Multi-modal data fusion retired lithium battery health state evaluation method and system

Abstract

The invention relates to the field of ex-service lithium battery health state evaluation, in particular to a multi-modal data fusion ex-service lithium battery health state evaluation method and a multi-modal data fusion ex-service lithium battery health state evaluation system. The method comprises the following steps: constructing a multi-modal acquisition session parameter group by acquiring a retired lithium battery cell, a test tool and a model version parameter group, synchronously acquiring a voltage, current and temperature sequence, a thermal image sequence and a deformation sequence, and performing unified time axis alignment, drift correction, modal quality scoring and low-quality fragment elimination on multi-modal original data; extracting modal degradation features and cross-modal contrast features, completing consistency analysis and evidence weight distribution, generating a joint representation vector through hierarchical attention fusion reasoning, realizing health state regression, confidence calculation and reason category discrimination, outputting a health state assessment result, and generating model update parameters. According to the invention, a multi-mode health state assessment link with consistent data aperture and traceable process can be formed.

Description

Technical Field

[0001] This invention relates to the field of health status assessment of retired lithium batteries, and more particularly to a method and system for assessing the health status of retired lithium batteries using multimodal data fusion. Background Technology

[0002] In the field of health status assessment of retired lithium batteries, existing solutions combining retired lithium battery cells with testing fixtures typically revolve around the acquisition and analysis of voltage, current, and temperature sequences. In some scenarios, thermal image sequences and deformation sequences are introduced to form multimodal raw data packets. However, these solutions suffer from limitations such as inconsistent parameter sets in multimodal acquisition sessions, difficulty in reusing unified time axis alignment and drift correction, and a lack of traceable constraints in modal quality scoring. Existing methods often rely on separate acquisition and offline alignment across different equipment links. Under acquisition constraints such as synchronous trigger parameter set registration and contact force threshold registration, issues arise such as difficulty in establishing common origins between aligned data sequences and quality scoring vectors, inconsistencies between low-quality segment removal and subsequent analysis, and an inability to stably generate health status assessment results across continuous links. For the joint processing of quality score vectors and aligned data sequences, existing technologies generally lack consistent data structure constraints between modality degradation feature sets, cross-modal contrast feature sets, and time window aggregation. Cross-modal consistency parameter sets are difficult to form a traceable link with evidence weight allocation, and evidence weight parameter sets are difficult to provide a stable input caliber for hierarchical attention fusion network inference. As a result, the correspondence between joint representation vectors and health status regression calculation, confidence calculation, and cause category discrimination is difficult to solidify. Consequently, it is not easy to keep the records and version connections of treatment strategy selection, incremental update parameter sets, and model version parameter sets consistent in the engineering process. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a multimodal data fusion method for assessing the health status of retired lithium batteries, comprising:

[0004] Acquire the parameter sets of retired lithium battery cells, test fixtures and model versions, perform cell surface mesh division, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter set registration and contact force threshold registration, and generate multimodal acquisition session parameter sets;

[0005] Based on the multimodal acquisition session parameter group, voltage, current, and temperature sequences, thermal image sequences, and deformation sequences are synchronously acquired to generate multimodal raw data packets. Then, unified time axis alignment, drift correction, modal quality scoring, and low-quality segment removal are performed to generate quality score vectors and aligned data sequences.

[0006] Based on the quality score vector and aligned data sequence, modal degradation feature set and cross-modal comparison feature set are extracted and time window aggregation is performed. Cross-modal consistency parameter set is calculated and evidence weight is assigned to generate evidence weight parameter set.

[0007] Based on the evidence weight parameter set, hierarchical attention fusion network reasoning is performed to generate a joint representation vector;

[0008] Based on the joint representation vector, health status regression calculation, confidence calculation and cause category discrimination are performed to generate health status assessment results. Then, the treatment strategy is selected based on the health status assessment results, and incremental update parameter set and model version parameter set are generated.

[0009] Furthermore, the infrared thermal imaging field of view arrangement includes:

[0010] The infrared thermal image alignment area is determined based on the cell surface grid, and the cell surface grid range corresponding to the infrared thermal image field of view is registered in the multimodal acquisition session parameter group.

[0011] Furthermore, the arrangement of deformation measuring points includes:

[0012] Multiple deformation measurement points are selected based on the grid on the cell surface, and the correspondence between the deformation measurement points and the deformation sequence is registered in the multimodal acquisition session parameter group.

[0013] Furthermore, the synchronous trigger parameter group registration includes:

[0014] The voltage, current, and temperature sequences, thermal image sequences, and deformation sequences are registered at the same trigger time, and the sampling frequency of each sequence is registered in the multimodal acquisition session parameter group.

[0015] Furthermore, the modal quality score includes:

[0016] Quality scores are calculated for voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on multimodal raw data packets, and the quality scores are then aggregated to generate a quality score vector.

[0017] Furthermore, the extraction of modal degradation feature sets includes:

[0018] Capacity increment curve features are extracted from voltage, current and temperature sequences; temperature rise rate and thermal unevenness features are extracted from thermal image sequences; and expansion curve morphology features are extracted from deformation sequences.

[0019] Furthermore, cross-modal contrast feature extraction includes:

[0020] Based on the time window aggregation results of thermal image sequences and deformation sequences, a correspondence between thermal non-uniformity features and expansion curve morphology features is generated, and the correspondence is incorporated into the cross-modal comparison feature group.

[0021] Furthermore, the options for handling the situation include:

[0022] Based on the cause category discrimination results, a target strategy is selected from the retest strategy, isolation strategy and sorting and grouping strategy, and the target strategy is written into the health status assessment results.

[0023] Furthermore, the incremental update parameter set generation includes:

[0024] The aligned data sequence, health status assessment results, and evidence weight parameter set are combined as an incremental update parameter set and input into the online sequence extreme learning machine, which outputs the model version parameter set.

[0025] Furthermore, a multimodal data fusion-based health status assessment system for retired lithium batteries, applied to any of the methods described above, includes:

[0026] The acquisition session parameter group generation module is used to obtain the parameter groups of retired lithium battery cells, test fixtures and model versions, perform cell surface mesh generation, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter group registration and contact force threshold registration, and generate multimodal acquisition session parameter groups;

[0027] The multimodal synchronous acquisition and raw data packet generation module is used to synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on the multimodal acquisition session parameter group, and generate multimodal raw data packets.

[0028] The time alignment and quality control module is used to perform unified timeline alignment, drift correction, modal quality scoring, and low-quality segment removal on multimodal raw data packets, generating quality score vectors and aligned data sequences;

[0029] The evidence weight parameter group generation module is used to extract modal degradation feature sets and cross-modal contrast feature groups based on quality score vectors and aligned data sequences, perform time window aggregation, calculate cross-modal consistency parameter groups, perform evidence weight allocation, and generate evidence weight parameter groups.

[0030] The hierarchical attention fusion network inference module is used to perform hierarchical attention fusion network inference based on evidence weight parameter groups and generate joint representation vectors.

[0031] The health status assessment and model version parameter group generation module is used to perform health status regression calculation, confidence calculation and cause category discrimination based on the joint representation vector, and generate health status assessment results.

[0032] The key innovations of this invention include:

[0033] (1) Under the unified input caliber of quality score vector and aligned data sequence, extract modal degradation feature set and cross-modal comparison feature set and perform time window aggregation, calculate cross-modal consistency parameter set and perform evidence weight allocation to generate evidence weight parameter set.

[0034] (2) Based on the multimodal acquisition session parameter group, voltage, current and temperature sequences, thermal image sequences and deformation sequences are synchronously acquired to generate multimodal raw data packets. Unified time axis alignment, drift correction, modal quality scoring and low quality segment removal are performed on the multimodal raw data packets to generate quality score vector and aligned data sequence.

[0035] (3) Based on the evidence weight parameter group, perform hierarchical attention fusion network reasoning to generate joint representation vector, and perform health status regression calculation, confidence calculation and cause category discrimination based on the joint representation vector to generate health status assessment results. At the same time, select the treatment strategy for the health status assessment results to generate incremental update parameter group and model version parameter group.

[0036] The following are its main beneficial effects:

[0037] (1) In view of the problems in the existing technology, such as the lack of consistent data structure constraints between modal degradation feature set, cross-modal comparison feature set and time window aggregation, and the difficulty of forming a traceable link between cross-modal consistency parameter set and evidence weight allocation, the cross-modal consistency parameter set calculation and evidence weight allocation driven by quality score vector and aligned data sequence are used to form a verifiable weight source association under the time window aggregation caliber, reduce the caliber drift of cross-modal comparison under the conditions of data arrival difference and fragment missing test, and enable the subsequent hierarchical attention fusion network reasoning based on evidence weight parameter set to have a stable input constraint boundary.

[0038] (2) In response to the problems in the existing technology, such as inconsistent caliber of multimodal acquisition session parameter groups, difficulty in stable reuse of unified time axis alignment and drift correction, and lack of traceable constraints in modal quality scoring, which make it difficult to correlate aligned data sequences and quality score vectors from the same source, synchronous acquisition under the constraints of multimodal acquisition session parameter groups and encapsulation of multimodal raw data packets are carried out. Unified time axis alignment, drift correction, modal quality scoring and low-quality segment removal are completed in the same link, so that the quality score vector and the aligned data sequence maintain a consistent reference relationship within the same session boundary. This provides a verifiable data governance foundation for subsequent modal degradation feature set extraction and evidence weight allocation based on quality score vector and aligned data sequence.

[0039] (3) In response to the problems in the existing technology that the evidence weight parameter group is difficult to provide a stable input caliber for fusion reasoning, the correspondence between the joint representation vector and the health status regression calculation and cause category discrimination is difficult to solidify, and the version connection between the treatment strategy selection and the model version parameter group is inconsistent, the joint representation vector is generated by the hierarchical attention fusion network reasoning driven by the evidence weight parameter group. The joint representation vector is then connected to the health status regression calculation, confidence calculation and cause category discrimination to form the health status assessment result. At the same time, the product of the treatment strategy selection, the incremental update parameter group and the model version parameter group are included in the same output link, so that the health status assessment result and the incremental update parameter group and the model version parameter group form a consistent record and call relationship in the same assessment session, which is suitable for engineering scenarios with multimodal raw data packet quality fluctuations and multi-source acquisition time-series disturbances. Attached Figure Description

[0040] Figure 1 A flowchart illustrating the multimodal data fusion method for assessing the health status of retired lithium batteries provided in this application embodiment;

[0041] Figure 2 This is a structural block diagram of the multimodal data fusion-based health status assessment system for retired lithium batteries provided in an embodiment of this application. Detailed Implementation

[0042] Example 1: Refer to Figure 1 This is a flowchart illustrating the multimodal data fusion method for assessing the health status of decommissioned lithium batteries provided in this embodiment of the invention. The process may include at least steps S100-S500:

[0043] S100: Obtain the parameter set of retired lithium battery cells, test fixtures and model versions; perform cell surface mesh division, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter set registration and contact force threshold registration; generate multimodal acquisition session parameter set.

[0044] S200: Based on the multimodal acquisition session parameter group, synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences to generate multimodal raw data packets. Then, perform unified time axis alignment, drift correction, modal quality scoring, and low-quality segment removal to generate quality score vectors and aligned data sequences.

[0045] S300: Based on the quality score vector and aligned data sequence, extract the modal degradation feature set and cross-modal comparison feature set and perform time window aggregation, calculate the cross-modal consistency parameter set and perform evidence weight allocation, and generate the evidence weight parameter set.

[0046] S400: Based on the evidence weight parameter set, perform hierarchical attention fusion network reasoning to generate a joint representation vector;

[0047] S500 performs health status regression calculation, confidence calculation, and cause category discrimination based on joint representation vectors, generates health status assessment results, selects treatment strategies based on health status assessment results, and generates incremental update parameter groups and model version parameter groups.

[0048] S100: Obtain the parameter set of retired lithium battery cells, test fixtures and model versions; perform cell surface mesh division, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter set registration and contact force threshold registration; generate multimodal acquisition session parameter set.

[0049] Specifically, the input sources for this step include retired lithium-ion battery cells, test fixtures, and model version parameter sets. The retired lithium-ion battery cells are the individual cells to be evaluated and serve as the same object carrier for subsequent thermal image sequences, deformation sequences, and voltage-current-temperature sequences. The test fixtures are a set of experimental devices used to apply charge-discharge conditions to the retired lithium-ion battery cells and output voltage-current-temperature sequences. The model version parameters are the set of parameters called by subsequent hierarchical attention fusion network inference and health status regression calculations, and include a version number, parameter activation flag, and rollback flag. In this step, the acquisition and control unit receives the identification information, rated capacity information, and external dimension information of the retired lithium-ion battery cells, as well as the channel number information, range information, and sampling reference clock information of the test fixtures. Simultaneously, the model management unit reads the model version parameters and completes the version activation flag verification. Under the trigger condition that the version activation flag is valid, the deployment process begins. Under the trigger condition that the version activation flag is invalid or the rollback flag is valid, the previous version of the model version parameter set is called, and a version switching log is recorded. The version switching log is stored in association with the model version parameter set and written to the multimodal acquisition session parameter set at the end of this step for subsequent traceability.

[0050] Furthermore, the cell surface mesh is established on the observable surface of the retired lithium battery cell to form a two-dimensional partitioned coordinate system and a cell surface mesh. The cell surface mesh consists of partition numbers, partition boundaries and partition reference points. The partition numbers are consecutively numbered partition identifiers, the partition boundaries are the set of boundary lines on the surface of the retired lithium battery cell, and the partition reference points are the geometric center or preset feature points of each partition. The implementation process of cell surface meshing is as follows: The calibration unit fixes the retired lithium battery cell in the clamping position of the test fixture. The acquisition and control unit reads the external dimension information of the retired lithium battery cell and divides the visible surface into several rectangular or polygonal partitions according to the preset mesh scale. Then, the partition boundaries are numbered and the partition reference point coordinates are written into the cell surface mesh. When there is an abnormal situation such as occlusion area or incomplete visible surface of the retired lithium battery cell, the calibration unit records the occlusion area and marks the partition number corresponding to the occlusion area as an unusable partition. The unusable partition mark is written along with the cell surface mesh and is called as a constraint condition in the subsequent infrared thermal imaging field of view arrangement and deformation measurement point arrangement, so that the acquisition range of the subsequent thermal image sequence and deformation sequence is consistent with the actual observable area.

[0051] After the cell surface mesh is divided, the infrared thermal imaging field of view and deformation measurement point arrangement are executed. The infrared thermal imaging field of view arrangement is completed by the thermal imaging deployment unit. The infrared thermal image is an infrared thermal imaging temperature measurement device and includes an imaging lens, a thermal detector array, and temperature calibration parameters. The input of the infrared thermal imaging field of view arrangement is the cell surface mesh, and the output is the registration information of the partition range of the infrared thermal image alignment area and the infrared thermal image field of view coverage area. During the implementation, the thermal imaging deployment unit adjusts the installation position and elevation angle of the infrared thermal image on the test fixture so that the infrared thermal image field of view coverage area includes the target partition set in the cell surface mesh except for the unusable partitions, and fixes the position of the infrared thermal image through the field of view locking mechanism. Then, the acquisition control unit triggers the infrared thermal image acquisition reference frame and maps the pixel area in the reference frame to the partition boundary of the cell surface mesh, obtaining the cell surface mesh range registration information corresponding to the infrared thermal image field of view and writing it into the subsequent multimodal acquisition session parameter group. The deformation measurement point arrangement is completed by the deformation deployment unit. The deformation measurement points are the set of measurement points corresponding to the displacement measuring device on the surface of the retired lithium battery cell. The displacement measuring device includes a laser displacement sensor or a contact displacement probe and has a ranging range and a sampling reference clock. The input of the deformation measurement point arrangement is the cell surface grid and the output is the correspondence registration information between the deformation measurement points and the deformation sequence. During the implementation, the deformation deployment unit selects multiple partitions as the partitions where the deformation measurement points are located according to the partition reference points of the cell surface grid, and aligns the measuring spot or probe contact point of the displacement measuring device with the corresponding partition reference point. Then, the acquisition control unit performs zero-point calibration on each deformation measurement point and records the zero-point baseline. When the displacement measuring device signal is lost or the ranging exceeds the range during the zero-point calibration process, the acquisition control unit marks the deformation measurement point as an invalid measurement point and triggers the re-arrangement process until the trigger condition of the effective measurement point number threshold is met, and then proceeds to the next registration process.

[0052] After completing the infrared thermal imaging field of view setup and deformation measurement point setup, the synchronous trigger parameter group registration and contact force threshold registration are performed. The synchronous trigger parameter group registration is completed by the synchronous control unit. This synchronous trigger parameter group is a set of trigger times, sampling frequencies, and timestamp sources, used in subsequent steps to synchronously acquire voltage-current-temperature sequences, thermal image sequences, and deformation sequences at the same trigger time and to achieve unified time axis alignment. During implementation, the synchronous control unit selects the sampling reference clock of the test fixture as the timestamp source and generates a trigger time marker. It then registers the sampling frequency of the voltage-current-temperature sequences, the frame rate of the infrared thermal image, and the sampling frequency of the deformation sequence. When the three sampling frequencies are inconsistent, the synchronous control unit registers a resampling marker in the synchronous trigger parameter group and registers the alignment reference sampling frequency. This alignment reference sampling frequency serves as the input for drift correction and alignment data sequence generation in the subsequent unified time axis alignment stage. The contact force threshold registration is completed collaboratively by the deformation deployment unit and the acquisition control unit. The contact force threshold is the upper limit of the contact force or the set of non-contact ranging thresholds when the deformation measuring device contacts a retired lithium battery cell. The implementation process is as follows: In contact displacement probe mode, the deformation deployment unit reads the calibration parameters of the probe's elastic element and sets the upper limit of the contact force, while the acquisition control unit records the probe contact status marker. In laser displacement sensor mode, the deformation deployment unit sets the non-contact ranging threshold, and the acquisition control unit records the ranging status marker. Abnormal triggering conditions exist for contact force threshold registration. When the contact status marker indicates that the contact force exceeds the contact force threshold, or the ranging status marker indicates that the ranging exceeds the non-contact ranging threshold, the acquisition control unit triggers the deformation measuring point to realign and update the zero-point baseline. Simultaneously, the abnormal event is written to the deployment log. This deployment log, along with the cell surface grid, the cell surface grid range registration information corresponding to the infrared thermal imaging field of view, and the correspondence registration information between the deformation measuring point and the deformation sequence, are summarized for subsequent traceability.

[0053] After completing the above processing steps, the acquisition control unit encapsulates the cell surface grid, the cell surface grid range registration information corresponding to the infrared thermal image field of view, the correspondence registration information between deformation measurement points and deformation sequences, the synchronization trigger parameter group, and the contact force threshold into a multimodal acquisition session parameter group. The test fixture channel number and model version parameter group version number are written into the multimodal acquisition session parameter group, ensuring that the multimodal acquisition session parameter group possesses the object identifier, spatial mapping relationship, and temporal synchronization relationship for the same acquisition session. The output of this step is the output field name "Multimodal Acquisition Session Parameter Group," and the section completes the connection description with subsequent steps. That is, the multimodal acquisition session parameter group serves as the input position for the next step in the synchronous acquisition stage and is used to synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences. Simultaneously, it retains a reference relationship as a basic configuration item across main steps in the subsequent evidence weight parameter group generation and model version parameter group generation stages.

[0054] In summary, the technical effects of this step are as follows: By spatially mapping and registering the grid division of the battery cell surface, the arrangement of the infrared thermal imaging field of view, and the arrangement of deformation measurement points, and combining the registration of the synchronous trigger parameter group and the time and boundary constraints of the contact force threshold, a multimodal acquisition session parameter group is formed, providing a traceable session basis for the subsequent synchronous acquisition and alignment processing of multimodal raw data packets.

[0055] S200: Based on the multimodal acquisition session parameter group, synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences to generate multimodal raw data packets. Then, perform unified time axis alignment, drift correction, modal quality scoring, and low-quality segment removal to generate quality score vectors and aligned data sequences.

[0056] Specifically, the input source for this step is the session configuration set whose output field is named "Multimodal Acquisition Session Parameter Group" from the previous main step. This multimodal acquisition session parameter group includes registration information for the cell surface grid, the cell surface grid range corresponding to the infrared thermal imaging field of view, the correspondence registration information between deformation measurement points and deformation sequences, the synchronization trigger parameter group, the contact force threshold, and the test fixture channel number information. It also carries the model version parameter group version number for session traceability. This step is executed collaboratively by the acquisition control unit, the synchronization control unit, and the alignment processing unit. The acquisition control unit establishes an acquisition link with the test fixture, infrared thermal imaging, and displacement measurement device. The synchronization control unit is responsible for the unified scheduling of trigger times and timestamp sources. The alignment processing unit is responsible for unified time axis alignment and drift correction. The quality management unit is responsible for modal quality scoring and low-quality segment removal. The triggering condition for data acquisition is that the multimodal acquisition session parameter group is in a valid state and the trigger time marker in the synchronous trigger parameter group has been generated. If the trigger time marker is missing or the test fixture channel number information is inconsistent with the device connection status, the acquisition control unit writes the acquisition anomaly log and stops the operation of this step. The acquisition anomaly log is stored in association with the multimodal acquisition session parameter group and serves as the input basis for subsequent session playback.

[0057] Furthermore, based on the multimodal acquisition session parameter group, synchronous acquisition of voltage, current, and temperature sequences, thermal image sequences, and deformation sequences is performed, and multimodal raw data packets are generated. Specifically, the acquisition control unit sends charging and discharging conditions to the test fixture and begins acquiring voltage, current, and temperature sequences when the trigger time marker issued by the synchronous control unit arrives. The voltage, current, and temperature sequences consist of a set of sampling points, with each sampling point carrying a timestamp, channel number, and measured value. Simultaneously, under the same trigger time marker, the acquisition control unit drives the infrared thermal imaging system to acquire thermal image sequences. These thermal image sequences consist of a set of thermal image frames, with each thermal image frame carrying a frame timestamp, thermal image temperature calibration parameter index, and frame integrity marker. Under the same trigger time marker, the displacement measurement device is driven to acquire deformation sequences. These deformation sequences consist of a set of multiple displacement sampling points, with each sampling point carrying a timestamp, measurement point number, and displacement value. The measurement point number matches the correspondence registration information between the deformation measurement point and the deformation sequence. The acquisition control unit encapsulates the voltage, current, and temperature sequences, thermal image sequences, and deformation sequences into multimodal raw data packets according to session numbers. The multimodal raw data packets contain at least the session number, timestamp sequence, device connectivity marker, and acquisition anomaly marker, and continuously updates the acquisition anomaly marker during the acquisition process. When the acquisition anomaly marker indicates that the infrared thermal image frame is lost, the displacement measurement device signal is lost, or the test fixture channel reading exceeds the range, the acquisition control unit writes the corresponding anomaly type into the multimodal raw data packet and records the time period of the anomaly occurrence. The time period of the anomaly occurrence is used as the input for candidate segments to be eliminated in the subsequent low-quality segment elimination process.

[0058] After generating the multimodal raw data packets, a unified time axis alignment and drift correction are performed to generate an aligned data sequence. Specifically, the alignment processing unit extracts the timestamp sequence from the multimodal raw data packets and identifies the timestamp sources registered by the synchronization trigger parameter group. The timestamps of the voltage, current, and temperature sequences are used as the alignment reference to establish a unified time axis. When a resampling marker is registered in the synchronization trigger parameter group, the alignment processing unit performs time resampling on the thermal image sequence and the deformation sequence according to the alignment reference sampling frequency. For the thermal image sequence, a frame index sequence is generated using frame timestamp proximity mapping, and the thermal image frames are mapped to the corresponding time slices on the same time axis. For the deformation sequence, linear interpolation or spline interpolation is used to generate equally spaced displacement sampling points and map them to the unified time axis. The drift correction process is as follows: the alignment processing unit performs cumulative offset estimation on the offset between the frame timestamp and the trigger time marker of the thermal image sequence, and writes the cumulative offset estimate into the drift correction parameters. Simultaneously, it performs the same processing on the offset between the timestamp and the trigger time marker of the deformation sequence. Under the trigger condition that the cumulative offset estimate exceeds the drift threshold registered in the synchronous trigger parameter group, the alignment processing unit performs segmented correction on the timestamp of the corresponding mode and writes the segmented correction marker into the alignment data sequence. The alignment data sequence consists of the aligned voltage, current, and temperature sequences, the aligned thermal image sequence, and the aligned deformation sequence, forming a three-modal alignment record on each unified time slice. The alignment processing unit writes the alignment data sequence into the output field and retains the session number association with the multimodal acquisition session parameter group, thereby supporting subsequent cross-modal comparison feature group time window aggregation calls.

[0059] After obtaining the aligned data sequence, the quality management unit performs modal quality scoring and low-quality segment removal based on the multimodal raw data packet and the aligned data sequence, and generates a quality score vector. Specifically, modal quality scoring calculates quality scores for voltage-current-temperature sequences, thermal image sequences, and deformation sequences respectively. The quality score for the voltage-current-temperature sequence is calculated based on the sampling point missing ratio, channel reading saturation marker, and temperature reading jump marker; the quality score for the thermal image sequence is calculated based on frame integrity marker, field-of-view stability marker, and thermal image temperature calibration parameter index consistency marker; the quality score for the deformation sequence is calculated based on the contact state marker or ranging state marker corresponding to the contact force threshold, signal loss marker, and zero-point baseline drift marker. The quality management unit summarizes the three types of quality scores along a unified time axis and writes them into the quality score vector. The quality score vector corresponds one-to-one with the unified time axis and carries a modal identifier. The process of removing low-quality segments is as follows: The quality management unit retrieves a set of time slices below a preset scoring threshold from the quality scoring vector, maps the set of time slices back to the aligned data sequence, and marks the corresponding intervals as low-quality segments. Then, the low-quality segments are removed from the aligned data sequence or marked as missing segments, and the index of the removed segments is recorded. When the trigger condition that the continuous length of low-quality segments exceeds a preset length threshold occurs, the quality management unit writes the session number into the retest trigger flag and appends the retest trigger flag to the multimodal original data packet. The retest trigger flag participates in the policy determination in the subsequent handling policy selection stage but does not change the output field of this step.

[0060] The outputs of this step include a quality score vector and an aligned data sequence. The quality score vector serves as the input for the next step, entering the cross-modal contrast feature group extraction and evidence weight allocation stage, and acting as the prior input for the subsequent evidence weight parameter group generation. The aligned data sequence serves as the input for the next step, entering the modality degradation feature set extraction and time window aggregation stage, supporting the time window aggregation of the cross-modal contrast feature group and the calculation of the cross-modal consistency parameter group. Understandably, this step retains the session number of the multimodal acquisition session parameter group at the cross-main step connection and binds it to the aligned data sequence and the quality score vector, providing a traceable session index for the subsequent hierarchical attention fusion network inference and model version parameter group generation stages.

[0061] In summary, the technical effects of this step are as follows: by completing the synchronous acquisition of three modalities under the same trigger time marker, and generating quality score vectors and aligned data sequences in the linked processing of unified time axis alignment, drift correction and modal quality scoring, a data foundation that can be directly called for subsequent cross-modal consistency parameter group calculation and evidence weight allocation is provided.

[0062] S300: Based on the quality score vector and aligned data sequence, extract the modal degradation feature set and cross-modal comparison feature set and perform time window aggregation, calculate the cross-modal consistency parameter set and perform evidence weight allocation, and generate the evidence weight parameter set.

[0063] Specifically, the input sources for this step are the quality score vector and aligned data sequence output from the previous main step. The aligned data sequence includes aligned voltage, current, and temperature sequences, aligned thermal image sequences, and aligned deformation sequences within the same charge / discharge session. The quality score vector includes the quality scores of the voltage, current, and temperature sequences, thermal image sequences, and deformation sequences corresponding to each time slice on the same unified time axis. This step is executed collaboratively by the feature extraction unit, time window aggregation unit, consistency calculation unit, and weight allocation unit, and the session number is used to complete the archiving index of the input data. When a missing time slice appears in the quality score vector or a continuous missing interval exists in the aligned data sequence, the time window aggregation unit registers the missing interval as an aggregation skip mark and writes it into the session log. The aggregation skip mark participates in the calculation as a weight suppression condition in the subsequent evidence weight allocation stage.

[0064] Furthermore, modal degradation feature sets are extracted based on the quality score vector and aligned data sequences, and minimum set parameter loading is completed. The modal degradation feature set refers to the set of degradation characterization elements extracted from voltage-current-temperature sequences, thermal image sequences, and deformation sequences under the same unified time axis, including capacity increment curve features, temperature rise rate features, thermal non-uniformity features, and expansion curve morphology features. The feature extraction unit reads the voltage-current-temperature sequences from the aligned data sequences, generates paired records of the voltage sequence and the capacity accumulation sequence under the current segment and sampling segment boundary conditions given by the test fixture, and performs differential sampling and local peak-valley retrieval on the paired records to obtain capacity increment curve features; wherein the capacity increment curve features include peak position interval markers, peak width interval markers, and peak value stability markers. The peak position interval markers are mapped from the unified time axis to the corresponding state of charge intervals and written into the feature field. The feature extraction unit reads the thermal image sequence from the aligned data sequence. Based on the cell surface grid range corresponding to the infrared thermal image field of view registered in the previous multimodal acquisition session parameter group, it divides each frame of thermal image into cell surface grids to generate partition temperature statistical records. It then performs differential slope estimation on the partition temperature statistical records of adjacent time slices to obtain the temperature rise rate feature. Simultaneously, it performs partition extreme value difference calculation and hot spot partition identification on the partition temperature statistical records within the same time slice to obtain thermal unevenness features. These features include hot spot partition numbers, hot spot duration segment markers, and partition temperature difference level markers. The feature extraction unit also reads the deformation sequence from the aligned data sequence. Based on the correspondence between the deformation measurement points registered in the previous multimodal acquisition session parameter group and the deformation sequence, it reconstructs the displacement sampling points of each measurement point into a measurement point displacement curve according to the measurement point number. It then performs baseline segment selection, inflection point retrieval, and rebound segment identification on the measurement point displacement curve to obtain the expansion curve morphology feature. This feature includes inflection point time slice markers, growth segment slope level markers, and rebound hysteresis segment markers. The minimum set parameters for this step include the time window length parameter, the time window step parameter, the quality score threshold, and the aggregation skip rule parameter. The time window length parameter and the time window step parameter are loaded in the time window aggregation unit, and the quality score threshold and the aggregation skip rule parameter are loaded in the weight allocation unit. The correspondence between the cell surface grid range corresponding to the infrared thermal image field of view and the deformation measurement points and deformation sequence is used as a registered session constant in the mapping operation of this step and is a basic input required for the operation of this step.

[0065] After obtaining the modal degradation feature set, cross-modal contrast feature groups are extracted and time-window aggregation is performed. The cross-modal contrast feature group refers to the set of contrast elements obtained by encoding the correspondence between thermal image sequences and deformation sequences within the same time window, at least including the correspondence between thermal heterogeneity features and expansion curve morphology features. The time-window aggregation unit generates a time-window sequence on a unified time axis according to the time-window length parameter and the time-window step parameter, and aggregates the modal degradation feature set within each time window. The aggregation methods include partition statistical merging, measurement point statistical merging, and time-slice stability merging, generating a time-window aggregated feature vector. When the quality score of any modality within a time window is lower than the quality score threshold or the aggregation skip mark hits that time window, the aggregated record of the corresponding modality in that time window is written with a low-quality segment marker. The generation of cross-modal comparison feature groups is performed by the comparison mapping unit. Within the same time window, the comparison mapping unit matches the hotspot partition number with the grid partition on the cell surface where the deformation measurement point is located, and writes the matching results into the corresponding relationship field. When multiple candidate matches exist for the hotspot partition number with the grid partition on the cell surface where the deformation measurement point is located, the comparison mapping unit introduces the growth segment slope level label and the hotspot persistence segment label within the same time window to perform consistent segment filtering, generating a unique corresponding relationship record and writing it into the cross-modal comparison feature group. Understandably, the cross-modal comparison feature group's aggregated feature vectors share the same time window number, facilitating subsequent consistency calculation units to complete the calculation of cross-modal consistency parameter groups under the same number.

[0066] Further, a cross-modal consistency parameter set is calculated and evidence weights are assigned to generate an evidence weight parameter set. The cross-modal consistency parameter set refers to a set of parameters obtained by encoding the response synchronicity, partition matching, and stable repeatability between voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on time window aggregation results. It includes partition consistency markers for thermal unevenness features and expansion curve morphology features; synchronous segment markers for temperature rise rate features and growth segment slope level markers; coupled segment markers for capacity increment curve features and hotspot persistence segment markers; and inherited records of low-quality segment markers. Within each time window, the consistency calculation unit reads the corresponding time slices of the cross-modal comparison feature set, the time window aggregated feature vector, and the quality score vector. It first performs partition consistency discrimination to obtain partition consistency markers, then performs synchronous segment discrimination to obtain synchronous segment markers, and maps the synchronous segment markers to the peak interval markers of the capacity increment curve features via time slices to obtain coupled segment markers. When low-quality segment markers exist, the consistency calculation unit writes the low-quality segment markers into the cross-modal consistency parameter set and adds a marker that does not participate in the consistency discrimination. After receiving the cross-modal consistency parameter set and the quality score vector, the weight allocation unit generates an evidence weight allocation record according to the modal dimension. The evidence weight allocation record includes the weights of the voltage, current, and temperature sequences, the weights of the thermal image sequence, and the weights of the deformation sequence. The triggering condition for weight allocation is that at least two modalities have quality scores higher than the quality score threshold within the same time window and the partition consistency is marked as valid. When the triggering condition is not met, the weight allocation unit registers the time window as a low-confidence time window and writes the three modal weights into the reduced-weight state mark. Further, the weight allocation unit performs weight merging processing on the evidence weight allocation record to ensure that the three modal weights within the same time window meet the unified dimensional constraint, and writes the merged weights into the evidence weight parameter set. The evidence weight parameter set simultaneously encapsulates the time window aggregated feature vector, the cross-modal consistency parameter set, and the evidence weight allocation record, thereby forming the input payload for subsequent hierarchical attention fusion network inference.

[0067] The output of this step is the evidence weight parameter group, which is an output field name, and retains an associated index with the session number within the segment. The evidence weight parameter group serves as the input location for the next step, which is called by the "evidence weight parameter group" of S400 and enters the processing link of the hierarchical attention fusion network inference to generate a joint representation vector. In the engineering embodiment, in the single-station testing scenario of the retired lithium battery sorting production line, the test fixture collects voltage, current and temperature sequences in the standard charge and discharge section, the infrared thermal imager continuously collects thermal image sequences within a fixed field of view, and the displacement measurement device continuously collects deformation sequences at the registered measurement points. After the session ends, this step is performed offline by the edge computing unit, which performs feature extraction, time window aggregation, consistency calculation and weight allocation, and writes the evidence weight parameter group and the session log together to the local cache, waiting for the subsequent inference unit to read.

[0068] In summary, the technical effects of this step are as follows: By forming a modal degradation feature set, a cross-modal contrast feature set, and a cross-modal consistency parameter set under a unified time axis, and by incorporating the quality score vector into the evidence weight allocation operation, a structured evidence weight parameter set is obtained, thereby supporting the input organization and session tracing of subsequent hierarchical attention fusion network inference.

[0069] S400: Based on the evidence weight parameter set, perform hierarchical attention fusion network reasoning to generate a joint representation vector;

[0070] Specifically, the input source for this step is the evidence weight parameter set output from the previous main step. This evidence weight parameter set includes at least the session number, time window number, time window aggregated feature vector, cross-modal consistency parameter set, evidence weight allocation record, and low-quality fragment marker. It is indexed and associated with the preceding multimodal acquisition session parameter set, quality score vector, and aligned data sequence through the session number. This step is collaboratively completed by the inference scheduling unit, input assembly unit, hierarchical attention fusion network execution unit, and output encapsulation unit. The hierarchical attention fusion network refers to a fusion network structure with multiple attention calculation layers within the same inference link, including an intra-time window attention layer, a cross-time window attention layer, and a cross-modal evidence gating attention layer. The inference process refers to the operation of performing forward computation on the input payload and outputting intermediate and final representations under the network parameter version identified by the model version parameter set. The inference scheduling unit first reads the model version identifier field in the evidence weight parameter group and matches it with the version list in the local model repository. When the version identifier field is missing or the match fails, the session is registered as a version rollback state and the default model version parameter group is loaded. At the same time, the version rollback state is written to the session log. The session log is archived along with the joint representation vector for subsequent incremental update parameter group generation and calling.

[0071] Furthermore, the input assembly unit performs structured decomposition and field completeness checks on the evidence weight parameter group. The field completeness checks include time window number continuity checks, evidence weight allocation record normalization constraint checks, and low-quality fragment label consistency checks. When a time window number has a breakpoint, the input assembly unit generates a missing time window placeholder and writes it to the time window sequence mapping table. When an evidence weight allocation record has an abnormal value, the input assembly unit writes a weight correction flag for the abnormal value and simultaneously writes it to the session log. Subsequently, the input assembly unit concatenates the time window aggregated feature vector with the cross-modal consistency parameter group according to the time window number to obtain the fused input tensor. Simultaneously, it converts the evidence weight allocation record into an evidence gating vector and aligns it to the time window dimension of the fused input tensor. The evidence gating vector contains the corresponding weight values ​​of voltage, current, and temperature sequence weights, thermal image sequence weights, and deformation sequence weights in each time window, and writes a suppression flag at the time window position where the low-quality fragment label is hit. Understandably, the fusion input tensor, evidence gating vector, and time window sequence mapping table together constitute the minimum input set of the hierarchical attention fusion network execution unit. The fusion input tensor carries cross-modal degradation information, the evidence gating vector carries evidence weight constraint information, and the time window sequence mapping table carries alignment information for missing and occupied parts.

[0072] In the execution unit of the hierarchical attention fusion network, attention calculation within a time window is first performed on the fusion input tensor. The attention layer within the time window treats the feature channels within the same time window as attention calculation objects, generates an attention mask based on the partition consistency marker, synchronization segment marker, and coupling segment marker in the cross-modal consistency parameter group, and performs suppression or amplification processing on the channels hit by the attention mask to obtain the representation sequence within the time window. Subsequently, cross-time window attention calculation is performed. The cross-time window attention layer inputs the representation sequence within the time window in chronological order, constructs a position code through the relative position information of the time window number, and writes it into the attention weight calculation path to form a cross-time window associated representation sequence. When there is a missing time window placeholder marker in the time window sequence mapping table, the cross-time window attention layer applies a mask to the placeholder position and writes the masking state into the intermediate state record. Next, cross-modal evidence-gated attention calculation is performed. The cross-modal evidence-gated attention layer injects the evidence gating vector into the attention score normalization process, applies gating coefficients to channels from different modalities, and thus obtains the evidence-constrained fusion representation sequence. The gating coefficients are directly mapped from the evidence weight allocation record, and the weight correction flag triggers the gating coefficient pruning process. The pruning result is written into the gating pruning record and archived with the session number. Finally, the hierarchical attention fusion network execution unit performs a global convergence operation on the evidence-constrained fusion representation sequence. The convergence operation includes convergence of the time dimension and compression of the channel dimension, outputting a joint representation vector. The joint representation vector includes a joint representation vector number, a vector dimension label, and a field associated with the session number, and retains a backtracking index field with the same time window number. The backtracking index field is generated by the time window sequence mapping table and written into the joint representation vector metadata area.

[0073] After receiving the joint representation vector, the output encapsulation unit performs output encapsulation and destination registration processing, encapsulating the joint representation vector along with the session log, version rollback status, gating pruning record, and intermediate state record into an inference output package, and writing the index pointer of the inference output package into the session index library. The joint representation vector is written as the output field name into the core payload area of ​​the inference output package, and serves as the input position for the subsequent main step's "joint representation vector" call, entering the processing chain of health status regression calculation, confidence calculation, and cause category discrimination. In the engineering embodiment, an inference server is deployed on the edge computing node of the retired lithium battery sorting production line. The inference server includes a graphics processing unit (GPU) and a model repository. After receiving the evidence weight parameter group, the hierarchical attention fusion network execution unit triggers the inference scheduling unit to load the matching model version parameter group, and registers the inference start timestamp and end timestamp in the session index library. When multiple sessions' evidence weight parameter groups arrive consecutively at a single workstation, the inference scheduling unit queues them according to the session number and performs batch processing assembly on sessions with the same model version parameter group, thereby completing the generation of joint representation vectors for multiple sessions and maintaining the consistency of the session backtracking index.

[0074] In summary, the technical effects of this step are as follows: by assembling the evidence weight parameter group into a fusion input tensor and introducing the evidence gating vector, the hierarchical attention fusion network completes multi-level attention calculation and outputs a joint representation vector. At the same time, it retains the backtracking index associated with the session number and time window number, which supports the subsequent main steps in inheriting the calculation link of the joint representation vector and tracing the version.

[0075] S500: Based on the joint representation vector, perform health status regression calculation, confidence calculation and cause category discrimination, generate health status assessment results, select treatment strategies based on health status assessment results, and generate incremental update parameter groups and model version parameter groups.

[0076] Specifically, the input source for this step is the joint representation vector output from the previous main step. Within the inference output package, the joint representation vector is associated with the session number, time window number backtracking index field, version rollback status, and session log. This step is collaboratively completed by the evaluation scheduling unit, regression calculation unit, confidence calculation unit, cause category discrimination unit, handling strategy selection unit, and incremental update encapsulation unit. Specifically, health status regression calculation refers to the numerical calculation link that maps the joint representation vector to a health status assessment quantity; confidence calculation refers to the calculation link that generates a confidence level label between the health status assessment quantity and the cause category discrimination result; and cause category discrimination refers to the classification inference link that generates the cause category discrimination result based on the joint representation vector. The evaluation scheduling unit reads the session number from the joint representation vector and retrieves the corresponding model version parameter group identifier field. When the model version parameter group identifier field does not match the local model repository, the evaluation scheduling unit writes the version rollback status and loads the default model version parameter group, while simultaneously writing the version rollback status to the session log. Subsequently, the joint representation vector along with the version rollback status is passed to the regression calculation unit, forming the minimum input set for this step.

[0077] Furthermore, the regression calculation unit performs input scale verification and missing test label verification on the joint representation vector. The input scale verification matches the vector dimension label of the joint representation vector with the input dimension registration field in the model version parameter group. When the matching fails, the regression calculation unit generates a dimension mismatch label and writes it to the session log, and sends the joint representation vector to the dimension aligner to complete the channel zero-padding or channel pruning operation. The dimension aligner outputs the dimension-aligned joint representation vector and retains the dimension alignment record. Subsequently, the regression calculation unit loads the health status regression model based on the model version parameter group and performs forward computation. The health status regression model consists of an input mapping layer, a nonlinear transformation layer, and a regression output layer. The input mapping layer receives the dimension-aligned joint representation vector and generates the regression input embedding. The nonlinear transformation layer performs multi-level transformations on the regression input embedding and outputs the regression implicit representation. The regression output layer performs numerical regression on the regression implicit representation and outputs the health status assessment quantity. The health status assessment quantity is written as the core output field name into the numerical payload area of ​​the health status assessment result and is bound to the session number, timestamp, and model version parameter group identifier fields. The dimension alignment record and regression implicit representation index field generated during the generation process of the health status assessment quantity are synchronously written into the retrospective metadata area of ​​the health status assessment result for subsequent incremental update parameter group generation.

[0078] In the confidence calculation unit, the unit reads the health status assessment, the regression implicit representation index field, and the version rollback status from the session log to construct a confidence input package. The confidence input package includes the health status assessment, the regression stability indicator field, and the data quality association field. The regression stability indicator field is obtained by comparing the regression implicit representation index field output by the regression calculation unit with the stability threshold registration field in the model version parameter group. The data quality association field is obtained by looking up the quality score vector from the session number and extracting the quality score statistics for the corresponding time window. The confidence calculation unit performs a normalized synthesis operation on the confidence input package to generate confidence scores. The normalized synthesis operation includes mapping the regression stability indicator field to a stability score, mapping the data quality association field to a quality score, and performing a weighted synthesis of the stability score and the quality score. The weighting parameters are provided by the confidence weight registration field in the model version parameter group. When the version rollback status is enabled, the confidence calculation unit writes a confidence rollback flag and completes the synthesis operation using the default confidence weight registration field. The generated confidence score is written as an output field name into the health status assessment result, and shares the session number and timestamp fields with the health status assessment quantity to form a traceable assessment payload.

[0079] In the cause category discrimination unit, the unit receives the dimension-aligned joint representation vector and loads it into the cause category discrimination model, which consists of a feature projection layer and a category scoring layer. The feature projection layer projects the joint representation vector into category discrimination features, and the category scoring layer outputs the score sequence for each cause category. To avoid introducing unregistered category names, this step generates a cause category mapping table upon first execution. The cause category mapping table contains category number and category label fields and is archived in association with the cause category version field in the model version parameter group. The cause category discrimination unit selects the category number with the highest score based on the category score sequence, combines it with the cause category mapping table, outputs the cause category discrimination result, and writes the cause category discrimination result into the health status assessment result. The cause category discrimination unit also generates a discrimination process record, which contains a category score sequence summary field and a cause category version field, and writes it into the session log for subsequent rule matching calls by the handling strategy selection unit.

[0080] The treatment strategy selection unit receives the health status assessment results and extracts the health status assessment quantity, confidence level, and cause category discrimination results. It then combines these with the treatment strategy rule set to complete the treatment strategy selection. In this step, the treatment strategy rule set is defined as a treatment strategy mapping table. This table contains mapping fields between the cause category discrimination results and retesting, isolation, and sorting / grouping strategies, as well as confidence and health status threshold fields. The treatment strategy mapping table is written into the model version parameter group and managed accordingly. The treatment strategy selection unit retrieves the treatment strategy mapping table based on the cause category discrimination results and performs a threshold comparison between the confidence level and the health status assessment quantity. If the threshold comparison is successful, a treatment strategy identifier is output. This identifier is written as an output field name into the health status assessment results and archived along with the session number. Simultaneously, the treatment strategy selection unit generates a treatment strategy selection record. This record contains the hit rule number, threshold comparison status field, and treatment strategy identifier, and is written to the session log. The treatment strategy selection record is then used as a component field of the incremental update parameter group in subsequent encapsulation processes.

[0081] Finally, the incremental update encapsulation unit aggregates the health status assessment results, session logs, and disposition strategy selection records of this step at the session number dimension, and retrieves the aligned data sequence and evidence weight parameter group associated with the session number, performing incremental update parameter group package processing. The incremental update parameter group is defined as a structured input package for model update, which at least includes an aligned data sequence index pointer, an evidence weight parameter group index pointer, a health status assessment result payload, and a disposition strategy identifier field, as well as a model version parameter group identifier field, a version rollback status field, and a time window number backtracking index field. The incremental update encapsulation unit initiates the model update process when trigger conditions are met. Trigger conditions include the cumulative number of sessions reaching the update threshold, the disposition strategy identifier hitting the retest strategy or isolation strategy, and the session log showing a dimension mismatch marker or a confidence rollback marker. After triggering, the incremental update encapsulation unit inputs the incremental update parameter group into the online sequence extreme learning machine to complete the parameter recursive update. The online sequence extreme learning machine consists of an input weight layer, a hidden node layer, and an output weight recursively recursively updates the output weights based on the incremental sample batch and outputs the model version parameter group. The generated model version parameter group includes a model version identifier field, a parameter effective timestamp field, a confidence weight registration field, and a cause category version field, and is written to the model repository and session index. The model version parameter group, as an output, is read in the next round of the process and input into the multimodal acquisition session parameter group generation link of the main step, forming a version connection closed loop across the main steps. Simultaneously, the health status assessment result generated in this step is written as an output field name into the business record library and used by subsequent retesting strategies, isolation strategies, or sorting and grouping strategies. In an engineering embodiment, an evaluation scheduling unit and an online sequence extreme learning machine are deployed on the edge server of the retired lithium battery sorting production line. The edge server receives the joint representation vector from the inference output packet and generates the health status assessment result. The disposal strategy selection unit synchronously writes the disposal strategy identifier into the workstation control interface. When the same workstation continuously hits the retesting strategy or isolation strategy, the incremental update encapsulation unit triggers a model update and generates a new model version parameter group. The new model version parameter group registers its effective timestamp in the model repository for subsequent sessions to read. The version rollback status and session logs are archived with the session index library, supporting version traceability.

[0082] Summary of the technical effects of this step: This step completes the health status regression calculation, confidence calculation and cause category discrimination based on the joint representation vector input, and writes the treatment strategy selection result into the health status assessment result; at the same time, it encapsulates the aligned data sequence, evidence weight parameter group and health status assessment result into an incremental update parameter group and outputs the model version parameter group, forming a version connection relationship with the previous main step.

[0083] Example 2: Figure 2A structural block diagram of a multimodal data fusion-based health status assessment system for retired lithium batteries according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0084] The acquisition session parameter group generation module 01 is used to acquire parameter groups for retired lithium battery cells, test fixtures, and model versions. It performs cell surface mesh generation, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter group registration, and contact force threshold registration to generate a multimodal acquisition session parameter group. Specifically, the acquisition session parameter group generation module receives the external dimensions, tab orientation, and description of the surface area to be tested from the retired lithium battery cells. It also receives the clamping posture of the test fixture, the infrared thermal imaging installation position, the deformation acquisition installation position, and the contact force loading method. Simultaneously, it receives the model version... This parameter group serves as the version constraint carrier for this session. When performing cell surface mesh generation, mesh generation rules are established according to the test surface area of ​​the retired lithium battery cell. The test surface is discretized into multiple mesh units, and the mesh unit identifier and spatial adjacency relationship are registered for each mesh unit. This yields the cell surface mesh and is written into the mesh field of the multimodal acquisition session parameter group. When performing infrared thermal imaging field of view setup, based on the infrared thermal imaging installation position and attitude parameters in the test fixture, the infrared thermal imaging field of view coverage is mapped and registered with the cell surface mesh, forming an infrared thermal imaging field of view that is consistent with the cell surface mesh. The correspondence of the core surface grid range is recorded and written into the field of view field of the multimodal acquisition session parameter group; when executing the deformation measurement point layout, based on the deformation acquisition installation position and the reachable area of ​​the measurement point in the test fixture, multiple grid cells are selected from the cell surface grid as the bearing position of the deformation measurement point, the deformation measurement point identifier and the mapping relationship between the deformation measurement point and the cell surface grid cell identifier are registered and written into the measurement point field of the multimodal acquisition session parameter group; when executing the synchronous trigger parameter group registration, the same trigger time description is established around the voltage, current and temperature sequence, thermal image sequence and deformation sequence, and each sequence is registered. The sampling frequency, sampling start and end conditions, and trigger alignment of the columns are used to form a synchronous trigger parameter group and written into the multimodal acquisition session parameter group. When performing contact force threshold registration, the contact force threshold is registered according to the contact force loading method of the test fixture and the synchronous trigger parameter group is associated and stored. The multimodal acquisition session parameter group is called as the input object of the subsequent multimodal synchronous acquisition and raw data packet generation module. At the same time, the model version parameter group maintains version association in the multimodal acquisition session parameter group for subsequent health status assessment and model version parameter group generation module to backfill and update.

[0085] The multimodal synchronous acquisition and raw data packet generation module 02 is used to synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on a multimodal acquisition session parameter set, generate multimodal raw data packets, and transmit the multimodal raw data packets to the time alignment and quality control module. Specifically, the multimodal synchronous acquisition and raw data packet generation module receives a multimodal acquisition session parameter set output from the acquisition session parameter set generation module, parses the cell surface grid, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter set, and contact force threshold registration content, and drives the voltage, current, and temperature sequences, thermal image sequences, and deformation sequences into a synchronous sampling state at the trigger time defined by the synchronous trigger parameter set. During the sampling process, the voltage, current, and temperature sequences generate sequence segments according to the sampling frequency registered in the synchronous trigger parameter set and bind the sampling time record, and the thermal image sequence is registered according to the infrared thermal imaging field of view arrangement. The field of view acquires image frames and binds them to sampling time records. The deformation sequence acquires displacement segments of measurement points according to the measurement point field registered by the deformation measurement points and binds them to sampling time records. At the same time, the contact force threshold registration is used to constrain the sampling permission conditions under the contact force loading state. When the contact force loading state does not meet the contact force threshold registration, the sampling suspension record is triggered and an abnormal segment mark is generated. The abnormal segment mark is written into the multimodal raw data packet along with the sampling time record. The multimodal raw data packet contains at least voltage, current and temperature sequences, thermal image sequences, deformation sequences and their sampling time records and abnormal segment marks, and retains the session association field with the multimodal acquisition session parameter group. Subsequently, the multimodal synchronous acquisition and raw data packet generation module transmits the multimodal raw data packet to the time alignment and quality control module as the input object for unified time axis alignment, drift correction, modal quality scoring and low-quality segment rejection.

[0086] The time alignment and quality control module 03 is used to perform unified time axis alignment, drift correction, modal quality scoring, and low-quality segment removal on the multimodal raw data packets, generating a quality score vector and an aligned data sequence, and transmitting the quality score vector and aligned data sequence to the evidence weight parameter group generation module. Specifically, the time alignment and quality control module receives multimodal raw data packets from the multimodal synchronous acquisition and raw data packet generation module, extracts sampling time records of voltage, current, and temperature sequences, thermal image sequences, and deformation sequences, constructs a unified time axis, and performs unified time axis alignment on each sequence, mapping sequence segments at different sampling frequencies to the time index of the unified time axis, thereby forming an aligned data sequence. During the drift correction process, the time alignment and quality control module calculates the clock offset and drift trend between sequences based on the sampling time records after unified time axis alignment, writes the offset and drift trend into the drift correction record, and performs drift correction on the aligned data sequence to maintain the time consistency relationship across sequences. During modal quality scoring, the time alignment and quality control module combines the abnormal segment markers in the multimodal raw data packet to perform quality scoring on the voltage, current, and temperature sequences, thermal image sequences, and deformation sequences respectively. The quality score includes a comprehensive evaluation of at least the proportion of missing markers, the proportion of noise mutations, the continuity of sampling time, and the coverage of abnormal segment markers. The quality scores of each sequence are then summarized to generate a quality score vector. During low-quality segment removal, the time alignment and quality control module performs segment-level screening based on the modal quality scoring results and low-quality segment judgment rules. The segments to be removed are removed from the aligned data sequence, and the time index record of the aligned data sequence is updated synchronously. The subsequently generated quality score vector and aligned data sequence are transmitted to the evidence weight parameter group generation module as input objects for the modal degradation feature set, cross-modal comparison feature group, cross-modal consistency parameter group, and evidence weight allocation. The drift correction record and the removal record maintain an associated field within the aligned data sequence for reference when generating the subsequent incremental update parameter group.

[0087] The evidence weight parameter group generation module 04 is used to extract modal degradation feature sets and cross-modal comparison feature groups based on the quality score vector and aligned data sequence, perform time window aggregation, calculate cross-modal consistency parameter groups and perform evidence weight allocation, generate evidence weight parameter groups, and transmit the evidence weight parameter groups to the hierarchical attention fusion network inference module. Specifically, the evidence weight parameter group generation module receives the quality score vector and aligned data sequence output from the time alignment and quality control module, and performs modal degradation feature set extraction under the time index constraint of the aligned data sequence. This extraction includes extracting capacity increment curve features from the voltage, current, and temperature sequences, temperature rise rate features and thermal unevenness features from the thermal image sequence, and expansion curve morphology features from the deformation sequence. The modal degradation feature set maintains a modal mapping relationship with the quality score vector. Subsequently, the evidence weight parameter group generation module constructs a cross-modal comparison feature group based on the aligned data sequence. The cross-modal comparison feature group at least includes the correspondence between thermal unevenness features and expansion curve morphology features, and binds the correspondence to the same time index range. During the time window aggregation process, the evidence weight parameter group... The weight parameter group generation module aggregates the modal degradation feature set and the cross-modal comparison feature set according to the time window aggregation rules, aggregating feature fragments within the time window into window-level feature fragments, and registering the mapping fields of the window start and end indices and the aligned data sequences. During the cross-modal consistency parameter group calculation, the evidence weight parameter group generation module calculates the cross-modal consistency parameter group based on the window-level feature fragments. The cross-modal consistency parameter group includes a description of the consistency degree of the correspondence within the cross-modal comparison feature group and establishes an association field with the quality score vector. During the evidence weight allocation process, the evidence weight parameter group generation module performs evidence weight allocation by combining the quality score vector and the cross-modal consistency parameter group, generating an evidence weight parameter group. The evidence weight parameter group at least registers the evidence weight of each modal feature fragment, the window index field, and the association field of the cross-modal consistency parameter group. The evidence weight parameter group is then transmitted to the hierarchical attention fusion network inference module as the input object for hierarchical attention fusion network inference, while retaining the session association field of the evidence weight parameter group and the aligned data sequences for subsequent incremental update parameter group generation.

[0088] The hierarchical attention fusion network inference module 05 is used to perform hierarchical attention fusion network inference based on the evidence weight parameter set, generate a joint representation vector, and transmit the joint representation vector to the health status assessment and model version parameter set generation module. Specifically, the hierarchical attention fusion network inference module receives the evidence weight parameter set output from the evidence weight parameter set generation module, and retrieves the associated fragments of the modality degradation feature set and cross-modality comparison feature set within the same window according to the window index field in the evidence weight parameter set. Then, during the hierarchical attention fusion network inference process, it performs weighted fusion on each modality feature fragment according to the evidence weight registered in the evidence weight parameter set. The hierarchical attention fusion network inference includes two layers: intra-modality attention fusion and cross-modality attention fusion. In the intramodal attention fusion stage, attention aggregation is performed on window-level feature fragments of the same modality according to the evidence weights to generate intramodal aggregated fragments. In the cross-modal attention fusion stage, cross-modal attention aggregation is performed on the intramodal aggregated fragments in combination with the association fields of the cross-modal consistency parameter group to generate a joint representation vector. The joint representation vector is generated in a way that maintains consistency with the session association field and window index field of the evidence weight parameter group, so that the joint representation vector has a traceable association with the quality score vector and the aligned data sequence. Subsequently, the hierarchical attention fusion network inference module transmits the joint representation vector to the health status assessment and model version parameter group generation module as the input object for health status regression calculation, confidence calculation and cause category discrimination.

[0089] The health status assessment and model version parameter group generation module 06 is used to perform health status regression calculation, confidence calculation, and cause category discrimination based on the joint representation vector, generate health status assessment results, and select treatment strategies based on the health status assessment results, generating incremental update parameter groups and model version parameter groups. Specifically, the health status assessment and model version parameter group generation module receives the joint representation vector output from the hierarchical attention fusion network inference module, and reads the corresponding evidence weight parameter group and the associated record of the aligned data sequence according to the session association field of the joint representation vector, thereby maintaining the session consistency relationship between the assessment link and the collected session parameter group; in the health status regression calculation process, the health status assessment and model version parameter group generation module performs regression mapping operation on the joint representation vector and outputs the health status regression calculation result; in the confidence calculation process, it performs confidence calculation on the weight distribution of the joint representation vector and the evidence weight parameter group and outputs the confidence calculation result; in the cause category discrimination process, it performs... The system performs category discrimination mapping and outputs the cause category discrimination result. Then, it writes the health status regression calculation result, confidence score calculation result, and cause category discrimination result into the health status assessment result and registers the correspondence between the session association field and the window index field. During the treatment strategy selection process, the health status assessment and model version parameter group generation module matches the treatment strategy selection rules based on the cause category discrimination result, forms the treatment strategy selection result, and writes it into the health status assessment result, ensuring the health status assessment result has a verifiable treatment strategy selection record. During the incremental update parameter group generation process, the health status assessment and model version parameter group generation module uses the aligned data sequence, evidence weight parameter group, and health status assessment result combination as the carrier for the incremental update parameter group and registers the session association field. It simultaneously generates the model version parameter group and writes it into the version association record. The model version parameter group provides a version constraint carrier to the acquisition session parameter group generation module, thereby closing the link backfeedback relationship from the multimodal acquisition session parameter group to the model version parameter group.

Claims

1. A method for assessing the health status of retired lithium batteries using multimodal data fusion, characterized in that, include: Acquire the parameter sets of retired lithium battery cells, test fixtures and model versions, perform cell surface mesh division, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter set registration and contact force threshold registration, and generate multimodal acquisition session parameter sets; Based on the multimodal acquisition session parameter group, voltage, current, and temperature sequences, thermal image sequences, and deformation sequences are synchronously acquired to generate multimodal raw data packets. Then, unified time axis alignment, drift correction, modal quality scoring, and low-quality segment removal are performed to generate quality score vectors and aligned data sequences. Based on the quality score vector and aligned data sequence, modal degradation feature set and cross-modal comparison feature set are extracted and time window aggregation is performed. Cross-modal consistency parameter set is calculated and evidence weight is assigned to generate evidence weight parameter set. Based on the evidence weight parameter set, hierarchical attention fusion network reasoning is performed to generate a joint representation vector; Based on the joint representation vector, health status regression calculation, confidence calculation and cause category discrimination are performed to generate health status assessment results. Then, the treatment strategy is selected based on the health status assessment results, and incremental update parameter set and model version parameter set are generated.

2. The method according to claim 1, characterized in that, The infrared thermal imaging field of view setup includes: The infrared thermal image alignment area is determined based on the cell surface grid, and the cell surface grid range corresponding to the infrared thermal image field of view is registered in the multimodal acquisition session parameter group.

3. The method according to claim 1, characterized in that, The arrangement of deformation measuring points includes: Multiple deformation measurement points are selected based on the grid on the cell surface, and the correspondence between the deformation measurement points and the deformation sequence is registered in the multimodal acquisition session parameter group.

4. The method according to claim 1, characterized in that, Synchronous trigger parameter group registration includes: The voltage, current, and temperature sequences, thermal image sequences, and deformation sequences are registered at the same trigger time, and the sampling frequency of each sequence is registered in the multimodal acquisition session parameter group.

5. The method according to claim 1, characterized in that, Modal quality scoring includes: Quality scores are calculated for voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on multimodal raw data packets, and the quality scores are then aggregated to generate a quality score vector.

6. The method according to claim 1, characterized in that, Modal degradation feature set extraction includes: Capacity increment curve features are extracted from voltage, current and temperature sequences; temperature rise rate and thermal unevenness features are extracted from thermal image sequences; and expansion curve morphology features are extracted from deformation sequences.

7. The method according to claim 1, characterized in that, Cross-modal contrast feature extraction includes: Based on the time window aggregation results of thermal image sequences and deformation sequences, a correspondence between thermal non-uniformity features and expansion curve morphology features is generated, and the correspondence is incorporated into the cross-modal comparison feature group.

8. The method according to claim 1, characterized in that, The options for handling strategies include: Based on the cause category discrimination results, a target strategy is selected from the retest strategy, isolation strategy and sorting and grouping strategy, and the target strategy is written into the health status assessment results.

9. The method according to claim 1, characterized in that, Incremental update parameter group generation includes: The aligned data sequence, health status assessment results, and evidence weight parameter set are combined as an incremental update parameter set and input into the online sequence extreme learning machine, which outputs the model version parameter set.

10. A multimodal data fusion-based health status assessment system for retired lithium batteries, applied to the method described in any one of claims 1 to 9, characterized in that, include: The acquisition session parameter group generation module is used to obtain the parameter groups of retired lithium battery cells, test fixtures and model versions, perform cell surface mesh division, infrared thermal imaging field of view arrangement, deformation measurement point arrangement, synchronous trigger parameter group registration and contact force threshold registration, and generate multimodal acquisition session parameter groups; The multimodal synchronous acquisition and raw data packet generation module is used to synchronously acquire voltage, current, and temperature sequences, thermal image sequences, and deformation sequences based on the multimodal acquisition session parameter group, and generate multimodal raw data packets; The time alignment and quality control module is used to perform unified timeline alignment, drift correction, modal quality scoring, and low-quality segment removal on multimodal raw data packets, generating quality score vectors and aligned data sequences; The evidence weight parameter group generation module is used to extract modal degradation feature sets and cross-modal contrast feature groups based on quality score vectors and aligned data sequences, perform time window aggregation, calculate cross-modal consistency parameter groups, perform evidence weight allocation, and generate evidence weight parameter groups. The hierarchical attention fusion network inference module is used to perform hierarchical attention fusion network inference based on evidence weight parameter groups and generate joint representation vectors. The health status assessment and model version parameter group generation module is used to perform health status regression calculation, confidence calculation and cause category discrimination based on the joint representation vector, and generate health status assessment results.

Images