Adaptive threshold fault early warning method for power battery based on kernel density estimation

Abstract

The application relates to the technical field of battery fault detection, and discloses a power battery adaptive threshold fault early warning method based on kernel density estimation, which comprises the following steps: collecting new energy automobile BMS power battery and fault battery pack single operation data, completing data segmentation through a time window; then extracting single body deviation features and performing standardization processing, and calculating abnormal scores of each single body according to the same; subsequently, an abnormal score history sample pool is built, a scoring distribution probability modeling is completed by using kernel density estimation, an adaptive threshold of a normal and abnormal area is determined by combining a second derivative change feature of a probability density function; real-time abnormal scores are compared with the threshold, and single body fault determination and early warning are realized; the system only updates the sample pool when an abnormality is detected, and drives the threshold to be dynamically iteratively optimized; and the technical problem that false positives and false negatives are caused by using a fixed threshold and static rules under complex non-stationary operation conditions of the power battery in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of battery fault detection technology, and specifically to an adaptive threshold fault early warning method for power batteries based on kernel density estimation. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the safety and reliability of the power battery, as the core energy unit of the vehicle, are directly related to the vehicle's operational quality and the safety of its occupants. Under complex actual operating conditions, due to multiple factors such as manufacturing inconsistencies, ambient temperature fluctuations, charge-discharge cycle aging, and accumulated mechanical stress, voltage inconsistencies, internal resistance differences, and capacity decay differences can easily occur between individual cells within the power battery pack. If early, subtle anomalies are not identified and warned in time, they can easily escalate into serious faults such as thermal runaway. Therefore, fault detection and early warning technologies at the individual power battery cell level have become a key requirement for battery management systems (BMS) and offline diagnostic platforms.

[0003] Existing power battery fault detection methods are mainly divided into two categories: physical model-based methods and data-driven methods.

[0004] Physical model-based methods typically rely on equivalent circuit models or electrochemical mechanism models to estimate the internal state of a battery by identifying model parameters (such as ohmic resistance, polarization capacitance, and diffusion coefficient), thereby determining the health of individual cells. However, such methods have extremely high requirements for model accuracy and parameter calibration. Under complex and variable actual operating conditions, model parameters are prone to drift, and it is difficult to take into account the differentiated characteristics of batteries with different chemical systems and at different aging stages, resulting in limited generalization ability and insufficient engineering applicability.

[0005] Data-driven methods utilize a large amount of measured data such as voltage, current, and temperature collected during battery operation to identify anomalies through statistical analysis or machine learning algorithms. While these methods reduce the reliance on precise physical models to some extent, most existing solutions still employ fixed thresholds or static judgment rules based on expert experience. Because the operating state of power batteries exhibits significant non-stationary characteristics—its voltage distribution, consistency level, and degradation trend continuously evolve with operating conditions, environment, and lifespan—fixed thresholds struggle to adapt to these dynamic changes, easily leading to numerous false alarms or missed alarms. Furthermore, some supervised learning-based methods rely on a large number of labeled samples for offline training, which is not only costly to label but also lacks online update capabilities, failing to reflect the temporal evolution of battery state in a timely manner and thus failing to meet the real-time and adaptive requirements of engineering applications. Summary of the Invention

[0006] The present invention aims to provide an adaptive threshold fault early warning method for power batteries based on kernel density estimation, so as to solve the technical problems of false alarms and missed alarms caused by the use of fixed thresholds and static rules in the complex and non-stationary operating conditions of power batteries.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive threshold fault early warning method for power batteries based on kernel density estimation, comprising: S1. Collect power battery data of the BMS operation of new energy vehicles, including the operation data of individual cells in multiple faulty battery packs, and process the data in segments according to time windows. S2. Extract features from the individual data within each window, construct deviation features that reflect the differences between individual data, and standardize the deviation features. S3. Construct individual anomaly characterization indicators based on standardized deviation features to obtain anomaly degree scores for each individual. S4. Construct a historical sample pool of abnormal scores, use the kernel density estimation method to model the probability density of the score distribution, and determine the adaptive threshold for the transition from the normal region to the abnormal region by analyzing the change of the second derivative of the density function. S5. Compare the abnormal score corresponding to the real-time data with the adaptive threshold. When the score meets the abnormal judgment condition, it is determined to be an abnormal single entity and an early warning is triggered. The corresponding abnormal score is updated to the historical sample pool only when an abnormality is detected, so as to realize the dynamic adjustment of the threshold.

[0008] The principle and advantages of this scheme are as follows: In practical applications, this scheme first uses time-windowed individual consistency deviation analysis to separate individual anomalies from global operating condition fluctuations and quantify them into anomaly scores. Then, kernel density estimation is used to perform non-parametric modeling of the score distribution, accurately depicting the data distribution shape without pre-setting the distribution form. Crucially, by analyzing the curvature sign characteristics of the second derivative of the probability density function, the transition boundary between normal and abnormal conditions is adaptively determined from the geometric structure of the distribution itself, completely eliminating the constraints of fixed thresholds and expert experience rules. Simultaneously, through an anomaly-driven selective update mechanism for the sample pool, the online dynamic evolution of the threshold is achieved while effectively blocking the dilution interference of normal data on the anomaly boundary. Therefore, this scheme solves the technical problems of fixed threshold failure, high false alarm and false negative rates, and insufficient online adaptive capability in existing technologies under complex non-stationary operating conditions, significantly improving the accuracy, robustness, and engineering applicability of power battery fault early warning.

[0009] Preferably, as an improvement, S1 specifically includes: collecting power battery data of the new energy vehicle BMS during normal operation, including time parameters and voltage parameters; collecting the operating data of all cells in the faulty power battery pack, including voltage data of normal cells and abnormal cells; and segmenting the collected battery data according to time windows.

[0010] The beneficial effects of this improvement are as follows: by simultaneously collecting individual operational data covering both normal and fault states, it ensures sufficient coverage and data integrity of abnormal samples, and transforms continuous data into structured analysis units according to time windows, providing a reliable data foundation for subsequent window-level feature extraction and anomaly detection; by using fixed-length sliding windows or overlapping sliding windows, the window length and step size can be flexibly configured according to the sampling frequency, which, while taking into account the timeliness of calculation, maintains data continuity through window overlap and reduces the risk of abnormal segments being truncated or missed.

[0011] Preferably, as an improvement, S2 specifically includes: extracting time series features from the voltage sequences of each individual unit within each time window to obtain a multi-dimensional feature vector corresponding to each individual unit; statistically aggregating the feature vectors of all individual units according to the feature dimensions to obtain window-level reference features; calculating the deviation between the feature vectors of each individual unit and the window-level reference features to construct individual unit difference features; and standardizing the individual unit difference features to eliminate the influence of different feature dimensions to obtain a standardized feature matrix.

[0012] The beneficial effects of this improvement are: by extracting multi-dimensional time series features from the individual cell voltage sequence and constructing a deviation quantity with the window-level aggregated reference features, it is possible to effectively separate individual cell anomalies from the global operating condition fluctuations of the battery pack, highlighting the consistency differences between individual cells; by standardizing the deviation features, the influence of the difference in dimensions between different feature dimensions is eliminated, so that the degree of anomalies of each individual cell can be compared horizontally on a unified scale, ensuring the accuracy and comparability of subsequent anomaly score calculations.

[0013] Preferably, as an improvement, S3 specifically includes: constructing an anomaly representation function for each individual based on a standardized feature matrix, and generating a corresponding anomaly metric; the anomaly metric is used to characterize the degree of deviation of the individual from the overall distribution within the current time window; and the anomaly metric of each individual is used as the anomaly score of that individual within the current time window.

[0014] The beneficial effects of this improvement are: by mapping the standardized multidimensional deviation vector to a single-dimensional anomaly score through the anomaly representation function, the quantitative representation of the degree of anomaly of a single entity is realized; this score directly reflects the degree of deviation of a single entity from the overall distribution of the current window, so that the anomaly judgment has a clear numerical basis, which facilitates efficient and objective quantitative comparison with the adaptive threshold.

[0015] Preferably, as an improvement, S4 specifically includes: storing the abnormal scores obtained in the historical window into a historical sample pool to construct a one-dimensional score distribution sample set; using a kernel density estimation method to perform probability density modeling on the score sample set to obtain a continuous probability density function; calculating the first and second derivatives of the probability density function to obtain the curvature change characteristics of the density function; and detecting the first zero crossover point where the second derivative changes from negative to positive in the low-value region of the score distribution, and determining the score value corresponding to that point as an adaptive threshold.

[0016] The beneficial effects of this improvement are as follows: By using kernel density estimation to perform nonparametric probability density modeling on historical samples of abnormal scores, no prior assumptions such as normal distribution are required regarding the distribution form. This allows for a true adaptation to the complex and varied distribution of abnormal scores under different operating conditions and aging stages of the power battery. Furthermore, by calculating the first and second derivatives of the probability density function, the first zero-crossing point where the second derivative changes from negative to positive is captured in the low-value region of the score. The adaptive boundary of the transition from the normal region to the abnormal region is accurately located using the curvature inflection point of the distribution itself. Compared with the traditional threshold setting method based on fixed quantiles or local minimum density values, this method has higher sensitivity and positioning accuracy to changes in the shape of non-stationary data distributions, thereby effectively improving the detection stability and threshold adaptation capability under complex operating conditions.

[0017] Preferably, as an improvement, S5 specifically includes: obtaining the abnormal score of each individual within the current time window and comparing it with the adaptive threshold; when the abnormal score is less than or equal to the adaptive threshold, determining that the individual is in an abnormal state; if there is an abnormal individual within the current time window, triggering a window-level early warning mechanism; updating the abnormal score sample corresponding to the time window that triggered the early warning to the historical sample pool, and recalculating the probability density function and the adaptive threshold.

[0018] The beneficial effects of this improvement are as follows: by dynamically comparing real-time anomaly scores with adaptive thresholds, online identification and window-level real-time early warning of abnormal units are achieved; by selectively adding the anomaly window score samples that trigger early warnings to the historical sample pool and recalculating the probability density function and threshold, the anomaly boundary can adaptively migrate with the time evolution of battery operating status; at the same time, the mechanism of updating the historical sample pool only when an anomaly is detected effectively avoids the dilution and interference of massive normal data on the anomaly distribution boundary, significantly improves the long-term robustness and cross-operating condition applicability of threshold estimation, and realizes a closed loop of online learning and adaptive updating.

[0019] Preferably, as an improvement, the historical sample pool is updated only when an abnormal window is detected, and the adaptive threshold is dynamically adjusted as the historical sample pool is updated.

[0020] The beneficial effects of this improvement are as follows: By strictly limiting the update of the historical sample pool to the window in which anomalies are detected, the dilution and contamination of the probability density distribution of anomaly scores by the influx of massive amounts of normal data are effectively prevented. This ensures the clarity of the morphology and structural stability of the anomaly distribution boundary and avoids the risk of false negatives caused by the shrinkage of the threshold towards dense areas due to excessive accumulation of normal samples. At the same time, the adaptive threshold only dynamically migrates with the directional accumulation of anomaly samples, which can truly reflect the time evolution trajectory of the deterioration of the operating state of the power battery, rather than causing unnecessary threshold oscillations due to random fluctuations in operating conditions. This significantly improves the long-term robustness of threshold estimation, the cross-operating-condition stability of detection and judgment, and the system's online learning ability for new anomaly patterns. Attached Figure Description

[0021] Figure 1 This is a flowchart of an embodiment of the present invention.

[0022] Figure 2 This is a graph showing the KDE and derivative analysis of the anomaly score for window 1 in an embodiment of the present invention.

[0023] Figure 3 This is a graph showing the KDE and derivative analysis of the anomaly score for window 70 in an embodiment of the present invention. Detailed Implementation

[0024] The following detailed description illustrates the specific implementation methods: Example The basics are as follows: Figure 1 As shown, the adaptive threshold fault early warning method for power batteries based on kernel density estimation includes: S1. Power battery data acquisition and time window segmentation processing: Acquire power battery data during the operation of the new energy vehicle BMS, and segment it according to time windows to provide structured input for subsequent feature extraction and anomaly analysis.

[0025] S11. Collect power battery data for normal operation of the BMS in new energy vehicles, including time parameters and voltage parameters. The time parameters are timestamp sequences of data sampling times; the voltage parameters are the voltage sampling values ​​of each battery cell at the corresponding sampling time.

[0026] S12. Collect all cell operation data in the faulty power battery pack, including voltage data of normal and abnormal cells, to cover the operation performance of cells in different health states within the battery pack.

[0027] S13. The collected battery data is segmented according to time windows to form a time-windowed voltage dataset:

[0028] in, Let i be the set of all individual cell voltage sequences within the t-th time window; t is the time window index, t=1,2,...,T, where T is the total number of windows; i is the individual cell index, i=1,2,...,N; N is the total number of individual cells in the battery pack, and N is a positive integer. This represents the voltage sequence of the i-th unit within the t-th time window. In this embodiment, the time window can be a fixed-length sliding window or an overlapping sliding window.

[0029] S2. Individual bias feature extraction and standardization: Within each time window, bias features that reflect differences between individuals are extracted, and the influence of dimensions is eliminated through standardization, providing a unified dimension of feature input for the construction of anomaly scoring.

[0030] S21. Extract time-series features from the voltage sequences of each individual unit within each time window to obtain a multi-dimensional feature vector corresponding to each unit:

[0031] in, This is the multidimensional feature vector obtained after feature extraction of the i-th individual within the t-th time window; This is a time series feature extraction function used to map the original voltage sequence of a single unit into a low-dimensional feature representation.

[0032] In this embodiment, the time series feature extraction function Statistical feature extraction can be used, such as extracting at least one of the following from the voltage sequence: mean, standard deviation, range, coefficient of variation, and first-order difference mean, to construct a multidimensional feature vector. Taking the extraction of mean and standard deviation as an example, then D=2. ,in This represents the arithmetic mean of the voltage of the i-th individual cell within the window. Let D be its standard deviation and D be the number of feature dimensions.

[0033] S22. Perform statistical aggregation on the feature vectors of all individual units according to the feature dimension to obtain window-level reference features. :

[0034] in, This is the window-level reference feature for the t-th time window; This is an aggregation function used to synthesize the feature vectors of all N individuals to represent the overall feature level of the window.

[0035] In this embodiment, the aggregation function can be implemented by taking the median of each dimension, taking the median of each component of the same dimension of each individual feature vector, and combining the medians of each dimension component into a D-dimensional window-level reference feature. .

[0036] S23. Calculate the deviation between each individual feature vector and the window-level reference feature. Construct individual differential features:

[0037] in, Let represent the individual difference characteristics, or deviation, of the i-th individual within the t-th time window.

[0038] S24. Standardize the individual differential features to eliminate the influence of different feature dimensions, and obtain a standardized feature matrix:

[0039] in, Let be the standardized feature vector of the i-th individual within the t-th time window; is the arithmetic mean vector of all individual differential features across all dimensions within the t-th time window; Let be the standard deviation vector of all individual differential features in each dimension within the t-th time window.

[0040] Among them, the arithmetic mean vector The calculation formula is:

[0041] Where N is the total number of individual cells in the battery pack; Let be the monomeric difference feature of the i-th monomer.

[0042] Standard deviation vector The calculation formula is:

[0043] S3. Construction of Individual Anomaly Characterization Indicators and Calculation of Anomaly Scores: Based on standardized deviation characteristics, a scoring index that can quantify the degree of individual anomalies is constructed.

[0044] S31. Based on the standardized feature matrix, construct anomaly representation functions for each individual entity and generate corresponding anomaly measurement values.

[0045] S32. The anomaly metric is used to characterize the degree of deviation of an individual from the overall distribution within the current time window; the smaller the anomaly metric, the worse the consistency of the behavior of the individual with that of most individuals within the window, and the higher the probability of an anomaly.

[0046] S33. Use the anomaly metric of each individual as the anomaly score for that individual within the current time window:

[0047] in, Let be the anomaly score of the i-th individual within the t-th time window, and be a scalar real number; This is an anomaly mapping function used to map a standardized multidimensional deviation vector into a single-dimensional anomaly score value; Let be the standardized feature vector of the i-th individual.

[0048] In this embodiment, the exception mapping function The negative Euclidean norm (negative L2 norm) of the standardized deviation vector can be used, i.e.:

[0049] in, It is the Euclidean norm; For standardized feature vectors The d-th component, d=1,2,...,D; D is the number of feature dimensions.

[0050] S4. Adaptive threshold determination based on kernel density estimation and second derivative of density function: Construct an abnormal rating historical sample pool, use kernel density estimation to perform nonparametric probability density modeling of the rating distribution, and determine the adaptive threshold for transition from normal region to abnormal region by analyzing the change of second derivative of density function.

[0051] S41. Store the abnormal scores obtained in the historical window to the historical sample pool. And construct a one-dimensional rating distribution sample set:

[0052] in, The historical sample pool for abnormal ratings contains a one-dimensional sample set of M historical abnormal ratings. This represents the abnormal rating value corresponding to the j-th sample in the historical sample pool; j is the sample index, j=1,2,...,M; M is the total number of rating samples in the historical sample pool, M is a positive integer, and it accumulates gradually as the system runs.

[0053] S42. The kernel density estimation method is used to model the probability density of the scoring sample set to obtain a continuous probability density function:

[0054] in, Let be the probability density function value at the rating value s, used to describe the distribution of abnormal ratings; s is the abnormal rating variable to be evaluated, and is a continuous real number; is the kernel function used to smooth the sample points; h is the bandwidth parameter, h>0, used to control the smoothness of the kernel function; the larger h is, the smoother the density curve, and the smaller h is, the closer the density curve is to the sample points; M is the total number of rating samples in the historical sample pool. This represents the anomaly score of the j-th sample in the historical sample pool.

[0055] In this embodiment, the kernel function A Gaussian kernel function can be used, and its expression is:

[0056] in, For the independent variable of the kernel function, It is a natural exponential function.

[0057] In this embodiment, the bandwidth parameter h can be initialized and determined using Silverman empirical rules:

[0058] in, Historical sample pool The standard deviation of the M abnormal rating samples; M is the total number of samples.

[0059] S43. Calculate the first and second derivatives of the probability density function to obtain the curvature change characteristics of the density function; The formula for calculating the first derivative is:

[0060] in, It is the first derivative of the probability density function p(s) at the rating value s, used to describe the increasing or decreasing trend of probability density as the rating changes.

[0061] The formula for calculating the second derivative is:

[0062] in, Let p(s) be the second derivative of the probability density function p(s) at the score value s, used to describe the curvature change of the probability density function curve.

[0063] S44. Within the low-value region of the score distribution, detect the first zero-crossing point where the second derivative changes from negative to positive, and determine the score value corresponding to this point as the adaptive threshold. :

[0064] in, The adaptive threshold is the boundary score value for the transition from the normal region to the abnormal region; The second derivative takes the value of zero at the score s, which is the inflection point; Let be the value of the second derivative at the left-hand limit at point s. Let be the value of the second derivative at the right limit of point s; condition and This indicates the curvature inflection point at s where the second derivative crosses zero from a negative value to a positive value, corresponding to the transition of the density distribution from a normal dense region to an abnormally sparse region.

[0065] The low-value region of the score distribution refers to the interval near the lowest score within the range of abnormal scores, i.e., the interval near the abnormal side. In actual calculations, the zero-crossing point can be searched within the lower quantile region of the entire score range, for example, in... Search within the interval, where The minimum score in the sample pool. Use lower quantiles for sample scoring, such as the 10th or 20th percentile, to ensure that the determined threshold is located at the transition boundary between normal and abnormal rather than in the main peak region of the distribution.

[0066] S5. Real-time anomaly detection, early warning triggering, and dynamic update of the sample pool: The anomaly score corresponding to the real-time data is compared with the adaptive threshold to realize the detection and early warning of anomalies. The historical sample pool is updated through feedback of anomalies to realize the dynamic adjustment of the threshold.

[0067] S51. Obtain the anomaly score of each individual within the current time window and compare it with the adaptive threshold.

[0068] S52. When the abnormal score is less than or equal to the adaptive threshold, the individual is determined to be in an abnormal state.

[0069] S53. If an abnormal cell exists within the current time window, trigger the window-level early warning mechanism and report the abnormal alarm information for that time window to the battery management system.

[0070] S54. Update the abnormal score samples corresponding to the time window that triggers the warning to the historical sample pool to achieve dynamic updating of the probability distribution and adaptive adjustment of the threshold.

[0071] Specifically, the anomaly scores corresponding to the individuals identified as abnormal within this window are added to the historical sample pool. Then, S42 to S44 are re-executed to calculate the updated probability density function and adaptive threshold.

[0072] As attached Figure 2As shown, this figure illustrates the kernel density estimation and derivative analysis results of the abnormal rating distribution in the initial stage (Window 1). This stage involves kernel density estimation based on M=40 abnormal rating samples from the historical sample pool. From top to bottom in the figure, the parameters are the probability density function p(s), the first derivative, and the... and second derivative The curve showing how the abnormality score s changes.

[0073] In the low-value region of the score distribution, by detecting the second derivative The first zero-crossing point when the value changes from negative to positive determines the adaptive threshold for this stage. The sub-figures are marked with red vertical dashed lines and dots. It can be seen that even with a limited initial sample size (M=40), this invention can still accurately locate the transition boundary between normal and abnormal regions using the curvature structure characteristics of the density function, thus establishing an effective initial threshold.

[0074] As attached Figure 3 As shown, it illustrates the kernel density estimation and derivative analysis results of the abnormal score distribution during the evolutionary stage (window 70). As the process progresses, the historical sample pool is continuously and dynamically updated when an abnormal window is detected, expanding the total number of samples from the initial 40 to M=760. At this point, the probability density distribution curve is smoother, and the left tail features (low-score region, corresponding to the abnormal score region) are more pronounced, indicating that the accumulation of abnormal samples leads to a more refined characterization of the distribution pattern. Based on the same second-order derivative zero-crossing rule, the recalculated adaptive threshold is transferred to... (The location is marked by the red vertical dashed line in the figure). The adaptive migration of this threshold reflects the time evolution characteristics of the power battery's operating state, demonstrating the invention's ability to dynamically characterize non-stationary data distributions and its online learning capability.

[0075] After obtaining the above adaptive threshold, the individual anomaly scores are calculated within each time window. Compared with the current threshold Perform a step-by-step enumeration comparison: when If a single entity is determined to be in an abnormal state, a window-level early warning mechanism is triggered. Simultaneously, the abnormal score samples corresponding to the abnormal window are added to the historical sample pool to achieve dynamic updates of the probability distribution and threshold. If there are no abnormal entities in the current window, the sample pool remains unchanged to maintain threshold stability.

[0076] In summary, this invention provides an adaptive threshold fault early warning method for power batteries based on kernel density estimation, which has the following advantages: First, at the probabilistic modeling level, this invention uses kernel density estimation to perform nonparametric probability density modeling of individual cell anomaly scores. It does not require prior assumptions about the data distribution form, such as normal distribution or Gaussian mixture, and can realistically depict the complex form of anomaly score distribution of power batteries under different life cycles and different operating conditions. This fundamentally avoids detection bias caused by model preset mismatch and improves the accuracy and stability of fault detection.

[0077] Secondly, regarding the threshold determination mechanism, this invention breaks through the traditional methods of setting fixed thresholds, static quantiles, or local minimum densities. It innovatively utilizes the sign-changing characteristic of the second derivative curvature of the probability density function to adaptively "grow" the transition boundary between normal and abnormal regions from the geometric structure of the data distribution itself. This threshold is determined by the inherent morphology of the distribution and can dynamically migrate following the non-stationary changes in battery operating conditions. This effectively solves the problem of fixed threshold failure under complex operating conditions and significantly reduces false alarm and false negative rates.

[0078] Third, regarding online updates and robustness, this invention constructs a historical sample pool of abnormal scores and employs an anomaly-driven selective update mechanism—adding the corresponding score sample to the pool only when an anomaly window is detected. This effectively prevents the dilution and interference of massive amounts of normal data on the boundaries of anomaly distributions. Under this mechanism, the probability density model and adaptive threshold can evolve progressively with the accumulation of abnormal samples, ensuring the long-term stability of threshold estimation and endowing the system with online learning capabilities for novel anomaly patterns. This gives anomaly judgment good adaptability and cross-condition applicability.

[0079] Fourth, at the engineering application level, this invention can establish an effective initial threshold based on limited historical samples in the initial stage. As operational data accumulates, the boundary judgment criteria are continuously optimized, maintaining high sensitivity to early weak anomalies while effectively suppressing false alarms caused by fluctuations in operating conditions and environmental changes. The entire method has a clear calculation process, requires no labeled data or complex model training, and is easy to embed into existing battery management systems (BMS) or offline diagnostic platforms, providing a technical solution for early warning of cell-level faults in power batteries that combines theoretical rigor with engineering practicality.

[0080] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A power battery adaptive threshold fault early warning method based on kernel density estimation, characterized in that, include: S1. Collect power battery data of the BMS operation of new energy vehicles, including the operation data of individual cells in multiple faulty battery packs, and process the data in segments according to time windows. S2. Extract features from the individual data within each window, construct deviation features that reflect the differences between individual data, and standardize the deviation features. S3. Construct individual anomaly characterization indicators based on standardized deviation features to obtain anomaly degree scores for each individual. S4. Construct a historical sample pool of abnormal scores, use the kernel density estimation method to model the probability density of the score distribution, and determine the adaptive threshold for the transition from the normal region to the abnormal region by analyzing the change of the second derivative of the density function. S5. Compare the abnormal score corresponding to the real-time data with the adaptive threshold. When the score meets the abnormal judgment condition, it is determined to be an abnormal single entity and an early warning is triggered. The corresponding abnormal score is updated to the historical sample pool only when an abnormality is detected, so as to realize the dynamic adjustment of the threshold.

2. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 1, characterized in that, S1 specifically includes: collecting power battery data of the new energy vehicle BMS during normal operation, including time parameters and voltage parameters; collecting the operating data of all cells in the faulty power battery pack, including voltage data of normal cells and abnormal cells; and processing the collected battery data in segments according to time windows.

3. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 2, characterized in that, S2 specifically includes: extracting time series features from the voltage sequences of each individual unit within each time window to obtain a multi-dimensional feature vector corresponding to each individual unit; statistically aggregating the feature vectors of all individual units according to the feature dimension to obtain window-level reference features; calculating the deviation between the feature vector of each individual unit and the window-level reference features to construct individual unit difference features; and standardizing the individual unit difference features to eliminate the influence of different feature dimensions to obtain a standardized feature matrix.

4. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 3, characterized in that, S3 specifically includes: constructing an anomaly representation function for each individual based on a standardized feature matrix, and generating a corresponding anomaly metric; the anomaly metric is used to characterize the degree of deviation of an individual from the overall distribution within the current time window; and the anomaly metric of each individual is used as the anomaly score of that individual within the current time window.

5. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 4, characterized in that, S4 specifically includes: storing abnormal scores obtained in the historical window into a historical sample pool to construct a one-dimensional score distribution sample set; using a kernel density estimation method to model the probability density of the score sample set to obtain a continuous probability density function; calculating the first and second derivatives of the probability density function to obtain the curvature change characteristics of the density function; and detecting the first zero crossover point where the second derivative changes from negative to positive in the low-value region of the score distribution, and determining the score value corresponding to that point as an adaptive threshold.

6. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 5, characterized in that, S5 specifically includes: obtaining the abnormal score of each individual within the current time window and comparing it with the adaptive threshold; when the abnormal score is less than or equal to the adaptive threshold, determining that the individual is in an abnormal state; if there is an abnormal individual within the current time window, triggering a window-level early warning mechanism; updating the abnormal score sample corresponding to the time window that triggered the early warning to the historical sample pool, and recalculating the probability density function and the adaptive threshold.

7. The adaptive threshold fault early warning method for power batteries based on kernel density estimation according to claim 6, characterized in that: The historical sample pool is updated only when an abnormal window is detected, and the adaptive threshold is dynamically adjusted as the historical sample pool is updated.

Images