Charging curve deviation detection method and system based on gaussian process regression

Abstract

The application relates to the technical field of intelligent operation and maintenance of new energy vehicle charging, and discloses a charging curve deviation detection method and system based on Gaussian process regression. Normal historical charging data and real-time charging data matched with the model identifier of a battery to be detected are acquired, a standard curve model of Gaussian process regression in stages is established through quality screening, time alignment and stage division, the real-time charging time window is mapped to the corresponding standard curve, a regularized covariance matrix is constructed in a unified scale space according to model prediction covariance and normal residual covariance, and a multi-parameter joint deviation score is calculated, graded thresholds are determined in combination with the charging stage, model prediction uncertainty, environmental temperature and battery health state, deviation grades and deviation types are identified, and when the deviation continuously meets corresponding conditions, recording, warning or charging protection is performed, and the model and the threshold are updated by using the charging data confirmed to be normal, so that short-time disturbance and continuous deviation can be distinguished.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for new energy vehicle charging, specifically to a method and system for detecting charging curve deviation based on Gaussian process regression. Background Technology

[0002] During the charging process of new energy vehicles, the battery management system, charging equipment, and monitoring platform need to continuously monitor parameters such as charging current, total charging voltage, individual cell voltage, and temperature. Existing methods typically set fixed upper and lower limits for individual parameters, which is difficult to adapt to the normal curve differences caused by vehicle model, battery specifications, state of charge, health status, ambient temperature, and charging stage.

[0003] Some monitoring methods compare real-time curves with template curves or historical average curves, and judge anomalies based on point value differences, Euclidean distances, or slope differences. Clock offsets between the vehicle and charging pile ends, variations in the stage length of different charging records, and sampling noise can easily be mistaken for parameter deviations. Meanwhile, independent judgment of single parameters is insufficient to express the correlation changes between current, voltage, temperature, and individual unit consistency, and some data-driven models do not uniformly handle prediction uncertainty, asymptotic shifts, and persistence. Therefore, existing technologies still need to address the problems of insufficient adaptability of standard curves, inadequate utilization of multi-parameter correlations, and difficulty in distinguishing between short-term disturbances and persistent deviations. Summary of the Invention

[0004] Existing charging curve deviation detection methods suffer from interrelated issues in terms of detection benchmarks, curve correspondence, and joint judgment: fixed limits or single templates cannot express normal differences caused by battery type, charging stage, ambient temperature, and health status; time offset between vehicle-side and charging pile-side data, as well as variations in the duration of the charging stage, can cause misalignment between real-time and standard curves; single-parameter residuals or direct superposition methods struggle to utilize the correlations between multiple charging parameters and fail to uniformly distinguish between model prediction uncertainty, short-term noise, and persistent deviations. This invention aims to establish a deviation detection path with conditional adaptability, time mapping constraints, and multi-parameter joint statistical basis without replacing the original safety protection of the charging equipment.

[0005] To address the aforementioned technical problems, one aspect of the present invention provides a charging curve deviation detection method based on Gaussian process regression. The method includes steps S1 to S6. Step S1: Acquire normal historical charging data and real-time charging data matching the model identifier of the battery to be tested, and perform quality screening, time alignment, and charging stage division on the data. Step S2: Establish Gaussian process regression standard curve models for the pre-charging stage, constant current charging stage, and constant voltage charging stage respectively. Step S3: Map the sliding time window of the real-time charging data to the corresponding standard curve and form a standardized residual vector. Step S4: Construct a regularized covariance matrix, calculate the squared Mahalanobis distance and joint deviation score. Step S5: Determine an adaptive grading threshold based on the normal verification score distribution under operating conditions, and determine the deviation level and deviation type. Step S6: Perform data recording, alarm, or charging protection according to the deviation level, and generate and verify candidate models and candidate thresholds by confirming normal data.

[0006] Specifically, the model identifier is formed based on the vehicle model, battery chemistry, battery rated capacity, module configuration, and charging protocol. Charging data includes charging time t, state of charge (SOC), charging current I, total charging voltage U, maximum single-cell voltage U_max, minimum single-cell voltage U_min, maximum temperature T_max, ambient temperature T_env, and battery state of health (SOH). The single-cell voltage range ΔU is calculated as ΔU = U_max - U_min. Normal historical charging data is derived from charging records where no charging protection events occurred, the charging process was complete, and the data quality met the training conditions.

[0007] The vehicle-side battery management system data and the charging pile-side data are aligned according to a unified time axis, and abnormal and missing sampling points are marked and processed. Based on the charging control state, state of charge, and current or voltage change characteristics, the charging process is divided into a pre-charging stage, a constant current charging stage, and a constant voltage charging stage, and a monotonically increasing normalized charging process τ is established for each stage.

[0008] The normalized charging process τ, state of charge (SOC), ambient temperature T_env, and battery health state (SOH) form the model input vector, while the charging current I, total charging voltage U, maximum temperature T_max, and single-cell voltage range ΔU form the model output vector. Gaussian process regression sub-models are established for each charging stage and each output parameter. The inputs and outputs are standardized according to the mean and positive scaling parameters of the training data, and the kernel function uses a radial basis function with an independent length scale. k_s,j(x_i′,x_q′)=σ_f,s,j^2 exp{-(1 / 2)(x_i′-x_q′)^T Λ_s,j^(-1)(x_i′-x_q′)}.

[0009] Where s represents the charging stage, j represents the output parameters, σ_f,s,j^2 represents the signal variance, and Λ_s,j represents the diagonal matrix formed by the squares of the length scale. The kernel function hyperparameters and observation noise variance are determined by maximizing the logarithmic marginal likelihood. For the standardized training output vector y_j′, let K_y,s,j=K_s,j+σ_n,s,j^2E_n, then the standardized prediction mean and prediction variance at the input x_*′ to be predicted are respectively: μ_j′(x_*)=k_*^T K_y,s,j^(-1)y_j′; v_j′(x_*)=k_s,j(x_*′,x_*′)-k_*^T K_y,s,j^(-1)k_*.

[0010] The prediction results are restored to the original scale based on the training mean m_y,j of the output parameters and the positive output scale s_y,j. The original scale prediction variances are grouped into Σ_GP(t), and A = diag(s_y,1,…,s_y,p). The dimensionless standardized model prediction covariance matrix is ​​obtained according to C_GP(t) = A^(-1)Σ_GP(t)A^(-1). The normal residual covariance matrix C_res,s is calculated based on the standardized residuals of the normal validation data, and constructed as follows: C(t)=C_GP(t)+C_res,s+λE_p.

[0011] Where λ is a positive dimensionless regularization coefficient, and E_p is the p-order identity matrix. By adding λE_p to the normalized space, the same scalar regularization term is avoided from being directly added to the original scale covariances with different physical units. The actual calculation is solved through matrix decomposition, without forming an explicit inverse matrix.

[0012] During curve mapping, an initial mapping is established based on timestamps, charging stage transition positions, state-of-charge interval boundaries, and current or voltage slope change positions. Then, dynamic time warping, constrained by time monotonicity, stage boundaries, and maximum scaling, corrects the local mapping. Let the real-time charging parameter vector at the mapping position be y_real(t), and the predicted mean vector be μ(t). Then the physical residual vector e(t) and the standardized residual vector r(t) are respectively: e(t) = y_real(t) - μ(t); r(t)=A^(-1)e(t).

[0013] Calculate the squared Mahalanobis distance based on the standardized residual vector and the regularized covariance matrix: D_M^2(t)=r(t)^TC(t)^(-1)r(t).

[0014] Time series aggregation is performed on a sliding time window of length W to obtain a joint bias score: J(t)=∑_(h=0)^(W-1)a_hD_M^2(t-h), Where a_h is a non-negative time weight, and the sum of all weights participating in the aggregation is 1. When a missing position is not included in the aggregation, the weights of the remaining valid positions are renormalized; when the number of valid positions is insufficient, the output of the joint bias score is paused.

[0015] A conditional score distribution is formed based on the charging stage, model prediction uncertainty range, ambient temperature range, and battery health state range of the normal verification data. First-level thresholds H_1, second-level thresholds H_2, and third-level thresholds H_3 are formed using progressively increasing conditional quantiles. State escalation is determined by consecutive out-of-limit counts, and state deterioration is determined by consecutive recovery counts below the recovery threshold. Signed standardized residuals are obtained using z_j(t)=r_j(t) / √C_jj(t), and the main deviation parameters are determined based on the absolute contribution ratio of each parameter to the Mahalanobis distance. Positive bias, negative bias, trend bias, and fluctuation bias are determined by combining the residual direction, the slope difference between the real-time curve and the predicted mean curve, and the variance ratio of the detrended residuals.

[0016] Another aspect of the present invention provides a charging curve deviation detection system based on Gaussian process regression. The system includes a data acquisition and stage processing module, a standard curve modeling module, a curve mapping and residual formation module, a joint deviation calculation module, an adaptive judgment module, and a response and model update module. The functional modules are connected in the order of charging data processing. The response and model update module only provides the validated model parameters and threshold parameters to the standard curve modeling module and the adaptive judgment module. The system also includes a vehicle-side data interface, a charging pile-side data interface, a processor, a memory, an alarm output interface, and a charging control interface. When data quality, curve mapping, or model integrity does not meet the detection conditions, the system outputs a detection unavailable status while retaining the original safety protections of the battery management system and charging equipment.

[0017] In summary, this invention uses normal historical charging data that matches the model identifier as a basis to establish a Gaussian process regression standard curve model for the output prediction mean and prediction uncertainty according to the charging stage; it determines the corresponding position of the real-time curve and the standard curve through constrained time mapping, integrates the model prediction covariance and the normal residual covariance in a unified dimensionless scale space, and uses Mahalanobis distance and its sliding time window aggregation results to characterize the joint bias of multiple parameters; further, it forms a graded threshold based on the normal score distribution under operating conditions, identifies the bias level, main bias parameters and bias type, executes graded response, and completes the validated model and threshold update by confirming normal data.

[0018] Compared to methods that rely on fixed limits, single templates, or single-parameter residuals, the aforementioned technical solution expresses the changes in the normal charging curve under varying operating conditions through model identification, phased Gaussian process regression, and conditional thresholds. It reduces the impact of data source clock skew and phase process scaling on curve comparisons through constrained dynamic time warping; preserves the normal correlation between multiple charging parameters using standardized covariance and Mahalanobis distance, while avoiding the mixing of dimensions between original-scale parameters and dimensionless scores; distinguishes between short-term disturbances and persistent deviations through sliding time window aggregation, continuous over-limit counting, and recovery hysteresis; and restricts abnormal data from entering the normal training set through graded response, record tracing, candidate model validation, and version rollback. This forms a closed-loop processing path from data access, standard curve prediction, joint deviation determination to charging protection and model maintenance. Attached Figure Description

[0019] Figure 1 This is a flowchart of the charging curve deviation detection method based on Gaussian process regression of the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the construction of the phased Gaussian process regression standard curve model of this invention.

[0021] Figure 3 This is a data flow diagram for real-time curve mapping and joint deviation calculation in this invention.

[0022] Figure 4 This is a schematic diagram illustrating the adaptive grading threshold and deviation response state transition of the present invention.

[0023] Figure 5 This is a block diagram of the charging curve deviation detection system of the present invention.

[0024] Figure 6 This is a schematic diagram illustrating the application scenarios and deployment relationship of the charging stations of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings. The following detailed embodiments are used to illustrate the technical concept and possible implementation methods of the present invention, and are not intended to limit the scope of protection of the present invention. Where there is no contradiction, the technical features of each embodiment can be combined with each other. Communication, security control, and data storage processes not described in detail can be implemented using conventional techniques in the art. The step names and module names in the accompanying drawings have the same meaning as the corresponding names in the specification.

[0026] Example 1 This embodiment combines Figure 1 This document describes the complete processing flow of the charging curve deviation detection method. The detection system is deployed in the edge computing unit of the charging station and connected to the vehicle battery management system, charging equipment controller, and charging monitoring platform. Figure 1 Steps S1 to S6 correspond to data acquisition and stage processing, standard curve modeling, curve mapping and residual formation, joint deviation calculation, adaptive judgment, and response and model update, respectively.

[0027] In step S1, the edge computing unit reads the vehicle model, battery chemistry, battery rated capacity, module configuration, charging protocol, and charging equipment output capability to form a model identifier for the battery to be tested. Charging records with the same model identifier are included in the same candidate dataset, while records with inconsistent model identifiers are not directly mixed for training. As a configuration example, 500 complete historical charging records without charging protection events can be selected, with 80% of the complete charging sessions used for training and the remaining complete charging sessions used for validation. The above quantities and proportions are for illustrating the data organization method and do not represent actual measured performance results; the actual quantities are determined based on the coverage of battery types.

[0028] Each charging record includes a charging session identifier, vehicle identifier, charging device identifier, sampling timestamp, charging status, state of charge (SOC), charging current (I), total charging voltage (U), maximum single-cell voltage (U_max), minimum single-cell voltage (U_min), maximum temperature (T_max), ambient temperature (T_env), and state of health (SOH). The single-cell voltage range (ΔU) is calculated using the formula ΔU = U_max - U_min. Both the vehicle-side and charging device-side records retain the original timestamps and the edge computing unit's reception time.

[0029] The basic sampling period in this embodiment is 100 ms. Data quality checks include field completeness, timestamp monotonicity, physical range, rate of change, and communication status checks. When timestamps are repeated, data with complete fields and high communication quality is retained; when no more than three consecutive sampling points are missing, amplitude-limited interpolation is performed using adjacent valid values ​​and the allowed rate of change; when more than three consecutive sampling points are missing, the corresponding segment is marked as incomplete data and joint bias calculation is suspended. Interpolated data retains the interpolation marker and is not used as candidate model update data.

[0030] Time alignment between the vehicle and the charging pile is achieved by establishing a correspondence between charging start-up, contactor closure, effective charging current establishment, and charging phase transition events to estimate a fixed time offset. When a slow change in clock offset is detected, piecewise linear correction is used. After time synchronization is complete, Kalman smoothing is applied to the charging current, total charging voltage, maximum temperature, and individual cell voltage range. The filtered data is used for online detection, while the original data remains in the charging session log. At candidate locations of phase transitions or abrupt changes, the filter's response weight to the current observation is increased to avoid over-smoothing.

[0031] The charging stages are defined based on the charging control status, state of charge, and changes in current and voltage. When a charging connection is established and the charging current is within a limited, low-current range, it is designated as the pre-charging stage. When the charging current remains around the target current and the total charging voltage increases with the process, it is designated as the constant-current charging stage. When the total charging voltage remains around the target voltage and the charging current decreases with the process, it is designated as the constant-voltage charging stage. Candidate stages are continuously maintained for a preset number of sampling periods before confirmation and conversion to avoid repeated stage switching caused by fluctuations in the control quantity near the target value.

[0032] In step S2, the edge computing unit reads the currently valid staged Gaussian process regression standard curve model based on the model identifier. The model takes the normalized charging process τ, state of charge (SOC), ambient temperature T_env, and battery health state (SOH) as inputs, and establishes sub-models for the charging current I, total charging voltage U, maximum temperature T_max, and single-cell voltage range ΔU. The model package stores the input and output standardized parameters, kernel function parameters, training data range, normal residual covariance matrix, conditional thresholds, model version, and validation information.

[0033] For historical records, the normalized charging progress at sampling time t_i within stage s is τ_i = (t_i - t_s,start) / (t_s,end - t_s,start), where t_s,end is greater than t_s,start, and τ_i is restricted to between 0 and 1. For real-time charging processes, τ is estimated based on the completed stage progress, changes in state of charge, and the length of historical stages identified by the same model; the estimation result remains monotonically increasing. When the stage duration cannot be reliably estimated, the state of charge relative to the progress is used as the initial position, and further correction is performed in step S3.

[0034] In step S3, a sliding time window of 5 s in the current charging phase is saved. First, an initial mapping is established based on a unified timestamp, phase transition position, state-of-charge interval boundaries, and the positions of low-frequency current and voltage slope changes. Then, dynamic time warping is performed, constrained by time monotonicity, phase boundaries, and maximum scaling. Taking a 100 ms sampling period as an example, the maximum warping width can be set to 10 sampling points. If the optimal warping path continuously lies on the warping boundary or the average local cost exceeds the alignment quality threshold, it is marked as insufficient curve mapping reliability, and the result is not used to improve the deviation level.

[0035] Based on the corrected mapping position, the real-time charging parameter vector y_real(t) and the predicted mean vector μ(t) are obtained, forming the physical residual vector e(t) = y_real(t) - μ(t). Let the output scale matrix in the model package be A = diag(s_y,1,…,s_y,p), where each s_y,j is greater than the corresponding minimum scale. Then the standardized residual vector is r(t) = A^(-1)e(t). This processing ensures that the charging current, total charging voltage, maximum temperature, and individual cell voltage range are included in subsequent calculations within a unified dimensionless space.

[0036] In step S4, the model outputs the original-scale prediction covariance matrix Σ_GP(t), and forms a standardized model prediction covariance matrix according to C_GP(t)=A^(-1)Σ_GP(t)A^(-1). This is combined with the normal residual covariance matrix C_res,s from the current charging stage and the regularization term to form C(t)=C_GP(t)+C_res,s+λE_p. The initial value of λ is determined based on the minimum eigenvalue and condition number of the normal validation location. λ is gradually increased when C(t) cannot be stably decomposed; if it still cannot be decomposed after reaching the regularization upper limit, the current score is marked as invalid.

[0037] The joint bias calculation is performed by solving C(t)v(t)=r(t) through Cholesky decomposition, and then calculating D_M^2(t)=r(t)^Tv(t), without directly forming the explicit inverse matrix of C(t). The squared Mahalanobis distance values ​​within the sliding time window are weighted and aggregated. When using exponentially decaying weights: a_h=(1-ρ)ρ^h / (1-ρ^W), 0<ρ<1; J(t)=∑_(h=0)^(W-1)a_hD_M^2(t-h).

[0038] Where h = 0 represents the current valid sampling point, and W represents the number of valid sampling points participating in the aggregation. When there are missing positions within the window that do not participate in the calculation, the remaining weights are renormalized; when the number of valid positions is lower than the calculation condition, the output of J(t) is paused.

[0039] In step S5, the first-level threshold H_1, the second-level threshold H_2, and the third-level threshold H_3 corresponding to the current charging stage, the model prediction uncertainty interval, the ambient temperature interval, and the battery health state interval are read from the threshold parameter area, where H_1 < H_2 < H_3. As a calibration configuration, the three thresholds can be the 95th, 99th, and 99.9% conditional quantiles of the normal verification score, respectively. The above quantile levels are configuration examples and do not represent the corresponding detection accuracy.

[0040] In this embodiment, the number of consecutive exceedances for Level 1, Level 2, and Level 3 are set to 5, 3, and 1, respectively. When J(t) reaches the corresponding threshold but does not meet the consecutive exceedance requirement, only a candidate deviation is formed. The recovery threshold is determined according to R_l = κ_lH_l, where 0 < κ_l < 1; the deviation level is reduced after the joint deviation score is continuously lower than the recovery threshold and the recovery count is met. The deviation direction, main deviation parameters, trend deviation, and fluctuation deviation are determined based on the signed standardized residual, absolute contribution ratio, slope difference, and detrended residual variance ratio, respectively.

[0041] In step S6, upon entering the first-level deviation state, data for a preset duration before and after the deviation time is saved, and a request to increase the sampling frequency is sent to the data interface that supports sampling period adjustment; if the data interface does not support adjustment, the original sampling period is maintained. Upon entering the second-level deviation state, an alarm message containing model identifier, model version, charging stage, main deviation parameters, deviation type, joint deviation score, threshold source, data quality status, and curve mapping reliability is sent to the charging monitoring platform, and the charging equipment is requested to limit the charging power. Upon entering the third-level deviation state, a charging stop control command is output, and the charging equipment reduces the output current and disconnects the charging output according to the original safety control procedure.

[0042] After charging is complete, only records that did not trigger deviation alarms, did not trigger equipment safety protection, meet the data quality update conditions, and whose charging results are confirmed as normal are added to the candidate training set. After passing checks on normal sample prediction error, prediction interval coverage, stage boundary continuity, threshold order, and numerical stability, the candidate model is switched to the current effective model before the start of the next charging session; the model before the switch is retained as a backup model that has recently passed validation. If a candidate model fails validation, the current model and threshold remain unchanged. This update path corresponds to... Figure 1 The feedback path in.

[0043] Example 2 This embodiment combines Figure 2 This section describes the construction, training, prediction, and covariance formation process of a phased Gaussian process regression standard curve model. Model training uses only normal historical charging records that match the model identifier and pass data quality screening; records with deviation alarms, equipment protection, long periods of missing data, or unconfirmed charging results are not used.

[0044] Model training uses complete charging sessions as the data partitioning unit. All sampling points within the same charging session are kept on the same side of the training or validation set, and are not randomly partitioned according to individual sampling points. For each normal historical charging record, data for the pre-charging stage, constant current charging stage, and constant voltage charging stage are extracted based on the confirmed stage boundaries. The normalized charging progress at sampling time t_i within stage s is calculated according to τ_i ​​= (t_i - t_s, start) / (t_s, end - t_s, start), and is limited to 0 to 1. Records whose stage endpoint cannot be reliably determined are not included in the corresponding stage training set.

[0045] Historical records are resampled according to a uniform stage process grid, ensuring that charging records of different durations form the same number of candidate training points in each stage. When there are too many candidate training points, stratified sampling is performed according to normalized charging process, state of charge, ambient temperature, and battery health status. The number of representative points in the same input space partition for a single charging session is limited to avoid single records with long durations or large local fluctuations occupying too high a weight.

[0046] The input vector and output vector of the i-th training point are respectively: x_i=[τ_i, SOC_i, T_env,i, SOH_i]^T; y_i=[I_i, U_i, T_max,i, ΔU_i]^T.

[0047] The d-th input dimension is standardized using x′_(i,d)=(x_(i,d)-m_x,d) / s_x,d, where m_x,d is the training mean, and s_x,d is the larger of the training standard deviation and the preset minimum scale. The j-th output parameter is standardized using y′_(i,j)=(y_(i,j)-m_y,j) / s_y,j, where s_y,j is guaranteed to be positive in the same way. The input and output means, scale parameters, and approximately constant dimension labels are saved with the model package, and real-time data is processed using the same rules.

[0048] Three charging stages are established, each with its own sub-model: charging current, total charging voltage, maximum temperature, and individual cell voltage range. Therefore, a complete model version includes twelve Gaussian process regression sub-models. Stage-by-stage modeling allows for separate fitting of the current holding characteristics of the constant current charging stage and the current decreasing characteristics of the constant voltage charging stage, avoiding the weakening of stage transition characteristics by smoothing across stages.

[0049] The radial basis kernel function of the j-th output parameter in stage s is: k_s,j(x_i′,x_q′)=σ_f,s,j^2 exp{-(1 / 2)(x_i′-x_q′)^TΛ_s,j^(-1)(x_i′-x_q′)}; Λ_s,j=diag(l_s,j,τ^2,l_s,j,SOC^2,l_s,j,T^2,l_s,j,SOH^2).

[0050] Where σ_f,s,j^2 is the signal variance, and l_s,j,τ, l_s,j,SOC, l_s,j,T, and l_s,j,SOH are the positive length scales of the corresponding input dimensions. The input has been standardized, so each length scale is used to express the relative influence of different input dimensions on the j-th output term. The element in the a-th row and b-th column of the training kernel matrix K_s,j is k_s,j(x_a′, x_b′). Considering observation noise, the training covariance matrix is: K_y,s,j=K_s,j+σ_n,s,j^2E_n, Where σ_n,s,j^2 is the positive observation noise variance, and E_n is the n-order identity matrix. The standardized training output vector is denoted as y_j′. The hyperparameter set includes the signal variance, four length scales, and observation noise variance, with upper and lower bounds set in the logarithmic domain. The logarithmic marginal likelihood of the j-th output parameter is: L_s,j=-(1 / 2)y_j′^TK_y,s,j^(-1)y_j′-(1 / 2)ln|K_y,s,j|-(n / 2)ln(2π).

[0051] Training employs multiple initial length scales to initiate bounded numerical optimization, with parameter combinations exhibiting high log-marginal likelihood and stable decomposition as candidate results. The objective function is computed through Cholesky decomposition and trigonometric equation solving, without calculating the explicit inverse matrix of K_y,s,j. When decomposition fails, a numerical stabilization term εE_n is added to the diagonal of K_y,s,j; if decomposition still fails even after ε reaches its upper bound, the current parameter combination is stopped. If all initial combinations fail, no publishable submodel is generated, and repeated training points, scale parameters, and the lower bound of noise variance are checked.

[0052] For the input to be predicted, x_*′, and the kernel function value vector between the training input and the input to be predicted, denoted as k_*, then the standardized prediction mean and the standardized latent function prediction variance are: μ_j′(x_*)=k_*^TK_y,s,j^(-1)y_j′; v_j′(x_*)=k_s,j(x_*′,x_*′)-k_*^TK_y,s,j^(-1)k_*.

[0053] If v_j′ has a value less than zero only due to floating-point error, it is corrected to zero; when the value is lower than the preset minimum prediction variance, the minimum value is used as the lower limit for subsequent calculations. The original scale prediction mean and prediction variance are as follows: μ_j(x_*)=m_y,j+s_y,jμ_j′(x_*); σ_GP,j^2(x_*)=s_y,j^2v_j′(x_*).

[0054] The four original-scale prediction means are combined to form μ(t), and the four original-scale prediction variances are combined to form Σ_GP(t) = diag[σ_GP,I^2(t), σ_GP,U^2(t), σ_GP,T^2(t), σ_GP,ΔU^2(t)]. Each sub-model predicts independently, therefore the off-diagonal elements of Σ_GP(t) are zero; the normal correlation between charging parameters is expressed by the validation residual covariance.

[0055] Let A = diag(s_y,1,…,s_y,p), then the standardized model prediction covariance matrix is ​​C_GP(t) = A^(-1)Σ_GP(t)A^(-1). For the q-th normal validation sampling position in stage s, the standardized validation residual is r_q = A^(-1)[y_q-μ_q]. Its sample mean and sample covariance are respectively: ; .

[0056] Where N_s is greater than 1. When the validation sample is insufficient or the condition number R_s exceeds the allowable range, a contraction estimation is used. C_res,s=(1-γ_s)R_s+γ_sdiag(R_s), 0≤γ_s≤1.

[0057] A smaller γ_s is used when the validation sample coverage is sufficient and the matrix is ​​stable; γ_s is increased when the correlation estimation is unstable. When the diagonal elements are below the preset lower limit of residual variance, the corresponding lower limit is used to avoid the normal residual variance of a single dimension approaching zero.

[0058] Real-time detection uses C(t) = C_GP(t) + C_res,s + λE_p. C_GP(t), C_res,s, and λE_p are all located in a dimensionless standardized space, thus avoiding the problem of directly adding covariances of different physical units to a single scalar. Before model deployment, the initial value, increase rules, and upper limit of λ are determined based on the minimum eigenvalue and condition number at normal validation locations.

[0059] For a single parameter, a reference boundary can be formed at the original scale: B_j^±(t)=μ_j(t)±z_(1-δ / 2)s_y,j√[C_GP,jj(t)+C_res,s,jj].

[0060] Where z_(1-δ / 2) is the quantile of the standard normal distribution, and δ is the preset tail probability. The reference boundary is used to display the position of individual parameters relative to the standard curve and to check the interval coverage of normal validation data, and is not used as a comparison threshold for joint bias scoring. The regularization term λE_p is only used for the numerical stability of the multi-parameter matrix and is not included in the above-mentioned individual reference boundary.

[0061] Model validation is performed on a per-complete charging session basis, checking the mean of prediction residuals, prediction error, reference boundary coverage, stage boundary continuity, and numerical stability of the kernel matrix. When the charging control strategy allows parameters to jump at stage boundaries, the model package stores the allowed jump direction and range; when there are no control jumps, the difference between the prediction means of adjacent stages should be within the allowable range of boundary continuity. Sub-models that fail validation are not included in the publishable model version.

[0062] Each validated model version saves the training representative points, kernel function parameters, input / output normalization parameters, matrix factorization results, normal residual covariance matrix, regularization parameters, model identifier, training data range, and validation result summary for each of the twelve sub-models. When the real-time input exceeds the training data coverage, an extrapolation position marker is output; when the extrapolation exceeds the allowed range, a status indicating insufficient model applicability is output, without increasing the bias level based on the model results at that position. Figure 2 The predicted mean vector, standardized model prediction covariance, normal residual covariance, and regularized covariance are formed by the above processes, respectively.

[0063] Example 3 This embodiment combines Figure 3 and Figure 4 This section explains real-time curve mapping, joint deviation calculation, adaptive grading threshold, persistence determination, and deviation type identification. Figure 3 The vehicle-side battery management system data and the charging pile data are initially aligned and dynamically time-normalized to form a mapping position. The model prediction results and standardized residuals are then used for joint bias calculation. Figure 4 The normal verification score distribution and operating conditions form a graded threshold, and the joint deviation score drives the state transition and graded response after continuous judgment.

[0064] Before curve mapping, timestamps, field integrity, and communication status are checked according to the charging session. When the vehicle battery management system timestamp is t_B and the charging device timestamp is t_C, the initial time offset Δt_0 is estimated based on events such as charging start, contactor closure, effective current establishment, and stage transition. A fixed correction is used when the offset is stable, and a piecewise linear correction is used when the offset changes slowly over time. Data that has completed time synchronization is merged according to a preset time tolerance, and missing markers are retained for sampling points that fail to be merged.

[0065] The initial curve position is determined based on the charging stage, state of charge (SOC), and normalized charging progress within the stage. Let the start time of the current stage be t_s, and the stage duration estimated as T_s, where T_s is greater than the minimum allowable value. Then τ(t) = (t - t_s) / T_s, and τ is restricted to between 0 and 1. If T_s cannot be reliably obtained, the initial position is determined based on the progress of the current SOC relative to the stage start and target SOC. As real-time data increases, the stage duration estimate is recursively updated, but the mapped position remains monotonic.

[0066] A real-time feature sequence P = {p_1, ..., p_m} is constructed within a real-time alignment window of length W_a, and candidate sequences Q = {q_1, ..., q_n} are truncated from the initial position of the standard sequence for the corresponding stage. Alignment features prioritize state of charge, normalized charging progress within the stage, and charging control state. Current and voltage slopes, smoothed by low-frequency processing, are used as low-weight auxiliary features. The instantaneous amplitude of the parameter to be detected does not have a major weight in the alignment cost to avoid the deviation amplitude being absorbed by time axis scaling.

[0067] The local costs of the standardized feature vectors p_a and q_b are: d(a,b)=∑_kω_k[p_(a,k)-q_(b,k)]^2, Where ω_k≥0 and ∑_kω_k=1. The cumulative cost is: L(a,b)=d(a,b)+min{L(a-1,b),L(a,b-1),L(a-1,b-1)}.

[0068] The warping path is not allowed to be reverse-mapped or cross the boundaries of the pre-charging stage, constant current charging stage, and constant voltage charging stage, and must satisfy |a-b|≤B, where B is the maximum warping width. The stage transition region is jointly confirmed by the charging control state and changes in current and voltage; after the new stage continuously meets the confirmation conditions, the warping path is re-initialized to prevent the accumulated offset of the previous stage from being passed to the next stage.

[0069] When the average local cost of the optimal path is lower than the alignment quality threshold and the mapping position increment satisfies the monotonic constraint, the current mapping position is output. When the average local cost exceeds the limit, the path is continuously located on the regularity boundary, or there is a stage information conflict, the output curve mapping reliability is insufficient. If this state lasts for a short period, the previous valid position is used and progress is made according to the sampling period; if it lasts for more than the set duration, the deviation level improvement is paused and the abnormal state of the curve mapping is output.

[0070] At the effective mapping position, obtain y_real(t), μ(t), Σ_GP(t), and A, calculate r(t) = A^(-1)[y_real(t)-μ(t)], and then construct C_GP(t) = A^(-1)Σ_GP(t)A^(-1) and C(t) = C_GP(t)+C_res,s+λE_p. Use Cholesky decomposition to solve C(t)v(t)=r(t), and calculate D_M^2(t)=r(t)^Tv(t). If the decomposition fails, increase λ according to the rules in the model package; if it still fails after reaching the upper limit, do not output the current score.

[0071] Model prediction uncertainty is represented by u(t) = tr[C_GP(t)] / p, where tr[·] represents the trace of the matrix and p represents the dimension of the detection parameters. Normal validation data are first grouped according to the charging stage, and then divided into condition intervals according to model prediction uncertainty, ambient temperature, and battery health status. Within each condition group, the empirical distribution of the normal joint deviation score J is statistically analyzed, and a threshold is formed according to the following formula: H_l(s, b_u, b_T, b_H)=Q_(α_l){J_normal|s, b_u, b_T, b_H}, l∈{1, 2, 3}.

[0072] Where b_u, b_T, and b_H represent the intervals of model prediction uncertainty, ambient temperature, and battery health status, respectively, and Q_(α_l) represents the quantile value of α_l, where α_1 < α_2 < α_3. When the number of samples in the condition group is insufficient, adjacent health status intervals, temperature intervals, or uncertainty intervals are merged according to a preset order; when the current operating condition is between the centers of adjacent groups, adjacent thresholds are linearly interpolated. After the thresholds are generated, a monotonicity check is performed to ensure that H_1 < H_2 < H_3.

[0073] Figure 4The normal monitoring state N_0, first-level deviation state N_1, second-level deviation state N_2, and third-level deviation state N_3 correspond to normal monitoring, data recording, alarm and power limiting, and charging interruption, respectively. A candidate deviation is formed when the joint deviation score first exceeds a certain level threshold; it only enters the corresponding level after the number of consecutive exceedances reaches n_l. Higher levels can be configured with smaller n_l; when J(t) directly reaches H_3 and meets the three-level data quality conditions, it can move from the current state to N_3.

[0074] To avoid repeated state switching caused by score fluctuations around the threshold, the recovery threshold for level l is R_l = κ_lH_l, where 0 < κ_l < 1. A lower state is allowed only when J(t) has m_l consecutive valid sampling points below R_l. When in N_3, the charging equipment must confirm that the charging stop action has been completed and the relevant parameters have recovered to the allowable range before entering a lower state. Figure 4 The paths from N_0 to N_1, from N_1 to N_2, and from N_2 to N_3 represent the upgrade process, while the reverse paths represent the state reduction process after the recovery conditions are met.

[0075] Bias type identification uses signed standardized residuals and the absolute contribution ratio of parameters. The signed standardized residual of the j-th parameter is: z_j(t)=r_j(t) / √C_jj(t).

[0076] Let g_j(t) = r_j(t)[C(t)^(-1)r(t)]_j, and let G(t) = ∑_(k=1)^p|g_k(t)|. When G(t) is greater than the preset lower limit, the absolute contribution ratio of the j-th parameter is c_j(t) = |g_j(t)| / G(t); when G(t) is lower than the lower limit, the main deviation parameter is uncertain. A weighted average of c_j(t) within the detection window is taken, and the parameter with the larger average absolute contribution ratio is determined as the main deviation parameter, while other parameters that meet the correlation conditions are determined as accompanying deviation parameters.

[0077] A weighted average of z_j(t) within the detection window is calculated. A positive bias is determined when the average value is consistently greater than the positive recognition threshold, and a negative bias is determined when it is consistently less than the negative recognition threshold. A weighted linear fit is performed with time within the window as the independent variable and z_j(t) as the dependent variable to obtain the residual change slope β_j. A trend bias is determined when |β_j| exceeds the normal residual slope threshold for the same period and consistently meets the condition.

[0078] Volatility deviation is determined by comparing the residual variance V_j (after removing the linear trend) with the normal reference variance V_ref,j. When V_ref,j is below the lower limit of variance, this lower limit is used, and the volatility ratio is F_j = V_j / V_ref,j. Volatility deviation is determined when F_j continuously exceeds the corresponding stage limit and positive and negative residuals alternate within the window. The same event can simultaneously possess directional, trend, and volatility attributes. The system determines the primary deviation type based on the average absolute contribution ratio and duration; other valid types are saved as associated deviation types.

[0079] The output types include deviation level, main deviation parameters, accompanying deviation parameters, direction attribute, trend attribute, fluctuation attribute, charging stage, standard curve mapping position, joint deviation score, threshold source, data quality status, and curve mapping confidence. This information, along with the original data, model version, matrix factorization status, and response action, is written into the same deviation event, enabling... Figure 3 Data processing path and Figure 4 The state transition path can be reviewed according to the charging session.

[0080] Example 4 This embodiment combines Figure 5 This section describes the charging curve deviation detection system. The system includes a data acquisition and stage processing module, a standard curve modeling module, a curve mapping and residual generation module, a joint deviation calculation module, an adaptive judgment module, and a response and model update module. The functions of each module are explained first, followed by their implementation methods.

[0081] The data acquisition and stage processing module is used to acquire normal historical charging data and real-time charging data that match the model identifier, and to complete field parsing, quality filtering, time alignment, filtering, scaling and charging stage division, and to attach charging session identifier, source identifier, validity identifier and stage identifier to the data.

[0082] The standard curve modeling module is used to establish Gaussian process regression standard curve models for the pre-charging stage, constant current charging stage, and constant voltage charging stage, respectively. It outputs the predicted mean vector and the standardized model predicted covariance matrix, forms the normal residual covariance matrix, and saves the model identifier, applicable range, kernel function parameters, output scale parameters, and model version.

[0083] The curve mapping and residual generation module is used to establish an initial mapping based on a unified timestamp, stage transition position, state of charge, and curve change position. It determines the position of the real-time sliding time window in the standard curve through dynamic time warping constrained by time monotonicity, stage boundaries, and maximum scaling, and forms the physical residual vector, standardized residual vector, and curve mapping credibility.

[0084] The joint bias calculation module is used to form a regularized covariance matrix based on the predicted covariance matrix, normal residual covariance matrix, and regularization term of the standardized model. It calculates the squared Mahalanobis distance through matrix decomposition, performs time-series aggregation on the squared Mahalanobis distance within the sliding time window, and outputs the joint bias score and numerical validity status.

[0085] The adaptive judgment module is used to select the grading threshold based on the charging stage, model prediction uncertainty, ambient temperature, battery health status and normal verification score distribution, perform continuous over-limit counting, recovery counting and state transition, and determine the deviation type based on the signed standardized residual, absolute contribution ratio, slope difference and detrended residual variance ratio.

[0086] The response and model update module is used to perform data recording, alarms, charging power limit requests, or charging stop control according to the deviation level, and to associate and save real-time data, model output, judgment results, and response actions; it is also used to filter and confirm normal data, generate candidate models and candidate thresholds, and complete verification, release, version recording, and anomaly rollback.

[0087] In its implementation, the data acquisition and stage processing module includes a vehicle-side communication adapter unit, a charging pile-side communication adapter unit, a session management unit, a time synchronization unit, a data quality check unit, a filtering unit, and a stage identification unit. The vehicle-side communication adapter unit receives data from the battery management system via the controller local area network or Ethernet, while the charging pile-side communication adapter unit receives output voltage, output current, charging stage, power command, and device status via the charging equipment's internal bus, Ethernet, or isolated serial interface.

[0088] The session management unit generates a charging session identifier based on the vehicle identifier, battery identifier, charging equipment identifier, and charging start time. The time synchronization unit reads the vehicle-side timestamp, the charging pile-side timestamp, and the local received time, and estimates the time offset through charging start, current establishment, and stage transition events. The data quality check unit checks field integrity, physical range, rate of change, and communication status; the filtering unit outputs detection data but does not overwrite the original data; the stage identification unit determines the three charging stages by comprehensively considering the control status, state of charge, and current and voltage changes.

[0089] The standard curve modeling module includes a sample selection unit, a training set generation unit, a sub-model training unit, a residual covariance estimation unit, a model validation unit, and a model packaging unit. The sample selection unit excludes records with vehicle malfunctions, charging equipment malfunctions, prolonged communication interruptions, deviation alarms, or abnormal terminations. The training set generation unit divides training and validation data based on complete charging sessions. The sub-model training unit trains twelve sub-models according to the kernel function, log-marginal likelihood, and numerical stability rules in Example 2. The model packaging unit encapsulates representative training points, standardized parameters, matrix factorization results, normal residual covariance, conditional thresholds, applicable scope, and validation information into a model package.

[0090] The curve mapping and residual generation module includes a stage position initialization unit, a real-time alignment cache unit, a constraint regularization unit, and a mapping quality evaluation unit. The stage position initialization unit generates the initial position within the stage; the real-time alignment cache unit saves the alignment features of the current stage in chronological order; the constraint regularization unit only calculates the cumulative cost within the maximum regularization width; and the mapping quality evaluation unit outputs the curve mapping reliability based on the average local cost, regularization boundary occupancy, and mapping position increment. After a stage switch, the regularization path of the previous stage is cleared and re-initialized.

[0091] The joint bias calculation module includes a model inference unit, a residual calculation unit, a covariance construction unit, a numerical solution unit, and a time window aggregation unit. The model inference unit obtains μ(t), Σ_GP(t), and A; the residual calculation unit forms e(t) and r(t); the covariance construction unit forms C(t); the numerical solution unit calculates D_M^2(t) using Cholesky decomposition; and the time window aggregation unit outputs J(t). In cases of missing data, failed matrix decomposition, or insufficient effective window length, this module outputs an invalid status instead of a numerical score.

[0092] The adaptive decision-making module includes a threshold matching unit, a state decision-making unit, and a type recognition unit. The threshold matching unit reads or interpolates H_1, H_2, and H_3 based on the operating conditions; the state decision-making unit saves the current state, candidate counts for each level, and recovery counts; the type recognition unit calculates z_j(t), c_j(t), β_j, and F_j. When the data quality, curve mapping reliability, or model applicability does not meet the conditions, the state decision-making unit pauses increasing the deviation level based on the current results and outputs the corresponding reason for detection unavailability.

[0093] The response and model update module includes an event generation unit, a data retention unit, an alarm output unit, a control request unit, an execution confirmation unit, a candidate model management unit, and a version rollback unit. The control request includes a charging session identifier, request level, request time, and validity period to prevent requests from the previous charging session from being used in the next. The execution confirmation unit reads the control results returned by the charging device and checks whether the power limiting or charging abort actions have taken effect. Candidate models are published using a "load first, verify later, switch later" approach. In case of model loading errors, consecutive numerical anomalies, or threshold matching anomalies, the system reverts to the most recently validated backup model.

[0094] like Figure 5 As shown, the functional modules are connected in the following order: data acquisition and stage processing module, standard curve modeling module, curve mapping and residual formation module, joint deviation calculation module, adaptive judgment module, and response and model update module. The standard curve modeling module provides model results to the curve mapping and residual formation module and the joint deviation calculation module. The response and model update module provides the validated model and threshold updates to the standard curve modeling module and the adaptive judgment module. The vehicle-side data interface and the charging pile-side data interface are connected to the data acquisition and stage processing module, and the alarm output interface and the charging control interface are connected to the response and model update module.

[0095] The system hardware includes a processor, volatile memory, non-volatile memory, a communication interface, a clock unit, and a power monitoring unit. The memory is divided into raw charging data area, alignment data area, standard curve model area, threshold parameter area, detection result area, and model version area. The processor calls each module according to the sampling period; when computing resources are insufficient, the priority of candidate model processing and historical data uploading is reduced, while real-time data access, joint deviation calculation, event logging, and security control interface processing are retained.

[0096] The real-time module can be deployed on edge computing devices at charging stations, while model training and candidate model validation can be deployed on a remote charging monitoring platform. When remote communication is interrupted, the edge computing device continues to use the currently valid local model for detection. Once the connection is restored, it uploads cached records and receives new models. When a model is unavailable, the system does not output normal conclusions based on the model and maintains the original safety protection functions of the battery management system and charging equipment.

[0097] Example 5 This embodiment combines Figure 6This invention describes its application in DC charging stations. The application scenario includes vehicles waiting to be charged, DC charging equipment, edge computing devices at the charging station, station monitoring terminals, and a remote charging monitoring platform. The vehicle battery management system and the charging equipment controller provide vehicle-side data and charging pile-side data to the edge computing device, respectively; the edge computing device performs real-time processing and returns a power limit request or charging stop control command to the charging equipment; the remote charging monitoring platform stores historical data, detection records, and model versions, and publishes validated models and thresholds to the edge computing device.

[0098] The application at the charging station requires that vehicle-side and charging pile-side data be correlated on a unified timeline according to the charging session. Local detection should continue to operate even when the remote network is interrupted. Model inference should not interrupt the original control tasks of the charging equipment. Deviation events should be traceable to the original data, model version, threshold source, and response results. When the detection system encounters data, model, or storage anomalies, it should output a definite degraded status, and should not indicate that the charging process is normal based on results lacking evidence.

[0099] As a hardware configuration, the site edge computing device employs an industrial computer with eight general-purpose computing cores, 32 GB of volatile memory, and 1 TB of industrial solid-state storage, and is equipped with independent Ethernet interfaces connecting the charging equipment's business network and the site monitoring network; a controller area network interface or an isolated serial interface is configured depending on the on-site communication method. The remote charging monitoring platform uses a general-purpose server to perform historical data management, model training, and version release. The above specifications are examples of implementable configurations and do not limit the processor architecture, storage capacity, or number of interfaces, nor do they represent performance test results.

[0100] Edge computing devices are configured with a ring-shaped real-time data cache, a current model area, a backup model area, a deviation event area, and an operation log area. The current model area stores the model package currently in use, while the backup model area stores the most recently validated backup model. The remote charging monitoring platform manages training data according to model identifiers, and does not directly include data with inconsistent battery chemistry systems, rated capacity ranges, or charging strategy versions into the same model.

[0101] This embodiment employs a 100 ms basic sampling period, a 5 s scoring sliding time window, and a maximum dynamic time warping width of 10 sampling points. Marked, limited interpolation is performed when no more than three consecutive sampling points are missing; joint bias scoring is paused when more than three sampling points are missing. The number of consecutive out-of-limit values ​​for the three levels is configured as 5, 3, and 1, respectively. The thresholds are calibrated based on the 95%, 99%, and 99.9% conditional quantiles of normal validation data, respectively, with the recovery threshold lower than the corresponding upgrade threshold. The above parameters are used to illustrate the configuration process; in actual applications, they should be recalibrated based on sampling capabilities and normal validation data.

[0102] After the vehicle connects to the DC charging equipment, the edge computing device generates a model identifier based on the vehicle's battery information and the charging equipment's charging strategy, and loads the corresponding local model. Once the model's integrity and applicability are verified, a charging session cache is established; if the verification fails or a matching model does not exist, a model unavailable status is output, and no other model is used as a substitute.

[0103] After charging begins, the data acquisition and stage processing module receives data from both the vehicle and the charging pile, completing time synchronization, quality checks, smoothing, and stage identification. The curve mapping and residual formation module performs constrained dynamic time warping based on the initial position within each stage. The joint deviation calculation module calls the corresponding stage sub-model to obtain the predicted mean and predicted variance, forming C(t), D_M^2(t), and J(t) in the standardized space. The adaptive judgment module obtains the grading threshold corresponding to the current operating conditions and updates the candidate count, deviation level, and deviation type.

[0104] In the first data state used to illustrate the processing logic, communication link resynchronization causes a short-term jump in the charging current of two consecutive sampling points, and the joint deviation score reaches the first-level threshold, but the number of consecutive sampling points does not reach five. The system records candidate deviations, clears the candidate count after the score drops, and the state remains N_0. This result is used to illustrate the persistence determination logic and is not used as false alarm rate statistics.

[0105] In the second data state, after the vehicle enters the constant voltage charging stage, the individual cell voltage range gradually increases relative to the predicted mean, and changes jointly with the total charging voltage residual. When J(t) reaches H_1 for five consecutive valid sampling points, it enters N_1, and the system saves the data before and after the deviation; when it reaches H_2 for three consecutive valid sampling points, it enters N_2. The identification result can be represented as a positive trend deviation of the individual cell voltage range as the main parameter, with the total charging voltage deviation as the accompanying parameter, and a power limiting request is sent to the charging equipment. After the limit takes effect, J(t) gradually decreases its state when it continuously meets the recovery conditions; when J(t) continues to reach H_3, it enters N_3 and outputs a charging stop control command.

[0106] In low-temperature operation, the model prediction uncertainty may be higher than that of the training dataset. The system evaluates the current score using C_GP(t) and the corresponding uncertainty condition threshold; when the real-time temperature exceeds the model's allowable range, it outputs a state indicating insufficient applicability of the model, and does not accept real-time residuals indefinitely with increased prediction variance. Therefore, model uncertainty only participates in scoring and threshold selection within the existing calibration range.

[0107] After charging is completed, the edge computing device summarizes data integrity, highest deviation level, vehicle malfunction, charging equipment malfunction, and termination reason. Sessions experiencing deviations, equipment protection issues, model mismatches, or prolonged data loss are not included in the candidate normal dataset. Sessions with consistently normal detection status, complete data, and normal charging completion are included in the candidate data area and uploaded to the remote charging monitoring platform. Candidate models are published after being checked using independent validation sessions. The edge computing device switches versions at the charging session boundary after completing file verification and trial loading.

[0108] In the above scenarios, short-term data jumps are handled by quality checks, time window aggregation, and continuous counting; local charging process differences are handled by constrained curve mapping; multi-parameter correlation deviations are handled by standardized covariance and Mahalanobis distance; and persistence is handled by graded thresholds and recovery hysteresis. Deviation events save the original data, mapping location, model version, score, threshold source, and control confirmation, enabling the detection process to be traced back to the charging session; confirmed normal data is validated by candidate models before participating in updates, ensuring that real-time deviation data does not directly change the current detection benchmark.

[0109] The above description is merely a specific embodiment of the present invention. Those skilled in the art can make adjustments to the data interface, sampling period, model training scale, condition interval, threshold quantile level, and deployment location without departing from the technical concept of the present invention; such adjustments should maintain the technical relationship between model identifier matching, phased Gaussian process regression, constrained curve mapping, standardized joint bias calculation, persistence grading determination, and validated model updates.

Claims

1. A method for detecting charging curve deviation based on Gaussian process regression, characterized in that, include: S1, acquire normal historical charging data and real-time charging data that match the model identifier of the battery to be tested, and perform quality screening, time alignment and charging stage division on the data; S2, establish a Gaussian process regression standard curve model for different charging stages, and output the predicted mean vector and the model predicted covariance matrix; S3 maps the sliding time window of real-time charging data to the standard curve of the corresponding charging stage, obtains the real-time charging parameter vector and the predicted mean vector according to the mapping position, and forms a standardized residual vector according to the output scale parameter. S4. Construct a regularized covariance matrix based on the model-predicted covariance matrix and the normal residual covariance matrix, calculate the squared Mahalanobis distance and aggregate it using a sliding time window to obtain the joint bias score. S5. An adaptive grading threshold is determined based on the charging stage, model prediction uncertainty, ambient temperature, battery health status, and the score distribution of normal verification data. The deviation level is determined by comparing the joint deviation score with the threshold for persistence, and the deviation type is determined based on the residual direction, slope change, and fluctuation change. S6. Perform data recording, alarm or charging protection according to the deviation level, associate and save the detection result with the real-time charging data, and generate candidate models and candidate thresholds using the confirmed normal charging data. After verification, update the standard curve model and adaptive grading threshold.

2. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In step S1, the model identifier is formed based on the vehicle model, battery chemistry system, battery rated capacity, module configuration, and charging protocol; the normal historical charging data is a charging record in which no charging protection event has occurred and the data integrity meets the training conditions; the charging data includes charging time, state of charge, charging current, total charging voltage, highest voltage of a single cell, lowest voltage of a single cell, highest temperature, ambient temperature, and battery health status, and the single cell voltage range is obtained based on the difference between the highest voltage of a single cell and the lowest voltage of a single cell; the vehicle-side battery management system data and the charging pile-side data are aligned according to timestamps and changes in charging status, abnormal sampling points and missing sampling points are marked and processed, and the charging process is divided into a pre-charging stage, a constant current charging stage, and a constant voltage charging stage based on the charging control mode, state of charge, and current or voltage change characteristics.

3. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In step S2, the normalized charging process τ, state of charge (SOC), ambient temperature T_env, and battery health state (SOH) during the charging stage are used to form the model input vector x, and the charging current I, total charging voltage U, maximum temperature T_max, and single-cell voltage range ΔU are used to form the model output vector y. Gaussian process regression sub-models are established for each charging stage and each output parameter, and the radial basis kernel function is used after standardizing the input vector. k_s,j(x_i′,x_q′)=σ_f,s,j^2 exp{-(1 / 2)(x_i′-x_q′)^T Λ_s,j^(-1)(x_i′-x_q′)}; Where s represents the charging stage, j represents the output parameters, x_i′ and x_q′ represent the standardized input vectors, σ_f,s,j^2 represents the signal variance, and Λ_s,j represents the diagonal matrix composed of the squares of the length scales of each input dimension. The signal variance, length scale, and observation noise variance are determined by maximizing the logarithmic marginal likelihood function of the training data, and the standard curve model is verified using normal charging records that are not involved in parameter optimization.

4. The charging curve deviation detection method based on Gaussian process regression according to claim 3, characterized in that, The training value of the j-th output parameter is standardized into a vector y_j′, and K_y,s,j=K_s,j+σ_n,s,j^2E_n is set, where K_s,j represents the training kernel matrix, σ_n,s,j^2 represents the observation noise variance, and E_n represents the n-order identity matrix. The standardized prediction mean μ_j′(x_*) and standardized prediction variance v_j′(x_*) are calculated at the input x_*′ to be predicted according to the following formula: μ_j′(x_*)=k_*^T K_y,s,j^(-1)y_j′; v_j′(x_*)=k_s,j(x_*′,x_*′)-k_*^T K_y,s,j^(-1)k_*; Where k_* represents the kernel function value vector between the training input and the input to be predicted; the original scale prediction mean and prediction variance are recovered based on the training mean m_y,j of the j-th output parameter and the positive output scale s_y,j, and the prediction variances of each term are used to form the original scale prediction covariance matrix Σ_GP(t). Let A = diag(s_y,1,…,s_y,p), and form the standardized model prediction covariance matrix according to C_GP(t) = A^(-1)Σ_GP(t)A^(-1); the normal residual covariance matrix C_res,s is formed based on the standardized residuals of the normal validation data, and the regularized covariance matrix is ​​constructed according to C(t) = C_GP(t) + C_res,s + λE_p, where λ is a positive dimensionless regularization coefficient and E_p is a p-order identity matrix.

5. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In steps S3 and S4, the boundaries of the state of charge interval, the transition positions of the charging stage, and the positions of changes in current or voltage slope are extracted from the vehicle-end data and the pile-end data. An initial time mapping is established based on the extracted positions, and the initial time mapping is corrected by dynamic time warping under the constraints of time mapping monotonicity, charging stage boundaries, and maximum expansion. A physical residual vector e(t) = y_real(t) - μ(t) and a standardized residual vector r(t) = A^(-1)e(t) are formed at the corrected mapping positions, and the squared Mahalanobis distance is calculated according to the following formula: D_M^2(t)=r(t)^TC(t)^(-1)r(t); The squared Mahalanobis distance values ​​within a sliding time window of length W are weighted and aggregated to obtain the joint bias score J(t)=∑_(h=0)^(W-1)a_hD_M^2(t-h), where a_h is a non-negative time weight, and the sum of the time weights participating in the aggregation is 1.

6. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In step S5, the joint deviation scores are conditionally grouped according to the charging stage, model prediction uncertainty range, ambient temperature range, and battery health status range to which the normal verification data belongs. First, second, and third thresholds are formed by using the first, second, and third quantiles of the joint deviation scores in each conditional group, respectively, resulting in progressively increasing thresholds. When the number of samples in a conditional group is less than the threshold calibration number, adjacent conditional groups are merged. When the current operating conditions are between the centers of adjacent groups, adjacent thresholds are interpolated. Continuous over-limit counting conditions and recovery counting conditions below the recovery threshold are set for the three levels. The deviation level is increased when the joint deviation score continuously reaches the corresponding level threshold, and decreased when the joint deviation score continuously falls below the corresponding recovery threshold.

7. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In step S5, the signed standardized residual of the j-th charging parameter is calculated according to z_j(t)=r_j(t) / √C_jj(t); the absolute contribution ratio of the j-th charging parameter is taken as the proportion of |r_j(t)[C(t)^(-1)r(t)]_j| to the sum of the corresponding absolute values ​​of each term; the main deviation parameter is determined according to the absolute contribution ratio; the positive or negative deviation is determined according to the signed standardized residual of the main deviation parameter; the trend deviation is determined according to the difference between the slope of the real-time curve and the slope of the predicted mean curve within the sliding time window; the fluctuation deviation is determined according to the ratio of the variance of the detrended residual to the variance of the normal residual; and the remaining deviation types that are simultaneously true are saved as associated deviation types.

8. The charging curve deviation detection method based on Gaussian process regression according to claim 1, characterized in that, In step S6, the first-level deviation corresponds to saving the charging data within a preset time range before and after the deviation time and requesting an increase in the data acquisition or detection frequency; the second-level deviation corresponds to sending alarm information containing deviation parameters, deviation type, joint deviation score and model version to the charging monitoring platform and requesting the charging equipment to limit the charging power; the third-level deviation corresponds to outputting a charging stop control command and saving the protection event record; after a charging process is completed, only charging records that have not generated deviation alarms, whose data quality meets the update conditions and are confirmed to be in a normal state by the charging results are added to the candidate training set. The candidate model replaces the current model after being verified by normal sample prediction error, prediction interval coverage, stage boundary continuity, threshold order and numerical stability, and the model parameters and threshold parameters before and after the replacement are saved for rollback.

9. A charging curve deviation detection system based on Gaussian process regression, characterized in that, include: The data acquisition and stage processing module is used to acquire normal historical charging data and real-time charging data that match the model identifier of the battery to be tested, and to perform quality screening, time alignment and charging stage division on the data. The standard curve modeling module is used to establish a Gaussian process regression standard curve model for different charging stages and output the predicted mean vector and the standardized model predicted covariance matrix. The curve mapping and residual generation module is used to map the sliding time window of real-time charging data to the standard curve of the corresponding charging stage, obtain the real-time charging parameter vector and the predicted mean vector, and form a standardized residual vector according to the output scale parameter. The joint bias calculation module is used to construct a regularized covariance matrix based on the predicted covariance matrix and the normal residual covariance matrix of the standardized model, calculate the squared Mahalanobis distance, and perform time-series aggregation of the squared Mahalanobis distance within the sliding time window to obtain the joint bias score. The adaptive judgment module is used to determine the adaptive grading threshold based on the joint deviation score distribution of charging stage, model prediction uncertainty, ambient temperature, battery health status and normal verification data. It determines the deviation level by comparing the persistence of the joint deviation score with the adaptive grading threshold, and determines the deviation type based on the residual direction, slope change and fluctuation change of each charging parameter. The response and model update module is used to perform data recording, alarm or charging protection according to the deviation level, associate and save the detection results with the corresponding real-time charging data, and generate and verify candidate models and candidate thresholds using the charging data that has been confirmed to be in a normal state. The modules are connected according to the processing order of the charging data, and the response and model update module forms an update connection with the standard curve modeling module and the adaptive judgment module under the constraint of verification conditions.

10. The charging curve deviation detection system based on Gaussian process regression according to claim 9, characterized in that, The system includes a vehicle-side data interface connected to the battery management system, a charging pile-side data interface connected to the charging equipment, a processor, a memory, an alarm output interface, and a charging control interface. The memory is configured with an original charging data area, an aligned data area, a standard curve model area, a threshold parameter area, a detection result area, and a model version area. The processor calls each functional module according to the sampling period to complete real-time deviation detection, and sends alarm information through the alarm output interface or sends charging power limit requests and charging stop control commands through the charging control interface. When the real-time data quality, curve mapping results, or model integrity do not meet the detection conditions, the system pauses the process of increasing the deviation level based on the current data and updating the model and threshold, retains the original voltage, current, temperature, and communication fault safety protections of the charging equipment, and continues detection using the most recently verified backup model or outputs a model unavailable status.

Images