A lithium battery state of health early warning method and system based on multi-sensor fusion

Abstract

The application provides a lithium battery health state early warning method and system based on multi-sensor fusion, and relates to the technical field of lithium battery health state monitoring and early warning. The method comprises the following steps: collecting multi-source sensing signals of a lithium battery in a strong dynamic working condition in real time, including electric signals, thermal signals and vibration signals; performing time synchronization and space alignment processing on the multi-source sensing signals to obtain an initial multi-dimensional sensing sequence; constructing a non-steady multi-physical field covariant decoupling model according to the initial multi-dimensional sensing sequence; performing motion trajectory reconstruction and multi-axis inertia moment orthogonal decomposition by the non-steady multi-physical field covariant decoupling model to remove dynamic disturbances generated by the superposition of pulse current and mechanical vibration, and obtaining a pure electrochemical characteristic signal set. The application can solve the problems of fault feature extraction distortion, high false alarm rate and high missed alarm rate under a strong dynamic coupling working condition, and can be stably generalized across cells, platforms and scenes.

Description

Technical Field

[0001] This invention relates to the field of lithium battery health status monitoring and early warning technology, and in particular to a lithium battery health status early warning method and system based on multi-sensor fusion. Background Technology

[0002] With the rapid advancement of electrification of non-road mobile equipment, commercialization of low-altitude transport devices, and large-scale deployment of specialized field energy storage platforms, lithium batteries, as the core power source and energy storage carrier, are increasingly expanding their application scenarios from conventional room temperature environments to highly dynamic coupled conditions involving high-rate pulse discharge, wide temperature range drastic changes, and strong mechanical vibrations. In these applications, real-time and accurate sensing of lithium battery health status has become a critical technological requirement for ensuring the safe and reliable operation of equipment.

[0003] Currently, mainstream lithium battery health status early warning technologies in the industry typically involve feature extraction and threshold determination of electrical, thermal, and vibration signals separately, or the use of machine learning models calibrated under fixed operating conditions for fault classification. However, most of these solutions suffer from the following technical shortcomings: they fail to establish a unified mapping model for operating condition disturbances, sensing features, and electrochemical degradation. For example, under the aforementioned highly dynamic coupled operating conditions, lithium batteries are subjected to alternating effects of pulse current impacts, wide-temperature-range thermal drift, and wide-frequency mechanical vibrations over long periods. Multiple disturbance components in the sensing signals superimpose and become deeply entangled with electrochemical degradation features. Due to the lack of such a unified mapping model, the sensing signals under highly dynamic conditions are subject to coupling interference from current impacts, temperature drift, and mechanical vibrations, leading to severe distortion in the extraction of fault features corresponding to electrochemical degradation. Furthermore, the difficulty in effectively distinguishing between operating condition disturbance signals and true electrochemical degradation signals results in persistently high false alarm and false alarm rates in traditional early warning systems. Summary of the Invention

[0004] This invention provides a lithium battery health status early warning method and system based on multi-sensor fusion, which can solve the problems of fault feature extraction distortion, false alarm rate and missed alarm rate under strong dynamic coupling conditions, and can be stably generalized across cells, platforms and scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, a method for early warning of the health status of lithium batteries based on multi-sensor fusion is provided, the method comprising:

[0007] Step 1: Real-time acquisition of multi-source sensor signals from the lithium battery under strong dynamic conditions, including electrical signals, thermal signals, and vibration signals; time synchronization and spatial alignment processing of the multi-source sensor signals to obtain an initial multi-dimensional sensor sequence.

[0008] Step 2: Based on the initial multidimensional sensing sequence, construct an unsteady multi-physics covariant decoupling model; based on the unsteady multi-physics covariant decoupling model, perform motion trajectory reconstruction and multi-axis inertial torque orthogonal decomposition to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, and obtain a pure electrochemical characteristic signal set.

[0009] Step 3: Based on the pure electrochemical characteristic signal set, and combined with the pre-set interface solvation potential well and stress relaxation coupling mechanism, construct a unified mapping network from dynamic operating conditions to electrochemical degradation state.

[0010] Step 4: Input the pure electrochemical feature signal set into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to the coordination field distortion and molecular potential field fluctuation, and obtain the deterioration feature vector set characterizing lithium plating, internal short circuit, tab aging, micro short circuit and seal failure.

[0011] Step 5: Perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status warning instruction.

[0012] Furthermore, multi-source sensor signals from the lithium battery under strong dynamic operating conditions are acquired in real time, including electrical signals, thermal signals, and vibration signals; the multi-source sensor signals are processed for time synchronization and spatial alignment to obtain an initial multi-dimensional sensing sequence, including:

[0013] Under strong dynamic conditions, the sensor locations corresponding to the electrical signal, thermal signal and vibration signal are mapped to a pre-established three-dimensional contour model of the battery cell through coordinate transformation, forming a set of dynamic monitoring nodes that spatially correspond to the electrode stacking structure inside the battery cell; the strong dynamic conditions include one or more of pulse current impact, wideband mechanical vibration and wide temperature range thermal drift.

[0014] Based on the dynamic monitoring node set, a multi-source field response topology cluster conforming to the stacking direction of the battery cell electrode is constructed, wherein the local response configuration of the multi-source field response topology cluster is kept in shape consistent with the physical layering of the battery cell.

[0015] The local interface stress distortion curvature at each node on the multi-source field response topology cluster is extracted, as well as the geodesic deviation between nodes along the potential field dissipation path. Based on the multi-source signal values ​​aligned with the pre-acquired timestamps, an initial multidimensional sensing sequence with unified spatiotemporal coordinates is formed on the topology cluster.

[0016] Furthermore, based on the initial multidimensional sensing sequence, an unsteady multi-physics covariant decoupling model is constructed. Using this model, motion trajectory reconstruction and orthogonal decomposition of multi-axis inertial torques are performed to remove dynamic disturbances caused by the superposition of pulse current and mechanical vibration, resulting in a pure electrochemical characteristic signal set, including:

[0017] The covariant temporal sequence of electrical signal components, thermal signal components and vibration signal components in a unified spatiotemporal coordinate is separated from the initial multidimensional sensing sequence. A multiphysics covariant tensor space is constructed to characterize the mutual coupling relationship between electrochemical response, thermal response and mechanical response. The multiphysics covariant tensor space is used as the solution basis of the unsteady multiphysics covariant decoupling model.

[0018] Based on the solution basis, the dynamic disturbance response trajectory excited by pulsed current impact and broadband mechanical vibration is traced in the multi-physics covariant tensor space to obtain the dynamic disturbance response trajectory; based on the dynamic disturbance response trajectory, multi-axis inertial torque orthogonal decomposition is performed along the electrochemical response principal axis and the mechanical vibration principal axis to obtain the independent projection of the dynamic disturbance component on each physical field principal axis;

[0019] From the initial multidimensional sensing sequence, the independent projections of the dynamic disturbance components on the principal axes of each physical field are filtered out to obtain a set of pure electrochemical characteristic signals stripped of operating condition coupling noise.

[0020] Furthermore, based on the pure electrochemical characteristic signal set, and combined with the pre-set interface solvation potential well and stress relaxation coupling mechanism, a unified mapping network from dynamic operating conditions to electrochemical degradation state is constructed, including:

[0021] The pre-set interface solvation potential well and stress relaxation coupling mechanism are obtained, and the pure electrochemical characteristic signal set is expanded under the coupling mechanism to analyze the evolution path of the intermolecular interaction potential well of the solvation sheath at the electrode interface and the temporal correlation of lattice stress relaxation, so as to obtain the interface potential well-stress relaxation temporal correlation mapping.

[0022] Based on the interface potential well-stress relaxation time-series correlation mapping, an interface potential well-stress relaxation constitutive relation characterizing the intrinsic response of electrochemical degradation is established, and this interface potential well-stress relaxation constitutive relation is used as the internal mapping kernel.

[0023] Based on the internal mapping kernel, a unified mapping network framework is constructed. In the unified mapping network framework, the impact of dynamic operating condition disturbances is encoded as boundary condition constraints, and the electrochemical degradation state is encoded as output layer state parameters.

[0024] The pure electrochemical feature signal set is input into a unified mapping network framework, and the internal mapping kernel projects the features onto it to obtain the feature projection result. The output layer state parameters are used to perform a nonlinear transformation on the feature projection result, so that the unified mapping network framework establishes the mapping relationship from dynamic operating condition disturbance to electrochemical degradation state, and the unified mapping network from dynamic operating condition to electrochemical degradation state is constructed.

[0025] Furthermore, the pure electrochemical feature signal set is input into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to ligand field distortion and molecular potential field fluctuations, thereby obtaining a set of deterioration feature vectors characterizing lithium plating, internal short circuit, tab aging, micro-short circuit, and potential seal failure, including:

[0026] The pure electrochemical feature signal set is input into the unified mapping network of the dynamic operating condition to the electrochemical degradation state. The unified mapping network calls the internal mapping kernel to perform feature projection on the pure electrochemical feature signal set, so as to map the pure electrochemical feature signal set to the hidden feature space with the coordinate field distortion and molecular potential field fluctuation as the coordinate axis, and obtain the hidden layer electrochemical degradation feature projection.

[0027] Based on the projection of the electrochemical degradation characteristics of the hidden layer, the output layer state parameters are nonlinearly transformed to obtain a nonlinear transformation response tensor. Based on the nonlinear transformation response tensor, electrochemical response characteristic components are extracted from the tensor along the coordination field distortion axis and the molecular potential field fluctuation axis of the hidden feature space to obtain an electrochemical degradation response characteristic set.

[0028] From the electrochemical degradation response feature set, the first feature subset associated with lithium plating, the second feature subset associated with internal short circuit, the third feature subset associated with tab aging, the fourth feature subset associated with micro-short circuit, and the fifth feature subset associated with sealing failure are separated and combined to form a degradation feature vector group.

[0029] Furthermore, the degraded feature vector group is subjected to cross-domain manifold covariance measurement with a preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; based on the cross-domain covariance measurement result, the deviation quantization value is calculated, including:

[0030] Based on the degraded feature vector group, a degraded feature manifold is constructed, and a preset cross-scene generalization threshold library is obtained. The preset cross-scene generalization threshold library stores a cross-scene reference manifold. The degraded feature manifold and the cross-scene reference manifold are registered together to establish a cross-domain manifold covariant correspondence and obtain a cross-domain manifold registration mapping.

[0031] Based on the cross-domain manifold registration mapping, the local manifold distance between each node on the degraded feature manifold and the corresponding node on the cross-scene benchmark manifold is calculated to obtain the node-level covariance deviation value.

[0032] Statistical aggregation is performed on all node-level covariance deviation values ​​to obtain cross-domain covariance measurement results; based on the cross-domain covariance measurement results, the deviation quantization value between the degraded feature vector group and the preset cross-scenario generalization threshold library is calculated.

[0033] Furthermore, based on the deviation quantification value, the early degradation type and health degradation degree of the lithium battery are determined, and a health status early warning instruction is obtained, including:

[0034] From the deviation quantification value, a first deviation component associated with lithium plating, a second deviation component associated with internal short circuit, a third deviation component associated with tab aging, a fourth deviation component associated with micro short circuit, and a fifth deviation component associated with seal failure are decomposed to obtain a set of categorized deviation quantities.

[0035] Each deviation component in the categorized deviation set is compared with the warning judgment threshold of the corresponding hidden danger type in the preset cross-scenario generalization threshold library to determine the confidence level of each hidden danger type, and the early deterioration type is determined based on the confidence level.

[0036] The evolution trend of the deviation component corresponding to the early deterioration type is analyzed to obtain the health decline characterization parameter; the health decline degree index is calculated based on the health decline characterization parameter; and a health status early warning instruction is obtained based on the early deterioration type and the health decline degree index.

[0037] Secondly, a lithium battery health status early warning system based on multi-sensor fusion includes:

[0038] The spatiotemporal alignment module is used to acquire multi-source sensing signals of lithium batteries under strong dynamic conditions in real time, including electrical signals, thermal signals and vibration signals; and to perform time synchronization and spatial alignment processing on the multi-source sensing signals to obtain an initial multidimensional sensing sequence.

[0039] The covariant decoupling module is used to construct an unsteady multi-physics covariant decoupling model based on the initial multi-dimensional sensing sequence. The unsteady multi-physics covariant decoupling model is used to reconstruct the motion trajectory and perform orthogonal decomposition of multi-axis inertial torque to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, thereby obtaining a pure electrochemical characteristic signal set.

[0040] The mapping network construction and feature extraction module is used to construct a unified mapping network from dynamic operating conditions to electrochemical degradation state based on a pure electrochemical feature signal set and a pre-set interface solvation potential well and stress relaxation coupling mechanism. The pure electrochemical feature signal set is input into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to ligand field distortion and molecular potential field fluctuations, thereby obtaining a set of degradation feature vectors characterizing lithium plating, internal short circuit, tab aging, micro short circuit and seal failure risks.

[0041] The measurement and early warning module is used to perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status early warning instruction.

[0042] The above-described solution of the present invention has at least the following beneficial effects:

[0043] By employing spatiotemporal alignment and topological configuration techniques for multi-source sensing signals, electrical, thermal, and vibration signals are mapped onto a multi-source field response topology cluster that conforms to the physical layering of the battery cell, forming an initial multidimensional sensing sequence with unified spatiotemporal coordinates. This overcomes the problems of spatial misalignment and temporal mismatch of multi-source signals under strong dynamic conditions, and achieves precise spatial correspondence between sensing information and the internal structure of the battery cell.

[0044] Based on this, an unsteady multi-physics field covariant decoupling method is adopted to reconstruct the trajectory and orthogonally decompose the dynamic disturbance generated by the superposition of pulse current impact and broadband mechanical vibration, filter out the operating condition coupling noise, and obtain a pure electrochemical feature signal set. This overcomes the problem of deep entanglement between the operating condition disturbance component and the electrochemical degradation feature, and realizes high-fidelity extraction of fault features.

[0045] Furthermore, by combining a pre-set interface solvation potential well with a stress relaxation coupling mechanism, a unified mapping network is constructed to cover five types of hidden dangers, including lithium plating, internal short circuit, tab aging, micro short circuit, and seal failure, from operating condition disturbances to electrochemical degradation. The pure electrochemical characteristic signal set is subjected to feature projection and nonlinear transformation to obtain a set of degradation feature vectors. This overcomes the problem in existing technologies where the lack of a unified mapping model leads to the inability to distinguish between operating condition disturbances and degradation signals, and achieves synchronous and accurate characterization of multiple hidden dangers.

[0046] Finally, a cross-domain manifold covariance metric and deviation quantization method is adopted to register and compare the degradation feature vector set with the cross-scenario benchmark manifold to determine the early degradation type and the degree of health decay. This overcomes the problems of traditional methods relying on fixed operating conditions for calibration and poor cross-scenario generalization ability. It achieves the technical effect of effectively reducing false alarm rate and missed alarm rate under strong dynamic coupling conditions, and realizing stable generalization across cells, platforms and scenarios. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a lithium battery health status early warning method based on multi-sensor fusion, provided by an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of a lithium battery health status early warning system based on multi-sensor fusion, provided by an embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0050] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a method for early warning of the health status of lithium batteries based on multi-sensor fusion, the method comprising the following steps:

[0051] Step 1: Real-time acquisition of multi-source sensor signals from the lithium battery under strong dynamic conditions, including electrical signals, thermal signals, and vibration signals; time synchronization and spatial alignment processing of the multi-source sensor signals to obtain an initial multi-dimensional sensor sequence.

[0052] Step 2: Based on the initial multidimensional sensing sequence, construct an unsteady multi-physics covariant decoupling model; based on the unsteady multi-physics covariant decoupling model, perform motion trajectory reconstruction and multi-axis inertial torque orthogonal decomposition to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, and obtain a pure electrochemical characteristic signal set.

[0053] Step 3: Based on the pure electrochemical characteristic signal set, and combined with the pre-set interface solvation potential well and stress relaxation coupling mechanism, construct a unified mapping network from dynamic operating conditions to electrochemical degradation state.

[0054] Step 4: Input the pure electrochemical feature signal set into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to the coordination field distortion and molecular potential field fluctuation, and obtain the deterioration feature vector set characterizing lithium plating, internal short circuit, tab aging, micro short circuit and seal failure.

[0055] Step 5: Perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status warning instruction.

[0056] In this embodiment of the invention, by synchronizing and aligning multi-source sensing signals in time and space to form an initial multidimensional sensing sequence, an unsteady multi-physics covariant decoupling model is constructed. The dynamic disturbance is reconstructed and orthogonally decomposed, which can effectively remove the operating condition coupling noise generated by the superposition of pulse current and mechanical vibration, and obtain a pure electrochemical characteristic signal set. On this basis, combined with the interface solvation potential well and stress relaxation coupling mechanism, a unified mapping network from dynamic operating conditions to electrochemical degradation is established. The pure characteristic signals are subjected to feature projection and nonlinear transformation, and the electrochemical response features corresponding to the coordination field distortion and molecular potential field fluctuation are accurately extracted. The degradation feature vector set characterizing multiple hidden dangers such as lithium plating, internal short circuit, tab aging, micro short circuit and sealing failure is obtained at one time.

[0057] Finally, by combining the degradation feature vector group with the cross-scenario generalization threshold library to perform cross-domain manifold covariance measurement and calculate the deviation quantization value, a reliable determination of early degradation type and health decay degree under strong dynamic coupling conditions is achieved, effectively reducing false alarm rate and missed alarm rate, and possessing strong generalization capability across cells, platforms, and scenarios.

[0058] In a preferred embodiment of the present invention, step 1 above may include:

[0059] Step 1.1: Under strong dynamic conditions, the sensor placement positions corresponding to the electrical, thermal, and vibration signals are mapped to a pre-established three-dimensional contour model of the battery cell through coordinate transformation, forming a set of dynamic monitoring nodes spatially corresponding to the internal electrode stacking structure of the battery cell. The strong dynamic conditions include one or more of pulse current impact, wideband mechanical vibration, and wide temperature range thermal drift. Specifically, under strong dynamic conditions, a pre-established three-dimensional contour model of the lithium battery cell is first acquired. The model establishes a spatial rectangular coordinate system with the geometric center of the battery cell as the origin. The layers are indexed according to the electrode stacking direction (i.e., the z-axis direction) to distinguish different physical layers such as the positive electrode layer, the separator layer, and the negative electrode layer; the spatial coordinates of the actual placement locations of the voltage sensor, temperature sensor, and vibration sensor obtained by laser ranging or structured light scanning are marked as follows: Where i is the sensor number, and its value ranges from 1 to 10. ; , , These represent the x, y, and vertices of the sensor in the 3D contour model of the battery cell, respectively, all in millimeters. This represents the total number of monitoring nodes.

[0060] Rigid body coordinate transformation is used to map the sensor coordinates to the spatial node positions corresponding to the electrode stack structure. The mapping formula is as follows: Where R is a 3×3 rotation matrix used to correct the angular deviation between the sensor coordinate system and the cell model coordinate system; t is a three-dimensional translation vector used to eliminate the origin offset between the two coordinate systems. The coordinates of the dynamically monitored nodes after mapping. These correspond to the horizontal, vertical, and angular coordinates of the nodes in the cell model space, respectively; for each node Associate it with its corresponding original sensing signal: voltage signal Temperature signal and vibration signals Where t is a continuous time variable in seconds; thus forming a dynamic monitoring node set, which takes the form:

[0061] ;

[0062] here The total number of nodes is denoted by , and each node carries one or more disturbance components from the pulse current impact, wideband mechanical vibration, and wide temperature range thermal drift that accompany the strong dynamic working conditions.

[0063] Step 1.2: Based on the dynamic monitoring node set, construct a multi-source field response topology cluster conformal to the stacking direction of the battery cell electrode sheets. Specifically, ensuring the local response configuration of the multi-source field response topology cluster matches the physical layering of the battery cell includes: based on the obtained dynamic monitoring node set N, first calculate the values ​​of each node... The spatial distribution density in the 3D contour model of the battery cell; a local weighted regression method is used to fit the node set to a surface, that is, the electrode stacking direction is set as the w-axis direction, and a multi-dimensional monitoring surface S(u,v) orthogonal to the w-axis is constructed; the w coordinate corresponding to each point (u,v) of this surface is obtained by weighted averaging the normal height of the neighboring nodes, and the calculation formula is as follows: In the formula For nodes The set of neighboring nodes, where K(⋅) is the Gaussian kernel function. Let be the Euclidean distance between the two nodes. For nodes The normal height corresponding to the curved surface.

[0064] Based on the fitted surface S(u,v), the Delaunay triangulation algorithm is executed to connect adjacent nodes into a triangular mesh. Specifically, all nodes in the node set N are first projected onto the parameter plane (u,v) of the surface to obtain a two-dimensional planar point set. Then, the Bowyer-Watson algorithm is used to construct the Delaunay triangular mesh of this planar point set, which satisfies the empty circle criterion: that is, the circumcircle of any triangle does not contain other nodes. Finally, the triangular connection relationships on the two-dimensional plane are mapped back to the nodes in the original three-dimensional space, forming a multi-source field response topological cluster. Where N is the set of nodes, Let be the edge set, representing one edge of each triangle; It is a set of facets, where each triangle is a facet.

[0065] During the construction of the topology cluster, the local response configuration of the topology cluster is forced to match the physical layered shape of the battery cell; the specific operation is as follows: first, the positive electrode layer, separator layer and negative electrode layer in the three-dimensional contour model of the battery cell are marked as layers. For triangular facets falling near the layer interface, i.e., facets with at least two vertices located in different layers, local resampling and edge swapping operations are performed. For the interface between the positive electrode layer and the separator layer, the side length of the triangular facets in the corresponding region is adjusted so that the edge length of the facet at the interface does not exceed 1.2 times the average side length. At the same time, the curvature of the facets is adjusted so that the normal vectors on both sides of the interface are continuous. For the interface between the separator layer and the negative electrode layer, the normal vector continuity constraint is also performed. In addition, for all regions located inside the layers, the second derivative of the facets (i.e., the rate of change of curvature) is forced to remain consistent when crossing the layer boundary. The above constraints are implemented through an iterative optimization algorithm: in each iteration, the angle of the normal vector and the curvature gradient of each interface facet are calculated. If they exceed a preset threshold, the angle of the normal vector is greater than 5° or the rate of change of curvature gradient is greater than 10, then the position of the relevant nodes and the facet connection relationship are fine-tuned. The iteration continues until the angle of the normal vector at all interfaces is less than 2° and the rate of change of curvature gradient is less than 5, and finally a multi-source field response topology cluster conforming to the stacking direction of the cell electrode is obtained.

[0066] Step 1.3: Extract the local interface stress distortion curvature at each node on the multi-source field response topology cluster, as well as the geodesic deviation between nodes along the potential field dissipation path. Based on the pre-acquired timestamp-aligned multi-source signal values, form an initial multidimensional sensing sequence with unified spatiotemporal coordinates on the topology cluster. Specifically, this includes: extracting the multi-source field response topology cluster obtained in Step 1.2. The local interface stress distortion curvature at each node, for the node The quadratic surface is fitted using the nodes in its neighborhood, and the formula is as follows:

[0067] ;

[0068] Here, (u,v) represents the coordinate parameters of the local tangent plane of the surface at the nodes, and a, b, c, d, e, f are coefficients obtained by least squares fitting. The first fundamental quantities E, F, and G of this quadratic surface are calculated using the following formulas:

[0069] , , ;

[0070] in Let be the first-order partial derivative of the surface along the u-direction. The first-order partial derivatives are defined along the v direction; the second fundamental quantities L, M, and N are defined as follows:

[0071] , , ;

[0072] in , , All are second-order partial derivatives; thus, the nodes are obtained. Gaussian curvature at the point and mean curvature The tensor combination of Gaussian curvature and mean curvature is defined as the local interface stress distortion curvature tensor. ;here The value reflects the overall curvature of the surface at the nodes; positive values ​​represent elliptic points, and negative values ​​represent hyperbolic points. H reflects the average curvature direction of the surface at that point. Further calculations are made of the geodesic deviation of the potential field dissipation paths between nodes. A topological cluster is assumed. The two points above and The geodesic distance between them is Its continuous form is calculated as follows: In the formula, For connection and For a differentiable curve, ds is the derivative of the curve with respect to the parameter s; the parameter , , ; ν is the Riemannian metric tensor induced on the topological family, whose components are determined by the first fundamental quantity of the surface; μ and ν are coordinate indices (which can be 1 and 2 corresponding to the u and v directions). Let μ be the component of the tangent vector of the curve in the μ direction; This indicates that the minimum value is taken as the least certain value for all possible curves γ.

[0073] In the discrete implementation, the edge set of the topological cluster Assign weights ;in, For nodes and Spatial Euclidean distance; For nodes at fixed times The signal vector contains three components: voltage, temperature, and vibration. The Euclidean distance between the signal vectors; To balance the weighting coefficients of spatial distance and signal difference, the values ​​are typically in the range of [0.5, 2.0]. The geodesic deviation is obtained by solving the shortest path problem, for example using Dijkstra's method. ,in From arrive The set of all paths.

[0074] Finally, the multi-source signal values ​​are aligned based on the pre-acquired timestamps. , , Generate an initial multidimensional sensing sequence with unified spatiotemporal coordinates on the topological cluster T; for each node At any moment The voltage signal, temperature signal, vibration signal, and local interface stress distortion curvature tensor are compared. and geodesic deviation vector Assemble into feature vectors: in This is the four-dimensional row vector obtained by expanding the curvature tensor row by row; all nodes are arranged in spatial order (i.e., a fixed order of node numbers from smallest to largest). This represents the total number of all sensing and monitoring nodes. The geodesic deviation of node i relative to node j forms the initial multidimensional sensing sequence: in For discrete time sampling points, k=1,2,…; the sequence as a whole has unified spatial coordinates (fixed node order) and time coordinates, and the same… All signal values ​​at the sampling time are obtained from synchronous acquisition.

[0075] In a preferred embodiment of the present invention, step 2 above may include:

[0076] Step 2.1 involves separating the covariant temporal sequence of electrical, thermal, and vibration signal components from the initial multidimensional sensing sequence under a unified spatiotemporal coordinate system. A multiphysics covariant tensor space characterizing the coupling relationship between electrochemical, thermal, and mechanical responses is constructed. This multiphysics covariant tensor space serves as the solution basis for the unsteady multiphysics covariant decoupling model. Specifically, this includes extracting the electrical, thermal, and vibration signal components from the obtained initial multidimensional sensing sequence according to a unified spatiotemporal coordinate system. The electrical signal components include the terminal voltage and charging / discharging current at each monitoring node, and the thermal signal components include the terminal voltage and charging / discharging current at each node. The surface temperature at the point and the vibration signal components include triaxial acceleration and vibration frequency. The time series of these three components are covariantly aligned: for each pair of signals (electric-thermal, electric-vibration), the cross-correlation coefficient between the two is calculated within a continuously sliding time window (window length is 0.1 seconds, step size is 0.02 seconds), and possible time delay offsets (from -0.5 seconds to 0.5 seconds) are traversed. The offset that makes the cross-correlation coefficient reach its maximum value is found as the relative delay of the pair of signals. The median of the delay offsets of each pair of signals at all nodes is taken as the final system-level delay compensation value, and the thermal signal and vibration signal are aligned to the time reference of the electrical signal.

[0077] Based on this, an unsteady multiphysics covariant decoupling model is constructed. This model is an improvement on a hybrid architecture of unsteady convolutional sparse coding and multimodal variational autoencoders. Standard variational autoencoders are suitable for stationary signals, while this model replaces the encoder part with temporal convolutional layers with different dilation rates to capture transient abrupt changes (lasting several milliseconds) caused by pulsed current impacts and long-period oscillations (lasting several seconds to tens of seconds) caused by broadband mechanical vibrations. The decoder part is designed with a multi-branch structure, corresponding to electrochemical perturbation reconstruction, mechanical perturbation reconstruction, and thermal drift compensation branches, with the three branches sharing the underlying latent variable space. This model is constructed... During the process, the time-series data of the aligned electrical, thermal, and vibrational components are used as input. The encoder maps these data to the latent variable space. One dimension of this space encodes the time-varying amplitude of the unsteady-state perturbation, and the other dimension encodes the frequency domain distribution of the perturbation. Each branch of the decoder reconstructs the electrochemical perturbation, mechanical perturbation, and thermal drift from the latent variables. By minimizing the KL divergence between the reconstruction error and the latent variable distribution, the latent variable space is forced to decouple into mutually independent orthogonal subspaces. The decoupled latent variable space is defined as the basis of the multiphysics covariant tensor space, that is, the three orthogonal axes of this tensor space correspond to the electrochemical perturbation latent subspace, the mechanical perturbation latent subspace, and the thermal drift latent subspace, respectively.

[0078] The specific training process of this model is as follows: First, data is collected from typical strongly dynamic coupled operating conditions, including the pulse discharge and severe vibration coupling process of non-road mobile equipment (such as mining dump trucks) during heavy-load uphill climbing, the current impact and broadband vibration superposition process of low-altitude transport devices (such as electric vertical take-off and landing aircraft) during hovering and transition phases, and the thermal drift and load change intertwining process of special field energy storage platforms (such as energy storage cabins in remote areas) under large day-night temperature difference environments. The sensor data collected under these operating conditions are divided into training set, validation set and test set in a ratio of 8:1:1. During training, Adam is used. The optimizer has an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. The loss function consists of three parts: the mean square error of electrical signal reconstruction, the mean square error of vibration signal reconstruction, and the KL divergence of the latent variable distribution. After every 10 training epochs, the decoupling index, i.e., the cross-correlation coefficient between the outputs of each perturbation branch, is calculated on the validation set. Training stops and the model parameters are saved when the validation set loss no longer decreases for 15 consecutive epochs. After training, the model can automatically separate the input composite sensing signal into mutually orthogonal electrochemical perturbation components, mechanical perturbation components, and thermal drift components.

[0079] The advantages of this model architecture are as follows: First, the convolutional structure can capture non-steady-state features at different time scales, overcoming the shortcomings of fixed-condition calibration methods that cannot adapt to dynamic conditions. Second, the multi-branch decoder achieves explicit decoupling of three types of disturbances, ensuring that the disturbance components remain orthogonal, fundamentally solving the problem of difficulty in distinguishing between operating condition disturbances and electrochemical degradation signals in the original technology. Third, the decoupling constraints of the latent variable space enable the model to have good transfer capabilities across multiple battery cells and platforms, requiring only a small amount of target domain data for fine-tuning to adapt to new scenarios, effectively improving the generalization performance across battery cells, platforms, and scenarios. After training, this multi-physics covariant tensor space is set as the solution basis for the non-steady-state multi-physics covariant decoupling model, and all subsequent decoupling operations are performed within this tensor space.

[0080] Step 2.2: Based on the solution basis, trace the dynamic disturbance response trajectory excited by the pulsed current impact and broadband mechanical vibration in the multiphysics covariant tensor space to obtain the dynamic disturbance response trajectory; based on the dynamic disturbance response trajectory, perform multi-axis orthogonal decomposition of inertial torque along the electrochemical response principal axis and the mechanical vibration principal axis to obtain the independent projection of the dynamic disturbance component on each physical field principal axis. Specifically, this includes: in the multiphysics covariant tensor space constructed in step 2.1, the basis of this tensor space has been orthogonally normalized through the model training and latent space decoupling constraints in step 2.1, where the unit vector of the electrochemical response principal axis is... and the unit vector of the principal axis of mechanical vibration It has been determined that, based on this, short-time Fourier transform analysis will be performed on the vibration signal components at each node. The specific operation is as follows: the continuous time series of the vibration signal is segmented into 0.2-second time windows, with 50% overlap of sample points between each two adjacent time windows; a Hanning window is applied to the signal within each time window to reduce spectral leakage, and then a fast Fourier transform is performed to obtain the power spectral density distribution of that time window. The power spectral densities of all time windows are arranged in chronological order to form a time-spectrum diagram. In this time-spectrum diagram, two types of features are identified: the first type is the transient high-frequency feature of pulsed current impulse excitation, which is characterized by a sudden concentration of energy in a wide frequency band from 500Hz to 2000Hz at a certain moment (this moment strictly corresponds to the current change point recorded by the current sensor), with the energy peak lasting for a very short time (less than 0.01 seconds), and the energy decaying rapidly after the peak; the second type is the continuous spectral distribution of broadband mechanical vibration excitation, which is characterized by the continuous existence of energy in the frequency range of 50Hz to 500Hz, with the energy changing slowly with changes in load or rotational speed, without obvious transient spikes.

[0081] After identifying the above features, the position coordinates of the transient high-frequency features at each moment are superimposed with the intensity values ​​of the continuous spectrum. That is, in the three-dimensional coordinates of the tensor space, a point is marked for each sampling moment. The electrochemical coordinates of the point are taken as the voltage and current values ​​at the current moment, the thermal coordinates are taken as the temperature value at the current moment, and the vibration coordinates are taken as the vibration energy amplitude at the current moment. These marked points are connected sequentially in time to form a dynamic disturbance response trajectory. This trajectory is represented in the tensor space as a spatial curve that extends continuously with time. Each point of the curve corresponds to the disturbance state at a sampling moment, and the direction of the tangent vector of the curve reflects the direction of change and distribution ratio of the disturbance in each physical field at that moment.

[0082] Furthermore, the trajectory is orthogonally decomposed along the principal axis of electrochemical response and the principal axis of mechanical vibration using multi-axis inertial torque. The principal axis of electrochemical response is defined as the direction spanned by the unit voltage vector and the unit current vector in the tensor space, and the principal axis of mechanical vibration is defined as the direction spanned by the unit vibration acceleration vector and the unit vibration frequency vector. Since the substrate has been orthogonally normalized using the Schmidt orthogonalization method when constructing the tensor space in step 2.1, the unit vectors of these two principal axes are perpendicular to each other, and both have a magnitude of 1. For clarity, these two unit vectors are denoted as... Electrochemical principal axis unit vector sum Unit vector of the principal axis of mechanical vibration.

[0083] To achieve decomposition, the tangent vector at each sampling point on the trajectory is first calculated. For two consecutive sampling points on the trajectory, they are designated as the previous point A and the subsequent point B according to their time sequence. The spatial coordinates of the previous point A are subtracted from the spatial coordinates of the subsequent point B to obtain a displacement vector. This displacement vector is then divided by the time interval between the two points to obtain the instantaneous velocity vector at sampling point A, which is an approximation of the tangent vector. For the last sampling point on the trajectory, backward difference is performed using its preceding neighboring point. For the first sampling point on the trajectory, forward difference is performed using its following neighboring point. This process is repeated to obtain the tangent vector at each sampling point. The tangent vector at the k-th sampling point (where k is the same as in step 1.3, representing the index of the discrete time sampling point) is denoted as... .

[0084] Based on this, the tangent vector for each sampling point is calculated separately. unit vector on the principal axis of electrochemistry and the unit vector of the principal axis of mechanical vibration The projection length is calculated using the following formula: ; where ・ represents the dot product of vectors, that is, the corresponding coordinate components are multiplied and then summed to calculate the result. It is a scalar and can be positive (indicating that the direction of the tangent vector is the same as the direction of the electrochemical principal axis) or negative (indicating that the direction is opposite). Similarly, after obtaining the projection length, multiply the projection length by the unit vector of each principal axis to obtain the projection vector of the tangent vector on the electrochemical principal axis and the projection vector on the mechanical vibration principal axis: ;

[0085] because and Since they are perpendicular to each other and have a magnitude of 1, these two projection vectors naturally satisfy an orthogonal relationship. This represents the portion of the tangent vector that belongs to the electrochemical response, i.e., the electrochemical perturbation component. This represents the component of the tangent vector attributable to mechanical vibration, i.e., the mechanical disturbance component. Since the physical meaning of the tangent vector is the change in the disturbance state per unit time, the above projection results directly give the magnitude and direction of the independent contribution of the dynamic disturbance on the electrochemical principal axis and the mechanical vibration principal axis at each sampling moment.

[0086] Finally, the electrochemical and mechanical perturbation component sequences at each sampling point are arranged in chronological order, and time integration is performed on each component sequence, i.e., the changes within each time interval are accumulated to obtain the independent projection values ​​of the dynamic perturbation components on each physical field principal axis. These independent projection values ​​reflect the total accumulated contribution of electrochemical and mechanical perturbations from the initial time to the current time. Thus, the original dynamic perturbation is successfully decomposed into mutually orthogonal electrochemical and mechanical perturbation components, providing accurate separation results for the filtering operation in step 2.3.

[0087] Step 2.3: From the initial multidimensional sensing sequence, the independent projections of the dynamic disturbance components on each physical field principal axis are filtered out to obtain a pure electrochemical feature signal set stripped of operating condition coupling noise. Specifically, this includes: subtracting the independent projections of the decomposed dynamic disturbance components on each physical field principal axis from the obtained initial multidimensional sensing sequence one by one; first, subtracting the electrochemical disturbance projection value at the corresponding time from the electrical signal components (voltage, current) of each node. At this time, the remaining signal part only contains slow-changing features related to the electrochemical degradation process, such as the slow decrease of the voltage plateau period and the abnormal inflection point of the charge-discharge curve; on this basis, subtracting the mechanical disturbance projection value at the corresponding time from the vibration signal components (acceleration, frequency) of each node. The remaining part mainly reflects features related to the structural integrity such as the deterioration of the electrode interface contact state, electrode loosening, and diaphragm damage.

[0088] Meanwhile, for the temperature signal component, its covariance with the vibration signal component is first calculated: that is, the correlation coefficient between temperature change and mechanical vibration energy in time. Then, this correlation coefficient is multiplied by the vibration disturbance projection value as an estimate of thermal drift. This estimate is subtracted from the original temperature signal to filter out the thermal drift component caused by mechanical friction or drastic changes in ambient temperature. After completing the above component-by-component and node-by-node filtering operation, all remaining signal components are reassembled according to the original spatiotemporal coordinates to obtain a pure electrochemical characteristic signal set stripped of operating condition coupling noise such as pulse current impact, broadband mechanical vibration, and wide-temperature-range thermal drift. This signal set only retains the intrinsic information in the original sensing signal that is directly related to degradation processes such as lithium plating, internal short circuit, tab aging, micro short circuit, and sealing failure. Its signal-to-noise ratio is effectively improved compared to the original signal, providing interference-free and high-fidelity input data for the construction of the unified mapping network and the extraction of degradation features in subsequent steps.

[0089] In a preferred embodiment of the present invention, step 3 above may include:

[0090] Step 3.1: Obtain the pre-set interface solvation potential well and stress relaxation coupling mechanism. Expand the pure electrochemical feature signal set under the coupling mechanism to analyze the evolution path of the intermolecular interaction potential well of the solvation sheath at the electrode interface and the temporal correlation of lattice stress relaxation, and obtain the interface potential well-stress relaxation temporal correlation mapping. Specifically, starting from the pure electrochemical feature signal set obtained in Step 2.3, first call the pre-stored interface solvation potential well and stress relaxation coupling mechanism. The mechanism is a set of mathematical models based on electrochemical principles. Its core idea is as follows: At the electrode / electrolyte interface of a lithium battery, before lithium ions enter the electrode material lattice from the electrolyte, they need to peel off the solvent molecule layer surrounding them. This peeling process needs to overcome a problem that depends on the electrode potential and The energy barrier at the interface temperature is called the solvation well. The well depth increases with increasing electrode potential, making it more difficult for lithium ions to desolvate. It decreases with increasing temperature, and thermal motion helps overcome the barrier. Simultaneously, lithium ion insertion into the electrode lattice causes local lattice expansion, generating compressive stress within the lattice. As lithium ions continue to diffuse or desorb, the stress is gradually released, forming an exponentially decaying relaxation process. This coupling mechanism links the well evolution and stress relaxation processes through lithium ion flux: when the lithium ion flux is high, the insertion rate is fast, the well is repeatedly overcome, and stress accumulates rapidly; when the lithium ion flux is low, the stress has sufficient time to release. This mechanism incorporates several experimentally calibrated parameters, such as standard well depth, stress relaxation characteristic time, and temperature influence coefficient.

[0091] In practice, for each sampling moment in the pure electrochemical characteristic signal set, the solvation well depth and lattice stress value are calculated according to the following process: Specifically, the solvation well depth is first calculated by extracting the voltage value at the current moment, obtaining the theoretical open-circuit voltage by looking up the table through the open-circuit voltage curve, and then subtracting the current ohmic polarization voltage (estimated by multiplying the current by the battery internal resistance) and concentration polarization voltage (estimated by integrating the current and the diffusion coefficient) to obtain the true electrode potential relative to the reference electrode; using this electrode potential, the pre-calibrated electrode potential-well depth baseline curve is consulted to obtain a baseline well depth value; further, the current value at the current moment is extracted, and the current is divided by the electrode area and the lithium ion molar concentration to estimate the interfacial lithium ion flux.

[0092] The higher the lithium-ion flux, the more lithium ions attempt to cross the interface per unit time, which dynamically increases the effective depth of the potential well because solvent molecules do not have enough time to completely rearrange. Therefore, an increment term proportional to the lithium-ion flux is added to the basic potential well depth, and the increment coefficient is obtained through experimental fitting. Finally, the temperature value at the current moment is extracted, and the reaction rate is corrected using the Arrhenius calculation method: that is, as the temperature increases, the thermal motion of lithium ions is enhanced, making it easier to overcome the potential barrier, so the effective potential well depth should decrease exponentially. The specific operation is as follows: calculate the ratio of the temperature to the reference temperature (e.g., 298 Kelvin), multiply the sum of the basic potential well depth and the increment by a temperature correction factor. The higher the temperature, the smaller the correction factor. The correction factor is pre-calculated into a lookup table using the Arrhenius calculation method and obtained directly from the temperature. After the above three steps, the solvation potential well depth at the current moment is obtained.

[0093] Further, the lattice stress value and its relaxation time constant are calculated: Accumulated capacity data is obtained from the battery management terminal. The current accumulated charge / discharge capacity is divided by the battery's rated capacity to obtain the lithium insertion depth (range 0 to 1). The greater the lithium insertion depth, the more severe the lattice expansion, and the higher the peak stress. A piecewise linear relationship is adopted: stress growth is slow when the lithium insertion depth is below 0.2, rises rapidly between 0.2 and 0.8, and slows down again when it is above 0.8. The theoretical peak stress at that moment is determined based on the lithium insertion depth. Simultaneously, the relaxation effect of stress over time needs to be calculated: the stress does not reach its peak instantaneously but accumulates gradually with the lithium-ion insertion rate and diffusion process; after the current is stopped, the stress will gradually decrease according to the time-dependent relaxation process. The stress decays according to a first-order exponential law. To address this, a stress state variable is maintained, recording the residual stress value at the previous moment. The actual stress value at the current moment equals the residual stress multiplied by a decay coefficient, determined by the relaxation time constant, plus the current elastoplastic increment, determined by the current current and the rate of change of lithium insertion depth. The relaxation time constant itself also varies with temperature: at high temperatures, atomic vibrations intensify, stress relaxation is faster, and the time constant is smaller; at low temperatures, relaxation is slower, and the time constant is larger. The relaxation time constant at the current temperature is obtained by consulting the temperature-relaxation time constant table. This time constant is then substituted into the first-order exponential decay law described above to calculate the decay coefficient, which is between 0 and 1. Finally, the lattice stress value at the current moment and the corresponding relaxation time constant are obtained.

[0094] The potential well depths calculated above are connected sequentially at each sampling time to obtain the evolution trajectory of the potential well depth over time; the stress values ​​at each sampling time are connected sequentially to obtain the stress decay and relaxation curve over time; based on this, the two curves are time-series correlated and aligned: firstly, the local maxima in the potential well depth curve are found by using the peak detection method (i.e., comparing three adjacent points, if the middle point is greater than the left and right neighboring values, it is determined to be a local maximum), these points correspond to the time when lithium ions undergo rapid desolvation events; the starting point of each falling segment in the stress relaxation curve is found, which is the first point where the stress continuously decreases from the peak value; the time offset between each pair of adjacent potential well maxima and the stress starting point is calculated, i.e., the stress starting point minus the potential well maxima, and the median of all offsets is taken as the fixed delay of the system, which reflects the inherent hysteresis between lithium ion desolvation and the stress starting to be released.

[0095] Next, cross-correlation analysis is performed on the entire curve: after normalizing the potential well depth sequence and stress sequence to zero mean, the cross-correlation coefficients under different time lags are calculated, and the lag value that maximizes the cross-correlation coefficient is found as the phase difference; based on the calculated delay and phase difference, the stress curve is shifted backward on the time axis by the delay amount, and then fine-tuned according to the phase difference, so that the change in potential well depth corresponds one-to-one with the change in stress relaxation in time; after the above analysis and alignment operations, the interface potential well-stress relaxation time series correlation mapping is obtained. This mapping records the pairing relationship between solvation potential well depth and lattice stress value at each time point from the start time to the current time, providing accurate alignment input data for establishing constitutive relations in step 3.2.

[0096] Step 3.2: Based on the interface potential well-stress relaxation time-series correlation mapping, establish the interface potential well-stress relaxation constitutive relation characterizing the intrinsic response of electrochemical degradation. Use this interface potential well-stress relaxation constitutive relation as the internal mapping kernel. Specifically, this includes: extracting the essential characteristics of the electrochemical degradation process from the interface potential well-stress relaxation time-series correlation mapping obtained in Step 3.1; first, numerically differentiating the potential well depth curve; then calculating the slope of each internal point using the three-point central difference method: that is, subtracting the previous point from the next point and dividing by twice the time interval; for the first and last endpoints, forward or backward difference is used; after obtaining the curve of slope changing with time, calculate the mean of the sliding window of the slope (window length is 0.5 seconds), and set an abnormal threshold (the mean of the slope under normal operating conditions plus 3 times the standard deviation); when the slope of 3 consecutive sampling points exceeds this threshold, it is judged as an abnormal increase, indicating that lithium plating or solvent decomposition may occur at the electrode interface.

[0097] Based on this, an exponential decay fitting is performed on the stress relaxation curve: in each descending segment of the stress relaxation curve, from the local peak to the next local trough, an exponential function is fitted using the least squares method; the form of this exponential function is as follows: ,in This indicates the relative time from the start point of the descent segment; relative time The stress value at the starting point is given by: A being the amplitude of the decay, i.e., the difference between the initial stress value and the final steady-state stress value; λ being the decay rate, the reciprocal of which is the decay time constant; and C being the steady-state baseline value, i.e., the residual stress after decay ends. By fitting using the least squares method, the sum of squared prediction errors at all sampling points is minimized, thus solving for the optimal λ value, and consequently obtaining the decay time constant for this descent segment. ;here The time required for the stress to decay to approximately 36.8% of the initial difference is represented by the fitted decay time constant. The time constant is compared with the standard time constant under healthy battery conditions (obtained through factory calibration). If the deviation exceeds 20% and lasts for more than 1 second, it is judged as abnormal, indicating that microcracks or phase transitions have occurred in the crystal lattice.

[0098] Furthermore, using the potential well depth as the abscissa and the stress value as the ordinate, all data points in the time-series correlation mapping are plotted on a two-dimensional plane. A smooth constitutive curve is fitted using a local weighted regression method. The specific process is as follows: For each target point on the abscissa, data points in its neighborhood are selected, with the neighborhood radius determined according to 15% of the total data range; each neighborhood point is assigned a weight, the weight value being inversely proportional to the abscissa distance from the target point, with the weight increasing as the distance decreases; the weights are then calculated; simultaneously, a first-order polynomial is fitted within the current neighborhood; the value of this polynomial at the target point is used as the ordinate of the corresponding point on the constitutive curve; after traversing all abscissa points, the fitted values ​​are connected to obtain a smooth constitutive curve; this curve defines the theoretical stress value corresponding to any given potential well depth. The constitutive curve is compared point-to-point with the standard constitutive curve of the health state measured at the time of battery delivery: the stress deviation value is calculated at the same potential well depth, and the magnitude of the deviation (maximum deviation value), the duration of the deviation (the duration of continuous deviation beyond the allowable range), and the trend of deviation change (the deviation value increases or decreases with time) are statistically analyzed to obtain a three-dimensional deviation feature vector; this vector is the interface potential well-stress relaxation constitutive relation characterizing the intrinsic response of electrochemical degradation; finally, this constitutive relation is used as the core processing unit of the subsequent unified mapping network, namely the internal mapping kernel, to convert the input pure feature signal into an intermediate representation related to degradation.

[0099] Step 3.3: Based on the internal mapping kernel, a unified mapping network framework is constructed. Within this framework, the effects of dynamic operating condition disturbances are encoded as boundary condition constraints, and the electrochemical degradation state is encoded as output layer state parameters. Specifically, this includes: using the established interface potential well-stress relaxation constitutive relation as the internal mapping kernel to construct the unified mapping network framework. This unified mapping network framework is a three-layer fully connected structure: an input layer, a hidden layer, and an output layer. The input layer receives two types of signals: the first type consists of voltage, current, and temperature extracted from the pure electrochemical characteristic signal set, with three nodes; the second type is the dynamic disturbance component serving as boundary condition constraints, including the following steps... 2.2 The decomposed components are electrochemical, mechanical, and thermal drift, with a total of 3 auxiliary nodes. The two types of signals are combined to form a total of 6 input nodes. The input layer does not perform any transformation on the signal and directly transmits the 6-dimensional features at each sampling time to the hidden layer unchanged. The hidden layer contains an internal mapping kernel and several auxiliary nodes. The internal mapping kernel performs mapping operations according to the constitutive relation defined in step 3.2: it receives voltage and current values, calculates the potential well depth, and then obtains the theoretical stress value by looking up the table according to the constitutive curve. Furthermore, it subtracts the theoretical stress value from the actual stress estimate implied in the input temperature signal to obtain the residual signal. This residual signal is the degradation characterization quantity.

[0100] Two auxiliary nodes are set in the hidden layer to receive mechanical disturbance components and thermal drift components, respectively, for dynamically adjusting boundary conditions. Specifically, the impact of dynamic operating condition disturbances is encoded as boundary condition constraints: a threshold is set for the mechanical disturbance component (e.g., an acceleration amplitude of 0.2g). When this threshold is exceeded, the activation bias of the neurons corresponding to the vibration-sensitive region in the hidden layer is forcibly increased, and the bias is proportional to the disturbance amplitude. A threshold is set for the thermal drift component, such as a temperature change rate of 2℃ / second. When this threshold is exceeded, the activation bias of the neurons corresponding to the temperature-sensitive region is forcibly increased. The output layer contains 5 neuron nodes, corresponding to five degradation types: lithium plating, internal short circuit, electrode aging, micro-short circuit, and sealing failure. The output value of each node needs to be compressed to between 0 and 1, representing the probability of this type of degradation occurring. The output layer state parameters include the weight matrix and bias vector of each node. These parameters are updated during subsequent training. After completing the above definitions, the prototype of the untrained unified mapping network framework is obtained.

[0101] Step 3.4: Input the pure electrochemical feature signal set into the unified mapping network framework, and project the features onto it using the internal mapping kernel to obtain the feature projection result; perform a nonlinear transformation on the feature projection result using the output layer state parameters, so that the unified mapping network framework establishes the mapping relationship from dynamic operating condition disturbance to electrochemical degradation state, and constructs the unified mapping network from dynamic operating condition to electrochemical degradation state. Specifically, this includes: inputting the pure electrochemical feature signal set into the constructed unified mapping network framework, first organizing the six-dimensional features (voltage, current, temperature, electrochemical disturbance component, mechanical disturbance component, and thermal drift component) at all sampling times into a sequence in the input layer according to time order, and passing them to the hidden layer frame by frame; the hidden layer... The internal mapping kernel performs point-by-point projection on the feature vector at each sampling time. For the nth sampling time, the voltage and current values ​​are extracted from the pure electrochemical feature signal set at that time. Using the voltage value, the theoretical open-circuit potential corresponding to the voltage is first obtained by looking up the table through the open-circuit voltage curve. Then, polarization correction is performed by combining the current value: the current is multiplied by the battery internal resistance to obtain the ohmic polarization voltage, and the current integral over a period of time is multiplied by the diffusion coefficient to obtain the concentration polarization voltage. These two polarization voltages are subtracted from the measured voltage to obtain the true electrode potential. Based on this electrode potential, the pre-calibrated electrode potential-potential well depth relationship table is consulted, and the theoretical potential well depth at the current time is obtained by linear interpolation between two adjacent points.

[0102] Furthermore, based on the theoretical well depth, interpolation is performed on the constitutive curve obtained in step 3.2: the two points on the constitutive curve closest to the theoretical well depth are found, and linear interpolation is performed using the ordinates (stress values) of these two points to calculate the corresponding theoretical stress values; simultaneously, the temperature signal is extracted from the feature vector at this moment, and combined with the current lithium insertion depth (obtained by dividing the cumulative capacity by the rated capacity) and the temperature value, an empirical calculation method pre-fitted through experiments is used to estimate the actual stress value. The basic logic of this empirical calculation method is: the actual stress value equals the reference stress (determined by the lithium insertion depth) multiplied by the temperature correction factor. The temperature correction factor is less than 1 in the low-temperature region and greater than 1 in the high-temperature region. The specific value can be obtained by referring to the temperature-stress correction factor table. After obtaining the theoretical stress value and the actual stress value, the difference between the two is calculated, that is, the actual stress value minus the theoretical stress value. If the difference is positive, it means that the actual stress is too large, and if the difference is negative, it means that the actual stress is too small. The absolute value of the difference reflects the degree of abnormal deviation caused by electrochemical degradation. This difference is used as the residual at the current moment. The above calculation process is repeated for each sampling moment. The residual values ​​of all sampling moments are arranged in chronological order to form a one-dimensional residual sequence. This sequence represents the electrochemical abnormal components remaining after removing the normal constitutive response.

[0103] Simultaneously, the output layer state parameters perform a nonlinear transformation on the residual sequence. First, the residual sequence is segmented according to a time window: the window length is 0.5 seconds, the sliding step size is 0.1 seconds, and a window is generated for each sliding step. The following statistical calculations are performed on the residual points within each window: the arithmetic mean of all residual values ​​within the window is calculated as the mean feature of the window; the average of the squared deviations of each residual point within the window from the mean is calculated to obtain the variance feature; the residual points within the window that continuously exceed the allowable error threshold (this threshold is set to 0.05 times the standard stress value according to the residual distribution under battery health conditions) are traversed, and the longest continuous time exceeding the threshold is recorded as the maximum deviation duration feature; these three statistical features (mean, variance, and maximum deviation duration) are combined into a three-dimensional feature vector.

[0104] Based on this, a pre-set logistic regression classifier is invoked to classify the three-dimensional feature vector. This logistic regression classifier is a multi-class configuration, and its internal parameters (weight matrix and bias vector) have been determined during training with historical data. The training data covers the statistical features of residual sequences collected under different degradation types (lithium plating, internal short circuit, tab aging, micro-short circuit, and seal failure). The input to this classifier is the aforementioned three-dimensional feature vector, and the output is five probability values ​​between 0 and 1, corresponding to the likelihood of each of the five degradation types. During the mapping process of this classifier, the boundary condition constraints of the dynamic operating condition disturbance coding will adjust the decision boundary of the classifier in real time. Specifically, the mechanical disturbance component (this component is determined by step 2.2) is read from the second type of auxiliary input node of the network input layer at the current sampling time. The decomposition yields a set of signals independent of pure electrochemical features. This set is compared with a preset mechanical disturbance threshold (e.g., 0.2 times the amplitude of gravitational acceleration). When the mechanical disturbance component exceeds this threshold, the decision weight in the classifier corresponding to tab aging is temporarily multiplied by a decay factor (value 0.8) to moderately reduce the sensitivity of this type of degradation and avoid misjudgment caused by vibration interference. Simultaneously, the thermal drift component is read from the auxiliary input node and compared with a preset thermal drift threshold (e.g., a change rate of 2 degrees Celsius per minute). When the thermal drift component exceeds this threshold, the decision weight in the classifier corresponding to sealing failure is temporarily multiplied by an enhancement factor (value 1.2) to improve the sensitivity to sealing problems caused by temperature anomalies. The weights corresponding to the other three degradation types remain unchanged.

[0105] After the aforementioned feature projection and nonlinear transformation, the classifier outputs five probability values. As time progresses, a new set of probability values ​​is obtained with each sliding window, and these probability value sequences gradually stabilize. When the change in probability values ​​output by 10 consecutive sampling windows (i.e., the maximum absolute difference between the probability values ​​of the current window and the previous window) is less than 0.05, the determination network has successfully established a stable mapping relationship from dynamic operating condition disturbance input to electrochemical degradation state output. At this point, the unified mapping network framework is completed, becoming a complete unified mapping network from dynamic operating conditions to electrochemical degradation state. It should be noted that the output of the aforementioned logistic regression classifier is mainly used to verify the convergence and stability of the mapping network during the training phase. In the actual deployment and online early warning after training, the network no longer directly uses the classifier to output the final early warning. Instead, it extracts a set of degradation feature vectors with clear physical meaning through the method in step 4 and uses the cross-domain manifold covariance metric method in step 5 for more generalizable and accurate determination.

[0106] In a preferred embodiment of the present invention, step 4 above may include:

[0107] Step 4.1: Input the pure electrochemical feature signal set into the unified mapping network from dynamic operating conditions to electrochemical degradation state. The unified mapping network calls its internal mapping kernel to perform feature projection on the pure electrochemical feature signal set, mapping the pure electrochemical feature signal set to the hidden feature space with ligand field distortion and molecular potential field fluctuation as coordinate axes, thus obtaining the hidden layer electrochemical degradation feature projection. Specifically, this includes: inputting the pure electrochemical feature signal set into the unified mapping network from dynamic operating conditions to electrochemical degradation state constructed in step 3.4; this unified mapping network first receives the 6-dimensional feature vector at each sampling time in the input layer; based on this, the unified mapping network calls its internal mapping kernel to perform the feature projection operation; the internal mapping kernel performs the feature projection operation based on the interface potential well established in step 3.2. The stress relaxation constitutive relation is used to perform a first-order mapping on the input voltage, current, and temperature to calculate the intermediate residual sequence characterizing electrochemical anomalies. Subsequently, a pre-trained decoupling projection matrix in the hidden layer of the unified mapping network is used to further project the residual sequence onto a two-dimensional hidden feature space. The two coordinate axes of this hidden feature space are defined as the coordination field distortion axis and the molecular potential fluctuation axis. The coordination field distortion axis reflects the degree of distortion of the coordination environment of the electrode material's crystal structure during lithium-ion insertion and extraction, and its value is obtained by nonlinear transformation of the stress anomaly component in the residual sequence. The molecular potential fluctuation axis reflects the dynamic fluctuation of the intermolecular potential of the solvent molecules at the electrode / electrolyte interface, and its value is obtained by the potential well depth anomaly component in the residual sequence.

[0108] The specific method of projection is as follows: the residual value at each sampling time is decomposed onto these two axes, and the decomposition coefficients are determined by the pre-trained decoupling matrix in the internal mapping kernel; after the above projection, the residual signal at each sampling time is converted into a two-dimensional coordinate point, the ligand field distortion value, and the molecular potential field fluctuation value. The coordinate points of all sampling times are arranged in chronological order to form a hidden layer electrochemical degradation feature projection sequence; this sequence depicts a trajectory that evolves over time in the hidden feature space, and the shape, curvature, and speed of the trajectory directly correspond to the development state of electrochemical degradation.

[0109] Step 4.2: Based on the projection of the hidden layer electrochemical degradation features, the output layer state parameters are nonlinearly transformed to obtain a nonlinear transformation response tensor. Based on the nonlinear transformation response tensor, electrochemical response feature components are extracted from the tensor along the coordination field distortion axis and the molecular potential field fluctuation axis of the hidden feature space to obtain an electrochemical degradation response feature set. Specifically, this includes: based on the obtained hidden layer electrochemical degradation feature projection sequence, the output layer state parameters of the unified mapping network are nonlinearly transformed. The output layer state parameters contain a set of pre-trained nonlinear mapping functions, which map points in the two-dimensional hidden feature space to a higher-dimensional response tensor space. Specifically, for each two-dimensional coordinate point (coordination field distortion value, molecular potential field fluctuation value) at each sampling time, the output layer first calculates the distance from the point to the origin of the hidden feature space and the direction angle of the line connecting the point to the previous time point; the output layer then calculates the distance from the point to the origin of the hidden feature space and the direction angle of the line connecting the point to the previous time point. The calculation results of each sampling window are denoted as two-dimensional geometric feature vectors. ,in The distance to the origin. Let the direction angle of the line connecting to the previous time point be denoted as . These geometric features are then further propagated forward through a multilayer perceptron network, the mathematical expression of which is:

[0110] ;

[0111] in, The input is a two-dimensional geometric feature vector; This is the first layer weight matrix, with a dimension of 4×2, and its elements are the learnable parameters obtained from network training; is the first layer bias vector with a dimension of 4; tanh is the hyperbolic tangent activation function, which calculates the hyperbolic tangent value element-wise over the input vector to introduce nonlinearity. This is the second-layer weight matrix, with a dimension of 6×4; This is the second-layer bias vector, with a dimension of 6; The output is a six-dimensional response tensor component, i.e., a column vector containing six components; the above All of these were learned during network training using a large amount of data containing different degradation types. The training objective was to minimize the cross-entropy loss between the output and the true degradation label. After the aforementioned nonlinear transformation, each sampling window generates a six-dimensional response tensor. The tensors of all windows are arranged in chronological order to form a sequence of nonlinear transformation response tensors. ,in This refers to the total number of windows in this area.

[0112] Next, electrochemical response feature components are extracted from the response tensor along two coordinate axes of the implicit feature space (ligand field distortion axis and molecular potential fluctuation axis). The extraction process is as follows: for the six-dimensional response tensor of each window... The first three components are defined as ligand field distortion response components, corresponding to the amplitude, duration, and rate of change of the distortion, respectively. The last three components are defined as molecular potential field fluctuation response components, corresponding to the frequency, phase shift, and energy dissipation of the fluctuation, respectively. After aligning these two response component sequences by time, they are merged to obtain an electrochemical degradation response feature set. This feature set is a two-dimensional matrix, with rows corresponding to the sampling window and columns corresponding to the six response feature components, comprehensively depicting the dynamic evolution of electrochemical degradation in both spatial (ligand field distortion) and temporal (potential field fluctuation) dimensions.

[0113] Step 4.3: From the electrochemical degradation response feature set, separate the first feature subset associated with lithium plating, the second feature subset associated with internal short circuits, the third feature subset associated with tab aging, the fourth feature subset associated with micro-short circuits, and the fifth feature subset associated with sealing failure, and combine them to form a degradation feature vector group. Specifically, this includes: separating feature subsets corresponding to five specific hidden danger types from the obtained electrochemical degradation response feature set, where the separation is based on the unique fingerprint characteristics of each hidden danger type in the response feature space; specifically, principal component analysis is first used to reduce the dimensionality of the response feature set, compressing the six-dimensional features into three dimensions and retaining the main variance components.

[0114] Further clustering analysis was performed in three-dimensional space using a density-based clustering algorithm. The neighborhood radius was adaptively determined based on the sample distribution, dividing the response points into five clusters, each corresponding to a type of degradation: The first feature subset corresponds to lithium plating risk, characterized by a slow increase in the amplitude of the coordination field distortion response component accompanied by high-frequency oscillations, while the energy dissipation in the molecular potential field fluctuation response component remains consistently high; the second feature subset corresponds to internal short circuit risk, characterized by transient spikes in both distortion amplitude and fluctuation frequency, with a fixed phase delay between them; the third feature subset corresponds to tab aging risk, characterized by a significantly prolonged distortion duration and a gradual decrease in fluctuation frequency; the fourth feature subset corresponds to micro-short circuit risk, characterized by periodic small jumps in distortion amplitude and a step-like increase in fluctuation energy dissipation; and the fifth feature subset corresponds to sealing failure risk, characterized by a sharp increase in fluctuation phase shift and an abnormal sign reversal in the distortion rate of change.

[0115] For each sampling time, the original six-dimensional feature vector of that time is assigned to the corresponding subset according to the cluster category to which the response point belongs. After the feature vectors of all times are reorganized by category, five feature subsets are obtained. Finally, these five feature subsets are spliced ​​into a complete degradation feature vector group in a fixed order of lithium plating, internal short circuit, tab aging, micro short circuit, and seal failure. This vector group retains the feature information of different hidden danger types at each sampling time, providing structured input data for the cross-domain manifold covariance measurement in step 5.

[0116] In a preferred embodiment of the present invention, step 5 above may include:

[0117] Step 5.1: Construct a degraded feature manifold based on the degraded feature vector group, obtain a preset cross-scene generalization threshold library, which stores a cross-scene reference manifold, and perform manifold registration between the degraded feature manifold and the cross-scene reference manifold to establish a cross-domain manifold covariant correspondence, thereby obtaining a cross-domain manifold registration mapping. Specifically, this includes: based on the degraded feature vector group, treating the six-dimensional feature vector of each sampling window as a point in a high-dimensional space, and the points of all windows constitute a high-dimensional point cloud; using the local linear embedding method to perform dimensionality reduction processing on the point cloud, for each point in the point cloud, calculating its Euclidean distance to all other points, and sorting them according to the distance from smallest to largest. Sort the data in ascending order and select the 15 nearest neighbors as the nearest neighbors of each point. Based on this, within the local neighborhood of each point, attempt to reconstruct the point itself using a linear combination of its nearest neighbors. Specifically, first determine the optimal combination weights of the 15 nearest neighbors for each point. This optimization process is constrained by minimizing the reconstruction error and ensuring that the sum of the weights of all nearest neighbors is 1. For each point, iteratively adjust the weight values ​​to minimize the difference between the weighted sum of the weights of the point and its nearest neighbors, thus obtaining a stable set of local reconstruction weights. After completing the above calculations for all points, a sparse weight matrix is ​​obtained, where each point has non-zero weights only with its nearest neighbors.

[0118] Furthermore, keeping these local reconstruction weights unchanged, a new set of three-dimensional coordinates is found in the low-dimensional space, such that the position of each point in the low-dimensional space can also be optimally reconstructed by its nearest neighbors with the same weight combination. This process is equivalent to calculating the eigenvalue decomposition of a sparse matrix. Specifically, a sparse matrix based on the local reconstruction weights is constructed, and then the eigenvectors corresponding to the three smallest non-zero eigenvalues ​​in the matrix are extracted by numerical methods. These eigenvectors are used as coordinate values ​​in three dimensions, and the six-dimensional eigenvector of each point is mapped to a three-dimensional coordinate point. All points constitute a point set in a three-dimensional space, and this point set is smoothly connected to form a degraded feature manifold (a three-dimensional surface). Each point on this manifold corresponds to the electrochemical degradation state of a sampling window, and the adjacency relationship between points is maintained by the original nearest neighbor relationship.

[0119] Simultaneously, a cross-scenario benchmark manifold is read from a preset cross-scenario generalization threshold library. This preset cross-scenario generalization threshold library is a comprehensive database that stores two types of data entities: cross-scenario benchmark manifolds and corresponding graded early warning judgment thresholds for each type of hidden danger. The benchmark manifold is a standard reference manifold established by averaging or clustering all low-dimensional embedding points obtained after performing the same local linear embedding process on a large number of battery samples (including healthy batteries and batteries with known degradation types under different cell models, different equipment platforms, and different environmental conditions). This benchmark manifold represents the standard geometric structure under normal operating conditions and typical degradation modes.

[0120] Next, manifold registration is performed between the degraded feature manifold and the cross-scene baseline manifold, using an iterative nearest-point algorithm. The specific process is as follows: First, for each point on the degraded feature manifold, find the point on the baseline manifold with the closest Euclidean distance to it, establishing an initial corresponding point pair; Second, based on all current corresponding point pairs, calculate an optimal rigid body transformation (including a rotation matrix and a translation vector) by minimizing the sum of squared distances between all corresponding point pairs. This transformation is obtained through iterative approximation: first, calculate the centroids of the two point sets, then decenter the point sets, then obtain the rotation component through singular value decomposition, and finally calculate the translation component; Third, apply the calculated rigid body transformation to all points on the degraded feature manifold. Obtain the new coordinates after transformation; fourth, find the nearest corresponding point pair between the new degraded feature manifold and the reference manifold, repeating steps two to four above, recording the average distance between all corresponding point pairs in each iteration; when the average distance change between two consecutive iterations is less than a preset threshold (e.g., 0.01 unit length), stop the iteration. At this point, the degraded feature manifold has been aligned to the vicinity of the reference manifold through rotation and translation, and a stable one-to-one correspondence is established between the points on the two manifolds: each point on the degraded feature manifold can uniquely find the closest point on the reference manifold; this correspondence is called cross-domain manifold registration mapping, which is essentially a mapping logic from points on the degraded feature manifold to points on the reference manifold.

[0121] Step 5.2: Based on the cross-domain manifold registration mapping, calculate the local manifold distance between each node on the degraded feature manifold and its corresponding node on the cross-scene reference manifold to obtain the node-level covariance deviation value. Specifically, this includes: based on the obtained cross-domain manifold registration mapping, for each point on the degraded feature manifold, directly obtain its corresponding point on the cross-scene reference manifold; now it is necessary to calculate the local manifold distance between these two corresponding points, which reflects the degree of actual degraded deviation from the standard reference in the current local region; the specific calculation process is as follows: taking the current point on the degraded feature manifold as the center, take all points in its neighborhood to form a local point set, where the neighborhood radius is selected by: calculating the bounding box size of the entire degraded feature manifold in three-dimensional space, that is, the maximum span in the three coordinate axis directions, and taking one-tenth of the longest side length of the bounding box as the neighborhood radius.

[0122] A tangent plane is fitted onto this local point set. Specifically, the local point set is centered (by subtracting the mean of the point set), and the covariance matrix of the centered point set is calculated. Eigenvalue decomposition is performed on this covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is taken as the normal vector of the tangent plane. The eigenvectors corresponding to the largest and second largest eigenvalues ​​span the tangent plane, which passes through the center point of the local point set. Further, the current point on the degraded feature manifold is projected onto this tangent plane: the coordinates of the projected point are equal to the coordinates of the current point minus its component in the direction of the normal vector. Simultaneously, the corresponding point on the reference manifold is also projected onto the same tangent plane using the same projection method. Based on this, the Euclidean coefficient between these two projected points is calculated. Distance, which is the straight-line distance between two points in three-dimensional space, is the local distance of the manifold at that location. Repeat the above steps for each point on the degraded feature manifold: select its local neighborhood, fit the tangent plane, project, and calculate the distance. Each point will obtain a corresponding local distance value, the magnitude of which represents the degree of geometric deviation between the actual degraded feature and the cross-scene benchmark under the sampling window. Arrange these distance values ​​in the time order of the sampling window to form a one-dimensional sequence, called the node-level covariant deviation value sequence. The length of this sequence is equal to the total number of sampling windows. Each value in the sequence corresponds to the local deviation intensity of a specific sampling window, providing basic data for the statistical aggregation in step 5.3.

[0123] Step 5.3: Statistically aggregate all node-level covariance deviation values ​​to obtain cross-domain covariance measurement results; based on the cross-domain covariance measurement results, calculate the quantified deviation value between the degraded feature vector group and the preset cross-scenario generalization threshold library, specifically including: statistically aggregating all obtained node-level covariance deviation values, first calculating the arithmetic mean of all deviation values, denoted as... Next, calculate the standard deviation of the deviation, denoted as . Further, find the largest peak in the deviation value sequence, and denote this peak as... And calculate the multiple of the peak value relative to the average value, denoted as . ,like When it is close to zero, take it directly. Finally, the deviation values ​​are segmented chronologically, with each segment consisting of 10 windows. The cumulative sum of the deviation values ​​within each segment is calculated, and the segment with the fastest increase in cumulative sum is identified. The slope of the fitted line for that segment is recorded as... The four statistics mentioned above are then summed using pre-trained weighting coefficients. The weighted summation formula is as follows:

[0124] ;

[0125] in The non-negative weight coefficients obtained through training with historical data satisfy the following conditions: ; This is the cross-domain covariance measurement result, which is a dimensionless scalar value. Further, based on this cross-domain covariance measurement result, the deviation quantization value between the degraded feature vector group and the preset cross-scene generalization threshold library is calculated. Specifically, the cross-domain covariance measurement result is input into a pre-fitted linear mapping function, the function expression of which is:

[0126] ;

[0127] in and These are the coefficients obtained through linear regression. , The intercept is... The deviation quantification value is limited to the range of 0 to 1. If the calculated value is less than 0, it is taken as 0, and if it is greater than 1, it is taken as 1. 0 represents no deviation and coincides with the baseline manifold; 1 represents the maximum deviation, which exceeds the upper limit of the threshold library.

[0128] In a preferred embodiment of the present invention, step 5 above may further include:

[0129] Step 5.4: Decompose the deviation quantification value into a first deviation component associated with lithium plating, a second deviation component associated with internal short circuits, a third deviation component associated with tab aging, a fourth deviation component associated with micro-short circuits, and a fifth deviation component associated with seal failure, to obtain a set of categorized deviation quantities. Specifically, this includes: decomposing the deviation quantification value into five deviation components corresponding to specific hidden danger types according to its temporal distribution and its association with the five feature subsets in Step 4.3; the specific method is: using the node-level consensus calculated in Step 5.2... The deviation value sequence is grouped according to the feature subset category (lithium plating, internal short circuit, tab aging, micro short circuit, and seal failure) to which each sampling window belongs in step 4.3. For the deviation values ​​in each group, the mean, maximum value, and growth trend of the group, i.e. the linear regression slope, are calculated. These group statistics are combined into a value representing the deviation intensity of the hazard type, which is called the deviation component of the type. Thus, five deviation components are obtained, corresponding to lithium plating, internal short circuit, tab aging, micro short circuit, and seal failure, respectively. They together form a five-dimensional set of categorized deviations.

[0130] Step 5.5 involves comparing each deviation component in the categorized deviation set with the corresponding early warning threshold in the preset cross-scenario generalization threshold library to determine the confidence level of each hazard type and to determine the early deterioration type based on the confidence level. Specifically, this includes comparing each of the five deviation components in the categorized deviation set with the corresponding early warning threshold in the preset cross-scenario generalization threshold library. The preset cross-scenario generalization threshold library stores, in addition to the cross-scenario baseline manifold, early warning thresholds for five hazard types, including yellow, orange, and red warning thresholds, corresponding to mild, moderate, and severe risks, respectively. For each hazard type, its real-time deviation component is compared with the corresponding threshold.

[0131] The comparison results are as follows: If the value of the deviation component is greater than the red threshold, the confidence level for this type of degradation is set to above 0.95; if it is between orange and red, the confidence level is set to 0.75 to 0.95; if it is between yellow and orange, the confidence level is set to 0.5 to 0.75; if it is below the yellow threshold, the confidence level is set to below 0.5. Simultaneously, if the deviation component exceeds the same threshold level in three consecutive sampling windows, the confidence level is increased by one level, i.e., from yellow to orange. Based on the confidence levels of each type, the early degradation type is determined: the hazard type with the highest confidence level is taken as the current primary degradation type; if the confidence levels of multiple types are close and differ by less than 0.1, it is determined to be a composite degradation type, and multiple types are output simultaneously.

[0132] Step 5.6: Analyze the evolution trend of the deviation component corresponding to the early degradation type to obtain the health degradation characterization parameter; calculate the health degradation degree index based on the health degradation characterization parameter; and obtain the health status early warning instruction based on the early degradation type and the health degradation degree index. Specifically, this includes: analyzing the evolution trend of the deviation component corresponding to the early degradation type determined in Step 5.5; extracting the time-varying sequence of the sub-categorical deviation of this degradation type from Step 5.4, i.e., the deviation component value of each sampling window, denoted as... The sequence was smoothed using a sliding window method with a window length of 5 sampling points and a step size of 1 to eliminate high-frequency noise, resulting in a smoothed sequence. Further calculate the first-order difference: The rate of change curve of the deviation component is obtained; then the second difference is calculated: The acceleration curve is obtained. This represents the deviation component value after smoothing in the (p+1)th window. This is the first-order difference corresponding to the (p+1)th window.

[0133] Extract the sign and magnitude of the rate of change at the current moment (last window) from the rate of change curve, and extract information on whether the trend is accelerating or decelerating from the acceleration curve. Match this information with a pre-set health decay feature library, which stores typical trend parameters corresponding to different degradation types and development stages, such as the rate of change threshold and duration. After matching, three health decay characterization parameters are obtained: decay rate... (The average increase of the deviation component per unit time, taken as the average rate of change of the last 10 windows), decay acceleration (The rate of increase, taken as the average of the second-order differences of the last 5 windows) and the cumulative decay. The total deviation integral from the start time to the current time, i.e., the integral of all windows. The sum of these three parameters is multiplied by the sampling period. Based on these three parameters, a weighted formula is used to calculate the health degradation index, which is:

[0134] ;

[0135] in The non-negative weighting coefficients are obtained by fitting the battery aging test results, satisfying... Furthermore, after normalization, the calculated values ​​within the parentheses are typically between 0 and 1. This is a health degradation index, with a value ranging from 0 to 1. 0 represents perfectly healthy, while 1 represents severely deteriorated condition requiring immediate replacement.

[0136] Finally, based on the early degradation type (from step 5.5) and the health decline index... Generate health status warning instructions: when Timely generate attention observation instructions; when Based on the type of degradation, planned maintenance instructions are generated, such as scheduling maintenance for aging tabs; when The system generates immediate alarm commands based on the type of degradation, such as "high risk of lithium plating, please stop operation"; the alarm command, along with the current degradation type and health degradation index, is output to the user interface or device management terminal.

[0137] like Figure 2 As shown, embodiments of the present invention also provide a lithium battery health status early warning system based on multi-sensor fusion, comprising:

[0138] The spatiotemporal alignment module is used to acquire multi-source sensing signals of lithium batteries under strong dynamic conditions in real time, including electrical signals, thermal signals and vibration signals; and to perform time synchronization and spatial alignment processing on the multi-source sensing signals to obtain an initial multidimensional sensing sequence.

[0139] The covariant decoupling module is used to construct an unsteady multi-physics covariant decoupling model based on the initial multi-dimensional sensing sequence. The unsteady multi-physics covariant decoupling model is used to reconstruct the motion trajectory and perform orthogonal decomposition of multi-axis inertial torque to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, thereby obtaining a pure electrochemical characteristic signal set.

[0140] The mapping network construction and feature extraction module is used to construct a unified mapping network from dynamic operating conditions to electrochemical degradation state based on a pure electrochemical feature signal set and a pre-set interface solvation potential well and stress relaxation coupling mechanism. The pure electrochemical feature signal set is input into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to ligand field distortion and molecular potential field fluctuations, thereby obtaining a set of degradation feature vectors characterizing lithium plating, internal short circuit, tab aging, micro short circuit and seal failure risks.

[0141] The measurement and early warning module is used to perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status early warning instruction.

[0142] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0143] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for early warning of lithium battery health status based on multi-sensor fusion, characterized in that, The method includes: Step 1: Real-time acquisition of multi-source sensor signals from lithium batteries under strong dynamic conditions, including electrical signals, thermal signals, and vibration signals; Time synchronization and spatial alignment processing of the multi-source sensor signals to obtain an initial multi-dimensional sensor sequence. Step 2: Based on the initial multidimensional sensing sequence, construct an unsteady multi-physics covariant decoupling model; based on the unsteady multi-physics covariant decoupling model, perform motion trajectory reconstruction and multi-axis inertial torque orthogonal decomposition to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, and obtain a pure electrochemical feature signal set. Step 3: Based on the pure electrochemical characteristic signal set, and combined with the pre-set interface solvation potential well and stress relaxation coupling mechanism, construct a unified mapping network from dynamic operating conditions to electrochemical degradation state. Step 4: Input the pure electrochemical feature signal set into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to the coordination field distortion and molecular potential field fluctuation, and obtain the deterioration feature vector set characterizing lithium plating, internal short circuit, tab aging, micro short circuit and sealing failure. Step 5: Perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status warning instruction.

2. The lithium battery health status early warning method based on multi-sensor fusion according to claim 1, characterized in that, Real-time acquisition of multi-source sensor signals from lithium batteries under highly dynamic operating conditions, including electrical signals, thermal signals, and vibration signals; The multi-source sensor signals are time-synchronized and spatially aligned to obtain an initial multidimensional sensor sequence, including: Under strong dynamic conditions, the sensor locations corresponding to the electrical signal, thermal signal and vibration signal are mapped to a pre-established three-dimensional contour model of the battery cell through coordinate transformation, forming a set of dynamic monitoring nodes that spatially correspond to the electrode stacking structure inside the battery cell; the strong dynamic conditions include one or more of pulse current impact, wideband mechanical vibration and wide temperature range thermal drift. Based on the dynamic monitoring node set, a multi-source field response topology cluster conforming to the stacking direction of the battery cell electrode is constructed, wherein the local response configuration of the multi-source field response topology cluster is kept in shape consistent with the physical layering of the battery cell. The local interface stress distortion curvature at each node on the multi-source field response topology cluster is extracted, as well as the geodesic deviation between nodes along the potential field dissipation path. Based on the multi-source signal values ​​aligned with the pre-acquired timestamps, an initial multidimensional sensing sequence with unified spatiotemporal coordinates is formed on the topology cluster.

3. The lithium battery health status early warning method based on multi-sensor fusion according to claim 2, characterized in that, Based on the initial multidimensional sensing sequence, an unsteady multi-physics covariant decoupling model is constructed. Using this model, motion trajectory reconstruction and orthogonal decomposition of multi-axis inertial torques are performed to remove dynamic disturbances caused by the superposition of pulsed current and mechanical vibration, resulting in a pure electrochemical characteristic signal set, including: The covariant temporal sequence of electrical signal components, thermal signal components and vibration signal components in a unified spatiotemporal coordinate is separated from the initial multidimensional sensing sequence. A multiphysics covariant tensor space is constructed to characterize the mutual coupling relationship between electrochemical response, thermal response and mechanical response. The multiphysics covariant tensor space is used as the solution basis of the unsteady multiphysics covariant decoupling model. Based on the solution basis, the dynamic disturbance response trajectory excited by pulsed current impact and broadband mechanical vibration is traced in the multi-physics covariant tensor space to obtain the dynamic disturbance response trajectory; based on the dynamic disturbance response trajectory, multi-axis inertial torque orthogonal decomposition is performed along the electrochemical response principal axis and the mechanical vibration principal axis to obtain the independent projection of the dynamic disturbance component on each physical field principal axis; From the initial multidimensional sensing sequence, the independent projections of the dynamic disturbance components on the principal axes of each physical field are filtered out to obtain a set of pure electrochemical characteristic signals stripped of operating condition coupling noise.

4. The lithium battery health status early warning method based on multi-sensor fusion according to claim 3, characterized in that, Based on a set of pure electrochemical characteristic signals, and combined with a pre-defined interface solvation potential well and stress relaxation coupling mechanism, a unified mapping network from dynamic operating conditions to electrochemical degradation states is constructed, including: The pre-set interface solvation potential well and stress relaxation coupling mechanism are obtained, and the pure electrochemical characteristic signal set is expanded under the coupling mechanism to analyze the evolution path of the intermolecular interaction potential well of the solvation sheath at the electrode interface and the temporal correlation of lattice stress relaxation, so as to obtain the interface potential well-stress relaxation temporal correlation mapping. Based on the interface potential well-stress relaxation time-series correlation mapping, an interface potential well-stress relaxation constitutive relation characterizing the intrinsic response of electrochemical degradation is established, and this interface potential well-stress relaxation constitutive relation is used as the internal mapping kernel. Based on the internal mapping kernel, a unified mapping network framework is constructed. In the unified mapping network framework, the impact of dynamic operating condition disturbances is encoded as boundary condition constraints, and the electrochemical degradation state is encoded as output layer state parameters. The pure electrochemical feature signal set is input into a unified mapping network framework, and the internal mapping kernel projects the features onto it to obtain the feature projection result. The output layer state parameters are used to perform a nonlinear transformation on the feature projection result, so that the unified mapping network framework establishes the mapping relationship from dynamic operating condition disturbance to electrochemical degradation state, and the unified mapping network from dynamic operating condition to electrochemical degradation state is constructed.

5. A lithium battery health status early warning method based on multi-sensor fusion according to claim 4, characterized in that, The pure electrochemical feature signal set is input into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to coordination field distortion and molecular potential field fluctuations, thereby obtaining a set of deterioration feature vectors characterizing lithium plating, internal short circuit, tab aging, micro-short circuit, and potential seal failure, including: The pure electrochemical feature signal set is input into the unified mapping network of the dynamic operating condition to the electrochemical degradation state. The unified mapping network calls the internal mapping kernel to perform feature projection on the pure electrochemical feature signal set, so as to map the pure electrochemical feature signal set to the hidden feature space with the coordinate field distortion and molecular potential field fluctuation as the coordinate axis, and obtain the hidden layer electrochemical degradation feature projection. Based on the projection of the electrochemical degradation characteristics of the hidden layer, the output layer state parameters are nonlinearly transformed to obtain a nonlinear transformation response tensor. Based on the nonlinear transformation response tensor, electrochemical response characteristic components are extracted from the tensor along the coordination field distortion axis and the molecular potential field fluctuation axis of the hidden feature space to obtain an electrochemical degradation response characteristic set. From the electrochemical degradation response feature set, the first feature subset associated with lithium plating, the second feature subset associated with internal short circuit, the third feature subset associated with tab aging, the fourth feature subset associated with micro-short circuit, and the fifth feature subset associated with sealing failure are separated and combined to form a degradation feature vector group.

6. The lithium battery health status early warning method based on multi-sensor fusion according to claim 5, characterized in that, The degraded feature vector group is subjected to cross-domain manifold covariance measurement with a preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result. Based on the cross-domain covariance measurement results, the deviation quantification value is calculated, including: Based on the degraded feature vector group, a degraded feature manifold is constructed, and a preset cross-scene generalization threshold library is obtained. The preset cross-scene generalization threshold library stores a cross-scene reference manifold. The degraded feature manifold and the cross-scene reference manifold are registered together to establish a cross-domain manifold covariant correspondence and obtain a cross-domain manifold registration mapping. Based on the cross-domain manifold registration mapping, the local manifold distance between each node on the degraded feature manifold and the corresponding node on the cross-scene benchmark manifold is calculated to obtain the node-level covariance deviation value. Statistical aggregation is performed on all node-level covariance deviation values ​​to obtain cross-domain covariance measurement results; based on the cross-domain covariance measurement results, the deviation quantization value between the degraded feature vector group and the preset cross-scenario generalization threshold library is calculated.

7. A method for early warning of lithium battery health status based on multi-sensor fusion according to claim 6, characterized in that, Based on the deviation quantification value, the early degradation type and health degradation degree of the lithium battery are determined, and a health status early warning instruction is obtained, including: From the deviation quantification value, a first deviation component associated with lithium plating, a second deviation component associated with internal short circuit, a third deviation component associated with tab aging, a fourth deviation component associated with micro short circuit, and a fifth deviation component associated with seal failure are decomposed to obtain a set of categorized deviation quantities. Each deviation component in the categorized deviation set is compared with the warning judgment threshold of the corresponding hidden danger type in the preset cross-scenario generalization threshold library to determine the confidence level of each hidden danger type, and the early deterioration type is determined based on the confidence level. The evolution trend of the deviation component corresponding to the early deterioration type is analyzed to obtain the health decline characterization parameter; the health decline degree index is calculated based on the health decline characterization parameter; and a health status early warning instruction is obtained based on the early deterioration type and the health decline degree index.

8. A lithium battery health status early warning system based on multi-sensor fusion, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The spatiotemporal alignment module is used to collect multi-source sensing signals of lithium batteries under strong dynamic conditions in real time, including electrical signals, thermal signals and vibration signals. The multi-source sensor signals are time-synchronized and spatially aligned to obtain the initial multidimensional sensor sequence. The covariant decoupling module is used to construct an unsteady multi-physics covariant decoupling model based on the initial multi-dimensional sensing sequence. The unsteady multi-physics covariant decoupling model is used to reconstruct the motion trajectory and perform orthogonal decomposition of multi-axis inertial torque to remove the dynamic disturbances generated by the superposition of pulse current and mechanical vibration, and obtain a pure electrochemical feature signal set. The mapping network construction and feature extraction module is used to construct a unified mapping network from dynamic operating conditions to electrochemical degradation state based on a pure electrochemical feature signal set and a pre-set interface solvation potential well and stress relaxation coupling mechanism. The pure electrochemical feature signal set is input into the unified mapping network for feature projection and nonlinear transformation to extract the electrochemical response features corresponding to ligand field distortion and molecular potential field fluctuations, thereby obtaining a set of degradation feature vectors characterizing lithium plating, internal short circuit, tab aging, micro short circuit and seal failure risks. The measurement and early warning module is used to perform cross-domain manifold covariance measurement on the degradation feature vector group and the preset cross-scenario generalization threshold library to obtain the cross-domain covariance measurement result; calculate the deviation quantization value based on the cross-domain covariance measurement result; determine the early degradation type and health degradation degree of the lithium battery based on the deviation quantization value, and obtain the health status early warning instruction.

Images