A Method for Predicting the State of Health of Lithium-Ion Batteries for Logistics Leasing

Abstract

The present invention provides a method for predicting the health state of lithium-ion batteries for logistics leasing. This method first extracts the performance parameters of each battery group in each charge-discharge cycle, and calculates the maximum value, minimum value, and average value of the performance parameters as statistical features. Clustering is performed on the statistical features of all battery charge-discharge segments to obtain a specified number of clusters, and a reference sequence representing the common pattern of each cluster is generated. The dynamic time warping distance between this reference sequence and the standardized charge-discharge segment features is calculated as the feature vector of each battery group in each charge-discharge cycle. Then, the remaining power at the end of each charge-discharge cycle of each battery group is extracted, and the degree of charge-discharge sufficiency is evaluated through probability density distribution, and the batteries are divided into three groups: fully charged and fully discharged, medium charge-discharge degree, and insufficient charge-discharge. The feature vector is used to predict the health state of batteries in different groups. This method improves the prediction accuracy and is applicable to scenarios with obvious changes in charge-discharge characteristics such as lithium-ion batteries for logistics leasing.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion batteries, and particularly to the application of a method for predicting the state of health of lithium-ion batteries by extracting the charging and discharging characteristics and grouping according to the degree of charging and discharging sufficiency in scenarios with significant differences in charging and discharging, such as lithium-ion batteries for logistics electric vehicle rental. Background Art

[0002] Lithium-ion batteries have been widely used in modern electronic devices and electric vehicles, and their charging and discharging characteristics directly affect the performance and lifespan of the batteries. However, the performance prediction of lithium-ion batteries is a complex problem, mainly because the charging and discharging behaviors of lithium-ion batteries under different usage conditions vary significantly, especially in the complex and changeable working scenarios of batteries for logistics electric vehicle rental, where the differences will be more obvious.

[0003] Existing battery lifespan prediction methods usually rely on single feature data or overall data for analysis. For example, only the actual values of current and voltage are used, or all monitoring data are directly used for prediction without feature selection. Such methods often cannot accurately reflect the state of health of the battery when dealing with complex and changeable usage conditions, resulting in a lack of accuracy in actual scenario predictions compared to laboratory test data.

[0004] With the widespread application of electric vehicles and energy storage systems in logistics scenarios, the usage frequency, load, and environmental conditions of leased lithium-ion batteries are becoming increasingly complex and changeable. Therefore, there is an urgent need for a method that can accurately extract the charging and discharging characteristics of lithium-ion batteries under complex usage conditions and group lithium-ion batteries according to the degree of charging and discharging sufficiency, so as to improve the accuracy and reliability of predicting the performance and lifespan of randomly used lithium-ion batteries under complex working conditions. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting the state of health of lithium-ion batteries for logistics electric vehicle rental based on charging and discharging feature extraction and grouping according to the degree of charging and discharging sufficiency, so as to solve the problem of low accuracy in predicting the performance and lifespan of lithium-ion batteries in actual complex application scenarios in the prior art.

[0006] To achieve the above purpose, the present invention provides a method for predicting the state of health of lithium-ion batteries for logistics rental, including the following steps:

[0007] S1: Extract the performance parameters of each lithium-ion battery in each charging and discharging cycle, including current, voltage, and temperature, and calculate the maximum value, minimum value, and average value of the current, voltage, and temperature respectively as statistical features. Cluster the statistical features of all batteries in each charging and discharging section. After clustering, obtain a specified number of clusters, and generate a reference sequence that can represent the common pattern of each cluster;

[0008] S2: For each charge-discharge cycle of each battery, calculate the minimum cumulative dynamic time warping distance between the standardized feature sequence of the battery in this charge-discharge cycle and the reference sequence of each cluster, and convert the minimum cumulative dynamic time warping distance into a vector, which serves as the dynamic time warping feature vector of the battery in this charge-discharge cycle finally;

[0009] S3: Respectively extract the remaining power of each battery at the end of each charge-discharge cycle for probability statistics, extract the peak value of the probability density distribution of the probability statistics, obtain the charge-discharge sufficiency degree of the battery, and group according to the charge-discharge sufficiency degrees of each battery;

[0010] S4: Using the dynamic time warping sequence feature vectors in the charge-discharge cycles of each group of lithium-ion batteries as features, and using the machine learning extreme gradient boosting tree algorithm, perform state-of-health prediction on the lithium-ion batteries in different charge-discharge degree groups respectively.

[0011] Preferably, in the said S1, it includes:

[0012] S11: Obtain the original feature sequence. Among them, the original feature sequence of the jth charging segment of the ith battery is in the following format, and the jth charging segment is the charging segment in the jth charge-discharge cycle:

[0013] , where is the maximum current value of the jth charging segment, is the minimum current value of the jth charging segment, is the average current value of the jth charging segment, is the maximum voltage value of the jth charging segment, is the minimum voltage value of the jth charging segment, is the average voltage value of the jth charging segment, is the maximum temperature value of the jth charging segment, is the minimum temperature value of the jth charging segment, is the average temperature value of the jth charging segment.

[0014] Preferably, in the said S2, it includes:

[0015] S21: Respectively perform feature standardization on the original feature sequence of each battery to obtain the standardized feature sequence;

[0016] S22: According to each standardized feature sequence, adopt principal component analysis for dimensionality reduction, and perform K-means clustering on the standardized feature sequences of all battery charging segments and discharging segments after dimensionality reduction respectively to generate a specified number of charging clusters and discharging clusters;

[0017] S23: For the generated charging clusters and discharging clusters, use the time series averaging algorithm based on dynamic time warping to generate a reference sequence for each cluster;

[0018] S24: Calculate the minimum cumulative dynamic time warping distance between the standardized feature sequence of each battery and the reference sequence of each cluster in the clustering, vectorize the dynamic time warping distance, and generate the final dynamic time warping feature vector;

[0019] S25: Obtain the dynamic time warping feature vectors of all batteries in each charge-discharge cycle. For a certain charge-discharge cycle c, calculate the minimum cumulative dynamic time warping distance between the standardized charging feature sequence and the reference sequences of each charging cluster in the charging section, and calculate the dynamic time warping distance between the standardized discharging feature sequence and the reference sequences of each discharging cluster in the discharging section. Concatenate the calculated dynamic time warping distances to obtain the final dynamic time warping feature vector of this charge-discharge cycle in the form of: , where is the feature of the charging section, is the feature of the discharging section.

[0020] Preferably, in S21, the standardized feature sequence includes the following steps:

[0021] For each battery, based on each feature of the charge-discharge section of the battery, calculate the global mean and the standard deviation , where k ranges from 1 to 9;

[0022]

[0023]

[0024] where is the number of all charging or discharging sections of the battery, is the value of the k-th feature in the l-th charging or discharging section;

[0025] After obtaining the global mean and the standard deviation of the battery, standardize each feature of the battery through the following formula:

[0026]

[0027] where is the standardized value of the k-th feature in the l-th charging or discharging section, and generate the standardized feature sequence of the battery , .

[0028] Preferably, in S23, generating a reference sequence includes the following steps:

[0029] For each cluster, randomly select an original feature sequence in a charging or discharging cycle of a certain lithium-ion battery in the cluster as the current reference sequence. For the original feature sequences of other batteries, use dynamic time warping to calculate the distance between the original feature sequence of the other battery and the current reference sequence, find the optimal alignment path, calculate the average value at each position after alignment, generate a new reference sequence to replace the current reference sequence, and repeat this step until the difference between the current reference sequence and the previous reference sequence is less than a certain set threshold, which is considered to converge, and obtain a reference sequence generated for the cluster.

[0030] Preferably, in S24, the following steps are included:

[0031] The dynamic time warping distance calculation formula is ,

[0032] where, represents the standardized charge-discharge feature sequence and the reference sequence of each cluster the dynamic time warping distance between them; for multi-dimensional feature vectors and , the distance is defined as:

[0033]

[0034] The value of the dynamic time warping distance matrix D at the position is calculated as:

[0035]

[0036] At a specific time point , the dynamic time warping distance is , and the minimum cumulative dynamic time warping distance between the standardized feature sequence of each battery and the reference sequence of each cluster in the clustering can be calculated. By starting from the first position (1,1) of the matrix and calculating, each position in the first row is accumulated from left to right, and each position in the first column is accumulated from top to bottom, and the distance of each step is accumulated in turn until the lower right corner D(P, Q) of the matrix is calculated. D(P, Q) represents the minimum dynamic time warping distance between the two sequences, which is obtained by continuously calculating backward from the first position of the matrix. P and Q are the lengths of the two sequences respectively. Thus, a complete dynamic time warping distance matrix can be obtained, and the element in the lower right corner of the matrix is the required minimum dynamic time warping distance. Therefore, the value of each position is obtained by selecting the minimum cumulative distance among the upper, left, or diagonal three positions.

[0037] Preferably, in S3, the charging and discharging sufficiency grouping method includes the following sub-steps:

[0038] S31: Calculate the probability density distribution function of the remaining power of each battery according to the remaining power at the end of the charging and discharging sections in each charging and discharging cycle of each battery.

[0039] S32: Based on the position of the peak in the probability density function of the remaining power of the battery, obtain the remaining power value corresponding to the peak of the probability density distribution of the remaining power. Based on the remaining power value corresponding to the peak of the probability density distribution of the remaining power and the experience of healthy charging and discharging, determine the grouping thresholds for the remaining power. The charging thresholds are 50% and 80% respectively, and the discharging thresholds are 40% and 50% respectively.

[0040] S33: For the charging section, if the peak of the probability density distribution of the remaining power is above 80% of the full charge, it is considered that the charging is full; if the peak of the probability density distribution of the remaining power is at medium power, that is, 50%-80% of the full charge, it is considered that the charging sufficiency is medium; if the peak of the probability density distribution of the remaining power is below 50% of the full charge, it is considered that the charging is insufficient.

[0041] S34: For the discharging section, if the peak of the probability density distribution of the remaining power is above 50% of the full charge, it is considered that the discharging is insufficient; if the peak of the probability density distribution of the remaining power is at medium power, that is, 40%-50% of the full charge, it is considered that the discharging sufficiency is medium; if the peak of the probability density distribution of the remaining power is below 40% of the full charge, it is considered that the discharging is sufficient.

[0042] S35: According to the comprehensive situation of the charging and discharging grading, the batteries are divided into three groups as a whole, namely fully charged and fully discharged, medium charging and discharging degree, and both insufficient charging and discharging.

[0043] The present invention also provides a health state prediction system for lithium-ion batteries used in logistics leasing, which executes the above-mentioned health state prediction method for lithium-ion batteries used in logistics leasing, and specifically includes a capacity calculation module, a data processing module, a feature extraction module based on charging and discharging, a grouping module based on charging and discharging sufficiency, and a lithium-ion battery grouping prediction module.

[0044] The capacity calculation module is used to calculate the specific capacity of the charge and discharge segments, as a characterization of the lithium-ion battery life attenuation; the data processing module is used to process the required features (such as current, voltage, temperature, etc.) and the target (capacity) according to certain rules, and the charge and discharge-based feature extraction module is responsible for extracting available features that can be used in the charge and discharge sufficiency-based grouping module and the prediction module through machine learning algorithms; the grouping module based on the charge and discharge sufficiency groups according to the remaining power of the extracted charge and discharge segments, divides the batteries into different groups for the grouping prediction module; the lithium-ion battery grouping prediction module. Finally, the remaining capacity and health status of the lithium-ion battery are predicted according to different groups, and accurate prediction results can be obtained for each group.

[0045] Further, the specific process of the capacity calculation module includes:

[0046] First, calculate the capacity change amount of the charge / discharge segment. The specific calculation formula is , where is the capacity change amount, and are the initial and end time points of the capacity segment to be calculated respectively, is the current.

[0047] Then, according to the initial capacity and the capacity change amount, the capacity value of the current state can be obtained. The specific calculation formula is , C is the current capacity, is the capacity at the end of the previous cycle.

[0048] Further, the specific process of the data extraction module includes:

[0049] Extract the performance parameters of each lithium-ion battery in each charge and discharge cycle, including current, voltage and temperature, and calculate the maximum value, minimum value and average value of the current, the voltage and the temperature respectively as statistical features,

[0050] Further, the specific steps of the feature extraction module include:

[0051] First, based on the existing data, separate the charge segment and the discharge segment to obtain the data of all charge segments and discharge segments. For each charge segment and discharge segment, by calculating the statistical features, a set of consistent feature sequences are obtained. The features are represented as

[0052] , where is the maximum current of the jth charge segment, is the minimum current of the jth charge segment, is the average current of the jth charge segment, is the maximum voltage of the jth charge segment, is the minimum voltage of the j-th charging segment, is the average voltage of the j-th charging segment, is the maximum temperature of the j-th charging segment, is the minimum temperature of the j-th charging segment, is the average temperature of the j-th charging segment.

[0053] Next, in order to eliminate the influence of different feature scales and ensure that each feature contributes equally to the calculation of the Dynamic Time Warping (DTW) distance, the original feature sequences of each battery are feature standardized to obtain the standardized feature sequences. To standardize the original features, first calculate the global mean and standard deviation for each feature of the charging and discharging segments of the battery during all usage cycles. The specific calculation method is as shown in the formula: , . Among them, is the number of all charging and discharging segments of the battery, is the value of the k-th feature in the l-th charging and discharging segment. After obtaining the mean and standard deviation, each feature is standardized through the following formula: , where is the standardized value of the k-th feature in the l-th charging or discharging segment, is the original feature value, generating the standardized feature sequence , .

[0054] Subsequently, in order to set the DTW reference sequence for each battery, cluster the statistical features of the charging and discharging segments of all batteries. After clustering, a specified number of clusters are obtained, and each cluster represents a similar charging and discharging behavior pattern of lithium-ion batteries. For the result of the generated clustering clusters, use the time series average algorithm based on dynamic time warping to generate a reference sequence for each cluster as the reference sequence for subsequent DTW matching. The time series average algorithm based on dynamic time warping (DTW Barycenter Averaging, DBA) is a method for generating the center of a time series, which makes the sequences in the cluster closest to its center through gradual iteration. Specifically, DBA can generate a representative center sequence that can capture the common features of all time series in the cluster to the greatest extent. For each clustering cluster, apply the DBA algorithm to calculate the reference sequence of the cluster. Specifically, randomly select an original feature sequence in a charging or discharging cycle of a lithium-ion battery in the current cluster as the initial reference sequence R 1 = [r 1 , r 2, …, r 9 , for each other charge-discharge sequence of the battery, use dynamic time warping to calculate its distance from the current original reference sequence, find the optimal alignment path, calculate the average value at each position after alignment, generate a new reference sequence, and repeat this step until the difference between the current reference sequence and the previous reference sequence is less than a certain set threshold, which is considered to converge, and obtain the final reference sequence corresponding to the original characteristic values standardized by the charge-discharge cycles of the lithium-ion batteries in each cluster R q。 The specific formula is , where is the charge-discharge segment sequence in this cluster , is the reference sequence generated by the i-th cluster

[0055] The dynamic time warping distance calculation formula is

[0056]

[0057] where represents the standardized charge-discharge characteristic sequence and the reference sequence of each cluster The dynamic time warping distance between them; for the multi-dimensional feature vectors and , the distance is defined as:

[0058]

[0059] The value of the dynamic time warping distance matrix D at the position is calculated and defined as:

[0060]

[0061] At a specific time point the dynamic time warping distance is , the minimum cumulative dynamic time warping distance between the standardized characteristic sequence of each battery and the reference sequence of each cluster in the clustering can be calculated. By starting from the first position (1,1) of the matrix, each position in the first row is accumulated from left to right, and each position It is cumulative from top to bottom, successively adding the distances of each step until the minimum dynamic time warping distance D(P, Q) at the lower right corner of the matrix is calculated. D(P, Q) represents the minimum dynamic time warping distance between two sequences and is obtained by continuously calculating backward from the first position of the matrix. P and Q are the lengths of the two sequences respectively. Thus, a complete dynamic time warping distance matrix can be obtained, where the element at the lower right corner of the matrix is the required minimum dynamic time warping distance. Therefore, the value of each position is obtained by selecting the minimum cumulative distance among the three positions above, to the left, or the diagonal. And the minimum cumulative distance is transformed into a vector to generate the final DTW feature vector, which contains the charge and discharge characteristics within this period.

[0062] For a certain charge and discharge cycle c, the final feature vector is in the form of , where is the charging section feature, is the discharging section feature.

[0063] Furthermore, the specific steps of the grouping module based on the remaining battery capacity are as follows:

[0064] According to the probability density distribution of the remaining battery capacity, the charge and discharge process is divided into different groups; first, calculate the probability density function of the remaining battery capacity of each battery. The specific calculation formula is: , where is the estimated density function, representing the probability density at ; is the number of samples, that is, the total number of measured values of the remaining battery capacity in the dataset; is each sample data point, that is, each measured remaining battery capacity value in the dataset; is the bandwidth (smoothing parameter), which controls the smoothness of the kernel density estimation. The larger the bandwidth, the smoother the estimated curve; is the kernel function. In this design, the Gaussian kernel function is used; the main role of the Gaussian kernel function is to assign a weighted value to each data point, and the size of the weight is determined according to the distance from the point to the target position. The closer the distance, the larger the weight; the farther the distance, the smaller the weight. The specific calculation formula is: , where is the value of the Gaussian kernel function; represents the target position and the standardized distance between the sample data point ; represents the exponential function;

[0065] Based on the probability density function of the remaining battery capacity of the battery and the healthy charge and discharge experience, the grouping thresholds of the remaining battery capacity are comprehensively determined. The grouping thresholds of the remaining battery capacity in the charging section are set to 50% and 80%, and the grouping thresholds of the remaining battery capacity in the discharging section are set to 40% and 50%.

[0066] For the charging stage, if the peak of the probability density distribution of the remaining power is above 80% of the full capacity, it is considered that the charging is full; if the peak of the remaining power distribution is at medium power (50%-80% of the full capacity), it is considered that the degree of charging is medium; if the peak of the remaining power distribution is below 50% of the full capacity, it is considered that the charging is insufficient; for the discharging stage, if the peak of the probability density distribution of the remaining power is above 50% of the full capacity, it is considered that the discharging is insufficient; if the peak of the remaining power distribution is at medium power (40%-50% of the full capacity), it is considered that the degree of discharging is medium, and if the peak of the remaining power distribution is below 40% of the full capacity, it is considered that the discharging is sufficient;

[0067] According to the comprehensive situation of charge and discharge grading, the overall battery is divided into three groups, namely full charge and full discharge, medium charge and discharge degree, and both charge and discharge are insufficient. They respectively represent the charge and discharge saturation degree of the battery, from high to low, that is, full charge and full discharge represent full charging and sufficient discharging, the medium charge and discharge degree group represents that one of charging or discharging is not sufficient, and the both charge and discharge are insufficient group represents that both charging and discharging are not sufficient, or even worse;

[0068] Furthermore, for the grouping prediction module, it specifically includes the following content:

[0069] For an ungrouped battery, first analyze it to determine which group it belongs to; then use machine learning algorithms Random Forest (RF), XGB, and Long Short-Term Memory (LSTM) to predict the battery life for the data of the three groups respectively, and train a most suitable model for each group; then use the model of this group to train the battery.

[0070] Through the above method, the accuracy and robustness of battery life and performance prediction can be effectively improved. The experimental results show that the method of the present invention significantly improves the accuracy of the prediction model, thereby prolonging the service life of the battery and improving safety.

[0071] The method and system of the present invention not only improve the accuracy of battery life and performance prediction, but also enhance the robustness of the model, and are applicable to complex and changeable charge and discharge working conditions such as logistics scenarios.

[0072] In summary, the present invention proposes a new method for predicting battery performance and life by extracting the charge and discharge characteristics of lithium-ion batteries and grouping the charge and discharge degrees, effectively solving the problems of prediction accuracy and robustness in the prior art, and having broad application prospects and significant practical significance.

[0073] The present invention has the following beneficial effects compared with the prior art:

[0074] 1. The present invention proposes a new feature extraction method that takes into account both dynamic time series features and automatic data pattern extraction. This method combines the dynamic matching ability of DTW and the pattern recognition ability of clustering methods, and can extract representative fixed-length feature vectors from a large number of variable-length sequences. This feature extraction method can effectively process large-scale variable-length sequence data and avoid information loss caused by interpolation or clipping, etc.

[0075] The time series averaging algorithm (DBA) for generating representative sequences by DTW clustering provides an objective, data-driven feature extraction scheme, and provides a new idea for extracting features for predicting the state of health of batteries in actual scenarios.

[0076] 2. When the present invention uses deep learning to predict the life of lithium-ion batteries, it statistically analyzes the probability density distribution of the remaining battery power at the end of each charge and discharge cycle of the battery. According to the charge and discharge degree, the batteries are divided into three groups: fully charged and fully discharged, medium charge and discharge degree, and insufficient charge and discharge. This helps to classify the attenuation situation of the batteries, so that the batteries with the same attenuation situation can be trained together, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, without creative efforts, other implementation drawings can be obtained according to the provided drawings.

[0078] Figure 1 It is a flowchart of a life prediction method based on charge and discharge feature grouping proposed by the present invention.

[0079] Figure 2 It is a flowchart of DTW feature extraction in Embodiment 1.

[0080] Figure 3 It is a probability density distribution diagram of the remaining battery power of 30 groups of battery charge and discharge segments in Embodiment 1.

[0081] Figure 4 It is a battery capacity attenuation diagram of 30 groups in Embodiment 1.

[0082] Figure 5 It is a schematic diagram of the battery grouping strategy.

[0083] Figure 6 It is a comparison diagram of the prediction effects of using DTW feature life prediction and using other features in Embodiment 1.

[0084] Figure 7It is a comparison result graph of the battery grouping life prediction and the overall prediction of the ungrouped battery in Example 1. Specific implementation method

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] Example 1:

[0087] The present invention provides a grouping method based on the charging and discharging characteristics of lithium-ion batteries, as Figure 1 shown, including the following steps:

[0088] In actual logistics scenarios, lithium-ion batteries are widely used as the batteries of distribution vehicles. The analysis of the battery state and the prediction of the battery life in the logistics scenario can provide a basis for battery management and has strong economic value. In this example, for 30 lithium-ion batteries in the actual operating logistics scenario, the battery management system (BMS) is used to obtain the characteristic data at the end of each charge and discharge cycle of the battery, including the remaining power, voltage, current, temperature, etc.

[0089] The specific steps of feature generation are as Figure 2 , first, based on the existing data, the charging section and the discharging section are separated to obtain the data of all charging sections and discharging sections. For each charging section and discharging section, the performance parameters in the charge and discharge cycle are extracted, including current, voltage and temperature, and the maximum value, minimum value and average value of the current, the voltage and the temperature are calculated respectively as statistical features to obtain a set of consistent feature sequences, and the features are represented as

[0090] , where is the maximum current value of the jth charging section, is the minimum current value of the jth charging section, is the average current value of the jth charging section, is the maximum voltage value of the jth charging section, is the minimum voltage value of the jth charging section,

[0091] is the average voltage value of the jth charging section, is the maximum temperature value of the jth charging section, is the minimum temperature value of the jth charging section, is the average temperature of the j-th charging segment. These feature arrays cover the key indicators during the charging and discharging processes, ensuring that the dimensionality of the feature vectors for each charging and discharging segment is consistent.

[0092] Next, to eliminate the influence of different feature scales and ensure that each feature contributes evenly to the DTW distance calculation, the features are standardized. To standardize the original features, first calculate the global mean of each feature for all battery charging and discharging segments within all usage cycles and the standard deviation , and the specific calculation method is as shown in the formula: , . Among them, is the number of all charging and discharging segments of this battery, is the value of the k-th feature in the l-th charging and discharging segment. After obtaining the mean and standard deviation, each feature is standardized through the following formula: , where is the standardized value of the k-th feature in the l-th charging or discharging segment, is the original feature value, generating a standardized feature sequence , .

[0093] Subsequently, to set the DTW reference sequence, cluster the statistical features of all battery charging and discharging segments to generate a specified number of clusters. In this example, the number of charging and discharging clusters is set to 5 each, and a reference sequence that can represent the common pattern of each cluster is generated as the reference sequence for subsequent DTW matching. DBA is a method for generating the center of a time series, which makes the sequences in the cluster closest to its center through gradual iteration. Specifically, DBA can generate a representative center sequence that can maximize the capture of the common features of all time series in the cluster. For each cluster, apply the DBA algorithm to calculate the reference sequence of the cluster. Specifically, first randomly select a sequence as the initial reference sequence R 1 = [r1, r2, …, r9], for each input time series, use DTW to calculate its distance from the current reference sequence, find the optimal alignment path, calculate the average value at each position after alignment, and generate a new reference sequence. Repeat the steps until the reference sequence converges. Thus, the final reference sequence for each cluster is obtained. The specific formula is , where, is the charging and discharging segment sequence in this cluster , is the reference sequence generated by the i-th cluster.

[0094] Calculate the DTW distance between the original feature vector and the reference sequence, specifically , where, Represents the standardized charge-discharge sequence (including the characteristics in the charge-discharge section) and the final reference sequence The DTW distance between them

[0095] For multi-dimensional feature vectors and , the distance is defined as: . The cumulative distance matrix The calculation is defined as: At a specific time point The dynamic time warping distance is , and the minimum cumulative dynamic time warping distance between the standardized feature sequence of each battery and the reference sequence of each cluster in the clustering can be calculated. By starting from the first position (1,1) of the matrix and calculating, each position in the first row is accumulated from left to right, and each position in the first column is accumulated from top to bottom. The distance of each step is accumulated in turn until the bottom right corner D(P, Q) of the matrix is calculated. D(P, Q) represents the minimum dynamic time warping distance between the two sequences, which is continuously calculated backward from the first position of the matrix. P and Q are the lengths of the two sequences respectively. Thus, a complete dynamic time warping distance matrix can be obtained, and the element in the bottom right corner of the matrix is the required minimum dynamic time warping distance. Therefore, the value of each position is obtained by selecting the minimum cumulative distance among the three positions above, to the left, or the diagonal, and vectorizing to generate the final DTW feature vector, which includes the charge-discharge characteristics during this period

[0096] For a certain charge-discharge cycle c, the final feature vector is in the form of , where is the charge section characteristic is the discharge section characteristic

[0097] Thus, each charge-discharge cycle will obtain a feature vector with a length of [1, 2q]. Specifically in this example, a feature vector of [1, 10] is obtained to train the model as the feature of each charge-discharge cycle

[0098] According to the probability density distribution of the remaining battery power, the charge-discharge process is divided into different groups. First, calculate the probability density function. Use software to calculate the probability density function of the remaining battery power. The specific calculation formula of the probability density is: , where is the estimated density function, representing the probability density at ; is the number of samples, that is, the total number of measured values of the remaining battery power in the dataset; is each sample data point, i.e., each measured remaining battery capacity value in the dataset; is the bandwidth (smoothing parameter), which controls the smoothness of the kernel density estimation. The larger the bandwidth, the smoother the estimated curve; is the kernel function. In this design, the Gaussian kernel function is used; the main role of the Gaussian kernel function is to assign a weighted value to each data point, and the weight size is determined according to the distance from this point to the target position. The closer the distance, the larger the weight; the farther the distance, the smaller the weight. The Gaussian kernel function helps to smoothly estimate the probability density, especially when the data points are sparse or there is noise. Kernel density estimation is a non-parametric method, which means it does not assume that the data follows a specific distribution (such as a normal distribution). The Gaussian kernel function is only used to smoothly estimate the distribution of the data, rather than assuming the true distribution of the data. The specific calculation formula of the Gaussian kernel function is: where is the value of the Gaussian kernel function; represents the target position and the sample data point the standardized distance between; represents the exponential function; the probability density distribution maps of 30 groups of batteries can be obtained, as shown in Figure 3 shown.

[0099] Based on the remaining battery capacity probability density function of the battery and the healthy charge and discharge experience, the grouping thresholds of the remaining battery capacity are comprehensively determined. The grouping thresholds of the remaining battery capacity in the charging section are set to 50% and 80%, and the grouping thresholds of the remaining battery capacity in the discharging section are set to 40% and 50%;

[0100] The specific logic of the overall grouping is as shown in Figure 5 shown. For the charging section, if the peak of the remaining battery capacity probability density distribution is above 80% of the full capacity, it is considered that the charging is full; if the peak of the remaining battery capacity distribution is at medium battery capacity (50% - 80% of the full capacity), it is considered that the charging adequacy is medium; if the peak of the remaining battery capacity distribution is below 50% of the full capacity, it is considered that the charging is insufficient; for the discharging section, if the peak of the remaining battery capacity probability density distribution is above 50% of the full capacity, it is considered that the discharging is insufficient; if the peak of the remaining battery capacity distribution is at medium battery capacity (40% - 50% of the full capacity), it is considered that the discharging adequacy is medium, and if the peak of the remaining battery capacity distribution is below 40% of the full capacity, it is considered that the discharging is sufficient;

[0101] According to the comprehensive situation of charge and discharge grading, the overall battery is divided into three groups, namely full charge and full discharge, medium charge and discharge degree, and insufficient charge and discharge. They respectively represent the charge and discharge saturation degree of the battery, from high to low. That is, full charge and full discharge means full charge and sufficient discharge, the medium charge and discharge degree group means that one of the charge or discharge is not sufficient, and the insufficient charge and discharge group means that both the charge and discharge are not sufficient, or even poor. There are 14, 6, and 7 batteries in the three groups respectively. Another three batteries were not put into operation during this period, and no valuable data was collected, so they were not grouped. The capacity attenuation curves of the three groups are as Figure 4 shown.

[0102] Use machine learning algorithms to predict the battery life and performance of data in different groups; the present invention selects Random Forest, Gradient Boosting Machine, and Long Short-Term Memory Network (LSTM) as machine learning algorithms; use the grouped data to train the machine learning model. In order to verify the effectiveness of DTW feature extraction, we respectively plotted the prediction results after grouping for the selected DTW features, all features, and the best feature subset obtained by the conventional method, and selected four groups of batteries as examples. They are from different groups respectively, and the prediction results are as Figure 6 shown, and it is found that the prediction results of DTW features are all optimal. Compared with other methods, the prediction effect of the DTW method has been improved by an average of 80%, and the average root mean square error has reached 1.48. Use the training set to train the model, and use the validation set to optimize the model to ensure the accuracy and robustness of the model; input the new data into the trained model to predict the remaining life and performance of the battery. The comparison of the prediction results of grouped prediction and ungrouped overall prediction is as Figure 7 shown, and it is found that the prediction effect of grouped prediction has almost doubled compared with overall prediction.

[0103] The evaluation index of each specific battery is shown in Table 1, with RMSE as the evaluation index:

[0104] Table 1

[0105] Battery / Method / RMSE (%) Battery No. 1 Battery No. 2 Battery No. 3 Battery No. 4 Original Features 9.81 9.15 7.18 11.59 Best Subset of Conventional Method 1.88 1.76 3.75 1.65 DTW Features 0.49 0.62 4.43 0.41

[0106] Calculate the average RMSE of each method:

[0107] Original features: (9.81 + 9.15 + 7.18 + 11.59) / 4 = 9.43

[0108] Best subset of conventional method: (1.88 + 1.76 + 3.75 + 1.65) / 4 = 2.26

[0109] DTW feature: (0.49 + 0.62 + 4.43 + 0.41) / 4 = 1.49

[0110] 1. Reduction ratio of RMSE of DTW compared to the original feature:

[0111] (9.43 - 1.49) / 9.43×100% = 84.2%

[0112] 2. Reduction ratio of RMSE of DTW compared to the best subset of the conventional method:

[0113] (2.26 - 1.49) / 2.26×100% = 34.2%

[0114] Result:

[0115] RMSE of DTW is reduced by 84.2% compared to the original feature

[0116] RMSE of DTW is reduced by 34.2% compared to the best subset of the conventional method

[0117] The above-described embodiments and / or implementation manners are only used to illustrate the preferred embodiments and / or implementation manners for implementing the technology of the present invention, and do not impose any form of limitation on the implementation manners of the technology of the present invention. Any person skilled in the art, without departing from the scope of the technical means disclosed in the content of the present invention, may make some changes or modifications to other equivalent embodiments, but should still be regarded as the same technology or embodiment as the present invention in essence.

Claims

1. A method for predicting the health status of lithium-ion batteries for logistics leasing, characterized in that: The following steps are involved: S1: extracting the performance parameters of each lithium-ion battery in each charge and discharge cycle, including current, voltage and temperature, and calculating the maximum, minimum and average values ​​of the current, voltage and temperature as statistical features, clustering the statistical features of all batteries in each charge and discharge segment, obtaining a specified number of clusters after clustering, and generating a reference sequence that can represent the common mode of each cluster; S2: for each charge and discharge cycle of each battery, calculate the minimum cumulative dynamic time warping distance between the standardized feature sequence of the battery in the charge and discharge cycle and the reference sequence of each cluster, and convert the minimum cumulative dynamic time warping distance into a vector as the final dynamic time warping feature vector of the battery in the charge and discharge cycle; S3: extract the remaining power of each battery at the end of each charge and discharge cycle for probability statistics, extract the probability density distribution peak of the probability statistics, obtain the charge and discharge sufficiency of the battery, and group the batteries according to the charge and discharge sufficiency of each battery; S4: Using the dynamic time warping feature vector of each group of lithium-ion batteries in the charge and discharge cycle as a feature, the machine learning extreme gradient boosting tree algorithm is used to predict the health status of lithium-ion batteries in different charge and discharge degree groups; The S2 includes: S21: The original feature sequence of each battery Perform feature standardization to obtain a standardized feature sequence; S22: according to each standardized feature sequence, principal component analysis is used for dimensionality reduction, and K-means clustering is performed on the standardized feature sequences of all battery charging segments and discharging segments after dimensionality reduction, to generate a specified number of charging clusters and discharging clusters; S23: for the generated charging clusters and discharging clusters, a reference sequence is generated for each cluster using a time series averaging algorithm based on dynamic time warping; S24: calculating the minimum cumulative dynamic time warping distance between the standardized feature sequence of each battery and the reference sequence of each cluster, vectorizing the dynamic time warping distance, and generating a final dynamic time warping feature vector; S25: Obtain the dynamic time warping feature vectors of all batteries in each charge and discharge cycle. For a certain charge and discharge cycle c, calculate the minimum cumulative dynamic time warping distance between the standardized charge feature sequence and the reference sequence of each charge cluster in the charge segment, and calculate the dynamic time warping distance between the standardized discharge feature sequence and the reference sequence of each discharge cluster in the discharge segment. Concatenate the calculated dynamic time warping distances to obtain the final dynamic time warping feature vector of the charge and discharge cycle. The form is: ,in, is the charging stage characteristic, It is the characteristic of the discharge stage.

2. The method according to claim 1, characterized in that The S1 includes: S11: Obtaining an original feature sequence, wherein the original feature sequence of the jth charging segment of the i-th battery is in the following format, and the jth charging segment is a charging segment in the jth charge-discharge cycle: ,in is the maximum current of the jth charging segment, is the minimum current value of the jth charging segment, is the average current of the jth charging segment, is the maximum voltage of the jth charging segment, is the minimum voltage of the jth charging segment, is the average voltage of the jth charging segment, is the maximum temperature of the jth charging segment, is the minimum temperature of the jth charging segment, is the average temperature of the jth charging segment.

3. The method according to claim 2, characterized in that In S21, the standardized feature sequence includes the following steps: For each battery, the global mean of the battery is calculated based on each feature of the battery's charge and discharge stages. and standard deviation , k ranges from 1 to 9; in, is the number of all charging or discharging stages of the battery, is the value of the kth feature in the lth charging or discharging segment; After obtaining the global mean and the standard deviation of the battery, each feature of the battery is standardized by the following formula: in The standardized value of the kth feature in the lth charging or discharging segment generates the standardized feature sequence of the battery , .

4. The method according to claim 2, characterized in that: In S23, generating a reference sequence includes the following steps: For each cluster, an original feature sequence in the charging or discharging cycle of a lithium-ion battery in the cluster is randomly selected as the current reference sequence. For the original feature sequences of other batteries, dynamic time warping is used to calculate the distance between the original feature sequences of other batteries and the current reference sequence, and the optimal alignment path is found. The average value at each position after alignment is calculated, and a new reference sequence is generated to replace the current reference sequence. This step is repeated until the difference between the current reference sequence and the previous reference sequence is less than a set threshold, which is considered to converge, and a reference sequence generated for the cluster is obtained.

5. The method according to claim 2, characterized in that: In S24, the following steps are included: The dynamic time warping distance calculation formula is: , in, Represents the standardized charge and discharge characteristic sequence The reference sequence of each cluster The dynamic time warping distance between them; for multidimensional feature vectors and , the distance is defined as: The dynamic time warping distance matrix D is at position Value The calculation of is defined as: At a specific point in time Under this condition, the dynamic time warping distance is , the minimum cumulative dynamic time warping distance between the standardized feature sequence of each battery and the reference sequence of each cluster can be calculated by starting from the first position (1,1) of the matrix, and each position in the first row It is accumulated from left to right, each position in the first column It is accumulated from top to bottom, accumulating the distance of each step in turn until the calculation reaches the lower right corner of the matrix D(P, Q). D(P,Q) represents the minimum dynamic time warping distance of the two sequences. It is calculated from the first position of the matrix and P and Q are the lengths of the two sequences respectively. Thus, a complete dynamic time warping distance matrix is ​​obtained, in which the element in the lower right corner of the matrix is ​​the required minimum dynamic time warping distance. Therefore, the value of each position is the minimum cumulative distance of the three positions above, left or diagonal.

6. The method according to claim 1, characterized in that In S3, the charge and discharge sufficiency degree grouping method includes the following sub-steps: S31: Calculate the probability density distribution function of the remaining power of each battery according to the remaining power at the end of the charging and discharging stages in each charging and discharging cycle of each battery; S32: according to the peak position of the battery's remaining power probability density function, obtain the remaining power value corresponding to the peak value of the remaining power probability density distribution, and determine the grouping threshold of the remaining power based on the remaining power value corresponding to the peak value of the remaining power probability density distribution and the healthy charging and discharging experience, the charging and discharging thresholds are 50% and 80% and 40% and 50% respectively; S33: For the charging stage, if the peak value of the probability density distribution of the remaining power is above 80% of the full charge, it is considered to be fully charged; if the peak value of the probability density distribution of the remaining power is at a medium charge, i.e., 50%-80% of the full charge, it is considered to be moderately charged; if the peak value of the probability density distribution of the remaining power is below 50% of the full charge, it is considered to be insufficiently charged; S34: For the discharge stage, if the peak value of the probability density distribution of the remaining power is above 50% of the full rating, the discharge is considered to be insufficient; if the peak value of the probability density distribution of the remaining power is at the medium power level, i.e. 40%-50% of the full rating, the discharge is considered to be medium sufficient; if the peak value of the probability density distribution of the remaining power is below 40% of the full rating, the discharge is considered to be sufficient; S35: According to the comprehensive situation of charge and discharge classification, the battery is divided into three groups, namely, fully charged and discharged, medium charge and discharge, and insufficient charge and discharge.

7. A health status prediction system for lithium-ion batteries for logistics leasing, executing the health status prediction method for lithium-ion batteries for logistics leasing as claimed in any one of claims 1 to 6, characterized in that: include: Charge and discharge feature extraction module: extracts the performance parameters of each lithium-ion battery in each charge and discharge cycle, including current, voltage and temperature, and calculates the maximum, minimum and average values ​​of the current, voltage and temperature as statistical features, clusters the statistical features of all batteries in each charge and discharge segment, obtains a specified number of clusters after clustering, and generates a reference sequence that can represent the common pattern of each cluster; Data processing module: for each charge and discharge cycle of each battery, calculate the minimum cumulative dynamic time warping distance between the standardized feature sequence of the battery in the charge and discharge cycle and the reference sequence of each cluster, and convert the minimum cumulative dynamic time warping distance into a vector as the final dynamic time warping feature vector of the battery in the charge and discharge cycle; Grouping module based on charge and discharge sufficiency: extract the remaining power of each battery at the end of each charge and discharge cycle for probability statistics, extract the probability density distribution peak of the probability statistics, obtain the charge and discharge sufficiency of the battery, and group the batteries according to the charge and discharge sufficiency of each battery; Lithium-ion battery grouping prediction module: Using the dynamic time warping feature vector of each group of lithium-ion batteries in the charge and discharge cycle as the feature, the machine learning extreme gradient boosting tree algorithm is used to predict the health status of lithium-ion batteries in different charge and discharge degree groups.

Images