Lithium ion battery RUL prediction method and system for performing centralized strategy learning based on multi-feature scale health factors

Abstract

The invention discloses a lithium ion battery RUL prediction method for centralized strategy learning based on a multi-feature scale health factor. The method comprises the following steps: acquiring multi-source battery operation data of a lithium ion battery for a full-life charge-discharge cycle experiment; respectively extracting capacity sequences according to the battery operation data; the obtained multi-feature scale health factor sequence is reserved according to availability; performing correlation analysis on the extracted multi-feature scale health factor sequence and the capacity sequence, and adaptively screening the health factor sequence as a feature factor sequence; a STAR model is constructed; dividing the characteristic factor sequence obtained by the operation data of each source battery into a first training set and a verification set according to a time sequence and a preset proportion; the first training set is used as STAR model input, the STAR model is trained, the trained first STAR model is predicted in a sliding window mode, and a verification set is adopted for verification; constructing a second training set according to a characteristic factor sequence obtained from the obtained operation data of the lithium ion battery to be predicted; training the STAR model according to the second training set to obtain a trained second STAR model; and utilizing the trained second STAR model to carry out step-by-step prediction through a sliding window mechanism to obtain a prediction result. The prediction efficiency of the model is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to a lithium ion battery RUL prediction method and system based on multi-feature scale health factors for centralized strategy learning, and belongs to the technical field of lithium ion battery state monitoring. BACKGROUND

[0002] Advocating green energy is always the main content of many issues in the world today, and new energy technologies related thereto are constantly innovated and improved. Lithium ion batteries are widely used in energy storage systems of electric vehicles due to their high energy, good charging and discharging performance, long cycle life and no memory effect. The battery management system as a key system for managing and monitoring the battery is applied to the battery management of electric vehicles. Among them, the remaining useful life (RUL) prediction of the battery is an important guarantee for the safe and effective work of the battery and is one of the performance indicators for evaluating the safety state of the battery. At present, the lithium ion battery RUL prediction method based on data driving mainly adopts a single health factor or a multi-health factor fusion method. Using a single health factor as a feature to predict the RUL of a lithium ion battery, this method has high prediction accuracy, but ignores the influence of other conditions on the lithium ion battery and is difficult to comprehensively reflect the dynamic changes existing in the actual complex system. The method of using multiple health factors for feature fusion has data redundancy between multiple data sequences and increases the calculation complexity. In terms of model construction, traditional time series prediction models have good prediction effect in short-term prediction, but have limited prediction modeling capability for long-term degradation sequences and multiple feature inputs.

[0003] In view of the above problems, a multi-feature scale health factor is needed as a feature input, and a centralized strategy learning method is used to establish an efficient time series prediction framework to realize the RUL prediction of the lithium ion battery. SUMMARY

[0004] The application provides a lithium ion battery RUL prediction method based on multi-feature scale health factors for centralized strategy learning. The method extracts multi-feature scale data sequences from a data set, combines the data sequences as feature inputs after adaptive filtering, greatly ensures the richness and interaction effect of the extracted data sequences, and effectively reduces the time complexity and calculation complexity in the data preprocessing stage. Further, a model network is built by combining the centralized learning strategy, which effectively reflects the dynamic changes existing in the actual system and significantly improves the prediction performance of the model.

[0005] The technical scheme of the application is as follows:

[0006] According to a first aspect of the application, a lithium ion battery RUL prediction method based on multi-feature scale health factors for centralized strategy learning is provided, comprising:

[0007] S1, acquire multi-source battery operation data of lithium ion battery for full-life charge-discharge cycle experiment;

[0008] S2, extract capacity sequence according to battery operation data; and extract preset multi-feature scale health factor sequence according to battery operation data, and obtain multi-feature scale health factor sequence according to availability reservation;

[0009] S3, the extracted multi-feature scale health factor sequence and the capacity sequence are respectively analyzed by the first coefficient and the second coefficient, and the health factor sequence is adaptively selected as the characteristic factor sequence;

[0010] S4, construct STAR model;

[0011] S5, combine steps S2-S4 to form MF-STAR model;

[0012] S6, the characteristic factor sequence obtained from each source battery operation data is respectively divided into first training set and validation set according to time sequence according to preset proportion; the first training set is used as the input of the STAR model, the STAR model is trained, and the trained first STAR model is obtained; the trained first STAR model is used to predict in a sliding window manner, and the validation set is used for verification, and the MF-STAR model is determined as an effective model;

[0013] S7, on the basis that the MF-STAR model is an effective model, the obtained lithium ion battery operation data to be predicted is used to obtain the characteristic factor sequence according to steps S2-S3 to construct a second training set; the STAR model is trained according to the second training set, and a trained second STAR model is obtained; the trained second STAR model is used to predict step by step through a sliding window mechanism to obtain a prediction result.

[0014] Further, the preset multi-feature scale health factor sequence includes three types: indirect health factor sequence, direct health factor sequence and similarity health factor sequence; the indirect health factor sequence includes at least one of the following: charge voltage current-time difference sequence, constant voltage charge-discharge time sequence, IC curve peak corresponding voltage value sequence, battery discharge temperature sequence; the direct health factor sequence includes: energy sequence; the similarity health factor sequence includes at least one of the following: Euclidean distance, Mahalanobis distance.

[0015] Further, the adaptive screening health factor sequence is specifically: weak correlation is represented by the absolute values of the first coefficient and the second coefficient being less than a first preset threshold; strong correlation is represented by the absolute values of the first coefficient and the second coefficient being greater than a second preset threshold; the health factor sequence with the first coefficient and the second coefficient both being weak correlation is taken as a first health factor sequence, and the number of the first health factor sequence is A; a health factor sequence with the first coefficient and the second coefficient both being strong correlation is randomly selected from the multiple feature scale health factor sequences, and the number of the health factor sequence is A, and the health factor sequence is taken as a second health factor sequence; the first health factor sequence and the second health factor sequence are taken as the feature factor sequence.

[0016] Further, the first coefficient is a Pearson correlation coefficient, and the second coefficient is a Spearman correlation coefficient, and the value ranges of the two are both [-1, 1], strong correlation is represented by the absolute value of the correlation coefficient being greater than 0.6, and weak correlation is represented by the absolute value of the correlation coefficient being less than 0.4.

[0017] Further, the STAR model first takes the feature factor sequence as the input of the STAR model; each sequence is respectively embedded through normalization; the recombined and fused information is sent through an MLP layer after centralized interactive learning by a core; and the final prediction of each sequence is generated through a linear layer.

[0018] According to a second aspect of the present application, a lithium-ion battery RUL prediction system based on multiple feature scale health factors for centralized strategy learning is provided, which comprises the modules of the lithium-ion battery RUL prediction method based on multiple feature scale health factors for centralized strategy learning in any one of the above.

[0019] According to a third aspect of the present application, a processor is provided, which is used for running a program, wherein the program performs the lithium-ion battery RUL prediction method based on multiple feature scale health factors for centralized strategy learning in any one of the above when running.

[0020] According to a fourth aspect of the present application, a computer readable storage medium is provided, which comprises a stored program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to perform the lithium-ion battery RUL prediction method based on multiple feature scale health factors for centralized strategy learning in any one of the above when the program runs.

[0021] The present application has the following beneficial effects:

[0022] Firstly, the present application proposes to extract multiple feature scale health factors from a battery data set, which greatly ensures the richness and interactive effect of the extracted data sequence.

[0023] Secondly, in the data preprocessing stage of the MF-STAR model proposed in this invention, correlation analysis is used to adaptively filter the extracted data sequences, and the filtering results are directly combined as the feature input of the model, which effectively reduces the time complexity and computational complexity of data preprocessing.

[0024] Third, the MF-STAR model proposed in this invention introduces a centralized learning strategy, which performs interactive learning by aggregating all embedded sequences, taking into account both the global degradation trend and the effective interaction of information.

[0025] In summary, this invention addresses the problems of neglecting complex dynamic changes in the system and high time and computational complexity in the data preprocessing stage in lithium-ion battery RUL prediction by adaptively filtering the dataset based on the extraction of health factors at multiple feature scales and combining it with the STAR model network structure design. It enhances the representation ability of features and the ability of the network structure to capture channel data by using a centralized learning strategy, thereby improving the prediction accuracy of lithium-ion battery RUL. Attached Figure Description

[0026] Figure 1 This is the overall flowchart of the present invention.

[0027] Figure 2 This is a scene diagram of the battery life aging accelerated experimental platform based on the present invention.

[0028] Figure 3 This is a structural diagram of the MF-STAR model of the present invention.

[0029] Figure 4 This is a comparison chart of the predictions made by the method of this invention and different models.

[0030] Figure 5 The graph shows the prediction results of the method of the present invention in multiple experimental datasets. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0032] Example 1: As Figures 1-5 As shown, according to a first aspect of the present invention, a method for predicting the relative uptime (RUL) of lithium-ion batteries based on centralized policy learning using multi-feature scale health factors is provided, comprising the following steps:

[0033] S1, acquire multi-source battery operation data of lithium ion battery for full life cycle charge-discharge cycle experiment;

[0034] The multi-source battery operation data includes at least two of the following: first source battery operation data, second source battery operation data, and third source battery operation data; for example, the first source battery operation data is derived from the NASA battery data set, the second source lithium ion battery operation data is derived from the Oxford University battery aging data set, and the third source lithium ion battery is derived from the battery life aging acceleration experiment platform data set.

[0035] S2, extract capacity sequences according to the battery operation data; and extract the preset multi-feature scale health factor sequence according to the battery operation data, and retain the obtained multi-feature scale health factor sequence according to availability;

[0036] If the data is multi-source battery operation data, then: extract capacity sequences according to the multi-source battery operation data; and extract the preset multi-feature scale health factor sequence according to the multi-source battery operation data, and retain the obtained multi-feature scale health factor sequence according to availability;

[0037] If the data is the acquired part of the lithium ion battery operation data to be predicted, then: extract capacity sequences according to the part of the lithium ion battery operation data to be predicted; and extract the preset multi-feature scale health factor sequence according to the part of the lithium ion battery operation data to be predicted, and retain the obtained multi-feature scale health factor sequence according to availability.

[0038] In the present application, the preset multi-feature scale health factor sequence includes three types: indirect health factor sequence, direct health factor sequence, and similarity health factor sequence; the indirect health factor sequence includes charge voltage current-time difference sequence, constant voltage charge-discharge time sequence, IC curve peak corresponding voltage value sequence, and battery discharge temperature sequence; the direct health factor sequence includes energy sequence; and the similarity health factor sequence includes Euclidean distance and Mahalanobis distance; the charge voltage current-time difference sequence and the constant voltage charge-discharge time sequence are extracted in multiple scales, for example, after extracting the charge voltage current-time difference sequence, the charge voltage time difference sequence and the charge current time difference sequence are extracted on this basis; after extracting the constant voltage charge-discharge time sequence, the constant voltage drop discharge time sequence and the constant voltage rise charge time sequence are extracted on this basis. That is, according to the above, the preset multi-feature scale health factor sequence specifically includes the IC curve peak corresponding voltage value sequence, the battery discharge temperature sequence, the energy sequence, the Euclidean distance, the charge voltage time difference sequence, the charge current time difference sequence, the constant voltage drop discharge time sequence, and the constant voltage rise charge time sequence.

[0039] In the above, the charging energy in the energy sequence is the sum of the constant current charging energy and the constant voltage charging energy, and the discharging energy is the constant current discharging energy; the charging voltage current-time difference sequence is that the charging current and the charging voltage change with time during the charging process of the battery, and by updating the current or voltage value at a certain point, the relationship between the charging current or the charging voltage and the time difference is shown; the isochronous charging and discharging time sequence is the time required for the voltage to decrease from a relatively high voltage to a relatively low voltage during the charging or discharging process of the battery; the IC curve peak value corresponding voltage value sequence is the voltage value corresponding to the peak position of the characteristic curve obtained by differentiating the battery charging and discharging curve; the Euclidean distance is the calculation of the geometric distance of the voltage, current and capacity data of multiple groups of data, measuring the difference between them; the Mahalanobis distance avoids manual selection of weights or standardization, and calculates the Mahalanobis distance of the voltage, current and capacity data of multiple groups of data.

[0040] Due to the different available multi-feature scale health factor sequences that can be extracted from each source battery operation data, only the actually available multi-feature scale health factor sequence is retained as the availability multi-feature scale health factor sequence for each source battery operation data. In the present application:

[0041] The availability multi-feature scale health factor sequence obtained from the first source battery operation data is the IC curve peak value corresponding voltage value sequence, the battery discharging temperature sequence, the Euclidean distance, the Mahalanobis distance, the charging voltage time difference sequence, the charging current time difference sequence, the isochronous discharging time sequence and the isochronous charging time sequence;

[0042] The availability multi-feature scale health factor sequence obtained from the second source battery operation data is the IC curve peak value corresponding voltage value sequence, the battery discharging temperature sequence, the Euclidean distance, the Mahalanobis distance, the isochronous discharging time sequence and the isochronous charging time sequence;

[0043] The availability multi-feature scale health factor sequence obtained from the third source battery operation data is the IC curve peak value corresponding voltage value sequence, the battery discharging temperature sequence, the energy sequence, the Euclidean distance, the Mahalanobis distance, the charging voltage time difference sequence, the charging current time difference sequence, the isochronous discharging time sequence and the isochronous charging time sequence.

[0044] S3, the extracted multi-feature scale health factor sequence and the capacity sequence are respectively analyzed for correlation by a first coefficient and a second coefficient, and a health factor sequence is adaptively selected as a characteristic factor sequence.

[0045] Further, the adaptive screening health factor sequence is specifically: weak correlation is indicated by the absolute values of the first coefficient and the second coefficient being less than a first preset threshold; strong correlation is indicated by the absolute values of the first coefficient and the second coefficient being greater than a second preset threshold; the health factor sequence with the first coefficient and the second coefficient both being weak correlation is taken as a first health factor sequence, and the number of the first health factor sequence is counted as A; a health factor sequence with the first coefficient and the second coefficient both being strong correlation is randomly selected from the multiple feature scale health factor sequence in a number of A and taken as a second health factor sequence; the first health factor sequence and the second health factor sequence are taken as the feature factor sequence.

[0046] Specifically, the first coefficient is a Pearson correlation coefficient, the second coefficient is a Spearman correlation coefficient, and the value ranges of the two are both [-1, 1]; strong correlation is indicated by the absolute value of the correlation coefficient being greater than 0.6, and weak correlation is indicated by the absolute value of the correlation coefficient being less than 0.4.

[0047] Through experiments, the usability multiple feature scale health factor sequences obtained from the operation data of each source battery all involve three categories, and the number of weakly correlated health factor sequences is less than the number of strongly correlated health factor sequences.

[0048] S4, constructing a STAR model; the STAR model first takes the feature factor sequence as an input of the STAR model; each sequence is respectively embedded through normalization; the recombined and fused information is sent through an MLP layer after centralized interactive learning by a core; and the final prediction of each sequence is generated through a linear layer.

[0049] Further, the STAR model is specifically:

[0050] S4-1, first embedding the input sequence in a Min-Max normalization embedding module, embedding the sequence in an MLP layer and a random pool to capture the learned representation and aggregate the connection to form a core. Further, the random pool and the core representation are as follows:

[0051] (1) The specific implementation of the random pool is as follows: first, calculate the probability P of each dimension j through SoftMax activation:

[0052]

[0053] In the formula, A ij represents the fusion feature intensity of the ith sequence in the jth dimension.

[0054] (2) The core representation of the multiple data sequences is:

[0055] o = f(s1, s2,..., s c )

[0056] where o denotes the vector generated by any function f, {s1, s2,..., s c} denotes the given multivariate sequence.

[0057] (3) Sampling from the multinomial distribution based on the probability P:

[0058] o j = A cj where c ~ P(p 1j , p 2j ,..., p Cj )

[0059] (4) Transforming the multivariate time series:

[0060] o i = Stoch Pool(MLP1(S i-1 ))

[0061] where MLP1: denotes the projection from a sequence of hidden dimensions d to a core dimension d', Stoch Pool denotes the stochastic pooling layer, and C aggregates the sequence to represent the generated core

[0062] S4-2, The core is built through the indirect interaction of sequences, i.e., the global representation of all channels. After construction, the steps of "repeat" and "connect" are performed, and the core is assigned to each sequence to form reorganized individual new sequences. Further, the fusion representation combines the core with all sequential representations as follows:

[0063] F i = Repeat Concat(S i-1 , o i )

[0064] S i = MLP2(F i ) + S i-1

[0065] where Repeat Concat denotes the operation of connecting the core representation to each series to obtain S i ∈ R C×d denotes the concatenated MLP2: fuses and projects the hidden layer backward.

[0066] S4-3, The information in the embedding system that is not captured by the MLP layer and the pooling layer is added to the core through a residual connection.

[0067] S4-4, during the fusion operation, the core and its corresponding series of residual information not captured are sent through the MLP layer to the subsequent linear layer for prediction.

[0068] S5, combining steps S2-S4, the MF-STAR model is composed to predict the RUL of the lithium ion battery.

[0069] In the MF-STAR model, the multi-feature scale data sequence is extracted from the data set, and the data sequence is combined as a feature input after adaptive screening, which effectively reduces the time complexity and computational complexity in the data preprocessing stage. The input feature factor is embedded by normalization, and the STAR model adopts a centralized learning strategy for interactive learning of the sequence. The recombined and fused information is sent through the MLP layer. Finally, the output of the STAR model generates the final prediction of each sequence through the linear prediction layer. The whole pipeline of "extraction-screening-learning-prediction" enhances the prediction ability of the model for battery RUL.

[0070] S6, the feature factor sequence obtained from the operation data of each source battery is divided into a first training set and a validation set according to a preset proportion according to the time sequence; the first training set is used as the input of the STAR model, and the STAR model is trained to obtain a trained first STAR model; the trained first STAR model is used to predict in a sliding window manner, and the validation set is used for verification to determine that the MF-STAR model is an effective model.

[0071] S7, on the basis that the MF-STAR model is an effective model, the feature factor sequence obtained from the operation data of the lithium ion battery to be predicted is obtained according to the steps of S2-S3 to construct a second training set (i.e. the operation data of the lithium ion battery to be predicted can be the first 20%-40% of the full-life charge-discharge cycle experiment); the STAR model is trained according to the second training set to obtain a trained second STAR model; the trained second STAR model is used to predict step by step through the sliding window mechanism to obtain the prediction result (each time the data of the current time window is used as the input of the model, the model outputs the feature factor and the capacity prediction value of the next time step; then the feature factor prediction value is included in the window data, the oldest data point in the window is removed, a new input window is formed, and the feature factor and the capacity prediction of the subsequent time step are continued).

[0072] According to a second aspect of the embodiments of the present application, there is provided a lithium-ion battery RUL prediction system based on multi-feature scale health factors for centralized strategy learning, comprising the modules of the lithium-ion battery RUL prediction method based on multi-feature scale health factors for centralized strategy learning according to any one of the above. Each module in the lithium-ion battery RUL prediction system based on multi-feature scale health factors for centralized strategy learning can be realized by software, hardware, and combinations thereof, in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0073] According to a third aspect of the embodiments of the present application, there is provided a processor for running a program, wherein the program performs the lithium-ion battery RUL prediction method based on multi-feature scale health factors for centralized strategy learning according to any one of the above when running.

[0074] According to a fourth aspect of the embodiments of the present application, there is provided a computer-readable storage medium comprising a stored program, wherein the program controls the device where the computer-readable storage medium is located to perform the lithium-ion battery RUL prediction method based on multi-feature scale health factors for centralized strategy learning according to any one of the above when running.

[0075] Embodiment 2: The optional embodiments of the present application are further described as follows in combination with simulation data:

[0076] S1, the data sets are selected from the public NASA battery data set, the public Oxford University battery aging data set, and the experimental data set of the full-life cycle experiment of the lithium-ion battery through the built battery life aging acceleration experiment platform. Specifically, the NASA battery data set adopts 18650 lithium-ion batteries, and the battery numbers are B0005-B0007, B0018; the Oxford University battery aging data set adopts Kokam lithium cobalt ion pouch batteries, and the battery numbers are recorded as Cell1-Cell4; the battery life aging acceleration experiment platform adopts a high-performance battery test system of CTE-4000-5V30A of Xinwei Company, selects lithium-ion batteries of 18650 of Delphi brand as experimental objects (battery numbers are recorded as 1-1, 1-2, and 1-3), uses a thermostat to provide a stable operating environment, and the battery life aging acceleration experiment platform is as shown in Figure 2 .

[0077] Further, the rated capacity of the lithium-ion battery is 2 Ah, the energy is 7400 mWh, the nominal voltage is 3.7 V, and the voltage range is 2.75-4.3 Ah. In the host computer control system, the corresponding steps performed by each battery are set, and all steps are run at room temperature 24℃:

[0078] (1) Constant current charging: charge at 1C / 2A constant current, and when the cut-off voltage is 4.2V, change to constant voltage charging;

[0079] (2) Constant voltage charging: continue charging at 4.2V constant voltage, and when the current gradually decreases to 0.1A, the battery is fully charged, and charging is stopped;

[0080] (3) Standby: after the battery is fully charged, it is standby for thirty minutes;

[0081] (4) Constant current discharge: after the standby is over, the battery is continuously discharged at 1C / 2A, and when the cut-off voltage is 2.75V, the discharge is stopped;

[0082] (5) Standby: after the battery discharge is stopped, it is standby for thirty minutes again;

[0083] (6) Cycle: take (1) as the starting point and (5) as the end of a charge-discharge cycle, and after 800 cycles, the experiment is stopped.

[0084] Record the battery operating conditions every 10s, and export the battery operating data after the experiment is stopped.

[0085] In the above, the two batteries numbered 1-1 and 1-2 are subjected to charge-discharge cycles according to the above steps, and the battery numbered 1-3 is different from the charge-discharge cycle steps of batteries 1-1 and 1-2 in that it is continuously discharged at 0.75C / 1.5A in the constant current discharge stage, and the rest of the process is the same as battery 1-1.

[0086] S2, extract the capacity sequence of B0005-B0007, B0018, Cell1-Cell4, 1-1, 1-2 and 1-3 respectively as the actual capacity value; and according to the multi-source battery operating data, extract the preset multi-characteristic scale health factor sequence, and obtain the multi-characteristic scale health factor sequence according to the availability, specifically:

[0087] The availability multi-characteristic scale health factor sequence obtained from the NASA battery data set is the voltage value sequence corresponding to the peak value of the IC curve, the battery discharge temperature sequence, the Euclidean distance, the Mahalanobis distance, the charging voltage time difference sequence, the charging current time difference sequence, the constant voltage drop discharge time sequence and the constant voltage rise charging time sequence;

[0088] The available multi-feature scale health factor sequences obtained from the Oxford University battery aging dataset are the IC curve peak voltage value sequence, battery discharge temperature sequence, Euclidean distance, Mahalanobis distance, constant voltage discharge time sequence, and constant voltage charging time sequence.

[0089] The available multi-feature scale health factor sequences obtained from the battery life aging acceleration experiment platform dataset are the IC curve peak voltage value sequence, battery discharge temperature sequence, energy sequence, Euclidean distance, Mahalanobis distance, charging voltage time difference sequence, charging current time difference sequence, constant voltage discharge time sequence, and constant voltage charging time sequence.

[0090] S3, the extracted multi-feature scale health factor sequence and the capacity sequence are analyzed for correlation by Pearson coefficient and Spearman coefficient, and a plurality of health factor sequences with strong correlation and weak correlation are adaptively selected as feature inputs. Further, taking the first source data NASA battery dataset B0005 battery as an example, the Pearson coefficient and the Spearman coefficient of the health factor sequence are shown in Table 1. Specifically, HF1-HF8 in Table 1 represent charging voltage time difference, charging current time difference, constant voltage discharge time, constant voltage charging time, IC curve peak voltage, discharge temperature, Euclidean distance and Mahalanobis distance, respectively.

[0091] Table 1 Correlation coefficient

[0092] Health Factors HF1 HF2 HF3 HF4 HF5 HF6 HF7 HF8 Pearson Coefficient -0.9586 0.8925 0.9983 0.9952 0.9779 -0.3929 -0.9897 0.1562 Spearman Coefficient -0.9584 0.8852 0.9943 0.9906 0.9845 -0.3380 -0.9786 0.2006

[0093] S4, a STAR model is constructed, and the overall structure is as shown in Figure 3 Further, the running parameters for prediction by the model are as follows: model learning rate 0.0001, early stopping rate 0.2, activation function GELU, fully connected layer network dimension 64, iteration number 100, and data dimension determined according to the number of feature factors adaptively selected.

[0094] Error evaluation index: coefficient of determination (Coefficient of determination), root mean square error (Root mean square error, RMSE) and mean absolute error (Mean absolute error, MAE):

[0095]

[0096] In the formula, X i , and respectively represent the true value, the average value and the predicted value of the battery capacity.

[0097] Based on the battery B0005, the extracted multi-feature scale health factor sequences are arranged in various combinations, and the model proposed in the application is used for RUL prediction; wherein the combination strategies include: Group1 (randomly selected single strong correlation health factor sequence), Group2 (randomly selected one strong correlation sequence and one weak correlation sequence), Group3 (two strong correlation health factor sequences and two weak correlation health factor sequences determined by the adaptive mode of the application), Group4 (all strong correlation health factor sequences) and Group5 (all extracted multi-feature scale health factor sequences), and Table 2 shows the accuracy evaluation of battery life prediction under different health factor sequence combination strategies. As can be seen from Table 2, all combinations have a high determination coefficient. Compared with other combinations, Group1 and Group3 have smaller errors, with RMSE of 1.3493% and 1.2469% respectively, and MAE of 0.7845% and 0.7611% respectively. Generally, a single health factor sequence with strong correlation has good prediction results, but it ignores the influence of other conditions on lithium-ion batteries, and it is difficult to comprehensively reflect the dynamic changes existing in actual complex systems. However, the application selects multiple health factor sequences with strong correlation coefficients and adaptive screening, and further uses the proposed model to predict the input of multiple sequence combinations with strong correlation and weak correlation, with smaller prediction error and better prediction result. Therefore, it can be known that the application can achieve better prediction performance under the consideration of complex dynamic changes existing in battery operation. It should be noted that the prediction result using only weakly correlated health factors shows a large error, which is not statistically significant, so these results are excluded from the comparative analysis.

[0098] Table 2 Accuracy evaluation of battery life prediction under different health factor sequence combination strategies

[0099] Combination Ways Group1 Group2 Group3 Group4 Group5 [R 2 ]]> 0.9819 0.9580 0.9836 0.9577 0.9783 RMSE (%) 1.3493 1.9423 1.2469 1.8634 2.8020 MAE (%) 0.7845 1.6713 0.7611 1.5029 1.5824

[0100] The battery data set processed by steps S1-S3 is divided, and the first 40% of the feature input data is used as a training set for centralized strategy interaction learning, and the remaining data is used as a validation set. The trained MF-STAR model is used to predict the validation set step by step through the sliding window mechanism (when using the trained model for prediction, the data in the sliding window is used as input, the model outputs the predicted value of the next time step, and the output predicted value is used as part of the window input data of the next time step, and the next time step prediction is continued). Further, taking the battery B0005 of the NASA data set as an example, the comparison chart of battery RUL prediction of the model proposed in the application and other models such as SVM, ELM, LSTM and Transformer is as follows Figure 4As shown in the table 3, the RMSE of the model is 1.2469% and the MAE is 0.7611%, which are higher than those of other models.

[0101] Table 3 Accuracy evaluation of different models for battery life prediction

[0102] Models SVM ELM LSTM Transformer MF-STAR [R 2 ]]> 0.8989 0.9680 0.9871 0.9760 0.9836 RMSE (%) 2.7263 1.5561 2.1514 2.9287 1.2469 MAE (%) 2.3206 1.4027 1.5548 2.2424 0.7611

[0103] Further, the overall prediction results of the battery aging data set of Oxford University and the data set obtained by using the battery life aging acceleration experiment platform are as shown in the table 4. Figure 5 As shown in the table 4, Figure 5 (a) Cell1 Predicted capacity and Cell1 Practical capacity represent the predicted capacity value of Cell1 and the actual capacity value of Cell1 respectively, Figure 5 (b) 1-1 Predicted capacity and 1-1 Practical capacity represent the predicted capacity value of 1-1 and the actual capacity value of 1-1 respectively). The experimental results show that the present application has a good prediction effect on the lithium ion battery, and the prediction accuracy of the present application is verified. It should be noted that the capacity value decreases with the increase of the number of uses, and it is generally considered that the battery reaches the end of life when the battery capacity decreases to 70% to 80% of the rated capacity. The rated capacity of the battery in this embodiment is 2 Ah, and the failure threshold is 1.4 Ah.

[0104] The specific embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the present application.

Claims

1. A lithium-ion battery RUL prediction method based on multi-feature scale health factor for centralized policy learning, characterized in that, The method comprises the following steps: S1, obtaining multi-source battery operation data of lithium ion batteries for full-life charge-discharge cycle experiments; S2, extracting capacity sequences according to the battery operation data; and extracting a pre-set multi-characteristic scale health factor sequence according to the battery operation data, and retaining the obtained multi-characteristic scale health factor sequence according to availability; S3, performing correlation analysis on the extracted multi-characteristic scale health factor sequence and the capacity sequence by using a first coefficient and a second coefficient, and adaptively selecting the health factor sequence as a characteristic factor sequence; S4, constructing a STAR model; S5, combining steps S2-S4 to form an MF-STAR model; S6, dividing a first training set and a validation set according to a preset proportion in time sequence for each characteristic factor sequence obtained from the multi-source battery operation data; using the first training set as the input of the STAR model to train the STAR model, obtaining a trained first STAR model, and using the validation set to verify the trained first STAR model in a sliding window manner to determine that the MF-STAR model is an effective model; S7, on the basis that the MF-STAR model is an effective model, obtaining a characteristic factor sequence for the obtained lithium ion battery operation data to be predicted according to steps S2-S3 to construct a second training set; training the STAR model according to the second training set to obtain a trained second STAR model, and using the trained second STAR model to gradually predict through a sliding window mechanism to obtain a prediction result.

2. The lithium-ion battery RUL prediction method based on multi-feature scale health factor centralized policy learning according to claim 1, characterized in that, The pre-set multi-characteristic scale health factor sequence includes three types: indirect health factor sequence, direct health factor sequence, and similarity health factor sequence; the indirect health factor sequence includes at least one of the following: charge voltage-current-time difference sequence, constant-voltage charge-discharge time sequence, IC curve peak corresponding voltage value sequence, and battery discharge temperature sequence; the direct health factor sequence includes an energy sequence; and the similarity health factor sequence includes at least one of the following: Euclidean distance and Mahalanobis distance.

3. The lithium-ion battery RUL prediction method based on multi-feature scale health factor centralized policy learning according to claim 1, characterized in that, The adaptive selection of the health factor sequence is specifically: weak correlation is indicated by the absolute values of the first coefficient and the second coefficient being less than a first preset threshold value; strong correlation is indicated by the absolute values of the first coefficient and the second coefficient being greater than a second preset threshold value; the health factor sequence with weak correlation of the first coefficient and the second coefficient is selected as a first health factor sequence, and the number of the first health factor sequence is counted as A; a health factor sequence with strong correlation of the first coefficient and the second coefficient is randomly selected from the multi-characteristic scale health factor sequence, and the number of the selected health factor sequence is A, and the selected health factor sequence is selected as a second health factor sequence; the first health factor sequence and the second health factor sequence are selected as the characteristic factor sequence.

4. The lithium-ion battery RUL prediction method based on multi-feature scale health factor centralized policy learning according to claim 3, characterized in that, The first coefficient is a Pearson correlation coefficient, the second coefficient is a Spearman correlation coefficient, the values of the two coefficients are both in the range of [-1, 1], strong correlation is indicated by the absolute value of the correlation coefficient being greater than 0.6, and weak correlation is indicated by the absolute value of the correlation coefficient being less than 0.

4.

5. The lithium-ion battery RUL prediction method based on multi-feature scale health factor centralized policy learning according to claim 1, characterized in that, The STAR model first takes the sequence of feature factors as the input of the STAR model; each sequence is respectively embedded by normalization; the recombined and fused information is sent through the MLP layer after centralized interactive learning by the core; And then the final prediction of each sequence is generated through the linear layer.

6. A lithium-ion battery RUL prediction system based on multi-feature scale health factor for centralized policy learning, characterized in that, The module of the lithium ion battery RUL prediction method based on multi-feature scale health factor centralized strategy learning according to any one of claims 1-5.

7. A processor, comprising: The processor is used for running a program, wherein the program performs the lithium ion battery RUL prediction method based on multi-feature scale health factor centralized strategy learning according to any one of claims 1-5 when running.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to perform the lithium ion battery RUL prediction method based on multi-feature scale health factor centralized strategy learning according to any one of claims 1-5 when the program runs.

Images