A method and system for assessing the health status of lithium batteries based on intelligent sensors

Abstract

This invention discloses a method and system for assessing the health status of lithium batteries based on intelligent sensors, belonging to the field of intelligent sensor technology. The method includes: collecting multi-source sensor data and generating a standardized dataset; performing time-series orchestration to generate time-series samples and operating condition identifiers; extracting sensor modes using a successive variational mode decomposition algorithm to generate a health-sensitive mode sequence; constructing an improved TSLANT model to generate multi-stage degradation features; performing health status mapping to generate health assessment results; generating health levels, lifespan intervals, and risk identifiers; and collecting feedback data to update the dataset and model parameters. This invention, by introducing a successive variational mode decomposition algorithm and an improved TSLANT model, achieves the extraction of health-sensitive degradation features from multi-source sensor data of lithium batteries, isolation of sensor drift interference, and accurate assessment of health status.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a method and system for assessing the health status of lithium batteries based on intelligent sensors. Background Technology

[0002] With the rapid development of new energy vehicles, energy storage power stations, portable electronic devices, and industrial backup power supplies, the application scale of lithium batteries in power supply and energy storage scenarios continues to expand. Lithium batteries exhibit degradation phenomena such as capacity decay, increased internal resistance, abnormal temperature rise, shortened discharge plateau, and decreased consistency under long-term charge-discharge cycles, complex temperature environments, and different rate conditions. Therefore, health status assessment has gradually become an important technical aspect of battery management systems. Existing lithium battery health status assessment methods typically collect operational data using voltage sensors, current sensors, temperature sensors, internal resistance detection units, pressure sensors, and vibration sensors. This data is then estimated based on capacity calibration, equivalent circuit models, empirical degradation models, signal decomposition algorithms, or machine learning models to assess battery state of health (SOH), remaining lifespan, and abnormal risks. Compared to simple manual inspection or periodic capacity testing, online health status assessment based on intelligent sensors can continuously acquire status information during lithium battery operation and perform real-time analysis of health changes at the cell, module, or battery pack level.

[0003] Current lithium battery health status assessment technologies still have certain shortcomings. Traditional assessment methods mostly treat multi-source sensor data as ordinary time-series data, lacking the process of establishing unified time-series samples according to cell number, sensor number, acquisition time, and operating condition stage. This leads to the easy mixing of degradation characteristics in the charging, discharging, and resting stages. Although some methods introduce variational mode decomposition or deep learning models, they usually focus on decomposing and predicting the original voltage, capacity, or temperature sequences, failing to distinguish sensor zero-point offset, reference drift, and slow measurement deviation during mode extraction. This easily leads to misjudging sensor drift as a battery degradation characteristic. At the same time, existing mode decomposition results are mostly limited to the level of a single cycle or a single sequence, lacking the inheritance relationship of degradation genealogy across charge-discharge cycles, making it difficult to describe the continuation, splitting, merging, and regeneration processes of degradation modes. Existing time-series network models also typically rely on general time-series feature extraction structures, which are insufficient in expressing specific degradation phenomena such as lithium battery charge-discharge hysteresis, capacity plateau breakage, and static recovery trail. This leads to problems such as insufficient stability, inaccurate remaining lifetime prediction range, and incomplete identification of abnormal risks in health status assessment results under complex operating conditions, sensor drift, and changes in degradation stages.

[0004] Therefore, how to provide a method and system for assessing the health status of lithium batteries based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for assessing the health status of lithium batteries based on intelligent sensors. This invention utilizes multi-source sensor data acquisition from lithium batteries, successive variational mode decomposition algorithms, an improved TSLANT model, health status mapping, and feedback update technology. It details the process of standardizing multi-source sensor data during lithium battery operation, time-series arrangement, sensor drift residual isolation, degradation spectrum mode inheritance, multi-stage degradation feature extraction, and health status assessment. This enables the comprehensive generation of lithium battery SOH estimates, health levels, remaining lifetime prediction ranges, abnormal risk indicators, and sensor anomaly records. Compared to traditional health status assessment methods based on single sensor data or ordinary time-series prediction models, this invention offers advantages such as strong sensor drift interference identification capability, good continuity in cross-cycle degradation mode tracking, accurate expression of charge / discharge hysteresis degradation features, stable identification of capacity plateau fracture states, and full utilization of health information during the resting recovery phase.

[0006] A method and system for assessing the health status of a lithium battery based on a smart sensor, according to an embodiment of the present invention, includes: Collect multi-source sensor data during the operation of lithium batteries, perform preprocessing on the multi-source sensor data, and generate a standardized battery sensor dataset; The standardized battery sensor dataset is time-series arranged according to cell number, sensor number, acquisition time and operating condition stage to generate multi-source sensor time-series samples and operating condition stage identifiers. Multi-source sensor time-series samples and operating condition stage identifiers are input into the successive variational mode decomposition algorithm. The sensor modes are extracted round by round and the residual signals are updated. Sensor drift residual isolation and degradation spectrum mode inheritance are performed to generate sensor drift residual records, degradation spectrum mode sequences and health-sensitive mode sequences. An improved TSLANT model is constructed, which includes a charge-discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer. Based on the operating condition stage identifier, the health-sensitive mode sequence is processed to generate hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation, respectively, and combined to generate a multi-stage degradation feature representation. Health status mapping is performed based on multi-stage degradation feature representation, sensor drift residual records, and degradation spectrum mode sequences to generate lithium battery health status assessment results; Based on the health status assessment results of lithium batteries, the health level is divided, the remaining life prediction range, abnormal risk indicators and sensor abnormality records are determined, and the lithium battery health status output results are generated. Collect lithium battery operation feedback data and maintenance verification data, and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

[0007] Optionally, generating the standardized battery sensing dataset includes: Collect multi-source sensor data during the operation of lithium batteries. The multi-source sensor data includes voltage data, current data, temperature data, internal resistance data, pressure data, vibration data, charge / discharge status data, cycle count data, and ambient temperature data. The multi-source sensor data are correlated according to the cell number, sensor number, acquisition time and data type to generate the original sensor record table; The original sensor record table is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and filling missing values ​​to generate a cleaned sensor record table. The cleaned sensor record table is timestamped according to a uniform sampling interval, and the values ​​corresponding to different data types are subjected to unit unification and numerical standardization to generate a standardized sensor record table. The standardized sensor record sheets are archived according to the cell number, sensor number, and acquisition time to generate a standardized battery sensor dataset.

[0008] Optionally, generating multi-source sensing time-series samples and operating condition stage identifiers includes: Read the cell number, sensor number, acquisition time, standardized sensor value, charge / discharge status field and current direction information from the standardized battery sensing dataset; Establish a sensor channel index table according to cell number, sensor number and data type, and arrange the standardized sensor values ​​corresponding to each sensor channel in the order of acquisition time. The charging stage, discharging stage and resting stage are divided according to the charging and discharging status field and current direction information, and the start and end times of the stage are written into the operating condition stage identifier. Extract sensing sequence segments according to cell number, sensor number and operating condition stage identifier to generate single-channel sensing timing segments; Single-channel sensing time series segments with the same cell number, the same acquisition time range, and different sensor numbers are combined and associated with the operating condition stage identifier to generate multi-source sensing time series samples.

[0009] Optionally, the generation of the sensing drift residual record, the degradation lineage mode sequence, and the health-sensitive mode sequence includes: Read the multi-source sensor timing samples and operating condition stage identifiers, establish the sensor sequence to be decomposed according to the cell number, sensor number and operating condition stage, and use the sensor sequence to be decomposed as the initial residual signal; Successive variational mode decomposition is performed on the initial residual signal, the current sensing mode is extracted round by round, the center frequency range, duration range, amplitude change direction and corresponding operating condition stage of the current sensing mode are recorded, and the current sensing mode is subtracted from the current residual signal; After subtracting the current sensing mode in each round, sensor drift residual isolation is performed on the updated residual signal. The zero offset residual, reference drift residual and slow measurement deviation residual are written into the sensor drift residual record, and the sensor drift residual components are removed from the updated residual signal to generate a drift isolation residual signal. The drift isolation residual signal is used as the current residual signal for the new round. The sensor mode extraction, residual signal update and sensor drift residual isolation are repeated until no new sensor mode is extracted from the current residual signal. The sensor modes extracted in each round are summarized to generate the sensor mode set corresponding to each charge and discharge cycle. Based on the set of sensing modes corresponding to each charge-discharge cycle, a mode registration table is established according to the order of charge-discharge cycles. The sensing modes in the current charge-discharge cycle are matched with the sensing modes in the adjacent charge-discharge cycles. Based on the migration of the center frequency interval, the change of the duration interval, the direction of amplitude change and the consistency of the operating condition stage, the continuation relationship, split relationship, merging relationship and new generation relationship of the sensing modes are marked. Based on the continuation, splitting, merging, and new generation relationships of the sensing modes, the degradation lineage mode inheritance is performed. The corresponding sensing modes are written into the same degradation lineage, degradation lineage branch, degradation lineage merging node, and new degradation lineage number, respectively, and arranged in the order of charge and discharge cycles to generate a degradation lineage mode sequence. The sensor modes associated with sensor drift residual records are removed from the degraded spectrum mode sequence, and the remaining sensor modes are arranged according to cell number, sensor number, operating condition stage and acquisition time to generate a health sensitive mode sequence.

[0010] Optionally, the generation of multi-stage degradation feature representations includes: An improved TSLANet model was constructed, which includes a charge / discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer. The charge / discharge hysteresis mirror spectrum layer reads the health-sensitive mode sequence and operating condition stage identifier, divides the health-sensitive mode sequence into charging stage mode, discharging stage mode and resting stage mode, performs spectral trajectory encoding on the charging stage mode and mirror spectral trajectory encoding on the discharging stage mode, and generates hysteresis deviation representation based on the offset relationship between the charging stage spectral trajectory and the discharging stage mirror spectral trajectory. The capacity platform fracture latch layer reads the discharge phase mode and hysteresis deviation representation, identifies the voltage platform holding interval, platform interruption position and platform end drop position in the discharge phase mode, writes the voltage platform holding interval, platform interruption position and platform end drop position into the platform state latch unit, and generates a capacity platform fracture state representation. The stationary recovery wake echo layer reads the stationary phase mode, capacity plateau fracture state representation and operating condition stage identifier, extracts the voltage rebound sequence, temperature drop sequence, pressure release sequence and internal resistance recovery sequence from the stationary phase mode, performs wake echo encoding on each recovery sequence, and generates the stationary recovery wake echo representation. Hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation are combined and multi-stage degradation feature representation is generated through temporal lightweight adaptive coding. The improved TSLANT model was trained using a combination of health status assessment error, hysteresis deviation representation error, capacity plateau fracture state error, static recovery wake echo error, and multi-stage degradation feature representation consistency error as joint optimization objectives. The network parameters of the charge-discharge hysteresis mirror spectrum layer, capacity plateau fracture latch layer, and static recovery wake echo layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved TSLANT model was considered to have completed convergence training.

[0011] Optionally, the generation of lithium battery health status assessment results includes: Read the multi-stage degradation feature representation, sensor drift residual record and degradation spectrum mode sequence, and establish feature correspondence according to cell number and acquisition time; Hysteresis deviation features, capacity plateau fracture features, static recovery trail features, and temporal degradation embeddings are extracted from the multi-stage degradation feature representation. Lineage continuation state, lineage branching state, lineage merging state, and new lineage state are extracted from the degradation lineage mode sequence to generate a health mapping feature set. The drift sensing channel is determined based on the sensor drift residual record, and the features corresponding to the drift sensing channel are marked as sensing reliability judgment features to generate a corrected health mapping feature set. Based on the modified health mapping feature set, perform health state mapping to generate SOH estimate, capacity decay state, internal resistance degradation state, thermal recovery state, and sensing reliability state; The estimated SOH, capacity decay status, internal resistance degradation status, thermal recovery status, and sensor reliability status are archived according to the cell number and collection time to generate lithium battery health status assessment results.

[0012] Optionally, the generation of the lithium battery health status output result includes: The health level is determined based on the estimated SOH value in the lithium battery health status assessment results. The degradation rate is obtained by dividing the difference in SOH estimates between adjacent acquisition times by the cycle number increment, and the remaining lifetime prediction interval is determined based on the current SOH estimate and degradation rate. Anomaly risk identifiers and sensing anomaly records are generated based on capacity decay status, internal resistance degradation status, thermal recovery status, and sensing reliability status. The health level, remaining life prediction range, abnormal risk indicators, and sensor anomaly records are combined according to the cell number and collection time to form the output result of the lithium battery health status.

[0013] Optionally, the step of collecting lithium battery operation feedback data and maintenance verification data and performing feedback updates includes: Collect lithium battery operation feedback data and maintenance verification data, and generate health assessment feedback records according to cell number, collection time, and maintenance time; Health assessment feedback records are written into a standardized battery sensor dataset, and maintenance verification data is correlated with lithium battery health status assessment results; Correct the degenerate lineage number, degenerate lineage branch, degenerate lineage merging node, and newly formed degenerate lineage number in the degenerate lineage mode sequence based on the maintenance and verification data; Update the sensor drift residual isolation parameters, degradation spectrum mode inheritance parameters, and improved TSLnet model parameters based on health assessment feedback records.

[0014] refer to Figure 3 A lithium battery health status assessment system based on intelligent sensors includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source sensor data of lithium batteries and perform preprocessing to generate a standardized battery sensor dataset. The time-series sample construction module is used to perform time-series orchestration on the standardized battery sensing dataset to generate multi-source sensing time-series samples and operating condition stage identifiers. The mode decomposition processing module is used to perform sensor drift residual isolation and degradation lineage mode inheritance, generating sensor drift residual records, degradation lineage mode sequences, and health-sensitive mode sequences; The degradation feature extraction module is used to construct an improved TSLANet model and generate multi-stage degradation feature representations; The health status assessment module is used to perform health status mapping and generate lithium battery health status assessment results. The health outcome output module is used to generate health level, remaining life expectancy prediction range, abnormal risk indicators, and sensor anomaly records. The feedback update module is used to collect lithium battery feedback data and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

[0015] The beneficial effects of this invention are: This invention proposes a method and system for assessing the state of health (SOH) of lithium batteries based on intelligent sensors. Through multi-source sensor data acquisition, standardized processing, time-series orchestration, successive variational mode decomposition, and an improved TSLANT model feature extraction process, it provides unified processing of data such as voltage, current, temperature, internal resistance, pressure, vibration, and charge / discharge status during lithium battery operation. Compared to methods relying solely on single sensor data or a single operating curve for SOH estimation, this invention organizes sensor data according to cell number, sensor number, acquisition time, and operating condition stage. This allows for the separate extraction and continuous representation of health degradation information at different operating conditions, improving the data completeness and operating condition adaptability of the health status assessment.

[0016] This invention employs a successive variational mode decomposition algorithm to extract modes from multi-source sensor time-series samples round by round, and performs sensor drift residual isolation and degradation spectrum mode inheritance. This enables the separation of sensor zero-point offset, reference drift, and slow measurement bias from health-related modes, reducing the interference of sensor drift on the evaluation results. Simultaneously, by recording the continuation, splitting, merging, and emergence relationships of sensor modes in different charge-discharge cycles through degradation spectrum mode sequences, the continuity and stability of cross-cycle degradation trend tracking are improved.

[0017] This invention employs an improved TSLanet model, comprising a charge / discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a resting recovery wake echo layer, to extract multi-stage degradation features from health-sensitive modal sequences. This model can separately express charge / discharge hysteresis deviation, capacity plateau fracture state, and resting recovery wake changes. Compared to general time-series prediction models, this invention can more fully utilize degradation information during the charging, discharging, and resting stages of lithium batteries, generating SOH estimates, health levels, remaining lifetime prediction ranges, abnormal risk indicators, and sensor anomaly records. Furthermore, feedback updates improve the accuracy, stability, and continuous adaptability of the health status assessment results. Attached Figure Description

[0018] 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: Figure 1 This is a flowchart of a lithium battery health status assessment method based on intelligent sensors proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved TSLANT model of the lithium battery health status assessment method based on intelligent sensors proposed in this invention. Figure 3 This is a schematic diagram of the structure of a lithium battery health status assessment system based on intelligent sensors proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A method and system for assessing the health status of lithium batteries based on intelligent sensors, comprising: Collect multi-source sensor data during the operation of lithium batteries, perform preprocessing on the multi-source sensor data, and generate a standardized battery sensor dataset; The standardized battery sensor dataset is time-series arranged according to cell number, sensor number, acquisition time and operating condition stage to generate multi-source sensor time-series samples and operating condition stage identifiers. Multi-source sensor time-series samples and operating condition stage identifiers are input into the successive variational mode decomposition algorithm. The sensor modes are extracted round by round and the residual signals are updated. Sensor drift residual isolation and degradation spectrum mode inheritance are performed to generate sensor drift residual records, degradation spectrum mode sequences and health-sensitive mode sequences. An improved TSLANT model is constructed, which includes a charge-discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer. Based on the operating condition stage identifier, the health-sensitive mode sequence is processed to generate hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation, respectively, and combined to generate a multi-stage degradation feature representation. Health status mapping is performed based on multi-stage degradation feature representation, sensor drift residual records, and degradation spectrum mode sequences to generate lithium battery health status assessment results; Based on the health status assessment results of lithium batteries, the health level is divided, the remaining life prediction range, abnormal risk indicators and sensor abnormality records are determined, and the lithium battery health status output results are generated. Collect lithium battery operation feedback data and maintenance verification data, and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

[0021] In this embodiment, generating a standardized battery sensing dataset includes: Collect multi-source sensor data during the operation of lithium batteries. The multi-source sensor data includes voltage data, current data, temperature data, internal resistance data, pressure data, vibration data, charge / discharge status data, cycle count data, and ambient temperature data. The multi-source sensor data are correlated according to the cell number, sensor number, acquisition time and data type to generate the original sensor record table; The original sensor record table is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and filling missing values ​​to generate a cleaned sensor record table, in which: The handling of anomalous jump values ​​and missing value imputation are as follows: According to the cell number, sensor number, and data type, the sensor data in the original sensor record table are arranged in chronological order of acquisition time. Calculate the variation range between sensor values ​​corresponding to adjacent acquisition times and compare it with the allowable variation range of a single sample for the corresponding data type; When the change exceeds the allowable change range of a single sample for the corresponding data type, the corresponding sensor value will be marked as an abnormal jump value. Read the sensor values ​​corresponding to the previous and next acquisition times of the abnormal jump value, calculate the average of the two sensor values, generate an abnormal replacement value, and replace the abnormal jump value. Missing sensor data is identified according to the acquisition time interval. When there are sensor values ​​of the same data type in both the previous and next acquisition times of the missing sensor data, the average of the two sensor values ​​is calculated to generate the missing filling value. When the same data type corresponding to the same sensor number is missing consecutively, read the same type of sensor data corresponding to adjacent sensor numbers under the same cell number and generate a continuous missing fill value. The original sensor record table is updated based on the abnormal jump value processing results, missing filling values, and continuous missing filling values ​​to generate a cleaned sensor record table. The cleaned sensor record table is timestamped according to a uniform sampling interval, and the values ​​corresponding to different data types are subjected to unit unification and numerical standardization to generate a standardized sensor record table. The standardized sensor record sheets are archived according to the cell number, sensor number, and acquisition time to generate a standardized battery sensor dataset.

[0022] In this embodiment, generating multi-source sensing time-series samples and operating condition stage identifiers includes: Read the cell number, sensor number, acquisition time, standardized sensor value, charge / discharge status field and current direction information from the standardized battery sensing dataset; Establish a sensor channel index table according to cell number, sensor number and data type, and arrange the standardized sensor values ​​corresponding to each sensor channel in the order of acquisition time. The system is divided into charging, discharging, and resting stages based on the charging / discharging status field and current direction information, and the start and end times of each stage are written into the operating condition stage identifier. The process is divided into three stages: charging, discharging, and resting. Specifically: Read the charging / discharging status field and current direction information, and establish a working condition judgment sequence according to the acquisition time order; When the charge / discharge status field is in the charging state and the current direction information corresponds to the charging direction, the corresponding acquisition time will be marked as the charging stage. When the charge / discharge status field is in discharge status and the current direction information corresponds to the discharge direction, the corresponding acquisition time will be marked as the discharge stage. When the charge / discharge status field is in a static state and the current direction information corresponds to a zero current state, the corresponding acquisition time will be marked as a static stage. The start and end times of consecutive identical phase markers are merged to generate the start and end times of the charging, discharging, and resting phases. Write the phase type, phase start time, and phase end time into the operating condition phase identifier. Sensing sequence segments are extracted based on cell number, sensor number, and operating condition stage identifier to generate a single-channel sensing timing segment, where: Generate a single-channel sensing timing segment, specifically as follows: Read the sensor channel index table and operating condition stage identifier to determine the collection time range corresponding to the same cell number, the same sensor number and the same operating condition stage; Extract standardized sensor values ​​corresponding to the sensor number within the acquisition time range from the standardized battery sensor dataset; The extracted standardized sensor values ​​are arranged in chronological order of acquisition time to generate a single-channel candidate sensor sequence. The single-channel candidate sensing sequence is bound to the corresponding cell number, sensor number, data type, stage type, stage start acquisition time, and stage end acquisition time to generate a single-channel sensing timing segment. Single-channel sensing time series segments with the same cell number, the same acquisition time range, and different sensor numbers are combined and associated with the operating condition stage identifier to generate multi-source sensing time series samples.

[0023] In this embodiment, the generation of sensor drift residual records, degradation lineage mode sequences, and health-sensitive mode sequences includes: Read the multi-source sensor timing samples and operating condition stage identifiers, establish the sensor sequence to be decomposed according to the cell number, sensor number and operating condition stage, and use the sensor sequence to be decomposed as the initial residual signal; Successive variational mode decomposition is performed on the initial residual signal, extracting the current sensing mode round by round. The center frequency range, duration range, amplitude change direction, and corresponding operating condition stage of the current sensing mode are recorded. The current sensing mode is then subtracted from the current residual signal, where: Perform successive variational mode decomposition on the initial residual signal, specifically as follows: Record the initial residual signal as the current residual signal, and initialize the current sensing mode, center frequency, and modal bandwidth parameters; The current residual signal is frequency domain transformed to extract the spectral energy distribution in the current residual signal, and the initial center frequency of the current round is determined based on the location of the concentrated spectral energy. Extract the concentrated components of the spectrum around the initial center frequency to generate the initial modal components of the current sensing mode; The initial modal components, center frequency, and modal bandwidth parameters of the current sensing mode are iteratively updated to limit the spectral distribution of the current sensing mode to the corresponding center frequency range. When the average absolute value of the difference between the current sensing mode components obtained from two adjacent iterations is less than 0.001, it is determined that the current sensing mode extraction is complete. Record the center frequency range based on the spectral distribution range of the current sensing mode, record the duration range based on the start and end positions of the current sensing mode on the acquisition time axis, record the amplitude change direction based on the amplitude change of the current sensing mode within the duration range, and read the operating condition stage at the corresponding acquisition time position. Subtract the current sensing mode from the current residual signal to generate an updated residual signal, and use the updated residual signal as the input for the next round of sensing mode extraction; After subtracting the current sensing mode in each round, sensor drift residual isolation is performed on the updated residual signal. The zero-point offset residual, reference drift residual, and slow measurement deviation residual are written into the sensor drift residual record, and the sensor drift residual component is removed from the updated residual signal to generate a drift isolation residual signal, where: Perform sensor drift residual isolation on the updated residual signal, specifically as follows: Read the updated residual signal and extract the offset direction, offset duration, and offset amplitude of the residual signal according to the acquisition time sequence; When the residual signal maintains the same offset direction in the charging, discharging and resting stages, the corresponding component of the residual signal is marked as the zero-point offset residual. When the residual signal accumulates and changes along the same offset direction during continuous charge and discharge cycles, the corresponding component of the residual signal is marked as the reference drift residual. When the amplitude of the residual signal change is less than the amplitude of the current sensing mode change, and the duration of the residual signal covers multiple acquisition time periods, the corresponding component of the residual signal is marked as slow measurement deviation residual. Write the zero-point offset residual, reference drift residual, and slow measurement deviation residual into the sensor drift residual record according to the cell number, sensor number, data type, acquisition time, and residual type; Zero-point offset residuals, reference drift residuals, and slow measurement deviation residuals are removed from the updated residual signal to generate a drift isolation residual signal. The drift isolation residual signal is used as the current residual signal for the new round. The sensor mode extraction, residual signal update and sensor drift residual isolation are repeated until no new sensor mode is extracted from the current residual signal. The sensor modes extracted in each round are summarized to generate the sensor mode set corresponding to each charge and discharge cycle. Based on the set of sensing modes corresponding to each charge-discharge cycle, a mode registration table is established according to the order of the charge-discharge cycles. The sensing modes in the current charge-discharge cycle are mapped to the sensing modes in adjacent charge-discharge cycles. Based on the shift in center frequency intervals, changes in duration intervals, direction of amplitude changes, and consistency of operating conditions, the continuity, splitting, merging, and emergence relationships of the sensing modes are marked. The continuation, splitting, merging, and emergence relationships of the labeled sensing modes are as follows: Read the set of sensing modes corresponding to each charge and discharge cycle, and group the sensing modes according to cell number, sensor number and operating condition stage; Match the sensing modes in the current charge / discharge cycle with the same set of sensing modes in the adjacent charge / discharge cycles; When a sensing mode in the current charge-discharge cycle matches only one sensing mode in an adjacent charge-discharge cycle, and the center frequency ranges of the two sensing modes overlap, the amplitude change direction is consistent, and the corresponding operating conditions are consistent, the two sensing modes are marked as having a continuous relationship. When a sensing mode in the current charge-discharge cycle matches two or more sensing modes in an adjacent charge-discharge cycle, and the two or more sensing modes retain the center frequency range and duration range of the current sensing mode respectively, the corresponding sensing mode is marked as a split relationship. When two or more sensing modes in the current charge-discharge cycle match the same sensing mode in an adjacent charge-discharge cycle, and the same sensing mode covers the center frequency range of the two or more sensing modes, the corresponding sensing modes are marked as merged. When the sensing mode in the current charge-discharge cycle does not match the same group of sensing modes in the adjacent charge-discharge cycle, the corresponding sensing mode is marked as a new relationship; Based on the continuation, splitting, merging, and new generation relationships of sensing modes, degenerate lineage mode inheritance is performed. Corresponding sensing modes are written into the same degenerate lineage, degenerate lineage branch, degenerate lineage merging node, and new degenerate lineage number, respectively, and arranged according to the charge / discharge cycle order to generate a degenerate lineage mode sequence, where: Perform degenerate lineage modal inheritance, specifically as follows: When a sensing mode is marked as a continuation relationship, the sensing mode in the current charge-discharge cycle is assigned the same degradation series number as that established in the adjacent charge-discharge cycle and written into the same degradation series. When a sensing mode is marked as a split relationship, the original degenerate lineage number is retained, and a degenerate lineage branch number is established for the split sensing mode. The split sensing mode is then written into the corresponding degenerate lineage branch. When a sensing mode is marked as a merged relationship, a degradation spectrum merge node is established, and the sensing mode number before merging, the sensing mode number after merging, and the corresponding charge-discharge cycle number are written into the degradation spectrum merge node. When a sensing mode is marked as a newborn relation, a newborn-degenerate lineage number is established, and the newborn sensing mode is written into the newborn-degenerate lineage number; The sensor modes associated with the same degradation lineage, degradation lineage branches, degradation lineage merging nodes, and newly generated degradation lineage numbers are arranged according to the charge-discharge cycle sequence to generate a degradation lineage mode sequence. The sensor modes associated with sensor drift residual records are removed from the degraded spectrum mode sequence, and the remaining sensor modes are arranged according to cell number, sensor number, operating condition stage and acquisition time to generate a health sensitive mode sequence.

[0024] In this embodiment, the generation of multi-stage degradation feature representation includes: An improved TSLANT model is constructed, which includes a charge / discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer, wherein: The improved TSLaneet model is constructed as follows: In the traditional TSLANT model, the network includes adaptive spectral blocks, interactive convolutional blocks, and an output mapping structure. Based on the traditional adaptive spectral blocks, the ordinary spectrum processing structure is replaced with a charge-discharge hysteresis mirror spectral layer. Spectral trajectory encoding and mirror spectral trajectory encoding are performed on the charging and discharging phase modes in the health-sensitive mode sequence, respectively. An offset correspondence between the charging phase spectral trajectory and the discharging phase mirror spectral trajectory is established. A capacity plateau break latch layer is added after the traditional interactive convolutional block to latch the voltage plateau holding interval, plateau interruption position, and plateau end drop position in the discharging phase mode. A stationary recovery wake echo layer is added after the capacity plateau break latch layer to encode the voltage rebound sequence, temperature drop sequence, pressure release sequence, and internal resistance recovery sequence in the stationary phase mode. The charge-discharge hysteresis mirror spectral layer, capacity plateau break latch layer, and stationary recovery wake echo layer are connected to the traditional TSLANT model in the order of spectrum encoding, plateau state latching, and stationary wake encoding to obtain the improved TSLANT model. The charge / discharge hysteresis mirror spectrum layer reads the health-sensitive mode sequence and operating condition stage identifier, dividing the health-sensitive mode sequence into charging stage mode, discharging stage mode, and resting stage mode. Spectral trajectory encoding is performed on the charging stage mode, and mirror spectral trajectory encoding is performed on the discharging stage mode. Hysteresis deviation representation is generated based on the offset relationship between the charging stage spectral trajectory and the discharging stage mirror spectral trajectory, where: The charge / discharge hysteresis mirror spectrum includes: Operating condition mode distribution table: Stores the charging phase mode, discharging phase mode, and resting phase mode in the health-sensitive mode sequence according to the operating condition stage identifier; Charging spectrum trajectory buffer: stores the center frequency range, amplitude change direction, and spectral energy distribution corresponding to the charging phase modes; Discharge image spectrum buffer: stores the center frequency range, image direction, and spectral energy distribution of the discharge phase modes after image direction conversion; Mirror spectrum registration lookup table: Records the correspondence between cell number, sensor number, operating condition stage, and acquisition time between charging phase modes and discharging phase modes; Spectrum trajectory encoder: Performs spectrum trajectory encoding on the charging phase mode to generate the charging phase spectrum trajectory; Mirror spectrum encoder: Performs mirror spectrum trajectory encoding on the discharge phase mode to generate the discharge phase mirror spectrum trajectory; Hysteresis Offset Calculator: Calculates the center frequency offset, spectral energy offset, direction offset, and duration offset between the spectral trajectory during the charging phase and the mirror spectral trajectory during the discharging phase; Hysteresis Deviation Output: Generates a hysteresis deviation representation based on center frequency deviation, spectral energy deviation, direction deviation, and duration deviation; In the charge / discharge hysteresis mirror spectrum layer, the operating condition mode splitting table splits the health-sensitive mode sequence into charging stage mode, discharging stage mode, and resting stage mode according to the operating condition stage identifier. The charging stage mode is written into the charging spectrum trajectory buffer, and the discharging stage mode is written into the discharging mirror spectrum buffer after mirror direction conversion. The mirror spectrum registration lookup table establishes the correspondence between the charging stage mode and the discharging stage mode according to the cell number, sensor number, operating condition stage, and acquisition time. The spectrum trajectory encoder reads the data in the charging spectrum trajectory buffer to generate the charging stage spectrum trajectory, and the mirror spectrum encoder reads the data in the discharging mirror spectrum buffer to generate the discharging stage mirror spectrum trajectory. The hysteresis offset calculator receives the charging stage spectrum trajectory and the discharging stage mirror spectrum trajectory and calculates the center frequency offset, spectrum energy offset, direction offset, and duration offset. The hysteresis offset output device generates a hysteresis offset representation based on the output of the hysteresis offset calculator. The health-sensitive mode sequence is divided into charging phase mode, discharging phase mode, and resting phase mode, specifically: Read the cell number, sensor number, acquisition time, duration range, and operating condition stage corresponding to each sensing mode in the health-sensitive mode sequence; Read the phase type, phase start time, and phase end time from the working condition phase identifier; The sensing modes whose duration falls within the start and end time of the charging phase are written into the charging phase mode. The sensing modes whose duration falls within the start and end time of the discharge phase are written into the discharge phase mode. The duration interval of the sensing mode is written into the static phase mode. When the duration interval of the sensing mode spans two operating conditions, calculate the overlap between the duration interval and the start and end times of each operating condition, and write the sensing mode into the stage mode corresponding to the operating condition with the largest overlap. The charging phase mode, discharging phase mode, and resting phase mode are arranged according to the cell number, sensor number, and acquisition time, respectively. Spectral trajectory encoding is performed on the charging phase modes, specifically as follows: Read the center frequency range, amplitude change direction, spectral energy distribution, and duration range of the charging phase mode; Arrange the charging phase modes corresponding to the same cell number and the same sensor number in chronological order of data collection time; The trajectory position of each charging stage mode in the spectrum space is determined based on the center frequency range; The energy response value of each charging stage mode at the corresponding trajectory position is determined based on the spectral energy distribution. Mark the direction of the spectral trajectory corresponding to each charging stage mode according to the direction of amplitude change; The trajectory position, energy response value, spectral trajectory direction, and duration interval are combined to generate the spectral trajectory during the charging phase. Mirror spectrum trajectory encoding is performed on the discharge phase modes, specifically as follows: Read the center frequency range, amplitude change direction, spectral energy distribution, and duration range of the discharge phase mode; Arrange the discharge phase modes corresponding to the same cell number and the same sensor number in reverse order of the acquisition time; The trajectory position of each discharge stage mode in the spectrum space is determined based on the center frequency range; The amplitude change direction of each discharge phase mode is converted into the mirror image direction corresponding to the charging phase mode; The mirror energy response values ​​of each discharge stage mode at the corresponding trajectory position are determined based on the spectral energy distribution. By combining the trajectory position, mirror energy response value, mirror direction, and duration interval, a mirror spectrum trajectory for the discharge phase is generated. The hysteresis deviation representation is generated as follows: Read the spectrum trajectory during the charging phase and the mirror spectrum trajectory during the discharging phase, and establish the trajectory correspondence according to the cell number, sensor number, and charge / discharge cycle number; Based on the trajectory correspondence, the charging phase spectrum trajectory and the discharge phase mirror spectrum trajectory under the same cell number and the same sensor number are registered; Calculate the offset of the two registered spectral trajectories in the center frequency range and generate a center frequency offset record. Calculate the difference in energy response values ​​between the two registered spectral trajectories to generate a spectral energy offset record; Compare the spectral trajectory directions and mirror directions corresponding to the two registered spectral trajectories to generate a direction deviation record; Compare the duration intervals corresponding to the two registered spectral trajectories to generate a duration deviation record; The center frequency offset record, spectral energy offset record, direction offset record, and duration offset record are combined in the order of acquisition time to generate a hysteresis deviation representation; The capacity plateau fracture latch layer reads the discharge phase mode and hysteresis deviation representation, identifies the voltage plateau holding interval, plateau interruption position, and plateau end drop position in the discharge phase mode, and writes the voltage plateau holding interval, plateau interruption position, and plateau end drop position into the plateau state latch unit to generate a capacity plateau fracture state representation, wherein: The capacity platform fracture latch layer includes: Discharge mode cache table: stores the discharge stage modes and their corresponding cell numbers, acquisition times, center frequency ranges, amplitude variation directions, and spectral energy distributions; Hysteresis Deviation Register: Stores the hysteresis deviation representation of the charge / discharge hysteresis mirror spectrum output; Platform candidate interval queue: Stores candidate platform intervals with continuous and stable voltage changes during the discharge phase; Platform hold discriminator: determines the voltage platform hold range; Platform interrupt locator: Determines the location of voltage platform interruption; End-of-phase fall detector: Determines the end-of-phase fall location of the voltage platform; Platform state latch unit: stores the platform holding interval, platform interruption position, and platform end-fall position; Fracture state encoder: Generates a capacity platform fracture state representation based on the platform state latch unit and hysteresis deviation representation; In the capacity platform fracture latch layer, the discharge mode cache table transmits the discharge stage mode and corresponding cell number, acquisition time, center frequency range, amplitude change direction and spectral energy distribution to the platform candidate interval queue. The platform candidate interval queue forms candidate platform intervals according to the continuous and stable voltage change state, and inputs the candidate platform intervals into the platform holding discriminator to generate the voltage platform holding interval. The hysteresis deviation register transmits the hysteresis deviation representation to the platform interruption locator and the fracture state encoder respectively. The platform interruption locator determines the voltage platform interruption position by combining the discharge stage mode and the hysteresis deviation representation. The end drop detector determines the voltage platform end drop position based on the discharge stage mode after the voltage platform holding interval. The platform state latch unit receives and stores the voltage platform holding interval, the voltage platform interruption position and the voltage platform end drop position. The fracture state encoder reads the information in the platform state latch unit and the hysteresis deviation register to generate the capacity platform fracture state representation. Identify the voltage plateau sustaining interval in the discharge phase mode, specifically: Read the voltage mode components, acquisition time, duration interval, and spectral energy distribution in the discharge phase modes; Calculate the voltage change amplitude between voltage mode components corresponding to adjacent acquisition times in the discharge phase mode, and generate a voltage change amplitude sequence according to the acquisition time order; Calculate the median and mean absolute deviation of the voltage change amplitude sequence, and determine the stable change range of the platform by subtracting the mean absolute deviation from the median to the sum of the median and the mean absolute deviation. The time length corresponding to 30 consecutive sampling points is determined as the lower limit of the platform duration; The time period during which the voltage change amplitude is continuously within the stable range of the platform is marked as the platform candidate interval; Read the duration interval and spectral energy distribution corresponding to the candidate interval of the platform, and remove the candidate intervals whose duration is less than the lower limit of the platform duration; The retained candidate platform intervals are arranged in order of acquisition time to generate voltage platform holding intervals; Identify the plateau interruption location in the discharge phase mode, specifically: Read the voltage plateau hold interval, discharge phase modes, and hysteresis deviation representations; Starting from the end acquisition time of the voltage platform holding interval, read the voltage mode components and spectral energy distribution in the order of acquisition time; Calculate the drop amplitude between voltage mode components corresponding to adjacent acquisition times, and generate a voltage drop sequence after plateau; When the voltage drop sequence after the platform changes from a stable state to a continuous state, and the spectral energy response value at the corresponding acquisition time deviates continuously from the average spectral energy value within the voltage platform holding interval, the acquisition time corresponding to the change position is marked as the candidate platform interruption position. If there is a center frequency offset record or spectrum energy offset record within the acquisition time range corresponding to the candidate platform interruption location, the candidate platform interruption location is retained, and the retained candidate platform interruption location is determined as the platform interruption location. The specific location of the platform drop-off at the end of the discharge phase mode is identified as follows: Read the discharge phase modes after the platform interruption location, and extract the voltage mode components and amplitude change direction according to the acquisition time sequence; The time length corresponding to 5 consecutive sampling points is determined as the lower limit of the duration of the final fall; Calculate the decrease amplitude between voltage mode components corresponding to adjacent acquisition times, and generate a platform end descent sequence; When the magnitude of the descent in the platform's terminal descent sequence increases continuously, and the direction of the magnitude change remains downward, the corresponding collection time period is marked as the candidate terminal descent interval. Read the duration range and spectral energy distribution corresponding to the candidate terminal drop interval, and filter out candidate terminal drop intervals whose duration is less than the lower limit of the terminal drop duration; The position with the earliest start time among the retained candidate end-fall intervals is determined as the end-fall position of the platform; The generation capacity platform fracture state representation is as follows: Write the voltage platform holding range, platform interruption position, and platform end drop position into the platform status latch unit, and associate them with the corresponding cell number, sensor number, acquisition time, and charge / discharge cycle number; Read the platform start time and platform end time corresponding to the voltage platform holding interval, and determine the time difference between the platform end time and the platform start time as the platform holding duration; Read the interruption acquisition time corresponding to the platform interruption location and the fall acquisition time corresponding to the platform end fall location, and determine the time difference between the fall acquisition time and the interruption acquisition time as the duration of platform fracture. The platform holding time, platform fracture duration, platform interruption location, platform end drop location, and hysteresis deviation are mapped to generate a platform fracture status record. The platform fracture status records are arranged according to cell number, charge / discharge cycle number, and acquisition time. The arranged platform fracture status records are input into the fracture status encoder in the order of the fields: platform holding time, platform fracture duration, platform interruption location, platform end drop location, and hysteresis deviation, to generate a capacity platform fracture status representation. The stationary recovery wake echo layer reads the stationary phase modes, capacity plateau fracture state representation, and operating condition stage identifier. It extracts the voltage rebound sequence, temperature drop sequence, pressure release sequence, and internal resistance recovery sequence from the stationary phase modes. Wake echo encoding is performed on each recovery sequence to generate a stationary recovery wake echo representation, where: The recovery of the wake echo layer after static placement includes: Static mode cache table: stores the static mode and its corresponding cell number, sensor number, acquisition time, center frequency range and amplitude change direction; Fracture Status Input Register: Represents the capacity plateau fracture status output from the capacity plateau fracture latch layer; Static Phase Index Table: Records the start and end times of the static phase data acquisition based on the phase identifier, along with the corresponding charge / discharge cycle number. Recovery sequence extraction queue: Extract the voltage rebound sequence, temperature drop sequence, pressure release sequence, and internal resistance recovery sequence from the resting phase mode respectively; Wreath origin locator: determines the recovery start acquisition time corresponding to each recovery sequence; Wake decay tracker: records the amplitude change direction and recovery duration interval of each recovery sequence during the resting phase; Echo state encoder: Encodes the wake changes corresponding to voltage rebound, temperature drop, pressure release, and internal resistance recovery; Wake echo outputter: Generates a stationary recovery wake echo representation based on the output of the echo state encoder and the capacity plateau fracture state representation; In the stationary recovery wake echo layer, the stationary mode buffer table transmits the stationary phase modes and corresponding cell numbers, sensor numbers, acquisition times, center frequency ranges, and amplitude change directions to the recovery sequence extraction queue. The stationary phase index table provides the recovery sequence extraction queue with the stationary phase start acquisition time, end acquisition time, and corresponding charge / discharge cycle number based on the operating condition phase identifier. The recovery sequence extraction queue extracts the voltage rebound sequence, temperature drop sequence, pressure release sequence, and internal resistance recovery sequence accordingly. Each recovery sequence is input to the wake start locator to determine the recovery start acquisition time and is transmitted to the wake attenuation tracker to record the amplitude change direction and recovery duration range. The echo state encoder receives the output of the wake attenuation tracker and generates the wake echo code. The fracture state input register transmits the capacity platform fracture state representation to the wake echo output device. The wake echo output device generates the stationary recovery wake echo representation based on the wake echo code and the capacity platform fracture state representation. The voltage rebound sequence, temperature drop sequence, pressure release sequence, and internal resistance recovery sequence are extracted from the static phase modes, specifically as follows: Read the static phase mode and working condition phase identifier, and determine the static phase acquisition time range based on the static phase start acquisition time and static phase end acquisition time; Based on the cell number, sensor number, and data type, voltage mode components, temperature mode components, pressure mode components, and internal resistance mode components are extracted from the static mode modes within the time range of the static stage acquisition. The voltage mode component corresponding to the initial acquisition time of the static phase is taken as the voltage recovery starting point, and the sequence formed by the change of voltage mode components with acquisition time during the static phase is denoted as the voltage rebound sequence. The temperature modal component corresponding to the initial acquisition time of the static phase is taken as the temperature recovery starting point, and the sequence formed by the change of temperature modal components with acquisition time during the static phase is recorded as the temperature fall-off sequence. The pressure modal component corresponding to the start time of the settling phase is taken as the pressure release starting point, and the sequence formed by the change of pressure modal components with the acquisition time during the settling phase is recorded as the pressure release sequence. The internal resistance modal component corresponding to the initial acquisition time of the static stage is taken as the starting point of internal resistance recovery, and the sequence formed by the change of internal resistance modal component with acquisition time during the static stage is denoted as the internal resistance recovery sequence. Wake echo coding is performed on each recovered sequence, specifically as follows: Calculate the difference between the recovery start value and the termination acquisition time value in each recovery sequence to generate the voltage rebound amplitude, temperature drop amplitude, pressure release amplitude, and internal resistance recovery amplitude; Calculate the direction and magnitude of change of values ​​corresponding to adjacent acquisition times in each recovery sequence according to the acquisition time sequence, and generate a recovery change sequence; Read the values ​​corresponding to the last 5 sampling points of each recovery sequence, calculate the average value of the last 5 sampling points, and generate the tail segment baseline value; Calculate the difference between the termination acquisition time value and the tail segment reference value to generate the tail segment residual deviation; The time length between the start and end acquisition times, in which the direction of change in the recovery sequence remains consistent, is defined as the recovery duration interval. The voltage rebound amplitude, temperature drop amplitude, pressure release amplitude, internal resistance recovery amplitude, recovery duration range, and tail residual deviation are encoded according to the cell number, sensor number, and the collection time during the resting stage to generate a tail echo code. The hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation are combined, and a multi-stage degradation feature representation is generated through temporal lightweight adaptive coding, wherein: Generate multi-stage degradation feature representations, specifically: By combining the same cell number, the same charge-discharge cycle number, and the hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation within the corresponding acquisition time range, a stage degradation combination representation is generated. Channel normalization is performed on the stage-degenerate combinatorial representation to map the values ​​in different representations to a unified feature scale; Arrange the normalized stage degradation combination representations in the order of collection time to generate the stage degradation time sequence representation; Lightweight temporal coding is performed on the stage-degraded temporal representation to extract degradation change features between adjacent charge-discharge cycles and generate temporal degradation embeddings; The temporal degradation embedding is mapped to the corresponding hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation to generate a multi-stage degradation feature representation; The improved TSLANT model was trained using a combination of health status assessment error, hysteresis deviation representation error, capacity plateau fracture state error, stationary recovery wake echo error, and multi-stage degradation feature representation consistency error as joint optimization objectives. The network parameters of the charge / discharge hysteresis mirror spectrum layer, capacity plateau fracture latch layer, and stationary recovery wake echo layer were continuously optimized. When the change in the joint loss value over five consecutive training rounds was less than 0.001, the improved TSLANT model was considered to have completed convergence training. The improved TSLaneet model is trained as follows: The health-sensitive modal sequence, operating condition stage identifier, SOH calibration value, capacity retest record, internal resistance retest record, and static recovery record are read from the training samples. The health-sensitive modal sequence and operating condition stage identifier are input into the improved TSLANT model, which passes through the charge-discharge hysteresis mirror spectrum layer, capacity plateau fracture latch layer, and static recovery wake echo layer in sequence. The outputs hysteresis deviation representation, capacity plateau fracture state representation, static recovery wake echo representation, multi-stage degradation feature representation, and health status prediction results. The training samples are divided according to cell number and charge-discharge cycle number. The training batch size is set to 64, and 64 charge-discharge cycle samples are read in each training round. The initial learning rate is set to 0.001, and the maximum number of training rounds is set to 200 rounds. The average of the squared differences between the SOH calibration value and the SOH prediction value in the health status prediction results is used to obtain the health status assessment error. The average of the squared differences between the calibration hysteresis deviation record between the spectral trajectory of the charging stage and the mirror spectral trajectory of the discharging stage in the training samples and the hysteresis deviation representation output by the charge-discharge hysteresis mirror spectrum layer is used to obtain the hysteresis deviation representation error. The average of the squared differences between the platform interruption mark and the platform end drop mark corresponding to the capacity retest record and the capacity platform fracture state representation output by the capacity platform fracture latch layer is used to obtain the capacity platform fracture state error. The average of the squared differences between the voltage rebound time, temperature drop time, pressure release residual value and internal resistance recovery deviation in the static recovery record and the static recovery wake echo representation output by the static recovery wake echo layer is used to obtain the static recovery wake echo error. The consistency error of the multi-stage degradation feature representation difference corresponding to adjacent charge-discharge cycles is obtained by comparing the change direction of the multi-stage degradation feature representation. The health status assessment error is multiplied by 0.30, the hysteresis deviation representation error by 0.20, the capacity plateau fracture state error by 0.20, the static recovery wake echo error by 0.15, and the multi-stage degradation feature representation consistency error by 0.15, and then summed to generate a joint loss value. The rate of change of the loss value corresponding to each network parameter in the charge-discharge hysteresis mirror spectrum layer, the capacity plateau fracture latch layer, and the static recovery wake echo layer is read to obtain the gradient value corresponding to each network parameter. The gradient value is multiplied by the current learning rate and then subtracted from the current network parameter value to obtain the updated network parameter value, which is then rewritten into the charge-discharge hysteresis mirror spectrum layer, the capacity plateau fracture latch layer, and the static recovery wake echo layer. When the change amplitude of the joint loss value in 5 consecutive training rounds is less than 0.001, the improved TSLnet model is considered to have completed convergence training.

[0025] In this embodiment, generating the lithium battery health status assessment result includes: Read the multi-stage degradation feature representation, sensor drift residual record and degradation spectrum mode sequence, and establish feature correspondence according to cell number and acquisition time; Hysteresis deviation features, capacity plateau fracture features, static recovery wake features, and temporal degradation embeddings are extracted from the multi-stage degradation feature representation. Lineage continuation state, lineage branching state, lineage merging state, and new lineage state are extracted from the degradation lineage mode sequence to generate a health mapping feature set, wherein: Generate a health mapping feature set, specifically as follows: Read the multi-stage degradation characteristics and arrange them according to cell number, charge / discharge cycle number and acquisition time; Extract the center frequency shift, spectral energy shift, direction shift and duration shift corresponding to the hysteresis deviation representation from the multi-stage degradation feature representation to generate hysteresis deviation features; Extract the platform holding time, platform fracture duration, platform interruption location, and platform end drop location corresponding to the capacity platform fracture state representation from the multi-stage degradation feature representation to generate capacity platform fracture features; Extract the voltage rebound amplitude, temperature drop amplitude, pressure release amplitude, internal resistance recovery amplitude, recovery duration range, and tail residual deviation corresponding to the static recovery wake echo representation from the multi-stage degradation feature representation to generate static recovery wake features; Extract the temporal degradation embedding between adjacent charge-discharge cycles from the multi-stage degradation feature representation; Read the degenerate lineage modal sequence and extract the lineage continuation state, lineage branch state, lineage merging state, and new lineage state according to the degenerate lineage number, degenerate lineage branch, degenerate lineage merging node, and new degenerate lineage number, respectively. Hysteresis deviation features, capacity plateau fracture features, static recovery wake features, temporal degradation embedding, lineage continuation status, lineage branching status, lineage merging status, and new lineage status are combined according to cell number, charge / discharge cycle number, and acquisition time to generate a health mapping feature set; The drift sensing channel is determined based on the sensor drift residual record, and the features corresponding to the drift sensing channel are marked as sensing reliability judgment features to generate a corrected health mapping feature set, wherein: Generate the corrected health mapping feature set, specifically as follows: Read the sensor drift residual records and determine the drift sensing channels with sensor drift residuals according to cell number, sensor number, data type and acquisition time; The drift sensing channel is mapped to the hysteresis deviation feature, capacity plateau fracture feature, static recovery wake feature and temporal degradation embedding in the health mapping feature set; When the features in the health mapping feature set originate from the drift sensing channel, the corresponding features are marked as sensing reliability judgment features. Based on the type of residual in the sensor drift residual record, the sensor reliability judgment feature is marked as zero-point offset correlation feature, reference drift correlation feature or slow measurement deviation correlation feature; Write the sensor reliability judgment features and corresponding residual type labels into the health mapping feature set to generate the corrected health mapping feature set; Based on the modified health mapping feature set, a health state mapping is performed to generate the SOH estimate, capacity decay state, internal resistance degradation state, thermal recovery state, and sensing reliability state, where: Perform health status mapping, specifically as follows: Hysteresis deviation features, capacity plateau break features, static recovery trail features, temporal degradation embedding, lineage continuation state, lineage branching state, lineage merging state, and new lineage state are extracted from the modified health mapping feature set, and a comprehensive degradation representation is generated based on the numerical values ​​corresponding to each feature. Based on the mapping relationship between the comprehensive degradation characterization and the SOH calibration values ​​in the training samples, the estimated SOH value for the current battery cell is generated. The capacity decay state is determined based on the platform holding time, platform fracture duration, platform interruption location, and platform end drop location in the capacity platform fracture characteristics. The internal resistance degradation state is determined based on the degradation change characteristics of adjacent charge-discharge cycles, the spectrum continuation state, the spectrum branching state, and the spectrum merging state characterized by the temporal degradation embedding. The thermal recovery state is determined based on the temperature drop amplitude, internal resistance recovery amplitude, recovery duration range, and residual deviation in the tail segment of the static recovery wake characteristics. The sensor reliability status is determined based on the zero-point offset correlation characteristics, reference drift correlation characteristics, and slow measurement deviation correlation characteristics corresponding to the sensor reliability judgment characteristics. The estimated SOH value, capacity decay status, internal resistance degradation status, thermal recovery status, and sensing reliability status are mapped according to the cell number, charge / discharge cycle number, and acquisition time to generate the lithium battery health status assessment result. The estimated SOH, capacity decay status, internal resistance degradation status, thermal recovery status, and sensor reliability status are archived according to the cell number and collection time to generate lithium battery health status assessment results.

[0026] In this embodiment, the step of generating the lithium battery health status output result includes: The health level is determined based on the estimated SOH value in the lithium battery health status assessment results, where: Health levels are determined based on estimated SOH values, specifically as follows: Read the estimated SOH value from the lithium battery health status assessment results and match it with the cell number and charge / discharge cycle number; When the estimated SOH value is greater than or equal to 90% and less than or equal to 100%, the corresponding cell will be marked as Grade A health status; When the estimated SOH value is greater than or equal to 80% and less than 90%, the corresponding cell will be marked as Grade B health status; When the estimated state of health (SOH) is greater than or equal to 70% and less than 80%, the corresponding cell will be marked as a Class C health condition. When the estimated SOH value is greater than or equal to 60% and less than 70%, the corresponding cell will be marked as Grade D health status; When the estimated SOH value is less than 60%, the corresponding cell will be marked as Grade E health status; The health status of Grade A, Grade B, Grade C, Grade D, and Grade E are archived according to the cell number and charge / discharge cycle number to generate a health rating. The degradation rate is obtained by dividing the difference in SOH estimates between adjacent acquisition times by the cycle number increment. The remaining lifetime prediction interval is then determined based on the current SOH estimate and the degradation rate, where: The remaining lifetime prediction range is determined as follows: Read the SOH estimate corresponding to the current acquisition time, the SOH estimate corresponding to the adjacent acquisition time, and the loop count increment; Subtract the estimated SOH value from the estimated SOH value of the next acquisition time to generate the SOH decrease value, and divide the SOH decrease value by the increment of the number of cycles to obtain the single-segment degradation rate. Read the single-segment degradation rate corresponding to the 10 consecutive charge-discharge cycles before the current acquisition time in the order of acquisition time, and retain the single-segment degradation rate with a value greater than 0. Extract the minimum and maximum degradation rates from the retained single-segment degradation rates to generate degradation rate intervals; The lifespan prediction boundary value is determined based on the lower limit SOH value of Grade D health status in the health level classification. Divide the difference between the current SOH estimate and the lifetime prediction boundary value by the maximum degradation rate to obtain the lower limit of the remaining lifetime prediction interval; Divide the difference between the current SOH estimate and the lifetime prediction boundary value by the minimum degradation rate to obtain the upper limit of the remaining lifetime prediction interval; The lower limit and upper limit of the remaining life prediction range are recorded according to the cell number and charge / discharge cycle number to generate the remaining life prediction range. Anomaly risk identifiers and sensing anomaly records are generated based on capacity decay status, internal resistance degradation status, thermal recovery status, and sensing reliability status, among which: Generate anomaly risk identifiers and sensor anomaly records, specifically as follows: A capacity decay anomaly flag is generated when any of the following conditions are met: the platform holding time corresponding to the capacity decay state is less than the platform holding time of the previous charge-discharge cycle, the platform breakage duration is greater than the platform breakage duration of the previous charge-discharge cycle, or the platform end drop position is earlier than the platform end drop position of the previous charge-discharge cycle. An internal resistance degradation anomaly flag is generated when any of the following conditions are met: the internal resistance recovery amplitude corresponding to the internal resistance degradation state is greater than the internal resistance recovery amplitude of the previous charge-discharge cycle; the internal resistance recovery duration is longer than the internal resistance recovery duration of the previous charge-discharge cycle; or the number of degradation spectrum branches is greater than the number of degradation spectrum branches of the previous charge-discharge cycle. A thermal recovery anomaly flag is generated when any of the following conditions are met: the temperature drop in the thermal recovery state is less than the temperature drop in the previous charge-discharge cycle; the recovery duration is longer than the recovery duration in the previous charge-discharge cycle; or the residual deviation at the end of the charge-discharge cycle is greater than the residual deviation at the end of the charge-discharge cycle. The abnormal capacity decay flag, abnormal internal resistance degradation flag, and abnormal thermal recovery flag are combined according to the cell number, charge / discharge cycle number, and data collection time to generate an abnormal risk flag. When the sensing reliability status includes any of the zero-point offset associated status, reference drift associated status, or slow measurement deviation associated status, read the corresponding cell number, sensor number, data type, and acquisition time. Write the corresponding cell number, sensor number, data type, acquisition time, and sensor reliability status into the sensor anomaly record; The health level, remaining life prediction range, abnormal risk indicators, and sensor anomaly records are combined according to the cell number and collection time to form the output result of the lithium battery health status.

[0027] In this embodiment, the step of collecting lithium battery operation feedback data and maintenance verification data and performing feedback updates includes: Collect lithium battery operation feedback data and maintenance verification data, and generate health assessment feedback records according to cell number, collection time, and maintenance time, including: The lithium battery operation feedback data and maintenance verification data are as follows: The lithium battery operation feedback data includes feedback voltage data, feedback current data, feedback temperature data, feedback internal resistance data, feedback pressure data, feedback vibration data, feedback charge / discharge status data, and feedback cycle count data. Maintenance and verification data include capacity retest data, internal resistance retest data, sensor verification data, anomaly handling records, and cell maintenance records; Health assessment feedback records are written into a standardized battery sensor dataset, and maintenance verification data is correlated with lithium battery health status assessment results; Correct the degenerate lineage number, degenerate lineage branch, degenerate lineage merging node, and newly formed degenerate lineage number in the degenerate lineage mode sequence based on the maintenance and verification data; Update the sensor drift residual isolation parameters, degradation spectrum mode inheritance parameters, and improved TSLnet model parameters based on health assessment feedback records.

[0028] refer to Figure 3 A lithium battery health status assessment system based on intelligent sensors includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source sensor data of lithium batteries and perform preprocessing to generate a standardized battery sensor dataset. The time-series sample construction module is used to perform time-series orchestration on the standardized battery sensing dataset to generate multi-source sensing time-series samples and operating condition stage identifiers. The mode decomposition processing module is used to perform sensor drift residual isolation and degradation lineage mode inheritance, generating sensor drift residual records, degradation lineage mode sequences, and health-sensitive mode sequences; The degradation feature extraction module is used to construct an improved TSLANet model and generate multi-stage degradation feature representations; The health status assessment module is used to perform health status mapping and generate lithium battery health status assessment results. The health outcome output module is used to generate health level, remaining life expectancy prediction range, abnormal risk indicators, and sensor anomaly records. The feedback update module is used to collect lithium battery feedback data and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a lithium battery module health status assessment scenario. The test object was a lithium battery module composed of 24 cells with a rated capacity of 50Ah. Each cell was equipped with a voltage sampling channel, a temperature sampling channel, and an internal resistance detection channel. The module side was equipped with a current sensor, a pressure sensor, and a vibration sensor. During the test, the lithium battery module operated according to a cycle of charging, resting, discharging, and resting, collecting data for a total of 320 charge-discharge cycles with a sampling interval of 1 second. The raw data included 2,764,800 voltage records, 115,200 current records, 2,764,800 temperature records, 23,040 internal resistance records, 115,200 pressure records, and 115,200 vibration records. To simulate typical problems in actual operation, a slow reference drift was set in the temperature sensing channel of the 9th cell, with the drift amplitude gradually increasing from 0.2℃ to 1.6℃; a zero-point offset was set in the voltage sampling channel of the 15th cell, with an offset amplitude of 7.5mV; and a capacity decay phenomenon was set in the 18th cell, with the discharge platform holding time decreasing from 1830 seconds in the early cycle to 1475 seconds in the later cycle.

[0030] After the data enters the processing flow, the system first preprocesses the multi-source sensor data. The original records contained 4386 duplicate records, 916 invalid voltage records, 1280 missing temperature records, and 342 records of abnormal vibration jumps. The system deletes duplicate and invalid records, fills in missing temperature records, and replaces records of abnormal vibration jumps. After processing, the voltage data missing rate decreased from 0.18% to 0, the temperature data missing rate decreased from 0.46% to 0, and the number of peaks in the vibration sequence decreased from 342 to 31. Subsequently, the system performs timestamp alignment according to a 1-second sampling interval and performs numerical standardization on the voltage, current, temperature, internal resistance, pressure, and vibration data to generate a standardized battery sensor dataset.

[0031] The system time-series organizes standardized battery sensor datasets according to cell number, sensor number, acquisition time, and operating condition stage. Taking the 260th cycle of the 18th cell as an example, this cycle includes 2140 sampling points in the charging stage, 600 sampling points in the first resting stage, 1985 sampling points in the discharging stage, and 720 sampling points in the second resting stage. The system divides the operating condition stages based on the charge / discharge status field and current direction information, and combines the single-channel sensing time sequence segments corresponding to voltage, temperature, internal resistance, pressure, vibration, and current to form a multi-source sensing time sequence sample. In this sample, the voltage plateau range in the discharging stage is from 3180 seconds to 4655 seconds, the starting point of the plateau drop is at 4656 seconds, the voltage rebound duration in the resting stage is 412 seconds, and the time for the temperature to drop back to the stable range is 538 seconds.

[0032] After inputting multi-source sensor timing samples and operating condition stage identifiers into the successive variational mode decomposition algorithm, the system extracts sensing modes for each sensing channel round by round. Taking the temperature sequence of the 9th cell as an example, the system extracts the temperature rise response mode in the first round, with a center frequency range of 0.006Hz to 0.011Hz; the static fallback mode is extracted in the second round, with a center frequency range of 0.002Hz to 0.005Hz; in the third round, a continuous unidirectional rise component appears in the residual signal, with a rise slope of 0.0048℃ per cycle. This component is written into the baseline drift residual record. If the traditional mode decomposition method is used, this residual will be incorporated into the temperature degradation mode, thus amplifying the thermal degradation degree of the 9th cell. After performing sensing drift residual isolation in this invention, this drift component no longer enters the health-sensitive mode sequence.

[0033] In the 15th cell voltage sequence, the system extracted the discharge plateau variation mode and zero-point offset residual. The discharge plateau variation mode persisted after the 220th cycle, with the average plateau voltage decreasing from 3.218V to 3.164V, and the duration shortening from 1768 seconds to 1512 seconds. The zero-point offset residual maintained an approximately fixed offset during the charging, discharging, and resting phases, with an average offset of 7.3mV. The system wrote the zero-point offset residual into the sensor drift residual record and retained the discharge plateau variation mode as a health-sensitive mode, thereby avoiding interference from voltage sampling zero-point offset in the capacity plateau breakage judgment.

[0034] During the inheritance of degradation spectrum modes, the system establishes a mode registration table according to the charge-discharge cycle sequence. Taking the 18th cell as an example, the discharge plateau mode appears consecutively in the 180th, 220th, 260th, and 300th cycles. The center frequency range gradually shifts from 0.004Hz to 0.007Hz to 0.003Hz to 0.006Hz, and the duration range decreases from 1830 seconds to 1475 seconds, with the amplitude changing in a decreasing direction. The system marks this mode as a continuation of the same degradation spectrum. Simultaneously, the resting temperature recovery mode splits into a recovery delay branch after the 240th cycle, with the temperature drop time increasing from 421 seconds to 563 seconds. This branch is written into the degradation spectrum branch. After processing, the system generates 96 sensor drift residual records, 184 degradation spectrum mode sequences, and 142 health-sensitive mode sequences.

[0035] After inputting the health-sensitive modal sequence and operating condition stage identifier into the improved TSLANT model, the charge / discharge hysteresis mirror spectrum layer encodes the charging and discharging modes, respectively. Taking the 260th cycle of the 18th cell as an example, the peak value of the voltage rise response in the charging phase spectrum trajectory is 0.72, and the peak value of the voltage plateau decay response in the discharging phase mirror spectrum trajectory is 0.81. The offset between the two is 0.19, which is higher than the average offset of 0.07 for healthy samples, and the system generates a hysteresis deviation representation. The capacity plateau break latch layer identifies the plateau interruption position at 4656 seconds and the duration of the plateau drop at the end of the plateau at 329 seconds, and writes the plateau breakage state into the plateau state latch unit. The stationary recovery wake echo layer encodes the voltage rebound, temperature drop, pressure release, and internal resistance recovery sequences of the stationary phase, obtaining the stationary recovery wake echo representation. The voltage rebound completion time is 412 seconds, the temperature drop completion time is 538 seconds, the pressure release residual deviation is 0.034, and the internal resistance recovery deviation is 0.012mΩ. The model combines the various representations to generate a multi-stage degradation feature representation.

[0036] The system performs health status mapping based on multi-stage degradation feature representation, sensor drift residual records, and degradation spectrum mode sequences. Taking the 260th cycle of the 18th cell as an example, the system extracts hysteresis deviation feature (0.19), capacity plateau fracture feature (0.76), static recovery wake feature (0.68), spectrum continuation state 1, spectrum branching state 1, and sensor reliability state (0.93), ultimately generating a SOH estimate of 84.6%, a capacity decay state of moderate decay, an internal resistance degradation state of mild degradation, a thermal recovery state of delayed recovery, and a sensor reliability state of reliable. During maintenance verification, the cell's capacity retest result was 42.1 Ah, corresponding to an SOH of 84.2%, with an error of 0.4 percentage points. Traditional methods directly use the original voltage, current, and temperature sequences for time-series prediction, outputting an SOH of 87.9% with an error of 3.7 percentage points, and failing to identify temperature sensor drift.

[0037] Based on the lithium battery health status assessment results, the system classifies the 18th cell as Grade B and calculates the degradation rate based on the difference in SOH estimates between adjacent acquisition times and the cycle count increment, obtaining an average decrease of 0.018 percentage points per cycle. This determines the remaining lifetime prediction range to be 214 to 268 cycles. Traditional methods yield a remaining lifetime prediction range of 168 to 341 cycles, with a range width of 173 cycles; this invention's range width is 54 cycles. For abnormal risk identification, this invention generates a capacity decay anomaly identifier, a thermal recovery anomaly identifier, and a sensor anomaly record for the 9th temperature channel; traditional methods only generate a capacity decay anomaly identifier.

[0038] In the comparative experiment, the number of training samples was 7680 cyclic samples, and the number of validation samples was 1920 cyclic samples.

[0039] The average absolute error of SOH using the traditional method is 2.91%, while that of this invention is 1.08%. The root mean square error of SOH using the traditional method is 3.36%, while that of this invention is 1.47%. The traditional method has 43 false alarms due to sensor drift, while this invention has 9. The accuracy rate of capacity plateau fracture identification using traditional methods is 80.6%, while that of this invention is 94.1%. The accuracy rate of anomaly identification using traditional methods after static recovery is 77.8%, while that of this invention is 90.9%. The average width of the remaining lifetime prediction interval for traditional methods is 136 cycles, while that of this invention is 69 cycles.

[0040] Regarding the average processing time for a single cycle sample, the traditional method takes 0.38 seconds, while this invention takes 0.52 seconds, meeting the requirements for online health status assessment.

[0041] As can be seen from Example 1, the present invention establishes a clear data change process in standardization, sensor drift residual isolation, degradation spectrum mode inheritance, improved TSLANT model feature extraction, and health status output. Compared with traditional health assessment methods, the present invention can distinguish between sensor drift and actual battery degradation, continuously track cross-cycle degradation modes, and extract charge / discharge hysteresis, capacity plateau breakage, and resting recovery wake information, demonstrating better accuracy and stability in SOH estimation, remaining lifetime prediction, and anomaly risk identification.

[0042] 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 assessing the health status of a lithium battery based on intelligent sensors, characterized in that, include: Collect multi-source sensor data during the operation of lithium batteries, perform preprocessing on the multi-source sensor data, and generate a standardized battery sensor dataset; The standardized battery sensor dataset is time-series arranged according to cell number, sensor number, acquisition time and operating condition stage to generate multi-source sensor time-series samples and operating condition stage identifiers. Multi-source sensor time-series samples and operating condition stage identifiers are input into the successive variational mode decomposition algorithm. The sensor modes are extracted round by round and the residual signals are updated. Sensor drift residual isolation and degradation spectrum mode inheritance are performed to generate sensor drift residual records, degradation spectrum mode sequences and health-sensitive mode sequences. An improved TSLANT model is constructed, which includes a charge-discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer. Based on the operating condition stage identifier, the health-sensitive mode sequence is processed to generate hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation, respectively, and combined to generate a multi-stage degradation feature representation. Health status mapping is performed based on multi-stage degradation feature representation, sensor drift residual records, and degradation spectrum mode sequences to generate lithium battery health status assessment results; Based on the health status assessment results of lithium batteries, the health level is divided, the remaining life prediction range, abnormal risk indicators and sensor abnormality records are determined, and the lithium battery health status output results are generated. Collect lithium battery operation feedback data and maintenance verification data, and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

2. The lithium battery health status assessment method based on intelligent sensors according to claim 1, characterized in that, The generation of the standardized battery sensing dataset includes: Collect multi-source sensor data during the operation of lithium batteries. The multi-source sensor data includes voltage data, current data, temperature data, internal resistance data, pressure data, vibration data, charge / discharge status data, cycle count data, and ambient temperature data. The multi-source sensor data are correlated according to the cell number, sensor number, acquisition time and data type to generate the original sensor record table; The original sensor record table is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and filling missing values ​​to generate a cleaned sensor record table. The cleaned sensor record table is timestamped according to a uniform sampling interval, and the values ​​corresponding to different data types are subjected to unit unification and numerical standardization to generate a standardized sensor record table. The standardized sensor record sheets are archived according to the cell number, sensor number, and acquisition time to generate a standardized battery sensor dataset.

3. The lithium battery health status assessment method based on intelligent sensors according to claim 1, characterized in that, The generation of multi-source sensing time-series samples and operating condition stage identifiers includes: Read the cell number, sensor number, acquisition time, standardized sensor value, charge / discharge status field and current direction information from the standardized battery sensing dataset; Establish a sensor channel index table according to cell number, sensor number and data type, and arrange the standardized sensor values ​​corresponding to each sensor channel in the order of acquisition time. The charging stage, discharging stage and resting stage are divided according to the charging and discharging status field and current direction information, and the start and end times of the stage are written into the operating condition stage identifier. Extract sensing sequence segments according to cell number, sensor number and operating condition stage identifier to generate single-channel sensing timing segments; Single-channel sensing time series segments with the same cell number, the same acquisition time range, and different sensor numbers are combined and associated with the operating condition stage identifier to generate multi-source sensing time series samples.

4. The lithium battery health status assessment method based on intelligent sensors according to claim 1, characterized in that, The generation of sensor drift residual records, degradation lineage mode sequences, and health-sensitive mode sequences includes: Read the multi-source sensor timing samples and operating condition stage identifiers, establish the sensor sequence to be decomposed according to the cell number, sensor number and operating condition stage, and use the sensor sequence to be decomposed as the initial residual signal; Successive variational mode decomposition is performed on the initial residual signal, the current sensing mode is extracted round by round, the center frequency range, duration range, amplitude change direction and corresponding operating condition stage of the current sensing mode are recorded, and the current sensing mode is subtracted from the current residual signal; After subtracting the current sensing mode in each round, sensor drift residual isolation is performed on the updated residual signal. The zero offset residual, reference drift residual and slow measurement deviation residual are written into the sensor drift residual record, and the sensor drift residual components are removed from the updated residual signal to generate a drift isolation residual signal. The drift isolation residual signal is used as the current residual signal for the new round. The sensor mode extraction, residual signal update and sensor drift residual isolation are repeated until no new sensor mode is extracted from the current residual signal. The sensor modes extracted in each round are summarized to generate the sensor mode set corresponding to each charge and discharge cycle. Based on the set of sensing modes corresponding to each charge-discharge cycle, a mode registration table is established according to the order of charge-discharge cycles. The sensing modes in the current charge-discharge cycle are matched with the sensing modes in the adjacent charge-discharge cycles. Based on the migration of the center frequency interval, the change of the duration interval, the direction of amplitude change and the consistency of the operating condition stage, the continuation relationship, split relationship, merging relationship and new generation relationship of the sensing modes are marked. Based on the continuation, splitting, merging, and new generation relationships of the sensing modes, the degradation lineage mode inheritance is performed. The corresponding sensing modes are written into the same degradation lineage, degradation lineage branch, degradation lineage merging node, and new degradation lineage number, respectively, and arranged in the order of charge and discharge cycles to generate a degradation lineage mode sequence. The sensor modes associated with sensor drift residual records are removed from the degraded spectrum mode sequence, and the remaining sensor modes are arranged according to cell number, sensor number, operating condition stage and acquisition time to generate a health sensitive mode sequence.

5. The lithium battery health status assessment method based on intelligent sensors according to claim 1, characterized in that, The generation of multi-stage degradation feature representations includes: An improved TSLANet model was constructed, which includes a charge / discharge hysteresis mirror spectrum layer, a capacity plateau fracture latch layer, and a static recovery wake echo layer. The charge / discharge hysteresis mirror spectrum layer reads the health-sensitive mode sequence and operating condition stage identifier, divides the health-sensitive mode sequence into charging stage mode, discharging stage mode and resting stage mode, performs spectral trajectory encoding on the charging stage mode and mirror spectral trajectory encoding on the discharging stage mode, and generates hysteresis deviation representation based on the offset relationship between the charging stage spectral trajectory and the discharging stage mirror spectral trajectory. The capacity platform fracture latch layer reads the discharge phase mode and hysteresis deviation representation, identifies the voltage platform holding interval, platform interruption position and platform end drop position in the discharge phase mode, writes the voltage platform holding interval, platform interruption position and platform end drop position into the platform state latch unit, and generates a capacity platform fracture state representation. The stationary recovery wake echo layer reads the stationary phase mode, capacity plateau fracture state representation and operating condition stage identifier, extracts the voltage rebound sequence, temperature drop sequence, pressure release sequence and internal resistance recovery sequence from the stationary phase mode, performs wake echo encoding on each recovery sequence, and generates the stationary recovery wake echo representation. Hysteresis deviation representation, capacity plateau fracture state representation, and static recovery wake echo representation are combined and multi-stage degradation feature representation is generated through temporal lightweight adaptive coding. The improved TSLANT model was trained using a combination of health status assessment error, hysteresis deviation representation error, capacity plateau fracture state error, static recovery wake echo error, and multi-stage degradation feature representation consistency error as joint optimization objectives. The network parameters of the charge-discharge hysteresis mirror spectrum layer, capacity plateau fracture latch layer, and static recovery wake echo layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved TSLANT model was considered to have completed convergence training.

6. The method for assessing the health status of a lithium battery based on intelligent sensors according to claim 1, characterized in that, The generated lithium battery health status assessment results include: Read the multi-stage degradation feature representation, sensor drift residual record and degradation spectrum mode sequence, and establish feature correspondence according to cell number and acquisition time; Hysteresis deviation features, capacity plateau fracture features, static recovery trail features, and temporal degradation embeddings are extracted from the multi-stage degradation feature representation. Lineage continuation state, lineage branching state, lineage merging state, and new lineage state are extracted from the degradation lineage mode sequence to generate a health mapping feature set. The drift sensing channel is determined based on the sensor drift residual record, and the features corresponding to the drift sensing channel are marked as sensing reliability judgment features to generate a corrected health mapping feature set. Based on the modified health mapping feature set, perform health state mapping to generate SOH estimate, capacity decay state, internal resistance degradation state, thermal recovery state, and sensing reliability state; The estimated SOH, capacity decay status, internal resistance degradation status, thermal recovery status, and sensor reliability status are archived according to the cell number and collection time to generate lithium battery health status assessment results.

7. The method for assessing the health status of a lithium battery based on intelligent sensors according to claim 1, characterized in that, The output result of the lithium battery health status includes: The health level is determined based on the estimated SOH value in the lithium battery health status assessment results. The degradation rate is obtained by dividing the difference in SOH estimates between adjacent acquisition times by the cycle number increment, and the remaining lifetime prediction interval is determined based on the current SOH estimate and degradation rate. Anomaly risk identifiers and sensing anomaly records are generated based on capacity decay status, internal resistance degradation status, thermal recovery status, and sensing reliability status. The health level, remaining life prediction range, abnormal risk indicators, and sensor anomaly records are combined according to the cell number and collection time to form the output result of the lithium battery health status.

8. The method for assessing the health status of a lithium battery based on intelligent sensors according to claim 1, characterized in that, The process of collecting lithium battery operation feedback data and maintenance verification data and performing feedback updates includes: Collect lithium battery operation feedback data and maintenance verification data, and generate health assessment feedback records according to cell number, collection time, and maintenance time; Health assessment feedback records are written into a standardized battery sensor dataset, and maintenance verification data is correlated with lithium battery health status assessment results; Correct the degenerate lineage number, degenerate lineage branch, degenerate lineage merging node, and newly formed degenerate lineage number in the degenerate lineage mode sequence based on the maintenance and verification data; Update the sensor drift residual isolation parameters, degradation spectrum mode inheritance parameters, and improved TSLnet model parameters based on health assessment feedback records.

9. A lithium battery health status assessment system based on intelligent sensors, comprising performing the lithium battery health status assessment method based on intelligent sensors as described in any one of claims 1 to 8, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source sensor data of lithium batteries and perform preprocessing to generate a standardized battery sensor dataset. The time-series sample construction module is used to perform time-series orchestration on the standardized battery sensing dataset to generate multi-source sensing time-series samples and operating condition stage identifiers. The mode decomposition processing module is used to perform sensor drift residual isolation and degradation lineage mode inheritance, generating sensor drift residual records, degradation lineage mode sequences, and health-sensitive mode sequences; The degradation feature extraction module is used to construct an improved TSLANet model and generate multi-stage degradation feature representations; The health status assessment module is used to perform health status mapping and generate lithium battery health status assessment results. The health outcome output module is used to generate health level, remaining life expectancy prediction range, abnormal risk indicators, and sensor anomaly records. The feedback update module is used to collect lithium battery feedback data and update the standardized battery sensor dataset, degradation spectrum mode sequence, successive variational mode decomposition algorithm parameters, and improved TSLANT model parameters.

Images