AI-based charging process safety monitoring method and system

Abstract

The application provides an AI-based charging process safety monitoring method and system, relating to the technical field of charging safety monitoring, comprising obtaining charging state data by collecting current, voltage and temperature data of a charging device, extracting an implicit feature vector using a neural network encoder and projecting it to a feature space to obtain a state representation point, forming a state evolution trajectory from the representation points at consecutive time points and performing shape matching with a reference trajectory to identify an abnormal deviation section, extracting current and temperature data sequences from the charging state data according to the abnormal deviation identifier, performing sliding window sampling to calculate a correlation peak value to determine a conduction time delay amount and a conduction coefficient, and then determining a prediction time window according to the conduction time delay amount and performing phase space reconstruction to calculate a compensation direction vector and a current compensation amount, and generating a control instruction and sending it to the charging device. The application realizes high-precision identification and active compensation control of charging process abnormalities, significantly improving charging safety and equipment operation stability.

Description

Technical Field

[0001] This invention relates to the field of charging safety monitoring technology, and in particular to an AI-based method and system for monitoring the safety of the charging process. Background Technology

[0002] With the rapid popularization of electric vehicles and energy storage devices, safety monitoring technology during the charging process has received widespread attention. Current conventional practices mainly rely on single-parameter threshold alarm mechanisms. This involves real-time monitoring of physical quantities such as current, voltage, or temperature, triggering protective actions when a parameter exceeds a preset safety threshold, such as cutting off the charging circuit or reducing charging power. In addition, some systems employ multi-parameter comprehensive judgment methods, such as jointly analyzing current and temperature data and assessing risk levels through weighted scoring or logical threshold combinations. These methods are typically based on fixed empirical rules, relying on engineers' summarization of historical fault data and the setting of parameter boundaries.

[0003] Existing technologies typically suffer from two significant drawbacks. First, due to time delays and coupling effects in parameter changes during charging—for example, temperature rise often lags by several seconds after a sudden current change—simply relying on instantaneous thresholds for a single parameter cannot capture this dynamic correlation, easily leading to missed or false alarms. Second, localized overheating may develop slowly even when the current is normal, approaching a dangerous state by the time the temperature threshold is triggered. Third, fixed thresholds and weighted scoring rules lack adaptability, making it difficult to handle parameter drift under different charging devices, ambient temperatures, or battery aging conditions. The same threshold may be too conservative at low temperatures but ineffective at high temperatures, resulting in insufficient reliability and robustness of safety monitoring. These shortcomings make it difficult for existing technologies to accurately predict the development trend of potential faults and to provide effective parameter compensation control in the early stages of a fault. Summary of the Invention

[0004] This invention provides an AI-based method and system for monitoring the safety of the charging process, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an AI-based method for safety monitoring of the charging process, comprising:

[0006] Collect current, voltage, and temperature data from the charging device to obtain charging status data;

[0007] The charging status data is input into the neural network encoder, and the hidden feature vector is extracted through multi-layer nonlinear transformation. The hidden feature vector is then projected onto the feature space of a preset dimension to obtain the state representation points in the feature space.

[0008] The state evolution trajectory is formed by acquiring state representation points at consecutive time points. The state evolution trajectory is then matched with the pre-stored reference trajectory in terms of shape. The local curvature deviation and spatial distance deviation of the state evolution trajectory are calculated. Abnormal deviation segments are identified and abnormal deviation indicators and deviation intensity values ​​are obtained.

[0009] When an abnormal deviation flag is received, the current data sequence and temperature data sequence are extracted from the charging status data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined according to the time offset corresponding to the correlation peak, and the conduction coefficient is determined according to the value of the correlation peak to obtain the cross-parameter conduction relationship.

[0010] The prediction time window is determined based on the conduction delay in the cross-parameter conduction relationship. Within the prediction time window, phase space reconstruction is performed, the compensation direction vector and current compensation are calculated, and the corresponding control commands are generated and sent to the charging device.

[0011] In one optional embodiment, charging state data is input into a neural network encoder, and latent feature vectors are extracted through multi-layer nonlinear transformations. These latent feature vectors are then projected onto a feature space of a preset dimension to obtain state representation points in the feature space, including:

[0012] The time delay cross-correlation coefficient is calculated for the current data and temperature data in the charging status data. The propagation delay is determined based on the time delay cross-correlation coefficient. The temperature data is time-aligned according to the propagation delay to obtain aligned temperature data. The covariance matrix is ​​calculated for the current data, voltage data and aligned temperature data. The covariance matrix is ​​eigenvalued to obtain the covariance principal axis direction. A rotation transformation matrix is ​​constructed based on the covariance principal axis direction. The coordinate rotation of the charging status data is performed using the rotation transformation matrix to obtain rotated data.

[0013] The rotational data is input into a neural network encoder, and the hidden feature vector is extracted through multiple nonlinear transformations.

[0014] The covariance matrix of the latent eigenvectors is calculated and eigenvalue decomposition is performed to obtain the eigenvalue sequence and eigenvector sequence. From the eigenvalue sequence, eigenvalues ​​with the same number of dimensions as the preset dimensions are selected according to their numerical values. The eigenvectors corresponding to the selected eigenvalues ​​are extracted to form the projection matrix. The latent eigenvectors and the projection matrix are multiplied to obtain the dimension-reduced vector. The dimensional component values ​​of each dimension of the dimension-reduced vector determine the coordinate position of the state representation point in the feature space, thus obtaining the state representation point in the feature space.

[0015] In one optional embodiment, obtaining state representation points at consecutive time points to form a state evolution trajectory, performing morphological matching between the state evolution trajectory and a pre-stored reference trajectory, calculating the local curvature deviation and spatial distance deviation of the state evolution trajectory, identifying abnormal deviation segments and obtaining abnormal deviation identifiers and deviation intensity values ​​includes:

[0016] The state representation points are obtained in a continuous time sequence. The tangent direction between adjacent state representation points is extracted to construct a tangent vector field. The curvature tensor at each state representation point is obtained by performing a differential operation on the tangent vector field. The principal curvature component is extracted as the trajectory curvature based on the curvature tensor. The state representation points and their corresponding trajectory curvatures are organized in a time sequence to form a state evolution trajectory.

[0017] Extract reference trajectories that match the current charging stage from the pre-stored reference trajectory library, perform dynamic time warping on the state evolution trajectory and the reference trajectory, establish a nonlinear time alignment mapping, match the corresponding reference point on the reference trajectory for each state representation point on the state evolution trajectory, construct the corresponding curvature deviation field, accumulate the curvature deviation field along the time dimension to determine the local curvature deviation, construct the spatial displacement field between the state representation point and the corresponding reference point, and accumulate the spatial displacement field along the time dimension to determine the spatial distance deviation.

[0018] A dual-dimensional deviation judgment criterion is constructed based on local curvature deviation and spatial distance deviation. The time continuous interval that simultaneously meets the preset curvature dimension deviation condition and the preset spatial dimension deviation condition is identified to obtain the abnormal deviation segment. The corresponding time boundary information is extracted to determine the abnormal deviation identifier. The local curvature deviation and spatial distance deviation within the abnormal deviation segment are fused to obtain the deviation intensity value.

[0019] In one optional embodiment, performing dynamic time warping on the state evolution trajectory and the reference trajectory to establish a nonlinear time alignment mapping includes:

[0020] Extract the current data sequence and temperature data sequence of the current charging stage, perform time-domain differentiation to obtain the current change rate sequence and temperature change rate sequence, and extract the corresponding statistical features to construct the stage feature vector;

[0021] The stage feature vector is input into the pre-trained cost weight calculation module to obtain curvature cost weight and distance cost weight. Curvature difference metric and spatial distance metric are calculated for the state representation point and the reference point on the reference trajectory, respectively. The curvature difference metric is weighted by the curvature cost weight to obtain the curvature matching cost. The spatial distance metric is weighted by the distance cost weight to obtain the distance matching cost. The curvature matching cost and the distance matching cost are added to obtain the comprehensive matching cost.

[0022] Construct a cumulative cost matrix based on the comprehensive matching cost, find the minimum cumulative cost value of the predecessor position for each position in the cumulative cost matrix, add it to the current comprehensive matching cost and fill it into the current position;

[0023] Starting from the termination position of the cumulative cost matrix, the backtracking operation is performed. The position with the minimum cumulative cost value among the predecessor positions of the current position is selected as the backtracking path node. The backtracking operation is performed iteratively until the starting position is reached to form an alignment path. Based on the alignment path, the mapping relationship between each state representation point on the state evolution trajectory and the corresponding reference point on the reference trajectory is established to obtain the nonlinear time alignment mapping relationship.

[0024] In one optional embodiment, when an abnormal deviation flag is received, current data sequences and temperature data sequences are extracted from the charging state data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined based on the time offset corresponding to the correlation peak, and the conduction coefficient is determined based on the value of the correlation peak. The cross-parameter conduction relationship is obtained as follows:

[0025] Based on the time boundary information in the abnormal deviation identifier, the current data sequence and temperature data sequence within the corresponding time range are extracted from the charging state data. The current gradient sequence and temperature gradient sequence are constructed by performing time domain differentiation respectively. The gradient abrupt change positions are identified to obtain the current abrupt change point set and the temperature abrupt change point set respectively. The corresponding time interval distribution is calculated, and the sliding window length is determined based on the statistical characteristics of the time interval distribution.

[0026] A fixed reference window is set for the current data sequence to extract the current reference window data. A sliding window is set for the temperature data sequence and time offset sampling is performed according to a preset time step. Temperature window data within the sliding window is extracted at each time offset position. The first information entropy value of the temperature window data and the second information entropy value of the current reference window data are calculated. The conditional mutual information between the first information entropy value and the second information entropy value is constructed as the cross-correlation metric value at that time offset position.

[0027] The cross-correlation metric sequence is obtained by traversing all time offset positions. A cumulative increment sequence is constructed from the cross-correlation metric sequence. The position of maximum curvature is identified in the cumulative increment sequence to determine the correlation peak position. The time offset corresponding to the correlation peak position is extracted as the propagation delay. The corresponding cross-correlation metric is extracted as the propagation coefficient. Cross-parameter propagation relationship is constructed.

[0028] In one optional embodiment, calculating the corresponding time interval distribution and determining the sliding window length based on the statistical characteristics of the time interval distribution includes:

[0029] Extract the time difference between each current mutation point in the current mutation point set and the temperature mutation point in the temperature mutation point set that is closest in time. Collect all time differences to form the original time interval sample set. Calculate the jump amplitude between each time interval sample in the original time interval sample set and its adjacent time interval samples. Remove abnormal time interval samples according to the jump amplitude to obtain the filtered time interval sample set.

[0030] The rate of change of values ​​between adjacent samples is calculated in chronological order for the filtered time interval sample set, and the rate of change of values ​​is used as the gradient value of the time interval gradient field.

[0031] In the time interval gradient field, identify the time region where the gradient value decreases continuously and the rate of decrease is the largest. Determine this time region as the time interval convergence region. Extract the median value of all samples in the time interval convergence region and determine this median value as the reference time interval.

[0032] A nonlinear adjustment coefficient is constructed based on the deviation intensity value in the abnormal deviation identifier. The adaptive time interval is obtained by multiplying the baseline time interval by the nonlinear adjustment coefficient, and this time interval is determined as the sliding window length.

[0033] In one optional embodiment, determining a prediction time window based on the conduction delay in the cross-parameter conduction relationship, performing phase space reconstruction within the prediction time window, calculating the compensation direction vector and current compensation amount, generating corresponding control commands, and sending them to the charging device includes:

[0034] The conduction delay and conduction coefficient are extracted from the cross-parameter conduction relationship. The prediction time window is determined based on the conduction delay. The historical current evolution trajectory in the charging state data is extracted within the prediction time window. The phase space of the historical current evolution trajectory is reconstructed to obtain the current state attractor. The topological deviation of the current state attractor is calculated. The topological deviation and the deviation intensity are weighted and fused to obtain the comprehensive instability metric.

[0035] The initial compensation intensity is obtained by nonlinearly mapping the comprehensive instability metric value and the conduction coefficient. The compensation direction vector is determined according to the topological characteristics of the current state attractor. The initial compensation intensity is decomposed along the compensation direction vector to obtain the vectorized current compensation amount. The amplitude of the vectorized current compensation amount is extracted to determine the current compensation amount.

[0036] A control command data structure containing current compensation amount and compensation direction vector is constructed. The conduction delay amount is embedded in the control command data structure as a delay execution identifier, and it is encapsulated into a control command and sent to the charging device.

[0037] A second aspect of this invention provides an AI-based charging process safety monitoring system, comprising:

[0038] The data acquisition unit is used to collect current data, voltage data, and temperature data of the charging device to obtain charging status data.

[0039] The feature extraction unit is used to input charging state data into the neural network encoder, extract hidden feature vectors through multi-layer nonlinear transformation, and project the hidden feature vectors onto a feature space of a preset dimension to obtain state representation points in the feature space.

[0040] The trajectory matching unit is used to acquire the state characterization points at continuous time points to form the state evolution trajectory, perform morphological matching between the state evolution trajectory and the pre-stored reference trajectory, calculate the local curvature deviation and spatial distance deviation of the state evolution trajectory, identify abnormal deviation segments and obtain abnormal deviation identifiers and deviation intensity values.

[0041] The conduction analysis unit is used to extract current data sequence and temperature data sequence from charging state data when an abnormal deviation flag is received, and to perform sliding window sampling with different time offsets, calculate correlation peak value, determine conduction delay based on the time offset corresponding to the correlation peak value, determine conduction coefficient based on the value of the correlation peak value, and obtain cross-parameter conduction relationship.

[0042] The compensation control unit is used to determine the prediction time window based on the conduction delay in the cross-parameter conduction relationship, perform phase space reconstruction within the prediction time window, calculate the compensation direction vector and current compensation amount, generate corresponding control commands and send them to the charging device.

[0043] A third aspect of the present invention provides an electronic device, comprising:

[0044] processor;

[0045] Memory used to store processor-executable instructions;

[0046] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0048] In this embodiment of the invention, by inputting charging state data into a neural network encoder to extract hidden feature vectors and projecting them into a feature space, the deep correlation features of current, voltage, and temperature can be accurately captured, avoiding the false negative problem of traditional threshold judgment. The morphological matching of the state evolution trajectory with the reference trajectory, combined with local curvature deviation and spatial distance deviation, can accurately identify abnormal deviation segments and quantify the deviation intensity, achieving sensitive detection of early, minor anomalies. The establishment of cross-parameter transmission relationships effectively reveals the propagation law of anomalies between current and temperature. By calculating the correlation peak value through sliding window sampling, the transmission delay and transmission coefficient can be accurately determined, providing a reliable basis for predictive control. Based on the prediction time window determined by the conduction delay, combined with the compensation direction vector and current compensation amount calculated by phase space reconstruction, precise control commands can be generated to proactively intervene in the charging equipment before the anomaly spreads, preventing the accident from worsening. The adaptive adjustment capability of the charging process is significantly enhanced, dynamically adapting to different charging scenarios and load conditions. The neural network encoder and trajectory matching mechanism work together, enabling the monitoring system to have high generalization recognition capability for abnormal patterns under complex operating conditions, while reducing the false alarm rate. The introduction of cross-parameter conduction relationships also optimizes the response speed of the control strategy, ensuring that compensation measures are executed at the optimal time, thereby improving the service life of the charging equipment and the stability of the system. Attached Figure Description

[0049] Figure 1 A flowchart illustrating an AI-based method for monitoring the safety of the charging process;

[0050] Figure 2 This is a flowchart for analyzing cross-parameter transmission relationships. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0052] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0053] Figure 1 This is a flowchart illustrating the AI-based charging process safety monitoring method according to an embodiment of the present invention. Figure 1 As shown, the AI-based charging process safety monitoring method includes:

[0054] Collect current, voltage, and temperature data from the charging device to obtain charging status data;

[0055] The charging status data is input into the neural network encoder, and the hidden feature vector is extracted through multi-layer nonlinear transformation. The hidden feature vector is then projected onto the feature space of a preset dimension to obtain the state representation points in the feature space.

[0056] The state evolution trajectory is formed by acquiring state representation points at consecutive time points. The state evolution trajectory is then matched with the pre-stored reference trajectory in terms of shape. The local curvature deviation and spatial distance deviation of the state evolution trajectory are calculated. Abnormal deviation segments are identified and abnormal deviation indicators and deviation intensity values ​​are obtained.

[0057] When an abnormal deviation flag is received, the current data sequence and temperature data sequence are extracted from the charging status data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined according to the time offset corresponding to the correlation peak, and the conduction coefficient is determined according to the value of the correlation peak to obtain the cross-parameter conduction relationship.

[0058] The prediction time window is determined based on the conduction delay in the cross-parameter conduction relationship. Within the prediction time window, phase space reconstruction is performed, the compensation direction vector and current compensation are calculated, and the corresponding control commands are generated and sent to the charging device.

[0059] In one optional embodiment, charging state data is input into a neural network encoder, and latent feature vectors are extracted through multi-layer nonlinear transformations. These latent feature vectors are then projected onto a feature space of a preset dimension to obtain state representation points in the feature space, including:

[0060] The time delay cross-correlation coefficient is calculated for the current data and temperature data in the charging status data. The propagation delay is determined based on the time delay cross-correlation coefficient. The temperature data is time-aligned according to the propagation delay to obtain aligned temperature data. The covariance matrix is ​​calculated for the current data, voltage data and aligned temperature data. The covariance matrix is ​​eigenvalued to obtain the covariance principal axis direction. A rotation transformation matrix is ​​constructed based on the covariance principal axis direction. The coordinate rotation of the charging status data is performed using the rotation transformation matrix to obtain rotated data.

[0061] The rotational data is input into a neural network encoder, and the hidden feature vector is extracted through multiple nonlinear transformations.

[0062] The covariance matrix of the latent eigenvectors is calculated and eigenvalue decomposition is performed to obtain the eigenvalue sequence and eigenvector sequence. From the eigenvalue sequence, eigenvalues ​​with the same number of dimensions as the preset dimensions are selected according to their numerical values. The eigenvectors corresponding to the selected eigenvalues ​​are extracted to form the projection matrix. The latent eigenvectors and the projection matrix are multiplied to obtain the dimension-reduced vector. The dimensional component values ​​of each dimension of the dimension-reduced vector determine the coordinate position of the state representation point in the feature space, thus obtaining the state representation point in the feature space.

[0063] In one specific implementation, in practical applications of charging monitoring, there is a physical conduction delay between current data and temperature data. When a change in current causes a change in internal resistance power consumption, the temperature response does not occur instantaneously, but rather gradually manifests through a heat conduction process. To accurately capture this time dependency, a time-delay cross-correlation analysis is performed on the current and temperature data in the charging state data. Specifically, the current data sequence is denoted as... The temperature data sequence is denoted as In time offset Calculate the cross-correlation coefficient within the range ,in and These are the means of the corresponding sequences. Let be the length of the data sequence. By iterating through different time offsets, we find the value that maximizes the cross-correlation coefficient. The value represents the propagation delay. In practice, the search range for the time offset is typically set to 0 to 30 seconds, with a step size that is an integer multiple of the sampling period, to ensure coverage of typical timescales for heat conduction.

[0064] After determining the transmission delay, time alignment is performed on the temperature data. The original temperature sequence... According to transmission delay Perform a time shift to obtain aligned temperature data. This alignment operation ensures that the time base of the temperature data is synchronized with the current data in terms of physical causality, eliminating phase deviation caused by conduction delay. The aligned data more accurately reflects the intrinsic relationship between current changes and temperature response. Subsequently, the current data... Voltage data and align temperature data Constructing a three-dimensional data vector Calculate the covariance matrix of the vector sequence. ,in The number of samples within the time window. This is the mean of the data vector.

[0065] The calculated covariance matrix is ​​subjected to eigenvalue decomposition to obtain three eigenvalues. and its corresponding eigenvectors Eigenvectors indicate the main directions of data variation in three-dimensional space, i.e., the directions of the principal axes of covariance. Eigenvectors are sorted in descending order of their eigenvalues, with the eigenvector corresponding to the largest eigenvalue representing the direction of maximum data variance. A rotation transformation matrix is ​​constructed using these three orthogonal eigenvectors. The matrix is ​​an orthogonal matrix that satisfies The charging state data is rotated using a rotation transformation matrix. This process transforms the data from the original coordinate system to the principal component coordinate system. The rotated data is independent along each coordinate axis, eliminating linear correlation between variables and enabling the subsequent neural network to extract independent feature components more efficiently.

[0066] Rotate data The input is a neural network encoder. The encoder employs a multi-layer fully connected neural network structure. The first layer contains 64 neurons, using a modified linear unit activation function to process the input 3D rotation data. The second layer reduces the number of neurons to 32 and continues with non-linear transformations. The third layer further compresses the data to 16 neurons, extracting abstract features layer by layer. Batch normalization is introduced between each layer to standardize the neuron outputs, accelerating training convergence and improving model stability. The fourth layer's output dimension is set to 8 dimensions to obtain the latent feature vector. This layer-by-layer dimensionality reduction network structure forces the model to learn a compressed representation of the data, where each component of the latent feature vector contains information from multiple original parameters after nonlinear combination. During network training, an autoencoder framework is used, with a symmetric decoder network connected after the encoder. The encoder parameters are optimized by minimizing the reconstruction error, ensuring that the latent feature vector retains the key information of the original data.

[0067] After obtaining the latent feature vectors, the optimal projection direction needs to be determined to achieve dimensionality reduction projection. This involves collecting a sequence of latent feature vectors over a time window. Calculate the covariance matrix ,in The number of vectors within the window. The mean of the latent eigenvectors. For the covariance matrix... Eigenvalue decomposition yields 8 eigenvalues. and the corresponding feature vector The magnitude of the eigenvalues ​​reflects the variance contribution of the latent features in various directions; the directions corresponding to larger eigenvalues ​​contain the main information about the changes in the data.

[0068] Based on the preset dimension setting, assuming the target dimension of the feature space is 3-dimensional, the top 3 largest eigenvalues ​​are selected from the eigenvalue sequence according to their numerical values. Extract the eigenvectors corresponding to these three eigenvalues. Arrange them in columns to form a projection matrix. Each column of the projection matrix represents a projection direction. The three directions are orthogonal to each other and correspond to the three principal components with the largest variances in the latent feature space. The latent feature vectors... Matrix multiplication with the projection matrix , to obtain the dimension-reduced vector The three components of the dimension-reduced vector These correspond to the positions of the state representation points on the three coordinate axes 0 in the feature space.

[0069] The component values ​​of each dimension of the dimensionality-reduced vector directly determine the coordinate position of the state representation point in the feature space. The x-axis coordinate is... The vertical axis is the coordinate. Using the vertical axis as the coordinate, a unique point is marked in the three-dimensional feature space. This point represents the charging state in the feature space. In this way, the original high-dimensional complex charging data is compressed into a low-dimensional feature space, preserving the main variation characteristics of the data while enabling visualization and comparative analysis of the states. The distribution of state representation points in the feature space differs for different charging states. Representation points for normal charging states cluster in specific regions, forming a stable distribution pattern, while representation points for abnormal states deviate from the normal region. The choice of feature space dimension balances information preservation and computational complexity; the 3D space can capture the main dynamic characteristics of the charging process and facilitates subsequent trajectory analysis and anomaly detection.

[0070] In practical applications, charging equipment collects current, voltage, and temperature data every second, forming a continuous charging status data stream. For fast charging scenarios for electric vehicles, the charging power can reach over 100 kW, the current fluctuates between 0 and 250 amps, the voltage remains around 400 volts, and the temperature gradually rises from ambient temperature to 60 degrees Celsius. Time-delay cross-correlation analysis typically reveals a transmission delay of 5 to 15 seconds between current and temperature, which is closely related to battery thermal capacity and heat dissipation conditions. After time alignment and coordinate rotation, redundant correlations in the data are eliminated, significantly improving the training efficiency of the neural network encoder. Encoder training uses historical charging data, covering the charging process under different ambient temperatures and different initial battery charge states, ensuring the model has good generalization ability. During dimensionality reduction projection, the first three principal components can typically explain more than 85% of the total variance of the latent features, with information loss controlled within an acceptable range. The obtained state representation points exhibit a smooth evolution trajectory in the feature space, providing a reliable data foundation for subsequent trajectory matching and anomaly detection.

[0071] In one optional embodiment, obtaining state representation points at consecutive time points to form a state evolution trajectory, performing morphological matching between the state evolution trajectory and a pre-stored reference trajectory, calculating the local curvature deviation and spatial distance deviation of the state evolution trajectory, identifying abnormal deviation segments and obtaining abnormal deviation identifiers and deviation intensity values ​​includes:

[0072] The state representation points are obtained in a continuous time sequence. The tangent direction between adjacent state representation points is extracted to construct a tangent vector field. The curvature tensor at each state representation point is obtained by performing a differential operation on the tangent vector field. The principal curvature component is extracted as the trajectory curvature based on the curvature tensor. The state representation points and their corresponding trajectory curvatures are organized in a time sequence to form a state evolution trajectory.

[0073] Extract reference trajectories that match the current charging stage from the pre-stored reference trajectory library, perform dynamic time warping on the state evolution trajectory and the reference trajectory, establish a nonlinear time alignment mapping, match the corresponding reference point on the reference trajectory for each state representation point on the state evolution trajectory, construct the corresponding curvature deviation field, accumulate the curvature deviation field along the time dimension to determine the local curvature deviation, construct the spatial displacement field between the state representation point and the corresponding reference point, and accumulate the spatial displacement field along the time dimension to determine the spatial distance deviation.

[0074] A dual-dimensional deviation judgment criterion is constructed based on local curvature deviation and spatial distance deviation. The time continuous interval that simultaneously meets the preset curvature dimension deviation condition and the preset spatial dimension deviation condition is identified to obtain the abnormal deviation segment. The corresponding time boundary information is extracted to determine the abnormal deviation identifier. The local curvature deviation and spatial distance deviation within the abnormal deviation segment are fused to obtain the deviation intensity value.

[0075] In one specific implementation, during the charging process of an electric vehicle, the charging state is not static but a dynamic process that evolves continuously over time. To accurately capture this dynamic characteristic, state representation points are obtained from the output of the neural network encoder at fixed sampling intervals (e.g., every 0.5 seconds). Each state representation point corresponds to a specific coordinate position in the feature space; for example, in a 128-dimensional feature space, at a certain moment... The state representation point can be represented as a vector. By continuously collecting data for 60 seconds, a sequence containing 120 state representation points can be obtained. .

[0076] For two adjacent state representation points and Calculate the difference vector between them. The difference vector, after normalization, yields the unit tangent vector. These unit tangent vectors form a continuously distributed vector field in space, called the tangent vector field. The tangent vector field describes the local directional characteristics of state evolution. To quantify the curvature of the trajectory, a second-order differential operation is performed on the tangent vector field. Specifically, the rate of change between adjacent tangent vectors is calculated, and a curvature tensor is constructed. In the discrete case, it can be calculated. The curvature tensor is approximated by dividing by the corresponding arc length increment. Since the feature space is multidimensional, the curvature tensor contains multiple components. The largest eigenvalue of the curvature tensor is extracted as the principal curvature component, denoted as . This value reflects the degree to which the state evolution trajectory bends at that point. Each state characterization point... Its corresponding trajectory curvature The pairs are organized chronologically into a composite data structure containing both location and morphological information. This constitutes a complete trajectory of state evolution.

[0077] During the system initialization phase, a large amount of state data from the normal charging process has been collected. Following the same feature extraction and trajectory construction process, a reference trajectory library is formed. The reference trajectory library is categorized and stored according to the charging stage; for example, constant current charging, constant voltage charging, and trickle charging each correspond to different reference trajectory clusters. The current charging stage is determined based on the ratio of the charging voltage to the battery's rated voltage: when the voltage is below 85% of the rated voltage, it is considered a constant current stage; when the voltage reaches 85% to 100% of the rated voltage, it is considered a constant voltage stage; and when the voltage stabilizes at the rated voltage and the current drops below 5% of the initial current, it is considered a trickle charging stage. Based on the determination of the current charging stage, the corresponding reference trajectory is extracted from the reference trajectory library. ,in Indicates the position of the reference point. This represents the curvature of the reference trajectory at that point.

[0078] Because the actual charging process is affected by factors such as ambient temperature, battery aging, and charging power settings, the state evolution trajectory and the reference trajectory may have a nonlinear scaling relationship in the time dimension. A dynamic time warping algorithm is used to establish the optimal time alignment mapping between the two trajectories. A cost matrix is ​​then constructed. , of which elements Represents the state evolution trajectory of the first Points On the reference trajectory, the first Points Euclidean distance between The dynamic programming algorithm is used to find the path with the minimum cumulative cost from the bottom left corner to the top right corner of the matrix. This path determines a one-to-one mapping relationship. For example, points on the state evolution trajectory... Points that may correspond to the reference trajectory ,in It is a mapping function. This mapping allows for one-to-many or many-to-one correspondences and can handle local scaling of time scales.

[0079] Based on the established time-aligned mapping, each state representation point on the state evolution trajectory is represented... Matching corresponding reference points on the reference trajectory Calculate the difference in curvature between the two. This difference reflects the deviation of the actual trajectory from the normal trajectory in terms of curvature characteristics. Arranging the curvature differences at all moments in chronological order constitutes a curvature deviation field. To quantify the overall degree of deviation, the curvature deviation field is accumulated. A sliding window accumulation method is used, selecting a window length of 20 sampling points (corresponding to a 10-second time span), and the sum of the absolute values ​​of curvature differences within the window is calculated as the local curvature deviation. This cumulative operation smooths out instantaneous fluctuations while amplifying persistent deviations, facilitating subsequent anomaly detection.

[0080] Simultaneously, a spatial displacement field is constructed to characterize the trajectory's deviation in the position dimension. The state representation point is calculated. Its corresponding reference point Euclidean distance between Forming a spatial displacement field Similarly, using a sliding window accumulation strategy, the sum of distances is calculated within the same 20-point window to obtain the spatial distance deviation. Spatial distance deviation reflects the degree to which the actual charging state deviates from the normal state cluster in the feature space; the larger the value, the more abnormal the state.

[0081] A two-dimensional deviation judgment criterion is established, simultaneously examining local curvature deviation and spatial distance deviation. Based on statistical analysis of historical data, a deviation threshold for the curvature dimension is determined. Deviation threshold of spatial dimension For example, by analyzing trajectory data from 1000 normal charging cycles, the 95th percentile of the local curvature deviation is calculated as... Calculate the 95th percentile of the spatial distance deviation as For time If satisfied and If a condition is met, the time interval is determined to be in an abnormal state. Scan the entire time series to identify time intervals that continuously meet the above conditions. To avoid misjudging short-term fluctuations as abnormalities, a minimum duration threshold is set, for example, requiring an abnormal state to last at least 3 seconds (corresponding to 6 sampling points) before it is considered an abnormal deviation segment.

[0082] Once a continuous time interval that meets the conditions is identified, the start time of that interval is extracted. and end time This constitutes the time boundary information for the abnormal deviation segment. An abnormal deviation identifier is generated, containing information such as a timestamp and segment length, and using a structured data format for easy reception by subsequent processing modules. For example, the identifier can be represented as a triplet. ,in This is an exception flag.

[0083] To quantify the severity of the anomaly, it is necessary to integrate the local curvature deviation and spatial distance deviation within the anomaly deviation segment. In the segment... Extract all time points. and Numerical values. The curvature deviation is normalized, and its excess factor relative to a threshold is calculated. Similar processing is applied to spatial distance deviation, and the calculation is performed. The deviation intensity value was calculated using a weighted fusion method. The weighting coefficient and The settings are configured according to the actual application scenario. For example, in the case of safety risks caused by abnormal temperatures, spatial distance deviation may be more critical and can be configured accordingly. , The larger the deviation intensity value, the more serious the deviation from the normal charging state, requiring more urgent intervention measures. This value, along with the abnormal deviation indicator, is transmitted to the subsequent safety response module to guide the generation of control strategies.

[0084] Through the complete trajectory morphology matching and deviation judgment process described above, accurate monitoring of the dynamic evolution characteristics of the charging process can be achieved. Early warnings can be issued in the early stages of abnormal conditions, providing reliable trigger signals and quantitative basis for subsequent cross-parameter transmission relationship analysis and control command generation.

[0085] In one optional embodiment, performing dynamic time warping on the state evolution trajectory and the reference trajectory to establish a nonlinear time alignment mapping includes:

[0086] Extract the current data sequence and temperature data sequence of the current charging stage, perform time-domain differentiation to obtain the current change rate sequence and temperature change rate sequence, and extract the corresponding statistical features to construct the stage feature vector;

[0087] The stage feature vector is input into the pre-trained cost weight calculation module to obtain curvature cost weight and distance cost weight. Curvature difference metric and spatial distance metric are calculated for the state representation point and the reference point on the reference trajectory, respectively. The curvature difference metric is weighted by the curvature cost weight to obtain the curvature matching cost. The spatial distance metric is weighted by the distance cost weight to obtain the distance matching cost. The curvature matching cost and the distance matching cost are added to obtain the comprehensive matching cost.

[0088] Construct a cumulative cost matrix based on the comprehensive matching cost, find the minimum cumulative cost value of the predecessor position for each position in the cumulative cost matrix, add it to the current comprehensive matching cost and fill it into the current position;

[0089] Starting from the termination position of the cumulative cost matrix, the backtracking operation is performed. The position with the minimum cumulative cost value among the predecessor positions of the current position is selected as the backtracking path node. The backtracking operation is performed iteratively until the starting position is reached to form an alignment path. Based on the alignment path, the mapping relationship between each state representation point on the state evolution trajectory and the corresponding reference point on the reference trajectory is established to obtain the nonlinear time alignment mapping relationship.

[0090] In one specific implementation, during real-time monitoring of the charging process, to accurately identify abnormal deviations in the state evolution trajectory, it is necessary to precisely match the real-time acquired state evolution trajectory with a pre-stored reference trajectory. Because the charging process is affected by various factors such as ambient temperature fluctuations, grid voltage fluctuations, and battery aging, even under normal operating conditions, the state evolution trajectories of different charging cycles will exhibit non-uniform scaling and deformation in the time dimension, leading to numerous misjudgments from simple point-to-point time synchronization comparisons. To address this issue, dynamic time warping technology is employed to establish a non-linear time alignment mapping, enabling trajectories with similar evolution patterns but different time scales to establish accurate correspondences.

[0091] Before performing dynamic time warping, it is necessary to first determine the current charging stage. The charging process of a lithium battery typically involves pre-charging, constant current charging, constant voltage charging, and trickle charging, with significant differences in current and temperature evolution characteristics at each stage. By extracting the current and temperature data sequences for the current charging stage from the charging state data, and recording the data changes within a recent time window (the length of which is determined by the sampling frequency of the charging device, typically set to include 100 to 300 sampling points), the extracted current data sequence is differentiated in the time domain to calculate the difference between current values ​​at adjacent moments, yielding a current rate of change sequence reflecting the rate of current change. Similarly, a time-domain differentiation operation is performed on the temperature data sequence to obtain the temperature rate of change sequence.

[0092] Based on the current rate of change and temperature rate of change sequences, statistical features characterizing the current charging stage are extracted. Specifically, these extracted features include: the mean, standard deviation, skewness, and kurtosis of the current rate of change sequence, used to describe the central tendency, fluctuation, and distribution pattern of current changes; the mean, standard deviation, maximum, and minimum values ​​of the temperature rate of change sequence, reflecting the rate of temperature increase or decrease; the cross-correlation coefficient between the current and temperature data sequences, characterizing the strength of their correlation; and the decay rate of the autocorrelation coefficient of the current rate of change sequence under different time lags, reflecting the persistence of current fluctuations. These statistical features are arranged in a fixed order to construct a multi-dimensional stage feature vector, typically with 8 to 15 dimensions.

[0093] The constructed stage feature vector is input into a pre-trained cost weight calculation module. This module employs a multilayer perceptron structure, comprising an input layer, two hidden layers, and an output layer. The input layer receives the stage feature vector; the first hidden layer contains 32 neurons using the ReLU activation function, and the second hidden layer contains 16 neurons also using the ReLU activation function. The output layer contains two neurons using the Sigmoid activation function, outputting curvature cost weights and distance cost weights, respectively. Since the output range of the Sigmoid function is between 0 and 1, to ensure the rationality of the two weights, the output results are normalized in practical applications, making the sum of the curvature cost weight and the distance cost weight equal to 1. The training process of this cost weight calculation module uses normal charging samples and abnormal charging samples from historical charging data, optimizing network parameters through supervised learning, enabling the module to adaptively adjust the weight allocation of the two costs according to the characteristics of the current charging stage.

[0094] After obtaining the curvature cost weight and distance cost weight, for each state representation point on the state evolution trajectory and each reference point on the reference trajectory, a curvature difference metric and a spatial distance metric are calculated respectively. The curvature difference metric measures the similarity of the trajectories in their local morphology. For the i-th state representation point on the state evolution trajectory, when calculating its local curvature, two adjacent points before and after this point are selected to form a local trajectory segment. The local curvature value is obtained by fitting a cubic spline curve and calculating the second derivative of the curve at that point. Similarly, the local curvature value of the j-th reference point on the reference trajectory is calculated. The curvature difference metric is defined as the absolute value of the difference between the two curvature values. The spatial distance metric directly calculates the Euclidean distance between the i-th state representation point and the j-th reference point in the feature space, that is, the square root of the sum of the squares of the differences in the coordinates of the two points in each dimension.

[0095] The curvature cost is obtained by weighting the curvature difference metric with the obtained curvature cost weights. The spatial distance metric is then weighted with the distance cost weights to obtain the distance matching cost. The curvature matching cost and the distance matching cost are added together to obtain the comprehensive matching cost between state representation point i and reference point j. This comprehensive matching cost considers both the morphological similarity of the trajectories and the spatial proximity; a smaller cost value indicates a higher degree of matching between the two points. The comprehensive matching cost is calculated for each state representation point on the state evolution trajectory and each reference point on the reference trajectory, forming a two-dimensional matching cost matrix.

[0096] A cumulative cost matrix is ​​constructed based on the comprehensive matching cost matrix. Each position in the cumulative cost matrix stores the minimum cumulative cost from the starting position to the current position. During initialization, the starting position of the cumulative cost matrix is ​​set to the corresponding comprehensive matching cost. Other positions in the cumulative cost matrix are filled sequentially from left to right and from top to bottom. When filling position (i, j), the predecessor position is searched, with three candidate positions: (i-1, j), (i, j-1), and (i-1, j-1). The position with the minimum cumulative cost is selected from these three candidate positions, and its cumulative cost is added to the comprehensive matching cost corresponding to the current position (i, j), and then filled into position (i, j) of the cumulative cost matrix. This process employs dynamic programming to ensure that each position stores the cumulative cost of the optimal path from the starting position to that position.

[0097] After constructing the cumulative cost matrix, a backtracking operation is performed starting from the ending position of the matrix to determine the optimal alignment path. The ending position corresponds to the matching of the last state representation point of the state evolution trajectory with the last reference point of the reference trajectory. At the current position (i, j), the cumulative cost values ​​of its three predecessor positions (i-1, j), (i, j-1), and (i-1, j-1) in the cumulative cost matrix are found. The predecessor position with the smallest cumulative cost value is selected as the next node of the backtracking path, and this node is added to the alignment path. The current position is updated with the selected predecessor position, and the backtracking operation continues. The above backtracking process is iterated until the starting position of the cumulative cost matrix is ​​reached. At this point, the obtained node sequence constitutes the complete alignment path.

[0098] The alignment path establishes a mapping relationship between each state representation point on the state evolution trajectory and its corresponding reference point on the reference trajectory. Each node (i, j) in the alignment path represents the match between the i-th state representation point on the state evolution trajectory and the j-th reference point on the reference trajectory. Since dynamic time warping allows for both one-to-many and many-to-one mappings, a state representation point may correspond to multiple consecutive reference points on the reference trajectory, or multiple consecutive state representation points may correspond to the same reference point on the reference trajectory. This nonlinear time alignment mapping relationship can adapt to the non-uniform changes in the time scale during charging, ensuring that a correct correspondence with the reference trajectory can still be found even when the charging speed is faster or slower in certain stages.

[0099] The established nonlinear time alignment mapping provides an accurate benchmark for subsequent anomaly deviation detection. By comparing each state representation point on the state evolution trajectory with its aligned reference point, the calculation of local curvature deviation and spatial distance deviation can eliminate errors caused by time scale differences, focusing only on the true morphological and positional deviations. When the curvature deviation or spatial distance deviation between a state representation point and its aligned reference point exceeds a preset threshold, the point is marked as an anomaly deviation point, thereby identifying the anomaly deviation segment and generating an anomaly deviation identifier and deviation intensity value, providing a reliable basis for charging safety monitoring.

[0100] like Figure 2 As shown, a flowchart for cross-parameter transmission relationship analysis is presented.

[0101] In one optional embodiment, when an abnormal deviation flag is received, current data sequences and temperature data sequences are extracted from the charging state data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined based on the time offset corresponding to the correlation peak, and the conduction coefficient is determined based on the value of the correlation peak. The cross-parameter conduction relationship is obtained as follows:

[0102] Based on the time boundary information in the abnormal deviation identifier, the current data sequence and temperature data sequence within the corresponding time range are extracted from the charging state data. The current gradient sequence and temperature gradient sequence are constructed by performing time domain differentiation respectively. The gradient abrupt change positions are identified to obtain the current abrupt change point set and the temperature abrupt change point set respectively. The corresponding time interval distribution is calculated, and the sliding window length is determined based on the statistical characteristics of the time interval distribution.

[0103] A fixed reference window is set for the current data sequence to extract the current reference window data. A sliding window is set for the temperature data sequence and time offset sampling is performed according to a preset time step. Temperature window data within the sliding window is extracted at each time offset position. The first information entropy value of the temperature window data and the second information entropy value of the current reference window data are calculated. The conditional mutual information between the first information entropy value and the second information entropy value is constructed as the cross-correlation metric value at that time offset position.

[0104] The cross-correlation metric sequence is obtained by traversing all time offset positions. A cumulative increment sequence is constructed from the cross-correlation metric sequence. The position of maximum curvature is identified in the cumulative increment sequence to determine the correlation peak position. The time offset corresponding to the correlation peak position is extracted as the propagation delay. The corresponding cross-correlation metric is extracted as the propagation coefficient. Cross-parameter propagation relationship is constructed.

[0105] In one specific implementation, during safety monitoring of the charging process, upon detecting an abnormal deviation, it is necessary to conduct in-depth analysis of the correlation characteristics between current and temperature parameters. During operation, changes in current in the charging device cause heat generation due to internal resistance, leading to temperature changes; however, this conduction process exhibits a time delay. To accurately describe this cross-parameter conduction relationship, it is necessary to extract current and temperature data sequences from the charging state data for in-depth analysis.

[0106] Time boundary information, including the start and end times of the anomaly, is extracted from the anomaly deviation markers. Based on this time range, current and temperature data sequences for the corresponding time periods are extracted from the continuously collected charging status data. The current data sequence records the continuous change of charging current over time, while the temperature data sequence records the continuous change of temperature in key parts of the device over time. Time-domain differentiation is performed on the extracted current data sequence to calculate the difference between current values ​​at adjacent moments, resulting in a current gradient sequence. This gradient sequence reflects the rate of current change and highlights locations of sharp current fluctuations. Similarly, time-domain differentiation is performed on the temperature data sequence to calculate the difference between temperature values ​​at adjacent moments, resulting in a temperature gradient sequence.

[0107] In the current gradient sequence, abrupt changes in gradient values ​​are identified. Specifically, the mean and standard deviation of the gradient sequence are calculated, and locations where the absolute gradient value exceeds the mean plus twice the standard deviation are marked as abrupt change points. The timestamps of all current gradient abrupt change points are collected to form a current abrupt change point set. The same method is used to identify gradient abrupt change locations in the temperature gradient sequence, resulting in a temperature abrupt change point set. The time intervals between adjacent abrupt change points in the current abrupt change point set constitute the current time interval sequence, and the time intervals between adjacent abrupt change points in the temperature abrupt change point set constitute the temperature time interval sequence. Statistical analysis is performed on these two time interval sequences to calculate the median, interquartile range, and mode of the time intervals. A weighted average is calculated using the medians of the current and temperature time interval sequences, with the weighting coefficients determined by the reciprocal of the variance of each sequence; the smaller the variance, the greater the weight. The weighted average result serves as the initial window length, which is then quantized and adjusted based on the data sampling frequency to ensure the window length corresponds to an integer number of sampling points, ultimately determining the sliding window length.

[0108] After determining the sliding window length, a fixed reference window is set on the current data sequence. The starting position of the fixed reference window is selected at the beginning of the time range indicated by the abnormal deviation indicator, and the window length adopts the previously determined sliding window length. Current reference window data is extracted from the fixed reference window, which contains all current sample values ​​within that window period. A sliding window is set for the temperature data sequence, with the same length as the fixed reference window. A preset time step is set, which is determined based on the data sampling frequency and the desired time delay resolution, typically an integer multiple of the sampling period. Starting from the beginning position of the temperature data sequence, time-off sampling is performed according to the preset time step. At each time offset position, the temperature data sequence segment covered by the sliding window is the temperature window data at that position.

[0109] For each time offset, the first information entropy value of the temperature window data is calculated. First, the temperature window data is discretized, mapping continuous temperature values ​​to a finite number of discrete intervals. The number of intervals is determined based on the number of data points within the window and the data distribution characteristics, typically calculated using the Sturgess formula. The frequency of temperature data points within each discrete interval is counted, and the probability distribution of each interval is calculated. The first information entropy value is obtained through probability distribution calculation, characterizing the uncertainty of the temperature window data. The calculation formula is to sum the products of the probabilities of each interval and their logarithms, and then take the negative value. The same method is used to discretize and calculate the probability distribution of the current reference window data to obtain the second information entropy value, which characterizes the uncertainty of the current reference window data.

[0110] A conditional mutual information quantity is constructed between the first and second information entropy values. This conditional mutual information quantity reflects the additional information provided by the temperature window data entropy, given the known information entropy of the current reference window data. The calculation method involves first jointly discretizing the temperature window data and the current reference window data to construct a two-dimensional joint probability distribution matrix. Each element in the matrix represents the joint probability that both temperature and current fall within their corresponding discrete intervals. The joint information entropy is calculated based on the joint probability distribution. Then, combining the first and second information entropy values, the conditional mutual information quantity is obtained using the mutual information calculation formula in information theory. This conditional mutual information quantity serves as a measure of cross-correlation at the current time offset; a larger value indicates a stronger correlation between temperature and current at that time offset.

[0111] Following a preset time step, the calculation process is repeated for each possible time offset position, yielding a series of cross-correlation metrics. These metrics are then arranged in order of time offset, forming a cross-correlation metric sequence. This sequence reflects the changing strength of the correlation between current and temperature at different time offsets. Cumulative increment analysis is performed on the cross-correlation metric sequence, calculating the difference between cross-correlation metrics at adjacent time offset positions to construct a cumulative increment sequence. Positive values ​​in the cumulative increment sequence indicate an increase in cross-correlation metrics, while negative values ​​indicate a decrease. The position of maximum curvature is located within the cumulative increment sequence. Curvature is calculated by performing a second-order difference on the sequence. The position with the maximum absolute value of the second-order difference is the position of maximum curvature, corresponding to the inflection point where the cross-correlation metric changes from a rapid increase to a slow increase or begins to decrease, marking the peak of the correlation.

[0112] After identifying the location of maximum curvature, this location is determined as the correlation peak location. The time offset corresponding to the correlation peak location is extracted; this time offset reflects the time delay required for current changes to be conducted to temperature changes, and is used as the conduction delay. The cross-correlation metric corresponding to the correlation peak location is extracted; this value reflects the strength of the correlation between current and temperature, with a larger value indicating higher conduction efficiency, and is used as the conduction coefficient. The conduction delay and conduction coefficient are associated and stored to construct a cross-parameter conduction relationship data structure. This data structure includes a conduction delay field and a conduction coefficient field, fully describing the cross-parameter conduction characteristics from current to temperature under the current charging state.

[0113] Establishing the cross-parameter conduction relationship provides a crucial basis for subsequent predictions. The conduction delay can be used to determine the time window in which temperature changes lag behind current changes, allowing for predictive analysis within this window. The conduction coefficient is used to quantify the impact of current changes on temperature changes and serves as an important parameter in calculating compensation amounts. The entire cross-parameter conduction relationship extraction process fully considers the physical and statistical characteristics of the charging process. By calculating information entropy and mutual information, it avoids the limitations of simple linear correlation analysis, capturing nonlinear conduction characteristics and providing an accurate time prediction basis and quantitative control criteria for proactive intervention in charging safety.

[0114] In one optional embodiment, calculating the corresponding time interval distribution and determining the sliding window length based on the statistical characteristics of the time interval distribution includes:

[0115] Extract the time difference between each current mutation point in the current mutation point set and the temperature mutation point in the temperature mutation point set that is closest in time. Collect all time differences to form the original time interval sample set. Calculate the jump amplitude between each time interval sample in the original time interval sample set and its adjacent time interval samples. Remove abnormal time interval samples according to the jump amplitude to obtain the filtered time interval sample set.

[0116] The rate of change of values ​​between adjacent samples is calculated in chronological order for the filtered time interval sample set, and the rate of change of values ​​is used as the gradient value of the time interval gradient field.

[0117] In the time interval gradient field, identify the time region where the gradient value decreases continuously and the rate of decrease is the largest. Determine this time region as the time interval convergence region. Extract the median value of all samples in the time interval convergence region and determine this median value as the reference time interval.

[0118] A nonlinear adjustment coefficient is constructed based on the deviation intensity value in the abnormal deviation identifier. The adaptive time interval is obtained by multiplying the baseline time interval by the nonlinear adjustment coefficient, and this time interval is determined as the sliding window length.

[0119] In one specific implementation, during safety monitoring of the charging process, upon detecting an abnormal deviation, it is necessary to analyze the conduction relationship between current and temperature in depth. Determining the sliding window length is crucial for accurately capturing this conduction characteristic. A sliding window that is too short will fail to fully cover the complete conduction process from current change to temperature response, while a window that is too long will introduce excessive irrelevant noise data, reducing the accuracy of correlation analysis.

[0120] A set of current abrupt change points and a set of temperature abrupt change points are extracted from the current data sequence and the temperature data sequence, respectively. Current abrupt change points are identified using a second-order difference method. The first-order difference is calculated for the current values ​​at consecutive time points. When the absolute value of the first-order difference exceeds a preset current change threshold, that time point is marked as a candidate abrupt change point. Second-order difference calculations are then performed on the candidate abrupt change points. When the sign of the second-order difference is flipped, the point is confirmed as a true current abrupt change point. Temperature abrupt change points are extracted using a similar method, but considering the response characteristics and thermal inertia of the temperature sensor, the temperature change threshold is set between 0.3 and 0.5 times the current change threshold. In practical applications, the current change threshold can be set to 8% to 12% of the normal operating current, and the temperature change threshold can be set to 0.5℃ to 1.2℃.

[0121] After obtaining the sets of current abrupt change points and temperature abrupt change points, for each current abrupt change point in the current abrupt change point set, the closest temperature abrupt change point in the temperature abrupt change point set is found in the temperature abrupt change point set. Specifically, using the timestamp of the current abrupt change point as a reference, the temperature abrupt change point set is searched backwards, and the time difference between each temperature abrupt change point and the current abrupt change point is calculated. The temperature abrupt change point with the smallest absolute value of the time difference and a positive value is selected as the matching point, and the corresponding time difference value is recorded. A positive time difference is required because physically, the temperature response always lags behind the current change and cannot occur earlier. All the matched time difference values ​​are collected to form the original time interval sample set. This sample set reflects the time distribution characteristics required for current changes to be transmitted to temperature changes in different charging stages.

[0122] The original time interval sample set may contain anomalous samples due to sensor noise, signal interference, or misidentification of abrupt changes. To improve the reliability of subsequent analysis, anomalous samples need to be removed. The original time interval sample set is arranged chronologically, and the jump amplitude between each time interval sample and its preceding and following time interval samples is calculated. The jump amplitude is calculated as the absolute value of the difference between the current sample value and the adjacent sample value. For each sample, the jump amplitude with the preceding sample and the jump amplitude with the following sample are calculated, and the larger of the two values ​​is taken as the maximum jump amplitude for that sample. The distribution of all maximum jump amplitudes in the entire sample set is statistically analyzed, and the median and interquartile range are calculated. Samples with a maximum jump amplitude exceeding the median plus three times the interquartile range are marked as anomalous samples and removed. This statistical distribution-based removal method can effectively identify isolated outliers while preserving the true physical change trend. After removing anomalous samples, a filtered time interval sample set is obtained, which retains stable and reliable current-temperature conduction delay information.

[0123] A deep analysis was performed on the filtered time interval sample set to extract its temporal evolution characteristics. The rate of change of numerical values ​​was calculated for adjacent time interval samples according to the sampling time order. Assume the... The value of each time interval sample is , No. The value of each sample is Then the rate of change of the values ​​between the two Calculated as This rate of change reflects the trend of propagation delay between adjacent time points and can be understood as the gradient value at that location in the time interval gradient field. A positive gradient value indicates that the propagation delay is increasing, while a negative gradient value indicates that the propagation delay is decreasing. By traversing the entire filtered time interval sample set, a complete time interval gradient field can be constructed.

[0124] We search for regions with special physical significance within the time interval gradient field. As the charging process gradually enters a stable phase, the conduction relationship between current and temperature tends to stabilize, and the corresponding conduction delay converges to a relatively fixed value. This convergence process is represented in the gradient field as a region where the gradient value is continuously negative and has a large absolute value. We traverse the time interval gradient field to identify time segments where the gradient value continuously decreases. For each identified decreasing segment, we calculate its average decreasing rate, which is the cumulative change in gradient value within that segment divided by the segment's duration. We select the time segment with the largest absolute value of the average decreasing rate and define it as the time interval convergence region. This region corresponds to the stage in the charging process where the current-temperature conduction relationship stabilizes most quickly, and the time interval samples in this stage best represent the true conduction delay under stable conditions.

[0125] All time interval samples within the convergence region were extracted and statistically analyzed. Considering potential minor fluctuations, the median was used instead of the arithmetic mean as the representative value. All sample values ​​were sorted from smallest to largest. When the number of samples was odd, the middle value was taken as the median; when the number of samples was even, the arithmetic mean of the two middle values ​​was taken as the median. This median reflects the typical time delay of current change propagation to temperature change under stable charging conditions, and it was determined as the reference time interval. The reference time interval is measured under normal and stable operating conditions, but when abnormal deviations occur, the dynamic characteristics of the system may change, requiring adaptive adjustment of the window length.

[0126] A nonlinear adjustment coefficient is constructed based on the deviation intensity value contained in the received abnormal deviation identifier. The deviation intensity value is typically a normalized quantity ranging from 0 to 1; a larger value indicates a more severe deviation. Nonlinear adjustment coefficient. The construction of the adjustment mechanism needs to reflect the following physical laws: when the deviation intensity is small, the conduction characteristics do not change significantly, and the adjustment range should be small; when the deviation intensity increases, a larger adjustment range is needed to more accurately capture rapidly changing abnormal processes. An exponential function is used to construct the adjustment coefficient, assuming the deviation intensity value is... ,but ,in The basic adjustment factor ranges from 0.2 to 0.4. This is a nonlinear enhancement coefficient, ranging from 1.5 to 2.5. This exponential form ensures that the adjustment coefficient is close to 1 when the deviation intensity is small, while the adjustment coefficient grows rapidly when the deviation intensity increases, allowing the window length to adapt to rapid changes in the dynamic characteristics of the system.

[0127] Reference time interval With nonlinear adjustment coefficient Multiply to obtain the adaptive time interval. ,Right now This adaptive time interval comprehensively considers steady-state transmission characteristics and the current degree of anomaly, preserving statistical regularities based on historical stable data while dynamically adjusting according to real-time anomaly states. The adaptive time interval... The sliding window length is determined and used for subsequent sliding window sampling operations. In practical applications, to ensure the integrity of sampling and the stability of calculation, the calculated sliding window length is also subject to boundary constraints, ensuring that it is not less than three times the data sampling period and not greater than one-fifth of the total charging process time. Through this adaptive determination method based on the statistical characteristics of time interval distribution, the sliding window length can accurately match the current-to-temperature conduction process, providing a reliable analytical basis for accurately calculating correlation peaks and conduction relationships.

[0128] In one optional embodiment, determining a prediction time window based on the conduction delay in the cross-parameter conduction relationship, performing phase space reconstruction within the prediction time window, calculating the compensation direction vector and current compensation amount, generating corresponding control commands, and sending them to the charging device includes:

[0129] The conduction delay and conduction coefficient are extracted from the cross-parameter conduction relationship. The prediction time window is determined based on the conduction delay. The historical current evolution trajectory in the charging state data is extracted within the prediction time window. The phase space of the historical current evolution trajectory is reconstructed to obtain the current state attractor. The topological deviation of the current state attractor is calculated. The topological deviation and the deviation intensity are weighted and fused to obtain the comprehensive instability metric.

[0130] The initial compensation intensity is obtained by nonlinearly mapping the comprehensive instability metric value and the conduction coefficient. The compensation direction vector is determined according to the topological characteristics of the current state attractor. The initial compensation intensity is decomposed along the compensation direction vector to obtain the vectorized current compensation amount. The amplitude of the vectorized current compensation amount is extracted to determine the current compensation amount.

[0131] A control command data structure containing current compensation amount and compensation direction vector is constructed. The conduction delay amount is embedded in the control command data structure as a delay execution identifier, and it is encapsulated into a control command and sent to the charging device.

[0132] In one specific implementation, the calculated conduction delay and conduction coefficient are extracted from the cross-parameter conduction relationship. The conduction delay reflects the time delay characteristic of current change propagating to temperature change, and the conduction coefficient reflects the magnitude of this conduction effect. The prediction time window is determined based on the conduction delay, specifically by using the conduction delay... Based on this, set the length of the prediction time window. ,in To allow for buffer time, the value is typically between 5 and 15 seconds. The start time of the prediction time window is set to the current time minus [the time factor]. The termination time is set to the current time plus... Such window settings can capture the current evolution characteristics before the conduction effect occurs, and also reserve sufficient response time for subsequent compensation control.

[0133] Within a defined prediction time window, current data sequences for that time period are extracted from the charging state data to form a historical current evolution trajectory. This trajectory contains time-series information on current changes, recording the dynamic evolution of the current during charging. Phase space reconstruction is performed on this historical current evolution trajectory, using a delayed coordinate embedding method to construct a multi-dimensional phase space. The embedding dimensions of the phase space reconstruction are as follows: Determined using the pseudo-nearest neighbor method, with typical values ​​ranging from 3 to 7 dimensions. Delay time. Determined by the first zero of the autocorrelation function or the first minimum point of the mutual information function. A one-dimensional current-time series... Reconstructed 3D phase space vector: original current value ,Delay Current value at time ,Delay Current value at time And so on until the delay The current value at time t, this Each component constitutes a state point in phase space. By performing the same reconstruction operation on all moments within the prediction time window, a series of phase space state points are obtained. The geometric structure formed by these state points in phase space is the current state attractor.

[0134] The current-state attractor reflects the dynamic characteristics of the charging current system; under normal charging conditions, the attractor possesses a stable topology. The topological deviation of the current-state attractor is calculated to quantify the topological difference between the current attractor and the normal attractor. First, the correlation dimension of the current attractor is calculated. This dimension reflects the filling characteristics of the attractor in phase space and is calculated using the correlation integral method. Simultaneously, the correlation dimension of the reference attractor under the corresponding operating condition is extracted from pre-stored normal charging reference data. Topological deviation Calculated as the relative deviation between the current correlation dimension and the reference correlation dimension: In addition to the correlation dimension, the maximum Lyapunov exponent of the attractor is also calculated. This index reflects the chaotic characteristics and predictability of the system. The maximum Lyapunov exponent reference value under normal charging conditions is... A negative value or a small positive value close to zero indicates that the system is in a stable or weakly chaotic state. In abnormal states... A significant increase indicates that the system has entered a state of strong chaos or divergence. Combining the correlation dimension deviation and the change in the Lyapunov exponent yields the complete topological deviation. ,in and These are weighting coefficients, and the sum of the two is 1. Typical values ​​are 0.6 and 0.4, respectively.

[0135] The calculated topology deviation The deviation intensity value obtained from the aforementioned identification By performing weighted fusion, a comprehensive instability metric is obtained. The fusion method is as follows: ,in and To achieve the fusion of weighting coefficients, the sum of the two is 1. These two weighting coefficients are dynamically adjusted based on the anomaly type: when the anomaly deviation is mainly manifested in the deviation of the feature space trajectory, Take the larger value, ranging from 0.6 to 0.8; when the abnormal deviation is mainly reflected in changes in the phase space topology, A larger value is chosen, ranging from 0.6 to 0.8. Through this weighted fusion, the comprehensive instability metric can fully reflect the degree of instability of the charging system across multiple observation dimensions, taking into account both the overall evolutionary deviation in the high-dimensional feature space and the intrinsic structural changes of the electrodynamic system.

[0136] Comprehensive instability measurement value Conduction coefficients extracted from transparameter conduction relationships Perform nonlinear mapping to obtain the initial compensation strength. Nonlinear mappings take the form of piecewise functions: when Less than the set threshold When using linear mapping ;when exist and When between, a quadratic mapping is used. To enhance the compensation response; when Exceed When using a saturation function To avoid overcompensation, among which The maximum compensation strength limit is determined based on the rated parameters of the charging equipment. This piecewise nonlinear mapping strategy can provide appropriate compensation strength under different degrees of instability, ensuring accurate adjustment for small deviations while avoiding control saturation for large deviations.

[0137] The compensation direction vector is determined based on the topological characteristics of the current-state attractor. In phase space, the deviation vector between the current attractor's centroid position and the reference attractor's centroid position is calculated; the direction of this deviation vector indicates the overall drift direction of the attractor. Simultaneously, the standard deviation of the attractor's distribution in each phase space dimension is calculated, and the principal deviation dimension with the largest standard deviation is identified. This dimension corresponds to the most significant dynamic characteristic of deviation during current evolution. The compensation direction vector is obtained by combining the attractor centroid deviation direction and the principal deviation dimension direction through a weighted average. Specifically, let the unit vector of the centroid deviation direction be... The unit vector in the main deviation dimension direction is Then the compensation direction vector ,in and This is the directional weighting coefficient, dynamically adjusted according to the importance of the two directions, with typical values ​​ranging from 0.3 to 0.7. This compensation direction vector points to the direction the attractor needs to revert to, i.e., to the attractor structure in its normal charging state.

[0138] The calculated initial compensation strength Vector along the compensation direction Decompose the data to obtain the vectorized current compensation amount. The decomposition method is as follows: ,in To compensate for the magnitude of the direction vector, normalization is used. The vectorized current compensation quantity has components in all dimensions of the phase space, reflecting the coordinated changes in current adjustment under different time delay states. The amplitude is extracted from the vectorized current compensation quantity, i.e., the calculation is performed. The Euclidean norm is used to obtain the scalar form of the current compensation. This current compensation amount represents the required adjustment range of the charging current, measured in amperes. To ensure the safety of the compensation control, a boundary constraint check is performed on the calculated current compensation amount: if... Exceeding the maximum allowable adjustment range of the charging device Then restrict it to ;like Less than the minimum effective adjustment range If so, it will be determined that no compensation is required. Set to zero.

[0139] Construct a control command data structure that includes the current compensation amount and the compensation direction vector. This data structure is organized in key-value pair format and includes the following fields: the compensation amplitude field stores the current compensation amount. The compensation direction field stores the compensation direction vector. The various dimensional components, the instability metric field stores the comprehensive instability metric value. As supplementary diagnostic information, the timestamp field records the time when the control command was generated. The propagation delay is embedded in this control command data structure. As a delayed execution identifier, the specific implementation is as follows: a delayed execution field is added to the data structure to indicate the propagation delay. The value is written into this field, and the delay type is set to "conduction delay compensation," instructing the charging device not to immediately adjust the current after receiving the control command, but to wait. The operation is executed after a certain period of time. This delayed execution mechanism matches the current-temperature conduction delay characteristics obtained from the previous analysis, ensuring that the current compensation action takes effect at the optimal time and maximizes the compensation effect.

[0140] The constructed control command data structure is encapsulated to form a control command data packet conforming to the charging equipment communication protocol. The encapsulation process includes: adding a command type identifier, marked as "Safety Compensation Control Command"; adding a device address field, specifying the unique identifier of the target charging equipment; adding a check field, using the CRC16 cyclic redundancy check algorithm to calculate the checksum of the data packet for the receiving end to verify data integrity; and adding a priority field, setting the safety compensation command to high priority to ensure priority transmission in the event of communication network congestion. The encapsulated control command is sent to the corresponding charging equipment through the communication interface of the charging monitoring system. After receiving the control command, the charging equipment parses the various fields in the data structure, extracts the current compensation amount, compensation direction vector, and delay execution identifier, and initiates its internal compensation control logic, including delay... After a certain period of time, the charging current is precisely adjusted according to the multi-dimensional coordination method indicated by the compensation direction vector, and the current adjustment range is controlled within... Within the specified range, it enables proactive intervention and safety correction for abnormal deviations during the charging process, allowing the charging system to return to its normal state evolution trajectory and stable dynamic attractor structure.

[0141] A second aspect of this invention provides an AI-based charging process safety monitoring system, comprising:

[0142] The data acquisition unit is used to collect current data, voltage data, and temperature data of the charging device to obtain charging status data.

[0143] The feature extraction unit is used to input charging state data into the neural network encoder, extract hidden feature vectors through multi-layer nonlinear transformation, and project the hidden feature vectors onto a feature space of a preset dimension to obtain state representation points in the feature space.

[0144] The trajectory matching unit is used to acquire the state characterization points at continuous time points to form the state evolution trajectory, perform morphological matching between the state evolution trajectory and the pre-stored reference trajectory, calculate the local curvature deviation and spatial distance deviation of the state evolution trajectory, identify abnormal deviation segments and obtain abnormal deviation identifiers and deviation intensity values.

[0145] The conduction analysis unit is used to extract current data sequence and temperature data sequence from charging state data when an abnormal deviation flag is received, and to perform sliding window sampling with different time offsets, calculate correlation peak value, determine conduction delay based on the time offset corresponding to the correlation peak value, determine conduction coefficient based on the value of the correlation peak value, and obtain cross-parameter conduction relationship.

[0146] The compensation control unit is used to determine the prediction time window based on the conduction delay in the cross-parameter conduction relationship, perform phase space reconstruction within the prediction time window, calculate the compensation direction vector and current compensation amount, generate corresponding control commands and send them to the charging device.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based method for safety monitoring of the charging process, characterized in that, include: Collect current, voltage, and temperature data from the charging device to obtain charging status data; The charging status data is input into the neural network encoder, and the hidden feature vector is extracted through multi-layer nonlinear transformation. The hidden feature vector is then projected onto the feature space of a preset dimension to obtain the state representation points in the feature space. The state evolution trajectory is formed by acquiring state representation points at continuous time points. The state evolution trajectory is then matched with the pre-stored reference trajectory in terms of shape. The local curvature deviation and spatial distance deviation of the state evolution trajectory are calculated. Abnormal deviation segments are identified and abnormal deviation indicators and deviation intensity values ​​are obtained. When an abnormal deviation flag is received, the current data sequence and temperature data sequence are extracted from the charging status data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined according to the time offset corresponding to the correlation peak, and the conduction coefficient is determined according to the value of the correlation peak to obtain the cross-parameter conduction relationship. The prediction time window is determined based on the conduction delay in the cross-parameter conduction relationship. Within the prediction time window, phase space reconstruction is performed, the compensation direction vector and current compensation are calculated, and the corresponding control commands are generated and sent to the charging device.

2. The method according to claim 1, characterized in that, The charging state data is input into a neural network encoder, and latent feature vectors are extracted through multi-layer nonlinear transformations. These latent feature vectors are then projected onto a feature space of a preset dimension to obtain the state representation points in the feature space, including: The time delay cross-correlation coefficient is calculated for the current data and temperature data in the charging status data. The propagation delay is determined based on the time delay cross-correlation coefficient. The temperature data is time-aligned according to the propagation delay to obtain aligned temperature data. The covariance matrix is ​​calculated for the current data, voltage data and aligned temperature data. The covariance matrix is ​​eigenvalued to obtain the covariance principal axis direction. A rotation transformation matrix is ​​constructed based on the covariance principal axis direction. The coordinate rotation of the charging status data is performed using the rotation transformation matrix to obtain rotated data. The rotational data is input into a neural network encoder, and the hidden feature vector is extracted through multiple nonlinear transformations. The covariance matrix of the latent eigenvectors is calculated and eigenvalue decomposition is performed to obtain the eigenvalue sequence and eigenvector sequence. From the eigenvalue sequence, eigenvalues ​​with the same number of dimensions as the preset dimensions are selected according to their numerical values. The eigenvectors corresponding to the selected eigenvalues ​​are extracted to form the projection matrix. The latent eigenvectors and the projection matrix are multiplied to obtain the dimension-reduced vector. The dimensional component values ​​of each dimension of the dimension-reduced vector determine the coordinate position of the state representation point in the feature space, thus obtaining the state representation point in the feature space.

3. The method according to claim 1, characterized in that, The state evolution trajectory is constructed by acquiring state representation points at consecutive time points. This trajectory is then morphologically matched with a pre-stored reference trajectory. Local curvature deviation and spatial distance deviation are calculated. Abnormal deviation segments are identified, and abnormal deviation markers and deviation intensity values ​​are obtained, including: The state representation points are obtained in a continuous time sequence. The tangent direction between adjacent state representation points is extracted to construct a tangent vector field. The curvature tensor at each state representation point is obtained by performing a differential operation on the tangent vector field. The principal curvature component is extracted as the trajectory curvature based on the curvature tensor. The state representation points and their corresponding trajectory curvatures are organized in a time sequence to form a state evolution trajectory. Extract reference trajectories that match the current charging stage from the pre-stored reference trajectory library, perform dynamic time warping on the state evolution trajectory and the reference trajectory, establish a nonlinear time alignment mapping, match the corresponding reference point on the reference trajectory for each state representation point on the state evolution trajectory, construct the corresponding curvature deviation field, accumulate the curvature deviation field along the time dimension to determine the local curvature deviation, construct the spatial displacement field between the state representation point and the corresponding reference point, and accumulate the spatial displacement field along the time dimension to determine the spatial distance deviation. A dual-dimensional deviation judgment criterion is constructed based on local curvature deviation and spatial distance deviation. The time continuous interval that simultaneously meets the preset curvature dimension deviation condition and the preset spatial dimension deviation condition is identified to obtain the abnormal deviation segment. The corresponding time boundary information is extracted to determine the abnormal deviation identifier. The local curvature deviation and spatial distance deviation within the abnormal deviation segment are fused to obtain the deviation intensity value.

4. The method according to claim 3, characterized in that, Performing dynamic time warping on the state evolution trajectory and the reference trajectory to establish a nonlinear time alignment mapping includes: Extract the current data sequence and temperature data sequence of the current charging stage, perform time-domain differentiation to obtain the current change rate sequence and temperature change rate sequence, and extract the corresponding statistical features to construct the stage feature vector; The stage feature vector is input into the pre-trained cost weight calculation module to obtain curvature cost weight and distance cost weight. Curvature difference metric and spatial distance metric are calculated for the state representation point and the reference point on the reference trajectory, respectively. The curvature difference metric is weighted by the curvature cost weight to obtain the curvature matching cost. The spatial distance metric is weighted by the distance cost weight to obtain the distance matching cost. The curvature matching cost and the distance matching cost are added to obtain the comprehensive matching cost. Construct a cumulative cost matrix based on the comprehensive matching cost, find the minimum cumulative cost value of the predecessor position for each position in the cumulative cost matrix, add it to the current comprehensive matching cost and fill it into the current position; Starting from the termination position of the cumulative cost matrix, the backtracking operation is performed. The position with the minimum cumulative cost value among the predecessor positions of the current position is selected as the backtracking path node. The backtracking operation is performed iteratively until the starting position is reached to form an alignment path. Based on the alignment path, the mapping relationship between each state representation point on the state evolution trajectory and the corresponding reference point on the reference trajectory is established to obtain the nonlinear time alignment mapping relationship.

5. The method according to claim 1, characterized in that, When an abnormal deviation flag is received, current and temperature data sequences are extracted from the charging state data, and sliding window sampling with different time offsets is performed to calculate the correlation peak. The conduction delay is determined based on the time offset corresponding to the correlation peak, and the conduction coefficient is determined based on the value of the correlation peak. The cross-parameter conduction relationship is obtained, including: Based on the time boundary information in the abnormal deviation identifier, the current data sequence and temperature data sequence within the corresponding time range are extracted from the charging state data. The current gradient sequence and temperature gradient sequence are constructed by performing time domain differentiation respectively. The gradient abrupt change positions are identified to obtain the current abrupt change point set and the temperature abrupt change point set respectively. The corresponding time interval distribution is calculated, and the sliding window length is determined based on the statistical characteristics of the time interval distribution. A fixed reference window is set for the current data sequence to extract the current reference window data. A sliding window is set for the temperature data sequence and time offset sampling is performed according to a preset time step. Temperature window data within the sliding window is extracted at each time offset position. The first information entropy value of the temperature window data and the second information entropy value of the current reference window data are calculated. The conditional mutual information between the first information entropy value and the second information entropy value is constructed as the cross-correlation metric value at that time offset position. The cross-correlation metric sequence is obtained by traversing all time offset positions. A cumulative increment sequence is constructed from the cross-correlation metric sequence. The position of maximum curvature is identified in the cumulative increment sequence to determine the correlation peak position. The time offset corresponding to the correlation peak position is extracted as the propagation delay. The corresponding cross-correlation metric is extracted as the propagation coefficient. Cross-parameter propagation relationship is constructed.

6. The method according to claim 5, characterized in that, Calculate the corresponding time interval distribution, and determine the sliding window length based on the statistical characteristics of the time interval distribution, including: Extract the time difference between each current mutation point in the current mutation point set and the temperature mutation point in the temperature mutation point set that is closest in time. Collect all time differences to form the original time interval sample set. Calculate the jump amplitude between each time interval sample in the original time interval sample set and its adjacent time interval samples. Remove abnormal time interval samples according to the jump amplitude to obtain the filtered time interval sample set. The rate of change of values ​​between adjacent samples is calculated in chronological order for the filtered time interval sample set, and the rate of change of values ​​is used as the gradient value of the time interval gradient field. In the time interval gradient field, identify the time region where the gradient value decreases continuously and the rate of decrease is the largest. Determine this time region as the time interval convergence region. Extract the median value of all samples in the time interval convergence region and determine this median value as the reference time interval. A nonlinear adjustment coefficient is constructed based on the deviation intensity value in the abnormal deviation identifier. The adaptive time interval is obtained by multiplying the baseline time interval by the nonlinear adjustment coefficient, and this time interval is determined as the sliding window length.

7. The method according to claim 1, characterized in that, The prediction time window is determined based on the conduction delay in the cross-parameter conduction relationship. Within the prediction time window, phase space reconstruction is performed, the compensation direction vector and current compensation are calculated, and corresponding control commands are generated and sent to the charging device, including: The conduction delay and conduction coefficient are extracted from the cross-parameter conduction relationship. The prediction time window is determined based on the conduction delay. The historical current evolution trajectory in the charging state data is extracted within the prediction time window. The phase space of the historical current evolution trajectory is reconstructed to obtain the current state attractor. The topological deviation of the current state attractor is calculated. The topological deviation and the deviation intensity are weighted and fused to obtain the comprehensive instability metric. The initial compensation intensity is obtained by nonlinearly mapping the comprehensive instability metric value and the conduction coefficient. The compensation direction vector is determined according to the topological characteristics of the current state attractor. The initial compensation intensity is decomposed along the compensation direction vector to obtain the vectorized current compensation amount. The amplitude of the vectorized current compensation amount is extracted to determine the current compensation amount. A control command data structure containing current compensation amount and compensation direction vector is constructed. The conduction delay amount is embedded in the control command data structure as a delay execution identifier, and it is encapsulated into a control command and sent to the charging device.

8. An AI-based charging process safety monitoring system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect current data, voltage data, and temperature data of the charging device to obtain charging status data. The feature extraction unit is used to input charging state data into the neural network encoder, extract hidden feature vectors through multi-layer nonlinear transformation, and project the hidden feature vectors onto a feature space of a preset dimension to obtain state representation points in the feature space. The trajectory matching unit is used to acquire the state characterization points at continuous time points to form the state evolution trajectory, perform morphological matching between the state evolution trajectory and the pre-stored reference trajectory, calculate the local curvature deviation and spatial distance deviation of the state evolution trajectory, identify abnormal deviation segments and obtain abnormal deviation identifiers and deviation intensity values. The conduction analysis unit is used to extract current data sequence and temperature data sequence from charging state data when an abnormal deviation flag is received, and to perform sliding window sampling with different time offsets, calculate correlation peak value, determine conduction delay based on the time offset corresponding to the correlation peak value, determine conduction coefficient based on the value of the correlation peak value, and obtain cross-parameter conduction relationship. The compensation control unit is used to determine the prediction time window based on the conduction delay in the cross-parameter conduction relationship, perform phase space reconstruction within the prediction time window, calculate the compensation direction vector and current compensation amount, generate corresponding control commands and send them to the charging device.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images