Lithium battery residual life prediction method and system based on virtual sample generation

Abstract

The invention provides a lithium battery residual life prediction method and system based on virtual sample generation, and relates to the technical field of lithium batteries, and the method comprises the steps: extracting samples from a historical charging and discharging curve of a lithium battery, and forming an original small sample set; through a virtual sample generation method constructed based on a differential evolution algorithm, according to features and distribution of sample data in an original small sample set, high-similarity virtual samples are generated, and a virtual sample set is formed; based on the virtual sample set, constructing a training set to train a capacity prediction model; based on the iterative prediction of the capacity prediction model, when the predicted capacity is lower than a preset failure threshold value for the first time, the number of corresponding iterations is the number of charge-discharge cycles which can be continued when the target lithium battery reaches the end life, so that a prediction result of the remaining life of the lithium battery is obtained; through the newly proposed virtual sample generation method, the information amount of the original sample is effectively increased, and more accurate lithium battery capacity prediction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium batteries, in particular to a lithium battery remaining useful life prediction method and system based on virtual sample generation. BACKGROUND

[0002] Under the influence of the green transformation of the economic society, new energy vehicles, photovoltaic industries and other specific fields have considerable growth potential, which undoubtedly brings great opportunities for the development of lithium batteries. As a highly complex electrochemical system, lithium batteries will inevitably undergo a series of chemical reactions and physical structure changes with the continuous charging and discharging cycles, resulting in gradual reduction of their performance and service life. If the aging lithium batteries are not replaced in time, it will increase a series of safety hazards, which may cause serious damage to the battery, and even cause fire or explosion accidents. Therefore, in order to avoid similar accidents as much as possible, it is necessary to replace and maintain the battery in time, and accurate prediction of the remaining useful life (RUL) of the lithium battery is the key to improving the performance and enhancing the safety of the battery.

[0003] The RUL of the lithium battery refers to the number of charging and discharging cycles that the lithium battery can continue to perform from the current charging and discharging cycle state to the end of life. The RUL prediction process of the lithium battery is actually a process of predicting the future degradation trend of the lithium battery according to the actual degradation data of the lithium battery. Since the cycle life of the lithium battery can be as high as 10,000 times or more, it is undoubtedly a time-consuming and laborious challenge to obtain the data of the entire life cycle. In practical applications, it is often difficult to collect enough aging data in a short time, which makes the RUL prediction of the lithium battery inaccurate under the condition of small sample, and ultimately affects the judgment of the battery replacement and maintenance cycle. SUMMARY

[0004] In order to solve the above problems, the present application proposes a lithium battery remaining useful life prediction method and system based on virtual sample generation, which effectively increases the information quantity of the original sample by the newly proposed virtual sample generation method, and realizes more accurate lithium battery capacity prediction.

[0005] According to some embodiments, the present application adopts the following technical scheme: A lithium battery remaining useful life prediction method based on virtual sample generation, comprising: extracting samples from the historical charging and discharging curve of the lithium battery to form an original small sample set, wherein the samples include charging and discharging cycle times and corresponding health factors and capacities; generating high-similarity virtual samples according to the characteristics and distribution of the sample data in the original small sample set by a virtual sample generation method based on a differential evolution algorithm to form a virtual sample set; Based on the virtual sample set, a training set is constructed to train a capacity prediction model, the capacity prediction model taking a time sequence composed of health factors and capacities as input and taking health factors and capacities corresponding to the next charging and discharging as output; From the charging and discharging curve of the target lithium battery, health factors and capacities corresponding to the last preset charging and discharging are extracted to form a time sequence composed of health factors and capacities, which is input into the trained capacity prediction model. Based on the predicted health factors and capacities corresponding to the next charging and discharging, the time sequence is reconstructed, and iterative prediction based on the capacity prediction model is performed. When the predicted capacity first falls below the preset failure threshold, the number of iterations corresponds to the number of charging and discharging cycles that the target lithium battery can still perform when reaching the end of life, thereby obtaining the lithium battery remaining life prediction result.

[0006] According to some embodiments, the present application adopts the technical scheme as follows: A lithium battery remaining life prediction system based on virtual sample generation, comprising: An original sample construction module configured to extract samples from the historical charging and discharging curve of the lithium battery to form an original small sample set, the samples including charging and discharging cycle numbers and corresponding health factors and capacities; A virtual sample generation module configured to generate high-similarity virtual samples according to the characteristics and distribution of sample data in the original small sample set by a virtual sample generation method based on a differential evolution algorithm to form a virtual sample set; A prediction model construction module configured to construct a training set based on the virtual sample set to train a capacity prediction model, the capacity prediction model taking a time sequence composed of health factors and capacities as input and taking health factors and capacities corresponding to the next charging and discharging as output; A remaining life prediction module configured to extract health factors and capacities corresponding to the last preset charging and discharging from the charging and discharging curve of the target lithium battery to form a time sequence composed of health factors and capacities, input into the trained capacity prediction model. Based on the predicted health factors and capacities corresponding to the next charging and discharging, the time sequence is reconstructed, and iterative prediction based on the capacity prediction model is performed. When the predicted capacity first falls below the preset failure threshold, the number of iterations corresponds to the number of charging and discharging cycles that the target lithium battery can still perform when reaching the end of life, thereby obtaining the lithium battery remaining life prediction result.

[0007] According to some embodiments, the present application adopts the technical scheme as follows: A computer program product comprising a computer program, the computer program being executed by a processor to implement the lithium battery remaining life prediction method based on virtual sample generation.

[0008] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned method for predicting the remaining life of a lithium battery based on virtual sample generation.

[0009] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the aforementioned method for predicting the remaining life of a lithium battery based on virtual sample generation.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a Virtual Sample Generation (VSG) method based on an improved Differential Evolution (DE) algorithm. This method can generate highly similar virtual samples to expand the data volume according to the characteristics and distribution of sample data. The virtual samples are then fed into an Extreme Learning Machine (ELM) for RUL prediction and evaluation.

[0011] The virtual sample generation method proposed in this invention generates virtual samples that have the same overall trend as the original small sample data and the same temporal attributes within an acceptable range. These virtual samples serve as output data that are highly similar to the original input data in structure and features, thereby enriching the sample data. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0013] Figure 1 This is a diagram showing the relationship between virtual samples and real samples in Example 1. Figure 2 The flowchart is shown in Example 1, illustrating the improved DE algorithm. Figure 3 This is a flowchart of the lithium battery RUL prediction process based on virtual samples in Example 1.

[0014] Figure 4 This is an example diagram of the upper and lower limits of the extended range in Example 1.

[0015] Figure 5 This is a virtual sample image from Example 1.

[0016] Figure 6 The image shows the prediction results of real samples of battery B05 in Example 1, where (a) is the prediction result without virtual samples and (b) is the error comparison chart. Figure 7 The image shows the prediction results of the virtual sample for battery B05 in Example 1, where (a) is the prediction result with added virtual samples and (b) is the error comparison chart.

[0017] Figure 8 The image shows the prediction results of real samples of battery B06 in Example 1, where (a) is the prediction result without virtual samples and (b) is the error comparison chart.

[0018] Figure 9 The image shows the prediction results of virtual samples for battery B06 in Example 1, where (a) is the prediction result with added virtual samples and (b) is the error comparison chart. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in these embodiments have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0022] Example 1 Due to insufficient training data, machine learning algorithms often exhibit poor performance and struggle to meet prediction accuracy requirements. To address the unsatisfactory prediction results caused by insufficient training samples in data-driven methods, this embodiment proposes a lithium battery remaining life prediction method based on virtual sample generation, comprising: Step S1: Extract samples from the historical charge-discharge curves of lithium batteries to form an original small sample set. The samples include the number of charge-discharge cycles and the corresponding health factors and capacity. Step S2: Using a virtual sample generation method based on differential evolution algorithm, highly similar virtual samples are generated according to the characteristics and distribution of sample data in the original small sample set, forming a virtual sample set; Step S3: Based on the virtual sample set, construct a training set to train the capacity prediction model. The capacity prediction model takes the time series of health factors and capacity as input and the health factors and capacity corresponding to the next charge and discharge as output. Step S4: Extract the health factor and capacity corresponding to the most recent preset charge and discharge cycle from the charge and discharge curve of the target lithium battery, form a time series of health factors and capacity, and input it into the trained capacity prediction model. Based on the predicted health factor and capacity corresponding to the next charge and discharge cycle, reconstruct the time series and perform iterative prediction based on the capacity prediction model. When the predicted capacity is lower than the preset failure threshold for the first time, the corresponding number of iterations is the number of charge and discharge cycles that the target lithium battery can continue to perform when it reaches the end of its life, thereby obtaining the remaining life prediction result of the lithium battery.

[0023] The virtual sample generation method first uses an expansion function to determine the expansion range. Then, it generates multiple sets of random initial virtual samples within this range and uses them as the initial particle swarm for iterative optimization in the DE algorithm until the most feasible virtual sample is generated in the search space. When the individual fitness and global fitness of the particles meet the requirements, the optimal virtual sample is generated. By adding virtual samples to fill the information gaps between the original small sample data, the amount of training sample data is increased, thus improving the accuracy of the prediction model. In summary, the virtual sample generation method is reasonable and effective in solving the small sample prediction problem. The specific implementation process is described in detail below.

[0024] First, let's explain the virtual sample generation method based on the improved differential evolution algorithm: I. Virtual Sample Generation Method Virtual samples lack a clear definition, but researchers have summarized them as: samples obtained through some kind of transformation. , the original sample Convert to Then it is called The virtual sample generated through transformation is as shown in formula (1): (1) The essence of virtual sample generation algorithms is to incorporate prior knowledge into the learning process to generate new samples that match the feature space of the original samples, thus filling the information gaps between samples. As shown in Figure 1, the relationship between the actual data space, the small sample space, and the virtual data space is illustrated.

[0025] The virtual sample generation method proposed in this embodiment mainly includes three steps: determining the expansion range, generating random virtual samples, and finding the best virtual sample.

[0026] 1. Determine the scope of expansion.

[0027] For the original sample, determining the effective expansion range is crucial for the generation of subsequent virtual samples. This embodiment uses the Information Expansion Function Considering Sample Density and Amount (IEDA) to determine the expansion range. The specific process is as follows: A real data point in the original small sample set is Here, the subscript i represents the i-th battery. Let's assume the number of batteries is n. The subscript j indicates the j-th cycle. Here, we assume the battery has been charged and discharged P times. The subscript k indicates the k-th feature. These features include five elements: constant voltage rise charging time, constant voltage drop discharging time, average voltage change rate, constant-time discharge voltage, and capacity. The central observation location of the original small sample set Defined as a two-dimensional matrix, where The element in the j-th row and k-th column of a matrix is ​​defined as: (2) make and These are the minimum and maximum values ​​of the original small sample set, respectively, both of which are two-dimensional matrices. and The element in the j-th row and k-th column of a matrix is ​​defined as: (3) (4) Considering skewness and The relative quantity and density relationship of the observations on both sides will influence the left skewness matrix. and right skewness matrix Defined as: (5) (6) in, and This represents the element in the j-th row and k-th column of the matrix. and These are all samples ( )exist The number of samples on the top and bottom sides; and The weights are used to determine the upper and lower boundary matrices of the extended range; therefore, the upper and lower boundary matrices of the extended range can be obtained. , This can be expressed as a formula: (7) (8) 2. Random virtual sample generation.

[0028] The previous step used the IEDA function to determine the extended range. Within the extended range, multiple sets of virtual samples were generated using Latin Hypercube Sampling (LHS). The core of LHS lies in the hierarchical technique, which divides the range of sample input attributes into multiple equal small intervals. Each small interval has the same probability of being sampled, which can effectively avoid the problem of sample "clustering" and expand the spatial coverage of the samples.

[0029] 3. Find the best virtual sample.

[0030] Since the generated random samples cannot guarantee that all of them meet the individual fitness and global fitness requirements, in order to improve the quality of virtual samples, this embodiment uses the Differential Evolution (DE) algorithm to process the virtual samples. DE is a heuristic optimization algorithm that generates new individuals by using the difference vectors between individuals in the population. Normalized random particles are used as the initial particle swarm for the differential algorithm for iterative optimization.

[0031] When the individual fitness and global fitness of particles within each feature range meet the requirements, it indicates that the optimal virtual sample has been found. The advantage of the Differential Evolution (DE) algorithm lies in its simplicity and good performance in handling nonlinear, nonconvex, and high-dimensional problems. The core operation of the algorithm is to create new individuals through mutation, and then use the fitness function to evaluate the quality of the individuals, thereby guiding the search process. Considering that lithium battery degradation data has temporal characteristics (dynamic patterns and correlations that change with the number of charge-discharge cycles or time), the traditional DE algorithm lacks the ability to handle temporal dependencies and is difficult to adapt to this characteristic, so it is not suitable for the virtual sample generation process of lithium batteries. Therefore, it is necessary to improve the traditional DE algorithm by designing the fitness function and using the mutation and selection operations of the differential evolution algorithm to ensure that the deviation between the virtual sample and the real sample is less than 1%. The specific process is shown in Figure 2. Definition of the first A virtual sample It is a two-dimensional matrix, where the rows represent the number of charge-discharge cycles and the columns represent the feature dimensions. The real value data in the j-th row and k-th column is ,in, Let represent the i-th sample, j represent the j-th charge / discharge cycle, and k represent the k-th feature. (1) Initialization: - Define the fitness function: The fitness function is directly linked to the degree of deviation between virtual and real samples; therefore, the fitness function is defined as follows: , indicating the first The fitness value of the overall trend of a virtual sample compared to the trend of the original small sample data is essentially a quantification of the degree of deviation between the virtual sample and the real sample: (9) in, Indicates the original sample number. During the first charge and discharge cycle The standard deviation of each dimension.

[0032] - Population initialization: Initial population defined as (10) in, This represents the initial matrix obtained after the 0th iteration of the i-th virtual sample. and The maximum and minimum values ​​of the samples from formulas (3) and (4), respectively. Let represent a random number that follows a uniform distribution. Generate an initial matrix after the 0th iteration by generating several (e.g., m) virtual samples according to formula (10).

[0033] The initial population is normalized using the following formula to scale the data values ​​of each feature dimension to the same range, avoiding optimization bias caused by differences in units: (11) Where i represents the sample number, j represents the number of charge-discharge cycles, and k represents the feature dimension number. It is the theoretical minimum value of the k-th dimension. It is the theoretical maximum value of the k-th dimension.

[0034] The fitness value of the initial virtual samples is calculated according to the fitness function shown in formula (9). Then, the virtual samples are sorted in ascending order according to the size of the fitness value. The sorted virtual samples have their own ranking index. For virtual samples, fitness is... The smaller the value, the better the virtual sample; virtual samples with a fitness of less than 1 are defined as dominant virtual samples, and virtual samples with a fitness of greater than 1 are defined as inferior virtual samples. (2) Variation: Improvements to the mutation strategy of the traditional DE algorithm: (12) in, Indicates the number of iterations. Indicates the first The virtual sample at the th The mutation matrix obtained after the nth iteration Indicates the first The matrix of a virtual sample after the g-th iteration and From the first Selected from the superior virtual samples after the next iteration. and From the first Selected from the inferior virtual samples after the next iteration; It is a scaling factor, which is generally chosen between [0,2], and is usually 0.5. The main impact is on the algorithm's global localization capability. The smaller the value, the stronger the local search capability; The larger the value, the better the algorithm can escape local minima, but the convergence speed will be slower. Since the fitness function is defined based on the degree of deviation, if the deviation between the virtual sample and the real sample is less than 1%, the fitness of the generated virtual sample is less than 1. The algorithm distinguishes between superior and inferior virtual samples based on whether the fitness is less than 1, and performs mutation based on the difference between the superior and inferior virtual samples to guide the virtual samples to evolve in the direction of fitness less than 1. Finally, through the selection operation, samples with better (smaller) fitness are continuously selected. When the number of iterations is large enough, the fitness can be made less than 1.

[0035] (3) Crossing: definition For the i-th virtual sample in the th... The cross matrix obtained after the iteration is used to generate a new solution through cross operations: (13) in, This represents a random number that follows a uniform distribution. This represents the crossover probability.

[0036] (4) Choose: For each virtual sample The fitness is less than or equal to fitness: (14) in, It is the first The matrix generated by the crossover operation in the next iteration It is the fitness function, and the calculation method is shown in formula (9).

[0037] In this embodiment, the maximum number of iterations is set to 100. When the maximum number of iterations is reached, the optimal virtual sample is output, i.e. .

[0038] As can be seen, to achieve a deviation of less than 1% between virtual and real samples, it is accomplished through two parts: 1) The 1% deviation degree is directly linked to the fitness function. The fitness function is defined as the fitness value between the overall trend of the i-th virtual sample data sequence and the trend of the original small sample data. In essence, it is to quantify the deviation degree between the virtual sample and the real sample.

[0039] 2) Mutation and selection ensure that the virtual samples evolve towards a fitness (i.e., deviation) of less than 1% during the iterative process. In the mutation operation of the improved difference algorithm proposed in this example, the mutation vector is calculated to distinguish between dominant samples (fitness less than 1) and inferior samples (fitness greater than 1), thus causing the fitness of the mutated samples to evolve towards a fitness of less than 1. Finally, the selection operation ensures that the fitness value of the generated samples is minimized.

[0040] The following describes a lithium battery RUL prediction method based on virtual sample generation: II. Lithium-ion Battery RUL Prediction Method Based on Virtual Sample Generation Based on existing research, let... For the obtained sample size, if If the value is less than a certain predetermined value, the sample is considered a small sample; this predetermined value is generally considered to be 50. This embodiment adds virtual samples to fill the information gaps between the original small sample data, increasing the amount of training sample data and achieving long-term accurate prediction of the remaining lifespan of lithium batteries. Figure 3 As shown, the specific implementation steps mainly include three parts: data preprocessing, virtual sample generation, and lithium battery RUL prediction.

[0041] (1) Construction of original samples To better determine the expansion range of the generated virtual samples, it is necessary to construct a multidimensional and effective dataset as the original small sample set. In this embodiment, by analyzing the historical charge and discharge curves of lithium batteries, four indirect health factors and capacity corresponding to each charge and discharge cycle are extracted: constant voltage rise charging time, constant voltage drop discharge time, average voltage change rate, and constant time discharge voltage. These are used as an original sample to form the original small sample set.

[0042] (2) Virtual Sample Generation The data space of the original small sample set obtained in the previous step is expanded using the IEDA function to obtain the upper limit of the expansion. and lower limit That is, to extend the scope. Figure 4 In the diagram, the blue solid line represents the lower limit of expansion, the black solid line represents the upper limit of expansion, and the red solid line represents the actual value, i.e., the capacity data. (This is achieved through...) Figure 4 It can be observed that, without considering a few fluctuating samples, the actual capacity curve lies perfectly within the extended range. This phenomenon indicates that, for degenerate time series containing random fluctuations, the extended range determined by the IEDA method not only fully preserves the information of the original data, but also generates highly similar data by learning the characteristics and distribution of the sample data, thereby enriching the number of samples and effectively enriching the information content of the sample set.

[0043] To obtain multidimensional virtual samples of sample data sequences with time attributes, 100 virtual samples are generated using LHS within an extended range. The core of LHS lies in the hierarchical technique, which divides the range of sample input attributes into multiple equal small intervals, with each small interval having the same probability of being sampled. This effectively avoids the sample "clustering" problem and expands the spatial coverage of the samples.

[0044] In summary, the specific steps for generating virtual samples are as follows: - Use formulas (2), (7), and (8) to obtain data such as CL, LB, and UB; - Using LHS, the probability range of the sample is divided into 100 equal intervals. A point is randomly selected from each interval, resulting in a sequence containing 100 points. Therefore... , where vector If satisfied , then refers to ; - Normalize 100 sets of random virtual samples and use them as the initial particle swarm for iterative optimization of DE.

[0045] When the individual fitness and global fitness of each virtual sample meet the requirements At that time, the first time point of all time points The sample data set is regarded as a new virtual sample data sequence, that is, a virtual sample set. The finally generated virtual samples are as Figure 5 shown.

[0046] In the process of generating virtual samples, by controlling the value of within 1%, that is, the deviation between the virtual sample and the real sample is less than 1%, it can also be proved that the generated virtual samples have certain practical significance and can accurately represent the inherent characteristics of the lithium battery capacity data. This method can constrain the generated virtual samples within an ideal confidence interval, accurately capture the inherent distribution of battery characteristics, generate extended range data highly similar to the original training data, but more abundant in quantity.

[0047] (3) Prediction model construction Based on the virtual sample set, a training set is constructed to train the capacity prediction model. The capacity prediction model takes the time series composed of the health factor and capacity as the input, and the health factor and capacity corresponding to the next charge and discharge as the output.

[0048] Specifically, based on the General Grey Wolf-Extreme Learning Machine (GGWO-ELM) model, a capacity prediction model is constructed. The model takes the time series composed of the health factor and capacity corresponding to p (p < P, set as the number of charge and discharge cycles before the prediction starting point) charge and discharge cycles as the input, and the health factor and capacity corresponding to the (p + 1)-th charge and discharge cycle as the output. The model is trained with the goal of minimizing the difference between the health factor, capacity predicted by the model and the corresponding true values in the samples.

[0049] The input of the model is a time series with a length of p , , m is the number of time series in the training set, and each element in the series is a D-dimensional vector . In this embodiment, the vector dimension D is 5, which are the constant pressure rise charging time, constant pressure drop discharging time, average voltage change rate, discharging voltage at equal time, and capacity respectively; the output is the health factor and capacity corresponding to the (p + 1)-th charge and discharge cycle, that is .

[0050] Based on the virtual sample set, a training set that meets the above model input and output is constructed. Specifically: Select p + 1 consecutive samples from the virtual sample set, and form a time series with the health factor and capacity of the previous p times , and the health factor and capacity of the (p + 1)-th time as the label of the time series . Take ( , ( ) is used as a training sample to obtain the training set.

[0051] The model is trained using a training set, with the training objective set as minimizing the difference between the model's predicted health factor (p+1)th time and the corresponding true value in the sample. This can be expressed by the formula:

[0052] in, The weight is the difference between the predicted value and the actual value in the k-th dimension.

[0053] After training the model using the training set, in order to verify the training effect, a test set is constructed based on the original small sample set using the same training set construction method, and the training effect is tested.

[0054] (4) Remaining life prediction The trained capacity prediction model is used to predict the battery's RUL, specifically: From the first p charge-discharge curves of the target lithium battery, the health factors and capacity of a total of p cycles are extracted to form a time series. The data is input into the trained capacity prediction model to obtain the health factor and capacity for the (p+1)th iteration, i.e. ; The health factors and capacity of the last p-1 times Compared with the predicted health factors and capacity at the p+1th time Forming a new time series Input the data into the model again to predict the health factors and capacity for the (p+2)th iteration. The process iterates to predict the health factors and capacity during several future charge and discharge cycles. During the iterative prediction process, when the predicted capacity is lower than the preset failure threshold for the first time, the corresponding iteration number is the number of charge and discharge cycles that the target lithium battery can continue to perform when it reaches the end of its lifespan, thus obtaining the prediction result of the remaining lifespan of the lithium battery.

[0055] To verify the effectiveness and applicability of this method, this embodiment selected NASA-provided B05, B06, and B07 batteries for experimental verification, and set the battery capacity failure threshold to be... The prediction starting point is set at the 30th cycle (p=29), that is, the number of charge-discharge cycles is 30. Starting from 30 cycles, the capacity of different charge-discharge cycles (30, 31, 32, etc.) is continuously predicted. By calculating the number of charge-discharge cycles required for the predicted capacity to reach failure, the RUL of the lithium battery is predicted. To effectively evaluate model performance, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Absolute Error (AE) are used as evaluation criteria. The calculation methods are as follows:

[0056] in, The maximum number of iterations, For the first The actual value of the capacity in the next loop. For the first Predicted capacity for the next cycle.

[0057] To verify the model's generalization performance, predictions were performed using different individual cells of the same type. Specifically, cross-validation was used to predict the RUL of cells B05, B06, and B07. The RUL prediction methods employed were Extreme Learning Machine (ELM), Gray Wolf Algorithm-Extreme Learning Machine (GWO-ELM), and Generalized Gray Wolf Algorithm-Extreme Learning Machine (GGWO-ELM), respectively. This embodiment is configured as follows: -Scenario 1: Using real data from batteries B06 and B07 as the training set and real data from battery B05 as the test set, the results of RUL prediction are shown in Figure 6. The MAE, RMSE, and AE values ​​for the three methods, ELM, GWO-ELM, and GGWO-ELM, are shown in Table 1. -Scenario 2: By generating virtual samples from battery data of B06 and B07, 100 complete virtual sample data sequences were formed. These 100 virtual samples were used as the training set, and the original sample data of battery B05 was used as the test set for RUL prediction. The RUL prediction curve is shown in Figure 7. The MAE, RMSE, and AE values ​​of the three methods, ELM, GWO-ELM, and GGWO-ELM, are shown in Table 1. -Scenario 3: Using real data from batteries B05 and B07 as the training set and real data from battery B06 as the test set, the results of RUL prediction are shown in Figure 8. The MAE, RMSE, and AE values ​​for the three methods, ELM, GWO-ELM, and GGWO-ELM, are shown in Table 2. -Scenario 4: By generating virtual samples from battery data B05 and B07, 100 complete virtual sample data sequences were formed. These 100 virtual samples were used as the training set, and the original sample data of battery B06 was used as the test set for RUL prediction. The RUL prediction curve is shown in Figure 9. The MAE, RMSE, and AE values ​​for the three methods, ELM, GWO-ELM, and GGWO-ELM, are shown in Table 2.

[0058] In Figure 6-9, the red curve represents the actual RUL value curve of the lithium battery, and the blue curve represents the predicted RUL curve of the lithium battery after training with different models. Figure 6 (a) and Figure 8 In the diagram, (a) represents the RUL prediction result without virtual sample generation. Figure 7 (a) and Figure 9 In the figure, (a) represents the RUL prediction result generated by the virtual sample. As can be seen from the figure, compared with the prediction result of the real sample, the prediction result of the sample generated by the virtual sample generation technology is closer to the real capacity curve, indicating that under small sample conditions, the virtual sample generation method used in this embodiment can more accurately predict the RUL of lithium batteries.

[0059] Table 1 Comparison of B05 Battery Prediction Results

[0060] Table 2 Comparison of Prediction Results for Battery B06

[0061] A comprehensive qualitative and quantitative analysis based on Figure 6-9 and Table 1-2 clearly shows that: 1) For battery B05, compared with the prediction results of real samples, the method of generating virtual samples reduced the three error evaluations by 0.026, 0.0592, and 0.0348 respectively under the GGWO-ELM model; by 0.0474, 0.1636, and 0.0684 respectively under the GWO-ELM model; and by 0.1704, 0.3304, and 0.197 respectively under the ELM model. 2) For battery B06, compared with the prediction results of real samples, the method of generating virtual samples reduced the three error evaluations by 0.0082, 0.0531, and 0.0088 respectively under the GGWO-ELM model; by 0.0603, 0.1948, and 0.0768 respectively under the GWO-ELM model; and by 0.0673, 0.2767, and 0.0972 respectively under the ELM model.

[0062] The above results demonstrate that by expanding the battery data, the resulting model can achieve high accuracy in predicting other cells of the same type of battery, effectively increasing the information content of the original sample and achieving more accurate lithium battery capacity prediction.

[0063] Example 2 One embodiment of the present invention provides a lithium battery remaining life prediction system based on virtual sample generation, comprising: The original sample construction module is configured to extract samples from the historical charge-discharge curves of lithium batteries to form an original small sample set, wherein the samples include the number of charge-discharge cycles and the corresponding health factors and capacity. The virtual sample generation module is configured to generate highly similar virtual samples and form a virtual sample set by using a virtual sample generation method based on the differential evolution algorithm, according to the characteristics and distribution of sample data in the original small sample set. The prediction model building module is configured to: build a training set based on a virtual sample set to train the capacity prediction model, wherein the capacity prediction model takes the time series of health factors and capacity as input and the health factors and capacity corresponding to the next charge and discharge as output; The remaining life prediction module is configured to: extract the health factor and capacity corresponding to the most recent preset charge and discharge cycle from the charge and discharge curve of the target lithium battery, form a time series of health factors and capacity, input it into the trained capacity prediction model, reconstruct the time series based on the predicted health factor and capacity corresponding to the next charge and discharge cycle, and perform iterative prediction based on the capacity prediction model. When the predicted capacity is lower than the preset failure threshold for the first time, the corresponding number of iterations is the number of charge and discharge cycles that the target lithium battery can continue to perform when it reaches the end of its life, thereby obtaining the remaining life prediction result of the lithium battery.

[0064] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for predicting the remaining life of a lithium battery based on virtual sample generation.

[0065] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned method for predicting the remaining life of a lithium battery based on virtual sample generation.

[0066] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the lithium battery remaining life prediction method based on virtual sample generation.

[0067] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a lithium battery based on virtual sample generation, characterized in that, include: Samples were extracted from the historical charge-discharge curves of lithium batteries to form an original small sample set, which included the number of charge-discharge cycles and the corresponding health factors and capacity. A virtual sample generation method based on differential evolution algorithm is used to generate highly similar virtual samples based on the characteristics and distribution of sample data in the original small sample set, forming a virtual sample set. Based on a virtual sample set, a training set is constructed to train the capacity prediction model. The capacity prediction model takes the time series of health factors and capacity as input and the health factors and capacity corresponding to the next charge and discharge as output. From the charge-discharge curve of the target lithium battery, the health factor and capacity corresponding to the most recent preset charge-discharge cycle are extracted, and a time series of health factor and capacity is formed. This time series is then input into the trained capacity prediction model. Based on the predicted health factor and capacity corresponding to the next charge-discharge cycle, the time series is reconstructed, and iterative prediction based on the capacity prediction model is performed. When the predicted capacity is lower than the preset failure threshold for the first time, the number of iterations is the number of charge-discharge cycles that the target lithium battery can continue to perform when it reaches the end of its lifespan, thus obtaining the remaining lifespan prediction result of the lithium battery.

2. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 1, characterized in that, The health factors include isobaric charging time, isobaric discharge time, average voltage change rate, and isobaric discharge voltage.

3. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 1, characterized in that, The virtual sample generation method comprises the following steps: Based on the charge and discharge curves of lithium batteries, an original small sample set was constructed. Each sample consists of the number of charge and discharge cycles, the constant voltage rise charging time, the constant voltage drop discharge time, the average voltage change rate, the constant time discharge voltage, and the capacity. Based on the original small sample set, an information expansion function that considers sample density and quantity is used to determine the expansion range; Multiple sets of virtual samples are generated within the extended range using Latin hypercube sampling; Based on the generated multiple sets of virtual samples, an improved differential evolution algorithm is used to generate the optimal virtual sample.

4. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 3, characterized in that, The determination of the extended range specifically includes: Calculate the center observation location of the original small sample set; Considering the relative quantity and density relationship between skewness and the observations on both sides of the central observation position, calculate the left skewness and right skewness; Calculate the minimum and maximum values ​​of the sample; Based on the central observation position, left skewness, right skewness, minimum value, and maximum value, the upper and lower boundaries of the extended range are calculated to form the extended range.

5. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 3, characterized in that, The improved differential evolution algorithm is used to generate the optimal virtual sample. The specific steps are as follows: Based on the overall trend of the virtual sample data sequence and the trend of the original small sample data, a fitness function is defined. Based on the minimum and maximum values ​​of the samples, initial virtual samples are generated and divided into two categories: dominant virtual samples and inferior virtual samples, thus completing the population initialization. Based on the initialized population, iterative mutation, crossover, and selection are performed until the maximum number of iterations is reached, and the optimal virtual sample is output. The improvement is based on the use of superior and inferior virtual samples to refine the mutation strategy.

6. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 5, characterized in that, The fitness function is expressed by the formula: in, Indicates the first The fitness value of the overall trend of the virtual sample compared to the trend of the original small sample data. Indicates the original sample number. During the first charge and discharge cycle The standard deviation of each dimension.

7. The method for predicting the remaining life of a lithium battery based on virtual sample generation as described in claim 5, characterized in that, The improved mutation strategy is expressed by the following formula: in, Indicates the number of iterations. Indicates the first The virtual sample at the th The mutation matrix obtained after the nth iteration Indicates the first The matrix of a virtual sample after the g-th iteration and From the first Selected from the superior virtual samples after the next iteration. and From the first Select from the inferior virtual samples after the next iteration. This is the scaling factor.

8. A lithium battery remaining life prediction system based on virtual sample generation, characterized in that, include: The original sample construction module is configured to extract samples from the historical charge-discharge curves of lithium batteries to form an original small sample set, wherein the samples include the number of charge-discharge cycles and the corresponding health factors and capacity. The virtual sample generation module is configured to generate highly similar virtual samples and form a virtual sample set by using a virtual sample generation method based on the differential evolution algorithm, according to the characteristics and distribution of sample data in the original small sample set. The prediction model building module is configured to: build a training set based on a virtual sample set to train the capacity prediction model, wherein the capacity prediction model takes the time series of health factors and capacity as input and the health factors and capacity corresponding to the next charge and discharge as output; The remaining life prediction module is configured to: extract the health factor and capacity corresponding to the most recent preset charge and discharge cycle from the charge and discharge curve of the target lithium battery, form a time series of health factors and capacity, input it into the trained capacity prediction model, reconstruct the time series based on the predicted health factor and capacity corresponding to the next charge and discharge cycle, and perform iterative prediction based on the capacity prediction model. When the predicted capacity is lower than the preset failure threshold for the first time, the corresponding number of iterations is the number of charge and discharge cycles that the target lithium battery can continue to perform when it reaches the end of its life, thereby obtaining the remaining life prediction result of the lithium battery.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a method for predicting the remaining life of a lithium battery based on virtual sample generation as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a method for predicting the remaining life of a lithium battery based on virtual sample generation as described in any one of claims 1-7.

Images