Deep visual inference method for battery binder performance correlation features

Abstract

The application discloses a deep visual inference method for battery binder performance correlation characteristics, and particularly relates to the technical field of battery detection, and the method comprises the following steps: applying a hierarchical micro-deformation load to an electrode sheet and collecting images, obtaining local peeling strength and failure proportion of each section through a peeling test; constructing a particle-bridge bearing diagram according to an image segmentation result, and extracting dynamic response characteristics of the binder under a stress state; establishing a deep visual inference model comprising an image coding branch and a bearing diagram coding branch, and training the model with a local performance label as supervision; collecting hierarchical deformation images of a to-be-detected electrode and inputting the images into the model, and outputting a local peeling strength distribution diagram and a weak area position; and performing a reliability check on the inference result, and triggering a real measurement fallback and model incremental update if necessary. The application can infer the local distribution of the electrode sheet binder performance without destructive testing, and provides a basis for binder quality evaluation in the battery manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and more specifically, to a deep visual inference method for the correlation characteristics of battery binder performance. Background Technology

[0002] Binders are key auxiliary materials in the manufacture of lithium-ion battery electrodes. Their distribution and interfacial bonding performance directly affect the cycle life, rate performance, and safety of the electrodes. Currently, the evaluation of binder performance mainly relies on destructive methods such as peel tests. These methods are not only time-consuming and resource-intensive, but also cannot be implemented for online testing during continuous electrode production. More importantly, traditional peel tests only provide macroscopic average peel strength, which is insufficient to reflect the non-uniform distribution of the binder at the microscale and its progressive damage process under stress. This leads to binder formulation optimization and process anomaly troubleshooting often relying on trial and error, resulting in low efficiency.

[0003] In recent years, some researchers have attempted to use scanning electron microscopy or optical microscopy images to analyze the microstructure of electrodes through manual observation or traditional image processing techniques, thereby inferring the state of the binder. However, these methods have significant limitations: on the one hand, conventional microscopic images cannot directly distinguish the binder phase, and can only rely on indirect characteristics of carbon-binder-rich regions for judgment, resulting in insufficient accuracy; on the other hand, most existing methods are based on static image analysis, ignoring the dynamic response characteristics of the binder under stress, which is precisely the load-bearing response behavior that determines the actual bonding effect of the binder. In addition, non-binder factors such as the particle morphology and pore structure of the electrode active material significantly interfere with visual characteristics, and existing methods cannot effectively eliminate these spurious factors, leading to a weak correlation between the inferred results and the actual binder performance. Summary of the Invention

[0004] To overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide a deep visual inference method for battery binder performance-related features.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep visual inference method for battery binder performance-related features includes the following steps: Establish graded micro-deformation visual data and local performance labels for the same area: acquire battery electrode sheets after coating, drying and rolling, cut visual inference samples from the electrode sheets, set positioning marks and calibration plates in the observation area of ​​the samples, and acquire unloaded reference images; apply multiple graded micro-deformation loading levels to the samples, and acquire images at each loading level; perform peel tests on the samples after visual acquisition, divide the observation area into multiple local segments along the peel direction, determine the local peel strength of each segment according to the peel force curve, and acquire images after peeling to generate the interface failure ratio and cohesive failure ratio of each segment; Forming binder load response characteristics and constructing particle-bridging load map: Based on the collected reference image and images of each loading level, segmentation is performed to obtain active material particles or particle clusters, pores, visible cracks, and candidate binder load areas; pixel displacement field and local strain field are calculated to establish particle-bridging load map, which includes active material particle nodes, particle cluster nodes, boundary nodes, and inter-particle edges and interface edges; Training a deep visual inference model: Establish a deep visual inference model that includes a multi-state image coding branch, a carrier map coding branch, a local performance output branch, and a confidence output branch. Use the local peeling intensity, interface failure ratio, and cohesion failure ratio of each segment as supervision labels, and train the model with a weighted loss function. Inferring the adhesive performance correlation features of the electrode under test: Acquire images of the electrode under test in the unloaded state and multiple graded micro-deformation states under the same imaging conditions. After registration and segmentation, establish the particle-bridging load map of the area under test. Input the trained model and output the local performance distribution map and the overall performance vector. Verify the credibility of the inference and perform experimental backtracking and model update: Perform multi-level credibility checks on the inference results, select to perform experimental backtracking based on the check results, add the backtracked samples to the incremental dataset, and retrain the model after the update conditions are met.

[0006] Specifically, the steps for establishing hierarchical micro-deformation visual data and local performance labels for the same region include: Visual inference samples and adjacent calibration samples are cut from the electrode sheet, and the electrode type, active material type, binder type and mass ratio, conductive agent ratio, areal density, compaction density, coating thickness, drying conditions and rolling conditions are recorded. Set up no fewer than three non-collinear positioning marks around the observation area of ​​the visual inference sample, and place a calibration sheet with known size and grayscale. Use the same imaging device, objective lens, exposure time, illumination angle and polarization state to acquire unloaded reference images. At least one coaxial light image and at least one oblique incident light image are acquired. The coaxial light image is used to extract particle boundaries, pores and surface cracks, while the oblique incident light image is used to enhance micro-protrusions, particle bridging regions and local peeling precursors.

[0007] Specifically, the steps for applying multiple levels of graded micro-deformation loading include: The specimen was mounted on a micro-bending loading stage, and the direction of the current collector side toward the inside or outside of the bend was kept consistent in the same set of data. Using a representative calibration specimen of the same type as the test specimen, the curvature is gradually increased in independent tests, and the initial curvature at the first appearance of identifiable microcracks or interface slip is recorded as a benchmark for calibrating the microdeformation response range of this type of specimen. Based on the reference curvature and combined with engineering experience, three fixed, progressively increasing absolute strain levels are pre-set as the first, second, and third loading levels for all test samples in this batch. The loading levels are lower than the curvature that causes macroscopic delamination of the entire coating. After each loading level is reached and a predetermined stable time is maintained, images are acquired without changing the field of view and lighting conditions; If large-area delamination occurs at the first loading level, the sample is marked as a poor sample and is not used for normal training. The loading level setting value for subsequent samples of the same formulation is also reduced.

[0008] Specifically, the steps for conducting peel tests and determining the local peel strength of each section include: A peeling test with constant peeling angle and constant peeling speed was performed on the sample after visual acquisition. The change of peeling force with the displacement of the peeling front was recorded. The time and position of the peeling front passing through each visual image partition were determined by the displacement encoder of the testing machine or the synchronous camera. The observation area is divided into multiple segments along the stripping direction, with each segment having a length of Δx. m The width is b, and the local peel strength of the m-th segment is determined by the integral of the peel force curve. The initial clamping segment, the final release segment, and the segment with obvious equipment impact in the peel force curve do not participate in local label generation.

[0009] Specifically, the steps for generating the interface failure ratio and cohesion failure ratio include: Images of the current collector surface after peeling and the fracture surface of the peeled coating were acquired separately. The fracture image was registered to the observation area before stripping using positioning markers, stripping direction, and cumulative displacement of the section. For each segment, the exposed current collector area, the current collector residual coating area, and the coating internal fracture area are divided to generate the interface failure ratio and the cohesive failure ratio. The local peel strength is used to represent the overall ability of the section to resist peeling, and the interface failure ratio and cohesive failure ratio are used to distinguish whether the performance deficiency mainly occurs at the interface between the active material layer and the current collector or occurs inside the active material layer.

[0010] Specifically, the steps for forming binder load response characteristics and constructing particle-bridged load maps include: The baseline image and images at each loading level are input into the phase and structure segmentation network to obtain active material particles or particle clusters, pores, visible cracks, and candidate carbon-binder enrichment regions. For images where the binder phase cannot be directly distinguished, regions with small distances between adjacent particles, continuous grayscale or texture, and exhibiting cooperative displacement under low-level loading are identified as candidate binder-bearing regions. Calculate the pixel displacement field and local strain field based on the registered multi-state image. For any two adjacent particles or particle clusters, calculate their relative displacement of centroid, change of the distance between opposing boundaries, change of continuity of candidate bridging zone, change of crack probability, and irreversible displacement increment between adjacent loading levels. Establish a particle-bridged load diagram. The nodes in the load diagram include active material particle nodes, particle cluster nodes, and boundary nodes used to represent the influence of the current collector interface. When there is a candidate binder load area between two particles, an inter-particle edge is established. When there is a continuous coating connection between the particle area near the current collector side and the boundary node, an interface edge is established. The node features include at least the particle area, equivalent particle size, shape, local porosity, position from the current collector direction, and displacement under each loading level. The edge features include at least the reference spacing, texture continuity, spacing increment under graded loading, bridging continuity decrease, crack precursor growth, and residual slip.

[0011] Specifically, the steps for training a deep vision inference model include: The multi-state image coding branch uses a weighted convolutional encoder to extract features from the baseline image and images at each loading level, and calculates the difference features of the same region between adjacent loading levels. The load-transfer branch uses a graph attention network to aggregate the load transfer information of granular nodes and candidate bridge edges; The outputs of the two branches are fused at the image segment level; The model was trained by weighting the loss of local peeling intensity regression, failure mode classification, global consistency, loading level order, and morphology matching comparison. The training set, validation set, and test set are divided according to the electrode manufacturing batch. Adjacent samples from the same roll or the same formula batch shall not be included in the training set and test set at the same time. For models that need to be used across active material systems, domain identifiers are set according to the material system, and independent batches that were not used in training are retained in the test set.

[0012] Specifically, the steps for inferring the adhesive performance correlation characteristics of the electrode under test include: Positioning marks were set on the electrode sheet to be tested according to the same imaging conditions, and micro-deformation state images were acquired sequentially under unloaded state and multiple preset fixed absolute strain levels. Perform grayscale and scale correction, rigid registration and non-rigid registration on the image to be tested, calculate the registration error, image signal-to-noise ratio, overexposure ratio and effective field of view ratio, and re-acquire the image if any image quality index exceeds the training calibration range. The particles, pores, cracks and candidate load-bearing areas are segmented, the displacement and structural changes under each loading level are calculated, the particle-bridging load-bearing map of the area to be tested is established, and the multi-state image and load-bearing map are input into the trained deep vision inference model. The model outputs a local peel strength distribution map, an interface failure probability map, a cohesive failure probability map, a weak area contribution map, and an overall performance vector. The overall performance vector includes at least the segment-weighted average of the local peel strength, the proportion of weak areas below the calibration qualification boundary, the proportion of high-risk areas for interface failure, the proportion of high-risk areas for cohesive failure, and the inference uncertainty. The weak zone contribution map is mapped to the physical coordinates of the electrode sheet, and the weak zone location, length, area, main failure type and corresponding load-bearing map anomaly cause are output. The anomaly cause includes rapid decline in particle bridging continuity, excessive local relative slip, crack precursor concentration or excessive cumulative irreversible displacement increment under high-level loading.

[0013] Specifically, the steps for verifying the credibility of the inference and performing experimental backtracking and model updates include: A four-level credibility check is established for the inference results: the first level checks image quality and registration error; the second level checks whether the damage response under graded loading satisfies sequential consistency; the third level checks the distance between the fusion features of the test sample and the feature distribution of the training sample; and the fourth level checks whether the prediction of local weak regions is supported by both image difference features and load map anomalies. Calculate the out-of-domain distance of the fused features of the sample to be tested relative to the training feature domain. When the out-of-domain distance exceeds a threshold, the model only outputs the locations of weak regions that need to be verified by actual testing. When any of the inference uncertainty, registration error, out-of-domain distance or physical consistency check exceeds the corresponding threshold, a sample is cut from the area to be tested or its adjacent location to perform a peel test. When the predicted weak area is small but the risk is high, the peel path is preferentially made to pass through the predicted weak area. When the model passes all confidence checks, the local performance map and the overall performance vector are used as the inference results of the adhesive performance correlation. When the overall average value meets the qualification requirements but the proportion of weak area exceeds the weak area limit, the electrode sheet is determined to require local review or process traceability. The multi-state images, load-bearing maps, local peel strength, and fracture failure modes of the measured backtracking samples are added to the incremental dataset. The model is retrained after the new samples meet the update conditions. The new model will only replace the original model if it simultaneously meets the following conditions on the new batch and the historical representative batch: the false negative rate of low-performance segments has not increased, the cross-batch error has not worsened, and the out-of-domain recognition ability has not decreased.

[0014] The technical effects and advantages of this invention are as follows: This invention, through a graded micro-deformation loading and image registration process, captures the progressive damage response of the binder under stress without causing macroscopic delamination of the electrode sheet, thus establishing a direct correlation between the visual inference results and the actual load-bearing capacity of the binder. Furthermore, by simultaneously acquiring visual images and local peel strength labels in the same observation area, interference caused by differences in microstructure between samples on the correlation modeling is eliminated. Combined with a structured description of the binder load transfer path using a particle-bridging load map, the model simultaneously considers static morphological features and dynamic mechanical response features, effectively avoiding spurious correlations arising from relying solely on static texture features, and significantly improving the consistency between the inferred results and the actual binder performance.

[0015] This invention also achieves a complete closed loop from data acquisition to model inference and credibility verification. By employing batch-based independent partitioning of the training and test sets, setting up an out-of-domain distance judgment mechanism, and implementing multi-level credibility checks, the generalization risk of the model across different batches and material systems is effectively controlled. When the inference results are unreliable, experimental backtracking and incremental updates can be automatically triggered, avoiding blind output of the model on out-of-domain samples and ensuring safety and traceability in engineering applications. This method can significantly reduce the workload of destructive sampling and testing while outputting local distribution maps of binder performance and information on abnormal locations, providing a direct reference for online quality control and root cause tracing of anomalies in electrode coating and rolling processes. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the steps of the deep visual inference method for battery binder performance-related features are as follows: S1. Establish hierarchical micro-deformation visual data and local performance labels for the same region: S11: Obtain the electrode sheet after coating, drying, and rolling. Cut a visual inference sample and an adjacent calibration sample from the electrode sheet, and record the electrode type, type of active material, type and mass ratio of binder, proportion of conductive agent, areal density, compaction density, coating thickness, drying conditions, and rolling conditions. The above process data is used for training set stratification and out-of-domain judgment, and does not directly replace visual features for performance inference.

[0019] S12, set no fewer than three non-collinear positioning marks around the observation area of ​​the visually inferred sample, and place a calibration sheet with known size and grayscale; acquire an unloaded reference image using the same imaging device, objective lens, exposure time, illumination angle, and polarization state. Coaxial light images and at least one oblique incident light image can be acquired separately, wherein the coaxial light image is used to extract particle boundaries, pores, and surface cracks, and the oblique incident light image is used to enhance micro-protrusions, particle bridging regions, and precursors of localized abrasion.

[0020] S13. Mount the specimen on a micro-bending loading stage, ensuring the current collector side faces the inside or outside of the bend consistently within the same data set. Using a representative calibration specimen of the same type as the test specimen, gradually increase the curvature during independent testing, recording the initial curvature at which identifiable microcracks or interface slip first appear. This initial curvature serves as a benchmark for calibrating the micro-deformation response range of this type of specimen. Based on this benchmark curvature and combined with engineering experience, pre-set three fixed, progressively increasing absolute strain levels as the first, second, and third loading levels for all test specimens in this batch. The loading level must be lower than the curvature that causes macroscopic delamination of the entire coating, and adjacent loading levels should increase progressively. After reaching each loading level and maintaining a predetermined stabilization time, acquire images without changing the field of view and illumination conditions. If a large-area delamination occurs in the test specimen at the first loading level, the specimen is marked as a poor sample and not used for normal training; the loading level setting value for subsequent specimens of the same formulation should be reduced.

[0021] S14: Rigid registration is performed using the positioning markers, followed by local non-rigid registration using particle boundaries and pore structures to obtain the pixel correspondence between each loaded state image and the reference image. The registration error is determined by the reprojection distance between the center of the positioning marker and the corner point of the stable particle; the registration error threshold is taken as the upper quantile of the error of the correctly registered sample in the calibration image. If the threshold is exceeded, re-acquisition or re-registration is performed; images that still fail to meet the requirements are not included in the training.

[0022] S15, Perform a peeling test at a constant peeling angle and constant peeling speed on the sample after visual acquisition, and record the peeling force. Displacement with the peeling front The changes in the data were observed, and the time and location of the peeling front passing through each visual image partition were determined using a displacement encoder on the testing machine or a synchronous camera. The observation area was divided into multiple sections of length [missing information]. Width is the width of the sample. The local section, the first Local peel strength of each section Determined according to the following formula: ,in, Indicates the first Local peel strength of individual sections; Indicates the starting stripping location of this section; Indicates the length of the section along the stripping direction; This indicates the sample width. The initial clamping section, the final release section, and the section showing significant equipment impact in the peel force curve do not participate in local label generation.

[0023] S16: Images of the current collector surface after peeling and the fracture surface of the peeled coating are acquired respectively. The fracture surface images are registered to the observation area before peeling using positioning marks, peeling direction, and cumulative displacement of the segment. For each segment, the exposed current collector area, the current collector residual coating area, and the internal fracture area of ​​the coating are divided, generating the interface failure ratio and the cohesive failure ratio. The local peel strength is used to represent the overall ability of the segment to resist peeling, and the failure ratio is used to distinguish whether the performance deficiency mainly occurs at the active material layer-current collector interface or occurs inside the active material layer. The two are not simply multiplied to replace the actual physical quantity.

[0024] S2. Forming binder load response characteristics and constructing particle-bridging load diagram: S21, the baseline image and images at each loading level are input into the phase and structure segmentation network to obtain active material particles or particle clusters, pores, visible cracks, and candidate carbon-binder enrichment regions. For images where the binder phase cannot be directly resolved, the true binder boundary is not forced to be output. Instead, regions with small distances between adjacent particles, continuous grayscale or texture, and exhibiting cooperative displacement under low-level loading are identified as candidate binder-bearing regions. The labels required to train this segmentation network can be generated from high-resolution microscopic images, elemental surface scanning, correlated microscopic imaging, or manual verification results.

[0025] S22, calculate the pixel displacement field and local strain field based on the registered multi-state image; for any two adjacent particles or particle clusters, calculate their relative centroid displacement, change in the distance between opposing boundaries, change in the continuity of the candidate bridging zone, change in crack probability, and irreversible displacement increment between adjacent loading levels. All of these changes are referenced to the baseline state of the same region; therefore, particle size, static porosity, and fixed illumination texture are retained as background conditions during the difference process and do not individually determine the binder performance results.

[0026] S23, Establish a particle-bridging load diagram. The nodes in the diagram include active material particle nodes, particle cluster nodes, and boundary nodes representing the influence of the current collector interface. Inter-particle edges are established when a candidate binder load zone exists between two particles; interface edges are established when there is a continuous coating connection between the particle region near the current collector side and the boundary node. Node features include at least particle area, equivalent particle size, shape, local porosity, position relative to the current collector direction, and displacement under each loading level. Edge features include at least the reference spacing, texture continuity, spacing increment under graded loading, bridging continuity decrease, crack precursor growth, and residual slip.

[0027] S24, correlate the bearing sub-image corresponding to each image segment with the local peel strength, interface failure ratio, and cohesive failure ratio obtained in step S1. To reduce spurious correlations caused by the particle morphology of the active material, calculate the multidimensional morphology feature vector (including particle size distribution width, average porosity, local coating thickness, and compaction density) for each segment on the same electrode, and use standardized Euclidean distance to measure the morphological differences between segments. Use the lower quantile value (e.g., 20th percentile value) of the distance between all normal segments in the calibration set as the matching threshold, and pair segments with distances less than this threshold to form morphology matching control pairs; in the morphology matching control pairs, segments with faster bridging continuity decline, larger residual slip, or more crack precursors should not be labeled as having higher local adhesion performance.

[0028] S3. Training a deep visual inference model constrained by region labels and physical consistency: S31, Establish a deep vision inference model. The model includes a multi-state image coding branch, a load map coding branch, a local performance output branch, and a confidence output branch. The multi-state image coding branch uses a convolutional encoder with shared weights to extract features from the baseline image and images at each loading level, and calculates the difference features of the same region between adjacent loading levels; the load map coding branch uses a graph attention network to aggregate the load transfer information of particle nodes and candidate bridge edges; the outputs of the two branches are fused at the image segment level.

[0029] S32: For each local segment, the model outputs predicted local peel strength, interface failure probability, cohesive failure probability, adhesive load-bearing correlation feature vector, and pixel-level or graph-edge-level weak area contribution map; for the entire sample, it outputs average peel resistance, weak area proportion, and inference uncertainty. The adhesive load-bearing correlation feature vector is not a pre-set single texture parameter, but an intermediate representation trained by graded load response, load map connection state, and local performance labels.

[0030] S33, the model is trained using a weighted sum of local peeling intensity regression loss, failure mode classification loss, global consistency loss, loading level order loss, and morphology matching comparison loss: ,in, Used to limit the predicted local peel strength of each section from being close to the actual measured value. ; This is used to limit the probability of interface failure and cohesive failure to be close to the fracture segmentation result; This is used to limit the predicted values ​​of each segment, after being weighted by the segment length, from being close to the average peel resistance of the entire sample. Used to limit the damage response, such as crack precursor, bridging continuity loss and residual slip, from exhibiting an unfounded inverse decrease with increasing loading level; This is used to prevent sections with similar shapes but poor dynamic load response from being predicted as having better performance. to The non-negative loss weights are determined through a validation set combination search and remain unchanged in the same model version.

[0031] S34. The training set, validation set, and test set are divided according to electrode manufacturing batches. Adjacent samples from the same roll or the same formulation batch must not be included in the training set and test set simultaneously to prevent the model from obtaining artificially high accuracy through batch texture memory. For models that need to be used across active material systems, domain identifiers are set according to the material system, and independent batches that did not participate in the training are retained in the test set.

[0032] S35, the model must pass at least the following conditions: the local peel strength prediction error meets the calibration requirements; the recall rate of low-performance sections meets the preset requirements; the confusion between interface failure and cohesion failure does not exceed the allowable range; there is no significant decrease in cross-batch test results; and the weak area contribution map has a spatial correspondence with the actual peel trough sections. The above thresholds are determined through an independent calibration set, with controlling the false negative rate of low-performance regions as the primary objective.

[0033] S4. Inferring the correlation characteristics of binder properties of the electrode under test: S41, set positioning marks for the electrode sheet to be tested according to the imaging conditions in step S1, and sequentially acquire images of the unloaded state and the micro-deformation state under multiple fixed absolute strain levels pre-set for the electrode type in step S13; the loading direction, these fixed strain level values, holding time, illumination, imaging magnification and pixel size should be the same as those during model training, or converted to the same scale after calibration.

[0034] S42 performs grayscale and scale correction, rigid registration and non-rigid registration on the image under test, and calculates the registration error, image signal-to-noise ratio, overexposure ratio and effective field of view ratio. When any image quality index exceeds the training calibration range, the performance result is not directly output, but the image is reacquired.

[0035] S43, segment particles, pores, cracks and candidate load-bearing areas, calculate displacement and structural changes under each loading level, establish particle-bridging load-bearing map of the area to be tested, and input the multi-state image and load-bearing map into the trained deep vision inference model.

[0036] S44, the model outputs a local peel strength distribution map, an interface failure probability map, a cohesive failure probability map, a weak area contribution map, and an overall performance vector. The overall performance vector includes at least the segment-weighted average of the local peel strength, the proportion of weak areas below the calibration acceptance boundary, the proportion of high-risk areas for interface failure, the proportion of high-risk areas for cohesive failure, and the inference uncertainty. The acceptance boundary is determined jointly based on the local peel strength distribution of qualified calibration samples of the same electrode type and engineering acceptance requirements.

[0037] S45 maps the weak zone contribution map to the physical coordinates of the electrode sheet, outputting the location, length, area, main failure types, and corresponding causes of load-bearing map anomalies in the weak zone. These anomalies include a rapid decrease in particle bridging continuity, excessive local relative slip, crack precursor concentration, or excessive cumulative irreversible displacement increments under high-level loading. Therefore, the inference results not only provide an overall numerical value but also the spatial location that can be used for sampling verification and process traceability.

[0038] S5. Verify the confidence of the inference and perform experimental backtracking and model updates: S51 establishes a four-level credibility check for the inference results. The first level checks image quality and registration error; the second level checks whether the damage response under graded loading satisfies sequential consistency; the third level checks the distance between the fusion features of the test sample and the feature distribution of the training sample; and the fourth level checks whether the prediction of local weak regions is simultaneously supported by image difference features and load map anomalies.

[0039] S52, the fusion features of the sample to be tested are denoted as The mean and covariance of features of similar samples in the training and calibration sets are denoted as follows: and Calculate the distance outside the domain: ,in, This represents the distance of the sample to be tested relative to the training feature domain. When irreversibility is not possible, the covariance matrix after adding a stabilizing term or the covariance in the principal component space can be used. The out-of-domain threshold is taken from normal samples in an independent calibration set. The upper quantile value. When the threshold is exceeded, the model only outputs "requires actual verification" and the location of the weak area, and does not output performance values ​​that can be used as the basis for release.

[0040] S53: When any of the inference uncertainty, registration error, out-of-domain distance, or physical consistency check exceeds the corresponding threshold, a sample is cut from the test area or its adjacent location to perform a peel test; when the predicted weak area is small but the risk is high, the peel path is preferentially made to pass through the predicted weak area. The measured results are used to determine whether the model has a false alarm, a missed alarm, or only an overall calibration deviation.

[0041] S54. When the model passes all confidence checks, the local performance map and the overall performance vector are used as the results of the adhesive performance correlation inference. When the overall average value meets the qualification requirements but the proportion of weak area exceeds the weak area limit, the local risk is not covered by the overall average value. The electrode sheet should be judged as needing local review or process traceability.

[0042] S55: Add the multi-state images, load-bearing maps, local peel strength, and fracture failure modes of the measured backtracking samples to the incremental dataset; retrain the model after the new samples meet the update conditions. The new model will only replace the original model if it simultaneously meets the following conditions on both the new batch and the historical representative batch: the false negative rate of low-performance segments has not increased, the cross-batch error has not worsened, and the out-of-domain recognition ability has not decreased; otherwise, the original model will be retained.

[0043] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0044] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0045] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0046] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0047] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0048] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0049] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0050] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep visual inference method for the correlation characteristics of battery binder performance, characterized in that, Includes the following steps: Establish graded micro-deformation visual data and local performance labels for the same area: acquire battery electrode sheets after coating, drying and rolling, cut visual inference samples from the electrode sheets, set positioning marks and calibration plates in the observation area of ​​the samples, and acquire unloaded reference images; apply multiple graded micro-deformation loading levels to the samples, and acquire images at each loading level; perform peel tests on the samples after visual acquisition, divide the observation area into multiple local segments along the peel direction, determine the local peel strength of each segment according to the peel force curve, and acquire images after peeling to generate the interface failure ratio and cohesive failure ratio of each segment; Forming binder load response characteristics and constructing particle-bridged load maps: Based on the collected baseline images and images of each loading level, segments are performed to obtain active material particles or particle clusters, pores, visible cracks, and candidate binder load zones; Calculate the pixel displacement field and local strain field, and establish a particle-bridging load map. The load map includes active material particle nodes, particle cluster nodes, boundary nodes, as well as inter-particle edges and interface edges. Training a deep visual inference model: Establish a deep visual inference model that includes a multi-state image coding branch, a carrier map coding branch, a local performance output branch, and a confidence output branch. Use the local peeling intensity, interface failure ratio, and cohesion failure ratio of each segment as supervision labels, and train the model with a weighted loss function. Inferring the adhesive performance correlation features of the electrode under test: Acquire images of the electrode under test in the unloaded state and multiple graded micro-deformation states under the same imaging conditions. After registration and segmentation, establish the particle-bridging load map of the area under test. Input the trained model and output the local performance distribution map and the overall performance vector. Verify the credibility of the inference and perform experimental backtracking and model update: Perform multi-level credibility checks on the inference results, select to perform experimental backtracking based on the check results, add the backtracked samples to the incremental dataset, and retrain the model after the update conditions are met.

2. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for establishing hierarchical micro-deformation visual data and local performance labels within the same region include: Visual inference samples and adjacent calibration samples are cut from the electrode sheet, and the electrode type, active material type, binder type and mass ratio, conductive agent ratio, areal density, compaction density, coating thickness, drying conditions and rolling conditions are recorded. Set up no fewer than three non-collinear positioning marks around the observation area of ​​the visual inference sample, and place a calibration sheet with known size and grayscale. Use the same imaging device, objective lens, exposure time, illumination angle and polarization state to acquire unloaded reference images. At least one coaxial light image and at least one oblique incident light image are acquired. The coaxial light image is used to extract particle boundaries, pores and surface cracks, while the oblique incident light image is used to enhance micro-protrusions, particle bridging regions and local peeling precursors.

3. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for applying multiple levels of graded micro-deformation loading include: The specimen was mounted on a micro-bending loading stage, and the direction of the current collector side toward the inside or outside of the bend was kept consistent in the same set of data. Using a representative calibration specimen of the same type as the test specimen, the curvature is gradually increased in independent tests, and the initial curvature at the first appearance of identifiable microcracks or interface slip is recorded as a benchmark for calibrating the microdeformation response range of this type of specimen. Based on the reference curvature and combined with engineering experience, three fixed, progressively increasing absolute strain levels are pre-set as the first, second, and third loading levels for all test samples in this batch. The loading levels are lower than the curvature that causes macroscopic delamination of the entire coating. After each loading level is reached and a predetermined stable time is maintained, images are acquired without changing the field of view and lighting conditions; If large-area delamination occurs at the first loading level, the sample is marked as a poor sample and is not used for normal training. The loading level setting value for subsequent samples of the same formulation is also reduced.

4. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for conducting peel tests and determining the local peel strength of each section include: A peeling test with constant peeling angle and constant peeling speed was performed on the sample after visual acquisition. The change of peeling force with the displacement of the peeling front was recorded. The time and position of the peeling front passing through each visual image partition were determined by the displacement encoder of the testing machine or the synchronous camera. The observation area is divided into multiple segments along the stripping direction, with each segment having a length of Δx. m The width is b, and the local peel strength of the m-th segment is determined by the integral of the peel force curve. The initial clamping segment, the final release segment, and the segment with obvious equipment impact in the peel force curve do not participate in local label generation.

5. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for generating the interface failure ratio and cohesion failure ratio include: Images of the current collector surface after peeling and the fracture surface of the peeled coating were acquired separately. The fracture image was registered to the observation area before stripping using positioning markers, stripping direction, and cumulative displacement of the section. For each segment, the exposed current collector area, the current collector residual coating area, and the coating internal fracture area are divided to generate the interface failure ratio and the cohesive failure ratio. The local peel strength is used to represent the overall ability of the section to resist peeling, and the interface failure ratio and cohesive failure ratio are used to distinguish whether the performance deficiency mainly occurs at the interface between the active material layer and the current collector or occurs inside the active material layer.

6. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for forming binder load response characteristics and constructing particle-bridging load maps include: The baseline image and images at each loading level are input into the phase and structure segmentation network to obtain active material particles or particle clusters, pores, visible cracks, and candidate carbon-binder enrichment regions. For images where the binder phase cannot be directly distinguished, regions with small distances between adjacent particles, continuous grayscale or texture, and exhibiting cooperative displacement under low-level loading are identified as candidate binder-bearing regions. Calculate the pixel displacement field and local strain field based on the registered multi-state image. For any two adjacent particles or particle clusters, calculate their relative displacement of centroid, change of the distance between opposing boundaries, change of continuity of candidate bridging zone, change of crack probability, and irreversible displacement increment between adjacent loading levels. Establish a particle-bridged load diagram. The nodes in the load diagram include active material particle nodes, particle cluster nodes, and boundary nodes used to represent the influence of the current collector interface. When there is a candidate binder load area between two particles, an inter-particle edge is established. When there is a continuous coating connection between the particle area near the current collector side and the boundary node, an interface edge is established. The node features include at least the particle area, equivalent particle size, shape, local porosity, position from the current collector direction, and displacement under each loading level. The edge features include at least the reference spacing, texture continuity, spacing increment under graded loading, bridging continuity decrease, crack precursor growth, and residual slip.

7. The deep visual inference method for battery binder performance correlation characteristics according to claim 1, characterized in that, The specific steps for training a deep vision inference model include: The multi-state image coding branch uses a weighted convolutional encoder to extract features from the baseline image and images at each loading level, and calculates the difference features of the same region between adjacent loading levels. The load-transfer branch uses a graph attention network to aggregate the load transfer information of granular nodes and candidate bridge edges; The outputs of the two branches are fused at the image segment level; The model was trained by weighting the loss of local peeling intensity regression, failure mode classification, global consistency, loading level order, and morphology matching comparison. The training set, validation set, and test set are divided according to the electrode manufacturing batch. Adjacent samples from the same roll or the same formula batch shall not be included in the training set and test set at the same time. For models that need to be used across active material systems, domain identifiers are set according to the material system, and independent batches that were not used in training are retained in the test set.

8. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for inferring the adhesive performance correlation characteristics of the electrode under test include: Positioning marks were set on the electrode sheet to be tested according to the same imaging conditions, and micro-deformation state images were acquired sequentially under unloaded state and multiple preset fixed absolute strain levels. Perform grayscale and scale correction, rigid registration and non-rigid registration on the image to be tested, calculate the registration error, image signal-to-noise ratio, overexposure ratio and effective field of view ratio, and re-acquire the image if any image quality index exceeds the training calibration range. The particles, pores, cracks and candidate load-bearing areas are segmented, the displacement and structural changes under each loading level are calculated, the particle-bridging load-bearing map of the area to be tested is established, and the multi-state image and load-bearing map are input into the trained deep vision inference model. The model outputs a local peel strength distribution map, an interface failure probability map, a cohesive failure probability map, a weak area contribution map, and an overall performance vector. The overall performance vector includes at least the segment-weighted average of the local peel strength, the proportion of weak areas below the calibration qualification boundary, the proportion of high-risk areas for interface failure, the proportion of high-risk areas for cohesive failure, and the inference uncertainty. The weak zone contribution map is mapped to the physical coordinates of the electrode sheet, and the weak zone location, length, area, main failure type and corresponding load-bearing map anomaly cause are output. The anomaly cause includes rapid decline in particle bridging continuity, excessive local relative slip, crack precursor concentration or excessive cumulative irreversible displacement increment under high-level loading.

9. The deep visual inference method for battery binder performance correlation features according to claim 1, characterized in that, The specific steps for verifying the credibility of the inference and performing experimental backtracking and model updates include: A four-level credibility check is established for the inference results: the first level checks image quality and registration error; the second level checks whether the damage response under graded loading satisfies sequential consistency; the third level checks the distance between the fusion features of the test sample and the feature distribution of the training sample; and the fourth level checks whether the prediction of local weak regions is supported by both image difference features and load map anomalies. Calculate the out-of-domain distance of the fused features of the sample to be tested relative to the training feature domain. When the out-of-domain distance exceeds a threshold, the model only outputs the locations of weak regions that need to be verified by actual testing. When any of the inference uncertainty, registration error, out-of-domain distance or physical consistency check exceeds the corresponding threshold, a sample is cut from the area to be tested or its adjacent location to perform a peel test. When the predicted weak area is small but the risk is high, the peel path is preferentially made to pass through the predicted weak area. When the model passes all confidence checks, the local performance map and the overall performance vector are used as the inference results of the adhesive performance correlation. When the overall average value meets the qualification requirements but the proportion of weak area exceeds the weak area limit, the electrode sheet is determined to require local review or process traceability. The multi-state images, load-bearing maps, local peel strength, and fracture failure modes of the measured backtracking samples are added to the incremental dataset. The model is retrained after the new samples meet the update conditions. The new model will only replace the original model if it simultaneously meets the following conditions on the new batch and the historical representative batch: the false negative rate of low-performance segments has not increased, the cross-batch error has not worsened, and the out-of-domain recognition ability has not decreased.

Images