New energy battery safety monitoring method and system based on artificial intelligence

Abstract

The application provides a new energy battery safety monitoring method and system based on artificial intelligence, and relates to the technical field of artificial intelligence. First, the multi-source physical field response data stream of the internal sensing array of the battery pack is acquired. Then, the cross-field state correlation analysis processing is performed to construct the coupled physical field collaborative evolution structure. The coupled physical field collaborative evolution structure is processed by calling the time sequence causal convolution twin network to extract features, locate the causal conduction chain and generate the hidden space trajectory manifold. Based on the projection condition of the hidden space trajectory manifold in the high-dimensional curvature boundary, the thermal runaway risk evolution stage identifier and the abnormal physical field space positioning area are determined. Finally, the differentiated protection instruction set is generated to adjust the charging and discharging power and start the directional thermal management loop. The application can comprehensively and accurately monitor the battery safety state and effectively protect.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a new energy battery safety monitoring method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of the new energy industry, new energy batteries have been widely used in electric vehicles, energy storage systems, and other fields. However, during the charging and discharging process, new energy batteries may experience a series of safety issues due to complex internal electrochemical reactions and physical changes. Among these, thermal runaway is the most serious. Once thermal runaway occurs, it can lead to serious consequences such as battery fire and explosion, posing a huge threat to the safety of people and property.

[0003] Currently, traditional battery safety monitoring methods mainly rely on monitoring a single physical quantity, such as only monitoring the battery's temperature or voltage. These methods have significant limitations because internal battery faults and anomalies are often not caused by a single factor, but rather are the result of the interaction and co-evolution of multiple physical fields. A single monitoring method cannot comprehensively and accurately reflect the actual internal state of the battery, making it difficult to detect potential safety hazards in a timely manner. It also falls short in early warning and precise location of thermal runaway, failing to meet the high requirements for safety monitoring of new energy batteries. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an artificial intelligence-based method for monitoring the safety of new energy batteries, the method comprising: Acquire a multi-source physical field response data stream deployed in a sensor array inside the battery pack. The multi-source physical field response data stream includes a thermodynamic distribution sequence, an acoustic response sequence, and an impedance spectrum evolution sequence with synchronous time stamps. Perform cross-field state correlation analysis processing on the multi-source physical field response data stream to construct a coupled physical field cooperative evolution structure; The temporal causal convolutional Siamese network is invoked to process the coupled physical field co-evolution structure, extract the long-term non-stationary evolution characteristics of the coupled physical field co-evolution structure, locate the corresponding causal transmission chain, and generate the latent space trajectory manifold characterizing the evolution trend of the safe state. Based on the projection offset direction and curvature abrupt change amplitude of the latent space trajectory manifold within the preset high-dimensional curvature boundary, the identification of the battery pack thermal runaway risk evolution stage and the associated abnormal physical field spatial location area are determined. A differentiated protection instruction set is generated based on the thermal runaway risk evolution stage identifier and the abnormal physical field spatial positioning area. The differentiated protection instruction set is used to adjust the charge and discharge power boundary and start the directional thermal management loop corresponding to the abnormal physical field spatial positioning area.

[0005] Furthermore, embodiments of the present invention also provide an artificial intelligence-based new energy battery safety monitoring system, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned artificial intelligence-based new energy battery safety monitoring method by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the artificial intelligence-based new energy battery safety monitoring system reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the artificial intelligence-based new energy battery safety monitoring system to execute the aforementioned artificial intelligence-based new energy battery safety monitoring method.

[0007] Based on the above, firstly, the multi-source physical field response data stream deployed in the sensor array inside the battery pack is acquired, covering the thermodynamic distribution sequence, acoustic response sequence and impedance spectrum evolution sequence with synchronous time stamps. Cross-field state correlation analysis processing is performed on the multi-source physical field response data stream to construct a coupled physical field co-evolution structure, and the intrinsic connection and synergistic effect between different physical fields are explored, which can more accurately grasp the complex change process inside the battery. By invoking a temporal causal convolutional twin network to process the coupled physical field co-evolution structure, long-term non-stationary evolution features are extracted and causal transmission chains are located. This generates a latent space trajectory manifold characterizing the evolution trend of the safe state. Based on the projection offset direction and curvature abrupt change amplitude of the latent space trajectory manifold within a preset high-dimensional curvature boundary, the stage identifiers of the battery pack's thermal runaway risk evolution and the associated abnormal physical field spatial positioning areas are determined. This achieves accurate classification of thermal runaway risk and precise positioning of abnormal regions. Finally, based on the thermal runaway risk evolution stage identifiers and abnormal physical field spatial positioning areas, a differentiated protection instruction set is generated to adjust the charge and discharge power boundaries and activate the corresponding directional thermal management loops. This enables targeted protective measures to be taken according to the actual safety state of the battery, effectively preventing the occurrence of thermal runaway and ensuring the safe operation of the battery. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the artificial intelligence-based new energy battery safety monitoring method provided in the embodiments of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of an artificial intelligence-based new energy battery safety monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0010] Figure 1 This is a flowchart illustrating an artificial intelligence-based safety monitoring method for new energy batteries according to an embodiment of the present invention, which will be described in detail below.

[0011] Step S110: Obtain a multi-source physical field response data stream deployed in the sensor array inside the battery pack. The multi-source physical field response data stream includes a thermodynamic distribution sequence, an acoustic response sequence, and an impedance spectrum evolution sequence with synchronous time stamps.

[0012] In a safety monitoring scenario applied to an electric vehicle power battery system, the battery pack consists of multiple series-connected individual cells, each with a distributed sensor array deployed inside and on its surface. This multi-source physical field response data stream includes a thermodynamic distribution sequence, an acoustic response sequence, and an impedance spectrum evolution sequence with synchronized time stamps. The thermodynamic distribution sequence is acquired by a thermocouple array deployed on the surface of the battery electrodes and at the electrode connections, recording the temperature values ​​of all distributed sensor nodes at each sampling moment. The acoustic response sequence is acquired by a piezoelectric ceramic sensor array attached to the surface of the battery casing, with each piezoelectric ceramic sensor recording the time-domain waveform data of guided wave propagation within a fixed sampling window. The impedance spectrum evolution sequence is acquired by an electrochemical impedance spectroscopy monitoring module, which applies a small AC excitation at multiple discrete frequency points and measures the response signal, calculating the real and imaginary parts of the impedance at each discrete frequency point. All thermodynamic distribution sequences, acoustic response sequences, and impedance spectrum evolution sequences are accompanied by high-precision synchronized time stamps provided by a Global Positioning System (GPS) clock or a Precision Time Protocol (PTP) clock.

[0013] After step S110, the method may further include: Step S1110: Perform time-series segmentation processing on the multi-source physical field response data stream, and divide it into multiple thermodynamic time-series segment units, multiple acoustic time-series segment units, and multiple impedance spectrum time-series segment units with the same time start and end boundaries according to the preset sliding time window length.

[0014] After acquiring the multi-source physical field response data stream deployed in the sensor array inside the battery pack in step S110, step S1110 performs time-series segmentation processing on the multi-source physical field response data stream. A preset sliding time window length is set, which covers multiple complete thermal distribution sampling periods, acoustic response sampling periods, and impedance spectrum scanning periods. According to the preset sliding time window length, the thermal distribution sequence is divided into multiple thermal time-series segment units with the same time start and end boundaries. Each thermal time-series segment unit contains the temperature values ​​of all distributed sensor nodes at all sampling times within the time window. At the same time, the acoustic response sequence is divided into multiple acoustic time-series segment units with the same time start and end boundaries. Each acoustic time-series segment unit contains the time-domain waveform data collected by all piezoelectric ceramic sensors within the time window. At the same time, the impedance spectrum evolution sequence is divided into multiple impedance spectrum time-series segment units with the same time start and end boundaries. Each impedance spectrum time-series segment unit contains the real part and imaginary part of the impedance corresponding to all frequency points within the time window. The overlap rate between adjacent time segment units is controlled by the sliding step size, which is less than the preset sliding time window length.

[0015] Step S1111: Perform spatial graph structure construction processing on the thermal time series segment units within the same time window, calculate the correlation coefficient value between the temperature time series of any two spatial sensing nodes in the thermal time series segment unit, and construct a thermal spatial correlation graph structure with sensing nodes as vertices and correlation coefficient values ​​as edge weights based on the correlation coefficient value.

[0016] For each time window corresponding to a thermodynamic time series segment, obtain the time sequence of temperature values ​​of all distributed sensing nodes within that time window. For any two distributed sensing nodes i and j, extract their respective temperature time series sequences Ti(t) and Tj(t). Calculate the Pearson correlation coefficient between these two temperature time series sequences. The formula for calculating the Pearson correlation coefficient is r_ij=Σ[(Ti(t)-μ_i)*(Tj(t)-μ_j)] / sqrt(Σ[(Ti(t)-μ_i)^2]*Σ[(Tj(t)-μ_j)^2]), where μ_i and μ_j are the means of the two temperature time series sequences, respectively. Treat all distributed sensing nodes as vertices of a graph structure. For each pair of vertices i and j, use the calculated correlation coefficient value r_ij as the weight of the edge connecting vertices i and j. If the correlation coefficient between two vertices is less than a preset correlation threshold, no edge connection is established. The final constructed thermal spatial relational graph structure includes a set of vertices, a set of edges, and a set of edge weights.

[0017] Step S1112: Perform spatial graph structure construction processing on the acoustic wave timing segment units within the same time window, calculate the correlation coefficient between the guided wave response timing of any two spatial sensing nodes in the acoustic wave timing segment unit, and construct an acoustic wave spatial correlation graph structure with sensing nodes as vertices and correlation coefficient values ​​as edge weights based on the correlation coefficient values.

[0018] For each time window corresponding to the acoustic wave timing segment unit, the guided wave response timing data collected by all piezoelectric ceramic sensors within that time window are acquired. Multiple characteristic parameters can be extracted from the guided wave response timing data of each piezoelectric ceramic sensor, including guided wave travel time offset, guided wave energy attenuation, and amplitude of each guided wave mode. For any two piezoelectric ceramic sensors p and q, their respective guided wave energy attenuation timing sequences Ap(t) and Aq(t) are extracted. The Pearson correlation coefficient between these two timing sequences is calculated, using the same formula as in step S1111. All piezoelectric ceramic sensors are used as vertices of the graph structure. For each pair of vertices p and q, the calculated correlation coefficient is used as the weight of the edge connecting vertices p and q. If the correlation coefficient between two vertices is less than a preset correlation threshold, no edge connection is established. The final constructed acoustic wave spatial correlation graph structure includes a vertex set, an edge set, and an edge weight set.

[0019] Step S1113: Perform frequency graph structure construction processing on the impedance spectrum time sequence segment unit within the same time window, calculate the correlation coefficient between the impedance imaginary part time sequence of any two frequency sampling points in the impedance spectrum time sequence segment unit, and construct a frequency correlation graph structure with frequency points as vertices and correlation coefficient values ​​as edge weights based on the correlation coefficient values.

[0020] For each impedance spectrum time-series segment unit corresponding to a time window, the sequence of impedance imaginary part values ​​changing with time for all frequency points within that time window is obtained. The impedance spectrum evolution sequence contains multiple discrete frequency points f1, f2, f3, ..., fM on the frequency axis. For any two frequency points fu and fv, their respective impedance imaginary part time-series sequences Zim_u(t) and Zim_v(t) are extracted. The Pearson correlation coefficient between these two time-series sequences is calculated, and the formula for calculating the Pearson correlation coefficient is the same as in step S1111. All frequency points are used as vertices of the graph structure. For each pair of vertices fu and fv, the calculated correlation coefficient value is used as the weight value of the edge connecting vertices fu and fv. If the correlation coefficient value between two frequency points is less than a preset correlation threshold, no edge connection is established. The final constructed frequency correlation graph structure contains a vertex set, an edge set, and an edge weight set. This frequency correlation graph structure reflects the correlation pattern between impedance responses at different frequency points, and different frequency points correspond to electrochemical processes at different time scales within the battery.

[0021] Step S1114: Combine the thermal spatial correlation graph structure, the acoustic spatial correlation graph structure, and the frequency correlation graph structure within the same time window into a multi-view correlation graph network. The multi-view correlation graph network includes a thermal view adjacency matrix, an acoustic view adjacency matrix, a frequency view adjacency matrix, and a set of cross-view connection edge weights connecting different view nodes.

[0022] For the same time window, the thermal spatial correlation graph structure constructed in step S1111, the acoustic spatial correlation graph structure constructed in step S1112, and the frequency correlation graph structure constructed in step S1113 are combined into a multi-view correlation graph network. Each sensor node in the thermal spatial correlation graph structure corresponds to a vertex in the thermal view; each piezoelectric ceramic sensor node in the acoustic spatial correlation graph structure corresponds to a vertex in the acoustic view; and each frequency point in the frequency correlation graph structure corresponds to a vertex in the frequency view. The thermal view adjacency matrix is ​​an N×N matrix, where N is the number of distributed sensor nodes, and the elements in the matrix are the correlation coefficient values ​​calculated in step S1111. The acoustic view adjacency matrix is ​​a P×P matrix, where P is the number of piezoelectric ceramic sensors, and the elements in the matrix are the correlation coefficient values ​​calculated in step S1112. The frequency view adjacency matrix is ​​an M×M matrix, where M is the number of frequency points, and the elements in the matrix are the correlation coefficient values ​​calculated in step S1113. The cross-view connection edge weight set includes the connection edge weights between the thermal view vertex and the acoustic view vertex, the connection edge weights between the thermal view vertex and the frequency view vertex, and the connection edge weights between the acoustic view vertex and the frequency view vertex. The above cross-view connection edge weights are preset according to prior physical knowledge. For example, the connection edge weight between thermocouples and piezoelectric ceramic sensors located in the same or adjacent spatial positions is set to a higher value.

[0023] Step S1115: Call the preset multi-view graph neural network to perform multi-round graph convolution message passing and cross-view feature fusion operations on the multi-view association graph network. The features of the heat view node are propagated along the heat view adjacency matrix and weighted and fused with the features of the sound wave view node and the frequency view node through the cross-view connection edge weight set to generate the heat view node update feature vector, the sound wave view node update feature vector and the frequency view node update feature vector.

[0024] A multi-view graph neural network is pre-defined, containing multiple graph convolutional layers and cross-view fusion layers. During initialization, an initial feature vector is assigned to each vertex in the thermal view, containing a time-series segment of the temperature value of the corresponding sensor node. An initial feature vector is assigned to each vertex in the acoustic view, containing the statistical characteristics of the guided wave response time-series segment of the piezoelectric ceramic sensor corresponding to that vertex. An initial feature vector is assigned to each vertex in the frequency view, containing a time-series segment of the imaginary part of the impedance at the corresponding frequency point. In the first round of graph convolutional message passing, for the thermal view, each vertex collects the feature vectors of its neighboring vertices along the thermal view's adjacency matrix, performs a weighted sum with its own feature vector, and obtains the intermediate feature vector of the first round of the thermal view after transformation by a nonlinear activation function. The same operation is performed for the acoustic and frequency views. In the cross-view feature fusion, each thermal view vertex collects the feature vectors of its connected acoustic and frequency view vertices, multiplies them by the corresponding cross-view connection edge weights, and then adds them to its own feature vector. After multiple rounds of graph convolution message passing and cross-view feature fusion, feature vectors for updating thermal view nodes, acoustic view nodes, and frequency view nodes are generated.

[0025] Step S1116: Concatenate the updated feature vectors of all thermal view nodes, all acoustic view nodes, and all frequency view nodes along the vector dimension to generate a multi-view fusion graph representation vector sequence, and replace the thermal distribution sequence, the acoustic response sequence, and the impedance spectrum evolution sequence with the multi-view fusion graph representation vector sequence.

[0026] The updated feature vectors of all thermal view nodes generated in step S1115 are arranged in node index order to form a two-dimensional matrix. The number of rows in the matrix corresponds to the number of thermal view vertices, and the number of columns corresponds to the feature vector dimension. Similarly, the updated feature vectors of all acoustic view nodes are arranged into another two-dimensional matrix, and the updated feature vectors of all frequency view nodes are arranged into a third two-dimensional matrix. These three two-dimensional matrices are horizontally concatenated along the feature vector dimension to obtain a larger two-dimensional matrix. Each row of this matrix corresponds to a node of a physical view, and each column corresponds to a fused feature dimension. Each row of this two-dimensional matrix is ​​used as a multi-view fusion graph representation vector. All rows of multi-view fusion graph representation vectors are arranged in time window order to form a multi-view fusion graph representation vector sequence. In subsequent steps S120 to S150, this multi-view fusion graph representation vector sequence is used to replace the original thermal distribution sequence, acoustic response sequence, and impedance spectrum evolution sequence as input data. That is, the thermal distribution sequence, acoustic response sequence, and impedance spectrum evolution sequence in steps S121 to S129 are all replaced with the corresponding multi-view fusion graph representation vector sequence.

[0027] Step S120: Perform cross-field state correlation analysis processing on the multi-source physical field response data stream to construct a coupled physical field co-evolution structure.

[0028] Step S121: Perform spatial interpolation reconstruction operation on the time sequence of temperature values ​​of each sensing node in the thermal distribution sequence to generate a continuous temperature field distribution function covering the geometric entity of the battery pack. Based on the continuous temperature field distribution function, calculate the tangential temperature gradient field along the battery electrode stacking direction and the normal temperature gradient field perpendicular to the battery electrode stacking direction.

[0029] Spatial interpolation reconstruction is performed on the temperature values ​​at each sampling time in the thermal distribution sequence. Using the finite element mesh nodes of the battery pack's 3D geometric model as interpolation target points, the radial basis function (RBF) interpolation method is employed. The spatial coordinates (xi, yi, zi) and temperature values ​​ti of each distributed sensing node are used as known scattered data to construct a continuous temperature field distribution function T(x, y, z) covering the entire battery pack's geometry. In the RBF interpolation method, the temperature value at any target point (x, y, z) is represented as a weighted sum of the temperature values ​​at each known scattered point and the radial basis function value. The radial basis function adopts a multiple quadratic function form φ(r) = sqrt(r^2 + c^2), where r is the Euclidean distance between the target point and the known scattered points, and c is a smoothing parameter. The directional derivatives along the battery electrode stacking direction z of this continuous temperature field distribution function are calculated to obtain the tangential temperature gradient field Gt = (dT / dx, dT / dy). The magnitude of the tangential temperature gradient field Gt reflects the intensity of lateral heat conduction within the battery electrode plane. Simultaneously, the directional derivative perpendicular to the battery electrode stacking direction is calculated to obtain the normal temperature gradient field Gn = dT / dz. The magnitude of the normal temperature gradient field Gn reflects the intensity of longitudinal heat conduction between the battery electrode layers.

[0030] Step S122: Extract the core line set of heat flow vortex caused by non-uniform heat generation in the battery pack surface according to the curl distribution law of the tangential temperature gradient field, and extract the profile line set of thermal barrier interface caused by the difference in thermal conduction between battery pack layers according to the divergence distribution law of the normal temperature gradient field. Combine the core line set of heat flow vortex with the profile line set of thermal barrier interface in spatial cross-linking to generate a heat flow transfer topology map inside the battery pack with node connectivity strength as edge weight.

[0031] A curl operation is performed on the tangential temperature gradient field Gt = (dT / dx, dT / dy), with the formula curl(Gt) = d(Gt_y) / dx - d(Gt_x) / dy. Regions with non-zero curl values ​​indicate the existence of closed-loop heat flow paths, corresponding to heat flow vortex structures caused by non-uniform heat generation within the battery pack surface. Local maxima of the curl values ​​are extracted and connected by spatial adjacency to form a set of heat flow vortex core lines. A divergence operation is performed on the normal temperature gradient field Gn = dT / dz, with the formula div(Gn) = dGn / dz. Interface locations where divergence values ​​change drastically indicate heat accumulation or divergence, corresponding to thermally barrier interfaces caused by differences in interlayer thermal conduction within the battery pack. Zero-crossing points of divergence values ​​are extracted and connected into closed curves according to spatial adjacency, forming a set of thermal barrier interface contour lines. The set of heat flux vortex core lines and the set of thermal barrier interface contour lines are projected onto the same two-dimensional coordinate system, and the intersection points of each heat flux vortex core line and each thermal barrier interface contour line are calculated. These intersection points are used as nodes in the topology graph, and the curve segments between adjacent intersection points are used as edges. The node connectivity strength of each edge is calculated. The formula for calculating node connectivity strength is the product of the average heat flux density and the temperature gradient within the region spanned by the edge. All nodes and edges constitute the heat flow transfer topology graph inside the battery pack.

[0032] Step S123: Perform multi-scale time-frequency domain decomposition on the acoustic response sequence, extract the group velocity dispersion curve family and phase velocity decay curve family corresponding to narrowband components of different frequencies, establish the acoustic waveguide propagation equation of the battery casing structure based on the group velocity dispersion curve family and the phase velocity decay curve family, and analyze the spatial distribution map of energy attenuation coefficients corresponding to each order of guided wave modes from the acoustic waveguide propagation equation.

[0033] Multi-scale time-frequency domain decomposition was performed on the time-domain waveform data acquired by each piezoelectric ceramic sensor in the acoustic response sequence. Using the continuous wavelet transform method with the Morlet wavelet as the mother wavelet function, the time-domain waveform signal was decomposed into wavelet coefficient matrices corresponding to narrowband components with different center frequencies. For each narrowband component, its group velocity and phase velocity values ​​in the time-frequency domain were extracted. The group velocity value was obtained by calculating the ratio of the peak time offset of the wavelet coefficient envelope to the propagation distance, and the phase velocity value was obtained by calculating the ratio of the phase change of the wavelet coefficient to the propagation distance. The group velocity values ​​of all narrowband components were arranged in frequency order to obtain a family of group velocity dispersion curves. The phase velocity values ​​of all narrowband components were also arranged in frequency order to obtain a family of phase velocity decay curves. Based on the family of group velocity dispersion curves and the family of phase velocity decay curves, the acoustic waveguide propagation equation for the battery casing structure was established. This acoustic waveguide propagation equation is an eigenvalue problem of the frequency domain wave equation in the direction of casing thickness. Solving this acoustic waveguide propagation equation yielded the wavenumber eigenvalues ​​corresponding to each order of guided wave modes. The energy attenuation coefficient is calculated based on the imaginary part of the wavenumber eigenvalue. The energy attenuation coefficients of all spatial locations are arranged by coordinates to generate a spatial distribution map of the energy attenuation coefficients corresponding to each guided wave mode.

[0034] Step S124: Based on the steep change in the energy attenuation coefficient of the guided wave mode between adjacent sensing points in the spatial distribution map of the energy attenuation coefficient, locate the micro-damage sensitive area of ​​the shell structure. Based on the travel time offset of the reflected echo of each guided wave mode at the micro-damage sensitive area, construct a microstructure response path topology map with the sensing node as the vertex and the travel time offset as the edge vector. Use the microstructure response path topology map as the microstructure path feature of the battery shell.

[0035] For each pair of adjacent piezoelectric ceramic sensor points in the spatial distribution map of energy attenuation coefficient, the absolute value of the difference in energy attenuation coefficient between the two points is calculated, and this absolute value is used as a measure of steepness. When the steepness measure exceeds a preset steepness threshold, the area between these two adjacent piezoelectric ceramic sensor points is marked as a micro-damage sensitive area. At the marked micro-damage sensitive area, the reflected echo signals of each guided wave mode are extracted, and the travel time offset of the reflected echo signal relative to the excitation signal is calculated. The travel time offset is calculated by subtracting the peak time of the excitation wave packet from the peak time of the reflected echo wave packet. Using each piezoelectric ceramic sensor as a vertex of the topology graph, and using the travel time offset between each pair of sensors as the edge vector value of the directed edge from the transmitting sensor vertex to the receiving sensor vertex, a microstructure response path topology graph is constructed. This microstructure response path topology graph is the microstructure path feature of the battery casing.

[0036] Step S125: Perform a deconvolution analysis operation on the impedance spectrum evolution sequence to transform the broadband impedance imaginary part response to the relaxation time constant domain, generate a relaxation time distribution function containing a continuous relaxation peak distribution, and identify the high-frequency relaxation peak movement traces corresponding to the charge transfer process and the mid-frequency relaxation peak morphology distortion trend corresponding to the solid-phase diffusion process in the relaxation time distribution function.

[0037] An analytical deconvolution operation on the relaxation time distribution is performed on the broadband impedance data at each sampling time in the impedance spectrum evolution sequence. The broadband impedance data consists of a set of data points where the real part Zre and the imaginary part Zim of the impedance vary with frequency f. The analytical deconvolution operation on the relaxation time distribution first constructs the integral equation Zim(f)=∫[g(τ)*(2πfτ) / (1+(2πfτ)^2)]dlnτ, where τ is the relaxation time constant and g(τ) is the relaxation time distribution function to be solved. The Tikhonov regularization method is used to solve this integral equation, transforming the imaginary part response of the broadband impedance to the relaxation time constant domain, generating a relaxation time distribution function g(τ) containing a continuous relaxation peak distribution. In this relaxation time distribution function, the peak positions located in short relaxation time intervals are identified. These peak positions correspond to the high-frequency relaxation peaks of the charge transfer process. The high-frequency relaxation peaks at all sampling times are connected in chronological order to obtain the high-frequency relaxation peak movement trace. Identify the peak shape located in the mid-relaxation time interval. This peak shape corresponds to the mid-frequency relaxation peak position in the solid-phase diffusion process. Calculate the full width at half maximum (FWHM) and peak height of the mid-frequency relaxation peak position at each sampling time. Arrange the FWHM and peak height values ​​of all sampling times in chronological order to obtain the distortion trend of the mid-frequency relaxation peak position shape.

[0038] Step S126: Construct a charge transfer activated state drift vector along the time evolution direction based on the high-frequency relaxation peak position movement trace, construct a solid-phase diffusion path tortuosity change along the frequency evolution direction based on the mid-frequency relaxation peak position morphological distortion trend, and fuse the charge transfer activated state drift vector and the solid-phase diffusion path tortuosity change to generate an ion transport interface state descriptor.

[0039] The high-frequency relaxation peak relaxation time values ​​are extracted from the high-frequency relaxation peak movement trajectory at each sampling time. The difference between the high-frequency relaxation peak relaxation time values ​​between adjacent sampling times is calculated, and all differences are arranged in chronological order to form a charge transfer activated state drift vector. Each element in this charge transfer activated state drift vector represents the change in the activation energy barrier of the charge transfer process per unit time. The half-width at half-maximum (WHM) values ​​of the mid-frequency relaxation peak are extracted from the mid-frequency relaxation peak morphology distortion trend at each sampling time. The rate of change of the WHM values ​​between adjacent sampling times is calculated, and all rates of change are arranged in chronological order to form a solid-phase diffusion path tortuosity change. Each element in this solid-phase diffusion path tortuosity change represents the change in the tortuosity of the lithium-ion diffusion path in the solid phase per unit time. The charge transfer activated state drift vector and the solid-phase diffusion path tortuosity change are aligned with the same time index, resulting in a two-dimensional vector at each time index. The first component of this two-dimensional vector is the charge transfer activated state drift, and the second component is the solid-phase diffusion path tortuosity change. The two-dimensional vectors at all time indices constitute the state descriptor of the ion transport interface.

[0040] Step S127: Map the node coordinates of the heat flow transfer topology to a first spatiotemporal projection plane with time as the horizontal axis and battery spatial coordinates as the vertical axis; map the node coordinates of the microstructure response path topology to a second spatiotemporal projection plane with time as the horizontal axis and the arc length of the shell surface as the vertical axis; and map the feature points of the ion transport interface state descriptor to a third spatiotemporal projection plane with time as the horizontal axis and the amplitude of the imaginary part of impedance as the vertical axis.

[0041] The spatial coordinates of each node in the heat flow transfer topology are extracted. The x and y coordinates of these spatial coordinates are mapped to the horizontal axis of the first spatiotemporal projection plane, and the z coordinate is mapped to the vertical axis of the first spatiotemporal projection plane. Simultaneously, the sampling time corresponding to each node is used as the depth dimension coordinate of the first spatiotemporal projection plane, forming a three-dimensional spatiotemporal point set. The position of the piezoelectric ceramic sensor corresponding to each vertex in the microstructure response path topology is extracted on the surface of the battery casing. The arc length distance along the casing surface from a fixed reference point is calculated, and this arc length distance is used as the vertical axis coordinate of the second spatiotemporal projection plane. The sampling time corresponding to each vertex is used as the horizontal axis coordinate of the second spatiotemporal projection plane, forming a two-dimensional spatiotemporal point set. The sampling time of each feature point in the ion transport interface state descriptor is extracted as the horizontal axis coordinate of the third spatiotemporal projection plane, and the imaginary part of the impedance corresponding to that feature point is extracted as the vertical axis coordinate of the third spatiotemporal projection plane, forming a two-dimensional spatiotemporal point set.

[0042] Step S128: Based on the time dimension alignment reference between the first spatiotemporal projection plane, the second spatiotemporal projection plane, and the third spatiotemporal projection plane, calculate the peak offset of the autocorrelation function within the sliding time window for the edge weight change sequence of the heat flow transfer topology graph, the edge vector evolution sequence of the microstructure response path topology graph, and the amplitude fluctuation sequence of the ion transport interface state descriptor, respectively, to generate a characteristic time delay estimation array within each physical field; based on the characteristic time delay estimation array within each physical field, calculate the alignment offset on the time axis between the characteristic sequences of different physical fields to generate a cross-physical field time alignment offset array.

[0043] The synchronization time stamp attached to the multi-source physical field response data stream in step S110 is used as the time dimension alignment reference. For the heat flow transfer topology, the sequence of edge weight values ​​changing with sampling time for all edges is extracted. A sliding time window method is applied to this sequence, and the autocorrelation function of the sequence is calculated within each time window. The formula for calculating the autocorrelation function is R(k)=Σ[(w(t)-μ)*(w(t+k)-μ)] / Σ[(w(t)-μ)^2], where w(t) is the edge weight value, μ is the mean of the edge weight values ​​within the window, and k is the time lag. The k value corresponding to the peak value of the autocorrelation function within each window is found, and this k value is the characteristic time delay estimate. The characteristic time delay estimates of all windows are arranged in chronological order to obtain the characteristic time delay estimate array inside the heat flow physical field. Similarly, the same autocorrelation function peak shift calculation is performed on the edge vector magnitude evolution sequence of the microstructure response path topology graph and the amplitude fluctuation sequence of the ion transport interface state descriptor to obtain the characteristic time delay estimation arrays within the acoustic and electrochemical physical fields, respectively. The characteristic time delay estimation arrays of the three physical fields are then compared pairwise to calculate the alignment offset on the time axis between the characteristic sequences of different physical fields. The formula for calculating the alignment offset is ΔT_ab = T_a - T_b, where T_a and T_b are the corresponding elements in the characteristic time delay estimation arrays of the two physical fields, respectively. All alignment offsets constitute a cross-physical field time alignment offset array.

[0044] Step S129: Based on the cross-physics field time alignment offset array, perform time axis alignment correction on the edge weight change sequence of the heat flow transfer topology graph, the edge vector evolution sequence of the microstructure response path topology graph, and the amplitude fluctuation sequence of the ion transport interface state descriptor. Then, perform tensor splicing and normalization fusion operations on the time-aligned edge weight change sequence, edge vector evolution sequence, and amplitude fluctuation sequence under a unified spatiotemporal framework to generate a coupled physics field co-evolution structure indexed by three-dimensional spatial coordinates and one-dimensional time coordinates.

[0045] Based on the alignment offset values ​​in the cross-physics-field time alignment offset array, time-axis translation correction is performed on the edge weight change sequence of the heat flow transfer topology graph, the edge vector evolution sequence of the microstructure response path topology graph, and the amplitude fluctuation sequence of the ion transport interface state descriptor. For each sequence, the corresponding alignment offset value is subtracted from its time index to align the features representing the same physical event in the three sequences on the time axis. The time-aligned corrected heat flow transfer topology graph edge weight change sequence is organized into a three-dimensional tensor, with the three dimensions corresponding to the spatial x-coordinate, spatial y-coordinate, and time index, respectively. The time-aligned corrected microstructure response path topology graph edge vector evolution sequence is also organized into a three-dimensional tensor, with the three dimensions corresponding to the shell surface arc length coordinate, the waveguide mode index, and the time index, respectively. The time-aligned corrected ion transport interface state descriptor amplitude fluctuation sequence is organized into a two-dimensional matrix, with the two dimensions corresponding to the frequency point and the time index, respectively. The above three tensors are concatenated along the feature dimensions to form a higher-dimensional joint tensor. Perform Z-score normalization on the joint tensor so that the mean of each feature dimension is 0 and the standard deviation is 1. The joint tensor after normalization is the coupled physical field co-evolution structure indexed by three-dimensional spatial coordinates and one-dimensional time coordinates.

[0046] Step S130: Call the temporal causal convolutional Siamese network to process the coupled physical field co-evolution structure, extract the long-term non-stationary evolution characteristics of the coupled physical field co-evolution structure, locate the corresponding causal transmission chain, and generate the latent space trajectory manifold characterizing the evolution trend of the safe state.

[0047] Step S131: Divide the coupled physical field co-evolution structure into multiple spatiotemporal state segment units with temporal continuity according to a preset time step. Perform independent channel dimension normalization transformation on the thermodynamic channel tensor, acoustic channel tensor, and electrochemical channel tensor in each spatiotemporal state segment unit. Input the normalized spatiotemporal state segment units into the first layer of the temporal causal convolutional Siamese network in sequence. The first layer of the dilated causal convolutional network has a first dilation rate parameter and performs unidirectional causal convolution operation only along the time dimension. Extract the local physical field evolution dependency features between adjacent time steps in the spatiotemporal state segment units to generate a first-level temporal feature mapping map.

[0048] The coupled physical field co-evolution structure is divided into multiple continuous spatiotemporal state segments along the time axis according to a preset time step. Each spatiotemporal state segment contains complete physical field data for a fixed number of time steps. For each spatiotemporal state segment, the quantum blocks corresponding to the thermodynamic channel, the acoustic channel, and the electrochemical channel are extracted. A normalization transformation is independently performed on each channel's quantum block, using the formula X_norm=(X-μ) / σ, where μ is the mean of the channel's quantum block in both the time and spatial dimensions, and σ is the standard deviation of the channel's quantum block in both the time and spatial dimensions. The normalized spatiotemporal state segments are then sequentially input into the first layer of the temporal causal convolutional Siamese network in chronological order. The first layer of the dilated causal convolutional network has an dilation rate parameter d=1 and performs unidirectional causal convolution operations only along the time dimension; that is, the output of each time step depends only on the input of the current time step and previous time steps, and not on the input of future time steps. The formula for unidirectional causal convolution is Y(t)=Σ(k=0toK-1)W(k)*X(td*k), where W(k) is the kernel weight, K is the kernel size, and d is the dilation rate parameter. The first layer of dilated causal convolution extracts the local physical field evolution dependency features between adjacent time steps in the spatiotemporal state segment unit, and outputs the first-level temporal feature map.

[0049] Step S132: Input the first-level temporal feature map into the second-layer dilated causal convolution computation layer of the temporal causal convolutional Siamese network. The second-layer dilated causal convolution computation layer has a second dilation parameter that is greater than the first dilation parameter and performs unidirectional causal convolution operation along the time dimension to extract the mid-range physical field evolution dependency features spanning multiple time steps in the first-level temporal feature map and generate the second-level temporal feature map.

[0050] The first-level temporal feature map output from step S131 is input into the second layer of the temporal causal convolutional Siamese network, which is then processed by an expanded causal convolutional layer. This second layer has an expansion rate parameter d=2 and performs unidirectional causal convolution operations only along the time dimension. Due to the increased expansion rate parameter, the equivalent receptive field of the convolutional kernel expands, allowing each output position to aggregate input information from the first-level temporal feature map across multiple time steps. The second layer extracts the mid-range physical field evolution dependency features spanning multiple time steps from the first-level temporal feature map and outputs the second-level temporal feature map.

[0051] Step S133: Input the second-level temporal feature map into the third-level dilated causal convolution computation layer of the temporal causal convolutional Siamese network. The third-level dilated causal convolution computation layer has a third dilation parameter greater than the second dilation parameter and performs unidirectional causal convolution operation along the time dimension. Extract the long-range physical field evolution dependency features that span long temporal intervals in the second-level temporal feature map to generate the third-level temporal feature map. Cascade the first-level temporal feature map, the second-level temporal feature map, and the third-level temporal feature map along the feature channel dimension to generate a multi-scale long-term non-stationary evolution feature set.

[0052] The second-level temporal feature map output in step S132 is input into the third-level dilated causal convolutional computation layer of the temporal causal convolutional Siamese network. This third-level dilated causal convolutional computation layer has a dilation rate parameter d=4 and also performs unidirectional causal convolution operations only along the time dimension. By further increasing the dilation rate parameter, the equivalent receptive field of the convolutional kernel is further expanded, allowing each output position to aggregate input information from the second-level temporal feature map spanning longer temporal intervals. The third-level dilated causal convolutional computation layer extracts long-range physical field evolution dependency features spanning long temporal intervals from the second-level temporal feature map, outputting the third-level temporal feature map. The first-level, second-level, and third-level temporal feature maps are then concatenated along the feature channel dimension, i.e., the three feature maps are joined end-to-end along the channel dimension, generating a multi-scale, long-term non-stationary evolution feature set with an increased number of channels.

[0053] Step S134: Input the multi-scale long-term non-stationary evolution feature set into the differential attention weighted aggregation layer of the temporal causal convolutional Siamese network, calculate the adjacent frame feature difference residual map of the multi-scale long-term non-stationary evolution feature set on the time axis in the differential attention weighted aggregation layer, and calculate the channel attention weight distribution vector along the feature channel dimension based on the adjacent frame feature difference residual map.

[0054] The multi-scale long-term non-stationary evolution feature set generated in step S133 is input into the differential attention weighted aggregation layer of the temporal causal convolutional Siamese network. In this differential attention weighted aggregation layer, for two adjacent time steps t and t+1 on the time axis, a differential residual map is calculated between the feature maps corresponding to these two time steps. The formula for calculating the differential residual map is D(t) = |F(t+1) - F(t)|, where F(t) is the feature map at time step t. This calculation is performed on all adjacent time steps to obtain a series of differential residual maps. Each differential residual map is then globally averaged along the spatial dimension to obtain a vector equal to the number of feature channels. Each element in this vector represents the intensity of change of the corresponding channel at time step t. The intensity vectors of change at all time steps are stacked to form a two-dimensional matrix. The mean is calculated for each column of this two-dimensional matrix (corresponding to a feature channel) to obtain the channel attention weight distribution vector. Each element in this channel attention weight distribution vector represents the importance of the corresponding feature channel in characterizing the evolution of the physical field.

[0055] Step S135: Use the channel attention weight distribution vector to perform weighted enhancement processing on the multi-scale long-term non-stationary evolution feature set, increase the activation response value of the channel corresponding to the gradient steep change region inside the heat flow transfer topology map, increase the activation response value of the channel corresponding to the transient singular response perturbation in the microstructure path feature, increase the activation response value of the channel corresponding to the spectral response drift trend in the ion transport interface state descriptor, and generate a weighted feature map set that enhances causal conduction features.

[0056] Each weight coefficient in the channel attention weight distribution vector obtained in step S134 is multiplied by all feature maps in the multi-scale long-term non-stationary evolution feature set corresponding to that feature channel. For feature channels corresponding to regions with abrupt gradient changes within the heat flow topology, the channel attention weight coefficient is greater than 1, thereby increasing the activation response value of that channel. For feature channels corresponding to transient singular response perturbations in microstructure path features, the channel attention weight coefficient is also greater than 1, thereby increasing the activation response value of that channel. For feature channels corresponding to spectral response drift trends in ion transport interface state descriptors, the channel attention weight coefficient is also greater than 1, thereby increasing the activation response value of that channel. For feature channels with insignificant changes, the channel attention weight coefficient is less than 1, thereby suppressing the activation response value of that channel. The multi-scale long-term non-stationary evolution feature set after weighted enhancement processing is the weighted feature map set for enhancing causal conduction features.

[0057] Step S136: Perform parallel pooling operations of spatial dimension global average pooling and global max pooling on the weighted feature map set, concatenate the parallel pooling results along the feature channel dimension to generate a spatially compressed feature vector, input the spatially compressed feature vector into the fully connected mapping layer of the temporal causal convolutional Siamese network, and generate the causal transmission chain feature vector corresponding to the first Siamese branch and the causal transmission chain feature vector corresponding to the second Siamese branch.

[0058] Two pooling operations are performed simultaneously on the weighted feature map. Global average pooling calculates the average of the activation responses at all spatial locations in each feature channel, resulting in a vector A. Global max pooling calculates the maximum of the activation responses at all spatial locations in each feature channel, resulting in a vector M. Vectors A and M are concatenated along the feature channel dimension to generate a spatially compressed feature vector C=[A, M]. This spatially compressed feature vector is input into the fully connected mapping layer of a temporal causal convolutional Siamese network. The fully connected mapping layer consists of two cascaded fully connected sublayers, each containing multiple neurons. Adjacent layers are operated using a linear transformation combined with a non-linear activation function. After transformation by the fully connected mapping layer, a low-dimensional causal propagation chain feature vector is output. The temporal causal convolutional Siamese network contains two branches with identical structures but independent weights: the first Siamese branch and the second Siamese branch. The same spatially compressed feature vector is input into the fully connected mapping layers of both branches, resulting in the causal propagation chain feature vector Z1 corresponding to the first Siamese branch and the causal propagation chain feature vector Z2 corresponding to the second Siamese branch.

[0059] Step S137: Calculate the contrast loss function value between the causal transmission chain feature vector corresponding to the first twin branch and the causal transmission chain feature vector corresponding to the second twin branch. Based on the contrast loss function value, update the network weight parameters of the dilated causal convolution calculation layer and the differential attention weighted aggregation layer in reverse, so that the feature vector distance between similar physical field evolution modes is reduced and the feature vector distance between dissimilar physical field evolution modes is increased.

[0060] For each training sample pair, the input of the first twin branch comes from the current coupled physical field co-evolution structure, and the input of the second twin branch comes from the same coupled physical field co-evolution structure after random data augmentation. This sample pair is a positive sample pair. Simultaneously, coupled physical field co-evolution structures from different time windows or different battery packs are considered negative sample pairs. The contrastive loss function is calculated as L=Y*D^2+(1-Y)*max(margin-D,0)^2, where Y represents whether the sample pair is a positive sample pair, D is the Euclidean distance between Z1 and Z2, and margin is a preset boundary threshold. After calculating the contrastive loss function, the gradient descent algorithm is used to backpropagate the partial derivatives of the loss function with respect to the network weight parameters layer by layer, sequentially updating the network weight parameters of the third layer of dilated causal convolution, the second layer of dilated causal convolution, the first layer of dilated causal convolution, and the differential attention weighted aggregation layer. By comparing the optimization of the loss function values, the Euclidean distance between the feature vectors of the causal transmission chain corresponding to similar physical field evolution modes gradually decreases, while the Euclidean distance between the feature vectors of the causal transmission chain corresponding to dissimilar physical field evolution modes gradually increases.

[0061] Step S138: Input the causal transmission chain feature vector corresponding to the first twin branch and the causal transmission chain feature vector corresponding to the second twin branch into the manifold learning dimensionality reduction layer of the temporal causal convolutional twin network. The manifold learning dimensionality reduction layer uses the local linear embedding algorithm to project the high-dimensional causal transmission chain feature vector to the low-dimensional manifold space to generate a low-dimensional manifold coordinate sequence that preserves the local neighborhood structure relationship and the global topological connection relationship.

[0062] The causal transmission chain feature vector Z1 corresponding to the first twin branch and the causal transmission chain feature vector Z2 corresponding to the second twin branch obtained in step S136 are averaged element-wise to obtain the fused causal transmission chain feature vector Z = (Z1 + Z2) / 2. The fused causal transmission chain feature vectors of all time windows are arranged in chronological order to form a high-dimensional feature vector sequence. This high-dimensional feature vector sequence is input into the manifold learning dimensionality reduction layer of the temporal causal convolutional twin network. The manifold learning dimensionality reduction layer uses the local linear embedding algorithm. The execution process of the local linear embedding algorithm is as follows: For each data point in the high-dimensional feature vector sequence, its K nearest neighbors in the Euclidean distance sense are found. The weight coefficients of each data point linearly reconstructed from its K nearest neighbors are calculated. The problem of minimizing the reconstruction error is solved by solving a system of linear equations to obtain the weight coefficients. In the low-dimensional manifold space, keeping the above weight coefficients unchanged, a set of low-dimensional coordinates is found to minimize the reconstruction error. This optimization problem is solved by solving the eigenvalue problem of the sparse matrix to obtain the low-dimensional manifold coordinate sequence. This low-dimensional manifold coordinate sequence preserves the local neighborhood structure and global topological connectivity of the original high-dimensional feature vector sequence.

[0063] Step S139: Construct a latent space trajectory manifold based on the continuous change trajectory of the low-dimensional manifold coordinate sequence in the time dimension. The manifold distance between any adjacent time points in the latent space trajectory manifold is used to characterize the evolution rate of the battery pack's safety state. The local curvature value of the latent space trajectory manifold is used to characterize the nonlinear severity of the evolution of the battery pack's safety state.

[0064] The low-dimensional manifold coordinates corresponding to each time window generated in step S138 are connected sequentially to form a continuous curve, which is the latent space trajectory manifold. For any two adjacent time points in the latent space trajectory manifold, the Euclidean distance between these two coordinate points is calculated. This Euclidean distance is used as the manifold distance to characterize the rate of evolution of the battery pack's safety state within that time interval. For each coordinate point on the latent space trajectory manifold, the rate of change of the angle between the two vectors formed by the point and its preceding and following coordinate points is calculated. This rate of change of the angle is used as the local curvature value of that point. The larger the local curvature value, the more nonlinear and severe the evolution of the battery pack's safety state near that time point. The latent space trajectory manifold, the manifold distance sequence, and the local curvature value sequence together constitute a quantitative characterization of the evolution trend of the battery pack's safety state.

[0065] After step S130, the method may further include: Step S13100: Obtain multiple historical hidden space trajectory manifold samples recorded by the new energy battery pack in historical charge and discharge cycles and the safety state category label corresponding to each historical hidden space trajectory manifold sample. The safety state category label includes a safe operation state label and a thermal runaway precursor state label.

[0066] After generating the latent space trajectory manifold characterizing the evolution trend of the safety state in step S130, step S13100 retrieves multiple historical latent space trajectory manifold samples recorded in the historical charge-discharge cycles of the new energy battery pack from the historical database of the battery management system. Each historical latent space trajectory manifold sample is generated from the multi-source physical field response data stream of the previous historical charge-discharge cycle using the methods in steps S110 to S139. Each historical latent space trajectory manifold sample is accompanied by a safety state category label, which includes a safe operating state label and a thermal runaway precursor state label. The safe operating state label indicates that the battery pack has not experienced any thermal runaway-related anomalies in the historical charge-discharge cycle, while the thermal runaway precursor state label indicates that the battery pack has exhibited thermal runaway precursor characteristics such as abnormal temperature rise, abnormal voltage drop, or internal short circuit in the historical charge-discharge cycle.

[0067] Step S13101: Calculate the first manifold spacing value between the latent space trajectory manifold and the center of the safe manifold sample cluster composed of all historical latent space trajectory manifold samples marked as safe operating states; calculate the second manifold spacing value between the latent space trajectory manifold and the center of the risk manifold sample cluster composed of all historical latent space trajectory manifold samples marked as thermal runaway precursor states; when the second manifold spacing value is less than the first manifold spacing value, extract the manifold segment interval in the latent space trajectory manifold where the local curvature value exceeds the preset curvature threshold.

[0068] Calculate the mean coordinates of all historical latent space trajectory manifold samples marked as safe operating states in manifold space to obtain the center of the safe manifold sample cluster. Calculate the mean coordinates of all historical latent space trajectory manifold samples marked as thermal runaway precursor states in manifold space to obtain the center of the risk manifold sample cluster. For the latent space trajectory manifold generated in the current period, calculate the geodesic distance between each coordinate point and the center of the safe manifold sample cluster, and take the average of all geodesic distances as the first manifold spacing value. Similarly, calculate the second manifold spacing value between the latent space trajectory manifold and the center of the risk manifold sample cluster. Compare the second manifold spacing value with the first manifold spacing value. If the second manifold spacing value is less than the first manifold spacing value, it indicates that the current latent space trajectory manifold is closer to the center of the risk manifold sample cluster corresponding to the thermal runaway precursor state category label. On the latent space trajectory manifold, the local curvature value of each coordinate point is traversed. The local curvature value is calculated as follows: for three consecutive points P(t-1), P(t), and P(t+1) on the trajectory, vectors a = P(t) - P(t-1) and b = P(t+1) - P(t) are calculated, and then the rate of change of the angle between vectors a and b is calculated as the local curvature value of that point. Coordinate points whose local curvature values ​​exceed a preset curvature threshold are marked, and multiple consecutive marked coordinate points are merged into a manifold segment interval.

[0069] Step S13102: Backtrack to locate the time axis interval corresponding to the manifold segment interval, and analyze the characteristic activation contribution values ​​of the thermodynamic channel tensor, acoustic channel tensor and electrochemical channel tensor in the coupled physical field co-evolution structure within the time axis interval.

[0070] The manifold segment intervals extracted in step S13101 are mapped back to the original time axis to obtain the corresponding time axis intervals. Within these time axis intervals, thermodynamic channel tensors, acoustic channel tensors, and electrochemical channel tensors are extracted from the coupled physical field co-evolution structure. For each channel in the thermodynamic channel tensor, the variance of that channel within the time axis interval is calculated, and this variance is used as the characteristic activation contribution of that thermodynamic channel. A larger variance indicates a more significant change in that channel during anomalous evolution. The same operation is performed on the acoustic channel tensors and the electrochemical channel tensors to obtain the characteristic activation contributions of the acoustic channel and the electrochemical channel, respectively.

[0071] Step S13103: Extract the thermodynamic channel index, acoustic channel index, and electrochemical channel index ranked in the top preset position in the feature activation contribution value sorting. Determine the thermodynamic anomaly space region based on the mapping relationship of the thermodynamic channel index in the three-dimensional space of the battery pack, determine the acoustic anomaly space region based on the mapping relationship of the acoustic channel index on the surface of the battery casing, and determine the electrochemical anomaly space region based on the mapping relationship of the electrochemical channel index in the battery electrode stack.

[0072] Sort the characteristic activation contribution values ​​of all thermodynamic channels calculated in step S13102 from largest to smallest, and extract the thermodynamic channel indices of the top-ranked preset positions. Query the pre-established mapping table between the thermodynamic channel indices and the three-dimensional solid coordinates of the battery pack to determine the spatial coordinate range corresponding to each thermodynamic channel index. The union of these spatial coordinate ranges is taken as the thermodynamic anomaly spatial region. Similarly, sort the characteristic activation contribution values ​​of acoustic channels from largest to smallest, extract the acoustic channel indices of the top-ranked preset positions, and query the mapping table between the acoustic channel indices and the sensor positions on the battery casing surface to determine the acoustic anomaly spatial region. Sort the characteristic activation contribution values ​​of electrochemical channels from largest to smallest, extract the electrochemical channel indices of the top-ranked preset positions, and query the mapping table between the electrochemical channel indices and the battery electrode stack positions to determine the electrochemical anomaly spatial region.

[0073] Step S13104: The thermodynamic anomaly space region, the acoustic anomaly space region, and the electrochemical anomaly space region are fused to generate the anomaly physical field spatial positioning region. The thermal runaway risk evolution stage identifier is determined based on the cumulative change of manifold distance and the local curvature peak value of the manifold segment interval.

[0074] The thermodynamic, acoustic, and electrochemical anomaly regions determined in step S13103 are superimposed in a three-dimensional spatial coordinate system, and the union of the three regions is taken as the spatial location region of the anomalous physical field. For each manifold segment interval extracted in step S13101, the sum of the manifold distances between all adjacent coordinate points within the interval is calculated to obtain the cumulative change in manifold distance. The maximum local curvature within the manifold segment interval is extracted and taken as the local curvature peak value. The cumulative change in manifold distance and the local curvature peak value are input into a preset risk stage classification function. This risk stage classification function contains two comparators and one encoder: the first comparator compares the cumulative change in manifold distance with a first threshold, and the second comparator compares the local curvature peak value with a second threshold; the encoder combines the outputs of the two comparators to generate the corresponding thermal runaway risk evolution stage identifier. The thermal runaway risk evolution stage identifiers include an early warning stage identifier, a mid-stage acceleration stage identifier, and a near-thermal runaway stage identifier.

[0075] Step S140: Based on the projection offset direction and curvature abrupt change amplitude of the latent space trajectory manifold within the preset high-dimensional curvature boundary, determine the battery pack thermal runaway risk evolution stage identifier and associated abnormal physical field spatial location area.

[0076] Step S141: Obtain multiple historical hidden space trajectory manifold samples recorded by the new energy battery pack in historical charge-discharge cycles. Each historical hidden space trajectory manifold sample corresponds to a known safe operating state category label or a known thermal runaway precursor state category label. Perform manifold space similarity measurement processing on the hidden space trajectory manifold and the multiple historical hidden space trajectory manifold samples, calculate the first feature distance value between the hidden space trajectory manifold and the center of the safe manifold sample cluster composed of all historical hidden space trajectory manifold samples labeled as safe operating states, and calculate the second feature distance value between the hidden space trajectory manifold and the center of the risk manifold sample cluster composed of all historical hidden space trajectory manifold samples labeled as thermal runaway precursor states.

[0077] Multiple historical latent space trajectory manifold samples recorded during historical charge-discharge cycles of the new energy battery pack are obtained from the historical database of the battery management system. Each historical latent space trajectory manifold sample is accompanied by a known safety state category label, including a safe operating state category label and a thermal runaway precursor state category label. The mean coordinates of all historical latent space trajectory manifold samples marked as safe operating states are calculated in the manifold space to obtain the center of the safe manifold sample cluster. The mean coordinates of all historical latent space trajectory manifold samples marked as thermal runaway precursor states are calculated in the manifold space to obtain the center of the risk manifold sample cluster. For the latent space trajectory manifold generated in the current cycle, the geodesic distance between each coordinate point and the center of the safe manifold sample cluster is calculated, and the average of the geodesic distances of all coordinate points is taken as the first characteristic distance value. Similarly, the second characteristic distance value between the latent space trajectory manifold and the center of the risk manifold sample cluster is calculated.

[0078] Step S142: When the second feature distance value is less than the first feature distance value, the latent space trajectory manifold is classified as a potential thermal runaway evolution trajectory category, and the manifold segment intervals in the latent space trajectory manifold where the local curvature value exceeds the preset curvature threshold are extracted.

[0079] The second feature distance value is compared with the first feature distance value. If the second feature distance value is less than the first feature distance value, it indicates that the current latent space trajectory manifold is closer to the center of the risk manifold sample cluster corresponding to the thermal runaway precursor state category label. Therefore, the latent space trajectory manifold is classified as a potential thermal runaway evolution trajectory category. On this latent space trajectory manifold, the local curvature value of each coordinate point is traversed, and coordinate points whose local curvature values ​​exceed a preset curvature threshold are marked. Multiple consecutive marked coordinate points are merged into a manifold segment interval, with each manifold segment interval corresponding to a period of time when the safe state undergoes a drastic nonlinear change.

[0080] Step S143: Backtrack to locate the time axis interval corresponding to the manifold segment interval, and parse the feature activation contribution ranking of the thermodynamic channel tensor, acoustic channel tensor and electrochemical channel tensor in the coupled physical field co-evolution structure within the time axis interval. Extract the thermodynamic channel index, acoustic channel index and electrochemical channel index that rank first in the feature activation contribution ranking.

[0081] The manifold segment intervals extracted in step S142 are mapped back to the original time axis to obtain the corresponding time axis intervals. Within these time axis intervals, thermodynamic channel tensors, acoustic channel tensors, and electrochemical channel tensors are extracted from the coupled physical field co-evolution structure. For each channel tensor, its variance value within the time axis interval is calculated, and this variance value is used as the feature activation contribution of that channel. The larger the variance value, the more significant the change in that channel during the anomalous evolution process, and the higher its contribution. All thermodynamic channels are sorted from largest to smallest according to their feature activation contribution, and the top N thermodynamic channel indices are extracted. Similarly, the top N acoustic channel indices and electrochemical channel indices are extracted.

[0082] Step S144: Determine the spatial location region of thermodynamic anomaly based on the spatial mapping relationship of the thermodynamic channel index in the three-dimensional solid structure of the battery pack; determine the spatial location region of acoustic anomaly based on the spatial mapping relationship of the acoustic channel index in the sensor network on the surface of the battery casing; determine the spatial location region of electrochemical anomaly based on the spatial mapping relationship of the electrochemical channel index in the battery electrode stack; and merge the spatial location regions of thermodynamic anomaly, acoustic anomaly, and electrochemical anomaly to generate an abnormal physical field spatial location region.

[0083] Based on the thermodynamic channel index extracted in step S143, a pre-established mapping table between the thermodynamic channel index and the three-dimensional solid coordinates of the battery pack is consulted to determine the spatial coordinate range corresponding to each thermodynamic channel index. The union of these spatial coordinate ranges is taken as the spatial location region of the thermodynamic anomaly. Based on the acoustic channel index, a mapping table between the acoustic channel index and the sensor position on the battery casing surface is consulted to determine the coordinate range of the casing surface corresponding to each acoustic channel index. The union of these coordinate ranges is taken as the spatial location region of the acoustic anomaly. Based on the electrochemical channel index, a mapping table between the electrochemical channel index and the battery electrode stack position is consulted to determine the coordinate range of the electrode stack corresponding to each electrochemical channel index. The union of these coordinate ranges is taken as the spatial location region of the electrochemical anomaly. The spatial location regions of the thermodynamic anomaly, acoustic anomaly, and electrochemical anomaly are superimposed in a three-dimensional spatial coordinate system, and the union of the three regions is taken as the spatial location region of the anomaly physical field.

[0084] Step S145: Calculate the thermal runaway risk evolution rate coefficient based on the cumulative change of manifold distance in the manifold segment interval, calculate the thermal runaway risk severity coefficient based on the local curvature peak of the manifold segment interval, input the thermal runaway risk evolution rate coefficient and the thermal runaway risk severity coefficient into a preset risk stage division mapping function, and generate a corresponding potential thermal runaway risk evolution stage category identifier.

[0085] For each manifold segment interval extracted in step S142, the sum of manifold distances between all adjacent coordinate points within that interval is calculated to obtain the cumulative change in manifold distance. This cumulative change in manifold distance is divided by the time length of the manifold segment interval to obtain the thermal runaway risk evolution rate coefficient. This coefficient represents the distance the battery's safe state moves within the manifold space per unit time. The maximum local curvature within the manifold segment interval is extracted and used as the thermal runaway risk severity coefficient. A risk stage classification mapping function is predefined. The inputs to this function are the thermal runaway risk evolution rate coefficient and the thermal runaway risk severity coefficient, and the output is a potential thermal runaway risk evolution stage category identifier. The potential thermal runaway risk evolution stage category identifiers include an early warning stage identifier, a mid-stage acceleration stage identifier, and an imminent thermal runaway stage identifier.

[0086] Step S146: Based on the category identifier of the potential thermal runaway risk evolution stage, retrieve the preset graded protection strategy mapping table, and extract the charge / discharge power limit ratio parameter, thermal management cycle start threshold parameter, and alarm notification target address list corresponding to the category identifier of the potential thermal runaway risk evolution stage.

[0087] A pre-defined hierarchical protection strategy mapping table is established. Each record in this table corresponds to a potential thermal runaway risk evolution stage category identifier, including the charge / discharge power limit ratio parameter, thermal management cycle initiation threshold parameter, and alarm notification target address list associated with that stage category identifier. Based on the potential thermal runaway risk evolution stage category identifier determined in step S145, a search is performed in this hierarchical protection strategy mapping table to extract the corresponding charge / discharge power limit ratio parameter, thermal management cycle initiation threshold parameter, and alarm notification target address list. The charge / discharge power limit ratio parameter is a proportional value between 0 and 1, indicating the percentage by which the current maximum allowable charge / discharge power is limited to its original value. The thermal management cycle initiation threshold parameter is a temperature threshold value; when the detected temperature exceeds this value, the thermal management cycle is initiated. The alarm notification target address list includes the network address of the vehicle's central gateway controller and the network address of the remote cloud monitoring service platform that need to receive safety alarm information.

[0088] Step S147: Write the charge / discharge power limit ratio parameter into the power boundary register of the battery management control unit to modify the upper and lower limits of the charge / discharge power boundary envelope, and write the thermal management cycle start threshold parameter into the threshold comparator of the thermal management control unit to trigger the adjustment of the circulation rate of the cooling medium in the directional thermal management loop corresponding to the abnormal physical field spatial positioning area.

[0089] The battery management control unit contains a power boundary configuration register, which stores the upper and lower limits of the current charge / discharge power boundary envelope. The charge / discharge power limit ratio parameter extracted in step S146 is written into this power boundary configuration register, and the battery management control unit recalculates the maximum allowable charging current limit and the maximum allowable discharging current limit based on the new power boundary values. The thermal management control unit contains a temperature threshold comparison register, which stores the temperature threshold used to trigger the cooling medium circulation start-up. The thermal management circulation start-up threshold parameter is written into this temperature threshold comparison register. Simultaneously, based on the abnormal physical field spatial positioning area determined in step S144, the address code of the flow regulating valve covering the directional thermal management circulation path within the spatial positioning area is retrieved from the preset cooling circuit topology mapping table. When the real-time temperature value collected by the temperature sensing node deployed within the abnormal physical field spatial positioning area exceeds the thermal management circulation start-up threshold parameter, the thermal management control unit generates a valve opening adjustment command, adjusting the valve opening of the corresponding flow regulating valve to the preset maximum opening value, increasing the circulation velocity of the cooling medium.

[0090] Step S148: Based on the alarm notification target address list, encapsulate the potential thermal runaway risk evolution stage category identifier, the abnormal physical field spatial location area, and the thermal runaway risk evolution rate coefficient into a preset format safety warning data message, and send it to the vehicle central gateway controller and remote cloud monitoring service platform.

[0091] Following the data frame format of the vehicle network communication protocol, the potential thermal runaway risk evolution stage category identifier determined in step S145, the coordinate range of the abnormal physical field spatial positioning area determined in step S144, and the thermal runaway risk evolution rate coefficient calculated in step S145 are encapsulated into a safety warning data message. This safety warning data message includes a message type identifier field, a timestamp field, a stage category field, a spatial positioning area coordinate array field, and a rate coefficient field. This safety warning data message is sent to the vehicle's central gateway controller, which forwards it to the instrument panel display unit for warning notification. Simultaneously, this safety warning data message is sent to the remote cloud monitoring service platform via the vehicle's wireless communication module for data recording and analysis by the remote monitoring center.

[0092] Step S149: Record the changes in the thermal distribution sequence, acoustic response sequence, and impedance spectrum evolution sequence of the new energy battery pack within a preset monitoring time window after the execution of the differentiated protection instruction set, and feed them back to the incremental learning sample buffer of the temporal causal convolutional twin network for online optimization of the network weight parameters of the temporal causal convolutional twin network.

[0093] Within a preset monitoring time window following the execution of the differentiated protection command set, data on the thermal distribution sequence, acoustic response sequence, and impedance spectrum evolution sequence of the new energy battery pack are continuously collected. The newly collected data is processed according to steps S120 to S130 to generate a new latent space trajectory manifold. This new latent space trajectory manifold and its corresponding protection command execution effect label (effective or ineffective) are combined to form an incremental learning sample. This incremental learning sample is added to the incremental learning sample buffer of the temporal causal convolutional twin network. When the number of samples in the incremental learning sample buffer reaches a preset batch size, an online incremental learning process is triggered. The new samples in the buffer are used to fine-tune and update the network weight parameters of the temporal causal convolutional twin network, enabling the network to adapt to changes in battery pack aging characteristics and operating conditions.

[0094] Step S150: Generate a differentiated protection instruction set based on the thermal runaway risk evolution stage identifier and the abnormal physical field spatial positioning area. The differentiated protection instruction set is used to adjust the charge and discharge power boundary and start the directional thermal management loop corresponding to the abnormal physical field spatial positioning area.

[0095] Step S151: Based on the thermal runaway risk evolution stage identifier, retrieve the preset stage classification strategy mapping table, and extract the charge / discharge power limiting ratio parameter, thermal management cycle start threshold parameter, and response priority code value corresponding to the thermal runaway risk evolution stage identifier.

[0096] A pre-defined stage-based strategy mapping table is established. Each record in this table corresponds to a thermal runaway risk evolution stage identifier, including the charge / discharge power limiting ratio parameter, thermal management cycle initiation threshold parameter, and response priority code value associated with that stage identifier. Based on the thermal runaway risk evolution stage identifier determined in step S140, the corresponding charge / discharge power limiting ratio parameter, thermal management cycle initiation threshold parameter, and response priority code value are retrieved from the stage-based strategy mapping table. The response priority code value is an integer value; a larger value indicates that the protection command should be executed with priority.

[0097] Step S152: Write the charge / discharge power limiting ratio parameter into the power boundary configuration register of the battery management control unit, so that the battery management control unit recalculates the maximum allowable charging current limit and the maximum allowable discharging current limit in the current charge / discharge cycle based on the charge / discharge power limiting ratio parameter.

[0098] The battery management control unit internally includes a power boundary configuration register. The charge / discharge power limiting ratio parameter extracted in step S151 is written into this power boundary configuration register. The battery management control unit reads the ratio parameter value from the register and multiplies the original maximum allowable charging current limit by this ratio parameter to obtain a new maximum allowable charging current limit. Similarly, the original maximum allowable discharge current limit is multiplied by this ratio parameter to obtain a new maximum allowable discharge current limit. During subsequent charge / discharge control, the battery management control unit limits the charging current to within the new maximum allowable charging current limit and the discharge current to within the new maximum allowable discharge current limit.

[0099] Step S153: Based on the spatial coordinate coverage of the abnormal physical field spatial positioning area in the three-dimensional solid structure of the battery pack, retrieve the path identifier of at least one target-oriented thermal management circulation path covering the spatial coordinate coverage area and the address code of the flow regulating valve corresponding to the target-oriented thermal management circulation path from the preset cooling circuit topology mapping table.

[0100] A pre-defined cooling circuit topology mapping table records the correspondence between each spatial coordinate region in the three-dimensional solid structure of the battery pack and the directional thermal management circulation path covering that region. Each directional thermal management circulation path includes a cooling medium circulation pipe, a flow regulating valve controlling the flow rate of that pipe, and the address code of the valve. Based on the spatial coordinate coverage range of the abnormal physical field spatial positioning area determined in step S144, a spatial query is performed in the cooling circuit topology mapping table to retrieve the path identifiers of all directional thermal management circulation paths covering that spatial coordinate coverage range and their corresponding flow regulating valve address codes.

[0101] Step S154: Write the thermal management cycle start threshold parameter into the temperature threshold comparison register of the thermal management control unit. When the real-time temperature value collected by the temperature sensing node deployed in the abnormal physical field spatial positioning area exceeds the thermal management cycle start threshold parameter, the thermal management control unit generates a cooling medium cycle start trigger signal.

[0102] The thermal management control unit contains a temperature threshold comparison register. The thermal management cycle start-up threshold parameter extracted in step S151 is written into this register. The thermal management control unit continuously reads the real-time temperature values ​​collected by each temperature sensing node deployed within the abnormal physical field spatial positioning area. Each real-time temperature value is compared with the thermal management cycle start-up threshold parameter stored in the temperature threshold comparison register. When the real-time temperature value of any temperature sensing node exceeds the thermal management cycle start-up threshold parameter, the thermal management control unit generates a cooling medium cycle start-up trigger signal.

[0103] Step S155: Generate a valve opening adjustment command based on the cooling medium circulation start-up trigger signal and the address code of the flow regulating valve. The valve opening adjustment command is used to adjust the valve opening of the flow regulating valve corresponding to the target directional thermal management circulation path to a preset maximum opening value.

[0104] After receiving the cooling medium circulation start trigger signal generated in step S154, the thermal management control unit generates a valve opening adjustment command based on the address code of the flow regulating valve retrieved in step S153. This valve opening adjustment command includes the address code field of the target flow regulating valve and a target valve opening value field, with the target valve opening value field set to a preset maximum opening value. The thermal management control unit sends this valve opening adjustment command to the corresponding flow regulating valve actuator via the controller area network bus. The flow regulating valve actuator adjusts the valve opening to the preset maximum opening value, allowing the cooling medium to flow through the directional thermal management circulation path at maximum flow rate.

[0105] Step S156: Based on the response priority encoding value, encapsulate the charge / discharge power limiting ratio parameter, the path identifier of the target directional thermal management circulation path, the valve opening adjustment command, and the spatial coordinate coverage of the abnormal physical field spatial positioning area into a differentiated protection command data frame.

[0106] A differentiated protection instruction data frame is created according to the data frame format of the vehicle network communication protocol. This data frame includes a header field, a data field, and a check field. The header field contains the response priority encoding value extracted in step S151. The data field contains the charge / discharge power limiting ratio parameter, the path identifier of the target-oriented thermal management circulation path, the complete content of the valve opening adjustment instruction, and an array of spatial coordinates covering the abnormal physical field spatial positioning area. The cyclic redundancy check value of the data frame is calculated and filled into the check field.

[0107] Step S157: The differentiated protection instruction data frame is sent to the battery management control unit and the thermal management control unit with a bus arbitration priority corresponding to the response priority encoding value, triggering the charge and discharge power boundary adjustment operation and the directional thermal management loop start operation.

[0108] Based on the response priority encoding value in the differentiated protection instruction data frame encapsulated in step S156, the message identifier field is set when the controller local area network (CLAN) sends the data. A larger response priority encoding value and a smaller message identifier value result in higher priority in CLAN arbitration. The differentiated protection instruction data frame is then sent to the battery management control unit and the thermal management control unit via the CLAN. Upon receiving the data frame, the battery management control unit parses the charge / discharge power limiting ratio parameter and executes the power boundary adjustment operation described in step S152. Upon receiving the data frame, the thermal management control unit parses the thermal management cycle start-up threshold parameter and the valve opening adjustment command and executes the directional thermal management loop start-up operation described in steps S154 and S155.

[0109] For example, step S160: Perform manifold topological persistence analysis on the multidimensional latent space trajectory manifold, and construct a manifold topological persistence barcode map based on the continuous homology group structure change of the multidimensional latent space trajectory manifold under a preset scale parameter sequence.

[0110] After generating the multidimensional latent space trajectory manifold representing the evolution trend of the security state in step S130, step S160 performs manifold topological persistence analysis on the multidimensional latent space trajectory manifold. The multidimensional latent space trajectory manifold is considered as a point cloud dataset, where each point corresponds to the low-dimensional manifold coordinates of a time window. A preset scale parameter sequence is defined, containing multiple scale parameter values ​​that are logarithmically uniformly distributed from minimum to maximum. For each scale parameter value, a Vietoris-Rips complex of the point cloud dataset is constructed with that scale parameter value as the radius. The persistent homology group of each Vietoris-Rips complex is calculated, including a zero-dimensional homology group, a one-dimensional homology group, and a two-dimensional homology group. The generator of the zero-dimensional homology group corresponds to the connected components in the point cloud dataset, the generator of the one-dimensional homology group corresponds to the loop structure in the point cloud dataset, and the generator of the two-dimensional homology group corresponds to the hole structure in the point cloud dataset. Record the scale parameter values ​​at the first appearance and the last appearance of each homology group generator. The difference between these two values ​​is taken as the lifetime length of that generator. Plot a barcode diagram with the lifetime length of each generator on the x-axis and the scale parameter value at its first appearance on the y-axis, with each barcode corresponding to one generator. The barcodes of all generators constitute a manifold topological persistence barcode diagram.

[0111] Step S161: Extract topological hole features and topological connectivity component features from the manifold topological persistence barcode map where the lifetime length exceeds a preset persistence threshold. The topological hole features are used to characterize non-connected abnormal regions existing in the physical field state space inside the battery pack, and the topological connectivity component features are used to characterize discrete state clusters existing in the physical field state space inside the battery pack.

[0112] Traverse all barcodes in the manifold topological persistence barcode graph. For each barcode, compare its lifetime length with a preset persistence threshold. If the lifetime length is greater than the preset persistence threshold, retain the generator corresponding to that barcode. For generators corresponding to one-dimensional homology groups, mark them as topological hole features. These topological hole features indicate the existence of disconnected anomalous regions in the physical field state space within the battery pack, meaning that certain combinations of physical states cannot be reached from normal state regions via continuous paths. For generators corresponding to zero-dimensional homology groups, mark them as topological connected component features. These topological connected component features indicate the existence of discrete state clusters in the physical field state space within the battery pack, meaning that physical field states are distributed across multiple isolated clustered regions.

[0113] Step S162: Map the topological hole feature back to the three-dimensional solid space coordinate system of the battery pack corresponding to the coupled physical field co-evolution structure, and locate the potential micro-short circuit occurrence region inside the battery pack corresponding to the spatial generator coordinates of the topological hole feature.

[0114] For each topological hole feature extracted in step S161, the set of simplex vertices in the Vietoris-Rips complex at the first occurrence of the generator corresponding to that topological hole feature is obtained. Each simplex vertex corresponds to a point in the point cloud dataset, which is associated with the coordinates of a low-dimensional manifold in the multidimensional latent space trajectory manifold. The corresponding spatiotemporal state segment unit is traced back through the low-dimensional manifold coordinates, and then the specific spatial locations of the thermodynamic channel tensor, acoustic channel tensor, and electrochemical channel tensor in the coupled physical field co-evolution structure are located through the spatiotemporal state segment unit. The union of all located spatial locations is taken as the potential micro-short circuit occurrence region inside the battery pack.

[0115] Step S163: Map the topological connectivity component features back to the three-dimensional physical space coordinate system of the battery pack corresponding to the coupled physical field co-evolution structure, and locate the potential thermal accumulation isolation region inside the battery pack corresponding to the spatial generator coordinates of the topological connectivity component features.

[0116] For each topologically connected component feature extracted in step S161, all vertices contained in the connected component of the Vietoris-Rips complex at the first occurrence of the generator corresponding to that topologically connected component feature are obtained. Each vertex corresponds to a point in the point cloud dataset, and its specific spatial location in the coupled physical field co-evolution structure is traced back through this point. The union of all located spatial locations is taken as the potential thermal accumulation isolation region inside the battery pack. This potential thermal accumulation isolation region indicates that within this spatial region, heat cannot effectively diffuse to the surroundings, forming local thermal accumulation.

[0117] Step S164: Perform spatial overlay analysis on the spatial coordinate range of the potential micro-short circuit occurrence area and the spatial coordinate range of the potential thermal accumulation isolation area. When the spatial overlap between the potential micro-short circuit occurrence area and the potential thermal accumulation isolation area exceeds a preset overlap threshold, mark the spatially overlapping area as a high-risk thermal runaway induction source area.

[0118] Calculate the intersection region between the spatial coordinate range of the potential micro-short circuit occurrence region and the spatial coordinate range of the potential thermal accumulation isolation region. Calculate the ratio of the volume of this intersection region to the volume of the potential micro-short circuit occurrence region, and the ratio of the volume of this intersection region to the volume of the potential thermal accumulation isolation region. Take the larger of the two ratios as the spatial overlap. Compare this spatial overlap with a preset overlap threshold. If the spatial overlap is greater than the preset overlap threshold, it indicates that the micro-short circuit risk region and the thermal accumulation risk region significantly overlap. Within this overlapping region, the local high temperature generated by the micro-short circuit cannot effectively diffuse, easily inducing thermal runaway. Mark this spatially overlapping region as a high-risk thermal runaway induction source region.

[0119] Step S165: Generate high-risk region spatial positioning information based on the coordinate boundary of the high-risk thermal runaway induction source region in the three-dimensional physical space of the battery pack, and add the high-risk region spatial positioning information to the abnormal physical field spatial positioning region, so that the abnormal physical field spatial positioning region includes the abnormal region determined by manifold curvature analysis and the high-risk induction source region determined by topological persistence analysis.

[0120] The six boundary coordinates of the minimum axis-aligned bounding box of the high-risk thermal runaway induction source region in the three-dimensional solid space of the battery pack are extracted, including the minimum, maximum, minimum, maximum, minimum, and maximum values ​​of the X-axis, Z-axis, and Z-axis. These boundary coordinate values ​​are used as the spatial location information of the high-risk region. This high-risk region spatial location information is added as a new entry to the anomalous physical field spatial location region generated in step S144. The updated anomalous physical field spatial location region simultaneously includes the anomalous region determined by manifold curvature analysis and the high-risk induction source region determined by topological persistence analysis.

[0121] Step S166: Use the decay rate of the life cycle length in the manifold topology persistence barcode map as a quantitative characterization index of the degree of irreversible evolution of the battery pack's safety state. When the decay rate of the life cycle length exceeds a preset decay rate threshold, generate an irreversible damage warning label for the battery pack.

[0122] The lifetime lengths of all barcodes in the manifold topology persistence barcode map are traversed, and the lifetime lengths are arranged in ascending order of scale parameter values ​​to obtain a sequence. The difference between two adjacent lifetime lengths in this sequence is calculated, and the difference is divided by the interval of the scale parameter values ​​to obtain the decay rate of the lifetime length. The decay rate of the lifetime length reflects the speed at which the manifold topology degrades as the scale parameter increases. The calculated decay rate is compared with a preset decay rate threshold. If the decay rate is greater than the preset decay rate threshold, it indicates that the topology of the physical field state space inside the battery pack is rapidly degrading, and this degradation is irreversible. At this time, an irreversible damage warning indicator for the battery pack is generated.

[0123] For example, in step S170: during the process of constructing the internal heat flow transfer topology map of the battery pack based on the thermal distribution sequence, the acoustic wave response sequence is subjected to acoustic wave propagation path tomography processing to extract the waveguide travel time offset and waveguide energy attenuation on different propagation paths in the battery casing structure, and an acoustic tomography grayscale map of the casing structure is constructed based on the waveguide travel time offset and waveguide energy attenuation.

[0124] In step S122, during the construction of the internal heat flow transfer topology map of the battery pack based on the thermal distribution sequence, step S170 simultaneously performs acoustic wave propagation path tomography processing on the acoustic wave response sequence. For each pair of piezoelectric ceramic sensors deployed on the battery casing surface, one acts as a transmitter to emit a broadband guided wave signal, and the other acts as a receiver to receive the propagated guided wave signal. The travel time offset of the received signal relative to the transmitted signal is extracted. The formula for calculating the travel time offset is ΔT = T_received - T_emitted, where T_received is the peak time of the received signal wave packet, and T_emitted is the peak time of the transmitted signal wave packet. Simultaneously, the energy attenuation of the received signal relative to the transmitted signal is extracted. The formula for calculating the energy attenuation is A = 10 * log10(E_received / E_emitted), where E_received is the energy of the received signal, and E_emitted is the energy of the transmitted signal. Arrange the travel time offsets and energy attenuation between all transmitter-receiver pairs into a two-dimensional matrix according to the transmitter and receiver positions. Use this two-dimensional matrix as a grayscale image of the acoustic tomography of the shell structure, where the grayscale value of each pixel represents the travel time offset or energy attenuation on the corresponding propagation path.

[0125] Step S171: Perform image edge enhancement and region growing segmentation processing on the grayscale image of the acoustic tomography of the shell structure to extract the contour boundary of the local structural weakening region with abnormal grayscale value changes in the grayscale image of the acoustic tomography of the shell structure.

[0126] Image edge enhancement processing is performed on the grayscale image of the acoustic tomography of the shell structure. First, a Gaussian filter is used to smooth and denoise the grayscale image. Then, the gradient magnitude and direction of each pixel are calculated. The gradient magnitude is calculated using the formula G = sqrt(Gx^2 + Gy^2), where Gx is the horizontal gradient and Gy is the vertical gradient. A non-maximum suppression algorithm is used to preserve local maxima along the gradient direction. Then, a double-threshold hysteresis connection method is used to connect strong and weak edge points into continuous edge lines. On the edge-enhanced grayscale image, regions with abnormal grayscale value changes are selected as seed points. The selection criterion for seed points is that the grayscale value of a pixel deviates from the grayscale mean of the entire image by more than a preset deviation threshold. Region growing segmentation is performed starting from the seed points, merging adjacent pixels with similar grayscale values ​​into the same region until merging is no longer possible. The contour boundaries of the segmented regions are extracted as the contour boundaries of the local structural weakening regions.

[0127] Step S172: Map the contour boundary of the local structural weakening region to the corresponding surface coordinate position of the battery shell surface in the three-dimensional solid space coordinate system of the battery pack, and generate a spatial positioning coordinate set of the shell structure weakening region.

[0128] A 3D solid model of the battery casing surface is obtained, represented by a mesh, where each mesh vertex has 3D spatial coordinates. The coordinates of each pixel in the contour boundary of the local structural weakening region extracted in step S171 are mapped onto the 3D solid model of the battery casing surface. The mapping method uses nearest neighbor projection; that is, for each pixel coordinate, the vertex with the closest Euclidean distance among the mesh vertices of the 3D solid model is found, and the 3D spatial coordinates of that vertex are used as the mapped coordinates of that pixel. All the mapped 3D spatial coordinates constitute the spatial positioning coordinate set of the weakened structural region of the casing.

[0129] Step S173: Perform partial derivative calculation of the impedance imaginary part with respect to frequency on the impedance spectrum evolution sequence to generate the impedance imaginary part frequency partial derivative curve, and extract the frequency point position and offset value of the local extreme value shift in the impedance imaginary part frequency partial derivative curve.

[0130] For the impedance spectrum data at each sampling time in the impedance spectrum evolution sequence, discrete data points showing the change of the imaginary part of impedance, Zim, with frequency f are extracted. The first-order partial derivative of the imaginary part of impedance with respect to frequency is calculated using the finite difference method. The formula for the first-order partial derivative is dZim / df=(Zim(f+Δf)-Zim(f-Δf)) / (2Δf), where Δf is the frequency sampling interval. The partial derivative values ​​corresponding to all frequency points are connected to form a curve, resulting in the frequency partial derivative curve of the imaginary part of impedance. The positions of local extrema are located on this curve, including local maxima and local minima. The frequency positions of the local extrema at the current sampling time are compared with those in the initial healthy state, and the difference between the two is calculated as the offset amplitude. The positions of the frequency points exhibiting local extremum offsets and their corresponding offset amplitudes are recorded.

[0131] Step S174: Determine the type of electrochemical interface in which the interfacial ion transport anomaly occurs inside the battery pack based on the frequency point position of the local extreme value shift, and determine the degree of interface state degradation of the electrochemical interface type based on the shift amplitude value.

[0132] A mapping table is established between frequency point locations and electrochemical interface types. Local extreme value shifts in the high-frequency region correspond to electrode / electrolyte interfaces related to charge transfer processes; local extreme value shifts in the mid-frequency region correspond to interfaces related to solid electrolyte membranes; and local extreme value shifts in the low-frequency region correspond to electrode bulk phase interfaces related to diffusion processes. Based on the frequency point locations of the local extreme value shifts extracted in step S173, this mapping table is consulted to determine the type of electrochemical interface where ion transport anomalies occur. The shift amplitude value is compared with a preset degradation level threshold: when the shift amplitude value is less than the first threshold, the interface degradation level is mild; when the shift amplitude value is between the first and second thresholds, the interface degradation level is moderate; and when the shift amplitude value is greater than the second threshold, the interface degradation level is severe.

[0133] Step S175: Map the electrochemical interface type and the degree of interface degradation to the spatial position of the electrode stack inside the battery pack, determine the spatial coordinate range of the electrode stack corresponding to the electrochemical interface type, and generate a spatial positioning coordinate set of the electrochemical interface degradation area.

[0134] A correspondence table is established between electrochemical interface types and the spatial positions of electrode stacks within the battery pack. The electrode / electrolyte interface corresponds to the electrode surface region, the solid electrolyte interface film corresponds to the electrode surface coating layer region, and the electrode bulk phase interface corresponds to the internal region of the electrode. Based on the electrochemical interface type determined in step S174, this correspondence table is consulted to obtain the spatial coordinate range of the electrode stack corresponding to that electrochemical interface type. This spatial coordinate range is then correlated with the interface degradation degree determined in step S174 to generate a spatial location coordinate set of electrochemical interface degradation regions. Each coordinate point in this set is accompanied by its corresponding degradation degree level.

[0135] Step S176: Based on the preset three-dimensional structural model of the battery pack, the spatial positioning coordinate set of the weakened area of ​​the shell structure and the spatial positioning coordinate set of the deteriorated area of ​​the electrochemical interface are mapped to the same global three-dimensional coordinate system or a common two-dimensional planar layout projection map respectively; on the unified coordinate system or projection map, the mapped areas are subjected to spatial correlation matching processing; when the mapped range of the weakened area of ​​the shell structure and the mapped range of the deteriorated area of ​​the electrochemical interface have spatial overlap, a mechanical-electrochemical coupling deterioration collaborative early warning mark is generated.

[0136] A preset three-dimensional structural model of the battery pack is obtained, which defines the spatial geometric relationships of components such as the battery casing, electrode stacks, and electrolyte chamber. All coordinate points in the spatial location coordinate set of the weakened casing structure region generated in step S172 are transformed into the global three-dimensional coordinate system of this three-dimensional structural model. All coordinate points in the spatial location coordinate set of the electrochemical interface degradation region generated in step S175 are also transformed into the same global three-dimensional coordinate system. In the global three-dimensional coordinate system, the intersection of the mapped range of the weakened casing structure region and the mapped range of the electrochemical interface degradation region is calculated. If the volume of the intersection is greater than zero, it indicates that the two regions spatially overlap, meaning that the location of the weakened casing structure coincides with the location of the electrochemical interface degradation. Within this spatially overlapping region, the weakening of the casing structure may lead to electrolyte leakage or intrusion of external impurities, accelerating the degradation of the electrochemical interface; simultaneously, the gas or heat generated by the degradation of the electrochemical interface may exacerbate the weakening of the casing structure. The above spatial overlap indicates the existence of a mechano-electrochemical coupled degradation effect, generating a synergistic early warning indicator for mechano-electrochemical coupled degradation.

[0137] Step S177: Add the mechanical-electrochemical coupling degradation collaborative early warning identifier and spatial overlap coverage range coordinates to the abnormal physical field spatial positioning area, so that the generation process of the differentiated protection instruction set integrates the coupling degradation information of the battery casing mechanical structure state and the internal electrochemical interface state.

[0138] The mechano-electrochemical coupling degradation collaborative early warning identifier and the coordinate boundary of the spatial overlap coverage area generated in step S176 are added to the abnormal physical field spatial location area generated in step S144. The updated abnormal physical field spatial location area simultaneously includes the thermodynamic anomaly spatial location region, the acoustic anomaly spatial location region, the electrochemical anomaly spatial location region, the high-risk thermal runaway induction source region, and the mechano-electrochemical coupling degradation collaborative early warning region. In step S150, when generating the differentiated protection instruction set, based on the comprehensive information in the updated abnormal physical field spatial location area, the charge and discharge power boundaries are adjusted simultaneously, the directional thermal management loop is activated, and corresponding joint protection measures are taken for the shell structure weakening region and the electrochemical interface degradation region.

[0139] Step S180: After the differentiated protection instruction set is sent to the battery management control unit and the thermal management control unit, the recovery period thermal distribution sequence, recovery period acoustic response sequence and recovery period impedance spectrum evolution sequence of the new energy battery pack are continuously collected during the preset monitoring and recovery period.

[0140] After the differentiated protection command set generated in step S150 is sent to the battery management control unit and the thermal management control unit, step S180 initiates a preset monitoring recovery period. During this preset monitoring recovery period, the recovery period thermal distribution sequence, recovery period acoustic response sequence, and recovery period impedance spectrum evolution sequence of the new energy battery pack are continuously collected at the same sampling frequency and sampling accuracy as in step S110. The recovery period thermal distribution sequence is collected by a thermocouple array deployed on the surface of the battery electrode and at the electrode connection point; the recovery period acoustic response sequence is collected by a piezoelectric ceramic sensor array attached to the surface of the battery casing; and the recovery period impedance spectrum evolution sequence is collected by the electrochemical impedance spectroscopy monitoring module. All recovery period sequence data are accompanied by a high-precision synchronization time stamp.

[0141] Step S181: Perform the cross-field state correlation analysis on the recovery period thermal distribution sequence, the recovery period acoustic response sequence, and the recovery period impedance spectrum evolution sequence to generate the recovery period coupled physical field co-evolution structure.

[0142] The recovery-period thermodynamic distribution sequence, recovery-period acoustic response sequence, and recovery-period impedance spectrum evolution sequence acquired in step S180 are subjected to cross-field state correlation analysis processing according to the methods described in steps S121 to S129. Specifically, this includes: performing spatial interpolation reconstruction on the recovery-period thermodynamic distribution sequence to generate a recovery-period continuous temperature field distribution function; calculating the recovery-period tangential temperature gradient field and the recovery-period normal temperature gradient field; extracting the recovery-period heat flux vortex core line set and the recovery-period thermal barrier interface contour line set; and generating a recovery-period heat flux transfer topology map. Multi-scale time-frequency domain decomposition is performed on the recovery-period acoustic response sequence to extract the recovery-period group velocity dispersion curve family and the recovery-period phase velocity decay curve family; and constructing a recovery-period microstructure response path topology map. Relaxation time distribution deconvolution analysis is performed on the recovery-period impedance spectrum evolution sequence to generate a recovery-period ion transport interface state descriptor. Finally, the topology diagram of heat flow transfer during the recovery period, the topology diagram of microstructure response path during the recovery period, and the state descriptor of ion transport interface during the recovery period are subjected to tensor splicing and normalization fusion operations under a unified spatiotemporal framework to generate the cooperative evolution structure of coupled physical fields during the recovery period.

[0143] Step S182: Call the temporal causal convolutional Siamese network to process the recovery period coupled physical field co-evolution structure, generate the recovery period latent space trajectory manifold, and calculate the manifold coordinate offset vector of the recovery period latent space trajectory manifold and the latent space trajectory manifold at each corresponding time point in the manifold space.

[0144] The recovery-period coupled physics field co-evolution structure generated in step S181 is input into the trained temporal causal convolutional Siamese network described in steps S130 to S139. This temporal causal convolutional Siamese network performs the same feature extraction and dimensionality reduction processing on the recovery-period coupled physics field co-evolution structure, outputting the recovery-period latent space trajectory manifold. The recovery-period latent space trajectory manifold has the same time axis range and the same manifold space dimension as the latent space trajectory manifold generated in step S139. For each corresponding time point on the time axis, the vector difference between the coordinate points of the recovery-period latent space trajectory manifold and the coordinate points of the original latent space trajectory manifold is calculated to obtain the manifold coordinate offset vector at that time point. The formula for calculating the manifold coordinate offset vector is V_offset(t) = P_recovery(t) - P_original(t), where P_recovery(t) is the recovery-period manifold coordinate, and P_original(t) is the original manifold coordinate.

[0145] Step S183: Based on the trend of the magnitude change of the manifold coordinate offset vector, determine whether the execution of the differentiated protection instruction set causes the battery pack's safe state to evolve towards the center direction of the safe manifold sample cluster corresponding to the safe operating state mark.

[0146] For each time point obtained in step S182, the manifold coordinate offset vector is calculated, and the formula for calculating the modulus is |V_offset(t)|=sqrt(V_x^2+V_y^2+...), where V_x, V_y, etc., are the components of the offset vector in each manifold dimension. The modulus is arranged in chronological order to obtain the trend curve of the modulus change. At the same time, the distance D_recovery(t) between the latent space trajectory manifold coordinate point during the recovery period and the center of the safe manifold sample cluster defined in step S141, and the distance D_original(t) between the original latent space trajectory manifold coordinate point and the center of the safe manifold sample cluster are calculated. If, within the preset monitoring and recovery period, the D_recovery(t) sequence shows a monotonically decreasing trend, and the terminal value of the D_recovery(t) sequence is less than the terminal value of the D_original(t) sequence, while the |V_offset(t)| sequence shows a trend of first rapidly increasing and then gradually decreasing, then it is determined that the execution of the differentiated protection instruction set causes the battery pack's safety state to evolve towards the center of the safety manifold sample cluster, i.e., the protection instruction is effective.

[0147] Step S184: When the magnitude of the manifold coordinate offset vector shows a decreasing trend within a consecutive preset number of time windows and the distance between the endpoint manifold coordinate of the recovery period latent space trajectory manifold and the center of the safe manifold sample cluster is less than a preset recovery distance threshold, a valid feedback signal for the execution of the protection command is generated.

[0148] Extract the modulus values ​​corresponding to the last consecutive preset number of time windows from the modulus change trend curve calculated in step S183. Check whether these modulus values ​​decrease sequentially, i.e., the modulus value of each subsequent time window is less than the modulus value of the previous time window. Simultaneously, obtain the manifold coordinates of the end point (last time point) of the latent space trajectory manifold during the recovery period, and calculate the Euclidean distance between the end point manifold coordinates and the center of the safe manifold sample cluster. Compare this Euclidean distance with a preset recovery distance threshold. If the modulus values ​​of the consecutive preset number of time windows all show a decreasing trend, and the distance between the end point manifold coordinates and the center of the safe manifold sample cluster is less than the preset recovery distance threshold, it indicates that the battery pack's safety state has been fully restored to near a safe operating state. At this point, a valid feedback signal for the execution of the protection command is generated.

[0149] Step S185: Encapsulate the contents of the effective feedback signal of the protection command execution, the hidden space trajectory manifold during the recovery period, and the differentiated protection command set into a protection effect evaluation data record, and store the protection effect evaluation data record in the non-volatile storage area of ​​the battery management control unit.

[0150] Create a data structure for the protection effectiveness evaluation data record. This data record contains three main fields: the first field stores the valid feedback signal of the protection command execution generated in step S184, which is a Boolean value or status code; the second field stores the coordinate sequence data of the complete recovery period latent space trajectory manifold generated in step S182; the third field stores the complete content of the differentiated protection command set generated in step S150, including the charge / discharge power limiting ratio parameter, the path identifier of the target-oriented thermal management cycle path, the valve opening adjustment command, and the spatial coordinate coverage of the abnormal physical field spatial positioning area. This protection effectiveness evaluation data record is written to the non-volatile storage area of ​​the battery management control unit via a serial peripheral interface or internal bus. This non-volatile storage area can be embedded flash memory or electrically erasable programmable read-only memory to ensure that the data is not lost after the battery pack is powered off.

[0151] For example, step S190: obtain the set of historical hidden space trajectory manifold samples corresponding to multiple historical charge and discharge cycles recorded by the new energy battery pack within a preset historical operating cycle, and the battery pack health status level label corresponding to each historical hidden space trajectory manifold sample.

[0152] After generating the latent space trajectory manifold in step S130, step S190 retrieves a set of historical latent space trajectory manifold samples corresponding to multiple historical charge-discharge cycles recorded within a preset historical operating period from the historical database of the battery management system. Each historical latent space trajectory manifold sample is generated from the multi-source physical field response data stream of the previous historical charge-discharge cycle using the methods described in steps S110 to S139. Each historical latent space trajectory manifold sample is accompanied by a battery pack health status level marker, which is determined by the calibration data or historical maintenance records at the time of battery pack leaving the factory. The battery pack health status level markers include health status level 1, health status level 2, health status level 3, and health status level 4, corresponding to a battery pack capacity retention rate greater than 90%, a capacity retention rate between 80% and 90%, a capacity retention rate between 70% and 80%, and a capacity retention rate less than 70%, respectively.

[0153] Step S191: Call a preset manifold trajectory contrast learning network to perform cross-period similarity measurement processing on the latent space trajectory manifold generated in the current period and the historical latent space trajectory manifold sample set. The manifold trajectory contrast learning network includes a first temporal coding branch and a second temporal coding branch. The first temporal coding branch is used to extract the current period manifold feature codebook vector of the latent space trajectory manifold in the current period. The second temporal coding branch is used to extract the historical period manifold feature codebook vector of each historical sample in the historical latent space trajectory manifold sample set.

[0154] A pre-defined manifold trajectory contrastive learning network is constructed, comprising two structurally identical but independently updated temporal coding branches: a first temporal coding branch and a second temporal coding branch. Each temporal coding branch consists of three gated recurrent unit (ROU) layers and a fully connected output layer. The first temporal coding branch receives the coordinate sequence of the latent space trajectory manifold generated in the current period as input, with the shape of the coordinate sequence being (time step, manifold dimension). The three gated ROU layers process this sequence sequentially, each containing update and reset gates to control information flow through a gating mechanism. The hidden state of the last time step in the last gated ROU layer is input to the fully connected output layer, which maps this hidden state into a fixed-length vector, which is the manifold feature codebook vector for the current period. The second temporal coding branch processes each historical sample in the historical latent space trajectory manifold sample set in the same manner, outputting the historical period manifold feature codebook vector corresponding to each historical sample.

[0155] Step S192: In the codebook vector matching layer of the manifold trajectory contrast learning network, calculate the codebook vector cosine similarity value between the current periodic manifold feature codebook vector and each historical periodic manifold feature codebook vector in the historical periodic manifold feature codebook vector set. Based on the codebook vector cosine similarity value, select the top preset number of matching historical periodic manifold feature codebook vectors with the highest similarity value from the historical periodic manifold feature codebook vector set.

[0156] The current periodic manifold feature codebook vector output from the first temporal coding branch obtained in step S191 is denoted as Q, and all historical periodic manifold feature codebook vectors output from the second temporal coding branch are denoted as the set {H1, H2, H3, ..., Hn}. In the codebook vector matching layer, for each historical periodic manifold feature codebook vector Hi, the cosine similarity value between it and the current periodic manifold feature codebook vector Q is calculated. The formula for calculating the cosine similarity value is sim(Q, Hi) = (Q·Hi) / (||Q||*||Hi||), where Q·Hi is the dot product operation, and ||Q|| and ||Hi|| are the Euclidean norms of Q and Hi, respectively. All calculated cosine similarity values ​​are sorted from largest to smallest, and the top preset number of historical periodic manifold feature codebook vectors with the highest similarity values ​​are selected as the matching historical periodic manifold feature codebook vectors.

[0157] Step S193: Extract the battery pack health status level marker associated with the historical coupled physical field co-evolution structure sample set corresponding to the matching historical periodic manifold feature codebook vector, statistically analyze the category frequency distribution of the battery pack health status level marker, and determine the battery pack health status level marker with the highest frequency as the current health status reference level of the new energy battery pack.

[0158] For each matching historical periodic manifold feature codebook vector selected in step S192, its corresponding historical latent space trajectory manifold sample is traced, and then its corresponding historical coupled physical field co-evolution structure sample is traced through the historical latent space trajectory manifold sample. Finally, the battery pack health status level label associated with the historical coupled physical field co-evolution structure sample is obtained. The frequency distribution of the battery pack health status level labels corresponding to all matching historical periodic manifold feature codebook vectors is statistically analyzed, that is, the number of occurrences of health status level 1, health status level 2, health status level 3, and health status level 4 are counted. The battery pack health status level label with the highest frequency is determined as the current health status reference level of the new energy battery pack.

[0159] Step S194: Based on the current health status reference level, retrieve the evolution rate mapping table corresponding to the preset health status, extract the latent space trajectory manifold baseline evolution rate interval corresponding to the current health status reference level, and calculate the actual manifold distance cumulative change rate of the latent space trajectory manifold within the preset observation time window.

[0160] A pre-defined evolution rate mapping table is established, where each row corresponds to a battery pack health status reference level, including the lower and upper limits of the implicit space trajectory manifold baseline evolution rate interval associated with that level. The implicit space trajectory manifold baseline evolution rate interval is obtained through statistical analysis of historical manifold trajectories of a large number of battery packs with the same health level, representing the normal evolution rate range for that health level. Based on the current health status reference level determined in step S193, the corresponding implicit space trajectory manifold baseline evolution rate interval is retrieved from this mapping table. Simultaneously, a coordinate sequence within a pre-defined observation time window is extracted from the implicit space trajectory manifold generated in step S139, and the cumulative change in manifold distance from the start to the end of the time window is calculated. This cumulative change is divided by the length of the observation time window to obtain the actual cumulative change rate of manifold distance.

[0161] Step S195: Compare the actual manifold distance cumulative change rate with the implicit space trajectory manifold reference evolution rate range. When the actual manifold distance cumulative change rate exceeds the upper limit of the implicit space trajectory manifold reference evolution rate range, generate a battery pack accelerated aging warning indicator, and write the battery pack accelerated aging warning indicator and the current health status reference level together into the status monitoring log register of the battery management control unit.

[0162] The actual cumulative rate of change of manifold distance calculated in step S194 is compared with the upper and lower limits of the implicit space trajectory manifold baseline evolution rate interval. If the actual cumulative rate of change of manifold distance is greater than the upper limit of the baseline evolution rate interval, it indicates that the current safe state evolution rate of the battery pack is significantly faster than the normal evolution rate of a battery pack of the same health level, indicating accelerated aging. At this time, an accelerated aging warning flag is generated. This accelerated aging warning flag is a predefined flag. The generated accelerated aging warning flag and the current health state reference level determined in step S193 are written together into the status monitoring log register inside the battery management control unit. The status monitoring log register is a set of read-only registers inside the battery management control unit. External diagnostic tools can read the contents of this register through the controller area network bus for fault diagnosis and life prediction.

[0163] Step S196: When the actual manifold distance cumulative change rate is within the range of the latent space trajectory manifold baseline evolution rate, the current health status reference level is associated with the latent space trajectory manifold and stored in the non-volatile storage area of ​​the battery management control unit for updating the historical coupled physical field co-evolution structure sample set.

[0164] If the cumulative rate of change of the actual manifold distance calculated in step S194 is greater than or equal to the lower limit of the implicit space trajectory manifold baseline evolution rate interval and less than or equal to the upper limit of the baseline evolution rate interval, it indicates that the current battery pack's safety state evolution rate is within the normal range. At this time, the current health state reference level determined in step S193 is associated with and stored in conjunction with the implicit space trajectory manifold generated in step S139. Specifically, firstly, the implicit space trajectory manifold of the current period is traced back in reverse according to the methods of steps S110 to S139 to obtain the corresponding coupled physical field co-evolution structure. Then, this coupled physical field co-evolution structure, the implicit space trajectory manifold, and the current health state reference level are written as a new historical sample entry into the historical coupled physical field co-evolution structure sample set in the non-volatile storage area of ​​the battery management control unit. This sample is used for the health state assessment in subsequent steps S190 to S195, enabling continuous expansion and updating of the sample library.

[0165] In one exemplary embodiment, an artificial intelligence-based new energy battery safety monitoring system is provided, which may be a terminal, server, etc., and its internal structure diagram may be as follows. Figure 2As shown, the system specifically includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements an artificial intelligence-based method for monitoring the safety of new energy batteries. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of an AI-based new energy battery safety monitoring system, or an external keyboard, touchpad, or mouse, etc.

[0166] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A new energy battery safety monitoring method based on artificial intelligence, characterized in that, The method includes: Acquire a multi-source physical field response data stream deployed in a sensor array inside the battery pack. The multi-source physical field response data stream includes a thermodynamic distribution sequence, an acoustic response sequence, and an impedance spectrum evolution sequence with synchronous time stamps. Perform cross-field state correlation analysis processing on the multi-source physical field response data stream to construct a coupled physical field cooperative evolution structure; The temporal causal convolutional Siamese network is invoked to process the coupled physical field co-evolution structure, extract the long-term non-stationary evolution characteristics of the coupled physical field co-evolution structure, locate the corresponding causal transmission chain, and generate the latent space trajectory manifold characterizing the evolution trend of the safe state. Based on the projection offset direction and curvature abrupt change amplitude of the latent space trajectory manifold within the preset high-dimensional curvature boundary, the identification of the battery pack thermal runaway risk evolution stage and the associated abnormal physical field spatial location area are determined. A differentiated protection instruction set is generated based on the thermal runaway risk evolution stage identifier and the abnormal physical field spatial positioning area. The differentiated protection instruction set is used to adjust the charge and discharge power boundary and start the directional thermal management loop corresponding to the abnormal physical field spatial positioning area.

2. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, The step of performing cross-field state correlation analysis on the multi-source physical field response data stream to construct a coupled physical field cooperative evolution structure includes: Spatial interpolation reconstruction is performed on the temperature values ​​of each sensing node in the thermal distribution sequence to generate a continuous temperature field distribution function covering the geometric entity of the battery pack. Based on the continuous temperature field distribution function, the tangential temperature gradient field along the battery electrode stacking direction and the normal temperature gradient field perpendicular to the battery electrode stacking direction are calculated. Based on the curl distribution law of the tangential temperature gradient field, the core line set of heat flow vortex caused by non-uniform heat generation in the battery pack is extracted. Based on the divergence distribution law of the normal temperature gradient field, the profile line set of thermal barrier interface caused by the difference in thermal conduction between battery pack layers is extracted. The core line set of heat flow vortex and the profile line set of thermal barrier interface are spatially cross-linked and combined to generate a heat flow transfer topology map inside the battery pack with node connectivity strength as edge weight. Multi-scale time-frequency domain decomposition is performed on the acoustic response sequence to extract the group velocity dispersion curve family and phase velocity decay curve family corresponding to narrowband components of different frequencies. Based on the group velocity dispersion curve family and the phase velocity decay curve family, the acoustic waveguide propagation equation of the battery casing structure is established. The spatial distribution map of energy attenuation coefficient corresponding to each order of guided wave mode is analyzed from the acoustic waveguide propagation equation. Based on the steep change in the energy attenuation coefficient of the guided wave mode between adjacent sensing points in the spatial distribution map of the energy attenuation coefficient, the micro-damage sensitive area of ​​the shell structure is located. Based on the travel time offset of the reflected echo of each guided wave mode at the micro-damage sensitive area, a microstructure response path topology map is constructed with the sensing node as the vertex and the travel time offset as the edge vector. The microstructure response path topology map is used as the microstructure path feature of the battery shell. Perform a deconvolution analysis operation on the relaxation time distribution of the impedance spectrum evolution sequence to transform the broadband impedance imaginary part response to the relaxation time constant domain, generate a relaxation time distribution function containing a continuous relaxation peak distribution, and identify the high-frequency relaxation peak movement traces corresponding to the charge transfer process and the mid-frequency relaxation peak morphology distortion trend corresponding to the solid-phase diffusion process in the relaxation time distribution function. A charge transfer activated state drift vector is constructed along the time evolution direction based on the high-frequency relaxation peak position movement trace. A solid-phase diffusion path tortuosity change is constructed along the frequency evolution direction based on the mid-frequency relaxation peak position morphological distortion trend. The charge transfer activated state drift vector and the solid-phase diffusion path tortuosity change are fused to generate an ion transport interface state descriptor. The node coordinates of the heat flow transfer topology are mapped to a first spatiotemporal projection plane with time as the horizontal axis and battery spatial coordinates as the vertical axis. The node coordinates of the microstructure response path topology are mapped to a second spatiotemporal projection plane with time as the horizontal axis and shell surface arc length as the vertical axis. The feature points of the ion transport interface state descriptor are mapped to a third spatiotemporal projection plane with time as the horizontal axis and impedance imaginary part amplitude as the vertical axis. Based on the time dimension alignment reference between the first, second, and third spatiotemporal projection planes, the peak offset of the autocorrelation function within the sliding time window is calculated for the edge weight change sequence of the heat flow transfer topology graph, the edge vector evolution sequence of the microstructure response path topology graph, and the amplitude fluctuation sequence of the ion transport interface state descriptor, respectively, generating a characteristic time delay estimation array within each physical field; based on the characteristic time delay estimation array within each physical field, the alignment offset on the time axis between the characteristic sequences of different physical fields is calculated, generating a cross-physical field time alignment offset array; Based on the cross-physics field time alignment offset array, the edge weight change sequence of the heat flow transfer topology graph, the edge vector evolution sequence of the microstructure response path topology graph, and the amplitude fluctuation sequence of the ion transport interface state descriptor are time-axis aligned and corrected. The time-aligned and corrected edge weight change sequence, edge vector evolution sequence, and amplitude fluctuation sequence are then subjected to tensor splicing and normalization fusion operations under a unified spatiotemporal framework to generate a coupled physics field co-evolution structure indexed by three-dimensional spatial coordinates and one-dimensional time coordinates.

3. The artificial intelligence-based new energy battery safety monitoring method according to claim 2, characterized in that, The process of calling a temporal causal convolutional Siamese network to process the coupled physical field co-evolution structure, extracting the long-term non-stationary evolution features of the coupled physical field co-evolution structure, locating the corresponding causal transmission chain, and generating a latent space trajectory manifold characterizing the evolution trend of the safe state includes: The coupled physical field co-evolution structure is divided into multiple spatiotemporal state segment units with temporal continuity according to a preset time step. Independent normalization transformation operations are performed on the channel dimensions of the thermodynamic channel tensor, acoustic channel tensor, and electrochemical channel tensor in each spatiotemporal state segment unit. The standardized spatiotemporal state segment units are then sequentially input into the first layer of the temporal causal convolutional Siamese network. The first layer of the dilated causal convolutional network has a first dilation rate parameter and performs unidirectional causal convolution operations only along the time dimension. Local physical field evolution dependency features between adjacent time steps in the spatiotemporal state segment units are extracted to generate a first-level temporal feature mapping map. The first-level temporal feature map is input into the second-level dilated causal convolution computation layer of the temporal causal convolutional Siamese network. The second-level dilated causal convolution computation layer has a second dilation parameter that is greater than the first dilation parameter and performs unidirectional causal convolution operation along the time dimension. The mid-range physical field evolution dependency features spanning multiple time steps in the first-level temporal feature map are extracted to generate the second-level temporal feature map. The second-level temporal feature map is input into the third-level dilated causal convolution computation layer of the temporal causal convolutional Siamese network. The third-level dilated causal convolution computation layer has a third dilation parameter greater than the second dilation parameter and performs unidirectional causal convolution operation along the time dimension. The remote physical field evolution dependency features spanning long temporal intervals in the second-level temporal feature map are extracted to generate the third-level temporal feature map. The first-level temporal feature map, the second-level temporal feature map and the third-level temporal feature map are cascaded and merged along the feature channel dimension to generate a multi-scale long-term non-stationary evolution feature set. The multi-scale long-term non-stationary evolution feature set is input into the differential attention weighted aggregation layer of the temporal causal convolutional Siamese network. In the differential attention weighted aggregation layer, the feature difference residual map of the multi-scale long-term non-stationary evolution feature set on the adjacent frame on the time axis is calculated. Based on the feature difference residual map of the adjacent frame, the channel attention weight distribution vector is calculated along the feature channel dimension. The channel attention weight distribution vector is used to perform weighted enhancement processing on the multi-scale long-term non-stationary evolution feature set, thereby increasing the activation response value of the channel corresponding to the gradient abrupt change region inside the heat flow transfer topology map, increasing the activation response value of the channel corresponding to the transient singular response perturbation in the microstructure path feature, and increasing the activation response value of the channel corresponding to the spectral response drift trend in the ion transport interface state descriptor, thus generating a weighted feature map set that enhances causal conduction features. Parallel pooling operations of spatial dimension global average pooling and global max pooling are performed on the weighted feature map set. The parallel pooling results are concatenated along the feature channel dimension to generate a spatially compressed feature vector. The spatially compressed feature vector is input into the fully connected mapping layer of the temporal causal convolutional Siamese network to generate the causal transmission chain feature vector corresponding to the first Siamese branch and the causal transmission chain feature vector corresponding to the second Siamese branch. Calculate the contrast loss function value between the causal transmission chain feature vector corresponding to the first twin branch and the causal transmission chain feature vector corresponding to the second twin branch. Based on the contrast loss function value, update the network weight parameters of the dilated causal convolution calculation layer and the differential attention weighted aggregation layer in reverse, so that the feature vector distance between similar physical field evolution modes is reduced and the feature vector distance between dissimilar physical field evolution modes is increased. The causal transmission chain feature vectors corresponding to the first twin branch and the causal transmission chain feature vectors corresponding to the second twin branch are input into the manifold learning dimensionality reduction layer of the temporal causal convolutional twin network. The manifold learning dimensionality reduction layer uses the local linear embedding algorithm to project the high-dimensional causal transmission chain feature vectors onto the low-dimensional manifold space, generating a low-dimensional manifold coordinate sequence that preserves the local neighborhood structure relationship and the global topological connection relationship. A latent space trajectory manifold is constructed based on the continuous change trajectory of the low-dimensional manifold coordinate sequence in the time dimension. The manifold distance between any adjacent time points in the latent space trajectory manifold is used to characterize the evolution rate of the battery pack's safety state, and the local curvature value of the latent space trajectory manifold is used to characterize the nonlinear severity of the evolution of the battery pack's safety state.

4. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, The generation of a differentiated protection instruction set based on the thermal runaway risk evolution stage identifier and the abnormal physical field spatial location area includes: Based on the thermal runaway risk evolution stage identifier, a preset stage classification strategy mapping table is retrieved to extract the charge / discharge power limiting ratio parameter, thermal management cycle start threshold parameter, and response priority code value corresponding to the thermal runaway risk evolution stage identifier; Write the charge / discharge power limiting ratio parameter into the power boundary configuration register of the battery management control unit, so that the battery management control unit recalculates the maximum allowable charging current limit and the maximum allowable discharging current limit in the current charge / discharge cycle based on the charge / discharge power limiting ratio parameter. Based on the spatial coordinate coverage of the abnormal physical field spatial positioning area in the three-dimensional solid structure of the battery pack, the path identifier of at least one target-oriented thermal management circulation path covering the spatial coordinate coverage area and the address code of the flow regulating valve corresponding to the target-oriented thermal management circulation path are retrieved from the preset cooling circuit topology mapping table. The thermal management cycle start threshold parameter is written into the temperature threshold comparison register of the thermal management control unit. When the real-time temperature value collected by the temperature sensing node deployed in the abnormal physical field spatial positioning area exceeds the thermal management cycle start threshold parameter, the thermal management control unit generates a cooling medium cycle start trigger signal. A valve opening adjustment command is generated based on the cooling medium circulation start-up trigger signal and the address code of the flow regulating valve. The valve opening adjustment command is used to adjust the valve opening of the flow regulating valve corresponding to the target directional thermal management circulation path to a preset maximum opening value. Based on the response priority encoding value, the charge / discharge power limiting ratio parameter, the path identifier of the target directional thermal management circulation path, the valve opening adjustment command, and the spatial coordinate coverage of the abnormal physical field spatial positioning area are encapsulated into a differentiated protection command data frame. The differentiated protection command data frame is sent to the battery management control unit and the thermal management control unit with a bus arbitration priority corresponding to the response priority encoding value, triggering the charge and discharge power boundary adjustment operation and the directional thermal management loop start operation.

5. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, After generating the latent space trajectory manifold characterizing the evolution trend of the security state, the method further includes: Multiple historical latent space trajectory manifold samples recorded during the historical charge-discharge cycles of the new energy battery pack are obtained. Each historical latent space trajectory manifold sample corresponds to a known safe operating state category label or a known thermal runaway precursor state category label. Manifold space similarity measurement processing is performed on the latent space trajectory manifold and the multiple historical latent space trajectory manifold samples. The first feature distance value between the latent space trajectory manifold and the center of the safe manifold sample cluster formed by all historical latent space trajectory manifold samples marked as safe operating states is calculated. The second feature distance value between the latent space trajectory manifold and the center of the risk manifold sample cluster formed by all historical latent space trajectory manifold samples marked as thermal runaway precursor states is calculated. When the second feature distance value is less than the first feature distance value, the latent space trajectory manifold is classified as a potential thermal runaway evolution trajectory category, and the manifold segment interval in the latent space trajectory manifold where the local curvature value exceeds the preset curvature threshold is extracted. Backtrack to locate the time axis interval corresponding to the manifold segment interval, and analyze the feature activation contribution ranking of thermodynamic channel tensor, acoustic channel tensor and electrochemical channel tensor in the coupled physical field co-evolution structure within the time axis interval. Extract the thermodynamic channel index, acoustic channel index and electrochemical channel index that rank first in the feature activation contribution ranking. The spatial location region of thermodynamic anomaly is determined based on the spatial mapping relationship of the thermodynamic channel index in the three-dimensional solid structure of the battery pack; the spatial location region of acoustic anomaly is determined based on the spatial mapping relationship of the acoustic channel index in the sensor network on the surface of the battery casing; the spatial location region of electrochemical anomaly is determined based on the spatial mapping relationship of the electrochemical channel index in the battery electrode stack; and the spatial location regions of thermodynamic anomaly, acoustic anomaly, and electrochemical anomaly are fused to generate an abnormal physical field spatial location region. The thermal runaway risk evolution rate coefficient is calculated based on the cumulative change of manifold distance in the manifold segment interval, and the thermal runaway risk severity coefficient is calculated based on the local curvature peak of the manifold segment interval. The thermal runaway risk evolution rate coefficient and the thermal runaway risk severity coefficient are input into a preset risk stage division mapping function to generate a corresponding potential thermal runaway risk evolution stage category identifier. Based on the category identifier of the potential thermal runaway risk evolution stage, a preset graded protection strategy mapping table is retrieved, and the charge and discharge power limit ratio parameter, thermal management cycle start threshold parameter and alarm notification target address list corresponding to the category identifier of the potential thermal runaway risk evolution stage are extracted. The charge / discharge power limit ratio parameter is written into the power boundary register of the battery management control unit to modify the upper and lower limits of the charge / discharge power boundary envelope. The thermal management cycle start threshold parameter is written into the threshold comparator of the thermal management control unit to trigger the adjustment of the circulation flow rate of the cooling medium in the directional thermal management loop corresponding to the abnormal physical field spatial positioning area. Based on the alarm notification target address list, the potential thermal runaway risk evolution stage category identifier, the abnormal physical field spatial positioning area, and the thermal runaway risk evolution rate coefficient are encapsulated into a preset format safety warning data message and sent to the vehicle central gateway controller and remote cloud monitoring service platform. The changes in the thermal distribution sequence, acoustic response sequence, and impedance spectrum evolution sequence of the new energy battery pack within a preset monitoring time window after the execution of the differentiated protection instruction set are recorded and fed back to the incremental learning sample buffer of the temporal causal convolutional twin network for online optimization of the network weight parameters of the temporal causal convolutional twin network.

6. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, After acquiring the multi-source physical field response data stream deployed within the sensor array inside the battery pack, the method further includes: The multi-source physical field response data stream is processed by time segmentation, and divided into multiple thermodynamic time segment units, multiple acoustic time segment units, and multiple impedance spectrum time segment units with the same time start and end boundaries according to a preset sliding time window length. For the thermal time series segment units within the same time window, spatial graph structure construction processing is performed. The correlation coefficient value between the temperature time series of any two spatial sensing nodes in the thermal time series segment unit is calculated. Based on the correlation coefficient value, a thermal spatial correlation graph structure with sensing nodes as vertices and correlation coefficient values ​​as edge weights is constructed. For acoustic wave time sequence segment units within the same time window, perform spatial graph structure construction processing, calculate the correlation coefficient between the guided wave response time of any two spatial sensing nodes in the acoustic wave time sequence segment unit, and construct an acoustic wave spatial correlation graph structure with sensing nodes as vertices and correlation coefficient values ​​as edge weights based on the correlation coefficient values. For impedance spectrum timing segment units within the same time window, frequency graph structure construction processing is performed. The correlation coefficient between the imaginary part of the impedance at any two frequency sampling points in the impedance spectrum timing segment unit is calculated. Based on the correlation coefficient, a frequency correlation graph structure is constructed with frequency points as vertices and correlation coefficient values ​​as edge weights. The thermal spatial correlation graph structure, the acoustic spatial correlation graph structure, and the frequency correlation graph structure within the same time window are combined into a multi-view correlation graph network. The multi-view correlation graph network includes a thermal view adjacency matrix, an acoustic view adjacency matrix, a frequency view adjacency matrix, and a set of cross-view connection edge weights connecting different view nodes. A preset multi-view graph neural network is invoked to perform multi-round graph convolution message passing and cross-view feature fusion operations on the multi-view association graph network. The features of the thermal view nodes are propagated along the thermal view adjacency matrix and weighted and fused with the features of the acoustic wave view nodes and the frequency view nodes through the cross-view connection edge weight set to generate the updated feature vectors of the thermal view nodes, the updated feature vectors of the acoustic wave view nodes, and the updated feature vectors of the frequency view nodes. The feature vectors of all thermal view nodes, all acoustic view nodes, and all frequency view nodes are concatenated along the vector dimension to generate a multi-view fusion graph representation vector sequence. This multi-view fusion graph representation vector sequence replaces the thermal distribution sequence, the acoustic response sequence, and the impedance spectrum evolution sequence.

7. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, After generating the latent space trajectory manifold characterizing the evolution trend of the security state, the method further includes: The new energy battery pack acquires multiple historical hidden space trajectory manifold samples recorded in historical charge-discharge cycles and a safety state category label corresponding to each historical hidden space trajectory manifold sample. The safety state category label includes a safe operation state label and a thermal runaway precursor state label. Calculate the first manifold spacing value between the latent space trajectory manifold and the center of the safe manifold sample cluster composed of all historical latent space trajectory manifold samples marked as safe operating states; calculate the second manifold spacing value between the latent space trajectory manifold and the center of the risk manifold sample cluster composed of all historical latent space trajectory manifold samples marked as thermal runaway precursor states; when the second manifold spacing value is less than the first manifold spacing value, extract the manifold segment interval in the latent space trajectory manifold where the local curvature value exceeds the preset curvature threshold. Backtrack to locate the time axis interval corresponding to the manifold segment interval, and analyze the characteristic activation contribution values ​​of the thermodynamic channel tensor, acoustic channel tensor and electrochemical channel tensor in the coupled physical field co-evolution structure within the time axis interval; The thermodynamic channel index, acoustic channel index, and electrochemical channel index are extracted and ranked in the top preset position in the feature activation contribution value sorting. The thermodynamic anomaly space region is determined according to the mapping relationship of the thermodynamic channel index in the three-dimensional space of the battery pack, the acoustic anomaly space region is determined according to the mapping relationship of the acoustic channel index on the surface of the battery shell, and the electrochemical anomaly space region is determined according to the mapping relationship of the electrochemical channel index in the battery electrode stack. The anomalous physical field spatial location region is generated by integrating the thermodynamic anomalous spatial region, the acoustic anomalous spatial region, and the electrochemical anomalous spatial region. The thermal runaway risk evolution stage is identified based on the cumulative change in manifold distance and the local curvature peak value of the manifold segment interval.

8. The artificial intelligence-based new energy battery safety monitoring method according to claim 1, characterized in that, The method further includes: After the differentiated protection instruction set is sent to the battery management control unit and the thermal management control unit, the recovery period thermal distribution sequence, recovery period acoustic response sequence and recovery period impedance spectrum evolution sequence of the new energy battery pack are continuously collected during the preset monitoring and recovery period. The cross-field state correlation analytical processing is performed on the recovery period thermal distribution sequence, the recovery period acoustic response sequence, and the recovery period impedance spectrum evolution sequence to generate a recovery period coupled physical field co-evolution structure; The temporal causal convolutional Siamese network is invoked to process the recovery period coupled physical field co-evolution structure, generating a recovery period latent space trajectory manifold, and calculating the manifold coordinate offset vector between the recovery period latent space trajectory manifold and the latent space trajectory manifold at each corresponding time point in the manifold space; Based on the trend of the magnitude change of the manifold coordinate offset vector, it is determined whether the execution of the differentiated protection instruction set causes the battery pack's safety state to evolve towards the center direction of the safety manifold sample cluster corresponding to the safety operating state mark. When the magnitude of the manifold coordinate offset vector decreases within a consecutive preset number of time windows and the distance between the endpoint manifold coordinates of the recovery period latent space trajectory manifold and the center of the safe manifold sample cluster is less than a preset recovery distance threshold, a valid feedback signal for the execution of the protection command is generated. The effective feedback signal of the protection command execution, the hidden space trajectory manifold during the recovery period, and the contents of the differentiated protection command set are encapsulated into a protection effect evaluation data record, and the protection effect evaluation data record is stored in the non-volatile storage area of ​​the battery management control unit.

9. A new energy battery safety monitoring system based on artificial intelligence, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the artificial intelligence-based new energy battery safety monitoring method according to any one of claims 1 to 8 by executing the machine-executable instructions.

10. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. The processor of the artificial intelligence-based new energy battery safety monitoring system reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the artificial intelligence-based new energy battery safety monitoring system to perform the artificial intelligence-based new energy battery safety monitoring method as described in any one of claims 1 to 8.

Images