Method for judging SOH (state of health) state attenuation trend of battery of electric vehicle

Abstract

The invention discloses a method for judging an SOH state attenuation trend of an electric vehicle battery, particularly relates to the field of vehicle battery state evaluation, and aims to solve the problem of insufficient stability of attenuation trend judgment. According to the method, fragment charging and discharging data is converted into an information structure with a recognizable trend in a continuous space, and interference of random factors in initial classification is reduced by means of a combination mode of steady-state centroid screening and topological connection clustering, so that an evolution trajectory of a health state has high stability; on the basis, a potential energy diagram driven feature extraction and machine learning judgment mechanism is further introduced, dynamic coordination between a macroscopic trend and microscopic disturbance is achieved, it is ensured that trend filtering response can be tightly attached to real capacity change instead of single-point errors, and therefore a continuous, credible and low-pseudo-fluctuation health probability curve is formed; and the judgment process can conform to the existing rule and can dynamically perceive the abnormal change risk.

Description

Technical Field

[0001] The present invention relates to the field of automobile battery state assessment, and more specifically, to a method for judging the state attenuation trend of an electric vehicle battery. Background Art

[0002] The capacity of power lithium batteries gradually decreases with cycle count, and fluctuations in their state of health (SOH) directly impact range and scheduling. The published paper "A Multi-Timescale SOH Assessment Method for Electric Vehicle Batteries" (Publication No. CN109752656B) proposes an iEKF-GPR and K-means aggregation strategy. This method tracks capacity changes by matching segmented charging curves with historical full-set data, and introduces a five-cluster classification result to provide a reference for subsequent uncertainty indices. This method utilizes multi-timescale information fusion of laboratory full-charge records and in-use segments. However, the clustering step uses randomly initialized centers, causing the classification results to drift with changes in noise distribution and sample sparsity.

[0003] In actual operating conditions, historical charge and discharge samples are often mixed with noise such as sudden temperature changes and differences in driving habits. The randomly set initial centroid often falls near the outlier fragment, and the clustering iteration stage is pulled in the abnormal direction. The final cluster boundary is difficult to faithfully depict the true distribution of the health status. This offset first weakens the similarity measure between historical samples and new fragments, and secondly amplifies the cumulative error of the uncertainty index in multiple rounds of estimation processes, resulting in a jump in the interpretation of the SOH decay trend. Maintenance decisions rely on stable decay curves. Once the classification fluctuations affect the smoothness of the trend line, the operation and maintenance end faces the dual risks of replacing the battery prematurely or missing the hidden danger warning. The above-mentioned technical gaps have prompted the present invention to focus on building a robust SOH decay trend judgment mechanism under the impact of randomness and noise to support more continuous and reliable health status tracking.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for judging the state attenuation trend of the SOH of electric vehicle batteries. By constructing a multi-scale attenuation identification mechanism based on a time-series state vector, the fragmented charge and discharge data are converted into a trend-recognizable information structure in a continuous space, and with the help of a combination of steady-state centroid screening and topological connection clustering, the interference of random factors in the initial classification is reduced, so that the evolution trajectory of the health state has a high degree of stability; on this basis, a potential energy graph-driven feature extraction and machine learning judgment mechanism is further introduced to achieve dynamic coordination between macro trends and micro disturbances, ensuring that the trend filtering response can closely follow the actual capacity changes rather than single-point errors, thereby forming a continuous, reliable, and low-pseudo-fluctuation health probability curve; so that the judgment process can not only conform to existing rules, but also dynamically perceive the risk of abnormal changes, so as to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: S1: Collect the charge and discharge segment curves one by one and map them into a multi-scale state vector set and write them into the timing buffer; S2: Reference the historical state vector library in the time series buffer, use density threshold screening to locate the steady-state centroid and remove outlier curves to form the initial centroid set; S3: Perform adaptive radius clustering with the initial centroid set as the topological anchor point, and output the classification index chain for the current state vector; S4: Generates a decay potential energy diagram based on the classified index chain aggregation time series, extracts macro and micro characterization features, and obtains trend adjustment factors through machine learning reasoning. The trend adjustment factors are then used to adjust the filter weights to track the capacity decay trajectory and generate a health state probability curve; S5: Use the probability curve curvature mutation event as a trigger to mark the mutation event and guide the full verification, and write back the new centroid to update the clustering benchmark.

[0007] In a preferred embodiment, step S1 includes the following contents: The voltage, current and temperature data in the charge and discharge segment curves are subjected to multi-scale wavelet decomposition respectively to generate multiple levels of low-frequency approximate coefficients and high-frequency detail coefficients. The low-frequency approximate coefficients and high-frequency detail coefficients of the voltage, current and temperature are spliced ​​in a predetermined order to form a multi-scale state vector. The decomposition level of the multi-scale wavelet decomposition is 3 to 5 levels, and the Daubechies wavelet basis is used. Writing into the timing buffer includes attaching the timestamp of the acquisition start time to the multi-scale state vector and writing it into the timing buffer in chronological order. When the buffer is full, the multi-scale state vector with the earliest timestamp is removed and a new multi-scale state vector is inserted.

[0008] In a preferred embodiment, step S2 includes the following: Retrieve a historical state vector library from a timing buffer; calculate the local density of each state vector in the historical state vector library by counting the number of state vectors within a predetermined radius; multiply the median of the local density by a predetermined coefficient as a density threshold; select state vectors whose local density is greater than or equal to the density threshold as candidate centroids; for each state vector in the historical state vector library, calculate its distance to the nearest candidate centroid; remove state vectors whose distance is greater than the outlier threshold to obtain a refined historical state vector library.

[0009] In a preferred embodiment, step S2 further includes the following: Recalculate the local density for each state vector in the refined historical state vector library; for each state vector in the refined historical state vector library, determine its minimum distance to another state vector with higher local density; calculate the product of the local density and the minimum distance for each state vector in the refined historical state vector library; and select a predetermined number of state vectors with the highest product as the initial centroid set.

[0010] In a preferred embodiment, step S3 includes the following contents: The initial centroid set is used as the topological anchor point; the Euclidean distance between the current state vector and each centroid vector in the initial centroid set is calculated; and the adaptive clustering radius is determined by combining the minimum value and the median of these Euclidean distances according to a predetermined weighting factor.

[0011] In a preferred embodiment, step S3 further includes the following: Perform adaptive radius clustering, specifically identifying the centroid vector with the minimum Euclidean distance to the current state vector, and comparing this minimum Euclidean distance with the adaptive clustering radius. If the minimum Euclidean distance is less than or equal to the adaptive clustering radius, the current state vector is assigned to the cluster represented by the corresponding centroid vector. If the minimum Euclidean distance is greater than the adaptive clustering radius, a new cluster is created with the current state vector as the centroid vector. Finally, the classification results are recorded in the classification index chain, which stores the classification indexes of all state vectors in chronological order.

[0012] In a preferred embodiment, step S4 includes the following contents: The attenuation potential energy map is generated based on the classified index chain aggregation time series. The attenuation phase skewness is calculated by performing multi-scale wavelet decomposition on the attenuation potential energy map and extracting the main scale phase spectrum. At the same time, the attenuation potential energy map is divided in a sliding window manner and the window information entropy is calculated. The relative uniformity distribution of the entropy sequence is obtained to calculate the local entropy dispersion.

[0013] In a preferred embodiment, step S4 further includes the following: The attenuation phase skewness and local entropy dispersion are input into the lightweight gradient boosting decision maker to output the trend adjustment factor. Finally, the trend adjustment factor is used to adjust the filter weight to track the capacity decay trajectory and generate a health state probability curve.

[0014] In a preferred embodiment, step S5 includes the following contents: Numerical differentiation is performed on the health state probability curve to calculate the first and second derivatives to determine the curvature, and the local peak points in the curvature sequence are identified and compared with a preset threshold to detect mutation events.

[0015] In a preferred embodiment, step S5 further includes the following: When a mutation event is detected and there is no full charge verification record within the preset time window, a full charge verification is triggered to obtain accurate battery capacity data. The battery status data generated by the full charge verification is used to update the historical state vector library, and a new initial centroid set is generated through density threshold screening and outlier curve removal.

[0016] The technical effects and advantages of the method for judging the SOH state decay trend of electric vehicle batteries of the present invention are as follows: The present invention transforms fragmented charge and discharge data into trend-identifiable information structures in continuous space by constructing a multi-scale attenuation identification mechanism based on time-series state vectors, and by means of a combination of steady-state centroid screening and topological connection clustering, significantly reduces the interference of random factors in the initial classification, making the evolution trajectory of the health state highly stable. On this basis, the potential energy graph-driven feature extraction and machine learning judgment mechanism are further introduced to achieve dynamic coordination between macro trends and micro disturbances, ensuring that the trend filter response can closely follow the actual capacity changes rather than single-point errors, thereby forming a continuous, reliable, low-pseudo-fluctuation health probability curve. The entire process realizes the organic coupling from historical classification experience, adaptive judgment to predictive feedback, so that the judgment process can not only conform to existing rules, but also dynamically perceive the risk of abnormal changes, providing more robust analytical support for trend judgment of battery status, and reducing the risk of excessive warning or delayed response due to local fragment misleading. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of a method for judging the state of charge (SOH) decay trend of an electric vehicle battery according to the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 The present invention provides a method for judging the state of charge (SOH) decay trend of an electric vehicle battery, comprising: S1: Collect the charge and discharge segment curves one by one and map them into a multi-scale state vector set and write them into the timing buffer.

[0020] S2: Reference the historical state vector library in the timing buffer, use density threshold screening to locate the steady-state centroid and remove outlier curves to form an initial centroid set.

[0021] S3: Perform adaptive radius clustering with the initial centroid set as the topological anchor point, and output the classification index chain for the current state vector.

[0022] S4: Generate a decay potential energy diagram based on the classification index chain aggregation time series, extract macro and micro characterization features and obtain the trend adjustment factor through machine learning reasoning. The trend adjustment factor is then used to adjust the filter weight to track the capacity decay trajectory and generate a health status probability curve.

[0023] S5: Use the probability curve curvature mutation event as a trigger to mark the mutation event and guide the full verification, and write back the new centroid to update the clustering benchmark.

[0024] The capacity of electric vehicle (EV) lithium-ion batteries gradually decays with the number of charge and discharge cycles. Accurately determining their state of health (SOH) is directly related to range prediction and vehicle scheduling. Existing techniques, such as iEKF-GPR and K-means aggregation strategies, can track capacity changes through multi-timescale information fusion. However, the random initial centroids are susceptible to noise and outliers, leading to fluctuations in classification results and, in turn, affecting the stability of SOH decay trends. This paper addresses this issue by proposing a robust SOH decay trend determination method. This method aims to overcome the impact of noise and randomness through steps such as multi-scale state vector construction, steady-state centroid screening, adaptive radius clustering, and potential energy graph feature extraction, achieving continuous and reliable SOH tracking. Step S1, the starting point of the entire solution, is responsible for converting the raw charge and discharge segment curves into a standardized multi-scale state vector set and storing them in a time series, laying the data foundation for subsequent analysis. The processing logic of this step must ensure accurate data acquisition, comprehensive feature extraction, and effective time series management to support the precise retrieval of historical and current data in subsequent steps.

[0025] The goal of step S1 is to collect the charge and discharge segment curves of the electric vehicle battery one by one, map them into a multi-scale state vector set, and write them into the time series buffer, providing a high-quality time series data foundation for judging the SOH decay trend. The following is a detailed expansion of the specific processing logic: S1-1, collecting charge and discharge segment curves; During electric vehicle operation, batteries undergo multiple charge and discharge cycles. Data acquisition equipment must record key parameters from these cycles in real time to form continuous charge and discharge segment curves. Specifically, the data collected includes the battery's voltage, current, and temperature. These parameters vary over time, reflecting the battery's performance over a specific period. During the acquisition process, voltage is measured in volts, current is recorded in amperes, with positive values ​​indicating charge and negative values ​​indicating discharge. Temperature is recorded in degrees Celsius. The data acquisition equipment uses high-precision sensors to record these parameters at regular intervals, for example, every 0.1 seconds. This high-frequency sampling fully captures the battery's dynamic behavior under rapidly changing operating conditions. Each set of continuously recorded parameters forms a charge and discharge segment curve, which can be viewed as a multivariate data sequence containing the time-varying variations of voltage, current, and temperature. Each curve corresponds to a complete or partial charge and discharge cycle.

[0026] The purpose of collecting voltage, current, and temperature is to ensure that these parameters can directly reflect the battery's electrochemical characteristics and thermal management status, while high-frequency sampling ensures data precision and integrity. The charge and discharge fragment curves generated in this way provide comprehensive raw data support for subsequent feature extraction, preserving detailed changes in battery performance.

[0027] S1-2, mapped to a multi-scale state vector; The collected charge and discharge segment curves require further processing to convert them into a mathematical form that is easy to analyze, namely a multi-scale state vector. This process extracts features from the curves and decomposes them into feature sets at different time and frequency scales, thereby comprehensively characterizing the performance changes of the battery. Specifically, multi-scale wavelet decomposition is first performed on the voltage, current, and temperature data in each charge and discharge segment curve. Wavelet decomposition is a signal processing technique that can decompose raw data into multiple levels of features, taking into account both time and frequency information. For each parameter, such as voltage data, the decomposition process begins with the raw data and gradually divides it into a low-frequency approximate portion and a high-frequency detail portion. The decomposition level is typically set to 3 to 5 levels. At each level of decomposition, a low-pass filter is used to extract the overall trend of the data to generate low-frequency approximate coefficients. Then, a high-pass filter is used to extract local fluctuations in the data to generate high-frequency detail coefficients. The low-frequency approximate coefficients of each level continue to serve as input for the next level of decomposition until the preset number of decomposition levels is reached.

[0028] After the decomposition is completed, a set of wavelet coefficients is generated for the voltage, current, and temperature of each charge and discharge segment curve, including low-frequency approximate coefficients and high-frequency detail coefficients at all levels. These coefficients are organized according to parameter type and decomposition level. For example, all voltage coefficients are placed in one group, all current coefficients are placed in another group, and all temperature coefficients are placed in a third group. Subsequently, these three groups of coefficients are spliced ​​together in a predetermined order to form a high-dimensional vector, which is the multi-scale state vector. Each part of the multi-scale state vector corresponds to the eigenvalue of voltage, current, or temperature at a specific decomposition level, and fully contains information about battery performance at different time and frequency scales.

[0029] The reason for using multi-scale wavelet decomposition is that the battery charging and discharging data has non-stationary characteristics, and a single time or frequency analysis cannot fully capture its behavior. Wavelet decomposition can simultaneously extract the overall trend and local details of the data.

[0030] S1-3, write timing buffer; The generated multi-scale state vectors need to be stored in the timing buffer in the order of acquisition time to support subsequent timing analysis. The timing buffer is a storage structure with limited capacity, designed to store a set of state vectors in the most recent period of time. In specific operations, after each multi-scale state vector is generated, it will be accompanied by a timestamp to mark the start time of acquisition of its corresponding charge and discharge segment curve. These state vectors are written into the timing buffer in order from earliest to latest timestamps. The capacity of the buffer is pre-set to a fixed value, for example, it can store 1,000 state vectors. When a newly generated state vector needs to be written, if the buffer is not full, it is directly added to the end; if the buffer is full, the state vector with the earliest timestamp is removed and the new state vector is written to the end. Through this first-in-first-out method, the latest set of state vectors is always stored in the buffer.

[0031] The multi-scale state vector set in the timing buffer generated in step S1 plays a key role in the entire analysis process. The state vector set in the timing buffer can be used as part of the historical data for subsequent steps. For example, in the subsequent steady-state centroid screening process, these state vectors are used to identify stable feature patterns and eliminate abnormal data. The state vector newly written to the timing buffer is used as the current data to participate in the adaptive clustering analysis and generate classification results. In addition, the state vector set in the timing buffer can also be aggregated in chronological order to generate a feature map of the attenuation trend, extract key features, and infer the influencing factors of the trend change.

[0032] Step S1 completes the standardization and time-series processing of battery performance data by collecting charge and discharge segment curves of electric vehicle batteries, converting them into a set of multi-scale state vectors, and writing them into a time series buffer. The acquisition process records the dynamic changes in voltage, current, and temperature, and multi-scale wavelet decomposition extracts multi-level features of the data. This provides high-quality time series data support for subsequent feature screening, cluster analysis, and trend inference, ensuring the accuracy and stability of the analysis results.

[0033] Step S1 has collected the charge and discharge fragment curves and mapped them into a multi-scale state vector set, and written them into the time series buffer, providing a standardized time series data foundation for subsequent analysis. This process effectively captures the multi-dimensional characteristics of battery performance over time. However, in actual operating scenarios, such as the high-frequency charging and discharging conditions of urban buses and shared travel, the historical state vector library is inevitably mixed with noise and outlier curves caused by factors such as sudden temperature changes and differences in driving habits. If these abnormal data are not processed, they will interfere with the accuracy of subsequent clustering analysis, causing the classification boundary to shift, thereby weakening the stability of the SOH decay trend judgment. Therefore, step S2 focuses on screening steady-state centroids from the historical state vector library and eliminating outlier curves to construct a robust initial centroid set, providing a reliable topological anchor point for subsequent adaptive radius clustering.

[0034] The complete processing of step S2 aims to filter out the core points representing the steady-state operation of the battery from the historical state vector library and eliminate abnormal data, thereby constructing a robust initial centroid set, providing a stable topological anchor point for subsequent adaptive radius clustering. The following is a detailed expansion of the specific processing logic: S2-1, references the historical state vector library; The construction of the historical state vector library is based on the timing buffer generated in step S1. A set of multiple multi-scale state vectors is extracted from the timing buffer as the basic data for analysis in this step. Each multi-scale state vector is composed of characteristic coefficients of voltage, current and temperature at multiple wavelet decomposition levels. These coefficients can reflect the performance status of the battery at a specific time point. All vectors are arranged in chronological order, fully recording the changes in the characteristics of the battery during long-term operation. The purpose of citing this historical state vector library is to use comprehensive information on the evolution of battery performance over time to provide data support for screening core points representing steady-state operation. This approach ensures that the analysis process can capture the long-term trends and typical patterns of battery operation.

[0035] S2-2, using density threshold screening to locate the steady-state centroid; In order to identify the core points reflecting the steady-state operation of the battery from the historical state vector library, it is first necessary to calculate the local density of each state vector. The local density is determined by counting the number of other vectors within a specific range centered on a certain state vector. The specific process is: first calculate the pairwise Euclidean distances between all vectors in the historical state vector library, and take the average of these distances as the radius of the hypersphere. This radius reflects the typical scale of the data distribution. With each state vector as the center, count the number of vectors contained in the range defined by this radius, and the resulting value is the local density of the state vector. Vectors with higher local density indicate that the data in their area is more concentrated, which usually corresponds to the typical operating state of the battery.

[0036] After calculating the local density of all state vectors, the median of these local densities is determined. The median is used as a baseline value and is amplified by multiplying it by an adjustment coefficient greater than 1 to generate a density threshold. The selection of the adjustment coefficient is based on the data distribution characteristics and is usually set to a fixed value to ensure the representativeness of the screening results. State vectors with local density greater than or equal to this density threshold are selected to form a set of steady-state centroid candidates. This screening method can highlight the areas in the data set that reflect the steady-state operation of the battery while avoiding interference from outliers or sparse data, thereby providing reliable core point candidates for subsequent analysis.

[0037] S2-3, remove outlier curves; After obtaining the set of candidate steady-state centroids, it is necessary to identify and remove anomalous data from the historical state vector library to improve the robustness of the initial centroid set. The determination of anomalous data is based on the distance between each state vector and the nearest vector in the candidate steady-state centroid set. Specifically, for each vector in the historical state vector library, calculate its Euclidean distance to all vectors in the candidate steady-state centroid set and record the minimum distance. Vectors with greater distances may represent abnormal battery operation or noisy data.

[0038] To set the criteria for removing anomalous data, all calculated minimum distance values ​​are collected, and the upper and lower quartiles of these distances are determined. The upper quartile plus 1.5 times the interquartile range is used as the outlier threshold. This threshold design is based on the outlier detection method commonly used in statistics and can effectively identify data that deviates from the normal range. State vectors with a minimum distance greater than this outlier threshold are removed from the historical state vector library, resulting in a new dataset that has been depleted of anomalous data. This distance analysis can filter out data that deviates from the core points of steady-state operation, reduce noise interference, and thus improve the accuracy of subsequent processing.

[0039] S2-4, forming an initial centroid set; In the historical state vector library after removing abnormal data, the core points representing steady-state operation are relocated to form an initial centroid set, which provides a topological anchor point for subsequent adaptive radius clustering. The specific process is to first recalculate the local density of each state vector for the data set after removing abnormal data. The calculation method is the same as before, that is, the average value of the Euclidean distance between all vectors is taken as the new hypersphere radius, and the number of other vectors contained in each vector within this range is counted to obtain the local density. Then, for each state vector, find the closest one among the vectors with a higher local density than itself, and record this distance as the minimum distance. For the state vector with the highest local density, its minimum distance is defined as the maximum distance between all vectors in the data set.

[0040] Next, the product of the local density of each state vector and its minimum distance is calculated. This product comprehensively reflects the concentration of the vector and its separation from other high-density areas. Based on a preset number of clusters, the vectors with the largest product values ​​are selected from all state vectors as the initial centroids, forming the initial centroid set. This selection method ensures that the initial centroids are both located in data-dense areas and representative in spatial distribution, providing a stable starting point for subsequent cluster analysis.

[0041] Step S2 references a historical state vector library containing long-term battery operation information, uses local density to filter out a set of candidate steady-state centroids, removes outliers based on distance analysis, and ultimately forms an initial set of centroids within the purified dataset using the density peak concept. This process effectively filters out the influence of noise and outliers, constructing a robust initial set of centroids that provides a stable topological anchor for subsequent adaptive radius clustering and ensures the reliability and accuracy of battery health decay trend assessment.

[0042] Step S2, based on the historical state vector library in the time series buffer, generates a steady-state initial centroid set through density threshold screening and eliminates outlier curves to ensure that the centroid can represent the typical operating state of the battery, providing a stable reference benchmark for classification. However, in actual operating scenarios, the battery operating state is affected by various factors such as temperature fluctuations and differences in driving habits, and exhibits dynamic changes. Traditional clustering methods with fixed radius or preset categories are difficult to adapt to this complexity. Therefore, step S3 uses the initial centroid set as the topological anchor point and introduces an adaptive radius clustering mechanism to accurately classify the current state vector and generate a classification index chain to support the continuity and reliability of subsequent attenuation trend analysis.

[0043] S3-1, reference to the initial centroid set; The initial centroid set is derived from the calculation results of step S2 and contains multiple centroid vectors, each of which represents a typical state of the battery under steady-state operation. These centroid vectors are used as starting reference points for cluster analysis in step S3. The purpose of using these centroid vectors is to allow the clustering process to start from the typical operating mode of the battery, avoiding the deviation of the classification results from the actual situation due to the random selection of the starting point. The reference process of the initial centroid set is to load the centroid vector generated in step S2 directly into the processing module of step S3 as a reference point for subsequent distance calculation and classification. In this way, the classification process can be carried out based on a stable reference point, thereby ensuring the reliability of the results.

[0044] S3-2, calculate the distance between the current state vector and the center of mass vector; After extracting the current state vector from the timing buffer, the Euclidean distance between the current state vector and each centroid vector in the initial centroid set must be calculated. The calculation process first subtracts each component of the current state vector from each component of the corresponding centroid vector, obtaining the difference between each component. The squares of all differences are then added together to obtain the sum of the squares. Finally, the square root of the sum of the squares is taken to obtain the Euclidean distance between the current state vector and the corresponding centroid vector. This calculation is repeated for each centroid vector in the initial centroid set, generating a set of distance values. This distance calculation method intuitively reflects the degree of similarity between the current state vector and each representative state, providing a quantitative basis for classification. By comprehensively calculating the distance between each pair of vectors, the classification process is based on precise numerical comparisons, ensuring the objectivity of the results.

[0045] S3-3, determine the adaptive clustering radius; In order to adapt to the dynamic changes in the battery's operating state, the adaptive clustering radius needs to be dynamically determined based on the distance distribution characteristics between the current state vector and the centroid vector. The specific calculation process is: first, find the minimum value and median from the Euclidean distance between the current state vector and all centroid vectors; then, use a preset adjustment coefficient to weightedly sum the minimum value and median to obtain the adaptive clustering radius. The specific method of weighted summation is to multiply the minimum value by the first part of the adjustment coefficient, multiply the median by the second part of the adjustment coefficient, and then add the two together. The setting of the adjustment coefficient is based on empirical values ​​and is used to balance the influence of local distance and global distance, so that the clustering radius can be flexibly adjusted according to the distribution characteristics of the current state vector. This dynamic adjustment mechanism can effectively cope with state fluctuations caused by changes in external conditions during battery operation, thereby making the classification process more adaptive.

[0046] S3-4, perform adaptive radius clustering; Using the initial centroid set as the topological anchor point, the current state vector is classified based on the adaptive clustering radius. The classification process first finds the minimum value from the Euclidean distance between the current state vector and all centroid vectors and determines the corresponding nearest centroid vector; then compares this minimum value with the adaptive clustering radius; if the minimum value is less than or equal to the adaptive clustering radius, the current state vector is assigned to the category represented by the nearest centroid vector; if the minimum value is greater than the adaptive clustering radius, the current state vector is regarded as a new category, and the current state vector itself is used as the centroid vector corresponding to the new category and added to the initial centroid set. Through this classification rule, the clustering process can not only classify the current state vector into existing typical states, but also identify new operating modes. This mechanism improves the flexibility of classification, enabling it to accurately capture the diverse changes in battery status.

[0047] S3-5, output classification index chain; The classification results of the current state vector are recorded in the classification index chain. The classification index chain is a data structure that stores the classification results of all state vectors in chronological order. The recording process is as follows: Based on the aforementioned classification results, the category number to which the current state vector belongs is determined and this number is appended to the end of the classification index chain. By continuously recording the classification results at each time point, the classification index chain fully presents the classification trajectory of the battery state over time. This time-series classification information provides a continuous reference for subsequent analysis, supporting in-depth research on the changing patterns of battery health status.

[0048] Step S3 accurately classifies the battery state by referencing the initial set of centroids, calculating the Euclidean distance between the current state vector and the centroid vector, determining the adaptive clustering radius, performing adaptive radius clustering, and outputting a classification index chain. This process effectively addresses external factors such as sudden temperature changes and differences in driving habits in complex scenarios such as high-frequency charging and discharging, ensuring that the classification results truly reflect the evolving characteristics of the battery's operating state. The generation of the classification index chain provides stable time series information support for subsequent analysis, significantly enhancing the robustness and reliability of the entire method.

[0049] The decay trend of a battery's state of health (SOH) is affected by noise such as sudden temperature changes and differences in driving habits. While the classification index chain can reflect state evolution, it struggles to directly represent continuous decay trends and is susceptible to fluctuations due to local disturbances. Step S4, based on this, converts the classification index chain into a continuous representation that reflects the SOH decay trend. Through feature extraction and dynamic adjustment, a stable state of health probability curve is generated to support operation and maintenance decisions.

[0050] S4-1, aggregate time series to generate decay potential energy diagram; To analyze battery health decay trends, the first step is to convert the chain of classification indices in the time series data into a continuous curve that reflects the time-varying capacity characteristics. Specifically, the classification index corresponding to each time point is extracted from the historical data. This index represents the battery's category at that point. Next, for each time point, all time points belonging to the same category in the historical data are retrieved and the corresponding capacity decay values ​​are obtained. Capacity decay refers to the reduction in battery capacity relative to its initial capacity at a specific time point. The capacity decay values ​​corresponding to these historical time points are then statistically processed by calculating the average of all these decay values ​​and defining this average as the decay potential energy value at the current time point. By repeating this process for each time point, a continuous curve consisting of the decay potential energy values ​​at each time point is generated, called a decay potential energy graph. This method leverages the statistical properties of historical data and effectively reduces random fluctuations in the data at individual time points, resulting in a smoother and more stable curve.

[0051] S4-2, calculate the attenuation phase skew; To analyze the overall tilt trend of the attenuation potential energy graph, a method that can decompose the signal's features at different scales is required. The continuous wavelet transform (CWT) was chosen for this step. By decomposing the attenuation potential energy graph into phase information at multiple scales, it captures the macroscopic variation trend of the curve. The specific calculation process first applies the CWT to the attenuation potential energy graph to obtain transform coefficients at different scales, with each coefficient corresponding to the energy value of a scale. From these scales, the main scale with the highest energy is selected, and the phase spectrum at that scale is extracted. The phase spectrum is a sequence representing the time-varying phase variation of the signal at that scale. Next, the phase spectrum values ​​are mapped to a standard range, adjusting the range from minimum to maximum to between 0 and 1. Subsequently, a uniformly distributed reference sequence is constructed, whose values ​​are linearly distributed within the same range. The absolute difference between the phase spectrum and this reference sequence is calculated at each corresponding point. These absolute differences are integrated over the entire time series to obtain a total difference, which is then divided by the length of the time series to obtain the attenuation phase skewness. This indicator quantifies the overall skew characteristics of the attenuation potential energy map on the main scale, providing a numerical basis reflecting the global characteristics for subsequent trend adjustments, enabling the processing process to adapt to trend changes in different time spans.

[0052] S4-3, calculate the local entropy dispersion; To further characterize the local fluctuation characteristics of the decay potential energy graph, a metric that can measure data uncertainty is needed. Information entropy is used in this step, and the decay potential energy graph is analyzed segmentally using a sliding window approach. The calculation process begins by setting the sliding window width and step size. The width determines the number of time points included in each window, and the step size determines the distance between windows. The decay potential energy graph is divided into multiple subsequences based on the set window width and step size. These subsequences may or may not overlap depending on the step size. For each subsequence, the distribution of decay potential energy values ​​within the subsequence is statistically analyzed, that is, the frequency of occurrence of each value is calculated as a probability. Based on this probability distribution, the information entropy of the subsequence is calculated. The information entropy is calculated by taking the negative logarithm of each probability value, multiplying it by the probability value, and summing all the results. This process is repeated to obtain the information entropy value for each window, forming an entropy sequence. Next, the average of the entropy sequence is calculated, and the absolute difference between each entropy value and the average is summed. This sum is then divided by the length of the entropy sequence to obtain the average deviation value. This average deviation is normalized by dividing it by the maximum possible deviation of the entropy sequence to obtain the local entropy dispersion. This indicator reflects the degree of dispersion of the decay potential energy map within a local range, allowing subsequent processing to be optimized and adjusted based on the microscopic fluctuation characteristics.

[0053] S4-4, trend adjustment factor obtained through machine learning reasoning; When dynamically adjusting subsequent filtering processes, it's necessary to comprehensively consider both the overall trend and local characteristics of the decay potential energy graph. Therefore, a lightweight gradient boosting decision maker is introduced for inference. Specifically, the decay phase skew and local entropy dispersion calculated in the previous steps are used as input features. The lightweight gradient boosting decision maker is a decision tree-based machine learning model that iteratively constructs multiple weak predictors and combines them into a strong predictor, capable of capturing the complex relationship between input features and output values. During the training phase, the model is trained using known decay phase skew and local entropy dispersion feature pairs from historical data, along with the corresponding optimal trend adjustment factor values. The goal of training is to ensure that the trend adjustment factor output by the model ensures that the subsequent filtering results closely match the true capacity decay curve. In real-time processing, the decay phase skew and local entropy dispersion calculated for the current time period are input into the trained model, which then outputs a trend adjustment factor based on the input features. This factor integrates the characteristics of both macro-trends and micro-fluctuations, enabling the subsequent filtering process to adaptively adjust to the data characteristics, improving the applicability of the processing results to different decay patterns.

[0054] S4-5, adjust the filter weight and generate the health status probability curve; To generate the final battery state-of-health curve, the decay potential energy map needs to be smoothed and converted into a probabilistic form. A Kalman filter is selected for this step, controlling the degree of smoothing by adjusting the process noise covariance. Specifically, the trend adjustment factor obtained in the previous step is applied to the Kalman filter to dynamically adjust the process noise covariance. The process noise covariance determines the filter's sensitivity to data changes. By increasing or decreasing this value, the trend adjustment factor achieves a balance between smoothness and filter response speed. The adjusted Kalman filter processes the decay potential energy map and outputs a smoothed capacity decay trajectory. Next, the maximum decay value corresponding to complete battery failure is obtained. This value is typically determined based on battery design parameters or historical data. Each decay value in the smoothed capacity decay trajectory is divided by the maximum decay value to obtain a ratio. This ratio is then subtracted from 1 to generate a health probability value ranging from 0 to 1. This calculation is repeated for each time point to form the health probability curve. This curve smoothly reflects the changing trend of battery health status over time. Its continuity and accuracy provide a reliable reference for operation and maintenance decisions, while adapting to data fluctuations in complex usage scenarios.

[0055] Through the above steps, starting with generating a decay potential energy graph from aggregated time series, the decay phase skewness and local entropy dispersion are gradually calculated. Machine learning is then used to infer the trend adjustment factor, which is used to adjust the filter weights, ultimately generating a health state probability curve. This entire process accurately tracks battery capacity decay trends. Under complex conditions such as high-frequency charging and discharging, this method effectively reduces the impact of noise on the results by comprehensively analyzing overall trends and local characteristics, ensuring the stability and credibility of the health state probability curve, and providing solid data support for the optimized operation of battery management systems.

[0056] Step S4 aggregates the time series to generate a decay potential energy graph and adjusts the filter weights through machine learning reasoning, ultimately forming a health state probability curve. This curve can reflect the evolution trend of SOH, but due to abnormal operating conditions such as noise, temperature mutations, or differences in driving habits, local sections of the curve may be unreliable, resulting in trend judgment deviations. To solve this problem, step S5 focuses on monitoring the curvature changes of the health state probability curve, identifying sections of mutation events and triggering full charge verification, while updating the clustering benchmark to ensure the stability and accuracy of trend judgment. However, frequent full charge verification is difficult to achieve in actual operation, so step S5 needs to design an intelligent mechanism to trigger verification only when necessary, and adapt to the long-term evolution of the battery state by dynamically adjusting the clustering model.

[0057] The complete processing of step S5 is intended to monitor the curvature change of the health status probability curve, identify and process the sections corresponding to the mutation events, and ensure the accuracy and reliability of the judgment of the health status attenuation trend of the electric vehicle battery.

[0058] S5-1, calculate the curvature of the health state probability curve; The health probability curve reflects the temporal trend of the battery's health status. To identify possible abnormal fluctuations, the curvature of the curve must be calculated to characterize the degree of curvature at each time point. Curvature reflects the geometric properties of the curve, particularly exhibiting higher values ​​during sharp changes, thus providing a basis for identifying unreliable sections. The calculation process begins by numerically differentiating the health probability curve to obtain first- and second-order derivatives. The first-order derivative is calculated by taking the difference between the health probability value at each time point and the previous time point, dividing it by the time interval between the two time points to obtain the rate of change of the health probability over time. The second-order derivative is calculated by applying the same process to the first-order derivative sequence: taking the difference between the first-order derivative value at each time point and the previous time point, dividing it by the corresponding time interval to obtain the acceleration of the rate of change. Curvature is calculated by dividing the absolute value of the second-order derivative by an adjustment term that includes the first-order derivative. This adjustment term is a function whose value is one plus the square root of the sum of the squares of all first-order derivatives. This ensures that the curvature accurately reflects the degree of curvature whether the curve is flat or sharply curved. In this way, the curvature value can accurately quantify the local variation characteristics of the curve, providing an objective numerical basis for subsequent analysis while avoiding excessive sensitivity to noise in the data.

[0059] S5-2, detection of sudden curvature events; After obtaining the curvature sequence of the health probability curve, it is necessary to detect sudden changes in the sequence to flag segments that may render trend determination unreliable. A sudden change is characterized by a sharp increase in the curvature value within a short period of time, typically corresponding to abnormal fluctuations in the health probability curve. To achieve this, the detection process first identifies local peaks in the curvature sequence. Specifically, the curvature value at each time point is compared with the curvature values ​​at the previous and next time points. If the curvature value at that time point is higher than both the previous and next points, it is considered a local peak. Next, a curvature threshold is set to filter out significant sudden changes. The curvature threshold is determined based on statistical analysis of historical data and is typically several times the average value of the curvature sequence to ensure that only time points with significant changes are captured. When the curvature value of a local peak exceeds this threshold, that time point is marked as a sudden change. During the detection process, not only the absolute magnitude of the curvature value is considered, but also the temporal trend of the curvature value is considered to ensure that the marked sudden change event truly reflects abnormal fluctuations in the health probability curve. This method can quickly identify unreliable segments and improves detection accuracy through the combined judgment of threshold and trend.

[0060] S5-3, trigger full charge verification; When a sudden change in curvature is detected, a full-charge calibration is required to correct the segment of the health probability curve marked as a sudden change. A full-charge calibration performs a complete charge-discharge cycle to obtain accurate battery capacity data to correct the abnormal portion of the curve. The conditions for triggering a full-charge calibration are first confirming the presence of a sudden change and then checking whether a full-charge calibration has already been performed within a preset time window around the time of the sudden change. The length of this time window is set based on actual operational needs, typically several hours or days, to avoid overly frequent calibration. Specifically, for each time point marked as a sudden change, calibration records within the time window are retrieved. If no calibration record exists within the time window, a full-charge calibration is initiated, performing a complete charge-discharge cycle. Upon completion, the calibration time and measured battery capacity data are recorded. The resulting capacity data is used to adjust the health probability curve near the sudden change, ensuring that the curve more closely reflects the actual battery condition. This conditional triggering mechanism reduces unnecessary calibration operations, balances the need for data correction with the rational use of operational resources, and ensures improved curve accuracy.

[0061] S5-4, write-back new centroid update clustering benchmark; After the full-charge verification is complete, the clustering model's initial centroid set is updated using the acquired accurate capacity data to adapt to the long-term evolution of the battery's state of health. The update process first converts the battery state data generated by the full-charge verification into a multi-scale state vector, which contains multi-dimensional information such as capacity and time. This state vector is then added to the historical state vector library. Next, the density threshold screening and outlier removal processes used during the initial clustering are re-implemented. This involves performing a density analysis on all vectors in the historical state vector library, retaining vectors with higher density as the clustering baseline, and removing outliers that deviate from the main trend. This process generates a new initial centroid set, which replaces the original one. This new initial centroid set reflects the battery's current true state, ensuring that the clustering model remains consistent with the battery's aging process. By dynamically incorporating the latest data, this update mechanism makes the clustering results more reliable and enhances the model's adaptability to battery state changes, providing a stable baseline for subsequent state of health analysis.

[0062] Step S5 identifies and corrects the sections with sudden changes by calculating the curvature of the health probability curve, detecting sudden changes in curvature, triggering a full charge check, and writing back the new centroid to update the clustering benchmark. This ensures the continuity and accuracy of determining the health decay trend of electric vehicle batteries. In complex scenarios such as high-frequency charging and discharging, this process effectively addresses the impact of noise and abnormal operating conditions through intelligent triggering mechanisms and dynamically updated model design, providing a stable and reliable decision-making basis for battery operation and maintenance.

[0063] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0064] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0065] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0066] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0067] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for judging the state of charge (SOH) decay trend of an electric vehicle battery, characterized in that: Including steps: S1: Collect the charge and discharge segment curves one by one and map them into a multi-scale state vector set and write them into the timing buffer; S2: Reference the historical state vector library in the time series buffer, use density threshold screening to locate the steady-state centroid and remove outlier curves to form the initial centroid set; S3: Perform adaptive radius clustering with the initial centroid set as the topological anchor point, and output the classification index chain for the current state vector; S4: Generates a decay potential energy diagram based on the classified index chain aggregation time series, extracts macro and micro characterization features, and obtains trend adjustment factors through machine learning reasoning. The trend adjustment factors are then used to adjust the filter weights to track the capacity decay trajectory and generate a health state probability curve; S5: Use the probability curve curvature mutation event as a trigger to mark the mutation event and guide the full verification, and write back the new centroid to update the clustering benchmark.

2. The method for judging the SOH decay trend of an electric vehicle battery according to claim 1, characterized in that: Step S1 includes the following contents: The voltage, current and temperature data in the charge and discharge segment curves are subjected to multi-scale wavelet decomposition respectively to generate multiple levels of low-frequency approximate coefficients and high-frequency detail coefficients. The low-frequency approximate coefficients and high-frequency detail coefficients of the voltage, current and temperature are spliced ​​in a predetermined order to form a multi-scale state vector. The decomposition level of the multi-scale wavelet decomposition is 3 to 5 levels, and the Daubechies wavelet basis is used. Writing into the timing buffer includes attaching the timestamp of the acquisition start time to the multi-scale state vector and writing it into the timing buffer in chronological order. When the buffer is full, the multi-scale state vector with the earliest timestamp is removed and a new multi-scale state vector is inserted.

3. The method for judging the SOH decay trend of an electric vehicle battery according to claim 2, characterized in that: Step S2 includes the following contents: Retrieve a historical state vector library from a timing buffer; calculate the local density of each state vector in the historical state vector library by counting the number of state vectors within a predetermined radius; multiply the median of the local density by a predetermined coefficient as a density threshold; select state vectors whose local density is greater than or equal to the density threshold as candidate centroids; for each state vector in the historical state vector library, calculate its distance to the nearest candidate centroid; remove state vectors whose distance is greater than the outlier threshold to obtain a refined historical state vector library.

4. The method for judging the SOH decay trend of an electric vehicle battery according to claim 3, characterized in that: Step S2 also includes the following: Recalculate the local density for each state vector in the refined historical state vector library; for each state vector in the refined historical state vector library, determine its minimum distance to another state vector with higher local density; calculate the product of the local density and the minimum distance for each state vector in the refined historical state vector library; and select a predetermined number of state vectors with the highest product as the initial centroid set.

5. The method for judging the SOH decay trend of an electric vehicle battery according to claim 4, characterized in that: Step S3 includes the following contents: The initial centroid set is used as the topological anchor point; The Euclidean distance between the current state vector and each centroid vector in the initial centroid set is calculated; and the adaptive clustering radius is determined by combining the minimum value and the median of these Euclidean distances according to a predetermined weighting factor.

6. The method for judging the SOH decay trend of an electric vehicle battery according to claim 5, characterized in that: Step S3 also Includes the following: Perform adaptive radius clustering, specifically identifying the centroid vector with the minimum Euclidean distance to the current state vector, and comparing this minimum Euclidean distance with the adaptive clustering radius. If the minimum Euclidean distance is less than or equal to the adaptive clustering radius, the current state vector is assigned to the cluster represented by the corresponding centroid vector. If the minimum Euclidean distance is greater than the adaptive clustering radius, a new cluster is created with the current state vector as the centroid vector. Finally, the classification results are recorded in the classification index chain, which stores the classification indexes of all state vectors in chronological order.

7. The method for judging the state of charge (SOH) decay trend of an electric vehicle battery according to claim 6, characterized in that: Step S4 includes the following contents: The attenuation potential energy map is generated based on the classified index chain aggregation time series. The attenuation phase skewness is calculated by performing multi-scale wavelet decomposition on the attenuation potential energy map and extracting the main scale phase spectrum. At the same time, the attenuation potential energy map is divided in a sliding window manner and the window information entropy is calculated. The relative uniformity distribution of the entropy sequence is obtained to calculate the local entropy dispersion.

8. The method for judging the state of charge (SOH) decay trend of an electric vehicle battery according to claim 7, characterized in that: Step S4 also includes the following: The attenuation phase skewness and local entropy dispersion are input into the lightweight gradient boosting decision maker to output the trend adjustment factor. Finally, the trend adjustment factor is used to adjust the filter weight to track the capacity decay trajectory and generate a health state probability curve.

9. The method for judging the state of charge (SOH) decay trend of an electric vehicle battery according to claim 8, characterized in that: Step S5 includes the following contents: Numerical differentiation is performed on the health state probability curve to calculate the first and second derivatives to determine the curvature, and the local peak points in the curvature sequence are identified and compared with a preset threshold to detect mutation events.

10. The method for judging the state of charge (SOH) decay trend of an electric vehicle battery according to claim 9, characterized in that: Step S5 also Includes the following: When a mutation event is detected and there is no full charge verification record within the preset time window, a full charge verification is triggered to obtain accurate battery capacity data. The battery status data generated by the full charge verification is used to update the historical state vector library, and a new initial centroid set is generated through density threshold screening and outlier curve removal.

Images