Data set partitioning method, testing method and system for lead-acid battery performance testing

Abstract

This invention belongs to the field of lead-acid battery performance testing technology, specifically providing a method, testing method, and system for dividing a lead-acid battery performance testing dataset. The method includes: data acquisition, obtaining a preset number of feature combinations through a traditional random partitioning method; calculating the time-series dependency correlation value, data distribution characteristic value, and non-abnormal noise value of the feature combinations; then calculating the mean of these values ​​as evaluation indicators to obtain the optimal training and test sets; and finally training a neural network model for lead-acid battery SOH prediction and performance testing. This invention evaluates different feature combinations from multiple dimensions, making the resulting target test dataset and target training data more suitable for training the lead-acid battery SOH prediction model, thereby improving the prediction robustness and accuracy of the SOH prediction model.

Description

Technical Field

[0001] This invention belongs to the field of lead-acid battery performance testing technology, specifically, it relates to a method for dividing a lead-acid battery performance testing dataset, a testing method, and a system. Background Technology

[0002] Performance testing of lead-acid batteries is a crucial step in ensuring battery quality, extending battery life, and guaranteeing the normal operation of equipment. Current technologies typically establish State of Health (SOH) prediction models to effectively analyze and test lead-acid battery performance. This is because lead-acid battery performance is directly related to the SOH index. By analyzing SOH, the degree of battery performance degradation can be quantitatively assessed, remaining battery life predicted, and charging / discharging strategies optimized. In lead-acid battery management and testing, SOH analysis is not only a technical means but also a key tool for improving battery efficiency and economy. As the scenario suggests, the process of using SOH for battery performance testing involves a large amount of complex data, such as voltage, current, temperature, and internal resistance. Due to the powerful data processing capabilities of neural networks, current technologies generally train neural networks to capture the complex dynamic characteristics in the data, achieving accurate prediction of battery SOH.

[0003] The key to neural networks lies in model data training. Traditionally, neural networks typically use random partitioning to divide the dataset into training and test sets. While simple and intuitive, this approach leads to poor robustness and accuracy of the neural network prediction model. Specifically, it results in poor accuracy in predicting the State of Harm (SOH) values ​​of lead-acid batteries, consequently affecting the reliability and accuracy of subsequent performance assessments based on these predictions. For example, this strategy of randomly partitioning training and test datasets might result in the training set containing more data from early charge-discharge cycles, while the test dataset contains more data from later cycles. This leads to poor performance of the prediction model on the test dataset, impacting the generalization ability of the SOH prediction model. Therefore, how to partition the training and test datasets to improve the prediction accuracy of the SOH prediction model and ensure the accuracy and reliability of subsequent performance testing of lead-acid batteries has become a pressing issue. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, detection method and system for dividing a lead-acid battery performance testing dataset. By evaluating different feature combinations from multiple dimensions, the resulting target test dataset and target training data can be made more suitable for training a lead-acid battery SOH prediction model, thereby improving the prediction robustness and accuracy of the SOH prediction model.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for dividing a lead-acid battery performance testing dataset, comprising the following steps: S1. Using the traditional random partitioning method, the sample time in the sample time set is randomly partitioned to obtain a preset number of feature combinations. Each feature combination includes a training sample time set and a test sample time set. S2. For any feature combination, the training sample time set and the test sample time set in the feature combination are respectively denoted as the first set and the second set. The first feature sequence of the first set and the second feature sequence of the second set are obtained. Based on the difference of SOH values ​​between adjacent sample times in the first feature sequence and the second feature sequence, the difference between the SOH values ​​belonging to the first feature sequence and the SOH values ​​belonging to the second feature sequence, and the proportion of noise monitoring data and abnormal monitoring data in the first set and the second set, the evaluation index value of the feature combination is obtained. The target test dataset and the target training dataset are obtained based on the evaluation index value.

[0006] In a preferred embodiment, in step S2, the first feature sequence of the first set is the result of arranging the sample times in the first set in chronological order, and the second feature sequence of the second set is the result of arranging the sample times in the second set in chronological order.

[0007] In a preferred embodiment, the method for obtaining the evaluation index value of the feature combination is as follows: calculate the temporal dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination, and then calculate the mean of the temporal dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination as the evaluation index value.

[0008] In a preferred embodiment, the method for calculating the temporal dependence correlation characterization value of the feature combination is as follows: obtaining the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence; and obtaining the temporal dependence correlation characterization value of the feature combination based on the correlation between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence.

[0009] In a preferred embodiment, the method for calculating the data distribution characteristic representation value of the feature combination is as follows: the first feature sequence and the second feature sequence are divided into subsequences of the first feature sequence and the second feature sequence using a preset length, respectively, to obtain subsequences of the first feature sequence and subsequences of the second feature sequence; the mean and variance of the SOH values ​​of the monitoring data vectors corresponding to all sample times in the subsequences are calculated and recorded as the mean and variance of the SOH of the corresponding subsequences; and the data distribution characteristic representation value of the feature combination is obtained based on the difference in mean and variance of SOH between the subsequences of the first feature sequence and the subsequences of the second feature sequence.

[0010] In a preferred embodiment, the method for calculating the non-abnormal noise characterization value of the feature combination is as follows: the set of monitoring data constructed from all monitoring data vectors belonging to the first set is denoted as the first dataset, and the set of monitoring data constructed from all monitoring data vectors belonging to the second set is denoted as the second dataset. The absolute value of the Z-score of each monitoring data in the first dataset and the second dataset is calculated as the outlier value of the corresponding monitoring data. Based on the outlier values ​​of each monitoring data in the first dataset and the second dataset, and the signal-to-noise ratio of the first dataset and the second dataset, the non-abnormal noise characterization value of the feature combination is obtained.

[0011] In a preferred embodiment, the method for constructing the SOH change characterization value sequence is as follows: the a-th SOH change characterization value in the SOH change characterization value sequence of any feature sequence is the ratio of the result obtained by subtracting the SOH value of the monitoring data vector corresponding to the a-th sample time from the SOH value of the monitoring data vector corresponding to the (a+1)-th sample time in the corresponding feature sequence to the SOH value of the monitoring data vector corresponding to the a-th sample time.

[0012] In a preferred embodiment, the time-dependent correlation characterization value of the feature combination is the absolute value of the Pearson correlation coefficient between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence.

[0013] In a preferred embodiment, the method for obtaining the data distribution characteristic representation value of the feature combination based on the difference in SOH mean and SOH variance between the subsequences of the first feature sequence and the subsequences of the second feature sequence is as follows: based on the difference in SOH mean and SOH variance between the subsequences of the first feature sequence and the subsequences of the second feature sequence, obtain the SOH mean difference sequence and the SOH variance difference sequence; calculate the mean between the mean of the SOH mean difference sequence and the mean of the SOH variance difference sequence as the feature mean; and use the result of negatively correlated mapping of the feature mean as the data distribution characteristic representation value of the feature combination.

[0014] In a preferred embodiment, the method for constructing the SOH mean difference sequence and the SOH variance difference sequence is as follows: the b-th SOH mean difference in the SOH mean difference sequence is the absolute value of the difference between the SOH mean of the b-th subsequence in the first feature sequence and the SOH mean of the b-th subsequence in the second feature sequence; the b-th SOH variance difference in the SOH variance difference sequence is the absolute value of the difference between the SOH variance of the b-th subsequence in the first feature sequence and the SOH variance of the b-th subsequence in the second feature sequence.

[0015] In a preferred embodiment, the method for obtaining the non-abnormal noise characterization value of the feature combination based on the outliers of each monitoring data in the first dataset and the second dataset, and the signal-to-noise ratio of the first dataset and the second dataset, is as follows: the total amount of data in the comprehensive dataset formed by the first dataset and the second dataset is recorded as the comprehensive quantity value; in the first dataset and the second dataset, the number of all monitoring data whose outliers do not belong to a preset interval is counted and recorded as the number of abnormal monitoring data; the ratio of the number of abnormal monitoring data to the comprehensive quantity value is used as the outlier characterization value; the normalized result of the mean of the signal-to-noise ratio of the first dataset and the signal-to-noise ratio of the second dataset is used as the non-noise characterization value; the complement of the outlier characterization value and the mean of the non-noise characterization value are recorded as the non-abnormal noise characterization value of the feature combination.

[0016] In a preferred embodiment, the method for obtaining the target test dataset and the target training dataset based on the evaluation index value is as follows: among all feature combinations, the feature combination with the largest evaluation index value is selected as the target combination, and the set of all monitoring data vectors corresponding to all sample times in the training sample time set of the target combination is used as the target training dataset, and the set of all monitoring data vectors corresponding to all sample times in the test sample time set of the target combination is used as the target test dataset.

[0017] This invention also provides a method for testing the performance of lead-acid batteries, comprising the following steps: Step 1: Obtain the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time. The monitoring data in the monitoring data vector belongs to the lead-acid battery. Step 2: Based on the above-described method for partitioning a lead-acid battery performance testing dataset, obtain the optimal training set and test set; Step 3: Train a neural network model based on the target test dataset and the target training dataset for lead-acid battery SOH prediction and performance testing.

[0018] The present invention also provides a lead-acid battery performance testing system for performing the lead-acid battery performance testing method described above, comprising: The data acquisition module acquires the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time, wherein the monitoring data in the monitoring data vector belongs to the lead-acid battery. The dataset partitioning and evaluation module divides the data to obtain the optimal training and test sets; The SOH prediction and performance evaluation module has a built-in neural network model that, after being trained using the optimal training and test sets, is used to predict the SOH value of lead-acid batteries in real time and evaluate battery performance accordingly.

[0019] The present invention provides a method for dividing a dataset, a testing method, and a system for performance testing of lead-acid batteries, which have the following beneficial effects: 1. To address the problems of temporal disorder, uneven data distribution, and concentration of outliers and noise in traditional random partitioning methods, this invention evaluates different feature combinations from multiple dimensions. This makes the target test dataset and target training data obtained from the partitioning more suitable for training the SOH prediction model of lead-acid batteries, thereby improving the prediction robustness and accuracy of the SOH prediction model and ensuring the accuracy and reliability of subsequent performance testing of lead-acid batteries.

[0020] 2. The dataset partitioning results are comprehensively evaluated through three core dimensions: temporal dependency correlation, data distribution characteristics, and non-abnormal noise. This ensures accurate adaptation to the core characteristics of lead-acid battery SOH data. On the one hand, it strengthens the temporal consistency of the SOH change rate between the training and test sets, conforming to the physical law of continuous battery aging, avoiding the temporal logic break caused by traditional random partitioning, and ensuring that the temporal evolution law learned by the model can be transferred. On the other hand, through the quantification and optimization of mean and variance differences, it ensures the data distribution homogeneity between the training and test sets, avoiding the problem of distribution imbalance. On the third hand, through outlier screening and signal-to-noise ratio quantification, it reduces the concentrated interference of extreme data and noise, improves the purity of the dataset, and fundamentally solves the quality defects of traditional random partitioning, such as temporal disorder, uneven distribution, and abnormal noise concentration.

[0021] 3. The neural network model is trained using the optimized target training dataset and test dataset. The training set can comprehensively cover the characteristics and patterns of batteries at different aging stages and under different operating conditions, avoiding underfitting of the model due to insufficient data representativeness. The high representativeness of the test set ensures that the model evaluation results are true and reliable, avoiding misjudgment of generalization ability due to test data bias. Ultimately, the prediction accuracy and robustness of the SOH prediction model are significantly improved, and prediction errors are reduced. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the lead-acid battery performance testing method in this embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0024] Example 1: A method for testing the performance of lead-acid batteries, such as Figure 1 As shown, it includes the following steps: S001. Obtain the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time.

[0025] This embodiment primarily improves the robustness and accuracy of the lead-acid battery SOH prediction model by optimizing the partitioning strategy of the training and test datasets used during model training. This enhances the prediction accuracy of the SOH prediction model, thereby ensuring the accuracy and reliability of subsequent performance testing of lead-acid batteries. The battery SOH value is related to the battery's charging and discharging data during operation. Therefore, the data used to train the SOH prediction model in this embodiment is the charging and discharging data of the lead-acid battery. This data includes, but is not limited to, data such as voltage, current, temperature, and internal resistance. The battery's operation process refers to its charging and discharging process.

[0026] This embodiment first selects a lead-acid battery for training and designates it as a sample lead-acid battery. Then, it acquires all the times when the sample lead-acid battery collects charge and discharge data during its historical operation and designates them as sample times. The set of the acquired sample times is designated as the sample time set. Finally, it retrieves all the charge and discharge data of the sample lead-acid battery collected at each sample time from the battery management system (BMS).

[0027] The acquired charging and discharging data are preprocessed, and the preprocessed data is recorded as monitoring data. Therefore, through the above process, all monitoring data of the sample lead-acid battery collected at each sample time can be obtained. The vector constructed from all the monitoring data of the sample lead-acid battery collected at the corresponding sample time is recorded as the monitoring data vector at the corresponding sample time. The sample times in the sample time set all belong to the time when the charging and discharging data of the same sample lead-acid battery is collected during operation. Moreover, the set constructed from the monitoring data vectors corresponding to all sample times in the sample time set is the sample dataset used for training the prediction model in this embodiment. Therefore, the target test dataset and target training dataset obtained later in this embodiment for training the prediction model belong to the sample dataset. In addition, data preprocessing includes, but is not limited to, data cleaning and data normalization of the acquired charging and discharging data. Data normalization refers to scaling all charging and discharging data to a uniform range, such as scaling to the range of 0 to 1, in order to accelerate the training of the neural network and improve the stability of the network.

[0028] In this embodiment, the type and quantity of charge and discharge data collected at each time point of charge and discharge data acquisition are consistent. For example, if the charge and discharge data collected at a certain time point is the voltage, current, temperature, and internal resistance of the sample lead-acid battery, then the charge and discharge data collected at all times of charge and discharge data acquisition in this embodiment are voltage, current, temperature, and internal resistance. Therefore, the type and quantity of monitoring data in the monitoring data vector corresponding to all sample times in this embodiment are consistent, and the data types at the same position in the monitoring data vector corresponding to different sample times are the same. For example, if the first monitoring data in the monitoring data vector corresponding to a certain sample time is the result of data preprocessing of the voltage of the sample lead-acid battery, then the first monitoring data in the monitoring data vector corresponding to all sample times is the result of data preprocessing of the voltage of the sample lead-acid battery.

[0029] Since the training data also needs to be labeled during model training, this embodiment needs to calculate the SOH value at each sample time using the internal resistance of the lead-acid battery collected at each sample time, and record the calculated SOH value at each sample time as the SOH value of the monitoring data vector corresponding to the sample time. The method of obtaining the SOH value of the battery through internal resistance is a known technology, so this embodiment will not describe it in detail.

[0030] Therefore, this embodiment obtains the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time through the above process, and the monitoring data in the monitoring data vector belongs to the sample lead-acid battery.

[0031] S002. Based on the data set partitioning method for lead-acid battery performance testing, obtain the optimal training set and test set.

[0032] When using neural networks to predict battery state of health (SOH), the partitioning of the training and test sets is crucial, especially considering the unique characteristics of SOH data. Improper partitioning can lead to inaccurate model performance evaluation, thus affecting the model's reliability in practical applications. The main reasons are: First, SOH data typically exhibits strong time-series characteristics, meaning the battery aging process is continuous, and later SOH values ​​are closely related to earlier charge-discharge histories. For example, capacity decay is a cumulative process, and the SOH of the current cycle is related to parameters such as voltage, current, and temperature of previous cycles. In this case, using a traditional random partitioning method might lead to temporal overlap or confusion between the training and test sets.

[0033] Second, the characteristics of data distribution differ; the aging process of a battery can cause the data distribution to change over time. For example, a battery's capacity decays more slowly in early cycles and accelerates in later cycles. This change in distribution can lead to performance differences in the model at different time points. Random partitioning can also result in inconsistent data distributions between the training and test sets. For example, the training set might contain more early data with slow decay, while the test set might contain more late data with fast decay, leading to poor model performance on the test set.

[0034] Third, the data in SOH is diverse and relatively complex, with a high probability of outliers or noise. The random partitioning in the traditional method leads to outliers or noise being concentrated in the training or test set, affecting the generalization ability of the model.

[0035] Therefore, this step presents a method for partitioning a lead-acid battery performance testing dataset, including the following steps: S1. Using the traditional random partitioning method, the sample time in the sample time set is randomly divided to obtain a preset number of feature combinations. Each feature combination includes a training sample time set and a test sample time set.

[0036] The sample time sets are randomly divided a preset number of times according to a preset ratio, resulting in feature combinations in each division. The number of feature combinations is a preset number, and each feature combination consists of a training sample time set and a test sample time set. The training and test sample time sets in any feature combination are obtained from the same division. The ratio of the total number of sample times in the training sample time set to the total number of sample times in the test sample time set of any feature combination is a preset ratio. In this embodiment, the preset ratio is set to an empirical value, such as a typical ratio of 70% for the training set and 30% for the test set. Therefore, the preset ratio in this embodiment is 7:3, meaning that the sample times in the training sample time set of any feature combination account for 70% of the sample time set, and the sample times in the test sample time set of any feature combination account for 30% of the sample time set. For example, for any feature combination, after dividing and determining the training sample time set and the test sample time set in the feature combination, the first 70% of the sample time in the sample time set can be used as the training sample time set in the feature combination, and the remaining 30% of the sample time in the sample time set can be used as the test sample time set in the feature combination. Moreover, this embodiment requires that the number of sample time in the training sample time set in different feature combinations be the same, but the sample time contained in the training sample time set in different feature combinations is different. The same applies to the test sample time set in different feature combinations. If the sample time in the training sample time set of a certain feature combination consists of the first 70% of the sample time in the sample time set, then the training sample time set in other feature combinations will not also consist of the first 70% of the sample time in the sample time set.

[0037] In practical applications, implementers also need to set the preset number of values ​​based on practical experience, the size of the set to be divided, the complexity of the prediction model, and other practical considerations. For example, if the sample time set is large and the trained prediction model is highly complex, the preset number can be set to 5 or 10 to save computing resources. If the sample time set is small and the trained prediction model is less complex, the preset number can be set to 20 to obtain more stable evaluation results. That is, for small sets, more partitions may be needed, i.e., a larger preset number of values, to obtain more stable evaluation results. This is because fewer partitions may result in less representative distributions of the training and test sets due to randomness. For large datasets, however, due to sufficient data volume, fewer partitions are already stable enough, so the preset number of values ​​can be relatively small.

[0038] S2. For any feature combination, the training sample time set and the test sample time set in the feature combination are respectively denoted as the first set and the second set. The first feature sequence of the first set and the second feature sequence of the second set are obtained. Based on the difference of SOH values ​​between adjacent sample times in the first feature sequence and the second feature sequence, the difference between the SOH values ​​belonging to the first feature sequence and the SOH values ​​belonging to the second feature sequence, and the proportion of noise monitoring data and abnormal monitoring data in the first set and the second set, the evaluation index value of the feature combination is obtained. The target test dataset and the target training dataset are obtained based on the evaluation index value.

[0039] The method for obtaining the evaluation index value of the feature combination is as follows: calculate the time-series dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination, and then calculate the mean of the time-series dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination as the evaluation index value.

[0040] This embodiment will then analyze different feature combinations from three dimensions: time-series dependency relevance characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value, thereby obtaining the evaluation result of each feature combination. Subsequently, feature combinations will be selected based on the evaluation results to determine the target test dataset and the target training dataset.

[0041] Since the evaluation process is the same for different feature combinations, this embodiment will use the process of obtaining the evaluation index value of any feature combination Q as an example for ease of subsequent description.

[0042] 1. Temporal dependence correlation representation value of feature combinations (1) Obtain the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence.

[0043] The method for constructing the SOH change characterization value sequence is as follows: the SOH change characterization value of the a-th value in the SOH change characterization value sequence of any feature sequence is the ratio of the SOH value of the monitoring data vector corresponding to the (a+1)-th sample time in the feature sequence minus the SOH value of the monitoring data vector corresponding to the a-th sample time in the feature sequence to the SOH value of the monitoring data vector corresponding to the a-th sample time in the feature sequence.

[0044] The a-th SOH change characterization value The calculation formula is: ; in, Let SOH be the SOH value of the monitoring data vector corresponding to the (a+1)th sample time in the first feature sequence. Let SOH be the SOH value of the monitoring data vector corresponding to the a-th sample time in the first feature sequence, where a is greater than or equal to 1.

[0045] This method of calculating the SOH change value can reflect the relative change of the SOH value at any given time relative to the previous time, thus more accurately reflecting the degradation rate of the battery's health status, making the evaluation index values ​​of the feature combinations obtained in subsequent analysis more valuable.

[0046] (2) Based on the correlation between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence, the time-dependent correlation characterization value of the feature combination Q is obtained.

[0047] In this embodiment, the temporal dependence correlation characterization value of the feature combination is the absolute value of the Pearson correlation coefficient between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence.

[0048] The specific process for obtaining the time-dependent correlation characterization value of feature combination Q based on the correlation between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence is as follows: Calculate the Pearson correlation coefficient between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence, and denote the absolute value of the calculated Pearson correlation coefficient between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence as the time-dependent correlation characterization value of the feature combination Q.

[0049] In battery state of health (SOH) prediction scenarios, SOH data itself exhibits strong time dependence. If the SOH values ​​in the test dataset and the training dataset show a high temporal dependence, it means that this data partitioning method better simulates the real situation. The knowledge and patterns learned by the prediction model on the training set can be more naturally transferred to the test set, resulting in predictions that are closer to reality or have higher accuracy. Furthermore, the larger the temporal dependence correlation value of feature combination Q (i.e., the larger the absolute value of the Pearson correlation coefficient), the stronger the temporal dependence correlation between the SOH change value sequence of the first feature sequence and the SOH change value sequence of the second feature sequence. The greater the temporal dependency between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence, the higher the temporal dependency between the SOH values ​​of the data in the training dataset determined by the training sample time set in feature combination Q and the SOH values ​​of the data in the test dataset determined by the test sample time set. This means that the subsequent test dataset and training dataset obtained based on the training sample time set and test sample time set in feature combination Q will have higher robustness and accuracy of the trained SOH prediction model. In other words, the greater the temporal dependency characterization value of feature combination Q, the more suitable the test dataset and training dataset obtained based on the training sample time set and test sample time set in feature combination Q are for training the SOH prediction model of lead-acid batteries.

[0050] 2. Data distribution characteristic representation value of feature combination (1) Divide the first feature sequence and the second feature sequence into subsequences on the first feature sequence and subsequences on the second feature sequence using a preset length.

[0051] In this embodiment, the specific process of dividing the first feature sequence and the second feature sequence into subsequences on the first feature sequence and the second feature sequence using preset lengths is as follows: For the first feature sequence, starting from the first sample time of the first feature sequence, the first feature sequence is divided into non-overlapping segments with a preset length to obtain all subsequences of the first feature sequence; For the second feature sequence, starting from the first sample time point, the second feature sequence is divided into non-overlapping subsequences of a preset length to obtain all subsequences. The purpose of dividing the feature sequence is to analyze the differences in the data distribution characteristics of the feature combination Q, that is, to analyze the distribution differences between the monitoring data in the training sample time set belonging to the feature combination Q and the monitoring data in the test sample time set belonging to the feature combination Q.

[0052] Except for the last subsequence in the feature sequence, the number of sample times in other subsequences is a preset length. However, the number of sample times in the last subsequence of the feature sequence may be less than or equal to the preset length. Furthermore, in practical applications, the implementer needs to set the value of the preset length based on actual conditions such as experimental statistics and the length of the feature sequence. For example, in this embodiment, the preset length can be set to 10.

[0053] (2) Calculate the mean and variance of the SOH values ​​of the monitoring data vectors corresponding to all sample times in the subsequence, and record them as the mean and variance of the SOH of the corresponding subsequence. Based on the difference in the mean and variance of SOH between the subsequences of the first feature sequence and the subsequences of the second feature sequence, obtain the data distribution characteristic characterization value of the feature combination.

[0054] Specifically, the set of SOH values ​​of the monitoring data vectors corresponding to all sample times in each subsequence is denoted as the SOH value set corresponding to the subsequence. Then, the mean and variance of the SOH value set corresponding to each subsequence are calculated and denoted as the SOH mean and SOH variance of the corresponding subsequence. Based on the difference in SOH mean and SOH variance between the subsequences on the first feature sequence and the subsequences on the second feature sequence, the data distribution feature characterization value of the feature combination Q is obtained. The specific process is as follows: First, based on the differences in the mean and variance of SOH between the subsequences of the first and second feature sequences, we obtain the SOH mean difference sequence and the SOH variance difference sequence. The b-th SOH mean difference in the SOH mean difference sequence is the absolute value of the difference between the SOH mean of the b-th subsequence of the first feature sequence and the SOH mean of the b-th subsequence of the second feature sequence. Similarly, the b-th SOH variance difference in the SOH variance difference sequence is the absolute value of the difference between the SOH variance of the b-th subsequence of the first feature sequence and the SOH variance of the b-th subsequence of the second feature sequence. The number of data points is consistent with the number of the fewest subsequences on the first and second feature sequences. That is, if the number of subsequences on the first and second feature sequences is inconsistent, the number of SOH mean differences and SOH variance differences is dominated by the subsequences on the feature sequence with fewer subsequences. Specifically, the maximum value of b is N, where N is Min(N1, N2), where Min() is the minimum value function, N1 is the number of subsequences on the first feature sequence, and N2 is the number of subsequences on the second feature sequence. Then, the mean between the mean of the SOH mean difference sequence and the mean of the SOH variance difference sequence is calculated and used as the feature mean. The calculation formula is as follows: ; in, F The characteristic mean, f 1 represents the mean of the SOH mean difference sequence. f 2 represents the mean of the SOH variance difference sequence; then, a negative correlation mapping is performed on the feature mean, and the result of the negative correlation mapping is recorded as the data distribution feature representation value of the feature combination Q. For example, exp(-F) can be used as the data distribution feature representation value of the feature combination Q, where exp() is an exponential function with a base of constant e; as another real-time method, (1- F Q is the data distribution characteristic representation value of the feature combination Q.

[0055] Because battery aging causes data distribution to change over time—for example, battery capacity decays more slowly in early charge-discharge cycles and accelerates in later cycles—if the resulting training dataset contains more data from early charge-discharge cycles and the test dataset contains more data from later cycles, the prediction model will perform poorly on the test dataset, thus affecting its robustness and accuracy. Furthermore, a larger data distribution feature value in feature combination Q indicates a smaller difference in distribution characteristics between the training dataset (defined by the training sample time set) and the test dataset (defined by the test sample time set). Conversely, a smaller difference in distribution characteristics indicates a greater variation in the distribution characteristics of data from different charge-discharge cycles in the training dataset. The more uniform the distribution in the training and test datasets, the more suitable they are for training the lead-acid battery SOH prediction model. Here, the training and test datasets refer to the training dataset determined by the training sample time set in feature combination Q and the test dataset determined by the test sample time set in feature combination Q, respectively. In other words, the smaller the difference in distribution characteristics, the less likely the training dataset determined by the training sample time set in feature combination Q contains more data from the early charge-discharge cycle stages, and the less likely the test dataset determined by the test sample time set in feature combination Q contains more data from the later charge-discharge cycle stages. Therefore, the larger the data distribution characteristic value of feature combination Q, the more suitable the test and training datasets obtained based on the training and test sample time sets in feature combination Q are for training the lead-acid battery SOH prediction model.

[0056] 3. Non-abnormal noise characterization values ​​of feature combinations (1) The set of monitoring data vectors corresponding to all sample times in the first set is called the first vector set. The set of monitoring data vectors in the first set is called the first dataset. That is, the first dataset contains data from all vectors in the first vector set. The set of monitoring data vectors corresponding to all sample times in the second set is called the second vector set. The set of monitoring data vectors in the second set is called the second dataset. That is, the second dataset contains data from all vectors in the second vector set.

[0057] (2) Obtain the signal-to-noise ratio of the first dataset and the signal-to-noise ratio of the second dataset.

[0058] (3) Use the Z-score method to detect outliers in the first and second datasets.

[0059] The specific process is as follows: In the first dataset, all monitoring data of the same data type are arranged in the order of collection to obtain multiple data sequences. The number of data sequences obtained is consistent with the number of types of charging and discharging data collected at the above sample time. All data in the same data sequence are of the same data type. Then, the signal-to-noise ratio (SNR) of each data sequence is calculated, and the mean of the SNR of all data sequences is taken as the SNR of the first dataset. For any monitoring data in the first dataset, the data sequence to which the monitoring data belongs is first obtained and denoted as sequence G. The ratio of the monitoring data minus the mean of sequence G to the standard deviation of sequence G is calculated and denoted as the Z-score value of the monitoring data. The absolute value of the Z-score value of the monitoring data is taken as the outlier of the monitoring data.

[0060] The method for obtaining outliers in the second dataset is the same as that for obtaining outliers in the first dataset, so it will not be described in detail again.

[0061] (4) Based on the outliers of each monitoring data in the first dataset, the outliers of each monitoring data in the second dataset, the signal-to-noise ratio of the first dataset, and the signal-to-noise ratio of the second dataset, the non-abnormal noise characterization value of the feature combination Q is obtained. The specific process is as follows: First, the new dataset formed by the first dataset and the second dataset is denoted as the comprehensive dataset, and the total amount of monitoring data in the comprehensive dataset is denoted as the comprehensive quantity value.

[0062] In the first dataset, the number of all monitoring data with outliers greater than a preset outlier threshold is counted and recorded as the first quantity value. In the second dataset, the number of all monitoring data with outliers greater than the preset outlier threshold is counted and recorded as the second quantity value. The sum of the first quantity value and the second quantity value is recorded as the number of outlier monitoring data in the first and second datasets. That is, monitoring data with outliers greater than the preset outlier threshold is defined as outlier monitoring data. In this embodiment, the preset outlier threshold is an empirical value. Since data is generally marked as outlier data when its standard deviation from the mean is greater than 3, and since outlier values ​​are the result of taking absolute values, the preset outlier threshold in this embodiment is 3.

[0063] Then, the ratio of the number of abnormal monitoring data to the total number of data is calculated and recorded as the abnormality characterization value.

[0064] The mean SNR of the first dataset and the second dataset are calculated, and the mean SNR of the first dataset and the second dataset are normalized using the normalization function Norm(). The normalization result is used as the non-noise characterization value. Finally, the complement of the outlier characterization value and the mean of the non-noise characterization value are calculated and used as the non-outlier noise characterization value of the feature combination Q. That is, (1-H) and the mean of the non-noise characterization value are the non-outlier noise characterization value of the feature combination Q, and H is the outlier characterization value.

[0065] The expression for calculating the non-abnormal noise characterization value of the feature combination Q is: ; in, P The non-abnormal noise characterization value of the feature combination Q, H These are abnormal characterization values. SNR1 The signal-to-noise ratio of the first dataset is... SNR2 Let a1 be the signal-to-noise ratio of the second dataset, a2 be the first weight, and a1 be the second weight. Since the two-term formula evaluates from different aspects, the scheme gives the same weight allocation, which is 0.5 in this embodiment.

[0066] When abnormal monitoring data or noisy data are concentrated in the training or test datasets, it will affect the model's generalization ability, thus affecting the robustness and accuracy of the prediction model. Furthermore, the smaller the abnormal characterization value, the less abnormal monitoring data is in the first and second datasets. This means that the proportion of abnormal monitoring data in the training dataset and test dataset determined by the training sample time set in feature combination Q is smaller. Conversely, the larger the non-noise characterization value, i.e., the higher the signal-to-noise ratio of the dataset, the less noisy data is in the first and second datasets. This means that the proportion of noisy data or the noise level in the training dataset and test dataset determined by the training sample time set in feature combination Q is smaller. Therefore, the smaller the abnormal characterization value and the larger the non-noise characterization value, i.e., the larger the non-abnormal noise characterization value of feature combination Q, the more suitable the test and training datasets obtained based on the training and test sample time sets in feature combination Q are for training the SOH prediction model of lead-acid batteries.

[0067] In this embodiment, the specific process of obtaining the evaluation index value of feature combination Q based on its temporal dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value is as follows: The mean of the temporal dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of feature combination Q is calculated and denoted as the evaluation index value of feature combination Q. The calculation formula is: ; in, r This represents the temporal dependency correlation characterization value of the feature combination Q; F The data distribution characteristic representation value of the feature combination Q; P This is the non-abnormal noise characterization value of the feature combination Q.

[0068] The above formula shows that when the time-series dependency correlation value, data distribution characteristic value, and non-abnormal noise value of the feature combination Q are larger, or when the evaluation index value of the feature combination Q is larger, it indicates that the test dataset and training dataset obtained based on the training sample time set and test sample time set in the feature combination Q are more suitable for training the SOH prediction model of lead-acid batteries, and the prediction model trained based on the test dataset and training dataset obtained based on the training sample time set and test sample time set in the feature combination Q has better robustness and accuracy.

[0069] Therefore, this embodiment can obtain the evaluation index value of each feature combination by using the above-described method for obtaining the evaluation index value of the feature combination Q. The larger the evaluation index value of the feature combination, the better the robustness and accuracy of the prediction model trained based on the monitoring data vectors corresponding to the sample times in the training sample time set and the monitoring data vectors corresponding to the sample times in the test sample time set of the feature combination.

[0070] Evaluation metrics were used to evaluate the generated training and test sets. The calculation is performed, and the training and test sets corresponding to the maximum values ​​are selected as the basis for subsequent neural network processing.

[0071] S003. Train a neural network model based on the optimal training set and test set for lead-acid battery SOH prediction and performance testing.

[0072] Neural network architectures can include feedforward neural networks (FNNs), suitable for simple feature inputs and SOH prediction; or convolutional neural networks (CNNs), suitable for extracting local features from signals such as voltage and current; or adaptive neural network fuzzy inference systems (ANFIS), which can handle complex uncertainties and nonlinear systems. ANFIS combines the self-learning of neural networks with the heuristic knowledge of fuzzy inference to achieve accurate prediction of battery SOH.

[0073] Model Design: Input layer: The feature dimension is consistent with the number of features after feature engineering.

[0074] Hidden layers: Choose 1-3 layers depending on the complexity, with the number of neurons in each layer typically being 1-2 times the input dimension.

[0075] Output layer: a single neuron that outputs the SOH value (usually between 0 and 1, indicating the battery health status).

[0076] Activation functions: ReLU is used in the hidden layer, and a linear activation function is used in the output layer (for regression problems).

[0077] Loss function: Mean Squared Error (MSE) or Mean Absolute Error (MAE), used to measure the difference between the predicted value and the actual value.

[0078] Optimizer: Adam optimizer, learning rate is usually set to 0.001.

[0079] The model is trained using mini-batch gradient descent with a batch size of 32 or 64.

[0080] Evaluation metrics can include mean square error, root mean square error, mean absolute error, and coefficient of determination.

[0081] Model optimization: Adjust hyperparameters (such as learning rate, batch size, number of hidden layer neurons). Use regularization techniques (such as L1 / L2 regularization, Dropout) to prevent overfitting. Try different neural network architectures (such as increasing the number of layers, changing the activation function).

[0082] Cross-validation is used to evaluate the model's generalization ability.

[0083] The trained model is deployed to the BMS to predict the SOH of lead-acid batteries in real time. The model's predictive performance is monitored periodically, and the model is retrained when the data distribution changes. Alternatively, other methods such as Kalman filtering and particle filtering can be combined to improve the robustness of the prediction.

[0084] This invention evaluates different feature combinations from multiple dimensions, making the resulting target test dataset and target training data more suitable for training the SOH prediction model of lead-acid batteries. This improves the prediction robustness and accuracy of the SOH prediction model, thereby ensuring the accuracy and reliability of subsequent performance testing of lead-acid batteries.

[0085] Example 2: This embodiment provides a lead-acid battery performance testing system for performing the lead-acid battery performance testing method described in Embodiment 1, including: The data acquisition module acquires the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time, wherein the monitoring data in the monitoring data vector belongs to the lead-acid battery. The dataset partitioning and evaluation module divides the data to obtain the optimal training and test sets; The SOH prediction and performance evaluation module has a built-in neural network model that, after being trained using the optimal training and test sets, is used to predict the SOH value of lead-acid batteries in real time and evaluate battery performance accordingly.

[0086] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for partitioning a dataset for lead-acid battery performance testing, characterized in that, Includes the following steps: S1. Using the traditional random partitioning method, the sample time in the sample time set is randomly partitioned to obtain a preset number of feature combinations. Each feature combination includes a training sample time set and a test sample time set. S2. For any feature combination, the training sample time set and the test sample time set in the feature combination are respectively denoted as the first set and the second set. The first feature sequence of the first set and the second feature sequence of the second set are obtained. Based on the difference of SOH values ​​between adjacent sample times in the first feature sequence and the second feature sequence, the difference between the SOH values ​​belonging to the first feature sequence and the SOH values ​​belonging to the second feature sequence, and the proportion of noise monitoring data and abnormal monitoring data in the first set and the second set, the evaluation index value of the feature combination is obtained. The target test dataset and the target training dataset are obtained based on the evaluation index value.

2. The method for dividing a lead-acid battery performance testing dataset according to claim 1, characterized in that, In step S2, the first feature sequence of the first set is the result of arranging the sample times in the first set in chronological order, and the second feature sequence of the second set is the result of arranging the sample times in the second set in chronological order.

3. The method for dividing a lead-acid battery performance testing dataset according to claim 1, characterized in that, The method for obtaining the evaluation index value of the feature combination is as follows: calculate the time-series dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination, and then calculate the mean of the time-series dependency correlation characterization value, data distribution characteristic characterization value, and non-abnormal noise characterization value of the feature combination as the evaluation index value.

4. The method for dividing a lead-acid battery performance testing dataset according to claim 3, characterized in that, The method for calculating the temporal dependence correlation characterization value of the feature combination is as follows: obtain the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence; based on the correlation between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence, obtain the temporal dependence correlation characterization value of the feature combination.

5. The method for dividing a lead-acid battery performance testing dataset according to claim 3, characterized in that, The method for calculating the data distribution characteristic representation value of the feature combination is as follows: the first feature sequence and the second feature sequence are divided into subsequences of the first feature sequence and the second feature sequence using a preset length, respectively. The mean and variance of the SOH values ​​of the monitoring data vectors corresponding to all sample times in the subsequences are calculated and recorded as the mean and variance of the SOH of the corresponding subsequences. The data distribution characteristic representation value of the feature combination is obtained based on the difference in mean and variance of SOH between the subsequences of the first feature sequence and the subsequences of the second feature sequence.

6. The method for dividing a lead-acid battery performance testing dataset according to claim 3, characterized in that, The method for calculating the non-abnormal noise characterization value of the feature combination is as follows: the set of monitoring data constructed from all monitoring data vectors belonging to the first set is denoted as the first dataset, and the set of monitoring data constructed from all monitoring data vectors belonging to the second set is denoted as the second dataset. The absolute value of the Z-score of each monitoring data in the first dataset and the second dataset is calculated as the outlier value of the corresponding monitoring data. Based on the outlier values ​​of each monitoring data in the first dataset and the second dataset, and the signal-to-noise ratio of the first dataset and the second dataset, the non-abnormal noise characterization value of the feature combination is obtained.

7. The method for dividing a lead-acid battery performance testing dataset according to claim 4, characterized in that, The method for constructing the SOH change characterization value sequence is as follows: the a-th SOH change characterization value in the SOH change characterization value sequence of any feature sequence is the ratio of the result obtained by subtracting the SOH value of the monitoring data vector corresponding to the a-th sample time from the SOH value of the monitoring data vector corresponding to the (a+1)-th sample time in the corresponding feature sequence to the SOH value of the monitoring data vector corresponding to the a-th sample time.

8. The method for dividing a lead-acid battery performance testing dataset according to claim 4, characterized in that, The temporal dependence correlation characterization value of the feature combination is the absolute value of the Pearson correlation coefficient between the SOH change characterization value sequence of the first feature sequence and the SOH change characterization value sequence of the second feature sequence.

9. The method for dividing a lead-acid battery performance testing dataset according to claim 5, characterized in that, The method for obtaining the data distribution characteristic representation value of the feature combination based on the difference in SOH mean and SOH variance between the subsequences of the first feature sequence and the subsequences of the second feature sequence is as follows: based on the difference in SOH mean and SOH variance between the subsequences of the first feature sequence and the subsequences of the second feature sequence, obtain the SOH mean difference sequence and the SOH variance difference sequence; calculate the mean between the mean of the SOH mean difference sequence and the mean of the SOH variance difference sequence as the feature mean; and use the result of negatively correlated mapping of the feature mean as the data distribution characteristic representation value of the feature combination.

10. The method for partitioning a lead-acid battery performance testing dataset according to claim 9, characterized in that, The method for constructing the SOH mean difference sequence and the SOH variance difference sequence is as follows: the b-th SOH mean difference in the SOH mean difference sequence is the absolute value of the difference between the SOH mean of the b-th subsequence in the first feature sequence and the SOH mean of the b-th subsequence in the second feature sequence; the b-th SOH variance difference in the SOH variance difference sequence is the absolute value of the difference between the SOH variance of the b-th subsequence in the first feature sequence and the SOH variance of the b-th subsequence in the second feature sequence.

11. The method for partitioning a lead-acid battery performance testing dataset according to claim 6, characterized in that, The method for obtaining the non-abnormal noise characterization value of the feature combination based on the outliers of each monitoring data in the first dataset and the second dataset, and the signal-to-noise ratio of the first dataset and the second dataset, is as follows: the total amount of data in the comprehensive dataset formed by the first dataset and the second dataset is recorded as the comprehensive quantity value; in the first dataset and the second dataset, the number of all monitoring data whose outliers do not belong to a preset interval is counted and recorded as the number of abnormal monitoring data; the ratio of the number of abnormal monitoring data to the comprehensive quantity value is used as the outlier characterization value; the normalized result of the mean of the signal-to-noise ratio of the first dataset and the signal-to-noise ratio of the second dataset is used as the non-noise characterization value; the complement of the outlier characterization value and the mean of the non-noise characterization value are recorded as the non-abnormal noise characterization value of the feature combination.

12. The method for dividing a lead-acid battery performance testing dataset according to claim 1, characterized in that, The method for obtaining the target test dataset and the target training dataset based on the evaluation index value is as follows: among all feature combinations, the feature combination with the largest evaluation index value is selected as the target combination, and the set of all monitoring data vectors corresponding to all sample times in the training sample time set of the target combination is used as the target training dataset, and the set of all monitoring data vectors corresponding to all sample times in the test sample time set of the target combination is used as the target test dataset.

13. A method for testing the performance of a lead-acid battery, characterized in that, Includes the following steps: Step 1: Obtain the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time. The monitoring data in the monitoring data vector belongs to the lead-acid battery. Step 2: According to any one of claims 1 to 12, a method for partitioning a lead-acid battery performance testing dataset is used to obtain the optimal training set and test set; Step 3: Train a neural network model based on the target test dataset and the target training dataset for lead-acid battery SOH prediction and performance testing.

14. A lead-acid battery performance testing system, characterized in that, A method for performing a lead-acid battery performance testing according to claim 13 includes: The data acquisition module acquires the monitoring data vector corresponding to each sample time in the sample time set and the SOH value of the monitoring data vector corresponding to each sample time, wherein the monitoring data in the monitoring data vector belongs to the lead-acid battery. The dataset partitioning and evaluation module divides the data to obtain the optimal training and test sets; The SOH prediction and performance evaluation module has a built-in neural network model that, after being trained using the optimal training and test sets, is used to predict the SOH value of lead-acid batteries in real time and evaluate battery performance accordingly.

Images