A power battery usage data repair system and method

Abstract

The application provides a power battery use data repairing system and method, which comprises a data cleaning module, a data quality preliminary evaluation module, a data repairing module and a data quality reevaluation module, which are arranged on a cloud battery management system big data platform and are sequentially connected; the data quality preliminary evaluation module performs data quality preliminary evaluation on the power battery use data after data cleaning, determines whether the data meets a lower limit of repairing and judges a data missing degree; the data repairing module performs data repairing on the data with three different missing degrees through a spline interpolation method, a time sequence prediction algorithm, a matrix decomposition algorithm and a deep learning algorithm; and finally, the data quality reevaluation module performs comprehensive evaluation on the data quality, so that the data repairing cost is reduced under the premise of ensuring the data repairing quality.

Description

Technical Field

[0001] This invention relates to the field of big data statistics and algorithms for power batteries, and specifically to a system and method for repairing power battery usage data. Background Technology

[0002] my country's new energy vehicle industry has entered a critical period of large-scale application. As a key node in the "three vertical and three horizontal" R&D layout, designing high-strength, lightweight, high-safety, low-cost, and long-life power batteries and management systems has become a critical challenge. Cloud-based battery management systems have emerged as an important technological direction for solving this problem. By establishing a digital twin model of the power battery in the cloud, it is possible to describe the battery's usage status, safety boundaries, performance levels, and other characteristics from multiple scales, levels, and dimensions. The calculation of these characteristics relies on the uploaded high-quality battery sensor data for parameter identification and model construction. However, in actual use, due to reasons such as board-to-cloud communication failures, sensor power outages, big data system component errors, and data anomalies caused by system cleanup, some sensor data may be lost. Therefore, how to recover the power battery data obtained during use has become an urgent problem to be solved.

[0003] The difficulty in data repair lies in the complexity and diversity of missing data generated by power batteries. Current industry-standard data repair methods can be broadly categorized into model-based and data similarity-based methods. Model-based methods typically establish a physical model of the power battery (such as an equivalent circuit model, electrochemical model, or empirical model), and use simulation to obtain the model output as the filler value. However, this method has two problems in practical use: First, the model uses current parameters as input; if current data is missing, the model cannot perform numerical simulation. Even if current is used as the output, the model may be unable to run if all data at a specific sampling time is missing. Second, power batteries undergo performance aging during use, while models are mostly calibrated in the initial stages of use, making it difficult to meet actual needs. Data similarity-based methods typically select similar pattern data fragments before the missing data point to replace it, or directly fill in the missing data with data near the missing time. This method may cause "conceptual drift" in the time series, resulting in another form of data failure.

[0004] In summary, existing technical solutions are not ideal for repairing data from power batteries and cannot meet the data repair needs of actual engineering projects. Summary of the Invention

[0005] To address the problem that current methods for repairing usage data in power batteries are unsatisfactory and fail to meet the data repair needs of practical engineering projects, this invention proposes a power battery usage data repair system. The system performs an initial quality assessment on cleaned usage data to determine whether the data meets the repair threshold and identifies the type of missing data. Then, for different types of missing data, it selects appropriate repair methods based on the algorithm's applicable scenario and computational complexity. This enables the output of reasonable, continuous, high-quality data, achieving high-quality and low-cost repair for different data types. Furthermore, it facilitates model parameter identification and the construction of a digital twin model of the power battery. This invention also relates to a method for repairing power battery usage data.

[0006] The technical solution of the present invention is as follows:

[0007] A power battery data repair system is characterized by comprising a data cleaning module, a data quality initial assessment module, a data repair module, and a data quality reassessment module, all deployed on a cloud-based battery management system big data platform and connected sequentially.

[0008] The data cleaning module is used to clean the power battery usage data obtained from the cloud-based battery management system big data platform, according to the data cleaning rules formulated by experts from battery manufacturers or vehicle manufacturers, to remove default values, abnormal values ​​and null values ​​from the data.

[0009] The data quality initial assessment module is used to evaluate the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. The data cleaning rate is the ratio of the length of the cleaned power battery data to the length of the original data. The sampling stability rate is the proportion of cleaned power battery data whose sampling interval is consistent with the standard sampling interval. The sampling continuity rate is the proportion of cleaned power battery data whose sampling interval is consistent with the standard sampling interval and whose consecutive number is greater than or equal to a preset length threshold. Based on the quality assessment results, the module determines whether the power battery usage data meets the data repair lower limit. For power battery usage data that meets the data repair lower limit, the module compares the data with the constant operating condition data threshold, the dynamic operating condition occasional threshold, and the dynamic operating condition interval threshold to confirm the data missing type. Based on the data missing type, the power battery usage data is then transmitted to the corresponding unit of the data repair module.

[0010] The data repair module includes a constant operating condition data unit, a dynamic operating condition intermittent data unit, and a dynamic operating condition interval data unit. The constant operating condition data unit receives power battery usage data under constant operating conditions transmitted by the initial quality assessment module and repairs it using a spline interpolation method. The dynamic operating condition intermittent data unit receives power battery usage data with individual missing data under dynamic operating conditions transmitted by the initial quality assessment module and repairs it using a time-series prediction algorithm. The dynamic operating condition interval data unit receives power battery usage data with intermittent missing data under dynamic operating conditions transmitted by the initial quality assessment module and repairs the data using a matrix factorization algorithm or a deep learning algorithm.

[0011] The data quality reassessment module is used to re-evaluate whether the data repaired by the data repair module can support the function implementation, and to evaluate the performance of the repaired data in practical applications.

[0012] Preferably, the quality assessment results in the data quality initial assessment module include the calculation results of the three parameters: data cleaning rate, sampling stability rate, and sampling continuity rate. When all the quality assessment results are within the range defined by the constant operating condition data threshold, the data quality initial assessment module confirms it as the first type of data missing and then transmits the power battery usage data to the constant operating condition data unit of the data repair module. When all the quality assessment results are within the range defined by the dynamic operating condition occasional threshold, the quality initial assessment module confirms it as the second type of data missing and then transmits the power battery usage data to the dynamic operating condition occasional data unit of the data repair module. When all the quality assessment results are within the range defined by the dynamic operating condition interval threshold, the quality initial assessment module confirms it as the third type of data missing and then transmits the power battery usage data to the dynamic operating condition interval data unit of the data repair module.

[0013] Preferably, the power battery usage data includes four dimensions: voltage, current, temperature, and time, and the data repair module performs data repair on one or more of these dimensions simultaneously.

[0014] Preferably, in the data quality initial assessment module, the constant operating condition data threshold satisfies a data cleaning rate greater than the limit of the power battery data repair system algorithm capability, and a data sampling continuity rate greater than 95% for any dimension; and / or, the dynamic operating condition occasional threshold simultaneously satisfies a data cleaning rate greater than the limit of the power battery data repair system algorithm capability, a sampling stability rate greater than 80%, and a data sampling continuity rate greater than 95% for any dimension; and / or, the dynamic operating condition interval threshold simultaneously satisfies a data cleaning rate greater than the limit of the power battery data repair system algorithm capability, a sampling stability rate greater than 80%, and a data sampling continuity rate less than 95% but greater than 80% for any dimension.

[0015] Preferably, the data quality initial assessment module calculates the data missing rate based on the quality assessment results and then determines whether the power battery usage data meets the data repair lower limit.

[0016] Preferably, the time-series prediction algorithm in the data repair module includes the ARIMA algorithm and the LSTM algorithm; the deep learning algorithm includes the neural network algorithm.

[0017] A method for repairing data from a power battery, characterized by comprising the following steps:

[0018] The data cleaning step is used to clean the power battery usage data obtained from the cloud-based battery management system big data platform, according to the data cleaning rules formulated by experts from battery manufacturers or vehicle manufacturers, to remove defaults, anomalies and null values ​​from the data.

[0019] The initial data quality assessment step evaluates the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. The data cleaning rate is the ratio of the cleaned power battery data length to the original data length. The sampling stability rate is the proportion of cleaned power battery data whose sampling interval matches the standard sampling interval. The sampling continuity rate is the proportion of cleaned power battery data whose sampling interval matches the standard sampling interval and whose consecutive counts are greater than or equal to a preset length threshold. Based on the quality assessment results, it is determined whether the power battery usage data meets the data repair lower limit. For power battery usage data that meets the data repair lower limit, the results are compared with constant operating condition data thresholds, dynamic operating condition occasional thresholds, and dynamic operating condition interval thresholds to confirm the data missing type.

[0020] The data repair steps involve adopting corresponding repair methods based on different types of missing data, the applicable scenarios of the algorithms, and the amount of computation. For power battery usage data under constant operating conditions, the spline interpolation method is used for repair; for power battery usage data with individual missing data under dynamic operating conditions, the time series prediction algorithm is used for repair; and for power battery usage data with intermittent missing data under dynamic operating conditions, the matrix factorization algorithm or deep learning algorithm is used for data repair.

[0021] The data quality reassessment step is used to re-evaluate whether the repaired data can support the implementation of functions, and to evaluate the performance of the repaired data in practical applications.

[0022] Preferably, the quality assessment results in the initial data quality assessment step include the calculation results of the three parameters: data cleaning rate, sampling stability rate, and sampling continuity rate. When all the quality assessment results are within the range defined by the constant operating condition data threshold, the initial data quality assessment step confirms the first type of data missing, and the data repair step uses the spline interpolation method to repair the power battery usage data under constant operating conditions. When all the quality assessment results are within the range defined by the occasional threshold under dynamic operating conditions, the initial quality assessment step confirms the second type of data missing, and the data repair step uses a time-series prediction algorithm to repair the power battery usage data with individual data missing under dynamic operating conditions. When all the quality assessment results are within the range defined by the interval threshold under dynamic operating conditions, the initial quality assessment step confirms the third type of data missing, and the data repair step uses a matrix factorization algorithm or a deep learning algorithm to repair the power battery usage data with intermittent missing data under dynamic operating conditions.

[0023] Preferably, the power battery usage data includes four dimensions: voltage, current, temperature, and time, and the data repair module performs data repair on one or more of these dimensions simultaneously.

[0024] Preferably, in the initial data quality assessment step, the constant operating condition data threshold satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition intermittent threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition interval threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is less than 95% but higher than 80%.

[0025] The beneficial effects of this invention are as follows:

[0026] This invention proposes a power battery usage data repair system. Deployed within a cloud-based battery management system big data platform, the system includes a data cleaning module, a data quality initial assessment module, a data repair module, and a data quality re-assessment module. The initial assessment module is responsible for performing a preliminary assessment of the data quality of the cleaned power battery usage data, determining whether the data meets the repair threshold and identifying the type of data loss. This invention creatively proposes three parameters—data cleaning rate, sampling stability rate, and sampling continuity rate—to evaluate data quality. These three parameters correspond one-to-one with the completeness, authenticity, and continuity of data quality evaluation. The data cleaning rate is the ratio of the cleaned data length to the original data length, used to assess the stringency of the data cleaning rules and the quality of the original data, primarily to determine the possibility of data repair. The sampling stability rate is the proportion of cleaned data with a sampling interval consistent with the standard sampling interval, used to assess the authenticity of the data. This parameter is chosen because the parameter identification algorithm requires the data sampling interval to remain constant, thus requiring examination of the number of sampling points whose sampling interval matches the system-set sampling interval. The sampling continuity rate is the proportion of cleaned data with a sampling interval consistent with the standard sampling interval, and the number of consecutive points greater than or equal to the minimum length range, used to evaluate the continuity of the data. Then, the combined effect of these three parameters is used to determine whether the data is worth repairing and to determine the repair method. Finally, based on the confirmed data missing type, the data is passed to the corresponding unit in the data repair module. The selection of the three parameters—data cleaning rate, sampling stability rate, and sampling continuity rate—can accurately and comprehensively reflect data quality, laying a solid foundation for subsequent data repair work.

[0027] The data repair module includes a constant-condition data unit, a dynamic-condition intermittent data unit, and a dynamic-condition interval data unit, each designed to repair data with different degrees of missing information. For sensor data exhibiting monotonous changes under constant conditions, the constant-condition data unit of this invention uses spline interpolation to achieve fast and accurate data repair without requiring computationally intensive methods, achieving very high repair accuracy. For data missing in a few frames under dynamic conditions, this invention considers the lack of interruptions or abrupt changes in the overall time series continuity and stability of state changes. Therefore, the dynamic-condition intermittent data unit uses time series prediction algorithms for data repair, achieving excellent repair results. For dynamic-condition data where missing values ​​appear every few sampling points, a more powerful algorithm is needed. The dynamic-condition interval data unit of this invention uses matrix factorization algorithms or deep learning algorithms for data repair. This invention proposes a data repair method for data generated during the use of power batteries. It comprehensively considers three actual data loss scenarios and achieves the optimal balance between repair quality and computational cost. By comprehensively considering the different characteristics of the data, identifying different data loss types, and combining the applicable scenarios of the algorithm with the amount of computation, multiple repair methods are selected to repair the data. This method balances the accuracy of the results with the amount of computation, and greatly reduces the amount of computation while ensuring accuracy.

[0028] The data quality reassessment module, tailored to the actual functional requirements of the cloud platform, examines whether the repaired data can support functional implementation, accurately assesses the data repair function of the system, and further examines the performance of the repaired data in specific future applications, whether it can approximate the true value, and thus better suited for industrial production. The data quality initial assessment and reassessment calculation schemes developed in this invention conform to the evaluation indicators of data generation and usage scenarios, and can provide support for the quantification of data quality.

[0029] This invention also proposes a method for repairing power battery usage data. This repair method corresponds to the aforementioned power battery usage data repair system and can be understood as a method for implementing the aforementioned power battery usage data repair system. This method is deployed in a cloud-based battery management system big data platform and employs data cleaning, initial data quality assessment, data repair, and final data quality assessment steps. After data cleaning, an initial data quality assessment is performed on the cleaned usage data to determine whether the data meets the repair threshold and to identify the type of data loss. For different types of data loss, the corresponding repair method is selected based on the algorithm's applicable scenario and computational complexity, including sample... The algorithms are categorized into three types: spline interpolation, time-series prediction, and matrix factorization / deep learning. Specifically, the algorithm determines whether the missing data rate meets the lower limit of its repair capability. If the data is under constant operating conditions, spline interpolation is used for data repair. If the data is under dynamic operating conditions and has a small number of missing frames, time-series prediction is used for repair. If the data is under dynamic operating conditions and has gaps in data, matrix factorization and deep learning algorithms are used for repair. Finally, a comprehensive evaluation of the data quality is conducted to achieve reasonable, continuous, and high-quality output data. Based on this, model parameter identification and the construction of a digital twin model for the power battery can be completed. Attached Figure Description

[0030] Figure 1 This is a structural block diagram of the power battery data repair system of the present invention.

[0031] Figure 2 It is a schematic diagram of missing power battery data.

[0032] Figure 3 This is a flowchart of the power battery data repair method of the present invention.

[0033] Figure 4 This is a diagram illustrating different types of missing data under different dynamic operating conditions.

[0034] Figure 5 This is a schematic diagram of the neural network algorithm repair principle. Detailed Implementation

[0035] The present invention will now be described with reference to the accompanying drawings.

[0036] This invention provides a power battery usage data repair system, such as... Figure 1As shown, the system is deployed on a cloud-based battery management system big data platform, comprising a data cleaning module, a data quality initial assessment module, a data repair module, and a data quality reassessment module connected in sequence. The data cleaning module cleans the power battery usage data obtained from the cloud-based battery management system big data platform, removing default values, anomalies, and null values ​​according to data cleaning rules established by experts from battery manufacturers or vehicle manufacturers. The data quality initial assessment module evaluates the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. Based on the quality assessment results, it determines whether the power battery usage data meets the data repair threshold. For power battery usage data that meets the data repair threshold, it compares the quality assessment results with constant operating condition data thresholds, dynamic operating condition occasional thresholds, and dynamic operating condition interval thresholds to confirm the data missing type. Based on the data missing type, the power battery usage data is then passed to the corresponding unit of the data repair module. The data repair module includes a constant operating condition data unit, a dynamic operating condition occasional data unit, and a dynamic operating condition interval data unit. The system comprises several data units. The constant operating condition data unit employs a constant operating condition repair algorithm, which receives power battery usage data under constant operating conditions from the initial quality assessment module and repairs it using a spline interpolation method. The dynamic operating condition intermittent data unit employs a dynamic operating condition repair algorithm for data with few missing frames, which receives power battery usage data with individual missing data under dynamic operating conditions from the initial quality assessment module and repairs it using a time-series prediction algorithm. The dynamic operating condition interval data unit employs a dynamic operating condition repair algorithm for data with data intervals missing, which receives power battery usage data with intermittently missing data under dynamic operating conditions from the initial quality assessment module and repairs it using a matrix factorization algorithm or a deep learning algorithm. The data quality reassessment module is used to re-evaluate whether the data repaired by the data repair module can support functional implementation and to evaluate the performance of the repaired data in practical applications.The actual usage data after data cleaning contains a considerable number of missing, abnormal, or null values. To determine if the data missing rate meets the lower limit of the algorithm's repair capability, if the data is under constant operating conditions, spline interpolation is used for data repair. Spline interpolation completes the data near missing points by fitting spline curves. Under relatively stable operating conditions and with few missing data points, this method can obtain near-realistic nonlinear results with less computation. If the data is under dynamic operating conditions and has a small number of missing data frames, time-series prediction algorithms are used for repair. Time-series prediction algorithms identify data with trends, seasonality, and domain-specific characteristics. The algorithm first identifies the patterns (cycles) in the data, gradually extracting regular information until only noise remains, ideally white noise. This algorithm performs well in predicting short-term data and can simultaneously predict multivariate variables with moderate computational cost. If the data is dynamic and has missing intervals, matrix factorization and deep learning algorithms are used for data repair. Matrix factorization learns the overall characteristics of the time-series matrix, reducing the dimensionality of high-dimensional time-series data and using a low-rank approximation with time-series characteristics to fill in the missing data. This algorithm has lower computational complexity than deep learning algorithms and can handle larger datasets. Deep learning algorithms can employ neural network algorithms, such as GANs. As a generative network with self-supervised learning capabilities, G can be trained adversarially for a specific data format, enabling the generative model G to generate that type of data. This algorithm performs well in repairing large-scale missing data. Finally, a comprehensive evaluation of data quality is performed, ultimately resulting in reasonable, continuous, high-quality output data.

[0037] The power battery usage data can include four dimensions: voltage, current, temperature, and time. The data repair module can repair one or more of these dimensions simultaneously.

[0038] The cloud-based battery management system big data platform described above serves as the supporting node connecting the physical entity and the twin model. On one hand, this platform stores multi-source historical data generated during the operation of the physical entity through sensors, system backends, and other means. This data is stored in the cloud database of the digital twin platform and can be used for model training and standard definition. On the other hand, it utilizes the high computing power of the cloud server to continuously calculate various parameters or indicators of various models, obtaining information predictions, analysis, and evaluations for the current or future time period, realizing the value realization and sharing of information. The data cleaning module in this invention is built into the cloud platform, and cleaning rules are typically formulated by data experts from battery manufacturers and vehicle manufacturers based on business needs.

[0039] Data loss or corruption can occur during transmission when the vehicle-mounted data encounters various issues such as communication interruptions at the board or cloud level, poor sensor contact, errors in big data components, abnormal data cleansing by the system, remote system attacks, or data center failures. Different causes and patterns of failure will result in different forms of data loss, requiring different completion methods to match them.

[0040] After data is uploaded to the platform, it undergoes parameter identification and feature extraction, and is used for status estimation and fault diagnosis. Therefore, complete, accurate, and continuous data is crucial for the platform's functional modules. Because the data, after being uploaded to the cloud platform, needs to undergo a data cleaning module to remove outliers or standardize the format according to expert-defined rules, the length of the cleaned data is generally reduced by a certain percentage. For example, if an outlier or system default value is found in data such as voltage, current, temperature, and time uploaded at the same time, the entire line of data needs to be deleted to ensure data quality meets standards. The more data cleaned, the less usable data is available, and the more limitations subsequent functional modules face. High-quality data segments can be repaired using simpler and more efficient methods, while low-quality segments require methods with strong data repair capabilities and high fidelity.

[0041] The data quality initial assessment module will perform an overall evaluation of the uploaded data, mainly including three indicators: data cleaning rate, sampling stability rate, and sampling continuity rate. These three evaluation indicators correspond one-to-one with the completeness, authenticity, and continuity of data quality assessment. The data cleaning rate is used to evaluate the overall missing data situation and whether there is still a possibility of data repair. The sampling stability rate determines whether the algorithm needs to be downsampled or resampled to meet its capabilities. Generally, algorithms require data to be sampled at equal intervals. For the spline interpolation method of this invention, considering time series with unequal sampling intervals, curves can be directly fitted, and then the possible true values ​​at the missing values ​​can be obtained by looking up a table. For time series prediction, matrix factorization, and deep learning methods, the data can be downsampled and then resampled to equal intervals. In this case, the data may have some distortion, but the data quality can meet the algorithm's requirements before data completion is performed. The sampling continuity rate is used to find missing segments between large data blocks. Generally, if such missing data is not too long, it can be repaired using an algorithm based on the actual missing data situation. If there are too many missing segments in the middle, there is no need to repair them, and they can be directly regarded as two data blocks.

[0042] Data cleaning rate R data_clean The ratio of the length of the cleaned data to the length of the original data is used to evaluate the rigor of the data cleaning rules and the quality of the original data.

[0043]

[0044] In the formula, lendata_origin len represents the length of the original data. data_clean The length of the data after data cleaning.

[0045] For algorithms such as parameter identification, the data sampling interval needs to remain constant. Therefore, it's necessary to examine the number of sampling points whose sampling interval matches the system's set interval. For segments with good sampling stability, data restoration algorithms can achieve good results. For segments with abrupt or unequal sampling intervals, it's necessary to downsample or resample the data to equal intervals before data restoration, or to consider the unequal interval condition in the restoration algorithm. Sampling stability R sample_stability The proportion of cleaned data sampled at intervals consistent with the specified sampling interval is used to evaluate the stability of the data during transmission.

[0046]

[0047] In the formula, len time_interval＝sample_interval len is the data length of the standard sampling interval. data_clean This represents the length of the cleaned data.

[0048] Data segments that are continuous for a given sampling interval and meet a certain length requirement can capture more features; therefore, the continuity of the data needs to be evaluated. Data gaps between continuous segments can lead to the selection of appropriate data repair algorithms based on the specific circumstances. The sampling continuity rate R... sample_continue The proportion of cleaned data with sampling intervals consistent with the specified sampling intervals and a number of consecutive data points greater than or equal to the minimum length range is used to evaluate the stability of the data during transmission.

[0049]

[0050] In the formula, The number of consecutive data points equal to the sampling interval and with a length greater than or equal to the preset length `size`; `size` is the shortest consecutive data length based on actual needs; `len` data_clean This represents the length of the cleaned data.

[0051] Preferably, in the initial data quality assessment module, the constant operating condition data threshold satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, and the data sampling continuity rate for any dimension is greater than 95%. The dynamic operating condition intermittent threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is greater than 95%. The dynamic operating condition interval threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is less than 95% but higher than 80%. It should be noted that the above parameter value ranges are only preferred embodiments and are not the only limitations; the thresholds of this invention can be set to other ranges of parameter values.

[0052] The data repair module selects the appropriate repair method for different types of data loss, based on the applicable scenario of the algorithm and the amount of computation. These include three types: constant operating condition repair algorithm, dynamic operating condition repair algorithm for a small number of missing data frames, and dynamic operating condition repair algorithm for missing data intervals. The appropriate algorithm needs to be selected based on the actual data loss situation. Constant operating conditions refer to the battery's data changes exhibiting a monotonically increasing or decreasing trend during constant current / constant voltage charging and discharging. For occasional data gaps, simple spline interpolation methods can be used to fill in the data, offering higher accuracy compared to context data imputation and linear interpolation methods, while saving computational resources compared to complex algorithms. Dynamic operating conditions may result in missing frames of data, such as voltage, current, or temperature data at a particular moment not being uploaded to the cloud, or data being anomaly-filled with default values ​​and subsequently cleaned. In these cases, the overall data trend remains stable, allowing for data repair using time-series prediction algorithms such as ARIMA and LSTM. Intermittent data gaps may also occur under dynamic operating conditions, which are more frequent and irregular than missing frames. Therefore, traditional interpolation and prediction algorithms struggle to achieve ideal results. Some matrix factorization algorithms, however, can recover data based on characteristics obtained from low-rank matrices. This method can satisfy data repair requirements with lower accuracy needs. Furthermore, deep learning models can be trained based on historical data of the power battery system, and loss functions can be established to train the model, offering higher accuracy compared to matrix factorization algorithms.

[0053] The data quality re-evaluation module includes all the functions of the initial data quality evaluation module and evaluates the performance of the data in practical applications. For example, battery discharge data can be used to identify the internal resistance parameters of the battery model. Under the same identification algorithm, the difference between the internal resistance results calculated from different data and the true value can be used to judge the effectiveness of the data in practical applications. In addition, a segment of normal data can be artificially missing for data repair, and the mean square error between the repaired result and the true value, goodness of fit, and other parameters can be calculated as evaluation indicators.

[0054] This invention provides a power battery usage data repair system that can repair voltage, current, temperature, and other data uploaded by sensors in a cloud-based battery management system big data platform to meet the data requirements for building a digital twin model of the power battery. Power battery usage data, after data cleaning, contains a large number of missing, abnormal, or null values. The data quality initial assessment module can determine whether data loss can be repaired and what data repair methods should be used. The data repair module uses data repair algorithms to propose different algorithms that balance result accuracy and computational complexity for different situations, including constant and dynamic operating conditions and interval loss of a small number of missing frames.

[0055] Taking the data repair of a certain power battery pack as an example, the data generated by the vehicle-side battery management system (BMS), vehicle controller (VCU), engine controller (ECU), and other modules are transmitted to the T-Box controller via CAN bus, vehicle Ethernet, and other network forms. This controller provides the vehicle with vehicle networking services such as driving data collection, remote query and control, and fault monitoring. The data is then uploaded to the cloud-based battery management system big data platform. Based on the cloud-based digital twin model, functions and services such as state estimation, energy management, smart charging, fault diagnosis, and grid interconnection can be extended.

[0056] However, in actual data transmission, it may be due to, for example... Figure 2 The reasons shown may cause data loss in the power battery system. On the vehicle side, this may be due to board / cloud communication failures, sensor power failures, etc. On the cloud side, it may be due to big data system component errors, abnormal data being cleaned by the system, remote attacks on the system, data center failures, etc. All of these can cause data quality deterioration and make the data unusable.

[0057] To address this situation, the working principle diagram of the power battery data repair system of the present invention (i.e., the flowchart of the power battery data repair method) is as follows: Figure 3As shown, after obtaining the power battery system usage data uploaded by the T-Box device, the cloud platform cleans the data according to the data cleaning rules formulated by experts from battery manufacturers or vehicle manufacturers, removing default values, anomalies, and null values. The cleaned data then enters the data quality initial assessment module, which evaluates the cleaned data based on three indicators: data cleaning rate, sampling stability rate, and sampling continuity rate. If the data quality is below the lower limit that the algorithm can handle, the data cannot be repaired because even if repaired, it may still deviate significantly from the true value, making it difficult to meet actual calculation needs. Preferably, the data quality initial assessment indicators are: data cleaning rate ≥ 85%, sampling stability rate ≥ 80%, and sampling continuity rate ≥ 80%. The data quality initial assessment module then determines the missing data type based on the indicators, sending it to the data repair module for the corresponding repair algorithm. The data repair module selects the appropriate repair method for different missing data types, considering the algorithm's applicable scenario and computational complexity. When all quality assessment results are within the range defined by the constant operating condition data threshold, the initial data quality assessment module identifies it as the first type of data missing and then transmits the power battery usage data to the constant operating condition data unit of the data repair module. The constant operating condition data unit then performs data repair using spline interpolation, addressing the data as being under constant operating conditions (constant current charge and discharge). When all quality assessment results are within the range defined by the dynamic operating condition occasional threshold, the initial quality assessment module identifies it as the second type of data missing and then transmits the power battery usage data to the dynamic operating condition occasional data unit of the data repair module. The dynamic operating condition occasional data unit then performs data repair, addressing the data as being under dynamic operating conditions and containing a few missing data points. If data is missing, it is repaired using a time-series prediction algorithm. When all quality evaluation results are within the range defined by the dynamic operating condition interval threshold, the initial quality evaluation module confirms it as the third type of missing data and then transmits the power battery usage data to the dynamic operating condition interval data unit of the data repair module. The dynamic operating condition interval data unit repairs the missing data using matrix factorization algorithms and deep learning algorithms (GAN / Transformer, etc.) when the data is in dynamic operating conditions and the data interval is missing. Finally, the data quality is comprehensively evaluated, and reasonable, continuous, high-quality data can be output. Based on this, model parameter identification and the construction of the power battery digital twin model are completed.

[0058] Specifically, under constant operating conditions, sensor data exhibits a basically monotonically changing pattern. That is, under constant current and constant voltage charging and discharging conditions, the physical quantities of voltage, current, and temperature are all monotonically increasing or monotonically decreasing. Since the battery itself is a complex nonlinear component, constructing a spline interpolation function based on polynomial combinations can achieve fast and accurate data repair. Taking cubic spline curve interpolation as an example, the region [ab] covered by the existing data is first divided into n intervals [(x0,x1),(x1,x2),…,(x... n-1,x n There are a total of n+1 points, where the two endpoints x0 = a and x1 are... n =b. Cubic spline interpolation assumes that the curve in each subinterval is a cubic equation, and the equation must satisfy the following conditions:

[0059] (a) On each segmented subinterval, S(x) = S i (x) are all cubic equations;

[0060] (b) The interpolation condition is satisfied, S(x) i )=y i (i = 0, 1, ..., n);

[0061] (c) The curve is smooth, i.e. S(x), S′(x) and 5″(x) are continuous.

[0062] The cubic equation is then constructed in the form y = a i +b i x+c i x 2 +d i x 3 In this form, the curve can describe the trend and attributes of the neighborhood of missing data with an accuracy of over 95%.

[0063] For example Figure 4 In the dynamic scenario shown on the left, if several frames of data are missing, but there is no interruption or abrupt change in the overall continuity of the time series and the stability of state changes, then time series prediction algorithms can be used for data repair. Taking the classic ARIMA algorithm as an example, this algorithm is composed of a "hybrid" of an autoregressive model (AR model) and a moving average model (MA model). In time series analysis, the model that describes the relationship between a certain moment in the time series and the previous p moments is called an autoregressive model. The moving average model treats the time series as a linear combination of white noise sequences, assuming that a certain value can be predicted by the average of the previous N values. That is, the actual value can be obtained by adding the error to the predicted value of the previous value. Therefore, the actual value can be represented by the sum of multiple error values. The AR model establishes the relationship between the current value and historical values, the MA model calculates the accumulated error of the AR part, and ARMA is the sum of two dimensions.

[0064] For example Figure 4 The dynamic working condition shown on the right, where missing values ​​appear every few sampling points, requires an algorithm with stronger repair capabilities. In this case, matrix factorization-based algorithms and neural network algorithms can be selected.

[0065] The principle of matrix factorization algorithm is that high-dimensional long-term data have varying degrees of correlation. However, the information contained in the actual data is unprocessed, which occupies a lot of storage space and has great complexity. The feature matrix obtained by using singular value decomposition can store a large amount of data information with a very small amount of data. The theoretical basis of this method is complete. However, since the feature matrix is ​​the result of dimensionality reduction of the original data, it cannot retain all the key information (in actual running data, sometimes noisy data is also key information). Therefore, it is difficult to improve the accuracy to around 80%.

[0066] Another representative of neural network algorithms is GAN (Generative Adversarial Networks), such as... Figure 5 As shown, the core of this algorithm consists of two parts: a battery running data generator (G), which learns a mapping G(Z) that maps random noise Z to a time series X. The data G(Z) generated by the generator is called fake data; and a battery data discriminator (D), which learns a classifier that can distinguish between generated data and real data (real battery running training data). The input of the discriminator contains missing real data and intact fake data. GAN aims to have the generator generate fake data with the same distribution as the original data as possible, while also making it as difficult as possible for the discriminator to distinguish between fake and real data. The following objective function (i.e., the loss function) is optimized through a game-theoretic approach:

[0067]

[0068] Where p data (.) represents the true data distribution. min G This indicates that the generator error is minimized, meaning it generates data that is as realistic as possible; max D This represents maximizing the discriminator error, i.e., making it as difficult as possible for the discriminator to distinguish between generated and real sample distributions. In the formula, E represents the expectation, the first expectation... The first expectation is the expected value of logD(x) when all x are real data; the second expectation is... It is the expectation of log(1-D(G(z))) when all data is generated data. When the sample points of the real data and the generated data are fixed, the expectation equals the mean. By sharing the above loss function, the generator and discriminator achieve mutual adversarial interaction during training, minimizing... G max D The essence of V(D,G) is to minimize the loss. G Maximize Loss DAt the start of training, since the data generated by the generator differs greatly from the real data, D(x) should be close to 1 and D(G(z)) should be close to 0. After training is complete, both D(x) and D(G(z)) should be close to 0.5.

[0069] The advantage of this neural network algorithm is that it can use all the data generated during the battery's historical operation as the training set, and continuously compete and compromise between generated data and repair data to obtain repair data that best matches the distribution of usage habits. When the data quality is good, the accuracy can reach more than 90%.

[0070] This invention also proposes a method for repairing power battery usage data. This repair method corresponds to the aforementioned power battery usage data repair system and can be understood as a method for implementing the aforementioned power battery usage data repair system. This method is deployed in a cloud-based battery management system big data platform (hereinafter referred to as the cloud platform) and employs data cleaning, initial data quality assessment, data repair, and final data quality assessment steps. Its flowchart is shown below. Figure 4 As shown, after acquiring data from the cloud platform, data cleaning is performed on missing / abnormal / null values. Following data cleaning, an initial quality assessment is conducted on the cleaned data, calculating the missing data rate to determine whether data repair is necessary (to confirm if the data meets the repair threshold). If the data quality is too poor, repair is not possible. For repairable data (meeting the repair threshold), the missing data type is determined. For different missing data types, the appropriate repair method is selected based on the algorithm's applicable scenario and computational complexity. These methods include spline interpolation, time-series prediction algorithms, and matrix factorization / deep learning algorithms. Specifically, if the data is under constant operating conditions (constant current charging and discharging), spline interpolation is used for data repair; if the data is under dynamic operating conditions and some data is missing, time-series prediction algorithms are used for repair; if the data is under dynamic operating conditions and data is missing at intervals, matrix factorization algorithms and deep learning algorithms (GAN / Transformer, etc.) are used for repair. Finally, a comprehensive evaluation of the data quality is performed, ultimately resulting in reasonable, continuous, high-quality output data. Based on this, model parameter identification and the construction of a digital twin model for the power battery are completed.

[0071] The data cleaning step involves cleaning the power battery usage data obtained from the cloud-based battery management system's big data platform, according to data cleaning rules established by experts from battery manufacturers or vehicle manufacturers, to remove default values, anomalies, and null values. The initial data quality assessment step evaluates the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. The data cleaning rate is the ratio of the cleaned power battery data length to the original data length; the sampling stability rate is the proportion of cleaned power battery data with sampling intervals consistent with the standard sampling interval; and the sampling continuity rate is the proportion of cleaned power battery data with sampling intervals consistent with the standard sampling interval and a number of consecutive occurrences greater than or equal to a preset length threshold. Based on the quality assessment results, it is determined whether the power battery usage data meets the data repair lower limit. For power battery usage data that meets the data repair lower limit, the results are compared with constant operating condition data thresholds, dynamic operating condition occasional thresholds, and dynamic operating condition interval thresholds to confirm the data missing type. The data repair steps involve employing appropriate repair methods based on the different types of missing data, the applicable algorithm scenarios, and the computational complexity. For power battery usage data under constant operating conditions, spline interpolation is used for repair; for power battery usage data with occasional missing data under dynamic operating conditions, time-series prediction algorithms are used; and for power battery usage data with intermittent missing data under dynamic operating conditions, matrix factorization algorithms or deep learning algorithms are used. The data quality reassessment step is used to re-evaluate whether the repaired data can support functional implementation and to evaluate the performance of the repaired data in practical applications.

[0072] Furthermore, the quality assessment results in the initial data quality assessment step include the calculation results of the three parameters: data cleaning rate, sampling stability rate, and sampling continuity rate. When all the quality assessment results are within the range defined by the constant operating condition data threshold, the initial data quality assessment step confirms the first type of data missing, and the data repair step uses the spline interpolation method to repair the power battery usage data under constant operating conditions. When all the quality assessment results are within the range defined by the dynamic operating condition occasional threshold, the initial quality assessment step confirms the second type of data missing, and the data repair step uses a time-series prediction algorithm to repair the power battery usage data with individual data missing under dynamic operating conditions. When all the quality assessment results are within the range defined by the dynamic operating condition interval threshold, the initial quality assessment step confirms the third type of data missing, and the data repair step uses a matrix factorization algorithm or a deep learning algorithm to repair the power battery usage data with intermittent missing data under dynamic operating conditions.

[0073] Furthermore, the power battery usage data includes four dimensions: voltage, current, temperature, and time. The data repair module performs data repair on one or more of these dimensions simultaneously.

[0074] Furthermore, in the initial data quality assessment step, the constant operating condition data threshold satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition intermittent threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition interval threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is less than 95% but higher than 80%.

[0075] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention patent.

Claims

1. A power battery usage data repair system, characterized in that, This includes a data cleaning module, a data quality initial assessment module, a data repair module, and a data quality reassessment module, all deployed on a cloud-based battery management system big data platform and connected sequentially. The data cleaning module is used to clean the power battery usage data obtained from the cloud-based battery management system big data platform, according to the data cleaning rules formulated by experts from battery manufacturers or vehicle manufacturers, to remove default values, abnormal values ​​and null values ​​from the data. The data quality initial assessment module is used to evaluate the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. The data cleaning rate is the ratio of the length of the cleaned power battery data to the original data length. The sampling stability rate is the proportion of cleaned power battery data whose sampling interval matches the standard sampling interval. The sampling continuity rate is the proportion of cleaned power battery data whose sampling interval matches the standard sampling interval and whose consecutive counts are greater than or equal to a preset length threshold. Based on the quality assessment results, the module determines whether the power battery usage data meets the data repair lower limit. For power battery usage data that meets the data repair lower limit, the module compares the data with constant operating condition data thresholds, dynamic operating condition occasional thresholds, and dynamic operating condition interval thresholds to confirm the data missing type. Based on the data missing type, the module then... The transmission of power battery usage data to the corresponding unit of the data repair module specifically includes: the results of the quality assessment include the calculation results of three parameters: data cleaning rate, sampling stability rate, and sampling continuity rate. When all the results of the quality assessment are within the range defined by the constant operating condition data threshold, the data quality initial assessment module confirms it as the first type of data missing and then transmits the power battery usage data to the constant operating condition data unit of the data repair module; when all the results of the quality assessment are within the range defined by the dynamic operating condition occasional threshold, the quality initial assessment module confirms it as the second type of data missing and then transmits the power battery usage data to the dynamic operating condition occasional data unit of the data repair module; when all the results of the quality assessment are within the range defined by the dynamic operating condition interval threshold, the quality initial assessment module confirms it as the third type of data missing and then transmits the power battery usage data to the dynamic operating condition interval data unit of the data repair module. The data repair module includes a constant operating condition data unit, a dynamic operating condition intermittent data unit, and a dynamic operating condition interval data unit. The constant operating condition data unit receives power battery usage data under constant operating conditions transmitted by the initial quality assessment module and repairs it using a spline interpolation method. The dynamic operating condition intermittent data unit receives power battery usage data with individual missing data under dynamic operating conditions transmitted by the initial quality assessment module and repairs it using a time-series prediction algorithm. The dynamic operating condition interval data unit receives power battery usage data with intermittent missing data under dynamic operating conditions transmitted by the initial quality assessment module and repairs the data using a matrix factorization algorithm or a deep learning algorithm. The data quality reassessment module is used to re-evaluate whether the data repaired by the data repair module can support the function implementation, and to evaluate the performance of the repaired data in practical applications.

2. The power battery usage data repair system according to claim 1, characterized in that, The power battery usage data includes four dimensions: voltage, current, temperature, and time. The data repair module performs data repair on one or more of these dimensions simultaneously.

3. The power battery usage data repair system according to claim 2, characterized in that, In the initial data quality assessment module, the constant operating condition data threshold satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition occasional threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition interval threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is less than 95% but higher than 80%.

4. The power battery usage data repair system according to claim 1, characterized in that, The initial data quality assessment module calculates the data missing rate based on the quality assessment results, and then determines whether the power battery usage data meets the lower limit for data repair.

5. The power battery usage data repair system according to claim 1, characterized in that, The time-series prediction algorithms in the data repair module include the ARIMA algorithm and the LSTM algorithm; the deep learning algorithms include neural network algorithms.

6. A method for repairing data from a power battery, characterized in that, Includes the following steps: The data cleaning step is used to clean the power battery usage data obtained from the cloud-based battery management system big data platform, according to the data cleaning rules formulated by experts from battery manufacturers or vehicle manufacturers, to remove defaults, anomalies and null values ​​from the data. The initial data quality assessment step evaluates the cleaned power battery data using parameters such as data cleaning rate, sampling stability rate, and sampling continuity rate. The data cleaning rate is the ratio of the cleaned power battery data length to the original data length. The sampling stability rate is the proportion of cleaned power battery data with sampling intervals consistent with the standard sampling interval. The sampling continuity rate is the proportion of cleaned power battery data with sampling intervals consistent with the standard sampling interval and a consecutive number greater than or equal to a preset length threshold. Based on the quality assessment results, it is determined whether the power battery usage data meets the data repair lower limit. For power battery usage data that meets the data repair lower limit, the results are compared with constant operating condition data thresholds, dynamic operating condition occasional thresholds, and dynamic operating condition interval thresholds to confirm data deficiencies. The data loss types specifically include: the results of the quality assessment include the calculation results of the three parameters: data cleaning rate, sampling stability rate, and sampling continuity rate. When all the results of the quality assessment are within the range defined by the constant operating condition data threshold, the data quality initial assessment module confirms it as the first type of data loss and then transmits the power battery usage data to the constant operating condition data unit of the data repair module; when all the results of the quality assessment are within the range defined by the dynamic operating condition occasional threshold, the quality initial assessment module confirms it as the second type of data loss and then transmits the power battery usage data to the dynamic operating condition occasional data unit of the data repair module; when all the results of the quality assessment are within the range defined by the dynamic operating condition interval threshold, the quality initial assessment module confirms it as the third type of data loss and then transmits the power battery usage data to the dynamic operating condition interval data unit of the data repair module. The data repair steps involve adopting corresponding repair methods based on different types of missing data, the applicable scenarios of the algorithms, and the amount of computation. For power battery usage data under constant operating conditions, the spline interpolation method is used for repair; for power battery usage data with individual missing data under dynamic operating conditions, the time series prediction algorithm is used for repair; and for power battery usage data with intermittent missing data under dynamic operating conditions, the matrix factorization algorithm or deep learning algorithm is used for data repair. The data quality reassessment step is used to re-evaluate whether the repaired data can support the implementation of functions, and to evaluate the performance of the repaired data in practical applications.

7. The method for repairing power battery usage data according to claim 6, characterized in that, The power battery usage data includes four dimensions: voltage, current, temperature, and time. The data repair module performs data repair on one or more of these dimensions simultaneously.

8. The method for repairing power battery usage data according to claim 7, characterized in that, In the initial data quality assessment step, the constant operating condition data threshold satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition occasional threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is greater than 95%; and / or, the dynamic operating condition interval threshold simultaneously satisfies the following conditions: the data cleaning rate is greater than the limit of the power battery data repair system algorithm capability, the sampling stability rate is higher than 80%, and the data sampling continuity rate for any dimension is less than 95% but higher than 80%.

Images