Lithium battery health state joint prediction method based on multi-modal timing characteristics

Abstract

The application relates to the technical field of battery state prediction, and discloses a lithium battery health state joint prediction method based on multi-modal time sequence characteristics, which comprises the following steps: extracting state parameters and health state indexes from historical state data of a lithium battery; arranging time sequences of each state parameter and health state index respectively; calculating the response lag amount of each state parameter and constructing the response sensitive curve of each state parameter respectively; constructing a training set based on the response lag amount and the response sensitive curve of each state parameter and training a health prediction model; acquiring state parameters of the lithium battery in real time, and predicting the health state indexes of the lithium battery based on the trained health prediction model; and identifying the abnormal evolution trend of the lithium battery health state based on the state parameters and the health state indexes. The application can improve the lithium battery health state prediction precision and enhance the adaptability of the prediction method to different working conditions.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of battery state prediction, and in particular to a lithium battery health state joint prediction method based on multi-modal time sequence features. BACKGROUND

[0002] In recent years, lithium battery health state prediction technology has evolved from early empirical models to data-driven models. The health state of a lithium battery is affected by multiple factors such as electrical performance, thermal response characteristics, and energy state. Single modal data cannot fully reflect the battery aging process, and multi-modal data fusion has become an important direction to improve prediction accuracy. Although significant progress has been made in lithium battery health state prediction technology, the industry still has the following shortcomings.

[0003] The collection frequency and response speed of different state parameters of a lithium battery differ significantly, making it difficult to align multi-modal data in the time dimension. For example, the changes in temperature and state of charge have a time delay effect on the health state. Existing technologies use simple time interpolation or fixed time window clipping methods to process multi-modal data, without considering the response lag characteristics of different parameters. This time sequence misalignment can produce false feature correlations, not only failing to take advantage of the complementary advantages of multi-modal data, but also introducing noise interference, reducing model prediction accuracy, especially under dynamic conditions. In practical applications, lithium batteries may experience abnormal evolution of health state due to manufacturing defects, overcharging, over-discharging, external collisions, and other sudden factors. Existing prediction models focus on health state prediction under normal aging trends and lack effective identification mechanisms for abnormal evolution. They also lack effective feedback adjustment mechanisms, making it difficult to adjust model parameters based on abnormal features and leading to a continuous decline in prediction accuracy under abnormal scenarios.

[0004] A lithium battery health state prediction method and device are disclosed in Chinese Patent No. CN119619893B. The method involves preprocessing a second predetermined number of health factor data to improve data quality and obtain a preprocessed health factor data set. Two convolution layers, a bidirectional long short-term memory network, a pre-set attention mechanism, a fully connected layer, and a pre-set encoder-decoder are used to sequentially process the preprocessed health factor data set, thereby achieving accurate prediction of the health state of a target battery.

[0005] A patent application with publication number CN120630012A discloses a lithium battery health state prediction method and system. The method includes the following steps: obtaining relevant original data sequences during lithium battery charging and discharging cycles; calculating health factor characteristic sequences based on the relevant original data sequences; constructing a prediction model, inputting residual component sequences, fluctuation component sequences, and indirect health factor characteristic sequences into the prediction model to obtain a health state prediction result of the battery. The health factor characteristic sequences are decomposed into fluctuation sequence components representing capacity regeneration fluctuations and trend sequence components representing overall attenuation trends by using empirical mode decomposition, which improves the influence of the prediction model caused by lithium battery "capacity regeneration" and improves the prediction accuracy of lithium battery health state estimation. By mining four indirect health factor characteristic sequences, the battery health state is predicted and estimated in multiple dimensions, so that the change trend of the battery health state can be captured comprehensively.

[0006] The above technical solutions all have the problems raised in the background art: the response lag characteristics of different parameters are not considered, and the complementary advantages of multi-modal data cannot be played.

[0007] The information disclosed in this background section is only intended to increase the understanding of the general background of the application and should not be considered as recognition or in any form as admitting that this information constitutes prior art known to those of ordinary skill in the art. SUMMARY

[0008] The technical problem to be solved by the present application is to overcome the defects of the prior art, provide a lithium battery health state joint prediction method based on multi-modal time sequence characteristics, improve the prediction accuracy of lithium battery health state, and enhance the adaptability of the prediction method to different working conditions.

[0009] To solve the above technical problems, the present application provides the following technical solutions:

[0010] A lithium battery health state joint prediction method based on multi-modal time sequence characteristics includes the following steps:

[0011] Extracting state parameters and health state indicators from historical state data of the lithium battery; arranging the time sequence of each state parameter and health state indicator respectively;

[0012] Calculating the response lag of each state parameter based on the time sequence of each state parameter and health state indicator, and constructing the response sensitive curve of each state parameter respectively;

[0013] Constructing a training set based on the response lag of each state parameter and the response sensitive curve, and training a health prediction model;

[0014] Real-time acquisition of the state parameters of the lithium battery, and prediction of the health state indicators of the lithium battery based on the trained health prediction model;

[0015] Identify the abnormal evolution trend of the lithium battery health state based on the state parameters and the health state indicators.

[0016] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, wherein: the state parameters at least include voltage, working current, internal resistance, cell temperature and state of charge of the lithium battery; the health state indicator is any one of SOH or remaining useful life;

[0017] The historical state data includes historical record values and corresponding time stamps of each state parameter and health state indicator; the original time sequence of each state parameter and health state indicator is sorted respectively through the time stamp, and the different original time sequences are time sequence aligned through sampling and interpolation to obtain the time sequence of each state parameter and health state indicator.

[0018] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, wherein: the method for calculating the response lag amount of the corresponding state parameter based on the time sequence of any state parameter and the time sequence of the health state indicator is as follows:

[0019] Set a first time window; the time sequence of any state parameter is intercepted through the first time window to obtain a reference subsequence of the corresponding state parameter;

[0020] The time sequence of the health state indicator is intercepted through the first time window to obtain a standard subsequence; the time stamps of the elements contained in the standard subsequence correspond one by one to the time stamps of the elements contained in the reference subsequence;

[0021] The cross-correlation coefficients of the reference subsequence and the standard subsequence under different lags are calculated respectively; wherein, the lag represents the lag of the standard subsequence relative to the reference subsequence;

[0022] The lag corresponding to the maximum cross-correlation coefficient of the reference subsequence and the standard subsequence is selected as the response lag of the corresponding state parameter.

[0023] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, wherein: the response sensitive curve is used to describe the response sensitivity of the corresponding state parameter under different offset amounts; the method for constructing the response sensitive curve of any state parameter is as follows:

[0024] Select one health state indicator from the time sequence of the health state indicator as a target indicator; locate the state parameter with the same time stamp as the target indicator in the time sequence of any state parameter as a reference parameter;

[0025] a second time window is set; and a state parameter subsequence under each offset is intercepted from the time series of state parameters according to the reference parameter and the second time window;

[0026] a target indicator prediction model is trained; the input of the target indicator prediction model is a state parameter subsequence of any state parameter under any offset, and the output is a predicted value of a corresponding target indicator;

[0027] based on the target indicator prediction model, a predicted value of a target indicator is calculated according to a state parameter subsequence of any state parameter under different offsets, and a response sensitivity of a corresponding state parameter under different offsets is calculated; and a response sensitivity curve is drawn based on the response sensitivity of different state parameters under different offsets.

[0028] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, the offset is represented by the number of consecutive time stamps; the method of intercepting a state parameter subsequence under any offset s is as follows: the initial center position of the second time window is set to the position corresponding to the reference parameter; in the time sequence of state parameters, the second time window is moved s time stamps in the direction of earlier time, and a state parameter subsequence at the corresponding position is intercepted;

[0029] The method of calculating the response sensitivity of any state parameter under any offset is as follows:

[0030] a state parameter subsequence under any offset is input into the target indicator prediction model to obtain a reference predicted value of the state parameter under the corresponding offset;

[0031] a state parameter subsequence under any offset is input into the target indicator prediction model to obtain a reference predicted value of the state parameter under the corresponding offset;

[0032] The response sensitivity of the state parameter under the corresponding offset is calculated based on the difference between the reference predicted value and the perturbed predicted value.

[0033] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, the input of the health prediction model includes a state parameter subsequence of each state parameter under a response lag, and the output is a corresponding health state indicator;

[0034] Each piece of training data in the training set includes an input part of the health prediction model and a labeled reference value of the health state indicator output by the health prediction model;

[0035] The method for constructing any training data is as follows: selecting one health status indicator from the time series of the health status indicators as a target indicator, and recording the target indicator as the labeled reference value;

[0036] Locating the state parameters with the same timestamp as the target indicator in the time series of each state parameter as the reference parameters of each state parameter, respectively;

[0037] For any state parameter, according to the corresponding reference parameter, the state parameter subsequence under the target offset is intercepted from the time series of the state parameter through the second time window; the target offset is equal to the response lag amount corresponding to the state parameter.

[0038] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, wherein: the input of the health prediction model and any training data further includes a fusion weight vector; the fusion weight vector contains the fusion weight of each state parameter; the method for constructing the fusion weight vector is as follows:

[0039] The lag sensitivity of each state parameter is calculated respectively, specifically including: for any state parameter, based on the corresponding response sensitivity curve, the response sensitivity when the offset amount is equal to the response lag amount is calculated as the lag sensitivity of the corresponding state parameter;

[0040] The lag sensitivity of each state parameter is normalized, and the fusion weight of each state parameter is assigned based on the normalized lag sensitivity, and the fusion weight of any state parameter is positively correlated with the corresponding lag sensitivity.

[0041] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics described in the present application, wherein: the health prediction model includes a fusion attention layer; the health prediction model processes the fusion weight vector based on the fused attention layer, thereby controlling the attention weight of different state parameters;

[0042] The input of the health prediction model and any training data further includes a time decay vector; a time decay function is constructed; the time decay function is uniformly sampled to obtain time decay factors with the same number of elements as any state parameter subsequence, and the time decay factors are combined to form a time decay vector;

[0043] The health prediction model processes the time decay vector based on the fusion attention layer, thereby controlling the attention weight of the elements of different timestamps in the state parameter subsequence of any state parameter.

[0044] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics provided in the application, wherein: the abnormal evolution trend of the lithium battery health state is identified based on the state parameters and the health state indicators, and specifically includes:

[0045] The health state indicators of the lithium battery at different time points are predicted based on the health prediction model, and a state evolution sequence of the lithium battery is constructed; any element in the state evolution sequence corresponds to a time point, and the element value is a vector composed of the health state indicator at the corresponding time point and each state parameter at the corresponding time point;

[0046] A reference state evolution sequence of the lithium battery is obtained; the distance between the state evolution sequence and the reference state evolution sequence is calculated by dynamic time warping and normalized as an abnormal evolution indicator of the lithium battery;

[0047] An abnormal evolution threshold is set; if the abnormal evolution indicator is greater than the abnormal evolution threshold, the health state of the lithium battery has an abnormal evolution trend.

[0048] As a preferred scheme of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics provided in the application, wherein: if the health state of the lithium battery has an abnormal evolution trend, the prediction of the health state indicator is adjusted by feedback, and specifically includes:

[0049] The decay rate of the time decay function is increased, the time decay function is uniformly resampled again, and the time decay vector is updated;

[0050] Based on the response lag of each state parameter, a weight adjustment factor is assigned to each state parameter, and the weight adjustment factor of any state parameter is negatively correlated with the corresponding response lag; the fusion weight of each state parameter is multiplied by the corresponding weight adjustment factor, and the fusion weight vector is updated;

[0051] The updated time decay vector and fusion weight vector are applied to predict the health state indicator of the lithium battery.

[0052] Compared with the prior art, the application has the following beneficial effects:

[0053] The application extracts multi-dimensional state parameters such as lithium battery voltage, working current and internal resistance to construct multi-modal time sequence characteristics, and calculates the response lag of each state parameter, realizes accurate alignment of multi-modal data on the time axis, avoids time sequence misalignment caused by different parameter response speed differences, eliminates false feature correlation and noise interference. On this basis, the contribution of each parameter under different offsets is quantified combined with the response sensitive curve, dynamic weights are given for multi-modal feature fusion, the model focuses on key parameters that have a significant impact on the health state, and the feature fusion efficiency and prediction accuracy are improved.

[0054] The application realizes the identification of the abnormal evolution trend of the health state by constructing a battery state evolution sequence. The model prediction is optimized through a feedback adjustment mechanism to reduce the interference of untrusted data and ensure that the model can still maintain high prediction accuracy in abnormal scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0056] Figure 1 The flowchart of the lithium battery health state joint prediction method based on multi-modal time sequence characteristics provided by the application;

[0057] Figure 2 The method flowchart for calculating the response lag amount of the state parameter provided by the application. DETAILED DESCRIPTION

[0058] The technical solutions of the application will be described in detail below by means of the drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solutions of the application, rather than limitations of the technical solutions of the application. In the case of no conflict, the technical features in the embodiments and the embodiments can be combined with each other.

[0059] This embodiment introduces a lithium battery health state joint prediction method based on multi-modal time sequence characteristics, which refers to Figure 1 The method comprises the following steps:

[0060] The state parameters and health state indicators are extracted from the historical state data of the lithium battery; the time sequence of each state parameter and health state indicator is sorted respectively;

[0061] The state parameters at least include the voltage, working current, internal resistance, cell temperature and state of charge of the lithium battery.

[0062] The above state parameters jointly constitute the multi-modal time sequence characteristics of the lithium battery operating state, which can reflect the electrical performance, thermal response characteristics and energy state change of the battery under different working conditions, and provide a data basis for identifying the prediction and evolution trend of the health state indicator.

[0063] The health status indicator is any one of SOH or remaining useful life; wherein SOH (state of health) represents the degree of attenuation of the current performance of the battery relative to the performance of a brand-new battery, expressed as a percentage, and is a key indicator of the degree of battery aging, directly determining the battery's endurance, charging and discharging efficiency, and remaining useful life. When SOH is below a certain threshold, the battery performance will decrease significantly, and needs to be repaired or replaced. The remaining useful life is the time or cycle number that the battery can normally work from the current state to the point where it cannot meet the predetermined performance requirements (such as the capacity decays to 80% of the initial value).

[0064] The historical state data includes historical record values of each state parameter and health status indicator and corresponding time stamps; the original time series of each state parameter and health status indicator are sorted respectively through the time stamps, and the different original time series are time-aligned through sampling and interpolation to obtain the time series of each state parameter and health status indicator. For example, since the health status indicator of lithium battery such as SOH cannot be continuously collected, the interpolation method is used to supplement the sparse SOH data in this embodiment, so that the SOH data form a time series that is time-aligned with the high-frequency sampled state parameters, so as to realize the subsequent response lag analysis and model training. Alternatively, the non-linear degradation process of SOH is simulated by spline interpolation or local regression to obtain SOH data consistent with the actual situation.

[0065] The response lag amount of each state parameter is calculated based on the time series of each state parameter and the health status indicator, and the response sensitive curve of each state parameter is constructed respectively;

[0066] Reference Figure 2 The method for calculating the response lag amount of the corresponding state parameter based on the time series of any one state parameter and the time series of the health status indicator is as follows:

[0067] A first time window is set; the time series of any one state parameter is intercepted through the first time window to obtain a reference subsequence of the corresponding state parameter;

[0068] The time series of the health status indicator is intercepted through the first time window to obtain a standard subsequence; the time stamps of the elements contained in the standard subsequence correspond one by one to the time stamps of the elements contained in the reference subsequence;

[0069] correlation coefficient between the reference subsequence and the standard subsequence under different lag amounts, wherein the lag amount represents a lag amount of the standard subsequence relative to the reference subsequence; for example, if the lag amount is t, when calculating the cross-correlation coefficient between the reference subsequence and the standard subsequence, the standard subsequence is moved in the time sequence to the direction of earlier time by t time stamps, so that the i+t th element in the standard subsequence is aligned with the i th element in the reference subsequence, i is a positive integer, and t is any integer.

[0070] The lag amount corresponding to the maximum cross-correlation coefficient between the reference subsequence and the standard subsequence is selected as the response lag amount of the corresponding state parameter.

[0071] In this embodiment, the response lag amount represents the time delay required for the change of the state parameter to have an impact on the health status indicator. By identifying the response lag amount of different state parameters, the multi-modal data can be effectively adjusted on the time axis, so that when the multi-modal state parameters are fused, the features of each modality can be aligned at the moment when they really have an impact on the health status indicator, improving the perception ability and fusion efficiency of the health prediction model to the change rule of the state parameter, and avoiding prediction deviation caused by time misalignment between modalities.

[0072] The response sensitivity curve is used to describe the response sensitivity of the corresponding state parameter under different offset amounts; the method for constructing the response sensitivity curve of any state parameter is as follows:

[0073] A health status indicator is selected as a target indicator from the time sequence of the health status indicator; a state parameter with the same time stamp as the target indicator is located in the time sequence of any state parameter as a reference parameter;

[0074] A second time window is set; according to the reference parameter, a state parameter subsequence under different offset amounts is intercepted through the second time window;

[0075] An indicator prediction model is trained; the input of the indicator prediction model is a state parameter subsequence of any state parameter under any offset amount, and the output is a predicted value of the corresponding target indicator;

[0076] Optionally, the indicator prediction model is a single-feature regression model based on LightGBM. In this embodiment, the indicator prediction model is only used to analyze the response sensitivity of each state parameter under different offset amounts, and is not used for actual deployment and prediction of the health status indicator, so the prediction accuracy is not required to be too high, and when training and applying, the model parameters (such as tree depth, learning rate, minimum sample leaf node number, etc.) remain the same when processing different state parameters.

[0077] Based on the index prediction model, a prediction value of the target index is calculated according to a state parameter subsequence under different offsets of any state parameter, and a corresponding response sensitivity of the state parameter under different offsets is calculated; and the response sensitivities of different state parameters under different offsets are plotted into a response sensitivity curve.

[0078] The offset is represented by the number of consecutive time stamps; and the method of intercepting the state parameter subsequence under any offset s is as follows: the initial center position of the second time window is set as the position corresponding to the reference parameter; and in the time sequence of the state parameter, the second time window is moved s time stamps in the direction of earlier time (moving in the direction of earlier time means that the time stamp corresponding to the element intercepted in the second time window after moving is earlier than before moving), and the state parameter subsequence at the corresponding position is intercepted;

[0079] The method of calculating the response sensitivity of any state parameter under any offset is as follows:

[0080] The state parameter subsequence under any offset is input into the index prediction model to obtain a reference prediction value of the state parameter under the corresponding offset;

[0081] The state parameter subsequence is input into the index prediction model after a perturbation amount is applied to the elements in the state parameter subsequence to obtain a perturbation prediction value of the state parameter under the corresponding offset;

[0082] The response sensitivity of the state parameter under the corresponding offset is calculated based on the difference between the reference prediction value and the perturbation prediction value.

[0083] Optionally, the embodiment provides a specific method for calculating the response sensitivity, comprising:

[0084] The standard deviation of all elements in the state parameter subsequence is calculated as the perturbation amount;

[0085] The perturbation amount is sequentially applied to each element in the state parameter subsequence; and after the perturbation amount is applied to any element, the state parameter subsequence is input into the index prediction model to obtain a corresponding perturbation prediction value;

[0086] The application of the perturbation amount includes the application of a positive perturbation and the application of a negative perturbation; for any element in the state parameter subsequence, the application of the positive perturbation means that the element value is added by the perturbation amount, and the application of the negative perturbation means that the element value is subtracted by the perturbation amount;

[0087] The response sensitivity R of the state parameter under the corresponding offset is calculated based on the reference prediction value and the perturbation prediction value obtained after each application of the perturbation, and the formula is as follows:

[0088] ;

[0089] wherein, N represents the length of the state parameter subsequence, represents the reference prediction value; represents the perturbation prediction value calculated after applying a positive perturbation to the jth element in the state parameter subsequence, represents the perturbation prediction value calculated after applying a negative perturbation to the jth element in the state parameter subsequence.

[0090] In this embodiment, for any state parameter, the higher the response sensitivity at any offset, the more sensitive the state parameter is to the prediction of the health status indicator at this offset, that is, the prediction of the health status indicator depends more strongly on the state parameter at this offset, and a higher fusion weight will be assigned to this period to strengthen the information utilization of the state parameter. On the contrary, if the response sensitivity at any offset is low, it means that the state parameter at this offset has a low prediction contribution to the health status indicator, and the state parameter at this period is redundant. A lower fusion weight will be assigned to suppress the prediction error caused by irrelevant information.

[0091] Based on the response lag and the response sensitivity curve of each state parameter, a training set is constructed and a health prediction model is trained;

[0092] The input of the health prediction model includes the state parameter subsequence of each state parameter at the response lag, and the output is the corresponding health status indicator;

[0093] Any training data in the training set includes the input part of the health prediction model, and the labeled reference value of the health status indicator output by the health prediction model;

[0094] The method of constructing any training data is as follows: select a health status indicator from the time series of the health status indicator as the target indicator, and record the target indicator as the labeled reference value;

[0095] Locate the state parameter with the same timestamp as the target indicator in the time series of each state parameter as the reference parameter of each state parameter;

[0096] For any state parameter, according to the corresponding reference parameter, the state parameter subsequence at the target offset is intercepted from the time series of the state parameter through the second time window; the target offset is equal to the corresponding response lag of the state parameter.

[0097] The input of the health prediction model and any training data further includes a fusion weight vector; the fusion weight vector contains the fusion weight of each state parameter; the method of constructing the fusion weight vector is as follows:

[0098] The hysteresis sensitivity of each state parameter is calculated respectively, specifically including: for any state parameter, the response sensitivity when the offset is equal to the response hysteresis is calculated based on the corresponding response sensitivity curve, as the hysteresis sensitivity of the corresponding state parameter;

[0099] The hysteresis sensitivity of each state parameter is normalized, and the fusion weight of each state parameter is assigned based on the normalized hysteresis sensitivity, and the fusion weight of any state parameter is positively correlated with the corresponding hysteresis sensitivity.

[0100] The health prediction model includes a fusion attention layer; the health prediction model processes the fusion weight vector based on the fused attention layer, thereby controlling the attention weight of different state parameters;

[0101] The input of the health prediction model and any training data further includes a time decay vector; a time decay function is constructed; the time decay function is uniformly sampled to obtain time decay factors with the same number of elements in any state parameter subsequence, and the time decay factors are combined to form a time decay vector;

[0102] The health prediction model processes the time decay vector based on the fusion attention layer, thereby controlling the attention weight of elements with different time stamps in the state parameter subsequence of any state parameter.

[0103] Optionally, the fusion attention layer of the health prediction model automatically trains and optimizes an attention weight matrix based on the input multi-modal state parameters during the training process, which is used to record the attention weight of each element corresponding to each timestamp in the state parameter sub-sequence of each state parameter; the attention weight is used to control the attention degree of the model to each state parameter, so as to pay more attention to the state parameters with significant influence when predicting the health status indicators. For example, let the attention weight matrix be an M x N matrix, M is the number of state parameters; N is the length of the state parameter sub-sequence, that is, the total number of timestamps corresponding to any item of state parameter input into the model, any row of the attention weight matrix corresponds to a state parameter, and any element in any column position of any row corresponds to a timestamp; in this way, the attention weight in the attention weight matrix corresponds to the state parameter input into the model one by one; when processing the state parameter sub-sequence of different state parameters, the health prediction model controls the prediction weight of each state parameter through the corresponding attention weight; before processing the state parameter sub-sequence of different state parameters, the attention weight in the attention weight matrix is adjusted by the fusion weight vector and the time decay vector; wherein the fusion weight vector has a length of M, and each element in the attention weight matrix is multiplied by the corresponding fusion weight in the fusion weight vector. Through the regulation and control of the fusion weight vector, the false synergy caused by the multi-modal response time difference can be eliminated, and the prediction accuracy of multi-modal data fusion can be improved. The time decay vector has a length of N, and each column element in the attention weight matrix is multiplied by the corresponding time decay factor in the time decay vector. Through the regulation and control of the time decay vector, the information pollution of early data to the current state can be avoided, and the adaptability of the model to the nonlinear change of the battery state can be improved.

[0104] Optionally, the health prediction model is a multi-modal time series prediction model based on the Transformer architecture, or an LSTM model that fuses attention mechanism.

[0105] The state parameters of the lithium battery are acquired in real time, and the health status indicators of the lithium battery are predicted based on the trained health prediction model; specifically including: continuously acquiring and recording each state parameter, organizing and updating the time sequence of each state parameter; acquiring the response lag of each state parameter; for any specified time of the health status indicator to be predicted, the state parameter sub-sequence is respectively intercepted from the time sequence of each state parameter based on the corresponding response lag; the fusion weight vector is acquired, and the time decay function is constructed; the time decay function is uniformly sampled to obtain the time decay vector; the state parameter sub-sequence of each state parameter, the fusion weight vector, and the time decay vector are input into the trained health prediction model to obtain the health status indicators at the specified time.

[0106] Identify the abnormal evolution trend of the lithium battery health state based on the state parameters and the health state indicators; specifically including:

[0107] Predict the health state indicators of the lithium battery at different times based on the health prediction model, and construct the state evolution sequence of the lithium battery; any element in the state evolution sequence corresponds to a time, and the element value is a vector composed of the health state indicator at the corresponding time and each state parameter at the corresponding time;

[0108] Obtain the reference state evolution sequence of the lithium battery; calculate the distance between the state evolution sequence and the reference state evolution sequence by dynamic time warping and normalize it as the abnormal evolution indicator of the lithium battery;

[0109] Set the abnormal evolution threshold; if the abnormal evolution indicator is greater than the abnormal evolution threshold, the health state of the lithium battery has an abnormal evolution trend. Those skilled in the art can set the abnormal evolution threshold based on actual needs. Alternatively, multiple reference state evolution sequences are set according to different use conditions of the lithium battery, for example, long-term tracking experiments are respectively conducted on lithium batteries with regular use frequency, intermittent use frequency, and high use frequency, and reference state evolution sequences corresponding to each condition are respectively established. Calculate the distance between the state evolution sequence and each reference state evolution sequence and normalize it, and take the minimum value as the abnormal evolution indicator.

[0110] If the health state of the lithium battery has an abnormal evolution trend, feedback adjustment is made to the prediction of the health state indicator, specifically including:

[0111] Increase the decay rate of the time decay function, and re-sample the time decay function uniformly, and update the time decay vector;

[0112] Based on the response lag of each state parameter, assign a weight adjustment factor to each state parameter, and the weight adjustment factor of any state parameter is negatively correlated with the corresponding response lag; multiply the fusion weight of each state parameter by the corresponding weight adjustment factor, and update the fusion weight vector;

[0113] Apply the updated time decay vector and fusion weight vector to predict the health state indicator of the lithium battery.

[0114] In the embodiment, the abscissa of the time decay function is a negative time, for example, the abscissa of the specified time to be predicted is taken as the origin of the abscissa, and as the abscissa increases, the time corresponding to the abscissa gradually becomes earlier, that is, the greater the abscissa, the longer the time before the specified time to be predicted. The ordinate of the time decay function is a time decay factor. The time decay function is a function that monotonically decreases as the abscissa increases, and the greater the decay rate, the faster the monotonically decreasing, and the smaller the time decay factor at the same time obtained by sampling. The time decay vector is used to weaken the influence of the state parameters with a relatively long time stamp on the prediction of the health state index. When the distance between the state evolution sequence and the reference state evolution sequence is small, the health state evolution of the lithium battery is relatively stable, a smaller decay rate is set, so that the time decay factor is larger, the influence of the earlier state parameters is retained, so as to enhance the stable trend modeling; when the health state of the lithium battery has an abnormal evolution trend, a larger decay rate is set, so that the time decay factor is smaller, and the recent data is emphasized. The existence of an abnormal evolution trend means that the state evolution mode of the lithium battery is quite different from the training phase or the typical working condition, and the time sequence dependence relationship established in the training phase has a risk of failure, and the data of the earlier time sequence has less significance to the model prediction. On the one hand, by increasing the decay rate, on the other hand, by assigning a smaller weight adjustment factor to the state parameter with a larger response lag to reduce its fusion weight, both of which can strengthen the influence of the recent data on the prediction result, improve the information effectiveness of the model input data, and suppress the interference of the untrusted data on the prediction result.

[0115] Those skilled in the art will appreciate that embodiments of the application can be supplied as a method, a system, or a computer program product. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) containing computer usable program code.

[0116] The embodiments of the application are described above with reference to the accompanying drawings, but the application is not limited to the specific embodiments described above, which are merely illustrative rather than restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the application without departing from the purpose and the scope of protection of the application.

Claims

1. A method for lithium battery state of health joint prediction based on multi-modal time series features, characterized in that, The method comprises the following steps: extracting state parameters and health state indicators from historical state data of the lithium battery; organizing time series of each state parameter and health state indicator respectively; calculating the response lag of each state parameter based on the time series of each state parameter and health state indicator, and constructing the response sensitivity curve of each state parameter respectively; constructing a training set based on the response lag of each state parameter and the response sensitivity curve, and training a health prediction model; obtaining state parameters of the lithium battery in real time, and predicting the health state indicators of the lithium battery based on the trained health prediction model; 2. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 1, characterized in that: identifying abnormal evolution trends of the health state of the lithium battery based on the state parameters and the health state indicators. The state parameters at least include voltage, working current, internal resistance, cell temperature and state of charge of the lithium battery; and the health state indicators are any one of SOH or remaining service life. The historical state data includes historical record values of each state parameter and health state indicator and corresponding time stamps.

3. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 2, characterized in that: The original time series of each state parameter and health state indicator are organized respectively through the time stamps, and the different original time series are time series aligned through sampling and interpolation, to obtain the time series of each state parameter and health state indicator. The method for calculating the response lag of the corresponding state parameter based on the time series of any state parameter and the time series of the health state indicator is as follows: setting a first time window; the time series of any state parameter is intercepted through the first time window to obtain a reference subsequence of the corresponding state parameter; the time series of the health state indicator is intercepted through the first time window to obtain a standard subsequence; the time stamps of the elements contained in the standard subsequence correspond one by one to the time stamps of the elements contained in the reference subsequence; the cross-correlation coefficients of the reference subsequence and the standard subsequence under different lags are calculated respectively; wherein the lag represents the lag of the standard subsequence relative to the reference subsequence; 4. The lithium battery state of health joint prediction method based on multi-modal timing features of claim 3, wherein: the lag corresponding to the maximum cross-correlation coefficient of the reference subsequence and the standard subsequence is selected as the response lag of the corresponding state parameter. The response sensitivity curve is used to describe the response sensitivity of the corresponding state parameter under different offsets; the method for constructing the response sensitivity curve of any state parameter is as follows: selecting a health state indicator from the time series of the health state indicator as a target indicator; locating a state parameter with the same time stamp as the target indicator in the time series of any state parameter as a reference parameter; setting a second time window; according to the reference parameter, the state parameter subsequence under different offsets is intercepted through the second time window; training an indicator prediction model; the input of the indicator prediction model is the state parameter subsequence of any state parameter under any offset, and the output is the predicted value of the corresponding target indicator; based on the indicator prediction model, the predicted value of the target indicator is calculated according to the state parameter subsequence of any state parameter under different offsets, and the response sensitivity of the corresponding state parameter under different offsets is calculated; the response sensitivities of different state parameters under different offsets are plotted into a response sensitivity curve.

5. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 4, characterized in that: The offset is expressed by the number of continuous timestamps; the method for intercepting the state parameter subsequence under any offset s is as follows: the initial center position of the second time window is set as the position corresponding to the reference parameter; in the time sequence of the state parameters, the second time window is moved by s timestamps in the direction of early time, and the state parameter subsequence at the corresponding position is intercepted; The method for calculating the response sensitivity of any state parameter under any offset is as follows: The state parameter subsequence under any offset is input into the index prediction model to obtain the reference prediction value of the state parameter under the corresponding offset; The state parameter subsequence under any offset is input into the index prediction model to obtain the reference prediction value of the state parameter under the corresponding offset; The response sensitivity of the state parameter under the corresponding offset is calculated based on the difference between the reference prediction value and the perturbation prediction value.

6. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 5, characterized in that: The input of the health prediction model includes the state parameter subsequence of each state parameter under the response lag, and the output is the corresponding health state index; Any training data in the training set includes the input part of the health prediction model and the labeled reference value of the health state index output by the health prediction model; The method for constructing any training data is as follows: a health state index is selected from the time sequence of the health state index as a target index, and the target index is recorded as the labeled reference value; The same state parameter as the target index timestamp is located in the time sequence of each state parameter as the reference parameter of each state parameter; For any state parameter, the state parameter subsequence under the target offset is intercepted from the time sequence of the state parameter by the second time window according to the corresponding reference parameter; the target offset is equal to the response lag of the state parameter.

7. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 6, characterized in that: The input of the health prediction model and any training data also includes a fusion weight vector; the fusion weight vector includes the fusion weight of each state parameter; the method for constructing the fusion weight vector is as follows: The lag sensitivity of each state parameter is calculated respectively, specifically including: for any state parameter, the response sensitivity when the offset is equal to the response lag is calculated based on the corresponding response sensitivity curve, as the lag sensitivity of the corresponding state parameter; The lag sensitivity of each state parameter is normalized, and the fusion weight of each state parameter is assigned based on the normalized lag sensitivity, and the fusion weight of any state parameter is positively correlated with the corresponding lag sensitivity.

8. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 7, characterized in that: The health prediction model includes a fusion attention layer; The health prediction model processes the fusion weight vector based on the fused attention layer, thereby controlling the attention weight of different state parameters; The input of the health prediction model and any training data also includes a time decay vector; a time decay function is constructed; the time decay function is uniformly sampled to obtain time decay factors with the same number of elements as any state parameter subsequence, and the time decay factors are combined to form a time decay vector; The health prediction model processes the time decay vector based on a fusion attention layer, thereby controlling attention weights of elements of different timestamps in a state parameter subsequence of any item state parameter.

9. The lithium battery state of health joint prediction method based on multi-modal timing features according to claim 8, characterized in that: The abnormal evolution trend of the lithium battery health state is identified based on the state parameters and the health state indicators, and specifically includes: The health state indicators of the lithium battery at different times are predicted based on the health prediction model, and a state evolution sequence of the lithium battery is constructed; any element in the state evolution sequence corresponds to a time, and the element value is a vector composed of the health state indicator at the corresponding time and each state parameter at the corresponding time; A reference state evolution sequence of the lithium battery is obtained; the distance between the state evolution sequence and the reference state evolution sequence is calculated by dynamic time warping and normalized, serving as an abnormal evolution indicator of the lithium battery; An abnormal evolution threshold is set; if the abnormal evolution indicator is greater than the abnormal evolution threshold, the health state of the lithium battery has an abnormal evolution trend.

10. The lithium battery state of health joint prediction method based on multi-modal timing features of claim 9, wherein: If the health state of the lithium battery has an abnormal evolution trend, feedback adjustment is performed on the prediction of the health state indicator, and specifically includes: The decay rate of the time decay function is increased, the time decay function is uniformly resampled, and the time decay vector is updated; Based on the response lag amount of each state parameter, a weight adjustment factor is assigned to each state parameter respectively, and the weight adjustment factor of any item state parameter is negatively correlated with the corresponding response lag amount; the fusion weight of each state parameter is multiplied by the corresponding weight adjustment factor, and the fusion weight vector is updated; The updated time decay vector and the fusion weight vector are applied to predict the health state indicator of the lithium battery.

Images