the random range of θ ∈ [0, 360] is defined as [-1, 1]; SCSO selects the attack angle for each sand cat through the roulette wheel selection algorithm, so that each sand cat attacks along the new random angle in the search space attack, effectively avoiding the occurrence of local optimum conditions. The position updates during the attack are shown in Equations (13) and (14).

[0125]

[0126] When reaching a specific position, an adaptive parameter R is defined to achieve the transition of the sand cat group from search to attack, which is defined as:

[0127]

[0128] When linearly decreasing from the iterative process, R randomly varies within the range of [-2rG, 2rG]. When |R| > 1, the sand cat is searching for prey, and when |R| ≤ 1, the state of the sand cat can be search or attack, as shown in Equation (16)

[0129]

[0130] In Equation (12), represents the sensitivity range of each sand cat. This setting can avoid local optimal capture. By setting the sensitivity ranges of different sand cats (receiving low-frequency signals from the prey) to different random values, it can effectively avoid falling into local optimum; S M = 2 represents the maximum value of the sand cat hunting range (which can be appropriately adjusted under different problems), iter c represents the current iteration, iter Max is the maximum number of iterations; represents the current position of the cat's capture, represents the optimal position of the cat's capture;

[0131] In Equation (16), It is a random position for the cat to catch prey, ensuring that the sand cat randomly explores the prey;

[0132] It is the optimal solution for the cat to catch prey, It is the current position of the cat to catch prey, It represents the change range of R in the linear descent during the iterative process

[0133] As shown in SCSO Figure 5 As shown, during the search phase (|R|>1), the sand cat group searches for prey in several different random directions and obtains the optimal position after t iterations The rest Using the search parameter to approach the prey to obtain iteration t+1; when |R|≤1, the sand cat enters the attack (or search) phase at a random speed and angle. The random variables of the sand cat are shown in the figure. Taking a confident sand cat as an example, the solid line represents that the sand cat receives low-frequency signals with both ears, the dotted line represents the search direction of the sand cat, and θ is the angle between the search direction and the prey. It should be noted that due to the random mechanism of SCSO, the sand cat at the optimal position in the iteration is not at the optimal position in the next iteration It can effectively avoid local optimality.

[0134] Step Five: Obtain the final parameter r:

[0135] Repeat Step Two to Step Five until SCN satisfies Equation (9), so as to obtain the optimal value of r and the minimum fitness;

[0136] Step Six: Obtain the output of SCN:

[0137] After calculating the optimal value of r, train SCN to obtain the predicted value of the battery RUL.

[0138] Step (3), use the parameter adaptive incremental neural network SC-SCN obtained in Step (2) to predict the adaptive incremental battery RUL.

[0139] To verify the effect of the present invention, the following verification is carried out:

[0140] 1.1 Battery dataset

[0141] The NASA Ames lithium-ion battery data is used in this case study. The operating data summary is shown in Table 1.

[0142] Table 1 Battery dataset

[0143]

[0144] The NASA dataset contains several compressed files in the ".mat" format, which record the aging information of 18650 LIBs under different operating conditions by simulating the accelerated aging process of LIBs through repeated charge-discharge cycles. During charging, the battery is first charged in a constant current (CC) mode at 1.5 A until the terminal cut-off voltage reaches 4.2 V, and then switched to a constant voltage (CV) mode to continue charging until the charging current drops to 20 mA, ending the charging process. The discharge experiment is carried out at a constant current of 2 A for 10 minutes. This investigation will be cancelled after the battery degrades to the end of its service life (EOL), i.e., the rated capacity drops by 30%. Dataset B#5, B#6, and B#7 are selected in this application.

[0145] The capacity degradation images of the NASA battery dataset are as Figure 7 shown. It can be clearly seen from Table 1 and Figure 7 that although the degradation experiments were carried out in the same experimental environment, due to the different internal structures of the cells, there are different aging processes.

[0146] 1.2 Differential Thermal Voltammetry

[0147] Accurate prediction of the battery's RUL requires obtaining an effective HI from the battery aging data. The changes in the charge-discharge cycle parameters of Battery B#5 for the first (Cycle 1) and last (Cycle 168) cycles are as Figure 8 and Figure 9 shown.

[0148] It can be seen from Figures 8-9 that battery aging affects the changes in battery voltage and surface temperature after multiple charge-discharge cycles, indicating that the changes in both are closely related to battery degradation. How to obtain an HI with a large number of degradation information sources as the actual input of this model has always been an important research direction in this industry. This fusion factor can effectively solve problems such as insufficient prediction accuracy and poor model generalization caused by existing single degradation variables. Therefore, in order to fully integrate information from multiple data sources to characterize the attenuation of RUL, this application uses the previously proposed DTV. The HI for a single charge-discharge cycle calculated by DTV is as Figure 10 shown.

[0149] The magnitude of the differential step size t is crucial in the fusion calculation of DTV. An appropriate value is needed to make the resulting fusion HI have the highest correlation with capacity degradation. Univariate experiments were used to obtain the HI at different step sizes and compare their correlations with the degraded capacity respectively. As shown in the correlation analysis of Figure 10 , the HI obtained at t = 20 ms has the strongest correlation with the battery life decay curve. Finally, the voltage value corresponding to the valley of the t = 20 ms tDTV curve is selected as the HI.

[0150] 1.3 CEEMDAN for HI preprocessing

[0151] The HI calculated by DTV combining multiple source degradation information is a non - linear, non - smooth time - varying signal; it is necessary to construct a mode decomposition method to effectively decompose the HI sequence. After the mode decomposition of the battery degradation information sequence, the degradation characteristics of HI can be further extracted, and the phenomena of signal denoising and capacity regeneration can be eliminated.

[0152] CEEMDAN effectively solves the problems of "mode fixation" and "end - point" in EMD, and manages to eliminate the influence of residual white noise on the decomposition of the signal sequence. In this subsection, HI is preprocessed using CEEMADN. The preprocessed HI reduces the interference of noise and capacity regeneration phenomena on the prediction model, facilitating the model to analyze more degradation characteristics. The components obtained by preprocessing are compared with the components obtained by EMD as Figure 12 shown.

[0153] 1.4 Comparative experiments and analysis

[0154] Battery usage is a real - time application. Accurately predicting the battery RUL can effectively prevent various failures during operation. The earlier the battery aging problem is detected and the more availability predictions are made, the safer its use can be ensured, which helps the health maintenance of related equipment. In this subsection, based on the NASA battery dataset, the SC - SCN model is used to achieve the early prediction of the battery RUL, and RMSE and MAE are recorded in tabular form to evaluate the prediction accuracy (average of 10 prediction results). The definition of the MAE index is as follows:

[0155]

[0156] where y k is the actual value of the RUL at the k - th data point, is the model output at the k - th data point, and N is the total number of samples.

[0157] The average value of the experimental results exceeds 10 predictions. The results and errors of the multi - step ahead prediction of RUL based on the SC - SCN model are as Figure 13 、 14, as shown in Figure 15. For the NASA battery datasets B#5, B#6, and B#7, the RUL was predicted in advance by 100, 90, 80, 70, 60, 50, 40, and 30 cycles. It can be seen from the figure that the algorithm processed the HI extracted by the DTV algorithm, effectively eliminating the influence of capacity regeneration and noise on the RUL prediction. The degree of projection fitting gradually increases with the increase in training data, which is a basic feature of neural networks. It should be noted that even in the early stage with only a small amount of prediction data, the proposed SC-SCN also has good prediction results. The error between the prediction result and the appropriate life degradation is visualized by a box plot, and its most significant advantage is that it can accurately and stably represent the discrete distribution of data without being affected by outliers. The dotted line in the box plot represents the average value of the error, and the height of the rectangle represents the fluctuation range of the error. From Figure 13 , 14 and 15, it is easy to find that SC-SCN has a small error in the early stage of multi-step ahead prediction, showing a good fit, indicating that the SC-SCN algorithm still has excellent approximation ability even with only a small amount of training data. With the increase in training data, the fluctuation of the prediction error gradually decreases, and the maximum error gradually decreases. Taking B#5 as an example, the maximum error gradually decreases from 0.0729 at 100 prediction cycles to 0.0203 at 30 prediction cycles. The ability to obtain such results shows that the same SC-SCN model still has strong non-linear prediction ability in the prediction stage, and there is no need to train a new model to adapt to data changes and feature addition. Due to the constructive network structure of SCN, it is very suitable for real-time process-oriented industrial requirements similar to battery RUL prediction.

[0158] The prediction performance evaluations are shown in Tables 2, 3, and 4. Under the same conditions, DNN, SVR, and FNN were introduced for multi-step ahead prediction comparison experiments. As shown in the tables, with the increase in the amount of training data, the RMSE and MAE of the prediction gradually decrease because SC-SCN learns more degradation information from the increasing data. The network currently has strong degradation fitting ability and low error. The prediction error and the fitting results of the degradation images show that the proposed SC-SCN is a more effective and accurate battery RUL prediction algorithm. When there is only a small amount of training data (100-step, 90-step, 80-step ahead prediction) and the training data gradually increases, SC-SCN has high prediction performance.

[0159] 1 Evaluation of the multi-cycle ahead RUL prediction performance of B5 battery

[0160]

[0161] 2 Evaluation of the multi-cycle ahead RUL prediction performance of B6 battery

[0162]

[0163] Table 3 Performance evaluation of multi-cycle early RUL prediction for B#7 battery

[0164]

[0165]

[0166] SC-SCN has the best overall prediction performance in B#5 battery. Except that the 40-step-ahead prediction is greater than that of DNN, the multi-step-ahead prediction errors are all smaller than those of the control network. In the small-sample prediction experiments of B#6 and B#7 batteries, SVR has smaller errors in the 100-step and 90-step-ahead predictions. However, in the later degradation after the training data increases, the prediction performance may be better than that of SC-SCN, which has better prediction stability. Compared with other artificial neural networks, the proposed SC-SCN has excellent performance because of the constructed network structure of SCN and the mechanism of strong network generalization; adding the meta-heuristic algorithm SCSO enables the hyperparameters in SCN to be adaptive, improving the prediction performance of the network under real-time battery applications.

[0167] Based on the analysis of the above experimental data and results, it can be concluded that the proposed SC-SCN is an efficient and flexible incremental battery RUL prediction algorithm, and combining SC-SCN with multi-source degradation characteristics can effectively meet the real-time application requirements of the multi-step-ahead battery RUL prediction process.

[0168] Under the strong demand for battery RUL prediction in the real-time industry, this paper adopts the adaptive incremental neural network SC-SCN to achieve advanced prediction of battery RUL. First, the multi-source characteristic information of battery voltage and temperature is calculated by differential thermovoltammetry (DTV) to obtain the fused HI describing the battery life degradation process, and the improved empirical mode decomposition algorithm - ceemdan is used to preprocess the HI data. Then, the sand cat swarm optimization (SCSO) is used to optimize the regularization parameters in SCN to obtain the parameter-adaptive prediction model SC-SCN; finally, the NASA battery datasets B#5, B#6, B#6, B#7 are used to verify this model, the prediction errors and images are given, and comparative experiments are carried out on various neural networks.

[0169] The results show that the constructed incremental neural network can effectively meet the industrial requirements of the battery RUL prediction process. SC-SCN can effectively achieve multi-step-ahead prediction of battery RUL, and can obtain good prediction performance even with only a small amount of training data, and the prediction errors are significantly smaller than those of the comparison networks.

[0170] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims described above.

Claims

1. An adaptive incremental battery RUL prediction method combining multi-source degradation features, characterized in that: The method includes the following steps: Step (1), using differential thermo-voltammetry to extract multi-source degradation features from surface temperature and terminal voltage data, and obtaining HI preprocessing with adaptive noise; Step (2), using the HI data matrix as the input of SC-SCN, and optimizing the regularization parameters in the random configuration network by the meta-heuristic algorithm sand cat swarm optimization to obtain the parameter adaptive incremental neural network SC-SCN; Step (3), using the parameter adaptive incremental neural network SC-SCN obtained in step (2) to predict the RUL of the adaptive incremental battery; The detailed steps in step (2) are as follows: Step 1: Parameter initialization: Initialize the overall size of Sand Cat, the number of iterations iter, the upper and lower bounds ub and lb of the search hyperparameters, the dimension D, the maximum number of hidden nodes L max , the candidate node T max , the fault tolerance ε, and initialize the parameter value r within the given range [lb, ub] as Sand r ; Step 2: Construct and train SCN: Use L max , T max , ε and Sand r to construct and train the SCN, and train the dataset; The dataset training method is as follows: Assume that L-1 SCNs with hidden nodes have been constructed, and the output results are as follows: where g represents an activation function β represents the output weight, and the input of the model training dataset is X = {x1, x2, …, x N}, x i = [x i,1 , …, x i,d T ; w j and b j are the random weight and bias of the j-th node:​ w j = λ × (2 × rand(n, T max ) - 1) (3) b j = λ × (2 × rand(1, T max ) - 1) (4) T max is the maximum value of the candidate nodes, SCN update: Y L = Y L-1 + β L g L (5) Calculate the error of the current network e L-1 : e L-1 = Y - Y L-1 (6) Before reaching the error range, adding hidden nodes in the SCN will introduce an inequality constraint, such as Equation (7), which adopts a self-monitoring mechanism to allocate hidden layer parameters: where h L (X) = [g L (x1), g L (x2), …, g L (x N )] T , h L (X) represents the output of the L-th hidden node after the input is activated; r ∈ (0, 1) is a random regularization parameter; μ L is a given sequence that satisfies It is a series of non-negative real numbers, ξ L The maximum value in Equation (7) is used to calculate the parameters of node L, which is related to the candidate nodes; β is the output weight of the hidden layer, calculated based on the least squares method: In Equation (8), ||·|| F represents the Frobenius norm, and H L = [h1, h2, …, h L is an inverse matrix representing the output of the hidden layer; There are two ways to end network training, namely, setting the error limit e or the maximum number of hidden nodes L max ; Adopt the method of early stopping. If one of the termination conditions in Equation (9) is satisfied, finally determine the network structure and parameters of the SCN, and obtain the relevant output according to these parameters and the test dataset. Otherwise, continue to add nodes; Step 3: Calculate the fitness of the sand cat: Use the root mean square error (RMSE) as the fitness function of SC-SCN, and RMSE can be calculated as: where y i is the true value of the i-th data point, is the SCN output of the i-th data point, and N represents the total number of samples; Step 4: Update the parameter r: Step 5: Obtain the final parameter r: Repeat steps 2 to 5 until the SCN meets the set conditions, so as to obtain the optimal value of r and the minimum fitness; Step 6: Obtain the output of the SCN: After calculating the optimal value of r, train the SCN to obtain the predicted value of the battery RUL.

2. The adaptive incremental battery RUL prediction method combining multi-source degradation features according to claim 1, characterized in that: The detailed method in step (1) is: by calculating the entropy increase ratio of the battery surface temperature and terminal voltage during the constant current charge and discharge process of the battery, DTV is obtained, and its discrete expression is as Equation (1): In the formula, T, t, V, and k represent temperature, time, voltage, and time step, respectively. V (t) represents the terminal voltage at time t, and T v(t) is the surface temperature.

3. The adaptive incremental battery RUL prediction method combining multi-source degradation features according to claim 1, wherein: The method for updating the parameter r in step 4 is: update r through Equations (12) and (16): In formula (12), represents the sensitivity range of each sand cat, S M = 2 represents the maximum value of the sand cat's hunting range, iter c represents the current iteration, iter Max is the maximum number of iterations; represents the current position of the cat's capture, represents the optimal position of the cat's capture; In formula (16), is a cat-catching random position to ensure that the sand cat randomly explores prey; Is the optimal solution for cat hunting, Is the current position of cat hunting, Indicates the change range of R during the linear decrease in the iteration process 4. The adaptive incremental battery RUL prediction method combining multi-source degradation features according to claim 1, characterized in that: The set condition in step 5 is Equation (9).

Images