A Machine Learning-Based Method for Early Lifetime Anomaly Detection in Lithium-ion Batteries

Abstract

This invention provides a machine learning-based method for detecting early life anomalies in lithium-ion batteries, belonging to the field of lithium-ion battery management technology. The method includes: 1) performing accelerated aging tests on multiple lithium-ion batteries and defining battery life labels; 2) applying data processing methods to the post-test dataset; 3) selecting a portion of voltage and current data from the processed dataset as feature vectors to obtain a matrix; 4) determining whether any two batteries have the same label based on the labels; 5) training a neural network model using the matrix and true / false values ​​as input, and binarizing the results to obtain a classifier; 6) selecting one battery and an unknown battery to generate a feature matrix, which is then input into the classifier to complete the detection of early life anomalies in lithium-ion batteries based on machine learning. The method proposed in this invention does not require additional processing of voltage and current data, and has high accuracy and low false alarm rate, possessing significant potential for engineering applications.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery management technology, and particularly relates to a method for detecting early life anomalies in lithium-ion batteries based on machine learning. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles and energy storage power stations due to their high energy density, low self-discharge rate, and long cycle life. However, their manufacturing process is always accompanied by high energy and raw material consumption, and a sufficiently long lifespan is crucial for achieving low carbon emissions and generating a positive environmental impact. The long lifespan of a single battery does not necessarily guarantee a satisfactory lifecycle performance for the battery pack. A battery pack contains hundreds of battery cells connected in series and parallel to meet the power and energy demands of applications such as electric vehicles and renewable energy storage. Individual differences between batteries can lead to variations in battery lifespan, sometimes significant ones. One or more batteries that age faster can have a substantial impact on the lifespan of a large battery pack.

[0003] Many previous studies have emphasized battery screening, the core idea of ​​which is to group batteries with similar key parameters. Currently, the most widely used screening method in the industry is the capacity-resistance (CR) method, which assumes that batteries with similar capacity and resistance values ​​have similar performance. In addition to these two indicators, incremental capacity peak, pulse charging response, voltage trajectory, and electrochemical impedance spectroscopy are also used for battery screening. These tests are fast, generally not exceeding 12 hours. Furthermore, they can effectively screen out observable anomalies, such as high resistance. However, lifetime anomalies, which involve the long-term degradation of the battery's future capacity, are not considered. Identifying battery lifetime anomalies is challenging, especially in the early stages of their lifespan. First, anomalous aging behavior is more easily perceived in the later stages of battery life, while much less information can be extracted in the first few cycles. Therefore, even the best existing algorithms still require data collected in the first 3-5 cycles of the aging process for anomaly detection. In addition to increasing testing time, their prediction errors are as high as 10% to 15%. Second, the low anomaly rate itself poses a challenge to the construction of the dataset. To collect sufficient anomaly samples, we must conduct long-term aging tests on a large number of batteries, which makes the experiment costly and time-consuming. Finally, the correctness of a classification (at the beginning of battery life) can only be experimentally verified after long-term battery use. Delayed feedback hinders algorithm development. Therefore, the early identification of battery life anomalies remains an unsolved problem in the field of battery manufacturing and management. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based method for detecting early life anomalies in lithium-ion batteries, characterized by the following steps:

[0005] S1: Select multiple lithium-ion batteries as sample batteries, conduct accelerated aging tests on the sample batteries, and define battery life labels based on the test results;

[0006] S2: Apply a data processing method to the dataset of accelerated aging test data of the sample batteries. The data processing method includes three steps: data segmentation, normal data reduction, and abnormal data augmentation.

[0007] S3: Select voltage and current data from the processed dataset as the feature vectors of the battery to obtain the matrix;

[0008] S4: Based on the battery life tag in S1, determine whether the tags of any two batteries are the same;

[0009] S5: Train the neural network model offline with the battery feature vector matrix and true / false values ​​as input, binarize the training results, and obtain a classifier to determine whether the battery belongs to the same category.

[0010] S6: Gradually select a battery from the test set of known categories and generate a feature matrix with unknown batteries, input it into the classifier, and complete the detection of early life anomalies in lithium-ion batteries based on machine learning.

[0011] Furthermore, it includes a training part and a testing part. The training part specifically involves: conducting accelerated aging tests on a batch of brand-new sample batteries under the same conditions; observing abnormal battery life through accelerated aging test data and capacity aging curves; obtaining the feature vector of the battery through the first cycle data of the accelerated aging test; then synthesizing the feature vectors of the two batteries into a feature matrix; binarizing the similarities and differences in life labels; and finally using the feature matrix and the binarized matrix as input to train the neural network model to obtain a classifier that determines whether the batteries belong to the same category.

[0012] The testing section specifically involves: progressively selecting one battery from the test set of known categories and generating a feature matrix with unknown batteries, inputting the matrix into the trained classifier, and obtaining the classifier's judgment result.

[0013] Furthermore, in the training section, a batch of brand-new sample batteries are subjected to the same accelerated aging test under the same conditions to simulate abnormal phenomena in the battery itself, thereby shortening the experimental testing time.

[0014] Furthermore, in S2, the data segmentation is as follows: since the proportion of "abnormal" batteries is too low, normal batteries are divided into training set and test set in half when the data is divided, and abnormal batteries are divided using the leave-one-out method.

[0015] Normal data reduction specifically involves: for each abnormal battery in the training set, by examining its aging trajectory, finding the N1 normal batteries that are closest to its aging trajectory, using the selected normal samples as "boundaries" to classify normal and abnormal batteries, thereby improving the network's generalization ability.

[0016] Abnormal data augmentation specifically involves enriching the training set by using data collected in the first N2 cycles to address abnormal batteries that have an excessively low absolute number, resulting in more anomalous data in the experimental data.

[0017] Furthermore, in S3, the selection of the central data point in the dataset specifically refers to: the voltage data during the constant current phase and the current data during the constant voltage phase in the constant current and constant voltage charging cycle of the battery in the first cycle.

[0018] The matrix is ​​represented as: [Γ i ,Γ j ] T Γ in the matrix k Let be the eigenvector matrix of the k-th battery. in L1 represents the voltage sequence during the constant current phase, and L1 represents the duration of the constant current phase. L1 represents the current sequence during the constant voltage phase, L2 represents the duration of the entire charging process, and i and j represent the battery numbers.

[0019] Furthermore, in S4, the judgment result is represented by binary representation as follows: when the two battery tags are the same, it is true, which is [1; 0]; when the two battery tags are different, it is false, which is [0; 1].

[0020] Furthermore, in S5, the neural network model is trained H times to obtain H classifiers, thereby improving the stability of the model and reducing the stacking prediction error.

[0021] Furthermore, in S6, the feature matrix is ​​input into the trained H classifiers. The result of the H classifiers is the weight of the feature matrix that is greater than 50%. The class with the weight of the final class that is greater than 50% is classified into the corresponding class.

[0022] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in:

[0023] 1. Compared with traditional anomaly detection algorithms, such as support vector machines, autoencoders, density-based noisy applied spatial clustering (DBSCAN), isolated forests, k-nearest neighbors and local outlier factor models, the method proposed in this invention does not require additional processing of voltage and current data, and has high accuracy and low false alarm rate, and has great potential for engineering applications.

[0024] 2. Compared with other conventional detection methods, this invention does not require the use of data from multiple cycles for anomaly detection, but only the data from the first cycle, resulting in lower time costs and great application prospects.

[0025] 3. Compared with traditional data partitioning methods, the data processing method proposed in this invention alleviates the problem of severe imbalance in category ratios and overcomes the problem of excessively low absolute number of "abnormal" batteries. Attached Figure Description

[0026] Figure 1 This is a flowchart of a machine learning-based method for detecting early life anomalies in lithium-ion batteries proposed in this invention.

[0027] Figure 2 This is a scatter plot of capacity-internal resistance and a schematic diagram of capacity aging curve for a batch of a certain type of battery.

[0028] Figure 3 This is a schematic diagram of the offline training and testing portion of the method of the present invention.

[0029] Figure 4 This is a schematic diagram of abnormal detection results for the same type of battery.

[0030] Figure 5 Capacity-internal resistance diagrams for misdiagnosed normal batteries and abnormal batteries. Detailed Implementation

[0031] The following will describe in more detail a machine learning-based method for detecting early life anomalies in lithium-ion batteries, with reference to the schematic diagrams, which illustrate preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving its advantageous effects. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.

[0032] like Figure 1 As shown, a machine learning-based method for detecting early life anomalies in lithium-ion batteries includes the following steps:

[0033] Step 1: Select a certain number of batteries as sample batteries, conduct accelerated aging tests on the sample batteries, and define battery life labels.

[0034] Step 2: Apply data processing methods to the training set, which consists of three steps: data splitting, "normal" data reduction, and "abnormal" data augmentation.

[0035] Step 3: Select the voltage data from the constant current phase and the current data from the constant voltage phase during the first cycle of constant current and constant voltage charging of the battery as the battery's feature vector, and obtain the matrix [Γ]. i ,Γ j ]T , of which L1 represents the duration of the constant current phase, L2 represents the duration of the entire charging process, and i and j represent the battery serial numbers.

[0036] Step 4: Using the battery life label from Step 1, determine whether any two batteries have the same label. If both are "normal" or "abnormal", output a true value, which is [1; 0]. Otherwise, output a false value, which is [0; 1].

[0037] Step 5: Train the neural network offline using the battery feature vector matrix and true / false values ​​as input. Binarize the training results to obtain a classifier that determines whether batteries belong to the same category. To improve model stability and reduce random prediction errors, train the neural network model H times to obtain H classifiers.

[0038] Step 6: Gradually select one battery from the test set of known categories and generate a feature matrix with the unknown battery. Input this matrix into H trained classifiers. The battery with a weight greater than 50% belongs to that category. Finally, classes with a weight greater than 50% are classified into the corresponding categories.

[0039] This invention provides a machine learning-based method for detecting early lifespan anomalies in lithium-ion batteries, and also has the following characteristics: the method consists of two parts: offline training and testing.

[0040] In the offline training section, a batch of sample batteries undergoes accelerated aging testing under the same temperature and conditions. Abnormal battery life is observed through accelerated aging test data and capacity aging curves. The feature vectors of the batteries are obtained from the first cycle of accelerated aging test data. Then, the feature vectors of two batteries are combined into a feature matrix, and lifespan labels are binarized for similarities and differences. The above tests require brand-new batteries of the same type and batch to be conducted under identical conditions, resulting in a large amount of battery lifespan test data. The feature matrix and binarized matrix are used as inputs to train the neural network offline, resulting in a classifier that determines whether batteries belong to the same category.

[0041] In the testing section, a feature matrix is ​​formed by progressively selecting the voltage and current feature vectors of the charging process of one battery and an unknown battery from the test set of known categories. This matrix is ​​then input into the trained classifier to obtain the classifier's judgment result. The results of multiple classifiers are then combined and judged, and the category with a weight greater than 50% is the final category of the battery.

[0042] In the offline training section, accelerated aging tests are conducted on brand-new batteries of the same type and batch under identical conditions to obtain a large amount of battery life test data. This data is then used to train the neural network to improve the accuracy and adaptability of the training model. The voltage data from the constant current phase and the current data from the constant voltage phase of the first cycle of constant current / constant voltage charging are selected as the battery's feature vectors, resulting in the dataset [Γ]. i ,Γ j ] T , of which L1 represents the duration of the constant current phase, L2 represents the duration of the entire charging process, and i and j represent the battery serial numbers.

[0043] In the testing section, the battery category can be determined by progressively selecting voltage and current feature vectors from the charging processes of one battery and the unknown battery from a test set of known categories to form a feature matrix. The model's input is voltage and current, requiring no additional processing or transformation, resulting in a low computational burden.

[0044] During offline model training, the classifier can be trained H times, resulting in H classifiers. During testing, for a feature matrix consisting of the same battery and unknown batteries, H classification results can be obtained online. The proportions of these classification results are then calculated, and the category with more than 50% of the results is taken as the final category for that battery.

[0045] Example

[0046] The following example, using a specific 18650 battery, illustrates the method for detecting battery "abnormalities." (Reference) Figure 2 and Figure 3 .

[0047] In the offline training section, accelerated aging tests are performed at the same temperature to identify abnormal battery life conditions. The battery's feature vector is obtained from the first cycle data of the accelerated aging test. This includes the constant current voltage in constant current and constant voltage charging, and then combining the feature vectors of the two batteries into a feature matrix [Γ]. i ,Γ j ] T Then, the lifetime labels are binarized to identify similarities and differences.

[0048] In the testing section, a feature matrix is ​​generated by progressively selecting one battery from the test set of known categories and an unknown battery. This matrix is ​​then input into the trained classifier to obtain the classifier's judgment result, thus completing the battery anomaly detection.

[0049] The model was trained offline using a brand-new battery, and then used to detect anomalies in other batteries. The voltage data from the constant current phase and the current data from the constant voltage phase of the first cycle of constant current / constant voltage charging were selected as the battery's feature vector. This vector was then used to detect anomalies in the test set. The results are as follows: Figure 4 As shown, the model can accurately identify seven abnormal batteries. Furthermore, 100 out of 104 normal batteries were identified, achieving an overall prediction accuracy of 96.4% and a false alarm rate of only 3.8%. Of these, 100 batteries achieved a normality rate greater than 50%, meaning they could be classified as "normal," and 93 batteries had a normality rate above 99%. The highest score among the seven abnormal batteries was only 44.6%, indicating that our prediction was highly confident. Figure 5 The diagram shows the capacity-internal resistance of batteries that were mistakenly identified as normal or abnormal. Analysis revealed that in cases of normal misidentification, 5 out of 7 abnormal batteries had a resistance between 13.45 and 13.57 mΩ, while 3 out of 4 false alarm batteries also fell within this range. Furthermore, 3 out of the 4 false alarm batteries had a capacity between 2.551 and 2.553 Ah. In short, these false alarm batteries are similar to abnormal batteries to some extent. Even though the false alarm rate for normal batteries is extremely low, the identification of abnormal batteries is 100%, ensuring that abnormal batteries will not enter the battery pack assembly process, demonstrating significant engineering potential and application prospects.

[0050] The advantage of this invention is that it can detect abnormalities in battery life early in a low-time cost, with high efficiency and accuracy, solving the problems of low accuracy and high false alarm rate, and is also very beneficial to the improvement of battery pack life, cost and capacity.

[0051] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for early life anomaly detection of lithium-ion batteries based on machine learning, characterized in that, The method comprises the following steps: S1: selecting a plurality of lithium ion batteries as sample batteries, performing accelerated aging test on the sample batteries, and defining a battery life label according to the test results; S2: applying a data processing method to a data set of the data of the accelerated aging test of the sample batteries, the data processing method comprising three steps, namely data segmentation, normal data reduction and abnormal data expansion; S3: selecting the constant voltage current data and the constant current voltage data in the first cycle test from the data set after data processing as the feature vector of the battery, and synthesizing the feature vectors of any two batteries into a feature matrix; S4: based on the battery life label in S1, it can be known whether any two batteries belong to the same type of battery by comparing the labels, and the judgment result is used as the output of the subsequent classifier model training; S5: training a neural network model offline with the feature matrix and true / false value as input, binarizing the training result to obtain a classifier for judging whether the battery is of the same type; S6: gradually selecting a battery from a test set of known categories and generating a feature matrix with an unknown battery, inputting the feature matrix into the classifier, and completing the early life abnormality detection of the lithium ion battery based on machine learning; In S2, the data segmentation is specifically: when dividing the data, the normal batteries are divided into a training set and a test set by half, and the abnormal batteries are divided by the leave-one-out method; The normal data reduction is specifically: for each abnormal battery in the training set, find the N1 closest normal batteries to its aging trajectory by checking the aging trajectory, and classify the normal batteries and abnormal batteries by selecting the normal samples as the boundary; The abnormal data expansion is specifically: the data collected in the first N2 cycles are used to enrich the training set; In the S4, the judgment result is expressed by binaryzation, expressed as: when the two battery labels are the same, it is true, and is ; when the two battery labels are different, it is false, and is . 2.The method of claim 1, wherein, The training part is specifically: a batch of new sample batteries are tested under the same accelerated aging condition, the battery life abnormality is observed through the accelerated aging test data and the capacity aging curve, the feature vector of the battery is obtained through the first cycle data of the accelerated aging test, then the feature vectors of any two batteries are synthesized into a feature matrix, the life label is binarized, and finally the feature matrix and the binarized matrix are used as input to train a neural network model to obtain a classifier for judging whether the battery is of the same type; The test part is specifically: a battery is gradually selected from a test set of known categories and a feature matrix is generated with an unknown battery, the feature matrix is input into the trained classifier, and the judgment result of the classifier is obtained. 3.The method of claim 2, wherein, In the training part, a batch of new sample batteries are tested under the same accelerated aging condition to achieve the abnormality of the battery itself, thereby shortening the experimental test time. 4.The method of claim 1, wherein, In S3, the data in the data set is selected as: the voltage data in the constant current stage and the current data in the constant voltage stage of the first cycle of the battery; The feature matrix is represented as: , where is the feature matrix of the kth section battery, where is the voltage sequence of the constant current stage, is the duration of the constant current stage; is the current sequence of the constant voltage stage, is the duration of the entire charging process; and is the serial number of the battery. 5.The method of claim 1, wherein, In S5, the neural network model is trained H times to obtain H classifiers, thereby improving the stability of the model and reducing the accumulated prediction error. 6.The method of claim 5, wherein, In the S6, the feature matrix is input into the trained H classifiers, and the category with a proportion greater than 50% is the judgment result of the H classifiers, and the category with a proportion greater than 50% is divided into the corresponding category.

Images