Electric two-wheeled vehicle charging thermal runaway early warning method and system based on ai

Abstract

The present application relates to the technical field of electric two-wheeled vehicle charging safety, and particularly relates to an electric two-wheeled vehicle charging thermal runaway early warning method and system based on AI. Multidimensional monitoring data collected by a charging pile sensor network in real time is obtained, correlation mining across time domains and space domains is performed based on a space-time attention mechanism to generate comprehensive risk features, and then an early warning decision result containing a risk level and a warning time node is generated, an adaptive multi-stage charging control strategy is executed according to the early warning decision result, and a timing charging adjustment instruction is dynamically generated, a multi-level cascading early warning information pushing mechanism is established, and the early warning content and the instruction are encapsulated and pushed through multiple channels simultaneously. The present application realizes early warning and accurate control of charging thermal runaway, and effectively improves charging safety.

Description

Technical Field

[0001] This invention relates to the field of electric two-wheeled vehicle charging safety technology, and in particular to an AI-based method and system for early warning of thermal runaway during electric two-wheeled vehicle charging. Background Technology

[0002] Fires caused by battery thermal runaway during charging of electric two-wheeled vehicles are becoming increasingly frequent, seriously threatening users' lives and property safety as well as public safety. Existing charging thermal runaway early warning technologies mostly rely on monitoring single sensor parameters, such as using temperature sensors to detect the real-time temperature of the battery surface or charging interface, combined with fixed thresholds to trigger alarms. Some solutions also incorporate abnormal fluctuations in voltage or current as auxiliary judgment criteria. These conventional approaches typically employ periodic data collection and simple logical comparisons; when the monitored value exceeds a preset threshold, the system executes protective actions such as cutting off the charging circuit or issuing audible and visual alarms. Furthermore, most charging piles deployed in the power distribution network only have local fault alarm functions, lacking real-time data connectivity with user terminals or management platforms, resulting in extremely limited dissemination of early warning information.

[0003] The aforementioned conventional early warning methods have significant shortcomings. Single-sensor monitoring is insufficient to capture the early evolutionary characteristics of battery thermal runaway, as thermal runaway is usually accompanied by gradual coupled changes in multiple physical quantities, such as increased internal resistance, minor air pressure fluctuations, and uneven temperature differences. Relying solely on isolated thresholds for temperature or electrical parameters is prone to missed or false alarms. Fixed threshold alarm mechanisms lack dynamic assessment of risk trends and often only trigger when danger is imminent or even when visible anomalies have occurred, failing to provide sufficient intervention time for users and safety management systems. Early warning information delivery also relies primarily on local notifications from the charging pile itself, lacking multi-channel synchronization capabilities for user mobile devices and centralized monitoring platforms. This hinders remote real-time early warning and coordinated response, limiting the timeliness and overall reliability of the early warning system and failing to effectively curb the occurrence of thermal runaway accidents during electric two-wheeler charging. Summary of the Invention

[0004] This invention provides an AI-based method and system for early warning of thermal runaway during charging of electric two-wheeled vehicles, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides an AI-based method for early warning of thermal runaway during charging of electric two-wheeled vehicles, comprising: Acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheelers, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data; Based on the spatiotemporal attention mechanism, the correlation of the multi-dimensional monitoring data is mined across the time domain and across the spatial domain to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway. Based on the comprehensive risk characteristics, an early warning judgment result is generated, which includes the thermal runaway risk level and the early warning time node; An adaptive multi-stage charging control strategy is executed based on the warning judgment result. A differentiated control threshold library is established for different risk levels. Time-sequential charging adjustment instructions are dynamically generated in combination with the warning time node and the remaining battery charging demand. A multi-level linkage early warning information push mechanism is established, which encapsulates the early warning judgment results and the time-sequential charging adjustment instructions into differentiated early warning content, and pushes them synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

[0006] Based on a spatiotemporal attention mechanism, the multi-dimensional monitoring data is subjected to cross-temporal and cross-spatial correlation mining to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway, including: The multi-dimensional monitoring data is divided into constant current charging segment data and constant voltage charging segment data according to the charging stage. For each charging stage, a stage feature tensor containing the charging interface temperature data and the charging circuit electrical characteristic data is constructed. By analyzing the mapping relationship between the temperature distribution at different locations of the charging interface and the current density distribution at the corresponding locations, a spatial correlation diagram characterizing the heat conduction path and current loop topology is generated. A sliding window analysis is performed on the stage feature tensor to calculate the covariance matrix of the temperature rise rate and voltage fluctuation amplitude within each time window, and time-domain sensitivity weights are generated based on the eigenvalue distribution of the covariance matrix. Graph convolution is performed on the spatial correlation graph to capture the spatial propagation pattern of temperature anomalies spreading to the surrounding area by aggregating the feature information of adjacent nodes, and to generate spatial propagation weights. Based on the time-domain sensitivity weight and spatial propagation weight, the stage feature tensor is weighted and fused to obtain a fused feature vector. The temperature evolution component and electrical fluctuation component are extracted, and environmental compensation is performed in combination with the charging environment state data. The compensated temperature evolution component is converted into a thermal runaway trigger probability, and the compensated electrical fluctuation component is converted into an evolution velocity prediction value to obtain a comprehensive risk feature.

[0007] By analyzing the mapping relationship between temperature distribution and current density distribution at different locations of the charging interface, a spatial correlation map characterizing the heat conduction path and current loop topology is generated, including: The contact area and conductor area of ​​the charging interface are spatially gridded. Temperature acquisition points are arranged at the center point of each grid cell, and current acquisition points are arranged at each conductor segment. An index table of acquisition point numbers and grid cell numbers is established. Temperature data of each grid cell and current data of each conductor segment are obtained. Current density data is calculated by combining the conductor cross-sectional area. The temperature data and current density data are associated with the corresponding grid cell according to the index table. A heat conduction equation based on Fourier's law and a heating equation based on Joule's law are established. The heat conduction equation describes the heat flow driven by the temperature gradient, and the heating equation uses the square of the current density to represent the volumetric heating rate. The theoretical temperature value is obtained by coupling the two equations and solving them on each grid element. By comparing the theoretical and measured temperature values ​​of each grid cell, grid cells with deviations exceeding the threshold are marked as risk cells. The spatial connectivity and series-parallel connections of the risk cells are statistically analyzed. The risk unit is used as a node to construct the graph. Heat conduction edges are added according to the spatial adjacency relationship of the grid unit, and current topology edges are added according to the series and parallel relationship of the current path. The weight of the heat conduction edge is calibrated by the temperature gradient, and the weight of the current topology edge is calibrated by the current distribution ratio, thus forming a spatial association graph.

[0008] Based on the aforementioned comprehensive risk characteristics, a warning determination result is generated, including the thermal runaway risk level and warning time point, comprising: Extract the predicted values ​​of thermal runaway trigger probability and evolution rate from the comprehensive risk features, construct a probability-velocity two-dimensional feature space, and determine the initial thermal runaway risk level based on the coordinate position; Based on the initial thermal runaway risk level, the evolution trajectory template library is retrieved, and the temporal change trajectory of the thermal runaway trigger probability is matched with the standard trajectory to obtain the trajectory deviation coefficient. When the trajectory deviation coefficient exceeds the safety threshold, the corrected thermal runaway risk level is obtained. The evolution trend correction factor is determined based on the trajectory deviation coefficient. The predicted evolution rate is multiplied by the evolution trend correction factor to obtain the corrected evolution rate. The evolution time required to evolve from the current trigger probability to the critical trigger probability is calculated. The predicted evolution rate is analyzed in the time domain and the rate of change is calculated as an acceleration index. The time reduction is calculated based on the acceleration index. The time reduction is subtracted from the evolution duration to obtain the corrected evolution duration. This corrected evolution duration is added to the current time to determine the warning time node. The corrected thermal runaway risk level, warning time node, trajectory deviation coefficient, and acceleration index are encapsulated to form the warning judgment result.

[0009] Based on the warning judgment result, an adaptive multi-stage charging control strategy is executed. A differentiated control threshold library is established for different risk levels. Combined with the warning time node and the remaining battery charging demand, a time-sequential charging adjustment command is dynamically generated, including: The differentiated control threshold library is divided into a prevention layer threshold set and a response layer threshold set. The corresponding threshold set is selected according to the corrected thermal runaway risk level. The trajectory deviation coefficient is used as the layer penetration factor. When the layer penetration factor exceeds the boundary value, the threshold sets are fused to generate a hybrid threshold set. Establish a constraint propagation network, setting the remaining battery charging demand as the target node, the early warning time node as the time constraint node, and the mixed threshold set as the intermediate constraint node. In the constraint propagation network, energy demand is propagated backward from the target node, time margin is propagated forward from the time constraint node, and the solution space of control parameters is generated by converging at the intermediate constraint nodes. Multi-objective optimization is performed on the solution space of the control parameters, with charging completion degree as the main objective and risk suppression degree as the constraint objective. The non-dominated solution set is obtained through Pareto front search. The optimal solution vector is extracted from the non-dominated solution set and decomposed into discrete control actions. Spline interpolation is performed on the discrete control actions to generate a continuous control instruction sequence, which is then encapsulated as a time-series charging control instruction.

[0010] Multi-objective optimization is performed on the solution space of the control parameters, with charging completion rate as the primary objective and risk suppression rate as the constraint objective. The non-dominated solution set is obtained through Pareto front search, including: Orthogonal decomposition is performed on the solution space of the control parameters, and the solution space of the control parameters is projected to the charging efficiency subspace and the safety margin subspace. In the charging efficiency subspace, the product of the current adjustment parameter and the warning time node is calculated as the available charging time window. The normalized difference between the theoretical charging amount and the remaining charging demand of the battery within the available charging time window is used as the charging completion degree. In the safety margin subspace, the power adjustment parameter and the trajectory deviation coefficient are weighted and fused as the risk suppression degree. The weighting coefficients for charging completion are determined based on the acceleration index, and the weighting coefficients for risk suppression are determined based on the corrected thermal runaway risk level. A weighted scalar objective function is constructed, and contour scanning is performed on the weighted scalar objective function to generate contour clusters in the solution space of the control parameters. The intersections of each contour line with the risk suppression constraint boundary are extracted to form the initial Pareto front. Local search optimization is performed on the initial Pareto front. Iterative updates are performed in the neighborhood of each intersection along the gradient direction of the charging completion evaluation function. When the iterative solution exceeds the risk suppression constraint boundary, it backs to the boundary and terminates the iteration. All converged solutions form a non-dominated solution set.

[0011] A second aspect of this invention provides an AI-based early warning system for thermal runaway during charging of an electric two-wheeled vehicle, comprising: The data acquisition unit is used to acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheeled vehicles, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data. The risk feature unit is used to perform cross-temporal and cross-spatial correlation mining on the multi-dimensional monitoring data based on the spatiotemporal attention mechanism to generate comprehensive risk features that characterize the evolution trend of battery thermal runaway. The early warning determination unit is used to generate an early warning determination result that includes the thermal runaway risk level and the early warning time node based on the comprehensive risk characteristics. The charging control unit is used to execute an adaptive multi-stage charging control strategy based on the warning judgment result, establish a differentiated control threshold library for different risk levels, and dynamically generate time-sequential charging adjustment instructions in combination with the warning time node and the remaining charging demand of the battery. The early warning push unit is used to establish a multi-level linkage early warning information push mechanism, which encapsulates the early warning judgment result and the time-sequential charging adjustment instruction into differentiated early warning content, and pushes it synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

[0012] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] 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.

[0014] The beneficial effects of the embodiments of the present invention are as follows: By employing a spatiotemporal attention mechanism to deeply explore the evolution patterns of multidimensional monitoring data in the temporal and spatial dimensions, the sensitivity of capturing early signs of thermal runaway and the accuracy of characterizing comprehensive risk features are significantly improved, effectively reducing the false alarm and missed alarm rates.

[0015] After generating a judgment result that includes risk level and warning time point, the adaptive multi-stage charging control strategy dynamically adjusts charging parameters based on a differentiated threshold library. In low-risk stages, it maintains normal charging efficiency, while in medium- and high-risk stages, it proactively reduces current or suspends charging to prevent continuous overcharging or exacerbated heat accumulation. The time-sequential charging adjustment instructions generated by combining the warning time point and remaining charging demand ensure that the charging process is completed within a safe margin, suppressing thermal runaway while also considering the user's charging timeliness.

[0016] The multi-level linkage early warning push mechanism encapsulates the early warning judgment results and charging adjustment instructions into differentiated content, and pushes them simultaneously to the local display unit of the charging pile, the user's mobile terminal, and the monitoring center management platform, ensuring that on-site charging personnel, remote users, and operation and maintenance management personnel can obtain risk information as soon as possible. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the AI-based early warning method for thermal runaway during charging of an electric two-wheeled vehicle, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the method for generating comprehensive risk features based on a spatiotemporal attention mechanism according to an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] Figure 1 This is a flowchart illustrating an AI-based early warning method for thermal runaway during charging of an electric two-wheeled vehicle, according to an embodiment of the present invention. The present invention provides an AI-based early warning method for thermal runaway during charging of an electric two-wheeled vehicle, comprising: Acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheelers, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data; Based on the spatiotemporal attention mechanism, the correlation of the multi-dimensional monitoring data is mined across the time domain and across the spatial domain to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway. Based on the comprehensive risk characteristics, an early warning judgment result is generated, which includes the thermal runaway risk level and the early warning time node; An adaptive multi-stage charging control strategy is executed based on the warning judgment result. A differentiated control threshold library is established for different risk levels. Time-sequential charging adjustment instructions are dynamically generated in combination with the warning time node and the remaining battery charging demand. A multi-level linkage early warning information push mechanism is established, which encapsulates the early warning judgment results and the time-sequential charging adjustment instructions into differentiated early warning content, and pushes them synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

[0021] Figure 2 This is a flowchart illustrating the method for generating comprehensive risk features based on a spatiotemporal attention mechanism according to an embodiment of the present invention. Based on the spatiotemporal attention mechanism, the multi-dimensional monitoring data undergoes cross-temporal and cross-spatial correlation mining to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway, including: The multi-dimensional monitoring data is divided into constant current charging segment data and constant voltage charging segment data according to the charging stage. For each charging stage, a stage feature tensor containing the charging interface temperature data and the charging circuit electrical characteristic data is constructed. By analyzing the mapping relationship between the temperature distribution at different locations of the charging interface and the current density distribution at the corresponding locations, a spatial correlation diagram characterizing the heat conduction path and current loop topology is generated. A sliding window analysis is performed on the stage feature tensor to calculate the covariance matrix of the temperature rise rate and voltage fluctuation amplitude within each time window, and time-domain sensitivity weights are generated based on the eigenvalue distribution of the covariance matrix. Graph convolution is performed on the spatial correlation graph to capture the spatial propagation pattern of temperature anomalies spreading to the surrounding area by aggregating the feature information of adjacent nodes, and to generate spatial propagation weights. Based on the time-domain sensitivity weight and spatial propagation weight, the stage feature tensor is weighted and fused to obtain a fused feature vector. The temperature evolution component and electrical fluctuation component are extracted, and environmental compensation is performed in combination with the charging environment state data. The compensated temperature evolution component is converted into a thermal runaway trigger probability, and the compensated electrical fluctuation component is converted into an evolution velocity prediction value to obtain a comprehensive risk feature.

[0022] Dividing multi-dimensional monitoring data according to charging stages is a prerequisite step for spatiotemporal attention analysis. During the charging process of electric two-wheelers, the charging pile sensor network continuously collects charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data. Based on the current and voltage variation patterns in the charging control protocol, the collected raw data is divided into two subsets: constant current charging stage data and constant voltage charging stage data. The constant current charging stage is characterized by a stable charging current and a continuous increase in battery terminal voltage. During this stage, there is a strong correlation between the change in battery internal resistance and the rate of temperature rise. The constant voltage charging stage is characterized by the charging voltage being maintained at a set upper limit while the charging current gradually decreases. During this stage, the battery's state of charge approaches saturation, and the risks of lithium plating and local overheating increase. For each charging stage, time-series samples are extracted from the charging interface temperature data and the charging circuit electrical characteristic data. Using sensor sampling points as the spatial dimension, time steps as the temporal dimension, and various physical quantity measurements as the feature dimension, a three-dimensional stage feature tensor is constructed, thus providing structured input for subsequent joint temporal and spatial analysis.

[0023] By analyzing the mapping relationship between temperature distribution and corresponding current density distribution at different locations of the charging interface, the heat conduction path within the charging interface and the topological structure of current in the loop can be revealed. Multiple temperature sensors are deployed at the charging interface, covering key locations such as the positive and negative contacts, the insulating shell, and connecting wires; simultaneously, electrical characteristic sensors provide current amplitude information for each branch. Each sensor node is considered a vertex in a graph structure. If a physical heat conduction path or electrical connection exists between two nodes, an edge is established between them, with the product of the temperature gradient and current density used as the edge weight, thus constructing a spatial relational graph. This graph structure can intuitively reflect the directionality of heat conduction from high-temperature regions to low-temperature regions, as well as the spatial correspondence between current concentration areas and temperature rise hotspots, providing a topological foundation for subsequent graph convolution operations.

[0024] A sliding window analysis is performed on the stage feature tensor, sliding a fixed-length time window along the time-series dimension. Each sliding step is a sampling interval. Within each time window, the time series of charging interface temperature data and charging circuit voltage data are extracted, and the temperature rise rate sequence and voltage fluctuation amplitude sequence are calculated respectively. Let the temperature rise rate sequence within the window be... The voltage fluctuation amplitude sequence is Then the covariance matrix within this window Defined as by and The sample covariance of the two-dimensional observation matrix constructed by splicing. For the covariance matrix... Perform eigenvalue decomposition to obtain the eigenvalue set. Among them, the larger eigenvalues Reflects the principal direction energy of the combined change in the rate of temperature rise and the amplitude of voltage fluctuation, with relatively small eigenvalues. This reflects the residual change energy in orthogonal directions. The relative distribution of eigenvalues ​​is used as a measure of the significance of this time window to calculate the time-domain sensitivity weight. Its expression is ,in To prevent small positive numbers with a denominator of zero. The more drastic the coordinated change between the rate of temperature rise and the amplitude of voltage fluctuations, The larger, The closer it is to 1, the more sensitive the time window is to thermal runaway evolution.

[0025] Graph convolution is performed on the spatial correlation graph to capture the spatial propagation pattern of temperature anomalies spreading to surrounding areas. The core operation of graph convolution is the weighted aggregation of the feature vector of each vertex and the feature vectors of its neighboring vertices. The aggregation weights are determined by the normalized adjacency matrix. Let the graph contain... Nodes, Nodes The initial feature vector consists of the temperature and current density measurements at that node. Through layer-by-layer propagation via the graph convolutional layers, the feature representation of each node gradually incorporates the heat conduction and current distribution information from its neighborhood. When the temperature of a node abnormally rises, the graph convolution operation can propagate this abnormal signal along the edge weights to adjacent nodes, thus forming a spatial propagation pattern in the feature space that reflects the range and intensity of the abnormal propagation. Global pooling is then performed on the graph convolutional output features of each node to obtain the spatial propagation weight vector for the entire charging interface region. Its various dimensional components reflect the degree of contribution of different spatial regions to the overall thermal runaway risk.

[0026] Based on time-domain sensitivity weights and airspace propagation weight A weighted fusion operation is performed on the stage feature tensor; specifically, the temporal sensitivity weights are... Applying to the temporal dimension, it performs a weighted summation of features from different time windows; it also propagates the spatial weights. The algorithm operates in the spatial dimension, weighting and aggregating features from different sensor nodes. The results of the two weighting operations are concatenated and linearly transformed to obtain a fused feature vector. Two types of components are extracted from the fused feature vector: the temperature evolution component, mainly composed of the weighted mean of the temperature rise rate within each time window and its trend slope, reflects the dynamic evolution of heat accumulation; and the electrical fluctuation component, mainly composed of the weighted statistics of voltage fluctuation amplitude and the frequency characteristics of abnormal current pulses, reflects the degree of electrochemical reaction abnormality in the charging circuit.

[0027] After obtaining the temperature evolution component and the electrical fluctuation component, environmental compensation is performed by combining charging environment status data, which includes information such as ambient temperature, ambient humidity, and ventilation status of the charging location. High ambient temperature can introduce background thermal noise into sensor measurements; therefore, the predicted value of the baseline temperature rise model based on the ambient temperature needs to be subtracted from the temperature evolution component to eliminate the impact of environmental thermal interference on thermal runaway detection. Excessive ambient humidity can affect the insulation resistance of the charging interface, introducing voltage fluctuations caused by non-electrochemical factors into the electrical fluctuation component; therefore, the electrical fluctuation component needs to be corrected based on the correlation between humidity and insulation resistance. After environmental compensation, the compensated temperature evolution component is input into a pre-trained thermal runaway trigger probability mapping function, which outputs the thermal runaway trigger probability. This probability value, ranging from 0 to 1, reflects the likelihood of thermal runaway occurring in the battery under the current charging state. The compensated electrical fluctuation component is input into the evolution rate prediction model, which outputs the predicted evolution rate value. This value characterizes the rate at which the risk of thermal runaway evolves from the current state to the critical state, expressed in risk levels per minute. It represents the probability of thermal runaway triggering. and evolution rate prediction Together, these factors constitute a comprehensive risk characteristic, providing a quantitative basis for subsequent assessment of thermal runaway risk levels and calculation of early warning time points. The comprehensive risk characteristic includes information from both the static probability of thermal runaway and the dynamic evolution rate, enabling early warning assessments to consider both the severity of the current risk state and the urgency of future risk evolution, thereby supporting more precise adaptive charging control strategies.

[0028] By analyzing the mapping relationship between temperature distribution and current density distribution at different locations of the charging interface, a spatial correlation map characterizing the heat conduction path and current loop topology is generated, including: Spatial gridding is performed on the contact area and conductor area of ​​the charging interface. Temperature acquisition points are arranged at the center point of each grid cell, and current acquisition points are arranged in each conductor segment. An index table of acquisition point number and grid cell number is established. Acquire temperature data for each grid cell and current data for each conductor segment, calculate current density data by combining the conductor cross-sectional area, and associate temperature data and current density data with the corresponding grid cell according to the index table. A heat conduction equation based on Fourier's law and a heating equation based on Joule's law are established. The heat conduction equation describes the heat flow driven by the temperature gradient, and the heating equation uses the square of the current density to represent the volumetric heating rate. The theoretical temperature value is obtained by coupling the two equations and solving them on each grid element. By comparing the theoretical and measured temperature values ​​of each grid cell, grid cells with deviations exceeding the threshold are marked as risk cells. The spatial connectivity and series-parallel connections of the risk cells are statistically analyzed. The risk unit is used as a node to construct the graph. Heat conduction edges are added according to the spatial adjacency relationship of the grid unit, and current topology edges are added according to the series and parallel relationship of the current path. The weight of the heat conduction edge is calibrated by the temperature gradient, and the weight of the current topology edge is calibrated by the current distribution ratio, thus forming a spatial association graph.

[0029] In practical operation, the charging interface simultaneously undertakes two physical processes: power transmission and heat exchange. These two processes are highly coupled in spatial distribution. To accurately describe this coupling relationship, the geometry of the charging interface needs to be spatially meshed. Specifically, the charging interface is divided into two functional sub-regions: a contact area and a conductor area. Each sub-region is meshed using a uniform or adaptive step size, and each mesh cell corresponds to a spatial location number. Temperature acquisition points are placed at the center point of each mesh cell to obtain the local temperature value at that location; current acquisition points are placed in each conductor segment (including the main conductor, branch conductors, and contact conductors) to obtain the current value flowing through that segment. An index table is established between the acquisition point numbers and the mesh cell numbers, ensuring that each temperature and current acquisition point uniquely corresponds to its corresponding mesh cell, providing a structured mapping basis for subsequent data association.

[0030] After obtaining the measured temperature data of each grid cell, the current acquisition value of each conductor segment is simultaneously read and combined with the known conductor cross-sectional area. (in (Number the conductor segments), and calculate the current density of each conductor segment. ,satisfy ,in For the first The current density values ​​of the conductor segments are collected. Based on the index table, the current density values ​​of each conductor segment are mapped to the corresponding grid cells, so that each grid cell simultaneously holds both measured temperature and current density values, forming a joint temperature-current density dataset, which provides input for subsequent physical equation modeling.

[0031] In the physical modeling phase, heat conduction equations based on Fourier's law and heating equations based on Joule's law are established respectively. The heat conduction equation describes the propagation process of heat flow inside a conductor driven by a temperature gradient, and its steady-state form is: ,in For the thermal conductivity of a conductor, This is the theoretical value of the temperature field. This is the volumetric heating rate. The heating equation is based on Joule's law, expressing the volumetric heating rate as the product of the square of the current density and the resistivity, i.e. ,in For conductor resistivity, Let be the current density at that location. Substituting the heating equation into the heat conduction equation forms a coupled equation set. Numerical solutions are then performed on each grid cell under given boundary conditions (including ambient temperature boundaries and adiabatic boundaries) to obtain the theoretical temperature value for each grid cell. ,in Number the grid cells.

[0032] The theoretical temperature values ​​of each grid cell Corresponding measured temperature value Perform unit-by-unit comparison and calculate the deviation. Set the deviation threshold. ,when When a deviation exceeds a threshold, the corresponding grid cell is marked as a risk cell. A deviation exceeding the threshold indicates a significant discrepancy between the actual thermal behavior at that location and the theoretical prediction. This is typically caused by factors such as abnormally high contact resistance, localized oxide layer formation, or conductor cross-section damage, and represents an early spatial manifestation of thermal runaway. After marking risk cells, the spatial connectivity of these risk cells is statistically analyzed to determine whether adjacent grid cells are simultaneously risk cells, thereby identifying continuous thermal anomaly regions. Simultaneously, the series and parallel connections of the conductor segments containing risk cells within the overall circuit are analyzed to determine whether they belong to a series main path or a parallel branch, providing a basis for subsequent edge type classification in graph construction.

[0033] Using risk cells as nodes in the graph, edges are added based on two types of physical relationships to construct a spatial association graph. The first type is heat conduction edges: if two risk cells are adjacent in the grid space (sharing a boundary or vertex), a heat conduction edge is added between the corresponding nodes to represent the probability of heat propagation along a spatial path. The weight of the heat conduction edge is determined by... Temperature gradient between two nodes According to the calibration, the larger the temperature gradient, the stronger the heat flow driving force, and the higher the corresponding edge weight. ,in The first type is the maximum temperature gradient among all heat conduction edges, used for normalization. The second type is current topology edges: if the conductor segments containing two risk units are connected in series or parallel in the circuit, a current topology edge is added between the corresponding nodes to characterize the functional association on the current path. The weight of the current topology edge... Based on current distribution ratio The calibration, or current distribution ratio, reflects the proportion of the current flowing through that conductor segment to the total current, i.e. In the serial path Take 1, in the parallel branch Take the ratio of the corresponding branch current to the total current.

[0034] The resulting spatial graph contains two semantically distinct types of edges: heat conduction edges and current topology edges. Node attributes include temperature deviation. and current density The edge attributes include normalized temperature gradient weights. and current distribution ratio weight This graph structure intuitively reflects the spatial orientation of heat conduction paths and the topological characteristics of current loops, enabling the assessment of thermal runaway risk to move beyond a single-point temperature exceedance. Instead, it reveals a multi-point collaborative evolution risk propagation pattern through graph connectivity and path analysis. For example, if multiple risk units form a connected subgraph via heat conduction edges, and the nodes in this subgraph are simultaneously connected by high-weight current topological edges, it indicates the presence of a thermal-electric coupling positive feedback path in this region, representing a highly concentrated area of ​​thermal runaway risk. This region requires a higher spatial propagation weight in the subsequent comprehensive risk feature generation stage. This spatial correlation graph, as a structured input for spatial correlation analysis in the spatiotemporal attention mechanism, together with the temporal monitoring sequence, supports the extraction of comprehensive risk features across both temporal and spatial domains.

[0035] Based on the aforementioned comprehensive risk characteristics, a warning determination result is generated, including the thermal runaway risk level and warning time point, comprising: Extract the predicted values ​​of thermal runaway trigger probability and evolution rate from the comprehensive risk features, construct a probability-velocity two-dimensional feature space, and determine the initial thermal runaway risk level based on the coordinate position; Based on the initial thermal runaway risk level, the evolution trajectory template library is retrieved, and the temporal change trajectory of the thermal runaway trigger probability is matched with the standard trajectory to obtain the trajectory deviation coefficient. When the trajectory deviation coefficient exceeds the safety threshold, the corrected thermal runaway risk level is obtained. The evolution trend correction factor is determined based on the trajectory deviation coefficient. The predicted evolution rate is multiplied by the evolution trend correction factor to obtain the corrected evolution rate. The evolution time required to evolve from the current trigger probability to the critical trigger probability is calculated. Perform time-domain analysis on the predicted evolution rate and calculate the rate of change as an acceleration index. Calculate the time reduction based on the acceleration index, subtract the time reduction from the evolution duration to obtain the corrected evolution duration, and add it to the current time to determine the warning time node. The corrected thermal runaway risk level, early warning time point, trajectory deviation coefficient, and acceleration index are encapsulated to form an early warning judgment result.

[0036] Extracting thermal runaway trigger probability from comprehensive risk characteristics and evolution rate prediction Using these two quantities as coordinate axes, a probability-velocity two-dimensional feature space is constructed. Within this space, several partition boundaries are pre-defined based on historical accident data and simulation results, dividing the entire space into four regions: low-risk, medium-risk, high-risk, and emergency-risk. When a feature coordinate point at a given moment falls into the corresponding region, the initial thermal runaway risk level at that moment is determined. For example, when… lower and When the risk level is low, the coordinates fall within the low-risk zone, and the initial risk level is Level 1; when... Exceeding the medium threshold or When the value increases significantly, the coordinate point enters the high-risk zone, and the initial level rises to level three. This two-dimensional joint judgment method can more comprehensively capture the early signs of thermal runaway compared to a single threshold judgment, avoiding misjudgment or missed judgment due to fluctuations in data of a certain dimension.

[0037] After obtaining the initial thermal runaway risk level, the corresponding set of standard trajectories is retrieved from a pre-built evolutionary trajectory template library based on this level. The evolutionary trajectory template library stores typical curves showing the evolution of thermal runaway trigger probability over time at different risk levels. These curves are derived from statistical fitting of a large amount of experimental data and the summarization of expert experience. The observed thermal runaway trigger probability time-series trajectory is matched with each curve in the set of standard trajectories for similarity. A dynamic time warping algorithm is used to calculate the distance between the two trajectories, and the standard trajectory with the smallest distance is taken as the reference benchmark. The normalized distance between the measured trajectory and the reference benchmark is defined as the trajectory deviation coefficient. .when Exceeding the preset security threshold When this occurs, it indicates that the actual evolutionary path has deviated from the normal pattern, requiring correction of the initial risk level. This involves raising the risk level by one or more increments to obtain the corrected thermal runaway risk level. The introduction of this correction mechanism can effectively identify potentially dangerous states where the absolute value of the trigger probability has not yet reached the high-risk range, but the evolutionary trajectory already shows an abnormal acceleration, thereby improving the foresight of early warnings.

[0038] Determine the trajectory deviation coefficient Then, the evolutionary trend correction factor is calculated based on this. . and Positive correlation; the greater the deviation in trajectory, A larger value reflects the degree of acceleration in the actual evolutionary rate relative to the standard model. Specifically, This is determined through the following relationship: ,in This is the sensitivity adjustment coefficient, obtained through experimental calibration, used to control the response strength of the correction factor to the degree of deviation. The predicted evolution rate is... Evolutionary trend correction factor Multiply to obtain the corrected evolution rate. This value reflects the actual rate of progression of thermal runaway evolution under the current trajectory deviation state.

[0039] Obtaining the corrected evolutionary rate Then, calculate the probability of triggering from the current position. Evolution to critical trigger probability Required evolution time Critical trigger probability This is a pre-defined probability boundary value for irreversible thermal runaway. When the trigger probability reaches this value, thermal runaway is considered to have entered an uncontrollable stage. The evolution duration is calculated based on the difference between the current probability value and the critical probability value, as well as the adjusted evolution rate. ,in This expression is meaningful. If It has approached or exceeded If this occurs, the highest level of warning will be triggered immediately, and emergency control measures will be implemented.

[0040] Evolution rate prediction Perform time-domain analysis and calculate within a sliding time window. The first difference yields the rate of change of the evolution velocity, which is defined as the acceleration index. . This reflects the increasing or decreasing trend of the thermal runaway evolution rate itself: when At this point, the evolutionary rate continues to accelerate, meaning that the actual time to reach the critical state will be shorter than [previous time]. The prediction; when When the evolutionary rate tends to stabilize or slow down, the time correction can be set to zero or a negative value. Based on the acceleration index... Calculation time reduction Its calculation method is as follows ,in This is the time-reduction sensitivity coefficient, whose value range has been fixed after experimental verification. It is used to map the acceleration effect into a lead time in the time dimension. The evolution duration is also considered. Subtract time reduction The corrected evolution time was obtained. This value represents the estimated remaining time from the current moment until the critical state of thermal runaway. With the current moment Add them together to determine the warning time point. This node marks the time when the risk of thermal runaway is most likely to occur, providing an accurate time reference for the time-sequential scheduling of subsequent charging control strategies.

[0041] This will correct the thermal runaway risk level and early warning time point. Trajectory deviation coefficient and acceleration index The four pieces of information are encapsulated into a complete early warning judgment result. The corrected thermal runaway risk level determines the responsiveness of subsequent control strategies; the early warning time node provides a time coordinate for the generation of time-sequential charging adjustment commands; and the trajectory deviation coefficient and acceleration index serve as auxiliary diagnostic information transmitted to the monitoring center for in-depth fault tracing and analysis by maintenance personnel. This combined encapsulation of the four pieces of information makes the early warning judgment result both immediate and predictive, supporting the coordinated response of multi-level linked early warning mechanisms at different time scales. In actual charging scenarios, this early warning judgment process is executed on a rolling basis within each sampling cycle, and as new data is continuously integrated… , , , The indicators are updated in real time, and the warning time points are dynamically adjusted accordingly, so as to ensure that the warning judgment results always reflect the latest battery thermal state evolution and avoid warning delays or false alarms due to early data failure.

[0042] Based on the warning judgment result, an adaptive multi-stage charging control strategy is executed. A differentiated control threshold library is established for different risk levels. Combined with the warning time node and the remaining battery charging demand, a time-sequential charging adjustment command is dynamically generated, including: The differentiated control threshold library is divided into a prevention layer threshold set and a response layer threshold set. The corresponding threshold set is selected according to the corrected thermal runaway risk level. The trajectory deviation coefficient is used as the layer penetration factor. When the layer penetration factor exceeds the boundary value, the threshold sets are fused to generate a hybrid threshold set. Establish a constraint propagation network, setting the remaining battery charging demand as the target node, the early warning time node as the time constraint node, and the mixed threshold set as the intermediate constraint node. In the constraint propagation network, energy demand is propagated backward from the target node, time margin is propagated forward from the time constraint node, and the solution space of control parameters is generated by converging at the intermediate constraint nodes. Multi-objective optimization is performed on the solution space of the control parameters, with charging completion degree as the main objective and risk suppression degree as the constraint objective. The non-dominated solution set is obtained through Pareto front search. The optimal solution vector is extracted from the non-dominated solution set and decomposed into discrete control actions. Spline interpolation is performed on the discrete control actions to generate a continuous control instruction sequence, which is then encapsulated as a time-series charging control instruction.

[0043] After obtaining the early warning judgment result (including the corrected thermal runaway risk level and the early warning time node), the core of implementing the adaptive multi-stage charging control strategy lies in constructing a threshold management and command generation system that can sense changes in risk level and respond dynamically. The differentiated control threshold library is divided into two subsets according to the risk response level: the prevention layer threshold set and the response layer threshold set. The prevention layer threshold set corresponds to the low-risk to medium-risk range, with relatively loose threshold boundaries, allowing the charging process to continue under controlled derating, prioritizing charging completion. The response layer threshold set corresponds to the high-risk and critical runaway range, with strictly tightened threshold boundaries, prioritizing risk suppression and triggering charging termination commands when necessary. Based on the corrected thermal runaway risk level, the corresponding threshold set is selected from the differentiated control threshold library as the current control constraint benchmark.

[0044] Introducing trajectory deviation coefficient As a hierarchical penetration factor, it is used to determine whether the current risk evolution trajectory crosses the boundary between the prevention layer and the response layer. Let the hierarchical boundary value be... ,when When, only the prevention layer threshold set is activated; when When it is determined that the risk evolution has penetrated the boundary of the prevention layer, the response layer threshold set cannot be relied upon alone. Instead, the prevention layer threshold set and the response layer threshold set are weighted and fused according to the degree of penetration to generate a hybrid threshold set. The hybrid threshold set contains the first... Threshold parameters The fusion computing method is ,in To prevent the first layer One threshold parameter, For the first in the response layer One threshold parameter, To achieve penetrating fusion weights, by Exceeding The amplitude normalization is obtained, satisfying The hybrid threshold set can create a smooth transition between two threshold layers, avoiding abrupt changes in control commands when risk levels shift, and improving the continuity and safety of charging control.

[0045] The constraint propagation network consists of three types of nodes, each undertaking different constraint semantics. The target node carries the remaining battery charging demand. This represents the total amount of energy still needed to complete the current charging task, calculated from the difference between the current battery state of charge and the target state of charge. The time constraint node carries the early warning time node. , indicating the current moment The maximum remaining time window available for safe charging afterwards. Intermediate constraint nodes carry a mixed threshold set. Electrical parameters (upper limits for charging current and charging voltage), thermal parameters (upper limits for interface temperature and temperature rise rate), and power parameters (upper limit for instantaneous charging power) are uniformly incorporated into the constraint system. The three types of nodes are connected by directed edges, forming a constraint propagation path from the target node to the time constraint node and then to the intermediate constraint node.

[0046] In a constrained propagation network, energy demand propagates from the target node along the reverse path. Decomposed into energy allocation for each charging stage ,in Number the charging stages to meet the requirements. The time slack propagates along the forward path from the time constraint node, and... Time quotas allocated to each charging phase ,satisfy At intermediate constraint nodes, the propagation results of energy demand and time margin converge, and combined with a mixed threshold set, the charging power range of each stage is limited, generating a solution space for control parameters. Solution space Each candidate solution vector This represents the target power sequence for each charging stage, where For the first Phase charging power, The total number of stages, each The upper and lower bound constraints imposed by the mixed threshold set must be satisfied.

[0047] Solution space of control parameters Perform multi-objective optimization, setting two optimization objectives: charging completion rate. and risk inhibition Charging completion rate Defined as the energy that can actually be transferred within the time quota according to the candidate power sequence and The higher the ratio, the higher the degree of completion of the charging task; maximizing this ratio is the optimal approach. Risk suppression level. Defined as the suppression effect of candidate power sequences on the risk of thermal runaway, by relating the power at each stage to the corresponding thermal runaway trigger probability. The predicted response is obtained by weighted integration; a larger value indicates a better risk mitigation effect, and maximizing this value is the optimization direction. With risk mitigation as the constraint objective, the following requirements are made: ,in The minimum acceptable risk suppression level is determined by mapping the current corrected thermal runaway risk level; the charging completion rate is the primary objective, and it should be maximized as much as possible while satisfying the constraints. A Pareto front search algorithm is employed. The candidate solutions in the pool are sorted non-dominatedly, and will be... and Candidate solutions that are not dominated by other solutions in either dimension are included in the non-dominated solution set. .

[0048] From the non-dominated solution set When extracting the optimal solution vector, the ideal point distance minimization criterion is adopted: calculate the distance between each non-dominated solution and the ideal point. The solution with the smallest normalized Euclidean distance between the two solutions is selected as the final optimal solution vector. Each element in the optimal solution vector corresponds to the target power value of a charging stage, forming a discrete control action sequence. Since there are power jumps between stages, directly switching by stage would cause charging current spikes, impacting the battery and charging circuit. Therefore, cubic spline interpolation is performed on the discrete control action sequence, using the start time of each stage as a node and the corresponding target power value as the node function value, generating a continuously differentiable power adjustment curve within the entire charging time window. Spline interpolation requires that the first derivative be continuous at each node to ensure a smooth power change rate without abrupt changes. The continuously controlled command sequence generated by interpolation is discretized and sampled according to a preset command issuance time step, forming a power command list with timestamps. This list is then encapsulated together with the corresponding current and voltage commands to form a time-sequential charging control command. Each command in the time-sequential charging control command includes the execution time, target charging power, corresponding current upper limit, and voltage upper limit. The charging pile controller executes the commands sequentially according to the command sequence, achieving refined dynamic control of the charging process. This ensures that the charging task is completed as much as possible while continuously suppressing the risk of thermal runaway within the safety boundary.

[0049] Multi-objective optimization is performed on the solution space of the control parameters, with charging completion rate as the primary objective and risk suppression rate as the constraint objective. The non-dominated solution set is obtained through Pareto front search, including: Orthogonal decomposition is performed on the solution space of the control parameters, and the solution space of the control parameters is projected to the charging efficiency subspace and the safety margin subspace. In the charging efficiency subspace, the product of the current adjustment parameter and the warning time node is calculated as the available charging time window. The normalized difference between the theoretical charging amount and the remaining charging demand of the battery within the available charging time window is used as the charging completion degree. In the safety margin subspace, the power adjustment parameter and the trajectory deviation coefficient are weighted and fused as the risk suppression degree. The weighting coefficients for charging completion are determined based on the acceleration index, and the weighting coefficients for risk suppression are determined based on the corrected thermal runaway risk level. A weighted scalar objective function is constructed, and contour scanning is performed on the weighted scalar objective function to generate contour clusters in the solution space of the control parameters. The intersections of each contour line with the risk suppression constraint boundary are extracted to form the initial Pareto front. Local search optimization is performed on the initial Pareto front. Iterative updates are performed in the neighborhood of each intersection along the gradient direction of the charging completion evaluation function. When the iterative solution exceeds the risk suppression constraint boundary, it backs to the boundary and terminates the iteration. All converged solutions form a non-dominated solution set.

[0050] In obtaining the non-dominated solution set, an orthogonal decomposition of the control parameter solution space is required, mapping the original high-dimensional solution space to two mutually orthogonal subspaces, called the charging efficiency subspace and the safety margin subspace. The purpose of orthogonal decomposition is to decouple the two optimization objectives, charging completion and risk suppression, which have different physical meanings, so that the subsequent Pareto front search can be efficiently carried out in a well-structured low-dimensional projection space. The charging efficiency subspace corresponds to the direction of change of the current regulation parameter, and the safety margin subspace corresponds to the coupled direction of change of the power regulation parameter and the risk state index. The two subspaces together span the effective optimization region of the original solution space.

[0051] In the charging efficiency subspace, current adjustment parameters and warning time nodes are used. The product definition can be used with a charging time window. ,Right now ,in These are the current regulation parameters for the current stage, with units matching the dimensions of the charging quantity. Available charging time window. This represents the upper bound of the theoretical charge that the current charging strategy can accumulate before the risk of thermal runaway is triggered. This theoretical charge is then compared with the remaining battery charging demand. Perform normalized difference calculation to obtain the charging completion rate. Its calculation method is as follows ,when When the value approaches zero or is positive, it indicates that the current charging strategy can complete the remaining charging needs within a safe window; when... When the value is negative, it indicates that the charging time window is insufficient to cover the remaining demand, and needs to be improved by adjusting the current parameters or extending the available window.

[0052] In the safety margin subspace, the power regulation parameters Deviation coefficient from trajectory Weighted fusion is performed to obtain the risk suppression degree. The weighted fusion method is as follows: ,in and These are the fusion weights for the power adjustment parameters and the trajectory deviation coefficient, respectively, and their sum is normalized to 1. A larger value indicates a higher charging power and a more significant risk of heat accumulation. A higher value indicates a more severe deviation of the thermal runaway evolution trajectory from the safety baseline. Risk suppression level. This comprehensively reflects the combined impact of power state and trajectory deviation on the safety margin. A smaller value indicates a more sufficient safety margin. Exceeding the minimum acceptable value of risk inhibition When this happens, the corresponding solution is determined to not satisfy the safety constraints.

[0053] According to acceleration index Weighting coefficients for determining charging completion , A higher value indicates a more significant acceleration of thermal runaway evolution. In this case, the weight of charging completion should be appropriately reduced to avoid excessively pursuing charging efficiency during the thermal runaway acceleration phase. The weighting coefficient of risk suppression is determined based on the corrected thermal runaway risk level. The higher the risk level, the better. The larger the value, the more likely it is to ensure that the safety margin constraint is met first under high-risk conditions. The two weighting coefficients satisfy... And through piecewise mapping functions respectively with A monotonic mapping relationship is established between the risk level and the correction level. The specific form of the mapping function can be obtained by calibration based on historical charging data.

[0054] Constructing a weighted scalarization objective function This function transforms a bi-objective optimization problem into a single-objective form, facilitating contour scanning within the solution space of the control parameters. The contour scanning process involves setting uniform step sizes within the effective region of the control parameter solution space. A set of contour levels is defined, and for each contour level, the trajectory of parameter combinations satisfying that level is calculated, forming a contour cluster. The contour cluster covers the complete trade-off interval from low risk and low charging completion to high charging completion and high risk. Each contour line corresponds to a set of candidate solutions with the same weighted objective value in the solution space.

[0055] Extract each contour line and the risk inhibition constraint boundary. The intersection points, under the premise of satisfying safety constraints, correspond to the optimal solutions at different charging completion levels, forming the initial Pareto front. Each intersection point in the initial Pareto front represents a feasible solution on the safety constraint boundary, with its charging completion level... It changes monotonically along the frontier direction, while the risk suppression degree Exactly equal to the minimum acceptable value This means that all initial Pareto front points are in a critical state of safety margin.

[0056] Local search optimization is performed on the initial Pareto front to improve the accuracy and uniformity of the front solution. For each intersection point, the charging completion evaluation function is used within its neighborhood. Iterative updates are performed on the gradient direction, and the gradient direction is updated by adjusting the gradient direction. Regarding the control parameter vector The partial derivatives are calculated, and an adaptive reduction strategy is adopted for the iteration step size. The initial step size is relatively large to quickly approach the local optimum, and it is gradually reduced as the number of iterations increases to ensure convergence accuracy. After each iteration update, it is checked whether the current iterative solution satisfies the given conditions. If the iterative solution exceeds the risk inhibition constraint boundary, it is backed up to the nearest feasible point on the constraint boundary, and the iteration process at that intersection point is terminated to prevent the search outside the safety constraints from generating infeasible solutions.

[0057] Summarize all solutions obtained after local search convergence at all intersection points, and remove solutions with dominance relationships, i.e., if a solution has a certain degree of charging completion... and risk inhibition If a solution is not inferior to the other solution in either of the two dimensions, the dominated solution is removed from the set. The final set of all convergent solutions constitutes the non-dominated solution set. Each solution in this solution set represents a Pareto optimal trade-off between charging completion and safety margin under the current charging state and thermal runaway risk conditions. (Non-dominated solution set) As a candidate library for subsequent optimal solutions, it is compared with the ideal point. The distance metric is used to select the final optimal solution vector. It is used to generate timed charging regulation commands.

[0058] Throughout the multi-objective optimization process, orthogonal decomposition ensures the decoupling of the two subspace objectives, weighted scaling and contour scanning ensure the initial coverage integrity of the Pareto front, and local search optimization ensures the local accuracy of the front solution. The three links work together to achieve an effective balance between computational efficiency and solution quality in the final non-dominated solution set, meeting the dual requirements of real-time performance and safety in the charging thermal runaway early warning scenario.

[0059] A second aspect of this invention provides an AI-based early warning system for thermal runaway during charging of an electric two-wheeled vehicle, comprising: The data acquisition unit is used to acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheeled vehicles, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data. The risk feature unit is used to perform cross-temporal and cross-spatial correlation mining on the multi-dimensional monitoring data based on the spatiotemporal attention mechanism to generate comprehensive risk features that characterize the evolution trend of battery thermal runaway. The early warning determination unit is used to generate an early warning determination result that includes the thermal runaway risk level and the early warning time node based on the comprehensive risk characteristics. The charging control unit is used to execute an adaptive multi-stage charging control strategy based on the warning judgment result, establish a differentiated control threshold library for different risk levels, and dynamically generate time-sequential charging adjustment instructions in combination with the warning time node and the remaining charging demand of the battery. The early warning push unit is used to establish a multi-level linkage early warning information push mechanism, which encapsulates the early warning judgment result and the time-sequential charging adjustment instruction into differentiated early warning content, and pushes it synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

[0060] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0061] 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.

[0062] 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.

[0063] 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 early warning method for thermal runaway during charging of electric two-wheeled vehicles, characterized in that, include: Acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheelers, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data; Based on the spatiotemporal attention mechanism, the correlation of the multi-dimensional monitoring data is mined across the time domain and across the spatial domain to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway. Based on the comprehensive risk characteristics, an early warning judgment result is generated, which includes the thermal runaway risk level and the early warning time node; An adaptive multi-stage charging control strategy is executed based on the warning judgment result. A differentiated control threshold library is established for different risk levels. Time-sequential charging adjustment instructions are dynamically generated in combination with the warning time node and the remaining battery charging demand. A multi-level linkage early warning information push mechanism is established, which encapsulates the early warning judgment results and the time-sequential charging adjustment instructions into differentiated early warning content, and pushes them synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

2. The method according to claim 1, characterized in that, Based on a spatiotemporal attention mechanism, the multi-dimensional monitoring data is subjected to cross-temporal and cross-spatial correlation mining to generate comprehensive risk features characterizing the evolution trend of battery thermal runaway, including: The multi-dimensional monitoring data is divided into constant current charging segment data and constant voltage charging segment data according to the charging stage. For each charging stage, a stage feature tensor containing the charging interface temperature data and the charging circuit electrical characteristic data is constructed. By analyzing the mapping relationship between the temperature distribution at different locations of the charging interface and the current density distribution at the corresponding locations, a spatial correlation diagram characterizing the heat conduction path and current loop topology is generated. A sliding window analysis is performed on the stage feature tensor to calculate the covariance matrix of the temperature rise rate and voltage fluctuation amplitude within each time window, and time-domain sensitivity weights are generated based on the eigenvalue distribution of the covariance matrix. Graph convolution is performed on the spatial correlation graph to capture the spatial propagation pattern of temperature anomalies spreading to the surrounding area by aggregating the feature information of adjacent nodes, and to generate spatial propagation weights. Based on the time-domain sensitivity weight and spatial propagation weight, the stage feature tensor is weighted and fused to obtain a fused feature vector. The temperature evolution component and electrical fluctuation component are extracted, and environmental compensation is performed in combination with the charging environment state data. The compensated temperature evolution component is converted into a thermal runaway trigger probability, and the compensated electrical fluctuation component is converted into an evolution velocity prediction value to obtain a comprehensive risk feature.

3. The method according to claim 2, characterized in that, By analyzing the mapping relationship between temperature distribution and current density distribution at different locations of the charging interface, a spatial correlation map characterizing the heat conduction path and current loop topology is generated, including: The contact area and conductor area of ​​the charging interface are spatially gridded. Temperature acquisition points are arranged at the center point of each grid cell, and current acquisition points are arranged at each conductor segment. An index table of acquisition point numbers and grid cell numbers is established. Temperature data of each grid cell and current data of each conductor segment are obtained. Current density data is calculated by combining the conductor cross-sectional area. The temperature data and current density data are associated with the corresponding grid cell according to the index table. A heat conduction equation based on Fourier's law and a heating equation based on Joule's law are established. The heat conduction equation describes the heat flow driven by the temperature gradient, and the heating equation uses the square of the current density to represent the volumetric heating rate. The theoretical temperature value is obtained by coupling the two equations and solving them on each grid element. By comparing the theoretical and measured temperature values ​​of each grid cell, grid cells with deviations exceeding the threshold are marked as risk cells. The spatial connectivity and series-parallel connections of the risk cells are statistically analyzed. The risk unit is used as a node to construct the graph. Heat conduction edges are added according to the spatial adjacency relationship of the grid unit, and current topology edges are added according to the series and parallel relationship of the current path. The weight of the heat conduction edge is calibrated by the temperature gradient, and the weight of the current topology edge is calibrated by the current distribution ratio, thus forming a spatial association graph.

4. The method according to claim 1, characterized in that, Based on the aforementioned comprehensive risk characteristics, a warning determination result is generated, including the thermal runaway risk level and warning time point, comprising: Extract the predicted values ​​of thermal runaway trigger probability and evolution rate from the comprehensive risk features, construct a probability-velocity two-dimensional feature space, and determine the initial thermal runaway risk level based on the coordinate position; Based on the initial thermal runaway risk level, the evolution trajectory template library is retrieved, and the temporal change trajectory of the thermal runaway trigger probability is matched with the standard trajectory to obtain the trajectory deviation coefficient. When the trajectory deviation coefficient exceeds the safety threshold, the corrected thermal runaway risk level is obtained. The evolution trend correction factor is determined based on the trajectory deviation coefficient. The predicted evolution rate is multiplied by the evolution trend correction factor to obtain the corrected evolution rate. The evolution time required to evolve from the current trigger probability to the critical trigger probability is calculated. The predicted evolution rate is analyzed in the time domain and the rate of change is calculated as an acceleration index. The time reduction is calculated based on the acceleration index. The time reduction is subtracted from the evolution duration to obtain the corrected evolution duration. This corrected evolution duration is added to the current time to determine the warning time node. The corrected thermal runaway risk level, warning time node, trajectory deviation coefficient, and acceleration index are encapsulated to form the warning judgment result.

5. The method according to claim 1, characterized in that, Based on the warning judgment result, an adaptive multi-stage charging control strategy is executed. A differentiated control threshold library is established for different risk levels. Combined with the warning time node and the remaining battery charging demand, a time-sequential charging adjustment command is dynamically generated, including: The differentiated control threshold library is divided into a prevention layer threshold set and a response layer threshold set. The corresponding threshold set is selected according to the corrected thermal runaway risk level. The trajectory deviation coefficient is used as the layer penetration factor. When the layer penetration factor exceeds the boundary value, the threshold sets are fused to generate a hybrid threshold set. Establish a constraint propagation network, setting the remaining battery charging demand as the target node, the early warning time node as the time constraint node, and the mixed threshold set as the intermediate constraint node. In the constraint propagation network, energy demand is propagated backward from the target node, time margin is propagated forward from the time constraint node, and the solution space of control parameters is generated by converging at the intermediate constraint nodes. Multi-objective optimization is performed on the solution space of the control parameters, with charging completion degree as the main objective and risk suppression degree as the constraint objective. The non-dominated solution set is obtained through Pareto front search. The optimal solution vector is extracted from the non-dominated solution set and decomposed into discrete control actions. Spline interpolation is performed on the discrete control actions to generate a continuous control instruction sequence, which is then encapsulated as a time-series charging control instruction.

6. The method according to claim 1, characterized in that, Multi-objective optimization is performed on the solution space of the control parameters, with charging completion rate as the primary objective and risk suppression rate as the constraint objective. The non-dominated solution set is obtained through Pareto front search, including: Orthogonal decomposition is performed on the solution space of the control parameters, and the solution space of the control parameters is projected to the charging efficiency subspace and the safety margin subspace. In the charging efficiency subspace, the product of the current adjustment parameter and the warning time node is calculated as the available charging time window. The normalized difference between the theoretical charging amount and the remaining charging demand of the battery within the available charging time window is used as the charging completion degree. In the safety margin subspace, the power adjustment parameter and the trajectory deviation coefficient are weighted and fused as the risk suppression degree. The weighting coefficients for charging completion are determined based on the acceleration index, and the weighting coefficients for risk suppression are determined based on the corrected thermal runaway risk level. A weighted scalar objective function is constructed, and contour scanning is performed on the weighted scalar objective function to generate contour clusters in the solution space of the control parameters. The intersections of each contour line with the risk suppression constraint boundary are extracted to form the initial Pareto front. Local search optimization is performed on the initial Pareto front. Iterative updates are performed in the neighborhood of each intersection along the gradient direction of the charging completion evaluation function. When the iterative solution exceeds the risk suppression constraint boundary, it backs to the boundary and terminates the iteration. All converged solutions form a non-dominated solution set.

7. An AI-based early warning system for thermal runaway during charging of an electric two-wheeled vehicle, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data acquisition unit is used to acquire multi-dimensional monitoring data collected in real time by the charging pile sensor network during the charging process of electric two-wheeled vehicles, including charging interface temperature data, charging circuit electrical characteristic data, and charging environment status data. The risk feature unit is used to perform cross-temporal and cross-spatial correlation mining on the multi-dimensional monitoring data based on the spatiotemporal attention mechanism to generate comprehensive risk features that characterize the evolution trend of battery thermal runaway. The early warning determination unit is used to generate an early warning determination result that includes the thermal runaway risk level and the early warning time node based on the comprehensive risk characteristics. The charging control unit is used to execute an adaptive multi-stage charging control strategy based on the warning judgment result, establish a differentiated control threshold library for different risk levels, and dynamically generate time-sequential charging adjustment instructions in combination with the warning time node and the remaining charging demand of the battery. The early warning push unit is used to establish a multi-level linkage early warning information push mechanism, which encapsulates the early warning judgment result and the time-sequential charging adjustment instruction into differentiated early warning content, and pushes it synchronously to the local display unit of the charging pile, the user's mobile terminal and the monitoring center management platform through multiple channels.

8. 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 6.

9. 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 6.

Images