A lithium battery state of charge monitoring method based on edge computing

Abstract

The application discloses a lithium battery charging state monitoring method based on edge computing, which comprises the following steps: collecting charging monitoring data, performing time alignment and normalization processing, and generating an edge monitoring sequence; inputting an improved DeepESN network, generating a multi-scale state matrix, and constructing a future accessible charging domain; extracting a state propagation feature set, counting effective and safe path changes, and generating a charging domain collapse state; constructing an energy information space-time graph, calculating a fission boundary index, and generating a charging domain fission front; constructing a recovery acceptance graph, searching for a minimum recovery path, and determining a future domain escape bridge; comprehensively analyzing the collapse state and the fission front, determining the recovery ability, and outputting a charging monitoring result. Through the construction of the future accessible charging domain based on the DeepESN network and the fission propagation and escape bridge analysis mechanism, the application realizes the advanced identification and recoverability determination of the lithium battery charging implicit instability state.

Description

Technical Field

[0001] This invention relates to the field of lithium battery state of charge monitoring technology, and in particular to a lithium battery state of charge monitoring method based on edge computing. Background Technology

[0002] With the rapid development of electric vehicles, energy storage power stations, and portable electronic devices, the demand for safety monitoring of lithium batteries during charging is constantly increasing. Existing lithium battery charging status monitoring typically relies on the battery management system to collect data such as cell voltage, charging current, temperature, equalization status, and charging control status, and then uses edge controllers or cloud platforms to judge the status of the charging process in order to achieve early warning of overvoltage, overcurrent, overtemperature, and charging anomalies.

[0003] Most existing methods rely on voltage, current, and temperature thresholds at the current sampling time for judgment, or use conventional time-series prediction models to predict single state variables shortly afterward. When lithium batteries are in the early stages of thermal inertia accumulation, polarization release imbalance, or single-cell divergence expansion, the current voltage and temperature have not yet exceeded limits. Existing methods struggle to detect the latent instability process where the future safe charging path is decreasing, leading to delayed warning results.

[0004] Existing edge-side monitoring methods typically only output normal, abnormal, or risk levels, lacking analysis of the future achievable charging state space, fission propagation boundaries, and recovery paths, and cannot determine whether the current risk state can still return to a stable charging process.

[0005] Therefore, how to provide a lithium battery charging state monitoring method based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a lithium battery charging state monitoring method based on edge computing. This invention fully utilizes edge computing, deep echo state networks, multi-timescale state propagation, and future state space evolution analysis technologies to monitor and determine the collapse of future safety paths, fission propagation boundaries, and recovery paths during the lithium battery charging process. This enables advanced identification and recoverability analysis of latent instability states, and has the advantages of high early warning lead time, strong edge real-time performance, low false judgment rate, and strong interpretability of charging state.

[0007] A lithium battery charging state monitoring method based on edge computing according to an embodiment of the present invention includes:

[0008] Collect charging monitoring data during the lithium battery charging process, perform preprocessing on the charging monitoring data, and generate edge charging monitoring sequences;

[0009] The edge charging monitoring sequence is input into an improved DeepESN network with spectral radius self-adjustment and cross-layer orthogonal coupling mechanism. Under the event-triggered sparse restart control, a fast dynamic state matrix, a medium dynamic state matrix, and a long dynamic state matrix are generated sequentially to obtain the future reachable charging domain.

[0010] Read the state propagation feature set in the future reachable charging domain, count the number of state paths that can maintain thermal stability, polarization release, single-unit consistency and charging propulsion in the future, and generate the future charging domain collapse state based on the change in the number of state paths.

[0011] In the future reachable charging domain, an energy-information dual-weight spatiotemporal graph is established, and the minimum cut-minimum cross-region ratio iterative splitting is performed. The entropy flow density and energy misalignment gradient of each split subgraph are calculated. When both exceed the preset threshold simultaneously, the corresponding backbone state path is marked as the fission propagation boundary, and the fission frontier of the charging domain is generated.

[0012] After the fission front of the charging domain is formed, a recovery and acceptance search is performed on each state connection path in the future reachable charging domain. The minimum state connection path that can simultaneously restore thermal stability, polarization release, individual consistency and charging propulsion continuity is selected and the minimum state connection path is determined as the future domain escape bridge.

[0013] Based on the future charging domain collapse state, the charging domain fission frontier, and the existence state of the future domain escape bridge, the system outputs the hidden instability level, future domain collapse level, fission propagation level, recovery and acceptance state, and unrecoverable collapse state corresponding to the current charging state of the lithium battery.

[0014] Optionally, the charging monitoring data includes individual cell voltage sampling values, total charging current sampling values, individual cell temperature sampling values, ambient temperature sampling values, equalization branch conduction status, charging stage control status, and corresponding timestamp information.

[0015] Optionally, the preprocessing of the charging monitoring data to generate an edge charging monitoring sequence includes:

[0016] The charging monitoring data are processed to align time according to a unified sampling period;

[0017] For missing sampling points, perform neighborhood interpolation to complete the sampling points; for abnormal and abrupt sampling points, perform sliding window elimination.

[0018] Synchronous association is performed on each charging monitoring data based on the unit number, sampling time, and charging stage identifier;

[0019] Normalization is performed on the sampled values ​​of individual cell voltage, total charging current, individual cell temperature, and ambient temperature.

[0020] The normalized charging monitoring data are written into the edge ring buffer in chronological order to generate the edge charging monitoring sequence.

[0021] Optionally, obtaining the future reachable charging domain includes:

[0022] Based on the current change rate, voltage change rate, temperature rise change rate, equalization branch switching state, and charging stage switching state in the edge charging monitoring sequence, an improved three-layer reservoir structure of the DeepESN network is constructed. The first reservoir layer is configured as a fast dynamic reservoir, the second reservoir layer is configured as a medium dynamic reservoir, and the third reservoir layer is configured as a long dynamic reservoir. Input connection weights, intra-layer recursive connection weights, and inter-layer transfer connection weights are established respectively.

[0023] The recursive connection weights within the three layers of the reservoir are read separately, and the maximum eigenvalue corresponding to the recursive connection weights of each layer is calculated. The maximum eigenvalue is compared with the preset target spectral radius of each layer. When the maximum eigenvalue deviates from the corresponding target spectral radius, the recursive connection weights of the layers are scaled proportionally according to the corresponding target spectral radius, so that the fast dynamic reservoir maintains the short memory response, the medium dynamic reservoir maintains the polarization recovery response, and the long dynamic reservoir maintains the thermal inertia and single-cell branching propagation response.

[0024] Orthogonalization is performed on the interlayer transfer connection weights between adjacent reserve pool layers. The state transfer components that are repeated in the next reserve pool layer are removed, and the state transfer components that do not overlap with the previous reserve pool layer are retained, forming an orthogonal coupling channel from the first reserve pool layer to the second reserve pool layer and an orthogonal coupling channel from the second reserve pool layer to the third reserve pool layer.

[0025] The edge charging monitoring sequence is input into the improved DeepESN network according to the sampling time. The fast dynamic reservoir records the transient changes in current, rapid voltage disturbances and transient actions of the equalization branch to generate a fast dynamic state matrix. The medium dynamic reservoir records the changes in polarization release, dynamic impedance recovery and local heat accumulation to generate a medium dynamic state matrix. The long dynamic reservoir records the continuous changes in thermal inertia, changes in individual branch expansion and changes in safe path maintenance to generate a long dynamic state matrix.

[0026] When a charging phase transition, a step change in charging current, or a switching of the equalization branch is detected, a sparse restart is performed on the state unit associated with the event in the corresponding storage pool layer. The historical memory of unassociated state units is retained. A gating time splicing order is generated based on the charging control state, thermal stability state, polarization release state, and individual unit consistency state. The fast dynamic state matrix, medium dynamic state matrix, and long dynamic state matrix are spliced ​​into a set of state propagation candidate paths. State propagation candidate paths that simultaneously satisfy thermal stability acceptance, polarization release acceptance, individual unit consistency acceptance, and charging advancement acceptance are selected. The selected set of state propagation candidate paths is determined as the future reachable charging domain.

[0027] Optionally, generating the future charging domain collapse state based on the change in the number of state paths includes:

[0028] Read each state propagation candidate path in the future reachable charging domain, and extract the state propagation feature set according to the sampling time order corresponding to the path. The state propagation feature set includes state propagation direction, state propagation continuity, state stable propagation region, state split propagation region and state collapse propagation region.

[0029] For each state propagation candidate path, the temperature rise change, polarization release change, single-cell voltage divergence change and charging propulsion change corresponding to the adjacent state nodes in the path are read in sequence. Based on the state succession relationship between adjacent state nodes, the thermal stability succession state, polarization release succession state, single-cell consistency succession state and charging propulsion succession state are marked.

[0030] Candidate paths for state propagation that simultaneously meet the following conditions are marked as valid safe paths: thermal stability acceptance state, polarization release acceptance state, single-unit consistency acceptance state, and charging propulsion acceptance state. Candidate paths for state propagation that do not simultaneously meet the above four conditions are marked as failed paths. The number of valid safe paths and the number of failed paths are counted in the order of sampling time.

[0031] The number of effective safe paths at adjacent sampling times is continuously compared. When the number of effective safe paths continues to decrease and the number of failed paths continues to increase, the corresponding time segment is marked as the future charging domain shrinkage segment. The collapse intensity mark is generated based on the decrease in effective safe paths, the increase in failed paths and the number of continuous samplings within the future charging domain shrinkage segment.

[0032] Based on the collapse intensity marker, the coverage area of ​​the state collapse propagation zone, and the continuation position of the state split propagation zone, the collapse states of the future reachable charging domain are classified to generate future charging domain collapse states. The future charging domain collapse states include non-collapse state, local contraction state, continuous collapse state, and irreversible collapse candidate state.

[0033] Optionally, the generation of the charging domain fission front includes:

[0034] Read the candidate paths for state propagation in the future reachable charging domain, take the state nodes in each path as spatiotemporal graph nodes, connect the state nodes that have thermal stability acceptance, polarization release acceptance, single-cell consistency acceptance or charging promotion acceptance relationships between adjacent sampling times as spatiotemporal graph edges, and record the sampling time, storage pool level and acceptance type corresponding to each spatiotemporal graph edge.

[0035] Energy weights are generated based on the temperature rise, polarization release, individual voltage divergence, and charging propulsion changes corresponding to each spatiotemporal graph edge. Information weights are generated based on the state propagation continuity, state splitting times, state collapse times, and path failure markers corresponding to each spatiotemporal graph edge. The energy weights and information weights are written into the corresponding spatiotemporal graph edges to obtain an energy-information dual-weight spatiotemporal graph.

[0036] Using the stable propagation region and the collapsed propagation region as the splitting endpoints, the minimum cut splitting is performed on the energy-information dual-weight spatiotemporal graph. The number of spatiotemporal graph edges crossing the stable propagation region and the collapsed propagation region, the number of cross-regional state paths, and the number of internal state paths of the split subgraph are counted after each splitting. The minimum cross-regional ratio iterative splitting is performed based on the ratio of the number of cross-regional state paths to the number of internal state paths of the split subgraph.

[0037] For each split subgraph, the increase in the number of failure paths, the increase in state splitting nodes, and the expansion of state collapse nodes within the subgraph are read in the order of sampling time to generate entropy flow density. The directional deviation between temperature rise, polarization release, individual voltage divergence, and charging propulsion changes within the subgraph is also read to generate energy misalignment gradient.

[0038] When the entropy flow density and energy misalignment gradient of the same split subgraph both exceed the corresponding preset threshold in the same sampling segment, the backbone state path connecting the stable propagation region, the state splitting propagation region and the state collapse propagation region in the split subgraph is extracted, and the backbone state path is marked as the fission propagation boundary. The charging domain fission front is generated according to the sampling time, propagation direction, coverage of the reserve pool level and connection of the fission propagation boundary.

[0039] Optionally, determining the minimum state connection path as a future domain escape bridge includes:

[0040] Read the fission propagation boundary, stable propagation region, state splitting propagation region and state collapse propagation region in the fission frontier of the charging domain. Mark the state nodes located at the edge of the state splitting propagation region and the state collapse propagation region as candidate escape starting point nodes. Mark the state nodes located in the stable propagation region and continuously satisfying thermal stability acceptance, polarization release acceptance, single-unit consistency acceptance and charging propagation acceptance as candidate escape target nodes.

[0041] Within the future reachable charging domain, a recovery continuity graph is constructed using escape start point candidate nodes, escape target candidate nodes, and state nodes. State nodes that can form state continuity between adjacent sampling times are connected as recovery continuity edges. Thermal recovery state, polarization release state, single-unit divergence convergence state, charging propulsion maintenance state, and the level of the corresponding storage pool are recorded in the recovery continuity edges.

[0042] Perform acceptance screening on each recovery acceptance edge in the recovery acceptance diagram, retain the recovery acceptance edges that do not have thermal stability due to temperature rise changes, polarization release changes, individual cell voltage divergence changes, and charging propulsion changes, and delete the recovery acceptance edges that do not simultaneously meet the above acceptance conditions.

[0043] Starting from each candidate escape origin node, search for a state connection path to the candidate escape target node along the preserved recovery receiving edge. Sort each state connection path according to the number of path edges, thermal recovery residual, polarization release residual, individual branching residual, and charging propulsion residual. The state connection path that ranks first and can connect the state split propagation region or the state collapse propagation region to the stable propagation region is determined as the minimum state connection path.

[0044] When a minimum state connection path exists, the minimum state connection path, the corresponding escape starting point candidate node, recovery receiving edge, escape target candidate node, and the corresponding reserve pool level are combined into a future domain escape bridge. When no minimum state connection path exists, the corresponding charging domain fission frontier is marked as an escape bridge missing state.

[0045] Optionally, the latent instability level, future domain collapse level, fission propagation level, recovery state, and unrecoverable collapse state corresponding to the current charging state of the output lithium battery include:

[0046] Read the changes in the number of effective safe paths, the changes in the number of failed paths, the collapse intensity marker, the coverage of the state collapse propagation zone, and the connection position of the state split propagation zone in the future charging domain collapse state, and generate the future domain collapse level according to the non-collapse state, the local contraction state, the continuous collapse state, and the irrecoverable collapse candidate state.

[0047] Read the fission propagation boundary, fission expansion speed, fission propagation direction, fission coverage, fission propagation level and continuous sampling number in the fission front of the charging domain, and generate the fission propagation level based on whether the fission propagation boundary continuously connects the stable propagation region, the state splitting propagation region and the state collapse propagation region.

[0048] Read the existence status of the future domain escape bridge, the candidate nodes of the escape start point, the recovery and acceptance edge, the candidate nodes of the escape target and the level of the reserve pool to which they belong. When the future domain escape bridge exists, a recoverable acceptance state is generated. When the future domain escape bridge does not exist and the corresponding charging domain fission front is in a continuous expansion state, a recovery and acceptance failure state is generated.

[0049] Based on the future domain collapse level, fission propagation level, and recovery acceptance status, a joint determination is performed on the current charging state of the lithium battery, and the implicit instability level, future domain collapse level, fission propagation level, recovery acceptance status, and unrecoverable collapse status are output. The corresponding sampling time, abnormal impact cell number, associated fission propagation boundary, and future domain escape bridge identifier are written into the edge monitoring result record.

[0050] The beneficial effects of this invention are:

[0051] This invention constructs a future-reachable charging domain based on a DeepESN network to continuously model the state propagation across multiple time scales during lithium battery charging. It is no longer limited to static monitoring of current voltage, current, and temperature values, but dynamically analyzes the ability to maintain a safe charging path in the future. This allows for early identification of the decreasing trend of the future safe path before thermal inertia accumulation, polarization release imbalance, and individual cell divergence expansion form explicit limits, thus improving the early warning capability of lithium battery latent instability states.

[0052] This invention establishes an energy-information dual-weighted spatiotemporal graph and combines it with a fission propagation boundary identification mechanism to decompose and analyze the state propagation path in the future reachable charging domain. It can identify the fission frontier expanding from the stable propagation region to the state collapse propagation region, and realize continuous tracking of the future charging domain collapse process. Compared with traditional methods that rely solely on threshold alarms or short-term predictions, it can more accurately detect the contraction process of the future safe state space and reduce the false positive and false negative rates.

[0053] This invention further utilizes a future domain escape bridge search mechanism to search and filter recovery paths in the future reachable charging domain. This allows for the determination of whether the current charging state still has the ability to recover to a stable charging domain, thus enabling the assessment of the recoverability of the lithium battery charging state. When no future domain escape bridge exists, it can identify unrecoverable collapse states in advance, providing a basis for edge-side charging control, equalization adjustment, and risk intervention, thereby improving the safety and stability of the lithium battery charging process. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1This is a flowchart of a lithium battery charging state monitoring method based on edge computing proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the structure of the future reachable charging domain based on the DeepESN network for the lithium battery charging state monitoring method based on edge computing proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1 and Figure 2 A lithium battery charging state monitoring method based on edge computing, comprising:

[0059] Collect charging monitoring data during the lithium battery charging process, perform preprocessing on the charging monitoring data, and generate edge charging monitoring sequences;

[0060] The edge charging monitoring sequence is input into an improved DeepESN network with spectral radius self-adjustment and cross-layer orthogonal coupling mechanism. Under the event-triggered sparse restart control, a fast dynamic state matrix, a medium dynamic state matrix, and a long dynamic state matrix are generated sequentially to obtain the future reachable charging domain.

[0061] Read the state propagation feature set in the future reachable charging domain, count the number of state paths that can maintain thermal stability, polarization release, single-unit consistency and charging propulsion in the future, and generate the future charging domain collapse state based on the change in the number of state paths.

[0062] In the future reachable charging domain, an energy-information dual-weight spatiotemporal graph is established, and the minimum cut-minimum cross-region ratio iterative splitting is performed. The entropy flow density and energy misalignment gradient of each split subgraph are calculated. When both exceed the preset threshold simultaneously, the corresponding backbone state path is marked as the fission propagation boundary, and the fission frontier of the charging domain is generated.

[0063] After the fission front of the charging domain is formed, a recovery and acceptance search is performed on each state connection path in the future reachable charging domain. The minimum state connection path that can simultaneously restore thermal stability, polarization release, individual consistency and charging propulsion continuity is selected and the minimum state connection path is determined as the future domain escape bridge.

[0064] Based on the future charging domain collapse state, the charging domain fission frontier, and the existence state of the future domain escape bridge, the system outputs the hidden instability level, future domain collapse level, fission propagation level, recovery and acceptance state, and unrecoverable collapse state corresponding to the current charging state of the lithium battery.

[0065] In this embodiment, the charging monitoring data includes individual cell voltage sampling value, total charging current sampling value, individual cell temperature sampling value, ambient temperature sampling value, equalization branch conduction status, charging stage control status, and corresponding timestamp information.

[0066] In this embodiment, the step of preprocessing the charging monitoring data to generate an edge charging monitoring sequence includes:

[0067] The charging monitoring data are processed to align time according to a unified sampling period;

[0068] For missing sampling points, perform neighborhood interpolation to complete the sampling points; for abnormal and abrupt sampling points, perform sliding window elimination.

[0069] Synchronous association is performed on each charging monitoring data based on the unit number, sampling time, and charging stage identifier;

[0070] Normalization is performed on the sampled values ​​of individual cell voltage, total charging current, individual cell temperature, and ambient temperature.

[0071] The normalized charging monitoring data are written into the edge ring buffer in chronological order to generate the edge charging monitoring sequence.

[0072] In this embodiment, obtaining the future reachable charging domain includes:

[0073] Based on the current change rate, voltage change rate, temperature rise change rate, equalization branch switching state, and charging stage switching state in the edge charging monitoring sequence, an improved three-layer reservoir structure for the DeepESN network is constructed. The first reservoir layer is configured as a fast dynamic reservoir, the second reservoir layer as a medium dynamic reservoir, and the third reservoir layer as a long dynamic reservoir. Input connection weights, intra-layer recursive connection weights, and inter-layer transitive connection weights are established respectively. Specifically, the improved three-layer reservoir structure for the DeepESN network is as follows:

[0074] The input connection weight matrix of the first storage pool layer is randomly generated. The channel weights of voltage change rate and current change rate are set to high density sparse distribution, while the channel weights of temperature rise change rate, equalization branch switching state and charging stage switching state are set to low density sparse distribution to enhance the transient sensitivity to fast dynamic signals.

[0075] Recursive connection weights are generated within the first reservoir layer at a fixed sparsity rate. Then, spectral radius scaling is applied to the recursive weights to keep the memory time of this layer constant at several dozen sampling periods and to record fast dynamic features such as current steps, pulse disturbances and voltage jitter.

[0076] The input connection weights of the second reservoir layer are randomly generated, and the output signal is mapped from the first reservoir layer to the second reservoir layer. The input components related to polarization release and impedance recovery changes are retained by orthogonal component screening, and the components that are repeated with fast dynamics are deleted.

[0077] Recursive connection weights are generated and the spectral radius is scaled within the second reservoir layer, so that the memory time of this layer is often extended to hundreds of sampling periods to capture dynamic features such as polarization release, dynamic impedance recovery and local thermal accumulation.

[0078] The input connection weights of the third reserve pool layer are randomly generated, and the state outputs from the second reserve pool layer are mapped to the third reserve pool layer. Orthogonalization is used to remove duplicate components and retain the components related to thermal inertia and individual branching expansion.

[0079] Recursive connection weights are generated and the spectral radius is scaled within the third reserve pool layer so that the memory time of this layer often covers thousands of sampling periods, which is used to capture information on long-period thermal inertial propagation, single-cell branching expansion and safe path evolution.

[0080] Establish inter-layer transfer connection weights between the three-layer reservoirs, perform weight normalization and channel masking processing on cross-layer transmitted signals, and ensure that fast dynamic, medium dynamic and long dynamic information are clearly transmitted and do not get confused at their respective time scales, thus completing the construction of the improved three-layer reservoir structure of the DeepESN network.

[0081] The intra-layer recursive connection weights of the three reservoir layers are read separately, and the maximum eigenmode value corresponding to each layer's recursive connection weight is calculated. The maximum eigenmode value is compared with the preset target spectral radius of each layer. When the maximum eigenmode value deviates from the corresponding target spectral radius, the layer recursive connection weights are scaled proportionally according to the corresponding target spectral radius. This ensures that the fast dynamic reservoir maintains a short memory response, the medium dynamic reservoir maintains a polarization recovery response, and the long dynamic reservoir maintains thermal inertia and single-cell branching propagation response. Specifically, the calculation of the maximum eigenmode value corresponding to each layer's recursive connection weight is as follows:

[0082] For the first reserve pool, a random unit vector is selected as the initial right eigenvector for the recursive connection weight matrix. The weight matrix is ​​multiplied by the vector and then the amplitude is normalized in a loop. The iteration stops when the Euclidean distance between two consecutive normalized vectors is lower than the preset error threshold. The amplitude of the weight matrix projected onto the normalized vector during the iteration process is taken as the maximum eigenvalue of the fast dynamic reserve pool.

[0083] The power iteration calculation of the recursive connection weight matrix of the second reserve pool is performed in the same way to obtain the maximum eigenvalue of the dynamic reserve pool and record the corresponding polarization recovery weight response amplitude.

[0084] Continue to perform exponential iteration calculation on the recursive connection weight matrix of the third reserve pool to obtain the maximum eigenvalue of the long dynamic reserve pool, and record the corresponding thermal inertia propagation weight response amplitude.

[0085] The maximum characteristic root modulus values ​​of the three-layer reservoirs are compared with the preset target spectral radius. When the maximum characteristic root modulus value of the fast dynamic reservoir is greater than the target spectral radius, its recursive connection weights are scaled proportionally to maintain the short memory response. When the maximum characteristic root modulus value of the medium dynamic reservoir deviates from the target spectral radius, its recursive connection weights are scaled proportionally to maintain the polarization recovery response. When the maximum characteristic root modulus value of the long dynamic reservoir deviates from the target spectral radius, its recursive connection weights are scaled proportionally to maintain the thermal inertia and single-cell branching propagation response.

[0086] Orthogonalization is performed on the inter-layer transfer connection weights between adjacent reserve pool layers. State transfer components that overlap with those of the preceding reserve pool layer in the subsequent layer are removed, while non-overlapping state transfer components are retained. This forms orthogonal coupling channels from the first reserve pool layer to the second reserve pool layer and from the second reserve pool layer to the third reserve pool layer. Specifically, the formation of these orthogonal coupling channels is as follows:

[0087] Between the first and second reserve pool layers, the output state matrix of the first reserve pool and the corresponding inter-layer transfer connection weights are read. The state matrix is ​​normalized by column and used as a projection basis. The projection components of the input weights of the second reserve pool on this basis are calculated.

[0088] The projected components are removed from the input weights of the second reserve pool by vector difference to obtain the remaining components that are orthogonal to the output of the first reserve pool. The remaining components are then used to reconstruct new interlayer transfer connection weights to form an orthogonal coupling channel from the first reserve pool layer to the second reserve pool layer.

[0089] The same column normalization process is performed on the output state matrix of the second reserve pool, which is used as the projection basis. The projection components of the input weights of the third reserve pool on this basis are calculated and vector difference is performed to remove them, so as to obtain the remaining components that are orthogonal to the output of the second reserve pool.

[0090] The inter-layer transfer connection is reconstructed using the remaining weights after removing duplicate components, forming an orthogonal coupling channel from the second reserve pool layer to the third reserve pool layer, and the two orthogonal coupling channels are written into the network structure configuration.

[0091] The edge charging monitoring sequence is sequentially input into the improved DeepESN network according to the sampling time. A fast dynamic state matrix is ​​generated by recording transient changes in current, rapid voltage disturbances, and transient actions of the equalization branch using a fast dynamic reservoir. A medium dynamic state matrix is ​​generated by recording polarization release changes, dynamic impedance recovery changes, and local heat accumulation changes using a medium dynamic reservoir. A long dynamic state matrix is ​​generated by recording continuous changes in thermal inertia, individual branch expansion changes, and safe path maintenance changes using a long dynamic reservoir. Where:

[0092] The fast dynamic state matrix is ​​generated as follows: the charging monitoring vector is input into the first storage tank layer according to the sampling time order, the transient activation value of all state units in the layer is recorded at each sampling time, and the continuously recorded activation vectors are stacked into a two-dimensional matrix in time order, where the row index corresponds to the sampling time and the column index corresponds to the state unit number. The matrix is ​​updated in real time to save the fast dynamic characteristics caused by current step, pulse disturbance and equalization branch switching.

[0093] The generation of the dynamic state matrix is ​​as follows: the output of the first reservoir is transmitted to the second reservoir layer through an orthogonal coupling channel, the activation value of the second layer state unit is recorded according to the same sampling rhythm, and the continuous activation vectors are stacked into a two-dimensional matrix in time order. The matrix column vectors continuously accumulate the dynamic characteristics corresponding to the polarization release fluctuation, the slow recovery of dynamic impedance, and the local heat accumulation process.

[0094] The long dynamic state matrix is ​​generated by transmitting the output of the second reservoir to the third reservoir layer through an orthogonal coupling channel, continuously recording the activation values ​​of the state units in the third layer and stacking them into a two-dimensional matrix in chronological order. The matrix column vectors maintain long dynamic characteristics such as increasing thermal inertia, individual branch expansion and future safe path changes over a long period of time, and finally obtain the fast dynamic state matrix, the medium dynamic state matrix and the long dynamic state matrix.

[0095] When a charging phase transition, a step change in charging current, or a switching of the equalization branch is detected, a sparse restart is performed on the state unit associated with the event in the corresponding reservoir layer. The historical memory of unassociated state units is retained. A gating time concatenation order is generated based on the charging control state, thermal stability state, polarization release state, and individual unit consistency state. The fast dynamic state matrix, medium dynamic state matrix, and long dynamic state matrix are concatenated to form a set of state propagation candidate paths. State propagation candidate paths that simultaneously satisfy thermal stability acceptance, polarization release acceptance, individual unit consistency acceptance, and charging propagation acceptance are selected. The selected set of state propagation candidate paths is determined as the future reachable charging domain. Specifically, the selection of the selected set of state propagation candidate paths as the future reachable charging domain is as follows:

[0096] For the spliced ​​set of candidate paths for state propagation, the temperature rise amplitude, voltage rebound amplitude, single-cell voltage difference amplitude and total charging current stability of each path are read sequentially during each sampling period to generate corresponding thermal stability markers, polarization release markers, single-cell consistency markers and charging advancement markers.

[0097] For each state propagation candidate path, a succession check is performed. When the thermal stability marker, polarization release marker, single-cell consistency marker, and charging propulsion marker are all valid during the continuous sampling period, the path is marked as a valid succession path.

[0098] Write all valid connection paths into the path index table in the order of sampling time, retain the path start time, end time, cross-level identifier and connection duration, and delete path records that do not meet the valid connection criteria.

[0099] All valid receiving paths in the path index table are combined into a filtered state propagation path set. The sampling period, reserve pool level, and number of paths covered by the set are recorded, and the set is determined as the future reachable charging domain.

[0100] In this embodiment, generating the future charging domain collapse state based on the change in the number of state paths includes:

[0101] Read each state propagation candidate path in the future reachable charging domain, and extract the state propagation feature set according to the sampling time order corresponding to the path. The state propagation feature set includes state propagation direction, state propagation continuity, state stable propagation region, state split propagation region and state collapse propagation region.

[0102] For each state propagation candidate path, the temperature rise change, polarization release change, single-cell voltage divergence change and charging propulsion change corresponding to the adjacent state nodes in the path are read in sequence. Based on the state succession relationship between adjacent state nodes, the thermal stability succession state, polarization release succession state, single-cell consistency succession state and charging propulsion succession state are marked.

[0103] Candidate paths for state propagation that simultaneously meet the following conditions are marked as valid safe paths: thermal stability acceptance state, polarization release acceptance state, single-unit consistency acceptance state, and charging propulsion acceptance state. Candidate paths for state propagation that do not simultaneously meet the above four conditions are marked as failed paths. The number of valid safe paths and the number of failed paths are counted in the order of sampling time.

[0104] The number of valid safe paths at adjacent sampling times is continuously compared. When the number of valid safe paths continuously decreases and the number of failed paths continuously increases, the corresponding time segment is marked as a future charging domain contraction segment. A collapse intensity marker is generated based on the decrease in valid safe paths, the increase in failed paths, and the number of continuous samplings within the future charging domain contraction segment. Specifically, the generation of the collapse intensity marker is as follows:

[0105] Within the future charging domain contraction zone, the number of effective safe paths at the beginning and end of the zone is recorded, and the difference between the two is calculated as the effective path reduction rate. The increase in the number of failed paths within the zone is recorded as the failed path increment. The number of continuous samplings within the zone is then counted. The effective path reduction rate and the failed path increment are weighted and summed according to a preset proportional coefficient, and then divided by the number of continuous samplings to obtain the collapse intensity used to measure the strength of the contraction. The collapse intensity is compared with the four threshold intervals set by the system and marked as weak collapse, medium collapse, strong collapse, or unrecoverable collapse level, respectively. The results are written into the collapse intensity label.

[0106] Based on the collapse intensity marker, the coverage area of ​​the state collapse propagation zone, and the continuation position of the state split propagation zone, the collapse states of the future reachable charging domain are classified to generate future charging domain collapse states. These future charging domain collapse states include non-collapsed states, partially contracted states, continuous collapse states, and unrecoverable collapse candidate states. Specifically, the generation of future charging domain collapse states is as follows:

[0107] The system reads the collapse intensity marker and, combined with the proportion of time the collapse propagation area occupies in the future charging domain and the number of connections between the split propagation area and the collapse propagation area, classifies the collapse state of the future reachable charging domain: when the collapse intensity is zero and the proportion of the collapse propagation area is zero, it is marked as a non-collapsed state; when the collapse intensity is at the lowest level, the proportion of the collapse propagation area is less than one-fifth, and the number of connections does not exceed one, it is marked as a local contraction state; when the collapse intensity is at a medium level or the proportion of the collapse propagation area is between one-fifth and about one-half, and the number of connections exceeds one, it is marked as a continuous collapse state; when the collapse intensity reaches the highest level or the proportion of the collapse propagation area exceeds about one-half, and the number of connections continues to increase, it is marked as an unrecoverable collapse candidate state, and the corresponding time period is written into the future charging domain collapse state record.

[0108] In this embodiment, the generation of the charging domain fission front includes:

[0109] Read the candidate paths for state propagation in the future reachable charging domain, take the state nodes in each path as spatiotemporal graph nodes, connect the state nodes that have thermal stability acceptance, polarization release acceptance, single-cell consistency acceptance or charging promotion acceptance relationships between adjacent sampling times as spatiotemporal graph edges, and record the sampling time, storage pool level and acceptance type corresponding to each spatiotemporal graph edge.

[0110] Energy weights are generated based on the temperature rise, polarization release, individual voltage divergence, and charging propulsion changes corresponding to each spatiotemporal graph edge. Information weights are generated based on the state propagation continuity, state splitting frequency, state collapse frequency, and path failure markers corresponding to each spatiotemporal graph edge. The energy weights and information weights are then written into the corresponding spatiotemporal graph edges to obtain an energy-information dual-weighted spatiotemporal graph. Specifically, the energy-information dual-weighted spatiotemporal graph is as follows:

[0111] For each spatiotemporal graph edge, read its corresponding temperature rise change, polarization release change, single-cell voltage divergence change and charging propulsion change, and linearly superimpose the four changes according to a preset ratio to obtain the energy weight of the edge.

[0112] For the same spatiotemporal graph edge, read its state propagation continuity marker, state splitting count, state collapse count, and path failure marker, and linearly superimpose the four indicators according to a preset ratio to obtain the information weight of the edge.

[0113] Write the energy weight to the energy weight field of the spatiotemporal graph edge, write the information weight to the information weight field of the same spatiotemporal graph edge, and update the double weight data of all spatiotemporal graph edges according to the time index.

[0114] Traverse the entire spatiotemporal graph to confirm that all spatiotemporal graph edges contain both energy weight field and information weight field, and generate a complete energy-information dual-weight spatiotemporal graph.

[0115] Using the stable propagation region and the collapsed propagation region as the splitting endpoints, a minimum cut split is performed on the energy-information dual-weight spatiotemporal graph. After each split, the number of spatiotemporal graph edges crossing the stable propagation region and the collapsed propagation region, the number of cross-regional state paths, and the number of internal state paths in the split subgraph are counted. A minimum cross-regional ratio iterative split is performed based on the ratio of the number of cross-regional state paths to the number of internal state paths in the split subgraph. Specifically, the minimum cross-regional ratio iterative split is performed as follows:

[0116] Select the stable propagation region and the collapsed propagation region as the source and sink ends, respectively. Perform a minimum cut operation on the energy-information dual-weighted spatiotemporal graph to cut off all spatiotemporal graph edges that cross the two types of regions, and obtain the first round of split subgraph.

[0117] The number of spatiotemporal graph edges that still connect the stable propagation region and the collapsed propagation region after splitting is counted, the number of remaining state paths between the two types of segments is counted, and the number of state paths within each split subgraph is also counted.

[0118] Calculate the ratio of the number of remaining state paths between the two types of segments to the number of state paths within the split subgraph. When the ratio is greater than the cross-segment ratio threshold set by the system, select the split subgraph with the highest cross-segment ratio as the new graph and repeat the minimum cut operation.

[0119] After each iteration, the number of cross-region edges and the number of cross-region state paths are recounted and compared with the results of the previous round. The iteration is terminated when the number of cross-region edges or the number of cross-region state paths no longer decreases and the cross-region ratio is lower than the threshold. Finally, the splitting result that meets the minimum cross-region ratio requirement is obtained.

[0120] For each split subgraph, the increase in the number of failure paths, the increase in state splitting nodes, and the expansion of state collapse nodes within the subgraph are read in chronological order of sampling time to generate entropy-increasing flow density. Furthermore, the directional deviations between temperature rise changes, polarization release changes, individual cell voltage divergence changes, and charging propulsion changes within the subgraph are read to generate energy misalignment gradients, where the entropy-increasing flow density... for:

[0121] ;

[0122] in, This represents the entropy flow density of the q-th split subgraph within a decision window; This represents the increase in the number of failed paths within the decision window for the q-th subgraph. This represents the total number of state paths of the q-th subgraph at the start of the decision window; This represents the increase in the number of state split nodes of the q-th split subgraph within the decision window; This represents the increase in the number of state collapse nodes of the q-th split subgraph within the decision window; This represents the total number of state nodes in the q-th subgraph. This indicates the number of sampling steps corresponding to the decision window; , , These represent the weight coefficients corresponding to the increase in the number of failed paths, the increase in the number of state split nodes, and the increase in the number of state collapse nodes, respectively. + + =1, in practical implementation Take 0.4, Take 0.3, Take 0.3;

[0123] Energy misalignment gradient for:

[0124] ;

[0125] in, This represents the energy misalignment gradient of the q-th split subgraph within the decision window; This indicates the number of sampling segments contained within the decision window for the q-th subgraph; This represents the energy change direction vector of the q-th subgraph in the m-th sampling segment; This represents the energy change direction vector of the q-th subgraph in the previous sampling segment; . Represents the dot product of two energy change direction vectors; | | and | | represent the magnitudes of the two energy change direction vectors, respectively; To prevent extremely small positive numbers with a denominator of zero, the value is set to 0.000001. Wherein, the energy change direction vector... Build it as follows:

[0126] ;

[0127] in, This shows the normalized temperature rise change of the qth subplot within the mth sampling segment; This indicates the change in normalized polarization release insufficiency; This represents the normalized change in individual cell voltage divergence. This represents the normalized charge propulsion attenuation.

[0128] Within each split subgraph, the hourly increment of the number of failed paths is counted in the order of sampling time, and the hourly increment of state split nodes and state collapse nodes is also counted.

[0129] The three types of hourly increments are accumulated over time to obtain the total increment of the failure path, the total increment of the state splitting node, and the total increment of the state collapse node. Then, the average growth rate of the three in the same time period is calculated, and the arithmetic mean of the average growth rate is taken as the entropy flow density.

[0130] The temperature rise, polarization release, individual cell voltage divergence, and charging propulsion changes were read in the order of sampling time, and the direction of increase or decrease of the four types of changes in adjacent time periods was determined.

[0131] Compare the four types of change directions in pairs within the same time period, and count the number of logs with inconsistent directions and the number of logs with consistent directions. The proportion of the number of logs with inconsistent directions to the total number of logs is the energy misalignment gradient.

[0132] When the entropy flux density and energy misalignment gradient of the same subgraph both exceed the corresponding preset thresholds within the same sampling segment, the backbone state path connecting the stable propagation region, the state splitting propagation region, and the state collapse propagation region in the subgraph is extracted. The backbone state path is marked as the fission propagation boundary. The charging domain fission front is generated according to the sampling time, propagation direction, coverage of the reservoir level, and connection of the state nodes of the fission propagation boundary. Specifically, the generation of the charging domain fission front is as follows:

[0133] Within the split subgraph that meets the threshold condition, the shortest connected path is traced along the continuous spatiotemporal graph edges of the connected state stable propagation region, state split propagation region, and state collapse propagation region, and the connected path is determined as the backbone state path.

[0134] Record the start and end sampling times of the backbone state path on the time axis. The start time represents the time when the fission propagation front appears, and the end time represents the latest position of the fission propagation front at the current time.

[0135] Based on the directional projection of the backbone state path in the spatiotemporal diagram, the direction of fission propagation is determined, and the reserve pool levels crossed by the backbone state path are counted, indicating the fast dynamic layer, medium dynamic layer or long dynamic layer involved.

[0136] The system summarizes all state node numbers, time ranges, propagation directions, and cross-level information contained in the backbone state path to generate a charging domain fission frontier record.

[0137] When the entropy flux density and energy misalignment gradient of the same subgraph both exceed the corresponding preset thresholds within the same sampling segment, the backbone state path connecting the stable propagation region, the state splitting propagation region, and the state collapse propagation region in the subgraph is extracted. The backbone state path is marked as the fission propagation boundary. The charging domain fission front is generated according to the sampling time, propagation direction, coverage of the reservoir level, and connection of the state nodes of the fission propagation boundary. Specifically, the generation of the charging domain fission front is as follows:

[0138] Read historical charging samples that are the same type of lithium battery to be monitored, have the same charging rate, or are adjacent to each other. Divide the historical charging samples into stable charging samples and confirmed fission samples. The stable charging samples are those that do not show future charging domain collapse, thermal runaway warning, continuous expansion of single cell pressure difference, or charging interruption during the entire charging process. The confirmed fission samples are those that show a continuous expansion from the stable propagation region to the state splitting propagation region or the state collapse propagation region during the charging process, and are verified by subsequent temperature rise, pressure difference expansion, or charging control derating.

[0139] For stable charging samples and confirmed fission samples, the following steps are performed: construction of the future reachable charging domain, construction of the energy-information dual-weight spatiotemporal graph, minimum cut-span iterative splitting, and calculation of the splitting subgraph index. This yields the entropy flow density sample set and energy misalignment gradient sample set corresponding to the stable charging samples, as well as the entropy flow density sample set and energy misalignment gradient sample set corresponding to the confirmed fission samples.

[0140] Calculate the mean and standard deviation of the entropy flux density sample set in the stable charging samples, and use the sum of the mean and three times the standard deviation as the initial threshold of the entropy flux density; calculate the mean and standard deviation of the energy dislocation gradient sample set in the stable charging samples, and use the sum of the mean and three times the standard deviation as the initial threshold of the energy dislocation gradient.

[0141] Calculate the 5th percentile of the entropy flow density sample set in the confirmed fission samples and compare it with the initial threshold of entropy flow density. When the initial threshold of entropy flow density is less than the 5th percentile of the entropy flow density sample set in the confirmed fission samples, the initial threshold of entropy flow density is determined as the preset threshold of entropy flow density. When the initial threshold of entropy flow density is not less than the 5th percentile of the entropy flow density sample set in the confirmed fission samples, the average of the 95th percentile of the entropy flow density sample set in the stable charging samples and the 5th percentile of the entropy flow density sample set in the confirmed fission samples is determined as the preset threshold of entropy flow density.

[0142] Calculate the 5th percentile of the energy misalignment gradient sample set in the confirmed fission samples and compare it with the initial threshold of the energy misalignment gradient. When the initial threshold of the energy misalignment gradient is less than the 5th percentile of the energy misalignment gradient sample set in the confirmed fission samples, the initial threshold of the energy misalignment gradient is determined as the preset threshold of the energy misalignment gradient. When the initial threshold of the energy misalignment gradient is not less than the 5th percentile of the energy misalignment gradient sample set in the confirmed fission samples, the average of the 95th percentile of the energy misalignment gradient sample set in the stable charging samples and the 5th percentile of the energy misalignment gradient sample set in the confirmed fission samples is determined as the preset threshold of the energy misalignment gradient.

[0143] A threshold configuration table is established according to battery type and charging rate. The preset threshold for entropy current density of lithium iron phosphate batteries under constant current charging conditions of 0.5C to 0.7C is set to 0.35 to 0.45, and the preset threshold for energy misalignment gradient is set to 0.35 to 0.48. The preset threshold for entropy current density of ternary lithium batteries under constant current charging conditions of 0.5C to 0.7C is set to 0.32 to 0.42, and the preset threshold for energy misalignment gradient is set to 0.34 to 0.46. When the charging rate is greater than 0.7C but not greater than 1C, the preset thresholds for entropy current density and energy misalignment gradient under the corresponding battery type are increased by 0.03 to 0.05 respectively.

[0144] During edge monitoring, the current lithium battery type, current charging rate, and current charging stage are read. The corresponding entropy flow density preset threshold and energy misalignment gradient preset threshold are retrieved from the threshold configuration table. When the current operating condition is not directly matched in the threshold configuration table, the threshold record of the same type of battery with the closest charging rate is selected, and linear correction is performed based on the difference between the current charging rate and the charging rate corresponding to the threshold record to obtain the entropy flow density preset threshold and energy misalignment gradient preset threshold under the current operating condition.

[0145] When the entropy flow density in the same sampling segment of the same split subgraph is greater than the preset threshold of the entropy flow density under the current working condition, and the energy misalignment gradient is greater than the preset threshold of the energy misalignment gradient under the current working condition, the backbone state path that crosses the stable propagation region, the state split propagation region and the state collapse propagation region in the split subgraph is extracted, and the fission propagation boundary is generated according to the sampling time, state node number, reserve pool level number, path propagation direction and connection edge weight of the path.

[0146] All fission propagation boundaries within the same sampling segment are merged according to propagation direction and coverage reservoir level. Boundary segments that repeatedly pass through the same state node and have the same propagation direction are deleted. Boundary segments at the connection positions of the coverage state split propagation region and the state collapse propagation region are retained. The retained fission propagation boundaries are combined to generate the charging domain fission frontier.

[0147] In this embodiment, determining the minimum state connection path as a future domain escape bridge includes:

[0148] Read the fission propagation boundary, stable propagation region, state splitting propagation region and state collapse propagation region in the fission frontier of the charging domain. Mark the state nodes located at the edge of the state splitting propagation region and the state collapse propagation region as candidate escape starting point nodes. Mark the state nodes located in the stable propagation region and continuously satisfying thermal stability acceptance, polarization release acceptance, single-unit consistency acceptance and charging propagation acceptance as candidate escape target nodes.

[0149] Within the future reachable charging domain, a recovery continuity graph is constructed using escape start point candidate nodes, escape target candidate nodes, and state nodes. State nodes that can form state continuity between adjacent sampling times are connected as recovery continuity edges. Thermal recovery state, polarization release state, single-unit divergence convergence state, charging propulsion maintenance state, and the level of the corresponding storage pool are recorded in the recovery continuity edges.

[0150] Perform acceptance screening on each recovery acceptance edge in the recovery acceptance diagram, retain the recovery acceptance edges that do not have thermal stability due to temperature rise changes, polarization release changes, individual cell voltage divergence changes, and charging propulsion changes, and delete the recovery acceptance edges that do not simultaneously meet the above acceptance conditions.

[0151] Starting from each candidate escape origin node, search for a state connection path to the candidate escape target node along the preserved recovery receiving edge. Sort each state connection path according to the number of path edges, thermal recovery residual, polarization release residual, individual branching residual, and charging propulsion residual. The state connection path that ranks first and can connect the state split propagation region or the state collapse propagation region to the stable propagation region is determined as the minimum state connection path.

[0152] When a minimum state connection path exists, the minimum state connection path, the corresponding escape starting point candidate node, recovery receiving edge, escape target candidate node, and the corresponding reserve pool level are combined into a future domain escape bridge. When no minimum state connection path exists, the corresponding charging domain fission frontier is marked as an escape bridge missing state.

[0153] In this embodiment, the latent instability level, future domain collapse level, fission propagation level, recovery acceptance state, and unrecoverable collapse state corresponding to the current charging state of the output lithium battery include:

[0154] The changes in the number of effective safe paths, the changes in the number of failed paths, the collapse intensity marker, the coverage area of ​​the state collapse propagation zone, and the continuation position of the state split propagation zone are read from the future charging domain collapse state. A future domain collapse level is generated according to the following: non-collapsed state, locally contracted state, continuously collapsed state, and unrecoverable collapse candidate state. Specifically, the generation of the future domain collapse level is as follows:

[0155] The system reads changes in the number of valid safe paths and the number of failed paths. When the number of valid safe paths remains stable and the number of failed paths does not increase continuously, the future domain collapse level is marked as non-collapse level. When the number of valid safe paths decreases in stages and the number of failed paths increases in stages, and the collapse intensity marker corresponds to weak collapse, the future domain collapse level is marked as local contraction level. When the number of valid safe paths decreases continuously and the number of failed paths increases continuously, and the coverage area of ​​the state collapse propagation zone continues to expand and forms a continuous connection with the state split propagation zone, the future domain collapse level is marked as continuous collapse level. When the number of valid safe paths decreases below the preset minimum retention number, the number of failed paths continues to occupy the majority of the path count, and the collapse intensity marker corresponds to unrecoverable collapse, the future domain collapse level is marked as an unrecoverable collapse candidate level.

[0156] The fission propagation boundary, fission expansion velocity, fission propagation direction, fission coverage, fission propagation level, and continuous sampling number in the fission frontier of the charging domain are read. Based on whether the fission propagation boundary continuously connects the stable propagation region, the state splitting propagation region, and the state collapse propagation region, a fission propagation level is generated. Specifically, the generation of the fission propagation level is as follows:

[0157] The system reads the positional changes of the fission propagation boundary within adjacent sampling sections, the fission expansion rate, propagation direction, coverage area, coverage reserve pool level, and the number of continuous samplings. When the fission propagation boundary does not continuously connect the stable propagation zone, the state splitting propagation zone, and the state collapse propagation zone, the fission propagation level is marked as an incomplete level. When the fission propagation boundary only connects the stable propagation zone and the state splitting propagation zone, and the fission expansion rate is within a preset low-speed range, the fission propagation level is marked as an initial fission level. When the fission propagation boundary continuously connects the stable propagation zone, the state splitting propagation zone, and the state collapse propagation zone, and the fission coverage area continues to expand, the fission propagation level is marked as a complete fission level. When the fission propagation boundary continuously penetrates all three types of propagation zones, the fission expansion rate continues to increase, the coverage reserve pool level expands to the long dynamic reserve pool, and the number of continuous samplings reaches a preset number, the fission propagation level is marked as a high-risk fission level.

[0158] Read the existence status of the future domain escape bridge, the candidate nodes of the escape start point, the recovery and acceptance edge, the candidate nodes of the escape target and the level of the reserve pool to which they belong. When the future domain escape bridge exists, a recoverable acceptance state is generated. When the future domain escape bridge does not exist and the corresponding charging domain fission front is in a continuous expansion state, a recovery and acceptance failure state is generated.

[0159] Based on the future domain collapse level, fission propagation level, and recovery acceptance status, a joint determination is performed on the current charging state of the lithium battery, and the implicit instability level, future domain collapse level, fission propagation level, recovery acceptance status, and unrecoverable collapse status are output. The corresponding sampling time, abnormal impact cell number, associated fission propagation boundary, and future domain escape bridge identifier are written into the edge monitoring result record.

[0160] Example 1: During continuous charging monitoring in a lithium battery charging cabinet, the system connects to a group of 96-cell lithium iron phosphate battery packs with a rated capacity of 160Ah and an initial SOC of 23.6%. The charging method is constant current to constant voltage. Edge computing nodes are deployed within the charging controller. The collected charging monitoring data includes 96 channels of individual cell voltage, 1 channel of total charging current, 96 channels of individual cell temperature, 1 channel of ambient temperature, 96 channels of equalization branch status, and charging control status. The sampling interval for individual cell voltage is 1 second, the temperature sampling interval is 2 seconds, and the charging current sampling interval is 1 second. At the start of charging, the individual cell voltage is distributed between 3.284V and 3.307V, the highest individual cell temperature is 27.6℃, the lowest individual cell temperature is 26.4℃, and the total charging current is stable at around 80A. Traditional threshold methods determine this as normal charging.

[0161] After the data enters the edge processing flow, the system aligns data from different sampling frequencies to a 1-second time axis and linearly fills in missing temperature points using two adjacent temperature samples. In the first 300 seconds of data, the system received 28,800 voltage samples, 300 current samples, 14,400 original temperature samples, and 14,400 supplemented temperature samples, removing 37 abnormal sampling points caused by communication jitter. After normalization, the individual voltage is mapped to the 0.18–0.44 range, the total charging current to the 0.79–0.81 range, and the temperature to the 0.22–0.31 range, and these values ​​are written sequentially to a 900-second edge ring buffer.

[0162] During the model building phase, the system used charging segments from similar battery packs to train the improved DeepESN network. A total of 28,600 training samples were generated, each 180 seconds long. These included 12,800 stable charging samples, 5,200 samples with insufficient polarization release, 4,300 samples with accumulated thermal inertia, 3,900 samples with single-cell bifurcation expansion, and 2,400 samples before irreversible collapse. In one stable sample, the highest temperature increased from 27.1℃ to 27.9℃ within 180 seconds, and the maximum pressure difference changed from 19mV to 22mV, with a future safe path retention rate of 91.6%. In one pre-collapse sample, the highest temperature increased from 28.5℃ to 32.4℃, and the maximum pressure difference expanded from 26mV to 48mV. In the subsequent 600 seconds, continuous temperature rise and pressure difference expansion occurred, and the future safe path retention rate decreased to 38.7%. After training, the first reservoir layer is configured with 384 state units to record current transients and rapid voltage disturbances; the second reservoir layer is configured with 256 state units to record polarization release and dynamic impedance recovery; and the third reservoir layer is configured with 192 state units to record thermal inertia propagation and individual node branching expansion. The average inference time per edge node is 31ms, and the memory usage is 46MB.

[0163] At 600 seconds into the charging process, the highest single-cell voltage was 3.391V, the lowest single-cell voltage was 3.363V, the highest temperature was 29.1℃, and the maximum voltage difference was 28mV. After inputting the improved DeepESN network, the average activation value of the current transient state units in the first reservoir layer was 0.34, the average activation value of the polarization release state units in the second reservoir layer was 0.29, and the average activation value of the thermal inertia propagation state units in the third reservoir layer was 0.26. The system spliced ​​together 184 future state propagation candidate paths, of which 171 were valid safe paths and 13 were failed paths. The future reachable charging domain coverage was 92.9%, indicating a non-collapsed state.

[0164] At 1260 seconds into the charging process, the total charging current was 80.2A, the highest voltage was 3.462V, the highest temperature was 33.7℃, and the maximum differential voltage was 46mV, still below the traditional over-temperature, over-voltage, and differential voltage alarm thresholds. At this point, the average activation value of the polarization release state units in the second reservoir layer increased to 0.57, and the average activation value of the thermal inertia propagation state units in the third reservoir layer increased to 0.63. There were 196 future state propagation candidate paths, with the number of effective safe paths decreasing to 139 and the number of failed paths increasing to 57. The system counted 171, 158, 147, and 139 effective safe paths within four consecutive 180-second windows, respectively. The future reachable charging domain coverage decreased from 92.9% to 70.9%, marking this stage as a local contraction state.

[0165] At 1680 seconds of charging, the highest temperature was 36.3℃ and the maximum voltage difference was 58mV, and the traditional threshold method still output normal values. The system established an energy-information dual-weighted spatiotemporal graph within the future reachable charging domain, forming 612 state nodes and 1846 spatiotemporal graph edges. Among these, there were 328 nodes in the stable propagation region, 176 nodes in the state splitting propagation region, and 108 nodes in the state collapse propagation region. After performing minimum cut-segment iterative splitting, the system obtained 7 split subgraphs. In the third subgraph, the entropy flux density increased from 0.21 to 0.48, and the energy misalignment gradient increased from 0.18 to 0.46, both exceeding the set threshold of 0.40. This subgraph contained 11 backbone state paths connecting the stable propagation region and the collapse propagation region, which were designated as fission propagation boundaries. At this point, the fission front covered the second and third reserve pool layers, and the fission expansion rate increased by 9 collapse nodes every 180 seconds. The system output a latent instability level of two.

[0166] At 2040 seconds into the charging process, the highest temperature reached 39.4℃, and the maximum voltage difference was 64mV, still not triggering the traditional 45℃ temperature threshold and 80mV voltage difference threshold. The system selected 18 candidate escape start points and 24 candidate escape target nodes from the future reachable charging domain after fission, constructing a recovery connection graph with 286 recovery connection edges. After connection screening, 73 recovery connection edges satisfying thermal recovery, polarization release, single-cell consistency, and charging continuity were retained. The system found 3 candidate escape paths, with the shortest path containing 6 recovery connection edges. The corresponding control action was to reduce the charging current from 80A to 54A and maintain this for 240 seconds, while pausing the 37th cell balancing branch for 90 seconds before reconnecting. After the control was executed, the temperature rise rate of cell No. 52 decreased from 0.017℃ / s to 0.005℃ / s, the voltage difference between cell No. 37 and the average voltage decreased from 64mV to 41mV, the number of effective safe paths increased from 96 to 148, the number of failed paths decreased from 100 to 48, and the system marked the recovery status as recoverable.

[0167] In the comparative experiment, the system was validated using the same 500 sets of charging test segments, including 260 sets of stable charging segments, 160 sets of latent instability segments, and 80 sets of segments before irreversible collapse. The traditional threshold method achieved a latent instability identification accuracy of 79.2%, with an average warning lead time of 138 seconds and 31 missed segments; the ordinary time-series prediction method achieved an accuracy of 85.6%, with an average warning lead time of 426 seconds and 18 missed segments; the method of this invention achieved an accuracy of 94.8%, with an average warning lead time of 792 seconds and 4 missed segments. For the 80 sets of segments before irreversible collapse, the traditional method could only trigger an alarm after explicit temperature rise and could not determine recovery capability; the ordinary time-series prediction method correctly identified 47 sets; the method of this invention correctly identified 74 sets, achieving a recovery capability determination accuracy of 92.5%.

[0168] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the state of charge of a lithium battery based on edge computing, characterized in that, include: Collect charging monitoring data during the lithium battery charging process, perform preprocessing on the charging monitoring data, and generate edge charging monitoring sequences; The edge charging monitoring sequence is input into an improved DeepESN network with spectral radius self-adjustment and cross-layer orthogonal coupling mechanism. Under the event-triggered sparse restart control, a fast dynamic state matrix, a medium dynamic state matrix, and a long dynamic state matrix are generated sequentially to obtain the future reachable charging domain. Read the state propagation feature set in the future reachable charging domain, count the number of state paths that can maintain thermal stability, polarization release, single-unit consistency and charging propulsion in the future, and generate the future charging domain collapse state based on the change in the number of state paths. In the future reachable charging domain, an energy-information dual-weight spatiotemporal graph is established, and the minimum cut-minimum cross-region ratio iterative splitting is performed. The entropy flow density and energy misalignment gradient of each split subgraph are calculated. When both exceed the preset threshold simultaneously, the corresponding backbone state path is marked as the fission propagation boundary, and the fission frontier of the charging domain is generated. After the fission front of the charging domain is formed, a recovery and acceptance search is performed on each state connection path in the future reachable charging domain. The minimum state connection path that can simultaneously restore thermal stability, polarization release, individual consistency and charging propulsion continuity is selected and the minimum state connection path is determined as the future domain escape bridge. Based on the future charging domain collapse state, the charging domain fission frontier, and the existence state of the future domain escape bridge, the system outputs the hidden instability level, future domain collapse level, fission propagation level, recovery and acceptance state, and unrecoverable collapse state corresponding to the current charging state of the lithium battery.

2. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The charging monitoring data includes individual cell voltage sampling values, total charging current sampling values, individual cell temperature sampling values, ambient temperature sampling values, equalization branch conduction status, charging stage control status, and corresponding timestamp information.

3. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The preprocessing of the charging monitoring data to generate the edge charging monitoring sequence includes: The charging monitoring data are processed to align time according to a unified sampling period; For missing sampling points, perform neighborhood interpolation to complete the sampling points; for abnormal and abrupt sampling points, perform sliding window elimination. Synchronous association is performed on each charging monitoring data based on the unit number, sampling time, and charging stage identifier; Normalization is performed on the sampled values ​​of individual cell voltage, total charging current, individual cell temperature, and ambient temperature. The normalized charging monitoring data are written into the edge ring buffer in chronological order to generate the edge charging monitoring sequence.

4. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The obtained future reachable charging domain includes: Based on the current change rate, voltage change rate, temperature rise change rate, equalization branch switching state, and charging stage switching state in the edge charging monitoring sequence, an improved three-layer reservoir structure of the DeepESN network is constructed. The first reservoir layer is configured as a fast dynamic reservoir, the second reservoir layer is configured as a medium dynamic reservoir, and the third reservoir layer is configured as a long dynamic reservoir. Input connection weights, intra-layer recursive connection weights, and inter-layer transfer connection weights are established respectively. The recursive connection weights within the three layers of the reservoir are read separately, and the maximum eigenvalue corresponding to the recursive connection weights of each layer is calculated. The maximum eigenvalue is compared with the preset target spectral radius of each layer. When the maximum eigenvalue deviates from the corresponding target spectral radius, the recursive connection weights of the layers are scaled proportionally according to the corresponding target spectral radius, so that the fast dynamic reservoir maintains the short memory response, the medium dynamic reservoir maintains the polarization recovery response, and the long dynamic reservoir maintains the thermal inertia and single-cell branching propagation response. Orthogonalization is performed on the interlayer transfer connection weights between adjacent reserve pool layers. The state transfer components that are repeated in the next reserve pool layer are removed, and the state transfer components that do not overlap with the previous reserve pool layer are retained, forming an orthogonal coupling channel from the first reserve pool layer to the second reserve pool layer and an orthogonal coupling channel from the second reserve pool layer to the third reserve pool layer. The edge charging monitoring sequence is input into the improved DeepESN network according to the sampling time. The fast dynamic reservoir records the transient changes in current, rapid voltage disturbances and transient actions of the equalization branch to generate a fast dynamic state matrix. The medium dynamic reservoir records the changes in polarization release, dynamic impedance recovery and local heat accumulation to generate a medium dynamic state matrix. The long dynamic reservoir records the continuous changes in thermal inertia, changes in individual branch expansion and changes in safe path maintenance to generate a long dynamic state matrix. When a charging phase transition, a step change in charging current, or a switching of the equalization branch is detected, a sparse restart is performed on the state unit associated with the event in the corresponding storage pool layer. The historical memory of unassociated state units is retained. A gating time splicing order is generated based on the charging control state, thermal stability state, polarization release state, and individual unit consistency state. The fast dynamic state matrix, medium dynamic state matrix, and long dynamic state matrix are spliced ​​into a set of state propagation candidate paths. State propagation candidate paths that simultaneously satisfy thermal stability acceptance, polarization release acceptance, individual unit consistency acceptance, and charging propagation acceptance are selected. The selected set of state propagation candidate paths is determined as the future reachable charging domain.

5. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The generation of future charging domain collapse states based on changes in the number of state paths includes: Read each state propagation candidate path in the future reachable charging domain, and extract the state propagation feature set according to the sampling time order corresponding to the path. The state propagation feature set includes state propagation direction, state propagation continuity, state stable propagation region, state split propagation region and state collapse propagation region. For each state propagation candidate path, the temperature rise change, polarization release change, single-cell voltage divergence change and charging propulsion change corresponding to the adjacent state nodes in the path are read in sequence. Based on the state succession relationship between adjacent state nodes, the thermal stability succession state, polarization release succession state, single-cell consistency succession state and charging propulsion succession state are marked. Candidate paths for state propagation that simultaneously meet the following conditions are marked as valid safe paths: thermal stability acceptance state, polarization release acceptance state, single-unit consistency acceptance state, and charging propulsion acceptance state. Candidate paths for state propagation that do not simultaneously meet the above four conditions are marked as failed paths. The number of valid safe paths and the number of failed paths are counted in the order of sampling time. The number of effective safe paths at adjacent sampling times is continuously compared. When the number of effective safe paths continues to decrease and the number of failed paths continues to increase, the corresponding time segment is marked as the future charging domain shrinkage segment. The collapse intensity mark is generated based on the decrease in effective safe paths, the increase in failed paths and the number of continuous samplings within the future charging domain shrinkage segment. Based on the collapse intensity marker, the coverage area of ​​the state collapse propagation zone, and the continuation position of the state split propagation zone, the collapse states of the future reachable charging domain are classified to generate future charging domain collapse states. The future charging domain collapse states include non-collapse state, local contraction state, continuous collapse state, and irreversible collapse candidate state.

6. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The generation of the charging domain fission front includes: Read the candidate paths for state propagation in the future reachable charging domain, take the state nodes in each path as spatiotemporal graph nodes, connect the state nodes that have thermal stability acceptance, polarization release acceptance, single-cell consistency acceptance or charging promotion acceptance relationships between adjacent sampling times as spatiotemporal graph edges, and record the sampling time, storage pool level and acceptance type corresponding to each spatiotemporal graph edge. Energy weights are generated based on the temperature rise, polarization release, individual voltage divergence, and charging propulsion changes corresponding to each spatiotemporal graph edge. Information weights are generated based on the state propagation continuity, state splitting times, state collapse times, and path failure markers corresponding to each spatiotemporal graph edge. The energy weights and information weights are written into the corresponding spatiotemporal graph edges to obtain an energy-information dual-weight spatiotemporal graph. Using the stable propagation region and the collapsed propagation region as the splitting endpoints, the minimum cut splitting is performed on the energy-information dual-weight spatiotemporal graph. The number of spatiotemporal graph edges crossing the stable propagation region and the collapsed propagation region, the number of cross-regional state paths, and the number of internal state paths of the split subgraph are counted after each splitting. The minimum cross-regional ratio iterative splitting is performed based on the ratio of the number of cross-regional state paths to the number of internal state paths of the split subgraph. For each split subgraph, the increase in the number of failure paths, the increase in state splitting nodes, and the expansion of state collapse nodes within the subgraph are read in the order of sampling time to generate entropy flow density. The directional deviation between temperature rise, polarization release, individual voltage divergence, and charging propulsion changes within the subgraph is also read to generate energy misalignment gradient. When the entropy flow density and energy misalignment gradient of the same split subgraph both exceed the corresponding preset threshold in the same sampling segment, the backbone state path connecting the stable propagation region, the state splitting propagation region and the state collapse propagation region in the split subgraph is extracted, and the backbone state path is marked as the fission propagation boundary. The charging domain fission front is generated according to the sampling time, propagation direction, coverage of the reserve pool level and connection of the fission propagation boundary.

7. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The step of determining the minimum state connection path as a future domain escape bridge includes: Read the fission propagation boundary, stable propagation region, state splitting propagation region and state collapse propagation region in the fission frontier of the charging domain. Mark the state nodes located at the edge of the state splitting propagation region and the state collapse propagation region as candidate escape starting point nodes. Mark the state nodes located in the stable propagation region and continuously satisfying thermal stability acceptance, polarization release acceptance, single-unit consistency acceptance and charging propagation acceptance as candidate escape target nodes. Within the future reachable charging domain, a recovery continuity graph is constructed using escape start point candidate nodes, escape target candidate nodes, and state nodes. State nodes that can form state continuity between adjacent sampling times are connected as recovery continuity edges. Thermal recovery state, polarization release state, single-unit divergence convergence state, charging propulsion maintenance state, and the level of the corresponding storage pool are recorded in the recovery continuity edges. Perform acceptance screening on each recovery acceptance edge in the recovery acceptance diagram, retain the recovery acceptance edges that do not have thermal stability due to temperature rise changes, polarization release changes, individual cell voltage divergence changes, and charging propulsion changes, and delete the recovery acceptance edges that do not simultaneously meet the above acceptance conditions. Starting from each candidate escape origin node, search for a state connection path to the candidate escape target node along the preserved recovery receiving edge. Sort each state connection path according to the number of path edges, thermal recovery residual, polarization release residual, individual branching residual, and charging propulsion residual. The state connection path that ranks first and can connect the state split propagation region or the state collapse propagation region to the stable propagation region is determined as the minimum state connection path. When a minimum state connection path exists, the minimum state connection path, the corresponding escape starting point candidate node, recovery receiving edge, escape target candidate node, and the corresponding reserve pool level are combined into a future domain escape bridge. When no minimum state connection path exists, the corresponding charging domain fission frontier is marked as an escape bridge missing state.

8. The lithium battery charging state monitoring method based on edge computing according to claim 1, characterized in that, The latent instability level, future domain collapse level, fission propagation level, recovery state, and unrecoverable collapse state corresponding to the current charging state of the output lithium battery include: Read the changes in the number of effective safe paths, the changes in the number of failed paths, the collapse intensity marker, the coverage of the state collapse propagation zone, and the connection position of the state split propagation zone in the future charging domain collapse state, and generate the future domain collapse level according to the non-collapse state, the local contraction state, the continuous collapse state, and the irrecoverable collapse candidate state. Read the fission propagation boundary, fission expansion speed, fission propagation direction, fission coverage, fission propagation level and continuous sampling number in the fission front of the charging domain, and generate the fission propagation level based on whether the fission propagation boundary continuously connects the stable propagation region, the state splitting propagation region and the state collapse propagation region. Read the existence status of the future domain escape bridge, the candidate nodes of the escape start point, the recovery and acceptance edge, the candidate nodes of the escape target and the level of the reserve pool to which they belong. When the future domain escape bridge exists, a recoverable acceptance state is generated. When the future domain escape bridge does not exist and the corresponding charging domain fission front is in a continuous expansion state, a recovery and acceptance failure state is generated. Based on the future domain collapse level, fission propagation level, and recovery acceptance status, a joint determination is performed on the current charging state of the lithium battery, and the implicit instability level, future domain collapse level, fission propagation level, recovery acceptance status, and unrecoverable collapse status are output. The corresponding sampling time, abnormal impact cell number, associated fission propagation boundary, and future domain escape bridge identifier are written into the edge monitoring result record.

Images