Electric vehicle power battery dynamic threshold early warning method based on double sliding window baseline

Abstract

The application discloses a kind of based on double sliding window baseline electric vehicle power battery dynamic threshold early warning method, this method receives and washes vehicle operation data, effective sample is divided into monitoring scene according to charge state, vehicle speed interval, total current direction and amplitude variable;Long-period baseline window and short-period correction window are simultaneously constructed in the same monitoring scene, long, short cycle statistics are calculated respectively, and final baseline is obtained by correction coefficient fusion;According to the early warning direction of target parameter, generate upper and lower limit dynamic threshold or double-sided dynamic threshold;The over-limit proportion of current monitoring segment, continuous over-limit duration, maximum deviation and segment quantile value are judged and the early warning grade is output;The segment that is reviewed as false alarm is backfilled to normal sample pool, and the segment of real anomaly is isolated from baseline calculation.The application can adapt to long-term aging drift and short-term working condition change simultaneously, reduce false alarm and omission, and is suitable for multi-index power battery online early warning.

Description

Technical Field

[0001] This invention relates to the fields of electric vehicle power battery system operation safety monitoring, cloud data analysis and fault early warning technology, specifically to a dynamic threshold early warning method for electric vehicle power batteries based on joint modeling of long-cycle baseline window and short-cycle correction window. Background Technology

[0002] During actual operation, electric vehicles continuously generate multi-dimensional operational data conforming to the vehicle-to-everything (V2X) standard format. This data includes information such as vehicle status, charging status, vehicle speed, total voltage, total current, state of charge (SOC), individual cell voltage, probe temperature, insulation resistance, and fault alarm indicators. This data reflects the changes in the safety boundaries of the power battery system under different operating environments, aging stages, and load conditions, and is an important foundation for battery system fault early warning.

[0003] Existing power battery early warning methods typically use national standard fixed thresholds, empirical thresholds, or single statistical window thresholds for judgment. While these methods are simple to implement, they have significant shortcomings: First, the normal fluctuation range varies depending on the vehicle, battery type, service life, and usage scenario, making it difficult for fixed thresholds to balance sensitivity and false alarm rate; second, a single fixed window cannot simultaneously describe long-term aging drift and short-term operating condition fluctuations, easily leading to problems such as threshold lag, thresholds that are too tight, or thresholds that are too loose.

[0004] In engineering applications, recent high-temperature operation, frequent fast charging, mountain driving, or prolonged low SOC parking can cause short-term deviations in target parameters. Meanwhile, battery system capacity decay, increased internal resistance, and changes in thermal management performance can lead to long-term baseline drift. If warning thresholds are built solely based on short-cycle data, they are easily affected by occasional noise and temporary operating conditions; if they are built solely based on long-cycle data, they cannot adapt to recent operating characteristics in a timely manner, causing early signs of soft faults to be masked.

[0005] Furthermore, existing methods often simply remove segments that have triggered warnings, lacking mechanisms for false alarm backtracking and threshold self-calibration. When some warnings are confirmed as normal fluctuations after review, if they are not properly included in the next round of baseline updates, the system will continue to generate repeated false alarms; while if real anomalies are included in the benchmark dataset, they will pollute the normal baseline, further reducing the accuracy of warnings. Therefore, existing technologies still struggle to meet the requirements of high stability and sustainable deployment in complex operating conditions, multi-vehicle models, and multi-indicator online warning scenarios for power batteries. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a dynamic threshold early warning method based on a dual sliding window baseline. By setting a long-period baseline window and a short-period correction window, the normal distribution of the target parameters is continuously modeled under the same operating conditions, enabling the early warning threshold to simultaneously adapt to long-term aging drift and short-term operating state changes.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows: This invention proposes a dynamic threshold early warning method for electric vehicle power batteries based on a dual sliding window baseline, comprising the following steps: S1: Receive and store real-time operating data of electric vehicles; S2: Perform quality control on the data of the target parameters to be monitored to obtain an effective sample set; S3: Divide the effective sample set into multiple monitoring scenarios based on vehicle operating state variables, and construct baselines and thresholds within the same or nearly the same monitoring scenarios; S4: Construct a long-cycle baseline window and a short-cycle correction window with the threshold calculation day as the endpoint, and extract normal and valid samples of the corresponding monitoring scenario within the long-cycle baseline window and the short-cycle correction window; S5: Calculate the long-cycle statistics of the long-cycle baseline window and the short-cycle statistics of the short-cycle correction window respectively, and fuse the long-cycle and short-cycle statistics through the correction coefficient to obtain the final baseline of the target parameter in the corresponding monitoring scenario. S6: Generate an upper limit dynamic threshold, a lower limit dynamic threshold, or a two-sided dynamic threshold based on the warning direction of the target parameters; S7: Calculate the over-limit ratio, continuous over-limit duration, maximum deviation and segment percentile for the current monitoring segment, and output the warning label, warning level and warning report when the preset warning conditions are met; S8: Review and provide feedback on the warning segments. False alarm segments are marked as normal samples and included in the next round of threshold calculation. Real abnormal segments are isolated from the normal benchmark dataset.

[0008] Furthermore, in step S1, the data platform receives and stores real-time operating data of electric vehicles in accordance with the vehicle operation data format requirements of GB / T 32960.3. The real-time operating data includes acquisition time, vehicle identification number, charging status, vehicle speed, total voltage, total current, SOC, individual cell voltage, probe temperature, insulation resistance, fault alarm flags and their derived statistics.

[0009] Furthermore, the target parameters to be monitored in step S2 include battery cell voltage, battery temperature, insulation resistance, SOC, drive motor temperature, and electronic control system temperature; the target parameters are classified into upper limit indicators, lower limit indicators, and two-sided indicators according to the warning direction. The quality control includes time alignment, deduplication of duplicate frames, null value processing, unit unification, and physical boundary verification, and removes data that is from sensors that are offline, has packet loss, has parsing errors, has fault data defined by national standards, and data that obviously does not conform to physical laws.

[0010] Furthermore, the vehicle operating state variables mentioned in step S3 include charging state, vehicle speed range, total current direction and amplitude, SOC range, ambient temperature range, and operating cycle number. Each complete driving segment, charging segment, or stationary segment is treated as a statistical unit; baselines and thresholds are established only within the same or nearly identical monitoring scenarios.

[0011] Furthermore, the long-cycle baseline window mentioned in step S4 is used to describe the stable baseline under individual vehicle, battery type, and aging stage; the short-cycle correction window is used to capture short-term deviations caused by recent operating environment, vehicle usage intensity, and operating condition distribution. Both the long-period baseline window and the short-period correction window are constructed by selecting a set length of time forward from the threshold calculation day as the endpoint, thus forming a long-period and short-period sliding window.

[0012] Furthermore, the long-term statistics mentioned in step S5 include the mean, standard deviation, and quantiles of the long-term window data; the short-term statistics are the mean or quantiles of the short-term window data. The expression for the final baseline is: in, This is the corrected final baseline mean. This is a short-cycle correction factor. This represents the mean of short-period window data. This represents the mean of data within a long-term window.

[0013] Furthermore, for upper limit indicators, a dynamic upper limit threshold is adopted, and the upper limit indicators include temperature, charging voltage, voltage difference, and excessively high SOC; For lower limit indicators, a lower limit dynamic threshold is used, and the lower limit indicators include discharge voltage and insulation resistance; For two-sided indicators, both upper and lower dynamic thresholds are generated simultaneously. The upper and lower dynamic thresholds are calculated using the baseline mean, confidence coefficient, and standard deviation of long-period window data, or obtained using the quantiles of the fragment quantile set.

[0014] Furthermore, the specific formulas for calculating the upper and lower dynamic thresholds using the baseline mean, confidence coefficient, and standard deviation of long-term window data are as follows: in, The upper limit is a dynamic threshold. The lower limit is the dynamic threshold. Here is the confidence coefficient. The standard deviation of long-period window data; The specific method for obtaining the upper and lower dynamic thresholds from the quantiles of the fragment quantile value set is as follows: The data is divided into multiple complete running segments according to the monitoring scenario, with each segment serving as a statistical unit. For upper limit indicators, the 95th percentile of the sampling points within a segment is taken as the upper quantile value of the segment. For lower limit indicators, the 5th percentile of the sampling points within a segment is taken as the lower quantile value of the segment. Then, a second statistical analysis is performed on the set of quantile values ​​of all normal segments under the same monitoring scenario to construct the initial dynamic threshold at the segment level.

[0015] Furthermore, the preset warning conditions mentioned in step S7 include: the proportion of sampling points exceeding the dynamic threshold in the current monitoring segment exceeds the proportion threshold, or the duration of continuous exceeding the limit exceeds the duration threshold, or the segment quantile statistical value exceeds the dynamic threshold. The warning levels are divided into three levels based on the degree of deviation: Level 1 Warning, Level 2 Attention, and Level 3 Risk. The output warning report includes vehicle identification number, target parameter name, monitoring scenario, warning start time, warning end time, dynamic threshold, segment statistics, over-limit ratio, warning level, and recommended review items.

[0016] Furthermore, a weekly rolling update mechanism is adopted, in which the long-term threshold is recalculated weekly using normal working condition data within the past long-term window, and the short-term correction within the most recent short-term window is superimposed to obtain the final dynamic threshold for the week. When a warning segment is verified as a false alarm by manual intervention, maintenance, or subsequent data, the warning segment is fed back into the normal sample pool to calibrate the threshold tightness. When a warning segment is verified as a real anomaly, the warning segment is removed from the baseline calculation to avoid abnormal segments raising or lowering the threshold.

[0017] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention employs a combined modeling approach using a long-cycle baseline window and a short-cycle correction window, enabling dynamic thresholds to possess both long-term stability and short-term adaptability. Modeling under the same operating conditions avoids threshold shifts caused by data mixing from different scenarios such as charging, driving, stationary, high load, and low load. Combining mean-standard deviation thresholds with quantile thresholds for operational segments reduces the impact of instantaneous noise. A false alarm backtracking and real anomaly isolation mechanism gradually reduces repeated false alarms and prevents anomalies from contaminating the normal baseline. Through a unified upper limit, lower limit, and two-sided threshold framework, it can be extended to multiple indicators such as voltage, temperature, insulation, voltage consistency, SOC anomalies, motor temperature, and electronic control temperature. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0019] Figure 1 This is a flowchart of the dynamic threshold early warning method based on a dual sliding window baseline in this invention; Figure 2 This is a schematic diagram illustrating the relationship between the long-period baseline window and the short-period correction window in this invention; Figure 3 This is a schematic diagram of effective data filtering under the same working conditions in this invention; Figure 4 This is a schematic diagram of the fusion of long-period baseline and short-period correction in this invention; Figure 5 This is an example diagram of the dynamic threshold early warning for upper limit type indicators in this invention; Figure 6 This is an example diagram of the dynamic threshold early warning for the lower limit indicator in this invention; Figure 7 This is a schematic diagram illustrating the construction of the initial threshold based on the quantile of the running segment in this invention; Figure 8 This is a schematic diagram of the weekly rolling dynamic threshold update in this invention; Figure 9 This is a schematic diagram of the false alarm backtracking and baseline self-calibration mechanism in this invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1As shown, this embodiment provides a dynamic threshold early warning method for electric vehicle power batteries based on a dual sliding window baseline, including the following steps: S1, the data platform receives and stores real-time operating data of electric vehicles in accordance with the vehicle operation data format requirements of GB / T 32960.3; among which, the real-time operating data includes the acquisition time, vehicle identification number, charging status, vehicle speed, total voltage, total current, SOC, individual cell voltage, probe temperature, insulation resistance, fault alarm signs and their derived statistics.

[0022] S2 performs time alignment, duplicate frame deduplication, null value processing, unit unification, and physical boundary verification on the target parameter data to be monitored, and removes data from sensors that are offline, have packet loss, have parsing errors, have faults defined by national standards, and data that obviously does not conform to physical laws; among them, the target parameters include battery cell voltage, battery temperature, insulation resistance, SOC, drive motor temperature, and electronic control system temperature, etc.; let the target parameter to be monitored be... The sequence of observations arranged in chronological order is denoted as Target parameters can be categorized into upper limit indicators, lower limit indicators, and two-sided indicators based on the monitored object. After quality control, a valid sample set is obtained. .

[0023] S3, based on variables such as charging status, vehicle speed range, total current direction and amplitude, SOC range, ambient temperature range, and operating cycle number, divides the effective samples into multiple monitoring scenarios c; each complete driving segment, charging segment, or stationary segment is treated as a statistical unit. Baselines and thresholds are established only within the same or nearly identical monitoring scenarios to reduce threshold shifts caused by the mixing of different operating conditions.

[0024] S4, with the threshold calculation day as the endpoint, selects a forward length of... Long-period sliding window In the window Normal valid samples corresponding to scenario c are extracted internally. The long-term window is used to describe the stable baseline under individual vehicle, battery type, and aging stage.

[0025] S5, with the threshold calculation day as the endpoint, selects a forward length of... short-period sliding window In the window Normal valid samples corresponding to scenario c are extracted internally. The short-period window is used to capture short-term offsets caused by recent operating environment, vehicle usage intensity, and operating condition distribution.

[0026] In this embodiment, as Figure 2As shown, a long-term baseline window is constructed simultaneously on each threshold calculation day. and short-cycle correction window Long-cycle windows cover a longer time range and are used to provide a stable baseline for individual vehicles; short-cycle windows cover recent operating data and are used to correct for recent operating condition deviations.

[0027] in, At the current threshold calculation time, For a long-term window, 90 days is preferred; For a short period window, 7 days is preferred.

[0028] To avoid interference from differences in parameter distribution under different operating conditions on threshold calculation, this embodiment extracts only valid samples within the same monitoring scenario. For example... Figure 3 As shown, after removing fault data, invalid sensor data, and obvious physical anomalies during the quality control process, the remaining samples are used as candidate samples for the dynamic baseline.

[0029] in, The sample at time t belongs to monitoring scenario c, and flag(t)=normal indicates that the sample has not been marked as invalid, faulty, or a real anomaly.

[0030] S6 calculates the mean, standard deviation, and quantiles of the long-term window data and the mean or quantiles of the short-term window data respectively, and then merges the long-term and short-term statistics through the correction coefficient to obtain the final baseline.

[0031] Specifically, within a long-term window, the mean and standard deviation of the target parameters are calculated to characterize the long-term normal fluctuation range of the vehicle in the current scenario.

[0032] in, The mean of the long-term window data. The standard deviation of long-period window data. This represents the number of valid samples within a long-term window. To reduce the impact of extreme values, quantiles such as P1, P5, P95, and P99 can be calculated simultaneously, and quantile filtering can be applied to the window data.

[0033] Within a short-period window, calculate the recent mean of the target parameter. Or recent quantile statistics, used to reflect parameter shifts caused by recent operating conditions. The long- and short-period baseline fusion process is as follows: Figure 4 As shown.

[0034] in, This is the corrected final baseline mean. This is the short-period correction coefficient, with a preferred value of 0.3. When the short-period sample size is insufficient or fluctuations are abnormally large, the value can be reduced. When recent operating conditions are stable and deviations from long-term baselines persist, improvements can be made. .

[0035] S7 generates an upper limit dynamic threshold for upper limit type indicators, a lower limit dynamic threshold for lower limit type indicators, and generates both upper and lower limit thresholds for two-sided type indicators. The threshold can be calculated from the baseline mean, confidence coefficient, and standard deviation of long-term window data, or it can be obtained from the quantiles of the fragment quantile set.

[0036] In this embodiment, based on the warning direction of the target parameter, the indicators are divided into upper limit type, lower limit type, and two-sided type. For upper limit type indicators such as temperature, charging voltage, voltage difference, and excessively high SOC, an upper limit dynamic threshold is used; for lower limit type indicators such as discharge voltage and insulation resistance, a lower limit dynamic threshold is used.

[0037] The specific formulas for calculating the upper and lower dynamic thresholds using the baseline mean, confidence coefficient, and standard deviation of long-period window data are as follows: in, The upper limit is a dynamic threshold. The lower limit is the dynamic threshold. is the confidence coefficient. The value can be set based on battery type, indicator type, historical false alarm rate, and safety requirements; under normal circumstances, a value of 2 to 3 is acceptable. For safety-sensitive indicators, a lower value can be used. To improve sensitivity; for indicators with large fluctuations, the value can be appropriately increased. This value is used to reduce false alarms.

[0038] The specific method for obtaining the upper and lower dynamic thresholds from the quantiles of the fragment quantile value set is as follows: The data is divided into multiple complete operational segments according to the monitoring scenario. Each segment is treated as a statistical unit. For upper limit indicators, the 95th percentile of the sampling points within a segment is used; for lower limit indicators, the 5th percentile of the sampling points within a segment is used. A second statistical analysis is then performed on the set of quantile values ​​for all normal segments in the same scene to construct the initial dynamic threshold at the segment level, such as... Figure 7 As shown.

[0039] in, These represent the values ​​of the four degree quantiles: 0.05, 0.10, 0.90, and 0.95. For the first Data from each running segment, For the first The 95th percentile of each running segment For the first The 5th percentile of each running segment, The 90th quantile of the set of 95th quantiles for the upper limit index across all segments. It is the 10th percentile of the set of 5th percentiles of the lower limit index in all segments.

[0040] This quantile mechanism can reduce the impact of instantaneous noise at a single sampling point on threshold calculation, making the threshold more aligned with the business objective of segment-level fault early warning.

[0041] S8, for each currently monitored segment Calculate the proportion of sampling points where the target parameter exceeds the threshold, the duration of continuous over-limit, the maximum deviation, and the segment quantile value; when the over-limit proportion, duration, maximum deviation, or segment quantile value meets the preset warning conditions, output the warning label and warning level.

[0042] The preset warning conditions include: the proportion of sampling points exceeding the dynamic threshold in the current monitoring segment exceeds the preset proportion threshold, or the duration of continuous exceeding the limit exceeds the preset duration threshold, or the segment quantile statistical value exceeds the preset dynamic threshold, or the maximum deviation exceeds the preset deviation threshold.

[0043] In this embodiment, the current monitoring segment When making an early warning judgment, the over-limit ratio, continuous over-limit time, maximum deviation degree and segment quantile value are calculated according to the target parameter type. When the proportion of sampling points exceeding the dynamic threshold in the current monitoring segment exceeds the preset ratio threshold, or the continuous over-limit duration exceeds the preset duration threshold, or the segment quantile statistical value exceeds the preset dynamic threshold, the segment is determined to trigger an early warning. The early warning level can be divided into Level 1 alert, Level 2 attention and Level 3 risk according to the maximum deviation degree. Figure 5 and Figure 6 Examples of dynamic threshold warnings for upper limit indicators and lower limit indicators are given respectively.

[0044] S9 outputs a warning report, which includes the vehicle identification number, target parameter name, monitoring scenario, warning start time, warning end time, dynamic threshold, segment statistics, over-limit ratio, warning level, and recommended review items.

[0045] S10: If a warning segment is verified as a false alarm by manual intervention, maintenance, or subsequent data, the segment is marked as a normal sample and included in the next round of threshold calculation; if it is verified as a real anomaly, it is isolated from the normal benchmark dataset to avoid baseline contamination; the threshold tightness is continuously corrected through a weekly rolling update mechanism.

[0046] like Figure 9 As shown, in this embodiment, the warning segment is fed back after being reviewed by the business team. If the review result is a false alarm, it means that the segment belongs to the normal fluctuations acceptable under the current vehicle and current scenario, and it is included as a normal sample in the next threshold update. If the review result is a real anomaly, it is isolated from the baseline sample pool to prevent abnormal segments from raising or lowering the threshold. Through the above feedback mechanism, this solution can gradually reduce repeated false alarms in long-term operation without sacrificing sensitivity to real anomalies.

[0047] This embodiment employs a weekly rolling update mechanism. Each week, the long-cycle threshold is recalculated using data from the past 90 days of normal operating conditions, and a short-cycle correction from the most recent 7 days is added to obtain the final dynamic threshold for that week. The update process is as follows: Figure 8 As shown.

[0048] in, For the first Zhou ultimately adopted the dynamic threshold. This is a long-term threshold calculated based on normal samples from the past 90 days. The short-cycle threshold is calculated based on normal samples under the same operating conditions over the past 7 days. The short-cycle weighting coefficient is preferably set to 0.3.

[0049] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0050] An electric vehicle platform collected historical operating data from a batch of vehicles, with a data sampling interval of 10 seconds per frame. The data included fields such as charging status, vehicle speed, total voltage, total current, SOC, individual cell voltage, temperature, insulation resistance, and fault alarm indicators. The highest battery temperature, the lowest individual cell voltage, and the individual cell voltage range were selected as target parameters, corresponding to upper limit type, lower limit type, and dual-sided type monitoring indicators, respectively.

[0051] First, the platform performs quality control on the raw data, deleting data with duplicate times, offline sensors, null values, data that has already triggered alarms according to national standards, and data with obvious physical anomalies. Then, it divides the monitoring scenarios according to charging status, vehicle speed range, total current direction, and SOC range, such as "driving discharge - SOC 40% to 80% - medium current range" and "parking charging - SOC 20% to 90% - fast charging range," etc.

[0052] For the highest battery temperature index of a certain vehicle under the scenario of "driving discharge - SOC 40% to 80%", the system uses normal samples from the 90 days prior to the threshold calculation date as a long-term window to calculate the long-term average. and standard deviation Then, take the most recent 7 days of normal operating conditions as a short-term window and calculate the recent mean. .according to Obtain the corrected baseline, and then according to The dynamic threshold for the upper temperature limit is obtained.

[0053] For the lowest single-cell voltage index of a certain vehicle, the system calculates the 5th percentile for each complete driving segment. Then, for all normal fragments within 90 days The 10th percentile is calculated as the initial lower threshold. If the proportion of samples in the current segment that are below the lower dynamic threshold exceeds a preset proportion and the continuous duration exceeds a preset duration, a minimum single-unit voltage warning is output.

[0054] During a recent weekly update, the platform detected a slight upward shift in the 7-day temperature baseline compared to the 90-day baseline, without any national standard fault alarms occurring during the same period. Further investigation confirmed that this change was caused by increased ambient temperature and increased operating load. The system weighted this short-term offset. β The values ​​are superimposed to the final threshold, which is then moderately increased to reflect recent normal operating conditions, thus avoiding repeated false alarms. If another vehicle experiences a persistent minimum single-cell voltage below the dynamic lower limit threshold, and this is subsequently determined to be a genuine anomaly, then that segment is marked as a genuine anomaly and excluded from the next round of baseline sample set.

[0055] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines, characterized in that, Includes the following steps: S1: Receive and store real-time operating data of electric vehicles; S2: Perform quality control on the data of the target parameters to be monitored to obtain an effective sample set; S3: Divide the effective sample set into multiple monitoring scenarios based on vehicle operating state variables, and construct baselines and thresholds within the same or nearly the same monitoring scenarios; S4: Construct a long-cycle baseline window and a short-cycle correction window with the threshold calculation day as the endpoint, and extract normal and valid samples of the corresponding monitoring scenario within the long-cycle baseline window and the short-cycle correction window; S5: Calculate the long-cycle statistics of the long-cycle baseline window and the short-cycle statistics of the short-cycle correction window respectively, and fuse the long-cycle and short-cycle statistics through the correction coefficient to obtain the final baseline of the target parameter in the corresponding monitoring scenario. S6: Generate an upper limit dynamic threshold, a lower limit dynamic threshold, or a two-sided dynamic threshold based on the warning direction of the target parameters; S7: Calculate the over-limit ratio, continuous over-limit duration, maximum deviation and segment percentile for the current monitoring segment, and output the warning label, warning level and warning report when the preset warning conditions are met; S8: Review and provide feedback on the warning segments. False alarm segments are marked as normal samples and included in the next round of threshold calculation. Real abnormal segments are isolated from the normal benchmark dataset.

2. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, In step S1, the data platform receives and stores real-time operating data of electric vehicles in accordance with the vehicle operation data format requirements of GB / T 32960.

3. The real-time operating data includes acquisition time, vehicle identification number, charging status, vehicle speed, total voltage, total current, SOC, individual cell voltage, probe temperature, insulation resistance, fault alarm flags and their derived statistics.

3. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, The target parameters to be monitored in step S2 include battery cell voltage, battery temperature, insulation resistance, SOC, drive motor temperature, and electronic control system temperature; the target parameters are classified into upper limit indicators, lower limit indicators, and two-sided indicators according to the warning direction. The quality control includes time alignment, deduplication of duplicate frames, null value processing, unit unification, and physical boundary verification, and removes data that is from sensors that are offline, has packet loss, has parsing errors, has fault data defined by national standards, and data that obviously does not conform to physical laws.

4. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, The vehicle operating state variables mentioned in step S3 include charging state, vehicle speed range, total current direction and amplitude, SOC range, ambient temperature range, and operating cycle number. Each complete driving segment, charging segment, or stationary segment is treated as a statistical unit; baselines and thresholds are established only within the same or nearly identical monitoring scenarios.

5. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, The long-cycle baseline window mentioned in step S4 is used to describe the stable baseline under individual vehicle, battery type and aging stage; the short-cycle correction window is used to capture short-term deviations caused by recent operating environment, vehicle usage intensity and operating condition distribution. Both the long-period baseline window and the short-period correction window are constructed by selecting a set length of time forward from the threshold calculation day as the endpoint, thus forming a long-period and short-period sliding window.

6. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, The long-term statistics mentioned in step S5 include the mean, standard deviation, and quantiles of the long-term window data; the short-term statistics are the mean or quantiles of the short-term window data. The expression for the final baseline is: in, This is the corrected final baseline mean. This is a short-cycle correction factor. This represents the mean of short-period window data. This represents the mean of data within a long-term window.

7. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 6, characterized in that, For upper limit indicators, a dynamic upper limit threshold is used. The upper limit indicators include temperature, charging voltage, voltage difference, and excessively high SOC. For lower limit indicators, a lower limit dynamic threshold is used, and the lower limit indicators include discharge voltage and insulation resistance; For two-sided indicators, both upper and lower dynamic thresholds are generated simultaneously. The upper and lower dynamic thresholds are calculated using the baseline mean, confidence coefficient, and standard deviation of long-period window data, or obtained using the quantiles of the fragment quantile set.

8. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 7, characterized in that, The specific formulas for calculating the upper and lower dynamic thresholds using the baseline mean, confidence coefficient, and standard deviation of long-period window data are as follows: in, The upper limit is a dynamic threshold. The lower limit is the dynamic threshold. Here is the confidence coefficient. The standard deviation of long-period window data; The specific method for obtaining the upper and lower dynamic thresholds from the quantiles of the fragment quantile value set is as follows: The data is divided into multiple complete running segments according to the monitoring scenario, with each segment serving as a statistical unit. For upper limit indicators, the 95th percentile of the sampling points within a segment is taken as the upper quantile value of the segment. For lower limit indicators, the 5th percentile of the sampling points within a segment is taken as the lower quantile value of the segment. Then, a second statistical analysis is performed on the set of quantile values ​​of all normal segments under the same monitoring scenario to construct the initial dynamic threshold at the segment level.

9. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, The preset warning conditions mentioned in step S7 include: the proportion of sampling points exceeding the dynamic threshold in the current monitoring segment exceeds the proportion threshold, or the duration of continuous exceeding the limit exceeds the duration threshold, or the segment quantile statistics value exceeds the dynamic threshold; The warning levels are divided into three levels based on the degree of deviation: Level 1 Warning, Level 2 Attention, and Level 3 Risk. The output warning report includes vehicle identification number, target parameter name, monitoring scenario, warning start time, warning end time, dynamic threshold, segment statistics, over-limit ratio, warning level, and recommended review items.

10. The method for dynamic threshold early warning of electric vehicle power batteries based on dual sliding window baselines according to claim 1, characterized in that, A weekly rolling update mechanism is adopted. Every week, the long-term threshold is recalculated using normal working condition data within the past long-term window, and the short-term correction within the most recent short-term window is added to obtain the final dynamic threshold for that week. When a warning segment is verified as a false alarm by manual intervention, maintenance, or subsequent data, the warning segment is fed back into the normal sample pool to calibrate the threshold tightness. When a warning segment is verified as a real anomaly, the warning segment is removed from the baseline calculation to avoid abnormal segments raising or lowering the threshold.

Images