Fault detection method and system for mobile energy storage charging pile

Abstract

The invention relates to the technical field of mobile energy storage charging piles, and discloses a fault detection method and system for a mobile energy storage charging pile, and the method comprises the steps: collecting a voltage and current original sequence during an operation period, and executing the Kalman filtering to obtain a smooth electrical parameter; adaptively adjusting a process noise covariance matrix by using numerical dispersion to obtain a stable internal resistance value sequence; extracting a fluctuation trend and matching a working condition mode to filter periodic disturbance to obtain a pure internal resistance offset sequence; extracting a health degradation rate slope, establishing a dynamic health reference range by using kernel density estimation, and judging an abnormal migration event according to the dynamic health reference range; the offset severity is calculated through severity probability weighted fusion, the fault early warning level is obtained, and an event trigger counter reset mechanism is driven to judge the overall health state. The method can solve the problem of low fault detection accuracy in the prior art.

Description

Technical Field

[0001] This invention relates to the field of mobile energy storage charging pile technology, and in particular to a fault detection method and system for mobile energy storage charging piles. Background Technology

[0002] Currently, mobile energy storage charging piles, as core equipment for flexible power dispatch, directly impact the safety and power supply quality of the power system due to their operational stability. With the increasing scale of equipment deployment, charging piles frequently operate in high-power cyclic charging and discharging states, making the health evolution of internal power battery packs and power devices increasingly complex. To ensure the safe operation of equipment under dynamic loads, it is crucial to implement high-precision fault prediction and health management for the key physical parameters of charging piles, enabling early detection and risk assessment of internal hidden faults.

[0003] In existing technologies, fault detection in mobile energy storage charging piles primarily relies on fixed threshold alarms or basic voltage and current monitoring schemes. The system typically uses sensors to collect real-time electrical parameters and compares them with preset static judgment standards. When the parameter amplitude exceeds the range, a protection action is triggered. However, during actual operation, sensitive indicators reflecting aging and faults, such as the internal resistance of the charging pile's battery, are coupled with multiple dynamic operating conditions, including temperature field distribution, state of charge (SOC), and charge / discharge rate, exhibiting strong nonlinear time-varying characteristics. Traditional static monitoring schemes struggle to distinguish between instantaneous parameter fluctuations caused by normal charge / discharge switching and actual physical deviations caused by loose connections, internal short circuits, or the early stages of thermal runaway. When background noise from operating condition disturbances is superimposed on weak fault characteristics, traditional algorithms, lacking deep feature extraction and denoising capabilities, easily lead to a severe disconnect between warning signals and the actual health status of the equipment.

[0004] Therefore, existing technologies suffer from low fault detection accuracy. Summary of the Invention

[0005] This invention provides a fault detection method and system for mobile energy storage charging piles to solve the problem of low fault detection accuracy in existing technologies.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a fault detection method for mobile energy storage charging piles, comprising: The original voltage and current sequences during the operation of the charging pile are collected. Based on the original voltage and current sequences, the system state equation and observation equation are established, and Kalman filtering is performed to obtain a smooth electrical parameter sequence. The internal resistance observation sequence is calculated based on the smoothed electrical parameter sequence. The preset process noise covariance matrix is ​​adaptively adjusted using the numerical dispersion of the internal resistance observation sequence. A stable internal resistance value sequence is obtained through point-by-point iterative correction. A point-by-point difference operation is performed on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and an adaptive Bayesian fusion processing based on the operating condition is performed on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. An observation residual sequence is generated based on the matching results between the internal resistance fluctuation trend sequence and the preset operating condition mode library. An observation noise covariance matrix is ​​constructed based on the observation residual sequence, and periodic disturbances caused by operating conditions are filtered out to obtain a pure internal resistance offset sequence. The cumulative statistical value of the offset and the slope of the health degradation rate are extracted from the pure internal resistance offset sequence, and a dynamic health benchmark range is established by kernel density estimation in combination with a preset set of historical degradation rate slopes. If the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark, then an abnormal offset persistence determination process is performed to determine the abnormal offset event. Based on the abnormal offset events, aging factors are extracted, and offset severity values ​​are calculated through fuzzy membership degree weighted fusion to map and determine the fault warning level. High-risk signal sequences are selected from the fault warning levels. The event trigger counter reset mechanism is driven by the high-risk signal sequences to generate warning instructions and call the preset health assessment model to determine the overall health status.

[0007] Secondly, the present invention provides a fault detection system for mobile energy storage charging piles, comprising: The filtering module is used to collect the original voltage and current sequences during the operation of the charging pile, establish the system state equation and observation equation based on the original voltage and current sequences, and perform Kalman filtering to obtain a smooth electrical parameter sequence. The internal resistance stabilization module is used to calculate the internal resistance observation sequence based on the smooth electrical parameter sequence, adaptively adjust the preset process noise covariance matrix using the numerical dispersion of the internal resistance observation sequence, and obtain a stable internal resistance value sequence through point-by-point iterative correction. The trend extraction module is used to perform point-by-point difference operation on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and to perform adaptive Bayesian fusion processing based on operating conditions on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. The disturbance filtering module is used to generate an observation residual sequence based on the matching result between the internal resistance fluctuation trend sequence and the preset operating condition mode library, construct an observation noise covariance matrix based on the observation residual sequence, and filter out the periodic disturbances caused by the operating conditions to obtain a pure internal resistance offset sequence. The benchmark construction module is used to extract the cumulative statistical value of the offset and the slope of the health degradation rate from the pure internal resistance offset sequence, and to establish a dynamic health benchmark range by combining the preset historical degradation rate slope set with kernel density estimation. An anomaly locking module is used to perform an anomaly offset persistence determination process to determine an anomaly offset event if the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark. The severity assessment module is used to extract aging factors based on the abnormal offset events, calculate the offset severity value through fuzzy membership weighted fusion, and map and determine the fault warning level. The status decision module is used to filter high-risk signal sequences from the fault warning levels, drive the event trigger counter reset mechanism according to the high-risk signal sequences, generate warning instructions, and call the preset health assessment model to determine the overall health status.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves effective filtering of periodic disturbances caused by operating conditions through operating condition mode matching and adaptive updating of the observation noise covariance matrix. This process can remove periodic interference components from the complex operating background and extract a pure internal resistance offset sequence that reflects the true performance changes of the battery, which significantly improves the identification efficiency of fault characteristics.

[0009] (2) This invention establishes a dynamic health benchmark range based on kernel density estimation and combines interquartile range rules and persistent statistics to determine abnormal offset events. This dynamic benchmark can adapt to the degradation characteristics of batteries at different life stages. Combined with the offset persistence confirmation mechanism, it effectively filters out instantaneous noise and false deviations caused by measurement jitter, thereby improving the robustness of health assessment.

[0010] (3) This invention integrates the accumulation of internal resistance offset, piecewise linear mapping of aging factors, and a severity probability weighted model to establish a hierarchical quantitative fault early warning system. Through the event-triggered counter reset mechanism and health assessment scoring, the system can reliably track continuously degrading battery cells, reducing false alarm rate while improving the accuracy of prediction and the scientific nature of decision-making regarding potential risks of mobile energy storage charging piles. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of a fault detection method for a mobile energy storage charging pile provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a fault detection system for a mobile energy storage charging pile provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a fault detection method for mobile energy storage charging piles, including the following steps: S11, Collect the original voltage and current sequences during the operation of the charging pile, establish the system state equation and observation equation based on the original voltage and current sequences, and perform Kalman filtering to obtain a smooth electrical parameter sequence; S12, calculate the internal resistance observation sequence based on the smoothed electrical parameter sequence, adaptively adjust the preset process noise covariance matrix using the numerical dispersion of the internal resistance observation sequence, and obtain a stable internal resistance value sequence through point-by-point iterative correction. S13, perform point-by-point difference operation on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and perform adaptive Bayesian fusion processing based on operating conditions on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. S14. Generate an observation residual sequence based on the matching result between the internal resistance fluctuation trend sequence and the preset working condition mode library. Construct an observation noise covariance matrix based on the observation residual sequence and filter out the periodic disturbances caused by the working condition to obtain a pure internal resistance offset sequence. S15, extract the cumulative statistical value of the offset and the slope of the health degradation rate from the pure internal resistance offset sequence, and establish a dynamic health benchmark range by combining the preset historical degradation rate slope set with kernel density estimation; S16, If the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark, then perform abnormal offset persistence determination processing to determine the abnormal offset event. S17. Extract the aging factor based on the abnormal offset event, and calculate the offset severity value through fuzzy membership degree weighted fusion to map and determine the fault warning level. S18, select high-risk signal sequences from the fault warning levels, drive the event trigger counter reset mechanism according to the high-risk signal sequences, generate warning instructions, and call the preset health assessment model to determine the overall health status.

[0014] In step S11, the original voltage and current sequences during the operation of the charging pile are collected. Based on the original voltage and current sequences, system state equations and observation equations are established, and Kalman filtering is performed to obtain a smooth electrical parameter sequence, including: Obtain the original voltage and current sequence of the charging pile port; The system state equation is established based on the original voltage and current sequence. The system state equation is then used to deduce the prior state estimation vector and the observation equation. Based on the observation equation, the Kalman filter gain is determined by combining the prior state estimation vector and the actual observation value, wherein the actual observation value is composed of the voltage sample value and the current sample value in the original voltage and current sequence. The prior state estimation vector is corrected using the Kalman filter gain to calculate the posterior state estimation vector; The corrected voltage and current components are extracted from the posterior state estimation vector to form the smoothed electrical parameter sequence.

[0015] First, the raw voltage and current sequences of the charging pile port are obtained. Specifically, high-precision sensors deployed at the charging pile output port synchronously collect electrical data during operation at a preset sampling frequency. The collected discrete data are aligned by timestamps to form raw voltage and current sequences containing voltage and current sample values. It is worth noting that the preset sampling frequency is determined by fully considering the operating characteristics of the high-frequency switching devices inside the charging pile and the Nyquist sampling theorem. Since the switching frequencies of the power conversion modules in common mobile energy storage charging piles are mostly concentrated in the 20kHz to 50kHz range, in order to effectively capture transient electrical distortions caused by high-frequency switching actions and prevent signal aliasing, the preset sampling frequency is set to 100kHz through experimental analysis. Those skilled in the art know that this frequency parameter can be dynamically adjusted within the range of 50kHz to 200kHz according to the power level of the specific charging pile and the actual switching frequency of the internal insulated gate bipolar transistor (IGBT) devices.

[0016] Secondly, the system state equations are established based on the original voltage and current sequences. These equations are then used to deduce the prior state estimation vector and the observation equations. Specifically, the current voltage and current are used as state variables to construct a discrete-time linear state transition model. The system state equations are expressed as follows: In the formula, Represents the current time step number. Let this be the state vector at the current moment. This is the state transition matrix, which is derived by discretization of the continuous-time state-space model of the charging pile output filter circuit. Specifically, it involves obtaining physical parameters such as the equivalent inductance and equivalent capacitance values ​​of the charging pile output filter, combined with the sampling period corresponding to the preset sampling frequency. By using the zero-order hold (ZOH) method or the bilinear transform method, the differential equations of a continuous system are transformed into discrete-time difference equations, from which the matrix can be calculated. The fixed coefficient values ​​in the data are used to accurately characterize the natural evolution of parameters of the charging pile under ideal conditions. This is the posterior state estimate vector from the previous time step. This represents the process noise vector. Based on the state information from the previous time step, the state transition matrix is ​​used to perform a forward time extrapolation, calculating the predicted state variables for the current time step, i.e., the prior state estimation vector. Simultaneously, observation equations are constructed to correlate the state variables with the actual measured values: In the formula, This is the vector of actual observed values ​​at the current moment, directly taken from the voltage and current sample values ​​at the corresponding time nodes in the original voltage and current sequence. The observation matrix depends on the physical measurement configuration of the sensor. Since the system state vector in this embodiment is set to include two-dimensional variables of voltage and current, and the high-precision sensor directly samples these two physical quantities synchronously and absolutely, there is no dimensional transformation or complex function mapping across physical quantities. Therefore, the observation matrix... It is fixed as a two-dimensional identity matrix (i.e., the main diagonal elements are 1s and the rest are 0s). matrix). The observation noise vector characterizes the sensor's inherent measurement error and high-frequency thermal noise in the electromagnetic environment.

[0017] Then, based on the observation equation, the Kalman filter gain is determined by combining the prior state estimation vector and the actual observed values, where the actual observed values ​​consist of voltage and current sampled values ​​from the original voltage and current sequences. Specifically, the gain calculation depends on the dynamic proportional relationship between the prior estimation error covariance matrix and the observation noise covariance matrix. The gain calculation formula is: In the formula, The calculated Kalman filter gain matrix is... The prior estimate error covariance matrix obtained through derivation. The superscript represents the preset observation noise covariance matrix. Represents the matrix transpose operation, superscript This represents the matrix inversion operation.

[0018] It is worth noting that the preset observation noise covariance matrix The sensor noise floor was determined through long-term statistical experiments under no-load and standard load conditions. Specifically, the charging pile was operated continuously for 2 hours at 50% of its rated power. High-frequency disturbance components of the steady-state voltage and current collected by the sensor during this period were extracted, and the statistical variances of the voltage and current disturbance components were calculated. The calculated voltage and current variances were then used as matrices. The main diagonal elements. This matrix parameter objectively reflects the true reliability of the measurement hardware environment, enabling the filtering algorithm to reasonably allocate the trust weights between actual observations and prior predictions when faced with drastic changes in external noise.

[0019] Next, the prior state estimation vector is corrected using the Kalman filter gain to calculate the posterior state estimation vector. Specifically, the residual between the actual observation and the prior state estimation vector is calculated, and this residual is multiplied by the Kalman filter gain and added as a compensation adjustment term to the prior state estimation vector. The state update formula is: In the formula, This is the updated posterior state estimation vector. This is the prior state estimation vector obtained through deduction. It is the deviation component between the actual observed value and the observed value predicted by the model (i.e., the predicted mapping value derived from the observation equation).

[0020] For example, if a voltage sensor experiences a sudden spike in voltage due to external electromagnetic interference at a certain moment, the calculated filter gain will automatically decrease due to the limitation of the preset observation noise covariance matrix. This makes the state update process more inclined to trust the prior prediction value, thereby effectively weakening the impact of spike bias on parameter estimation.

[0021] Finally, the corrected voltage and current components are extracted from the posterior state estimation vector to form a smooth electrical parameter sequence. Specifically, the above time extrapolation and observation update steps are repeated for each sampling time, continuously outputting the optimal posterior state estimation vector for each discrete time. The voltage components extracted from all times are concatenated according to the time series to form a smooth voltage sequence, and the extracted current components are concatenated according to the time series to form a smooth current sequence. The smooth voltage and smooth current sequences are then combined to form a complete smooth electrical parameter sequence.

[0022] In step S12, the internal resistance observation sequence is calculated based on the smoothed electrical parameter sequence. The preset process noise covariance matrix is ​​adaptively adjusted using the numerical dispersion of the internal resistance observation sequence. A stable internal resistance value sequence is obtained through point-by-point iterative correction, including: Extract the smoothed voltage and smoothed current components from the smoothed electrical parameter sequence, and perform mapping calculations using Ohm's law discrete form to obtain the internal resistance observation sequence; For the internal resistance observation sequence at the current sampling time, the numerical dispersion is calculated. If the numerical dispersion is greater than a preset stability threshold, an adaptive adjustment coefficient is determined based on the degree of deviation of the numerical dispersion from the preset stability threshold. The adaptive adjustment coefficient is then used to perform scaling processing on the preset process noise covariance matrix to obtain the adaptively adjusted process noise covariance matrix. The Kalman gain is calculated using the adaptively adjusted process noise covariance matrix to obtain the dynamic filter gain corresponding to the current sampling time in real time. The dynamic filtering gain is applied to perform a weighted correction on the deviation between the observed internal resistance sequence and the prior internal resistance estimate at the current sampling time to obtain the stable internal resistance value at the current sampling time. The stable internal resistance values ​​at each sampling time are then concatenated in chronological order to obtain the stable internal resistance value sequence. The prior internal resistance estimate at the current sampling time is obtained by performing forward time extrapolation processing on the stable internal resistance value at the previous sampling time.

[0023] First, the smoothed voltage and smoothed current components are extracted from the smoothed electrical parameter sequence. Ohm's law is then applied in a discrete form for mapping calculation to obtain the internal resistance observation sequence. Specifically, from the smoothed electrical parameter sequence generated by the previous filtering process, the smoothed voltage and smoothed current components are extracted synchronously according to discrete timestamps. For each discrete sampling time, the smoothed voltage component value is divided by the corresponding smoothed current component value, and the instantaneous internal resistance observation value at that time is calculated using Ohm's law. The instantaneous internal resistance observation values ​​from multiple consecutive sampling times are arranged sequentially in chronological order to construct the internal resistance observation sequence.

[0024] Secondly, the numerical dispersion of the internal resistance observation sequence at the current sampling time is calculated. Specifically, a fixed-length sliding time window is set, and the statistical standard deviation of all instantaneous internal resistance observations within this window is calculated. The calculated statistical standard deviation is used as the numerical dispersion characterizing the severity of the current internal resistance fluctuation. The sliding time window gradually moves forward as the sampling time progresses, and a numerical dispersion corresponding to the current sampling time is calculated and output for each step. The size of the fixed-length sliding time window is determined based on a combination of the data sampling frequency and the typical transient response time constant of the electrochemical reaction inside the battery. For example, selecting a window length containing 20 consecutive sampling points ensures that the standard deviation calculation has sufficient statistical significance while avoiding severe lag in capturing the abrupt changes in internal resistance characteristics due to an excessively large time span.

[0025] It is worth noting that the preset stability threshold is not randomly set, but determined through statistical analysis of internal resistance fluctuation data of standard fault-free battery packs of the same model during historical constant current charging tests. Specifically, an internal resistance observation sequence of the standard battery pack is collected under normal constant current conditions at rated power, and its cumulative standard deviation distribution function (CDF) is calculated within the same sliding time window. The standard deviation value corresponding to a cumulative probability of 95% is selected as the preset stability threshold. In one application scenario of this embodiment, the specific value of this stability threshold is... Those skilled in the art will understand that this threshold can be adjusted based on the specific characteristics of the battery material (such as lithium iron phosphate or ternary lithium) and the degree of aging. to The calibration adjustment is performed within the empirical range. If the calculated numerical dispersion is greater than the preset stability threshold, it indicates that the current internal resistance observation value has been subjected to a non-negligible external sudden disturbance or a switch between charging and discharging conditions has occurred.

[0026] If the numerical dispersion exceeds a preset stability threshold, the adaptively adjusted process noise covariance matrix is ​​determined. Specifically, an adaptive adjustment coefficient is constructed that is correlated with the current degree of numerical dispersion deviation. The formula for calculating the adaptive adjustment coefficient is as follows: In the formula, For adaptive adjustment coefficient, This represents the numerical dispersion calculated within the current sliding time window. The preset stability threshold, This is the preset amplification penalty factor. It's worth noting that the preset amplification penalty factor... Used to control the sensitivity of matrix amplification. This factor is determined by parameter optimization based on historical internal resistance jump data under different operating conditions, enabling the filter to quickly track real internal resistance jumps without causing severe oscillations due to excessive noise amplification. It typically takes a value between 1.0 and 2.0. For example, when... for , for and When the value is 1.05, the calculated adaptive adjustment coefficient is approximately 1.7.

[0027] Specifically, the determination of the amplification penalty factor adopts a combination of offline simulation and experimental calibration. By collecting internal resistance observation data of charging piles under different operating conditions, such as constant current charging, pulse load, and load switching, a dataset containing normal fluctuations and simulated fault mutations is constructed. Candidate factors are traversed in the range of 1.0 to 2.0 with a step size of 0.01, and an adaptive filtering algorithm is run to calculate the root mean square error between the stable internal resistance value after filtering and the true reference value. The factor value that minimizes the root mean square error and can quickly track the real mutation is selected as the final set value. In typical applications, this factor is taken as 1.05.

[0028] Subsequently, a scalar multiplication operation is performed between the preset process noise covariance matrix and the calculated adaptive adjustment coefficient. It is worth noting that the preset process noise covariance matrix is ​​determined based on the statistical error of the system modeling of the charging pile under standard constant current conditions. In specific implementation, the statistical covariance of the difference between the theoretical internal resistance value derived from the system state equation and the experimentally measured value is calculated by comparing them under interference-free conditions, and this covariance is pre-stored in the storage unit as a static reference parameter. Specifically, the adaptively adjusted process noise covariance matrix is ​​obtained by proportionally multiplying the adaptive adjustment coefficient by each element in the preset process noise covariance matrix, particularly by proportionally enlarging its main diagonal elements.

[0029] Next, the Kalman gain is calculated using the adaptively adjusted process noise covariance matrix to determine the dynamic filter gain corresponding to the current sampling time in real time. Specifically, at each discrete sampling time, the adaptively amplified process noise covariance matrix is ​​substituted into the derivation formula of the prior estimation error covariance matrix. Based on the updated prior estimation error covariance matrix, combined with the current observation matrix and observation noise covariance matrix, the Kalman filter gain coefficient corresponding to the current sampling time is calculated in real time.

[0030] Finally, the dynamic filtering gain is applied to weighted correct the deviation between the observed internal resistance sequence and the prior internal resistance estimate at the current sampling time, obtaining the stable internal resistance value at the current sampling time. The stable internal resistance values ​​from each sampling time are then concatenated chronologically to obtain the stable internal resistance value sequence. Specifically, the prior internal resistance estimate at the current sampling time is obtained by performing forward time extrapolation processing on the stable internal resistance value at the previous sampling time. An internal resistance state update formula is established: In the formula, This is the stable internal resistance value obtained after updating at the current sampling time. This is the prior internal resistance estimate obtained by performing forward time extrapolation on the stable internal resistance value at the previous sampling time. This represents the internal resistance observation sequence value extracted at the current sampling time. This corresponds to the dynamic filter gain coefficient calculated at the current sampling time. The stable internal resistance values ​​at each discrete time after iterative updates are concatenated according to the time flow to obtain the final stable internal resistance value sequence.

[0031] For example, when the internal resistance dispersion increases significantly, causing the process noise covariance matrix to be adaptively amplified, the value of the dynamic filter gain coefficient increases accordingly, which increases the weighting of the actual observation deviation allocated to the state update process, thereby accelerating the algorithm's response speed to the true abrupt change trend of internal resistance. Conversely, in the stable stage where the dispersion is below the stability threshold, the gain automatically decreases, and the update process relies more on the prior internal resistance estimate, thereby outputting a highly smooth and stable internal resistance value sequence, effectively suppressing high-frequency random disturbances.

[0032] In step S13, a point-by-point difference operation is performed on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and an adaptive Bayesian fusion processing based on operating conditions is performed on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence, including: Perform point-by-point difference operation on the stable internal resistance value sequence to obtain the internal resistance gradient vector sequence; The internal resistance gradient vector sequence is input into a preset single-step trend evolution model to perform forward extrapolation calculations, thereby obtaining a priori estimation sequence of fluctuations. The deviation component between the internal resistance gradient vector sequence and the fluctuation prior estimation sequence is calculated, and the Bayesian adjustment weight is calculated using the preset inference variance and observation variance. The deviation component is compensated by the Bayesian adjustment weight to generate the fluctuation posterior estimation sequence. Based on the current operating conditions, a fusion weight matrix is ​​extracted from a preset mapping table. The fusion weight matrix is ​​then used to perform a weighted summation on the prior estimation sequence of the fluctuation and the posterior estimation sequence of the fluctuation to determine the internal resistance fluctuation trend sequence.

[0033] First, a point-by-point difference operation is performed on the stable internal resistance value sequence to obtain the internal resistance gradient vector sequence. Specifically, the instantaneous internal resistance values ​​of two adjacent sampling points in the stable internal resistance value sequence are read sequentially in discrete time order. The algebraic difference obtained by subtracting the internal resistance value at the previous sampling time from the current internal resistance value is calculated, and this algebraic difference is used as the internal resistance gradient value characterizing the rate and direction of change at the current time. All internal resistance gradient values ​​calculated sequentially within a continuous time period are concatenated in chronological order to finally obtain the internal resistance gradient vector sequence.

[0034] Secondly, the internal resistance gradient vector sequence is input into a pre-defined single-step trend evolution model for forward extrapolation calculation to obtain a priori estimation sequence of fluctuations. Specifically, a pre-defined single-step trend evolution model characterizing the time evolution law of the internal resistance gradient is established. The actual internal resistance gradient value at the previous moment is multiplied by a pre-defined state decay factor, and forward extrapolation calculation is performed to obtain the priori estimation value of fluctuations at the current moment. The pre-defined state decay factor is used to characterize the natural continuation of the internal resistance fluctuation trend under the absence of severe external disturbances. Its specific value is determined based on the autocorrelation statistical analysis of the internal resistance gradient of the same type of battery in the stable charge and discharge phase. By fitting the first-order autocorrelation coefficient of the internal resistance gradient sequence data in the long-term normal phase, multiple sets of internal resistance gradient data of the same type of battery under stable operating conditions are selected, and the first-order autoregression coefficient is estimated using the least squares method. The median of each group of estimation results is taken as the final factor. For example, according to the statistical data of measured data, the typical value of this factor is 0.92. The model is extrapolated one by one for each moment in the sequence, and the priori estimation values ​​of fluctuations at each moment are arranged in order to obtain the priori estimation sequence of fluctuations.

[0035] It is worth noting that the preset single-step trend evolution model is a first-order autoregressive model. Its state decay factor is determined by autocorrelation analysis of the internal resistance gradient sequence under historical stable operating conditions. The first-order autocorrelation coefficients of the gradient sequences of multiple normal aging batteries are calculated, and the median of each sequence coefficient is taken as the fixed decay factor. If the factor is greater than 1 or less than 0, it is truncated to the range of 0 to 1.

[0036] Then, the deviation component between the internal resistance gradient vector sequence and the prior estimation sequence of fluctuation is calculated, and the Bayesian adjustment weight is calculated using the preset inference variance and observation variance. This weight is used to compensate for the deviation component and generate the posterior estimation sequence of fluctuation. Specifically, within the Bayesian inference framework, the prior estimation value of fluctuation is used as the prior mean, and the actual input internal resistance gradient value is used as the observation evidence. The deviation component between the actual observed internal resistance gradient value and the prior estimation value of fluctuation is calculated. At the same time, the trace of the prior estimation error covariance matrix output from the preceding Kalman filter is extracted, i.e., the sum of the elements on the main diagonal of the matrix, and this is used as the preset inference variance reflecting the reliability of the model inference; and the statistical variance of the internal resistance observation sequence within the current sliding time window is extracted as the observation variance reflecting the reliability of the actual difference gradient; the sliding window length is consistent with the window length used to calculate the numerical dispersion in step S12. The Bayesian adjustment weight for this deviation component is calculated by dividing the inference variance by the sum of the inference variance and the observation variance. Finally, the bias component is multiplied by a Bayesian adjustment weight and added as a compensation term to the fluctuation prior estimate to update the fluctuation posterior estimate at the current time step. When the actual gradient deviates significantly from the prediction and the observation variance is small, the inference process will cause the fluctuation posterior estimate to significantly converge with the actual observed gradient. The updated posterior means at each time step are summarized to generate the fluctuation posterior estimate sequence.

[0037] Next, based on the current operating conditions, a fusion weight matrix is ​​extracted from a pre-set mapping table. This fusion weight matrix is ​​then used to perform a weighted summation of the prior and posterior fluctuation estimation sequences to determine the internal resistance fluctuation trend sequence. Specifically, for each discrete moment, the first weight coefficient of the corresponding prior estimate and the second weight coefficient of the corresponding posterior estimate are extracted from the pre-set mapping table, taking into account the current operating conditions, ensuring that their sum is always 1. The prior fluctuation estimate at the current moment is multiplied by the first weight coefficient, and the posterior fluctuation estimate is multiplied by the second weight coefficient. A weighted summation is then performed, and the sum is taken as the final value of the internal resistance fluctuation trend at the current moment.

[0038] It should be noted that the current operating condition is obtained by parsing the status flags of the charging pile control system or by real-time monitoring of electrical parameters. The current charging mode, such as constant current, constant voltage, standby, load status (light load, heavy load), and relay status are read from the battery management system or charger controller; or the operating condition switching is identified by calculating the voltage and current change rate. For example, when the current change rate exceeds a preset threshold, it is determined to be a load change condition. The construction of the preset mapping table is based on the principle of operating condition classification and error minimization. First, the historical operating data is classified according to the operating condition type, such as constant current charging, constant voltage charging, load switching, etc. For each operating condition, the prior weights and posterior weights are traversed within the range of 0 to 1, and the sum of the two is 1. The root mean square error between the fusion result and the benchmark under different weight combinations is calculated. The weight combination with the smallest error is selected as the fusion weight corresponding to the operating condition and stored in the mapping table.

[0039] It is worth noting that the internal coefficients of the preset fusion weight matrix are not statically fixed, but dynamically adjusted based on the current operating mode index of the charging pile (such as the current charging stage and load change state). In specific implementation, the system embeds a two-dimensional weight mapping table (i.e., the preset mapping table) constructed through historical feature matching experiments. During the middle stage of constant current charging when the internal resistance is relatively stable, in order to maintain the smoothness of the trend sequence, a higher first weight coefficient is retrieved from the table, for example, the first weight coefficient is set to 0.65 and the second weight coefficient is set to 0.35; when a step change in the charging current command is detected, i.e., approaching full charge or a sudden load change stage, the table lookup mechanism automatically responds and retrieves an increased second weight coefficient, for example, the first weight coefficient is set to 0.40 and the second weight coefficient is set to 0.60, so that the fused trend value is tilted towards the observation evidence. All discrete point values ​​obtained by dynamic weighting calculation are sequentially spliced ​​together to finally determine the internal resistance fluctuation trend sequence that has both anti-disturbance capability and abrupt change sensitivity.

[0040] In step S14, an observation residual sequence is generated based on the matching result between the internal resistance fluctuation trend sequence and the preset operating condition mode library. An observation noise covariance matrix is ​​constructed based on the observation residual sequence, and periodic disturbances caused by operating conditions are filtered out to obtain a pure internal resistance offset sequence, including: The similarity features of the internal resistance fluctuation trend sequence are calculated by traversing the preset working condition mode library to determine the matching working condition mode index. Retrieve the periodic disturbance component template corresponding to the matching working condition mode index, calculate the difference between the internal resistance fluctuation trend sequence and the periodic disturbance component template, and generate the observation residual sequence; An observation noise covariance matrix is ​​constructed using the observed residual sequence. The observation noise covariance matrix is ​​then input into a preset state observer to update the state of the internal resistance fluctuation trend sequence. Periodic disturbance estimates are removed to obtain the pure internal resistance offset sequence.

[0041] First, the similarity features of the internal resistance fluctuation trend sequence are calculated by traversing the preset operating condition pattern library to determine the matching operating condition pattern index. Specifically, the waveform feature vector of the internal resistance fluctuation trend sequence within the current sliding time window is extracted. Using a dynamic time warping algorithm or the Pearson correlation coefficient calculation method, the similarity score between this waveform feature vector and the waveform feature templates of various typical operating conditions stored in the operating condition pattern library is calculated. The waveform type with the highest similarity score exceeding the preset matching threshold is selected, and its corresponding category identifier is determined as the matching operating condition pattern index. This preset matching threshold is set based on the strong correlation statistical standard of the Pearson correlation coefficient, typically ranging from 0.80 to 0.95. In this embodiment, 0.85 is selected as the preset matching threshold to ensure that only operating conditions with highly consistent waveform evolution trends are considered successfully matched, preventing the retrieval of incorrect disturbance templates due to incorrect matching.

[0042] It is worth noting that the operating condition pattern library is constructed using the K-means clustering algorithm. It collects waveforms of internal resistance fluctuations under typical scenarios, such as during constant current charging, step pulse charging, full charge constant voltage charging, and load switching. Feature vectors, mean, variance, peak value, and waveform entropy are extracted for each waveform. The K-means algorithm is used to cluster the waveforms into several classes. For each class, the waveform closest to the cluster center is selected as the template for the periodic perturbation component, and each template is labeled with an operating condition type identifier. The number of clusters is determined according to the elbow rule, taking the number of classes corresponding to the inflection point of the sum of squared errors. This operating condition pattern library not only contains feature identifiers for various operating conditions but also corresponding periodic perturbation component templates. These templates objectively record the regular, minute fluctuations in internal resistance caused by the battery's own electrochemical polarization characteristics or the high-frequency ripple of the charging pile during specific charging stages.

[0043] Secondly, the periodic disturbance component template corresponding to the matching operating condition mode index is retrieved, and the difference between the internal resistance fluctuation trend sequence and the periodic disturbance component template is calculated to generate the observation residual sequence. Specifically, using the determined matching operating condition mode index, the periodic disturbance component template of the corresponding time length is accurately addressed and extracted from the operating condition mode library. On the time axis, the currently input internal resistance fluctuation trend sequence is phase-aligned with this template, and the algebraic difference between the actual internal resistance fluctuation trend value and the corresponding benchmark value in the template is calculated point by point. All algebraic differences calculated within consecutive time periods are sequentially concatenated to generate the observation residual sequence. This sequence initially isolates the main operating condition background and reflects the random or anomalous components in the actual fluctuations that deviate from the known periodic patterns.

[0044] Then, an observation noise covariance matrix is ​​constructed using the observed residual sequence. Specifically, the statistical variance of the observed residual sequence within a set time window is calculated. The calculated statistical variance values ​​are used as the main diagonal elements to construct a diagonal matrix, thus forming the observation noise covariance matrix. This matrix dynamically and quantitatively describes the statistical discreteness of noise fluctuations in the current residual sequence; the higher the discreteness, the larger the values ​​of the diagonal elements of the matrix.

[0045] Next, the observation noise covariance matrix is ​​input into a preset state observer to update the state of the internal resistance fluctuation trend sequence, removing the estimated periodic disturbance value to obtain a pure internal resistance offset sequence. Specifically, the preset state observer is a Luneburg state observer. The prior dynamic equations inside this observer are constructed using a discrete sinusoidal harmonic oscillator model to characterize the sinusoidal periodic fluctuations induced by switching frequency or charge / discharge ripple under specific operating conditions. Using the input observation noise covariance matrix, combined with the pole placement method, the state feedback gain vector inside the observer is dynamically calculated and adjusted in real time. The adjusted state feedback gain is used to iteratively estimate the state of the internal resistance fluctuation trend sequence, accurately fitting the actual periodic disturbance dynamic component at the current moment. Then, a subtraction operation is performed to directly subtract the estimated periodic disturbance value from the internal resistance fluctuation trend sequence, completely removing the background disturbance signal strongly correlated with the specific operating condition. The remaining non-periodic offsets, which only reflect the degradation of battery physical impedance or the deterioration of connection state, are spliced ​​together in chronological order to finally obtain the pure internal resistance offset sequence.

[0046] It is worth noting that the preset state observer does not employ a fixed parameter structure, but rather adaptively adjusts based on the matching operating condition mode index. In practical applications, when the matching operating condition mode index indicates a drastically fluctuating operating condition with frequent load switching, the observer automatically increases the state update step size and sensitivity parameter of the residual sequence to accelerate convergence and capture rapid transient changes in internal resistance. When the index indicates a stable constant current operating condition, the sensitivity parameter automatically decreases to enhance resistance to accidental transient noise. Through this adaptive mechanism combined with the operating condition index, it ensures that the extracted pure internal resistance offset sequence accurately reflects the equipment degradation trend and avoids being masked by background fluctuations during regular charge-discharge cycles.

[0047] In step S15, the cumulative statistical value of the offset and the slope of the health degradation rate are extracted from the pure internal resistance offset sequence, and a dynamic health benchmark range is established by kernel density estimation in combination with a preset set of historical degradation rate slopes, including: Perform numerical integration on the pure internal resistance offset sequence to generate the cumulative statistical value of the offset; The cumulative statistical value of the offset is input into the pre-trained regression model for linear fitting, and the slope coefficient of the fitted line is extracted as the slope of the health degradation rate. Calculate the Euclidean distance between the health degradation rate slope and a preset set of historical degradation rate slopes to generate a multi-cycle offset amplitude sequence; The probability density function is obtained by performing kernel density estimation on the multi-cycle offset amplitude sequence. The upper and lower boundary values ​​corresponding to the confidence interval are extracted based on the probability density function to establish the dynamic health benchmark range.

[0048] First, numerical integration is performed on the pure internal resistance offset sequence to generate cumulative statistical values ​​of the offset. Specifically, within a set continuous monitoring period, the discrete time-domain accumulation method is used to numerically integrate all discrete data points in the pure internal resistance offset sequence. The integration formula is as follows: In the formula, For the current number The cumulative statistical value of the offset generated at the end of each sampling period The first in the pure internal resistance offset sequence The internal resistance offset amplitude at each sampling point The sampling interval is fixed. Through this discrete integration operation, the short-term, minute internal resistance shift is converted into a cumulative amount that increases monotonically over time, objectively reflecting the cumulative degree of damage as the battery impedance continues to deviate from the healthy baseline.

[0049] Secondly, the cumulative offset statistics are input into the pre-trained regression model for linear fitting, and the slope coefficient of the fitted line is extracted as the slope of the health degradation rate. Specifically, the operating cycle time of the charging pile or the number of charge-discharge cycles is used as the independent variable, and the calculated cumulative offset statistics are used as the dependent variable. Multiple sets of cumulative value data of the same model of battery over the entire life cycle are collected to construct a training sample set. A univariate linear regression equation is constructed, and the least squares method is used to construct the objective loss function, which minimizes the sum of squared residuals from all data points to the fitted line, thereby solving for the optimal slope parameter and intercept parameter as model parameters. The obtained slope parameter is extracted as the health degradation rate slope, which quantifies the average acceleration rate of internal resistance cumulative degradation per unit time.

[0050] Then, the Euclidean distance between the current healthy degradation rate slope and a preset set of historical degradation rate slopes is calculated to generate a multi-cycle offset amplitude sequence. Specifically, the Euclidean distance between the currently calculated healthy degradation rate slope and each historical slope reference value in the set is calculated. All the calculated distance values ​​are arranged in the order of their corresponding historical cycles to generate a multi-cycle offset amplitude sequence, which characterizes the deviation distribution of the current battery degradation rate relative to historical typical levels.

[0051] It's worth noting that the pre-set historical degradation rate slope set is not a single value, but rather constructed through long-term statistical extraction of data from a large number of vehicle-to-vehicle or bench-tested battery packs of the same model throughout their entire lifecycle. The set includes slope reference values ​​corresponding to different physical aging stages, such as the slope reference value for the normal aging stage (e.g., ...). ), and the slope reference value for the mild abnormal stage (e.g. ) and the slope reference value for the accelerated degradation phase (e.g. This set provides a complete historical reference system for the current degradation trend. Specifically, by collecting the cumulative statistical values ​​of the offset of the same type of battery throughout its entire life cycle, the cumulative value sequence of each battery is linearly fitted to obtain the slope of the degradation rate at each stage. All slope data are summarized, outliers other than three times the standard deviation are removed, and the data are binned equally according to the numerical value. The median of each bin is taken as the typical slope reference value to form the set.

[0052] Finally, kernel density estimation is performed on the multi-period offset amplitude sequence to obtain the probability density function. Based on the probability density function, the upper and lower boundary values ​​corresponding to the confidence interval are extracted to establish a dynamic health benchmark range. Specifically, the Gaussian kernel function, a non-parametric statistical method, is used to smooth the multi-period offset amplitude sequence. The formula for kernel density estimation is: In the formula, The obtained probability density function is used for estimation; The offset magnitude is the independent variable; The total number of samples in the multi-period offset amplitude sequence; These are specific distance sample values ​​in the sequence; To control the smoothing level, the bandwidth parameter is automatically determined using the Silverman rule. Specifically, it is the smaller of 0.9 times the standard deviation and the interquartile range divided by 1.34, and then multiplied by the sample size to the power of -0.2. This rule can ensure the estimation accuracy while avoiding over-smoothing. The standard Gaussian kernel function, i.e. .

[0053] Furthermore, after obtaining the probability density function, it is integrally processed to obtain the cumulative distribution function (CDF). Based on the preset confidence level, the corresponding quantile values ​​are found in reverse on the CDF curve. The lower quantile is extracted as the lower boundary value, and the upper quantile is extracted as the upper boundary value. The numerical interval formed by these two boundary values ​​is established as the dynamic health benchmark range.

[0054] It is worth noting that the preset confidence interval level is set based on the statistical tolerance standard for power battery safety monitoring. To filter out reasonable fluctuations under normal operating conditions and accurately capture statistically significant abnormal shifts, the confidence level is typically set to 95%. At this point, the system extracts the value at the 2.5% cumulative distribution as the lower boundary value and the value at the 97.5% cumulative distribution as the upper boundary value. This dynamic health benchmark range can dynamically evolve with the expansion of historical data, effectively adapting to the nonlinear degradation characteristics of batteries at different life stages and avoiding rigid misjudgments caused by using fixed thresholds.

[0055] In step S16, if the pure internal resistance offset sequence exceeds the dynamic health benchmark range, then an abnormal offset persistence determination process is performed to determine the abnormal offset event, including: The pure internal resistance offset sequence is compared point by point with the dynamic health benchmark range to generate an out-of-limit state mapping sequence. Extract the pure internal resistance offset amplitude corresponding to the over-limit state in the over-limit state mapping sequence, identify outliers for the pure internal resistance offset amplitude using the interquartile range rule, update the over-limit state mapping sequence, and generate a corrected over-limit data stream; The duration of consecutive out-of-limit events in the corrected out-of-limit data stream is statistically analyzed to generate a continuous statistical sequence. If the duration statistical sequence is greater than a preset duration threshold, then the offset duration interval is locked. The abnormal offset event is determined by extracting a data segment from the pure internal resistance offset sequence based on the offset persistence interval.

[0056] First, the pure internal resistance offset sequence is compared point-by-point with the dynamic health benchmark range to generate an out-of-limit state mapping sequence. Specifically, discrete amplitude points in the pure internal resistance offset sequence are read one by one to determine whether the amplitude falls within the dynamic health benchmark range established in the previous step. A binary mapping function is constructed: if the internal resistance offset amplitude of the current sampling point is greater than the upper boundary value or less than the lower boundary value, the mapping state value of the sampling point is assigned as 1, indicating that it is in an out-of-limit state; if the offset amplitude is within the upper or lower boundary range, the mapping state value is assigned as 0, indicating that it is in a normal fluctuation state. The mapping state values ​​of all discrete sampling points are arranged in chronological order to generate an out-of-limit state mapping sequence consisting only of 0 and 1.

[0057] Secondly, the pure internal resistance offset amplitude corresponding to the out-of-limit state in the out-of-limit state mapping sequence is extracted. Outliers are identified using the interquartile range (IQR) rule for these pure internal resistance offset amplitudes, and the out-of-limit state mapping sequence is updated to generate a corrected out-of-limit data stream. Specifically, considering the possibility of instantaneous measurement abrupt changes caused by sensor flypoints or strong electromagnetic pulses in the test environment, the statistical interquartile range (IQR) method is used to clean the sequence of outliers. All corresponding pure internal resistance offset amplitudes marked as 1 (i.e., out-of-limit state) in the out-of-limit state mapping sequence are extracted to form an out-of-limit amplitude set. The algebraic difference between the third quartile and the first quartile of this set is calculated to obtain the interquartile range.

[0058] Furthermore, a reasonable upper limit for fluctuation is set as the third quartile plus 1.5 times the interquartile range. If the offset amplitude of a certain outlier in the set exceeds this upper limit, it is identified as an outlier (i.e., an outlier) that does not conform to the laws of physical evolution. The outlier state mapping sequence is updated accordingly, forcibly flipping the state value corresponding to the outlier in the mapping sequence from 1 to 0. This smoothing and cleaning process filters out false outlier signals caused by extreme value disturbances, generating a corrected outlier data stream.

[0059] Then, the continuous over-limit duration of the corrected over-limit data stream is statistically analyzed to generate a persistent statistical sequence. Specifically, the corrected over-limit data stream is traversed, and a dynamic accumulator counter is configured. When a status value of 1 is read, the accumulator counter increments; when a status value changes abruptly from 1 to 0, the current value of the accumulator counter is extracted as the continuous over-limit duration reflecting the duration of that over-limit event, and the counter is reset to zero. The extracted multiple relatively independent continuous over-limit duration values ​​are arranged in chronological order of occurrence to generate a persistent statistical sequence.

[0060] Next, if the persistent statistical sequence exceeds a preset duration threshold, the offset persistence interval is locked. Specifically, the magnitude of each value in the persistent statistical sequence is examined one by one. If a value in the sequence exceeds the preset duration threshold, it is determined that the out-of-limit phenomenon is not a random disturbance, but a persistent performance degradation with substantial physical significance. The timestamp indices of the start and end points of this continuous out-of-limit segment are then recorded, and it is locked as the offset persistence interval.

[0061] It is worth noting that the preset duration threshold is set through actual vehicle operation testing, statistically analyzing the upper limit of the duration of the maximum transient background noise under normal operating conditions, and incorporating a certain safety redundancy factor. For example, in a system with a sampling frequency of 1Hz, the preset duration threshold is set to 600 sampling points (equivalent to 10 minutes). Those skilled in the art will understand that this threshold can be adaptively adjusted within the sampling point range corresponding to 5 to 30 minutes, depending on the capacity level and thermal volume of the specific energy storage unit.

[0062] Finally, data segments are extracted from the pure internal resistance offset sequence based on the offset persistence interval and identified as anomalous offset events. Specifically, the start and end timestamps of the locked offset persistence interval are used as the boundaries of the extraction window. The system returns to the pure internal resistance offset sequence with high numerical accuracy and extracts the corresponding complete continuous data sequence within this time window. This extracted data segment fully preserves the high-frequency characteristics of the entire process of the occurrence, development, and maintenance of the anomalous internal resistance offset. It is packaged and identified as an anomalous offset event so that it can be directly output to the subsequent evaluation module for severity quantification and rating.

[0063] In step S17, an aging factor is extracted based on the abnormal offset event, and the offset severity value is calculated by fuzzy membership degree weighted fusion to map and determine the fault warning level, including: Extract the pure internal resistance data segment corresponding to the abnormal offset event, and perform discrete time-domain integration on the pure internal resistance data segment to obtain the cumulative internal resistance offset value; The accumulated value of internal resistance offset is substituted into a preset piecewise linear mapping function for mapping calculation, and the corresponding aging factor is extracted. The aging factor is input into a preset fuzzy membership function for fuzzification processing to generate a severity membership vector; The severity membership vector is weighted and summed using a preset risk weight coefficient to calculate the offset severity value. The offset severity value is then input into a preset fault warning level mapping table for numerical range comparison to determine the fault warning level.

[0064] First, extract the clean internal resistance data segment corresponding to the abnormal offset event, and perform discrete-time integration on the clean internal resistance data segment to obtain the cumulative internal resistance offset value. Specifically, extract the clean internal resistance data segment within the locked abnormal offset event time window, and accumulate the product of the internal resistance offset amplitude and the sampling time interval at each sampling moment within the segment in chronological order. The integration formula is: In the formula, The calculated cumulative value of internal resistance offset, This represents the total number of discrete sampling points contained in the abnormal offset event data segment. For the first data segment Pure internal resistance offset amplitude at each sampling point The sampling time interval is fixed. Through this time-domain integration operation, the transient offset amplitude is converted into a cumulative quantity, objectively quantifying the total energy or overall degradation of the internal resistance deviation from the normal baseline in this abnormal event.

[0065] Secondly, the cumulative internal resistance offset is substituted into a preset piecewise linear mapping function for mapping calculation to extract the corresponding aging factor. It is worth noting that the inflection point thresholds of the preset piecewise linear mapping function, such as the boundary values ​​dividing slow aging from accelerated aging and the slope parameters of each segment, are determined by fitting the actual internal resistance decay curves obtained from destructive accelerated aging tests on the same battery pack, such as high-temperature overcharge cycling and high-rate discharge cross-tests. This segmented mechanism accurately reflects the true nonlinear physical evolution law of the battery: extremely slow degradation in the early stages of health damage, followed by rapid deterioration after crossing the critical point.

[0066] Then, the aging factor is input into a preset fuzzy membership function for fuzzification processing to generate a severity membership vector. Specifically, a fuzzy logic algorithm is introduced to establish three overlapping triangular membership functions corresponding to mild, moderate, and severe risks, respectively. Assuming the effective range of the aging factor is 0 to 1.0, the vertex parameters of the triangular membership function for mild risk are defined as [0,0,0.4], for moderate risk as [0.2,0.5,0.8], and for severe risk as [0.6,1.0,1.0], ensuring smooth overlap between adjacent risk levels at the boundaries. The fuzzy membership function can be obtained by collecting a large number of historical abnormal events' cumulative internal resistance offset values ​​and manually labeled severity levels (mild, moderate, severe), statistically analyzing the distribution range of the cumulative values ​​under each level, and taking the boundary of the overlapping area of ​​each level's distribution as the vertex of the triangular membership function, ensuring smooth transition between adjacent levels at the boundaries. The extracted aging factor values ​​are input as independent variables into these three membership functions to calculate the membership values ​​of the aging factor belonging to mild, moderate, and severe risks, respectively. These three membership values ​​are then concatenated into a one-dimensional vector containing three elements, generating a severity membership vector representing the current uncertainty state. For example, the data generated at a certain moment might be a vector [0.1, 0.7, 0.2], indicating that the current state has a membership degree of 0.7, belonging to moderate risk.

[0067] Finally, the severity membership vector is weighted and summed using preset risk weight coefficients to calculate the offset severity value. This offset severity value is then input into a preset fault warning level mapping table for numerical range comparison to determine the fault warning level. Specifically, each risk level is pre-assigned a corresponding fixed preset risk weight coefficient. This risk weight coefficient is determined using the analytic hierarchy process (AHP). By comparing the importance of each severity level pairwise, a judgment matrix is ​​constructed. The relative weights of each level are calculated and a consistency check is performed. The normalized weights are used as preset coefficients; for example, a slight risk is assigned a weight of 0.1, a moderate risk of 0.5, and a severe risk of 0.9. Each membership value in the severity membership vector is multiplied by its corresponding risk weight coefficient and summed to obtain a continuous scalar value between 0 and 1, which is the offset severity value. The calculation formula is: In the formula, This represents the severity of the offset. For the severity membership vector corresponding to the first The membership value of each risk level. For the first Each risk level corresponds to a preset risk weight coefficient.

[0068] Next, the calculated offset severity value is input into a preset fault warning level mapping table for value range lookup and comparison to determine the fault warning level. Specifically, this mapping table clearly defines the correspondence between severity value ranges and discrete warning levels. By collecting offset severity values ​​and actual consequences of downtime or maintenance from historical abnormal events, the severity value ranges under different consequence levels are statistically analyzed. An equally spaced binning method is used to divide the value ranges into several levels, each corresponding to a warning level. For example, the value range [0, 0.3) corresponds to warning level 1; [0.3, 0.6) corresponds to warning level 2; and [0.6, 1.0] corresponds to warning level 3. The warning level labels corresponding to the intervals into which the severity values ​​fall are extracted to obtain the final output fault warning level.

[0069] In step S18, high-risk signal sequences are selected from the fault warning levels. An event-triggered counter reset mechanism is driven by the high-risk signal sequences to generate a warning instruction. A preset health assessment model is then invoked to determine the overall health status, including: The fault warning level is mapped to a preset risk threshold range, and the high-risk signal sequence is obtained by filtering. The trigger counter is used to accumulate the high-risk signal sequence. When the trigger counter meets the preset alarm triggering conditions, the event trigger counter reset mechanism is activated to generate the warning command. The warning instruction is transmitted to the charging pile management system to update the preset status register, and the status identifier of the status register, the cumulative value of the trigger counter, and the cumulative statistical value of the offset are extracted and input into the preset health assessment model to perform comprehensive scoring calculation to obtain the system operation score; Based on the scoring range in which the system's operating score falls, the overall health status is determined and output.

[0070] First, fault warning levels are mapped to preset risk threshold ranges, and high-risk signal sequences are filtered out. Specifically, the previously calculated discrete fault warning levels (e.g., level 1, level 2, or level 3) are obtained and numerically compared and matched with the pre-defined risk threshold ranges. If the current warning level falls within the threshold range representing medium to high risk, the signal timestamp and feature data corresponding to that level are extracted and arranged sequentially according to their occurrence time to construct a high-risk signal sequence. Regular fluctuation signals falling into the low-risk range are directly filtered out. It is worth noting that the division of the preset risk threshold range is based on statistical analysis of historical charging pile failure cases. In a warning system with three levels, level 1 is typically set as the low-risk range, level 2 as the medium-risk range, and level 3 as the high-risk range. To ensure high sensitivity to potential accelerated degradation or sudden failures, levels 2 and 3 are usually included in the threshold range triggering high-risk signals. In one implementation, if the same battery cell experiences two consecutive level 2 warnings within a very short period, it will be exceptionally upgraded and filtered into the high-risk signal sequence.

[0071] Secondly, the trigger counter accumulates high-risk signal sequences. When the trigger counter meets the preset alarm trigger conditions, the event trigger counter reset mechanism is activated to generate an early warning command. Specifically, an independent trigger counter is configured in the logic control unit, with its initial value set to zero. When a valid high-risk signal is added to the high-risk signal sequence, a step pulse is sent to the trigger counter, driving its count value to increment by one. Simultaneously, the current accumulated value of the trigger counter is monitored in real time. When this value reaches the critical number set by the preset alarm trigger conditions, an early warning command data packet containing core fields such as battery number, fault level code, and trigger timestamp is immediately generated. Simultaneously with the generation of the early warning command, the hardware or software logic activates the reset mechanism, forcibly resetting the trigger counter to zero, so as to start the next round of independent monitoring and calculation, preventing redundant alarms for the same fault event. It is worth noting that the preset alarm trigger conditions include a count value threshold and a sliding time window limit. These conditions are determined based on engineering experience and a trade-off between false alarm rates, aiming to filter out isolated false alarms caused by brief strong electromagnetic interference or power grid surges. For example, the sliding time window is set to 72 hours, and the count value threshold is set to 3 times. The alarm trigger condition is met only when the counter accumulates three high-risk signals within a continuous 72-hour time window. If the number of high-risk signals does not reach three within the 72-hour time window, the previous invalid counts are automatically decayed or cleared, thereby significantly improving the accuracy of fault prediction.

[0072] Then, the warning command is propagated to the charging pile management system to update the preset status register. The system extracts the status register's status identifier, the cumulative value of the trigger counter, and the cumulative offset statistics, and inputs these into the preset health assessment model to perform a comprehensive score calculation, obtaining the system operation score. Specifically, the generated warning command is sent to the upper-level charging pile management system via a control area network (CAN bus) or Ethernet communication interface. The charging pile management system parses the fault level code in the command and updates the corresponding energy storage battery module's status register flag from the normal state (e.g., logic 0) to the alarm state (e.g., logic 1 or 2). Subsequently, the health assessment model extracts the updated status register's status identifier, the cumulative value of the trigger counter, and the cumulative offset statistics calculated earlier (e.g., in step S15), and performs a comprehensive score calculation using a deduction system rule to obtain the current module's system operation score.

[0073] It is worth noting that the health assessment model is a weighted deduction rule model. The model inputs include the alarm level identifier in the status register, the cumulative value of the trigger counter, and the cumulative offset statistics. The model starts with a maximum score of 100 points and calculates deductions based on input characteristics. The alarm level identifiers are divided into normal, moderate, and severe alarms, corresponding to deductions of 0, 15, and 40 points respectively. Ten points are deducted if the cumulative trigger counter value reaches three times within 72 consecutive hours; otherwise, no points are deducted. Ten points are deducted if the cumulative offset statistics exceed a preset threshold. The model parameters are determined by statistically analyzing the correlation between alarm levels and remaining lifespan in historical fault cases. The calculation is as follows: The trigger counter threshold adopts the six sigma principle, which counts the number of times a normal device generates a high-risk signal due to random interference within 72 hours, and sets the threshold as the mean plus three times the standard deviation; The offset cumulative threshold is determined by conducting accelerated aging tests on batteries of the same model and recording the lower limit of the offset cumulative value when the capacity decays to 80%; The model output is a system operation score, and the overall health status is determined according to the score range: above 90 points is healthy, 70 to 89 points is sub-healthy, and below 70 points is unhealthy; The score range is divided according to the correspondence between the battery capacity decay curve and the score, ensuring that every 10-point decrease in score corresponds to approximately 5% capacity decay.

[0074] Finally, based on the scoring range of the system's operational score, the overall health status is determined. Specifically, the calculated system operational score is compared with a preset scoring range to determine its health category, and the corresponding status label, human-machine interaction prompt, or physical disconnection control command is output to the maintenance terminal. It's worth noting that the scoring range is determined based on the mapping relationship between the battery performance degradation curve and available capacity. For example, a score of 90 to 100 is set as a healthy range, indicating that full-power operation is allowed; a score of 70 to 89 is set as a sub-healthy range, indicating that a derating operation strategy is output and an increase in inspection frequency is prompted; and a score below 70 is set as an unhealthy range, indicating that a forced shutdown maintenance suggestion is output and the charging relay is physically disconnected. This hierarchical quantification and multi-verification mechanism can promptly filter low-confidence alarms, reliably track continuously degrading batteries, and improve the overall operational safety of the charging pile equipment.

[0075] In summary, this invention discloses a fault detection method for mobile energy storage charging piles, which solves the problem of low fault detection accuracy in existing technologies.

[0076] Reference Figure 2 The second embodiment of the present invention provides a fault detection system for mobile energy storage charging piles, comprising: The filtering module is used to collect the original voltage and current sequences during the operation of the charging pile, establish the system state equation and observation equation based on the original voltage and current sequences, and perform Kalman filtering to obtain a smooth electrical parameter sequence. The internal resistance stabilization module is used to calculate the internal resistance observation sequence based on the smooth electrical parameter sequence, adaptively adjust the preset process noise covariance matrix using the numerical dispersion of the internal resistance observation sequence, and obtain a stable internal resistance value sequence through point-by-point iterative correction. The trend extraction module is used to perform point-by-point difference operation on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and to perform adaptive Bayesian fusion processing based on operating conditions on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. The disturbance filtering module is used to generate an observation residual sequence based on the matching result between the internal resistance fluctuation trend sequence and the preset operating condition mode library, construct an observation noise covariance matrix based on the observation residual sequence, and filter out the periodic disturbances caused by the operating conditions to obtain a pure internal resistance offset sequence. The benchmark construction module is used to extract the cumulative statistical value of the offset and the slope of the health degradation rate from the pure internal resistance offset sequence, and to establish a dynamic health benchmark range by combining the preset historical degradation rate slope set with kernel density estimation. An anomaly locking module is used to perform an anomaly offset persistence determination process to determine an anomaly offset event if the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark. The severity assessment module is used to extract aging factors based on the abnormal offset events, calculate the offset severity value through fuzzy membership weighted fusion, and map and determine the fault warning level. The status decision module is used to filter high-risk signal sequences from the fault warning levels, drive the event trigger counter reset mechanism according to the high-risk signal sequences, generate warning instructions, and call the preset health assessment model to determine the overall health status.

[0077] It should be noted that the fault detection system for a mobile energy storage charging pile provided in this embodiment of the invention is used to execute all the process steps of the fault detection method for a mobile energy storage charging pile in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0078] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0079] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault detection method for a mobile energy storage charging pile, characterized in that, include: The original voltage and current sequences during the operation of the charging pile are collected. Based on the original voltage and current sequences, the system state equation and observation equation are established, and Kalman filtering is performed to obtain a smooth electrical parameter sequence. The internal resistance observation sequence is calculated based on the smoothed electrical parameter sequence. The preset process noise covariance matrix is ​​adaptively adjusted using the numerical dispersion of the internal resistance observation sequence. A stable internal resistance value sequence is obtained through point-by-point iterative correction. A point-by-point difference operation is performed on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and an adaptive Bayesian fusion processing based on the operating condition is performed on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. An observation residual sequence is generated based on the matching results between the internal resistance fluctuation trend sequence and the preset operating condition mode library. An observation noise covariance matrix is ​​constructed based on the observation residual sequence, and periodic disturbances caused by operating conditions are filtered out to obtain a pure internal resistance offset sequence. The cumulative statistical value of the offset and the slope of the health degradation rate are extracted from the pure internal resistance offset sequence, and a dynamic health benchmark range is established by kernel density estimation in combination with a preset set of historical degradation rate slopes. If the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark, then an abnormal offset persistence determination process is performed to determine the abnormal offset event. Based on the abnormal offset events, aging factors are extracted, and offset severity values ​​are calculated through fuzzy membership degree weighted fusion to map and determine the fault warning level. High-risk signal sequences are selected from the fault warning levels. The event trigger counter reset mechanism is driven by the high-risk signal sequences to generate warning instructions and call the preset health assessment model to determine the overall health status.

2. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The system collects the original voltage and current sequences during the operation of the charging pile, establishes system state equations and observation equations based on these sequences, and performs Kalman filtering to obtain a smooth electrical parameter sequence, including: Obtain the original voltage and current sequence of the charging pile port; The system state equation is established based on the original voltage and current sequence. The system state equation is then used to deduce the prior state estimation vector and the observation equation. Based on the observation equation, the Kalman filter gain is determined by combining the prior state estimation vector and the actual observation value, wherein the actual observation value is composed of the voltage sample value and the current sample value in the original voltage and current sequence. The prior state estimation vector is corrected using the Kalman filter gain to calculate the posterior state estimation vector; The corrected voltage and current components are extracted from the posterior state estimation vector to form the smoothed electrical parameter sequence.

3. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The step of calculating the internal resistance observation sequence based on the smoothed electrical parameter sequence, adaptively adjusting the preset process noise covariance matrix using the numerical dispersion of the internal resistance observation sequence, and obtaining a stable internal resistance value sequence through point-by-point iterative correction includes: Extract the smoothed voltage and smoothed current components from the smoothed electrical parameter sequence, and perform mapping calculations using Ohm's law discrete form to obtain the internal resistance observation sequence; For the internal resistance observation sequence at the current sampling time, the numerical dispersion is calculated. If the numerical dispersion is greater than a preset stability threshold, an adaptive adjustment coefficient is determined based on the degree of deviation of the numerical dispersion from the preset stability threshold. The adaptive adjustment coefficient is then used to perform scaling processing on the preset process noise covariance matrix to obtain the adaptively adjusted process noise covariance matrix. The Kalman gain is calculated using the adaptively adjusted process noise covariance matrix to obtain the dynamic filter gain corresponding to the current sampling time in real time. The dynamic filtering gain is applied to perform a weighted correction on the deviation between the observed internal resistance sequence and the prior internal resistance estimate at the current sampling time to obtain the stable internal resistance value at the current sampling time. The stable internal resistance values ​​at each sampling time are then concatenated in chronological order to obtain the stable internal resistance value sequence. The prior internal resistance estimate at the current sampling time is obtained by performing forward time extrapolation processing on the stable internal resistance value at the previous sampling time.

4. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The step of performing point-by-point difference operations on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and performing adaptive Bayesian fusion processing based on operating conditions on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence includes: Perform point-by-point difference operation on the stable internal resistance value sequence to obtain the internal resistance gradient vector sequence; The internal resistance gradient vector sequence is input into a preset single-step trend evolution model to perform forward extrapolation calculations, thereby obtaining a priori estimation sequence of fluctuations. The deviation component between the internal resistance gradient vector sequence and the fluctuation prior estimation sequence is calculated, and the Bayesian adjustment weight is calculated using the preset inference variance and observation variance. The deviation component is compensated by the Bayesian adjustment weight to generate the fluctuation posterior estimation sequence. Based on the current operating conditions, a fusion weight matrix is ​​extracted from a preset mapping table. The fusion weight matrix is ​​then used to perform a weighted summation on the prior estimation sequence of the fluctuation and the posterior estimation sequence of the fluctuation to determine the internal resistance fluctuation trend sequence.

5. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The process of generating an observation residual sequence based on the matching results between the internal resistance fluctuation trend sequence and a preset operating condition model library, constructing an observation noise covariance matrix based on the observation residual sequence, and filtering out periodic disturbances caused by operating conditions to obtain a pure internal resistance offset sequence includes: The similarity features of the internal resistance fluctuation trend sequence are calculated by traversing the preset working condition mode library to determine the matching working condition mode index. Retrieve the periodic disturbance component template corresponding to the matching working condition mode index, calculate the difference between the internal resistance fluctuation trend sequence and the periodic disturbance component template, and generate the observation residual sequence; An observation noise covariance matrix is ​​constructed using the observed residual sequence. The observation noise covariance matrix is ​​then input into a preset state observer to update the state of the internal resistance fluctuation trend sequence. Periodic disturbance estimates are removed to obtain the pure internal resistance offset sequence.

6. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The step of extracting the cumulative statistical value of the offset and the slope of the health degradation rate from the pure internal resistance offset sequence, and establishing a dynamic health benchmark range by kernel density estimation in combination with a preset set of historical degradation rate slopes, includes: Perform numerical integration on the pure internal resistance offset sequence to generate the cumulative statistical value of the offset; The cumulative statistical value of the offset is input into the pre-trained regression model for linear fitting, and the slope coefficient of the fitted line is extracted as the slope of the health degradation rate. Calculate the Euclidean distance between the health degradation rate slope and a preset set of historical degradation rate slopes to generate a multi-cycle offset amplitude sequence; The probability density function is obtained by performing kernel density estimation on the multi-cycle offset amplitude sequence. The upper and lower boundary values ​​corresponding to the confidence interval are extracted based on the probability density function to establish the dynamic health benchmark range.

7. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, If the pure internal resistance offset sequence exceeds the dynamic health benchmark range, then an abnormal offset persistence determination process is performed to determine the abnormal offset event, including: The pure internal resistance offset sequence is compared point by point with the dynamic health benchmark range to generate an out-of-limit state mapping sequence. Extract the pure internal resistance offset amplitude corresponding to the over-limit state in the over-limit state mapping sequence, identify outliers for the pure internal resistance offset amplitude using the interquartile range rule, update the over-limit state mapping sequence, and generate a corrected over-limit data stream; The duration of consecutive out-of-limit events in the corrected out-of-limit data stream is statistically analyzed to generate a continuous statistical sequence. If the duration statistical sequence is greater than a preset duration threshold, then the offset duration interval is locked. The abnormal offset event is determined by extracting a data segment from the pure internal resistance offset sequence based on the offset persistence interval.

8. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The step of extracting aging factors based on the abnormal offset events and calculating the offset severity value through fuzzy membership degree weighted fusion to map and determine the fault warning level includes: Extract the pure internal resistance data segment corresponding to the abnormal offset event, and perform discrete time-domain integration on the pure internal resistance data segment to obtain the cumulative internal resistance offset value; The accumulated value of internal resistance offset is substituted into a preset piecewise linear mapping function for mapping calculation, and the corresponding aging factor is extracted. The aging factor is input into a preset fuzzy membership function for fuzzification processing to generate a severity membership vector; The severity membership vector is weighted and summed using a preset risk weight coefficient to calculate the offset severity value. The offset severity value is then input into a preset fault warning level mapping table for numerical range comparison to determine the fault warning level.

9. The fault detection method for a mobile energy storage charging pile according to claim 1, characterized in that, The process of selecting high-risk signal sequences from the fault warning levels, driving an event-triggered counter reset mechanism based on the high-risk signal sequences, generating warning instructions, and calling a preset health assessment model to determine the overall health status includes: The fault warning level is mapped to a preset risk threshold range, and the high-risk signal sequence is obtained by filtering. The trigger counter is used to accumulate the high-risk signal sequence. When the trigger counter meets the preset alarm triggering conditions, the event trigger counter reset mechanism is activated to generate the warning command. The warning instruction is transmitted to the charging pile management system to update the preset status register, and the status identifier of the status register, the cumulative value of the trigger counter, and the cumulative statistical value of the offset are extracted and input into the preset health assessment model to perform comprehensive score calculation to obtain the system operation score; Based on the scoring range in which the system's operating score falls, the overall health status is determined and output.

10. A fault detection system for a mobile energy storage charging pile, characterized in that, include: The filtering module is used to collect the original voltage and current sequences during the operation of the charging pile, establish the system state equation and observation equation based on the original voltage and current sequences, and perform Kalman filtering to obtain a smooth electrical parameter sequence. The internal resistance stabilization module is used to calculate the internal resistance observation sequence based on the smooth electrical parameter sequence, adaptively adjust the preset process noise covariance matrix using the numerical dispersion of the internal resistance observation sequence, and obtain a stable internal resistance value sequence through point-by-point iterative correction. The trend extraction module is used to perform point-by-point difference operation on the stable internal resistance value sequence to generate an internal resistance gradient vector sequence, and to perform adaptive Bayesian fusion processing based on operating conditions on the internal resistance gradient vector sequence to determine the internal resistance fluctuation trend sequence. The disturbance filtering module is used to generate an observation residual sequence based on the matching result between the internal resistance fluctuation trend sequence and the preset operating condition mode library, construct an observation noise covariance matrix based on the observation residual sequence, and filter out the periodic disturbances caused by the operating conditions to obtain a pure internal resistance offset sequence. The benchmark construction module is used to extract the cumulative statistical value of the offset and the slope of the health degradation rate from the pure internal resistance offset sequence, and to establish a dynamic health benchmark range by combining the preset historical degradation rate slope set with kernel density estimation. An anomaly locking module is used to perform an anomaly offset persistence determination process to determine an anomaly offset event if the pure internal resistance offset sequence exceeds the range of the dynamic health benchmark. The severity assessment module is used to extract aging factors based on the abnormal offset events, calculate the offset severity value through fuzzy membership weighted fusion, and map and determine the fault warning level. The status decision module is used to filter high-risk signal sequences from the fault warning levels, drive the event trigger counter reset mechanism according to the high-risk signal sequences, generate warning instructions, and call the preset health assessment model to determine the overall health status.

Images