Battery Defect Monitoring, Identification, and Life Prediction Methods

Abstract

This invention provides a method for monitoring, identifying, and predicting the lifespan of batteries, relating to the field of battery parameter analysis technology. The invention constructs a standard judgment model based on a state dataset obtained from a laboratory. Based on the performance of the standard judgment model in scrapped samples, it classifies and adjusts the decision tree according to the error rates of different label types, forming an optimized judgment model. The state dataset of the target battery is obtained and input into the optimized judgment model to identify defect types, which are then input into the standard judgment model. Based on the different performances of the two models, a lifespan adjustment coefficient is obtained to adjust the lifespan prediction. This invention constructs a judgment model by differentiating the electrical parameters of plate corrosion and sulfation defects. The model is optimized based on the electrical parameters under actual usage conditions, increasing the robustness and accuracy of the judgment model. Simultaneously, the model's performance provides guidance for lifespan prediction.

Description

Technical Field

[0001] This invention relates to the field of battery parameter analysis technology, specifically to a method for monitoring and identifying battery defects and predicting battery life. Background Technology

[0002] With the promotion and popularization of new energy vehicles, the use of batteries is becoming increasingly widespread. Significant progress has also been made in battery defect identification and lifespan prediction. Manufacturers use Battery Management Systems (BMS) to monitor various battery parameters and determine battery defects based on parameter performance. Furthermore, changes in these parameters can predict battery lifespan, providing guidance for engineers to understand battery status. However, in certain specific areas, current detection methods are not accurate. For example, regarding mild plate corrosion and sulfation defects in batteries, these defects occur internally and cannot be visually detected externally. Moreover, these internal chemical defects exhibit similar electrical parameters, and mild defects can still allow the battery to remain usable. This situation makes defect identification difficult, thus affecting the BMS's assessment of the battery's remaining lifespan.

[0003] In the prior art, document CN115600143A discloses a method and apparatus for fault diagnosis of energy storage lithium batteries. The method includes: S1, obtaining the original parameters of the energy storage lithium battery under test; S2, constructing a feature sequence of the energy storage lithium battery under test; S3, inputting the feature sequence as the current sample into multiple isolated forest models for isolation, and repeatedly searching for the current sample in each isolated forest model using a binary tree algorithm to obtain the binary tree feature corresponding to the current sample in each isolated forest model; S4, calculating the anomaly probability of the current sample under each isolated forest model based on the binary tree feature; S5, repeating steps S3-S4 until a preset number of iterations is reached to obtain multiple anomaly probabilities of the current sample; S6, calculating the average of the multiple anomaly probabilities as the final anomaly probability of the current sample belonging to each type of anomaly, and selecting the anomaly type corresponding to the highest final anomaly probability as the actual anomaly type of the current sample.

[0004] While the publicly available technical documents demonstrate the ability to identify anomaly types through isolated forests, they fail to consider the different characteristics of specific defects under ideal and real-world conditions, do not optimize the judgment model based on the internal characteristics of changes in battery electrical parameters, and cannot provide practical guidance for lifespan prediction based on the identification results.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for monitoring and identifying battery defects and predicting battery life, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for battery defect monitoring, identification, and life prediction, comprising the following steps:

[0009] Step 1: Conduct charge-discharge experiments in the laboratory on batteries with plate corrosion defects, sulfation defects, and normal conditions respectively. Obtain the state dataset of the batteries at equal time intervals through BMS. Set the sliding window length and obtain samples in the time sequence of the state dataset. Filter out samples with single states, label the samples and input them into the random forest model for training to obtain the standard judgment model.

[0010] Step 2: Obtain batteries with plate corrosion, sulfation defects and normal condition through recycling points, obtain state datasets through BMS and form scrap samples, label the scrap samples and input them into the standard judgment model, obtain the accuracy of each decision tree for different label types, judge the decision trees with abnormal accuracy based on the accuracy threshold and filter out the wrong decision trees and abnormal decision trees.

[0011] Step 3: Summarize all samples to form a sample set. Retrain all incorrect and abnormal decision trees using the sample set. Replace the retrained decision trees in the standard judgment model. During the retraining process, set priority adjustment conditions for plate corrosion and sulfidation defects. After replacement, form an optimized judgment model.

[0012] Step 4: Obtain the actual state dataset through the BMS of the target battery, input the actual state dataset into the optimization judgment model to obtain the defect type of this battery, and at the same time obtain the number and number of decision trees with the judgment result of normal in the optimization judgment model. Input the actual state dataset into the standard judgment model to obtain the number of decision trees with the judgment result of normal in the standard judgment model. Construct a life adjustment coefficient based on the different performance of the two models, and adjust the life judgment of the BMS with the life adjustment coefficient.

[0013] Furthermore, the batteries were tested in the laboratory. Three types of batteries were used: a normal battery, a battery with plate corrosion defects, and a battery with sulfation defects. The severity of plate corrosion and sulfation was determined by experts to meet the criteria for "defect." Continuous charge-discharge experiments were conducted on the batteries, and a state dataset for each battery was acquired using a BMS. This state dataset included internal resistance, open-circuit voltage, discharge voltage, and charging voltage. The state datasets were acquired at a fixed frequency and numbered sequentially according to time. The formula used to generate the state datasets is as follows:

[0014] ;

[0015] in, Indicates the first A state dataset, For the retrieval variables of the state dataset, , The total number of state datasets, For internal resistance, Open circuit voltage, This is the discharge voltage. This is the charging voltage. The rate of change of internal resistance, The rate of change of open-circuit voltage. The rate of change of discharge voltage, This represents the rate of change of the charging voltage.

[0016] Furthermore, a sliding window length is set, and the sliding window length is randomly slid across the sequence of the state dataset. Subsequences within the sliding window are selected as samples, and the samples are filtered. The logic is as follows:

[0017] The samples in a single state are retained. In a single state, only one of the three data types, namely open circuit voltage, discharge voltage and charging voltage, has a reading. The state data set with only internal resistance and one of the voltage data has a reading, and the other two voltage data are 0, is a sample of a single state. Samples in non-single states are discarded.

[0018] Each sample is labeled with the type of battery that generated it: plate corrosion, sulfation defect, and normal. The labels are then one-hot encoded to annotate the samples. All data within each sample is normalized according to type. All samples are then divided into training and validation sets in an 8:2 ratio. The training set is input into a random forest model for training. The random forest model has 100 decision trees, a maximum depth of 10, and a minimum number of sample splits of 5. 70% of the samples are randomly selected from the training set, and features are randomly chosen as nodes in a single decision tree. The features for each node are determined based on maximizing information gain. Each decision tree is trained separately. The validation set is input into the trained random forest model. When the accuracy exceeds 90%, the random forest model is considered successfully built and is then labeled as the standard judgment model.

[0019] Furthermore, batteries scrapped due to plate corrosion defects and sulfation defects, as well as normally scrapped batteries, are obtained from battery recycling points. These batteries are those identified through disassembly. Their respective status datasets are obtained through their Battery Management System (BMS). The status datasets of the scrapped batteries are adjusted to obtain scrapped samples. The logic is as follows:

[0020] Data on open-circuit voltage, discharge voltage, and charging voltage were extracted and plotted as voltage curves. The horizontal axis of the voltage curve represents time, and the vertical axes represent the values ​​of open-circuit voltage, discharge voltage, and charging voltage, respectively. Resistance data were extracted and plotted as resistance curves. The horizontal axis of the resistance curve is the same as that of the voltage curve. State datasets of scrapped batteries were obtained using the frequency of state dataset acquisition in the laboratory environment. Similarly, state datasets of a single state with a sliding window length were obtained through filtering, and the datasets of this single state were marked as scrapped samples.

[0021] Furthermore, the scrapped samples are normalized, and each sample is labeled with its battery type, which is the same as that of the samples obtained in the laboratory. The scrapped samples are then input into a standard judgment model, and the accuracy rate of each decision tree for different label types is obtained. That is, the accuracy rate of each decision tree in judging scrapped samples labeled as plate corrosion, sulfation defects, and normal is set. An accuracy threshold is set, and each decision tree is filtered based on the relationship between its accuracy rate and the accuracy threshold, as follows:

[0022] Decision trees and their numbers for scrapped samples with labels of electrode corrosion, sulfidation defect, and normal that have an accuracy rate lower than the accuracy rate threshold are obtained separately. These are then summarized into sets of abnormal decision trees for electrode corrosion, sulfidation defect, and normal types, respectively. The elements in each set are the numbers of the decision trees. Decision trees that appear in two or more sets are labeled as incorrect decision trees, while decision trees that appear in only one set are labeled as abnormal decision trees of this type, and are respectively labeled as corrosion abnormal decision trees, sulfidation abnormal decision trees, and normal abnormal decision trees.

[0023] Furthermore, priority adjustment conditions are set respectively. The priority conditions for adjusting the plate corrosion are internal resistance and internal resistance change rate, and the priority conditions for adjusting the sulfidation defect are open circuit voltage and open circuit voltage change rate.

[0024] The scrapped samples are combined with the samples used to train the standard judgment model to form a sample set. The incorrect decision trees and normal judgment abnormal decision trees are retrained in the sample set. When the accuracy of all incorrect decision trees and normal judgment abnormal decision trees exceeds 90%, the retrained decision trees are replaced with the incorrect decision trees and normal judgment abnormal decision trees in the original standard judgment model.

[0025] The corrosion anomaly decision tree and the sulfidation anomaly decision tree are adjusted separately, with the following logic:

[0026] The entire sample set is input into the decision tree that needs adjustment for identification. Each element in the sample set receives information gain at each node. The average information gain of all elements in the sample set at that node is obtained, as is the average information gain of the validation set at that node when training the standard judgment model. The information gain difference is then calculated using the following formula: ;

[0027] in, Indicates the first The first decision tree that needs adjustment Information gain difference between nodes Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the validation set. Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the total sample set. Retrieve variables for decision trees that need adjustment for this type. , To retrieve variables for nodes in a decision tree, , This represents the total number of nodes in the decision tree.

[0028] A threshold for information gain difference is set, and the adjustment method for nodes whose information gain difference is less than or equal to the threshold is determined as the modified segmentation point. The modified segmentation point method is to use the total sample set for training and take the value of maximizing information gain as the segmentation point.

[0029] The adjustment method for nodes with information gain difference greater than the information gain difference threshold is determined as the replacement feature, and the replacement feature is preferentially replaced as the priority replacement condition for this decision tree type. Then, the split point of the decision tree is determined. When the accuracy of this decision tree in this type exceeds 90%, it is determined that the adjustment of this decision tree is complete and it is replaced in the standard judgment model. Otherwise, this decision tree is retrained in the total sample set until the accuracy exceeds 90%.

[0030] The adjusted standard judgment model is calibrated as the optimized judgment model.

[0031] Furthermore, the actual state dataset is obtained by monitoring the BMS of the target battery, and the acquisition frequency of the actual state dataset is the same as that of the state dataset in the laboratory. The actual state dataset is then filtered to obtain the actual state dataset under a single state within the sliding window length. The actual state datasets are then input into the optimization judgment model, and the defect type of the battery is determined based on the output of the optimization judgment model.

[0032] Furthermore, the number and index of decision trees whose judgment results are normal in the optimized judgment model are obtained. The actual state dataset is input into the standard judgment model to obtain the index of the decision trees whose judgment results are normal in the standard judgment model, forming the lifetime adjustment coefficient. The logic is as follows:

[0033] When the optimization judgment model result is normal, the life adjustment coefficient is 1;

[0034] When the optimized judgment model results in an abnormal outcome, decision trees that are judged as normal in both the standard judgment model and the optimized judgment model are removed. The number of decision trees with normal judgment results remaining in the optimized judgment model at this point is then used to obtain the lifetime adjustment coefficient, based on the following formula: ;

[0035] in, This is the lifespan adjustment factor. To optimize the number of remaining decision trees in the judgment model that are considered normal, .

[0036] Furthermore, the remaining lifespan is obtained through the battery's BMS. The method for obtaining the remaining lifespan is that the BMS adjusts the remaining lifespan using a lifespan adjustment coefficient based on a preset remaining lifespan determination method from the manufacturer according to product information. The adjusted remaining lifespan is then taken as the remaining lifespan of the battery. The formula used is as follows: ;

[0037] in, The adjusted remaining lifespan, The remaining lifespan before adjustment. This is the lifespan adjustment factor.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention constructs a standard judgment model using a random decision forest model based on a state dataset obtained from a laboratory. A new state dataset is obtained from recycled batteries to form scrap samples. Based on the performance of the standard judgment model on these scrap samples, decision trees are classified and filtered according to the error rate for different label types. The selected decision trees are adjusted using all samples, with priority adjustment conditions set for corrosion anomaly and sulfation anomaly decision trees to guide their adjustment. The adjusted decision trees are then replaced in the standard judgment model to form an optimized judgment model. The state dataset of the target battery is obtained and input into the optimized judgment model to identify defect types. The target battery's state dataset is then input into the standard judgment model. Lifetime adjustment coefficients are obtained based on the different performances of the two models, and these coefficients are used to adjust the lifetime judgment of the BMS. This invention constructs a judgment model based on the differences in electrical parameters between plate corrosion and sulfation defects, and then optimizes the model based on the electrical parameters under actual usage conditions, increasing the robustness and accuracy of the judgment model. Simultaneously, the severity of defects shown by the model provides guidance for lifetime judgment. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0043] Example:

[0044] Please see Figure 1 The present invention provides a technical solution:

[0045] A method for battery defect monitoring, identification, and life prediction, comprising the following steps:

[0046] Step 1: Conduct charge-discharge experiments in the laboratory on batteries with plate corrosion defects, sulfation defects, and normal conditions respectively. Obtain the state dataset of the batteries at equal time intervals through BMS. Set the sliding window length and obtain samples in the time sequence of the state dataset. Filter out samples with single states, label the samples and input them into the random forest model for training to obtain the standard judgment model.

[0047] Step 1 includes the following:

[0048] Step 101: Conduct experiments on the batteries in the laboratory. Three types of batteries are used: a normal battery, a battery with plate corrosion defects, and a battery with sulfation defects. The severity of plate corrosion and sulfation is determined by experts to meet the criteria for "defect." Continuous charge-discharge experiments are conducted on the batteries, and a state dataset for each battery is acquired using a BMS. This state dataset includes internal resistance, open-circuit voltage, discharge voltage, and charging voltage. The state datasets are acquired at a fixed frequency and numbered sequentially according to time. The formula used for the state datasets is as follows:

[0049] ;

[0050] in, Indicates the first A state dataset, For the retrieval variables of the state dataset, , The total number of state datasets, For internal resistance, Open circuit voltage, This is the discharge voltage. This is the charging voltage. The rate of change of internal resistance, The rate of change of open-circuit voltage. The rate of change of discharge voltage, This represents the rate of change of the charging voltage.

[0051] In the laboratory, charge-discharge experiments were conducted on batteries in three states (normal batteries, batteries with plate corrosion defects, and batteries with sulfation defects). State datasets were collected at a fixed frequency using a BMS system, establishing a feature database for different health states. Conducting experiments under controlled conditions ensured the accuracy, repeatability, and consistency of the data, laying a solid foundation for subsequent sample selection and model training. The state dataset includes core parameters such as internal resistance, open-circuit voltage, discharge voltage, and charging voltage. These parameters comprehensively reflect the battery's operating state and potential defect characteristics, serving as key evidence for judging its health status. Collecting data at a fixed frequency and numbering it chronologically not only facilitates subsequent time-series analysis but also provides structured data support for the application of sliding window techniques. The selection of three types of batteries in the experiment covered normal and typical defect states, realistically simulating various situations that might be encountered in practical applications, enhancing the representativeness of the data. Furthermore, this step, through efficient data collection using the BMS system, avoided human error, ensuring data quality met the requirements for model training. This step directly provides raw data for sample extraction and selection in step 102 and is a crucial data acquisition link in the entire technical solution.

[0052] Step 102: Set the sliding window length, randomly slide the sliding window across the sequence of the state dataset, select the subsequence within the sliding window as samples, and filter the samples as follows:

[0053] The samples in a single state are retained. In a single state, only one of the three data types, namely open circuit voltage, discharge voltage and charging voltage, has a reading. The state data set with only internal resistance and one of the voltage data has a reading, and the other two voltage data are 0, is a sample of a single state. Samples in non-single states are discarded.

[0054] Each sample is labeled with the type of battery that generated it: plate corrosion, sulfation defect, and normal. The labels are then one-hot encoded to annotate the samples. All data within each sample is normalized according to type. All samples are then divided into training and validation sets in an 8:2 ratio. The training set is input into a random forest model for training. The random forest model has 100 decision trees, a maximum depth of 10, and a minimum number of sample splits of 5. 70% of the samples are randomly selected from the training set, and features are randomly chosen as nodes in a single decision tree. The features for each node are determined based on maximizing information gain. Each decision tree is trained separately. The validation set is input into the trained random forest model. When the accuracy exceeds 90%, the random forest model is considered successfully built and is then labeled as the standard judgment model.

[0055] By setting a sliding window length, samples of a single state (i.e., only one of the open-circuit voltage, discharge voltage, and charging voltage readings is available) are randomly slid across the state dataset sequence to extract and filter samples. These samples are then labeled and input into a random forest model for training, constructing a standard judgment model. The sliding window technique captures the temporal characteristics of the battery state, ensuring the temporal continuity of the samples and thus improving the model's ability to recognize dynamic changes in the battery. Filtering samples to a single state effectively eliminates data interference from multiple states, improving sample purity and classification accuracy, enabling the model to more accurately learn feature patterns under different defect types. A random forest model is used, with parameters such as 100 decision trees, a maximum depth of 10, and a minimum number of sample splits of 5. Ensemble learning enhances the model's stability and generalization ability. After model training, a validation set is used for evaluation. When the accuracy exceeds 90%, the model is considered complete, ensuring its reliability and effectiveness. This step not only follows up on the data collection work in step 101 but also provides a standard judgment model for model evaluation and optimization in subsequent step 2, making it a crucial link in the model construction of the technical solution.

[0056] Step 2: Obtain batteries with plate corrosion, sulfation defects and normal condition through recycling points, obtain state datasets through BMS and form scrap samples, label the scrap samples and input them into the standard judgment model, obtain the accuracy of each decision tree for different label types, judge the decision trees with abnormal accuracy based on the accuracy threshold and filter out the wrong decision trees and abnormal decision trees.

[0057] Step 2 includes the following:

[0058] Step 201: Obtain batteries scrapped due to plate corrosion defects and sulfation defects, as well as batteries that are normally scrapped, from the battery recycling point. These batteries are those identified through disassembly. Obtain their respective status datasets through their BMS (Battery Management System). Adjust the status datasets of the scrapped batteries to obtain scrapped samples. The logic is as follows:

[0059] Data on open-circuit voltage, discharge voltage, and charging voltage were extracted and plotted as voltage curves. The horizontal axis of the voltage curve represents time, and the vertical axes represent the values ​​of open-circuit voltage, discharge voltage, and charging voltage, respectively. Resistance data were extracted and plotted as resistance curves. The horizontal axis of the resistance curve is the same as that of the voltage curve. State datasets of scrapped batteries were obtained using the frequency of state dataset acquisition in the laboratory environment. Similarly, state datasets of a single state with a sliding window length were obtained through filtering, and the datasets of this single state were marked as scrapped samples.

[0060] End-of-life batteries (including those with plate corrosion defects, sulfation defects, and those normally scrapped) are collected from recycling points. Their state datasets are collected using a BMS system, and this data is processed to conform to the laboratory environment, forming end-of-life samples. Introducing real-world end-of-life data enhances the model's generalization ability and practical application value, making the technical solution closer to actual operating conditions. By plotting voltage and resistance curves and employing the same data acquisition frequency and screening methods as the laboratory, consistency in data structure and quality between the end-of-life samples and laboratory samples is ensured, facilitating subsequent model evaluation and comparison. Furthermore, the introduction of end-of-life samples can reveal potential limitations in laboratory data, helping to identify shortcomings in the model based on real-world data. This step provides the necessary end-of-life samples for the accuracy analysis and model diagnosis in step 202, and is a crucial part of data preparation and model evaluation.

[0061] Step 202: Normalize the scrapped samples and label each sample with the battery type, which is the same as the sample obtained in the laboratory. Input the scrapped samples into the standard judgment model and obtain the accuracy rate of each decision tree for different label types, that is, the accuracy rate of each decision tree in judging scrapped samples labeled as plate corrosion, sulfation defects, and normal. Set an accuracy threshold and filter each decision tree according to the relationship between the accuracy rate of each decision tree and the accuracy threshold, as follows:

[0062] Decision trees and their numbers for scrapped samples with labels of electrode corrosion, sulfidation defect, and normal that have an accuracy rate lower than the accuracy rate threshold are obtained separately. These are then summarized into sets of abnormal decision trees for electrode corrosion, sulfidation defect, and normal types, respectively. The elements in each set are the numbers of the decision trees. Decision trees that appear in two or more sets are labeled as incorrect decision trees, while decision trees that appear in only one set are labeled as abnormal decision trees of this type, and are respectively labeled as corrosion abnormal decision trees, sulfidation abnormal decision trees, and normal abnormal decision trees.

[0063] The scrapped samples are input into the standard judgment model, and the accuracy of each decision tree for different defect types (plate corrosion, sulfidation defects, and normal) is analyzed. By setting an accuracy threshold, incorrect and abnormal decision trees are identified. This step provides an in-depth evaluation of the standard judgment model's performance on real scrapped data, revealing the differences in the model's recognition capabilities across different defect types, and providing a concrete basis for subsequent optimization. By distinguishing between incorrect decision trees (low accuracy across multiple defect types) and abnormal decision trees (low accuracy on specific defect types), strategies and directions are provided for subsequent targeted adjustments. For example, incorrect decision trees may require overall retraining, while abnormal decision trees may require feature adjustments or split point optimization for specific defect types. This step not only connects to the scrapped sample preparation in step 201 but also provides crucial diagnostic results for model optimization in step 3, serving as a bridge between model evaluation and optimization in the technical solution.

[0064] Step 3: Summarize all samples to form a sample set. Retrain all incorrect and abnormal decision trees using the sample set. Replace the retrained decision trees in the standard judgment model. During the retraining process, set priority adjustment conditions for plate corrosion and sulfidation defects. After replacement, form an optimized judgment model.

[0065] Step 3 includes the following:

[0066] Priority adjustment conditions are set respectively. The priority conditions for adjusting the plate corrosion are internal resistance and internal resistance change rate, and the priority conditions for adjusting the sulfidation defect are open circuit voltage and open circuit voltage change rate.

[0067] The scrapped samples are combined with the samples used to train the standard judgment model to form a sample set. The incorrect decision trees and normal judgment abnormal decision trees are retrained in the sample set. When the accuracy of all incorrect decision trees and normal judgment abnormal decision trees exceeds 90%, the retrained decision trees are replaced with the incorrect decision trees and normal judgment abnormal decision trees in the original standard judgment model.

[0068] The corrosion anomaly decision tree and the sulfidation anomaly decision tree are adjusted separately, with the following logic:

[0069] The entire sample set is input into the decision tree that needs adjustment for identification. Each element in the sample set receives information gain at each node. The average information gain of all elements in the sample set at that node is obtained, as is the average information gain of the validation set at that node when training the standard judgment model. The information gain difference is then calculated using the following formula: ;

[0070] in, Indicates the first The first decision tree that needs adjustment Information gain difference between nodes Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the validation set. Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the total sample set. Retrieve variables for decision trees that need adjustment for this type. , To retrieve variables for nodes in a decision tree, , This represents the total number of nodes in the decision tree.

[0071] Indicates the first The first of the decision trees that needs adjustment The information gain difference of each node specifically reflects the difference in its splitting performance on two different datasets, i.e., the inconsistency between the validation set used to train the standard judgment model and the total sample set containing discarded battery samples. Its meaning lies in quantifying the change in the classification ability of decision tree nodes on laboratory data and real-world discarded battery data, helping to identify potential problems with the model in practical applications. In a real-world environment, This reflects whether the node can effectively adapt to actual battery defect characteristics, such as plate corrosion or sulfation defects, because laboratory data is usually based on controlled conditions, while scrap data reflects more complex actual aging patterns. If A large difference indicates that the node's performance on the validation set differs significantly from that on the total sample set, which may suggest that the model is overfitting the laboratory data or has failed to capture the true defect characteristics of the scrapped batteries. It is based on the average information gain of the validation set, reflecting the splitting effect of the node during training, and is usually related to the idealized characteristics of the battery state under laboratory conditions; This is based on the average information gain of the total sample set, reflecting the splitting effect of nodes when real scrap data is included, and may involve more noise or atypical defect patterns. In practical problems, these two independent variables correspond to the model training effect in a laboratory environment and the generalization ability in a real environment, respectively. The value reveals the fitness gap between the two.

[0072] and and The correlation stems from the calculation logic of the information gain difference, which is directly defined by the difference between the two, reflecting the performance comparison of nodes under different data backgrounds. and The ability of a node to reduce classification uncertainty after splitting on the validation set and the total sample set was measured, respectively. By comparing these two, the effectiveness of node splitting features was revealed. When it increases, A tendency for the node to increase indicates that the node performs better at splitting on the validation set, and vice versa. When it increases, A tendency for the node to decrease indicates that it is more adapted to the overall sample set and may be closer to the actual defect characteristics. This relationship helps determine whether a node needs adjustment in a real-world environment: a positive value may indicate overfitting risk, while a negative value may indicate increased adaptability of the node to the actual data. The direct impact lies in… It is a linear combination of two independent variables, whose values ​​depend entirely on their relative magnitudes, reflecting the transition effect of the model from the laboratory to practical application.

[0073] A threshold for information gain difference is set, and the adjustment method for nodes whose information gain difference is less than or equal to the threshold is determined as the modified segmentation point. The modified segmentation point method is to use the total sample set for training and take the value of maximizing information gain as the segmentation point.

[0074] The adjustment method for nodes with information gain difference greater than the information gain difference threshold is determined as the replacement feature, and the replacement feature is preferentially replaced as the priority replacement condition for this decision tree type. Then, the split point of the decision tree is determined. When the accuracy of this decision tree in this type exceeds 90%, it is determined that the adjustment of this decision tree is complete and it is replaced in the standard judgment model. Otherwise, this decision tree is retrained in the total sample set until the accuracy exceeds 90%.

[0075] The adjusted standard judgment model is calibrated as the optimized judgment model.

[0076] By aggregating scrapped samples with samples from the standard judgment model training to form a sample set, the incorrect decision trees and normal-to-abnormal decision trees are retrained. Targeted adjustments are made to the corrosion and sulfidation anomaly decision trees, ultimately forming an optimized judgment model. Integrating laboratory data and real scrapped data improves the model's adaptability and recognition accuracy, making it more suitable for practical applications. Retraining the incorrect and normal-to-abnormal decision trees using the sample set enhances their classification capabilities. For the corrosion and sulfidation anomaly decision trees, adjustments such as modifying split points or replacing features are made by analyzing information gain differences, specifically improving the model's recognition accuracy for specific defect types. Furthermore, priority adjustment conditions are set for electrode corrosion and sulfidation defects (e.g., prioritizing internal resistance and internal resistance change rate for electrode corrosion, and prioritizing open-circuit voltage and open-circuit voltage change rate for sulfidation defects), ensuring the targeted effectiveness of the adjustments. This step not only builds upon the model evaluation and diagnostic results of step 2 but also provides an optimized judgment model for practical application in step 4, making it the core of the model optimization in this scheme.

[0077] Step 4: Obtain the actual state dataset through the BMS of the target battery, input the actual state dataset into the optimization judgment model to obtain the defect type of this battery, and at the same time obtain the number and number of decision trees with the judgment result of normal in the optimization judgment model. Input the actual state dataset into the standard judgment model to obtain the number of decision trees with the judgment result of normal in the standard judgment model. Construct a life adjustment coefficient based on the different performance of the two models, and adjust the life judgment of the BMS with the life adjustment coefficient.

[0078] Step 4 includes the following:

[0079] Step 401: Obtain the actual state dataset by monitoring the BMS of the target battery, and the acquisition frequency of the actual state dataset is the same as the acquisition frequency of the state dataset in the laboratory. Filter the actual state dataset to obtain the actual state dataset under a single state within the sliding window length. Input the actual state dataset into the optimization judgment model respectively, and determine the defect type of the battery based on the output of the optimization judgment model.

[0080] By monitoring the BMS of the target battery to obtain actual condition datasets, processing them, and inputting them into an optimized judgment model, the defect types of the battery are identified. The optimized judgment model is directly used for defect identification in actual batteries, realizing the practical application of the technical solution. The acquisition frequency and processing method of the actual condition dataset are consistent with those in the laboratory, ensuring data consistency and model applicability. Through the output of the optimized judgment model, it is possible to accurately determine whether the battery has plate corrosion or sulfation defects, providing basic information on defect types for subsequent life adjustment. This step provides the necessary defect judgment results for the construction of the life adjustment coefficient in step 402 and is a key step in the application of the technology.

[0081] Step 402: Obtain the number and index of the decision trees whose judgment result is normal in the optimized judgment model. Input the actual state dataset into the standard judgment model, obtain the index of the decision trees whose judgment result is normal in the standard judgment model, and form the lifetime adjustment coefficient. The logic is as follows:

[0082] When the optimization judgment model result is normal, the life adjustment coefficient is 1;

[0083] When the optimized judgment model results in an abnormal outcome, decision trees that are judged as normal in both the standard judgment model and the optimized judgment model are removed. The number of decision trees with normal judgment results remaining in the optimized judgment model at this point is then used to obtain the lifetime adjustment coefficient, based on the following formula: ;

[0084] in, This is the lifespan adjustment factor. To optimize the number of remaining decision trees in the judgment model that are considered normal, .

[0085] This represents the lifespan adjustment factor, which specifically reflects the degree to which the battery's health status adjusts the remaining lifespan prediction. Its meaning is that by comparing the differences in decision tree judgments between the optimized judgment model and the standard judgment model, the impact of battery defects on lifespan is quantified, thereby correcting the remaining lifespan estimated by the BMS (Battery Management System) to make the prediction results closer to reality. In a real-world environment... This reflects the degree of performance degradation of the battery due to defects (such as plate corrosion or sulfation defects). For example, when the optimization judgment model determines that the battery is abnormal, The lifespan prediction will be adjusted based on the model's evaluation results to reflect the battery's health status in actual use. Defined as the number of decision trees in the optimized judgment model that, after specific elimination, still determine the battery as normal. Specifically, after the optimized model determines the battery is abnormal, and after eliminating decision trees that were simultaneously judged as normal in both the standard and optimized judgment models, the remaining optimized model still classifies the battery as normal. This variable reflects the inconsistency between the optimized and standard models in assessing battery health. In practical applications, The optimized model (adjusted based on real scrap data) has some subtle defects or health characteristics that are not identified in the standard model (based on laboratory data). and The common challenge is how to transform the inconsistencies in model judgments into a basis for lifespan adjustments, helping users to more accurately understand the remaining battery life and indicating potential maintenance or replacement needs.

[0086] and The correlation stems from the exponential function design of the formula, that is... The value is directly determined by pass The calculations show that the difference in normal judgment between the two models affects the lifespan adjustment. The number of decision trees in the optimization model that still make normal judgments reflects the unique ability of the optimization model to judge the battery health status, which may stem from its higher sensitivity to actual defect and degradation characteristics. and A positive correlation is observed: with Increase, index part The absolute value of decreases (negative values ​​become smaller), thereby causing to decrease in absolute value (negative values ​​become smaller Increase. The significance of this positive correlation design is that when the number of decision trees judged as normal in the optimization model increases (i.e., ... (Larger), but less deteriorated battery, therefore the lifespan adjustment range is smaller. For dependent variable The effect exhibits a monotonically increasing trend. This relationship ensures the sensitivity of the lifetime adjustment coefficient to model discrepancies, enabling it to flexibly adapt to the uncertainty of defect impacts in real-world environments while maintaining the rationality of the adjustment.

[0087] In the standard judgment model, the decision tree distinguishes defects through typical defect features and also through the degree of abnormality of electrical parameters. The degree of abnormality of electrical parameters represents the degree of degradation of defects in the battery. The optimized judgment model adjusts the standard judgment model, increases its robustness, and makes it more adaptable to the electrical parameter performance in actual use. The standard judgment model is constructed from standard laboratory data, which contains more typical defect features. After adjustments to scrapped samples and priority adjustment conditions, the optimized judgment model reduces the number of decision trees in the standard model that identify defects based on typical features, and adds decision trees that judge defects from the perspective of battery degradation. Therefore, by comparing abnormal decision trees in the two models and proposing decision trees that simultaneously judge defects, and retaining the decision trees in the optimized judgment model that reflect defects from the degree of battery degradation, we can, to a certain extent, reflect the degree of battery degradation caused by defects, thereby guiding the determination of remaining life. Although the decision trees removed from the optimized judgment model also include decision trees that identify defects from the perspective of battery degradation, from an overall perspective, they are still a minority. What is retained are more instructive decision trees that can judge defects from the perspective of battery degradation. Therefore, this step is of progressive significance.

[0088] A lifespan adjustment coefficient is constructed by comparing the number of decision trees that indicate "normal" in the optimized judgment model and the standard judgment model. By analyzing the consistency of the decision trees and synthesizing the model's judgment results, a reasonable lifespan adjustment coefficient is constructed, improving the accuracy of lifespan prediction. Specifically, when the optimized judgment model determines the battery is normal, the lifespan adjustment coefficient is 1, indicating no adjustment is needed; when it determines the battery is abnormal, decision trees that are judged as normal in both the standard and optimized judgment models are removed, and the number of remaining normal decision trees is calculated to determine the lifespan adjustment coefficient. This method effectively reflects the model's judgment of the battery's health status and adjusts the lifespan prediction value accordingly, ensuring the rationality and accuracy of the adjustment. This step not only connects the defect identification results in step 401 but also provides a crucial adjustment coefficient for lifespan adjustment in step 403, making it an important link in lifespan prediction optimization.

[0089] Step 403: Obtain the remaining lifespan through the battery's BMS. The remaining lifespan is obtained by the BMS using a lifespan adjustment coefficient based on a preset remaining lifespan determination method by the manufacturer according to product information. The adjusted remaining lifespan is then taken as the remaining lifespan of the battery. The formula used is as follows: ;

[0090] in, The adjusted remaining lifespan, The remaining lifespan before adjustment. This is the lifespan adjustment factor.

[0091] This represents the adjusted remaining lifespan, specifically reflecting the actual expected usage time of the battery under its current health condition. It means that by introducing a lifespan adjustment factor, the original lifespan estimated by the BMS (Battery Management System) is corrected to more accurately reflect the impact of defects (such as plate corrosion or sulfation defects) on lifespan. In real-world environments, This helps users understand the remaining usable time of the battery under real-world usage conditions, predict when the battery will need maintenance or replacement, and avoid unexpected failures due to defects. In contrast, This is the remaining battery life before adjustment, calculated by the BMS according to the manufacturer's preset method. It is usually based on ideal conditions and does not consider the impact of actual defects, which may overestimate the battery life; another independent variable. This is the lifetime adjustment factor, calculated from the difference in decision tree judgments between the optimized judgment model and the standard judgment model, reflecting the actual impact of defects on lifetime. In real-world problems, Solved Inaccuracy when faced with actual aging or defects, and This quantifies the specific impact of changes in health status.

[0092] The remaining lifespan obtained from the BMS is adjusted by a lifespan adjustment factor to obtain the final prediction of the battery's remaining lifespan. The model's judgment results are then transformed into specific lifespan prediction values, realizing the final application of the technology. By applying the lifespan adjustment factor to the original lifespan judgment of the BMS, the accuracy of the lifespan prediction is optimized, ensuring the reliability and practicality of the prediction results. The adjustment method is simple, clear, and easy to implement, enabling the rapid determination of the adjusted remaining lifespan in practical applications. This step builds upon the achievements of the preceding steps (defect identification and the construction of the lifespan adjustment factor), forming a complete battery monitoring and lifespan prediction system, and is a key link in the application of the technical solution.

[0093] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0094] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0095] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring, identifying, and predicting the lifespan of a storage battery, characterized in that, The specific steps include: Step 1: Conduct charge-discharge experiments in the laboratory on batteries with plate corrosion defects, sulfation defects, and normal conditions respectively. Obtain the state dataset of the batteries at equal time intervals through BMS. Set the sliding window length and obtain samples in the time sequence of the state dataset. Filter out samples with single states, label the samples and input them into the random forest model for training to obtain the standard judgment model. Step 2: Obtain batteries with plate corrosion, sulfation defects and normal condition through recycling points, obtain state datasets through BMS and form scrap samples, label the scrap samples and input them into the standard judgment model, obtain the accuracy of each decision tree for different label types, judge the decision trees with abnormal accuracy based on the accuracy threshold and filter out the wrong decision trees and abnormal decision trees. Step 3: Summarize all samples to form a sample set. Retrain all incorrect and abnormal decision trees using the sample set. Replace the retrained decision trees in the standard judgment model. During the retraining process, set priority adjustment conditions for plate corrosion and sulfidation defects. After replacement, form an optimized judgment model. Step 4: Obtain the actual state dataset through the BMS of the target battery, input the actual state dataset into the optimization judgment model to obtain the defect type of this battery, and at the same time obtain the number and number of decision trees with the judgment result of normal in the optimization judgment model. Input the actual state dataset into the standard judgment model to obtain the number of decision trees with the judgment result of normal in the standard judgment model. Construct a life adjustment coefficient based on the different performance of the two models, and adjust the life judgment of the BMS with the life adjustment coefficient. Set a sliding window length, randomly slide the sliding window across the sequence of the state dataset, select a subsequence within the sliding window as a sample, and filter the samples. The logic is as follows: The samples in a single state are retained. In a single state, only one of the three data types, namely open circuit voltage, discharge voltage and charging voltage, has a reading. The state data set with only internal resistance and one of the voltage data has a reading, and the other two voltage data are 0, is a sample of a single state. Samples in non-single states are discarded. Each sample is labeled with the type of battery that generated it: plate corrosion, sulfation defect, and normal. The labels are then one-hot encoded to annotate the samples. All data within each sample is normalized according to type. All samples are then divided into training and validation sets in an 8:2 ratio. The training set is input into a random forest model for training. The random forest model has 100 decision trees, a maximum depth of 10, and a minimum number of sample splits of 5. 70% of the samples are randomly selected from the training set, and features are randomly chosen as nodes in a single decision tree. The features of each node are determined based on maximizing information gain. Each decision tree is trained separately. The validation set is input into the trained random forest model. When the accuracy exceeds 90%, the random forest model is considered to be successfully built and is then labeled as the standard judgment model. Priority adjustment conditions are set respectively. The priority conditions for adjusting the plate corrosion are internal resistance and internal resistance change rate, and the priority conditions for adjusting the sulfidation defect are open circuit voltage and open circuit voltage change rate. The scrapped samples are combined with the samples used to train the standard judgment model to form a sample set. The incorrect decision trees and normal judgment abnormal decision trees are retrained in the sample set. When the accuracy of all incorrect decision trees and normal judgment abnormal decision trees exceeds 90%, the retrained decision trees are replaced with the incorrect decision trees and normal judgment abnormal decision trees in the original standard judgment model. The corrosion anomaly decision tree and the sulfidation anomaly decision tree are adjusted separately, with the following logic: The entire sample set is input into the decision tree that needs adjustment for identification. Each element in the sample set receives information gain at each node. The average information gain of all elements in the sample set at that node is obtained, as is the average information gain of the validation set at that node when training the standard judgment model. The information gain difference is then calculated using the following formula: ； in, Indicates the first The first decision tree that needs adjustment Information gain difference between nodes Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the validation set. Indicates the first The first decision tree that needs adjustment The average information gain obtained by each node based on the total sample set. Retrieve variables for decision trees that need adjustment for this type. , , To retrieve variables for nodes in a decision tree, , , This represents the total number of nodes in the decision tree. A threshold for information gain difference is set, and the adjustment method for nodes whose information gain difference is less than or equal to the threshold is determined as the modified segmentation point. The modified segmentation point method is to use the total sample set for training and take the value of maximizing information gain as the segmentation point. The adjustment method for nodes with information gain difference greater than the information gain difference threshold is determined as the replacement feature, and the replacement feature is preferentially replaced as the priority replacement condition for this decision tree type. Then, the split point of the decision tree is determined. When the accuracy of this decision tree in this type exceeds 90%, it is determined that the adjustment of this decision tree is complete and it is replaced in the standard judgment model. Otherwise, this decision tree is retrained in the total sample set until the accuracy exceeds 90%. The adjusted standard judgment model is calibrated as the optimized judgment model.

2. The battery defect monitoring, identification, and lifespan prediction method according to claim 1, characterized in that: In the laboratory, three types of batteries were tested: a normal battery, a battery with plate corrosion defects, and a battery with sulfation defects. The severity of plate corrosion and sulfation was determined by experts to meet the criteria for "defect." Continuous charge-discharge experiments were conducted on the batteries, and a state dataset for each battery was acquired using a Battery Management System (BMS). This state dataset included internal resistance, open-circuit voltage, discharge voltage, and charging voltage. The state datasets were acquired at a fixed frequency and numbered sequentially according to time. The formula used to generate the state datasets is as follows: ； in, Indicates the first A state dataset, For the retrieval variables of the state dataset, , , The total number of state datasets, For internal resistance, Open circuit voltage, This is the discharge voltage. This is the charging voltage. The rate of change of internal resistance, The rate of change of open-circuit voltage. The rate of change of discharge voltage, This represents the rate of change of the charging voltage.

3. The battery defect monitoring, identification, and lifespan prediction method according to claim 1, characterized in that: Batteries scrapped due to plate corrosion defects and sulfation defects, as well as normally scrapped batteries, are collected at battery recycling points. These batteries were identified through disassembly. Their respective status datasets are obtained through their Battery Management System (BMS). The status datasets of the scrapped batteries are adjusted to obtain scrapped samples. The logic is as follows: Data on open-circuit voltage, discharge voltage, and charging voltage were extracted and plotted as voltage curves. The horizontal axis of the voltage curve represents time, and the vertical axes represent the values ​​of open-circuit voltage, discharge voltage, and charging voltage, respectively. Resistance data were extracted and plotted as resistance curves. The horizontal axis of the resistance curve is the same as that of the voltage curve. State datasets of scrapped batteries were obtained using the frequency of state dataset acquisition in the laboratory environment. Similarly, state datasets of a single state with a sliding window length were obtained through filtering, and the datasets of this single state were marked as scrapped samples.

4. The battery defect monitoring, identification, and lifespan prediction method according to claim 3, characterized in that: The scrapped samples were normalized, and each sample was labeled with its battery type, which was the same as that of the samples obtained in the laboratory. The scrapped samples were then input into a standard judgment model, and the accuracy rate of each decision tree for different label types was obtained. Specifically, the accuracy rate of each decision tree in judging scrapped samples labeled as having plate corrosion, sulfation defects, and normal was calculated. An accuracy threshold was set, and each decision tree was filtered based on the relationship between its accuracy rate and the accuracy threshold. The logic is as follows: Decision trees and their numbers for scrapped samples with labels of electrode corrosion, sulfidation defect, and normal that have an accuracy rate lower than the accuracy rate threshold are obtained separately. These are then summarized into sets of abnormal decision trees for electrode corrosion, sulfidation defect, and normal types, respectively. The elements in each set are the numbers of the decision trees. Decision trees that appear in two or more sets are labeled as incorrect decision trees, while decision trees that appear in only one set are labeled as abnormal decision trees of this type, and are respectively labeled as corrosion abnormal decision trees, sulfidation abnormal decision trees, and normal abnormal decision trees.

5. The battery defect monitoring, identification, and lifespan prediction method according to claim 1, characterized in that: The actual state dataset is obtained by monitoring the BMS of the target battery, and the acquisition frequency of the actual state dataset is the same as that of the state dataset in the laboratory. The actual state dataset is filtered to obtain the actual state dataset under a single state within the sliding window length. The actual state datasets are then input into the optimization judgment model, and the defect type of the battery is determined based on the output of the optimization judgment model.

6. The battery defect monitoring, identification, and lifespan prediction method according to claim 5, characterized in that: Obtain the number and index of decision trees whose judgment result is normal in the optimized judgment model. Input the actual state dataset into the standard judgment model, obtain the index of the decision tree whose judgment result is normal in the standard judgment model, and form the lifetime adjustment coefficient. The logic is as follows: When the optimization judgment model result is normal, the life adjustment coefficient is 1; When the optimized judgment model results in an abnormal outcome, decision trees that are judged as normal in both the standard judgment model and the optimized judgment model are removed. The number of decision trees with normal judgment results remaining in the optimized judgment model at this point is then used to obtain the lifetime adjustment coefficient, based on the following formula: ； in, This is the lifespan adjustment factor. To optimize the number of remaining decision trees in the judgment model that are considered normal, .

7. The battery defect monitoring, identification, and lifespan prediction method according to claim 6, characterized in that: The remaining lifespan is obtained through the battery's BMS. The method for obtaining the remaining lifespan involves the BMS adjusting the remaining lifespan using a lifespan adjustment coefficient based on a preset remaining lifespan determination method from the manufacturer according to product information. The adjusted remaining lifespan is then taken as the remaining lifespan of the battery. The formula used is as follows: ； in, The adjusted remaining lifespan, The remaining lifespan before adjustment. This is the lifespan adjustment factor.

Images