A visual positioning and detection method for a battery assembly process

Abstract

The application discloses a visual positioning and detection method for a battery assembling process, and particularly relates to the field of power battery production and detection, and is used for solving the problems that the main edge of the pole piece and the texture of the diaphragm are difficult to be stably aligned under light fluctuation and working condition switching, the deviation causes are difficult to be accurately determined, and the disposal action is easy to be mis-triggered; a unified feature caliber is constructed through homologous image acquisition and same-piece state binding, a unique deviation cause label is output by using the fusion determination of the sequence mapping inverse bit rate and the law distance spectrum, a disposal instruction is issued when the direction, position and trigger condition are consistent, the same caliber re-inspection after disposal, verification result alignment and template recursive update are combined, integrated closed-loop management and control of positioning detection, action execution, result review and version traceability are realized, and the assembling consistency and disposal certainty are improved.

Description

Technical Field

[0001] This invention relates to the field of power battery production and testing, and more specifically, to a visual positioning and testing method for the battery assembly process. Background Technology

[0002] In the assembly process of power batteries and energy storage batteries, lamination, tab alignment, casing positioning, and process re-inspection are usually carried out continuously. Production lines generally use visual methods to identify electrode boundaries, separator positions, tab relationships, and local abnormal morphologies online, and the detection results are directly used for workstation release, interception, and correction. Existing solutions can complete target area positioning and anomaly detection during cycle production, and can also output misalignment or pass / fail conclusions, which has become the basic technical path for battery assembly quality control.

[0003] However, existing visual inspection systems still have a key shortcoming in the assembly line: the inspection results remain at the level of "discretions were observed," failing to generate actionable conclusions on "how to correct them." This manifests as a lack of differentiation when similar misalignment conclusions are applied to different workstations, difficulty in consistently adapting the same handling strategy to different assembly states, and inconsistencies in the judgment criteria for the same cell between different workstations. Ultimately, this creates a gap between inspection and correction, leading to repeated adjustments, poor coordination between release and interception, and insufficient stability in quality control on-site. These problems directly affect the continuity and consistency of the battery assembly process, constituting the key background technical problem that this invention aims to address.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a visual positioning and detection method for the battery assembly process. This method constructs a unified feature caliber by acquiring images from the same source and binding them to the state of the same component. It uses the fusion of sequential mapping inversion rate and normal distance spectral clustering to determine and output a unique deviation cause label. When the direction, position, and triggering conditions are consistent, a handling instruction is issued. Combined with post-handling re-inspection of the same caliber, verification result alignment, and template recursive updating, this method achieves integrated closed-loop management of positioning detection, action execution, result verification, and version traceability, improving assembly consistency and handling certainty, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: When the workpiece is stationary after lamination and pressing, collect the same source image of the station and read the station action status, complete the workpiece identification binding, extract the main edge line of the electrode and the center line of the diaphragm texture, and form a key feature set. S2: Based on the main edge of the pole piece, establish a sampling reference and generate a mapping sequence. Extract the mapping order mismatch characterization and normal fluctuation clustering characterization. Map the two types of characterizations to support and perform conflict suppression evidence fusion. Output the unique bias cause label and confidence level. S3: Search the treatment template library by the unique deviation cause label and generate a unique treatment instruction set. Perform three consistency checks. If the checks pass, issue the treatment instruction set. If the checks fail, output a re-check instruction and write it into the traceable execution record. S4: After treatment, re-extract the same source image from the same work station and from the same perspective, reconstruct the key feature set according to the same caliber and recalculate the results, align with the results before treatment to generate verification results, update the work station baseline template or retain the original template based on the verification results and output the re-inspection instruction.

[0007] Furthermore, when the lamination and pressing are completed and the workpiece is stationary, the same source image of the station is acquired and the station action status is read synchronously. The binding is completed according to the same workpiece identifier. The effective area of ​​the electrode is constructed according to the action constraint window and the acquisition aperture is fixed to eliminate background noise and reflection disturbance.

[0008] Furthermore, continuous edges are extracted within the effective area of ​​the electrode and the main edge line of the electrode is selected. Texture ridges are extracted within the same area and the midline of the diaphragm texture is selected. The coordinate sequence of the main edge line of the electrode, the coordinate sequence of the midline of the diaphragm texture, and the station action status are uniformly encapsulated into a key feature set.

[0009] Furthermore, a unified reference is established based on the coordinate sequence of the main edge line of the electrode and a normal sampling line is generated. The coordinate sequence of the middle line of the diaphragm texture is projected onto the normal sampling line to form a mapping point pair sequence and a mapping sequence. Mapping conflicts are resolved according to the rules of priority of tangential offset and priority of sequence number. The mapping index is limited to the sampling sequence number range.

[0010] Furthermore, the order consistency check is performed on the mapping sequence and the inverse relationship is extracted to generate the order-mapped inverse rate. The normal distance sequence is extracted from the mapping point pairs and detrending processing and spectral decomposition are performed. The energy is converged according to the fold-related frequency band defined by the workstation baseline template to generate the normal spectral clustering. The mapping order mismatch characterization is quantified by the order-mapped inverse rate, and the normal fluctuation clustering characterization is quantified by the normal spectral clustering.

[0011] Furthermore, the sequence mapping inverse rate and normal distance spectrum clustering are mapped to the bias cause label support distribution. Conflict suppression evidence fusion is used to generate unique bias cause labels and confidence levels, while simultaneously outputting bias direction markers and bias location segments.

[0012] Furthermore, templates are filtered by deviation cause tags in the handling template library, and template versions are determined by the proximity relationship of the confidence level and the overlap relationship of the deviation location segment. A unique handling instruction set is generated and the template version and field order are locked.

[0013] Furthermore, the system performs consistency checks on deviation direction, deviation location, and trigger condition based on the unique handling instruction set. When the consistency check passes, the handling action is issued and written into the execution record. When the consistency check fails, a re-check instruction is output and the failure reason code and image evidence index are written into the record.

[0014] Furthermore, after the handling action is completed, the work station and viewpoint are locked according to the execution record and the same source image of the re-inspection work station is collected. The key feature set is reconstructed according to the same caliber in step S1 and the sequence inversion rate, normal distance spectrum clustering, deviation cause label and confidence level are recalculated according to the same caliber in step S2. At the same time, deviation direction mark and deviation location segment are generated and the comparison results before and after the handling are formed.

[0015] Furthermore, the comparison results before and after the treatment are written into the verification process and combined with the verification coefficient, the change status of the deviation cause label, the change status of the confidence level, and the convergence status of the key features to generate a verification pass mark. When the verification passes, the workstation baseline template is updated and the template version is released. When the verification fails, the original template version is maintained and a re-inspection instruction is issued. The handover fields are called by steps S1 to S3 in a fixed order.

[0016] The technical effects and advantages of the visual positioning and detection method for battery assembly process of the present invention are as follows: This invention integrates image acquisition, deviation identification, handling execution, and re-inspection and update in battery assembly into a single closed loop, eliminating the separation of detection and adjustment into two separate sets of standards. The front end uses the same station and identifier to constrain the acquisition standard; the middle section converts the electrode main edge line and the separator texture center line into a comparable mapping relationship; and then, through dual-parameter comprehensive analysis, it provides a unique deviation cause label. The rear end only executes handling actions when the direction, position, and triggering conditions are simultaneously consistent. Therefore, it can simultaneously reduce false detections and false actions, avoid mistaking local texture disturbances for assembly deviations, and prevent the true deviation from being sent to the wrong application location.

[0017] After execution, this invention performs a re-inspection using the same criteria and directly aligns the results before and after the treatment. Then, based on the verification conclusions, it selects to update the template or roll back the re-inspection, ensuring that the on-site strategy is always driven by traceable evidence rather than relying on repeated trial and error based on human experience. For power battery and energy storage battery assembly scenarios, this end-to-end collaborative mechanism can stabilize assembly consistency, reduce judgment drift caused by batch changes and light fluctuations, and improve the certainty and verifiability of production line treatment. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a visual positioning and detection method for battery assembly according to the present invention. Detailed Implementation

[0019] 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.

[0020] Example 1: Figure 1 This invention provides a visual positioning and detection method for a battery assembly process, comprising: S1: When the workpiece is stationary after lamination and pressing, collect the same source image of the station and read the station action status, complete the workpiece identification binding, extract the main edge line of the electrode and the center line of the diaphragm texture, and form a key feature set.

[0021] S2: Based on the main edge of the pole piece, establish a sampling reference and generate a mapping sequence. Extract the mapping order mismatch characterization and normal fluctuation clustering characterization. Map the two types of characterizations to support and perform conflict suppression evidence fusion. Output the unique bias cause label and confidence level.

[0022] S3: Search the handling template library by unique deviation cause label and generate a unique handling instruction set. Perform three consistency checks. If the checks pass, issue the handling instruction set. If the checks fail, output a re-check instruction and write it into the traceable execution record.

[0023] S4: After treatment, re-extract the same source image from the same work station and from the same perspective, reconstruct the key feature set according to the same caliber and recalculate the results, align with the results before treatment to generate verification results, update the work station baseline template or retain the original template based on the verification results and output the re-inspection instruction.

[0024] This invention integrates image acquisition, deviation identification, handling execution, and re-inspection and update in battery assembly into a single closed loop, eliminating the separation of detection and adjustment into two separate sets of standards. The front end uses the same station and identifier to constrain the acquisition standard; the middle section converts the electrode main edge line and the separator texture center line into a comparable mapping relationship; and then, through comprehensive analysis, provides a unique deviation cause label. The back end only executes handling actions when the direction, position, and triggering conditions are simultaneously consistent. Therefore, it can simultaneously reduce false detections and false actions, avoid mistaking local texture disturbances for assembly deviations, and prevent the true deviation from being sent to the wrong application location.

[0025] After execution, this invention performs a re-inspection using the same criteria and directly aligns the results before and after the treatment. Then, based on the verification conclusions, it selects to update the template or roll back the re-inspection, ensuring that the on-site strategy is always driven by traceable evidence rather than relying on repeated trial and error based on human experience. For power battery and energy storage battery assembly scenarios, this end-to-end collaborative mechanism can stabilize assembly consistency, reduce judgment drift caused by batch changes and light fluctuations, and improve the certainty and verifiability of production line treatment.

[0026] The determination of deviations at the battery assembly site relies on whether the image evidence corresponds one-to-one with the workstation status. If the acquisition time, workpiece identification, and viewing angle are inconsistent, subsequent calculations will mistakenly treat non-assembly factors as structural deviations, leading to distorted positioning results. Step S1 first binds and fixes the same source acquisition and the same part, and then extracts the main edge line of the electrode and the center line of the separator texture into the same reference, providing a comparable and reproducible input basis for subsequent mapping analysis.

[0027] S101. Data collection trigger and binding with the same component.

[0028] To ensure that the source image of the workstation corresponds to the same workpiece as the workstation action state, the stacking and pressing end state, the workpiece stationary state, and the acquisition permission state are read first. Acquisition is triggered when all three states are enabled. Then, the vision acquisition device outputs the source image of the workstation, and the workstation controller synchronously outputs the workstation action state. The workstation is bound with the same workpiece identifier to form a single workpiece input record. The workstation action state includes the end state, stationary state, acquisition permission state, fixture position quantity, and fixture attitude quantity. The end state, stationary state, and acquisition permission state are binary states. The fixture position quantity is expressed using a normalized position interval, and the fixture attitude quantity is expressed using a complete circumferential interval. Example: After the pressing mechanism stops, the workpiece remains stationary. The acquisition permission state is enabled. The vision acquisition device and the workstation controller write the same record under the same workpiece identifier. When there is a record conflict, the conflict record is discarded and acquisition is retried.

[0029] S102. Image normalization and illumination shaping.

[0030] To reduce the disturbance of edge positioning caused by local reflections and texture noise, bilateral filtering is first performed on the source images of the workstations. The logic is as follows: first, calculate the spatial distance attenuation of each neighboring pixel in the spatial neighborhood, then calculate the gray-level difference attenuation, then multiply the two attenuations to obtain the neighborhood contribution, and finally normalize the neighborhood contribution and sum the neighborhood gray levels to obtain the filtering result. Then, illumination shaping is performed. The operation order is as follows: first, smooth the background estimation of the filtering result, then divide the filtering result by the background estimation and add a small stabilizing term to obtain the illumination normalized map, and then map the illumination normalized map to the normalized gray-level range. The normalized gray-level range is expressed as a closed interval from the minimum gray level to the maximum gray level. The output is a normalized version of the source images of the workstations, and the edge and texture contrast remains stable under different lighting conditions.

[0031] S103. Construction of the effective region of the electrode.

[0032] To ensure that geometric extraction is performed only within a reliable region, the fixture reference window is first mapped to the source image of the current station based on the fixture position and attitude values ​​in the station's motion state, thus obtaining the motion constraint window. Then, the gradient magnitude map is calculated within the motion constraint window. Subsequently, a quantile threshold is selected according to the cumulative distribution of gradient magnitudes, and pixels above the quantile threshold are grouped into candidate regions. Finally, morphological closing operations are performed on the candidate regions to obtain the effective pole piece region. The proportion of the effective pole piece region is defined as the ratio of the area of ​​the effective pole piece region to the area of ​​the motion constraint window, and the value is in the range of greater than zero and not greater than one. Example: After the roll material batch is switched, the reflective area changes, the gradation threshold is automatically adjusted according to the gradient distribution, the effective area of ​​the electrode still covers the continuous area of ​​the main edge line, and subsequent extraction is not misled by local glare.

[0033] S104. Extraction of the main edge line of the electrode and the midline of the diaphragm texture.

[0034] To ensure that the dual-line features are computable under the same geometric semantics, edge thinning and connected contour extraction are first performed within the effective region of the pole piece. Then, the contour length and orientation consistency are calculated for each contour. The contour length is obtained by accumulating the distances between adjacent contour points, and the orientation consistency is obtained by the ratio of the resultant length of the unit tangent vector to the number of tangent vectors. The orientation consistency value is in the range of zero to one. The main edge lines of the pole piece are selected by comparing the contour length first and then the orientation consistency. Subsequently, the analytical expression of the main edge lines of the pole piece is obtained by fitting the total least squares of straight lines. Then, second derivative ridge line extraction is performed within the effective region of the pole piece, and the first step is to... The imaging rules of the workstation light source determine the texture polarity, and then the connected skeletonization is performed to obtain candidate midlines of the diaphragm texture. For each candidate, the coherence and tangential consistency are calculated. The coherence is obtained by the ratio of the effective skeleton length to the effective skeleton length plus the break length. The tangential consistency is obtained by averaging the degree of same direction between the local tangential of the diaphragm texture midline and the tangential of the electrode main edge line. Both indicators are in the range of zero to one. The diaphragm texture midline is selected by comparing the coherence first and then the tangential consistency. After this processing, the directional relationship between the electrode main edge line and the diaphragm texture midline is clear under the same reference, and S2 can directly establish a mapping sequence.

[0035] S105. Encapsulation and handover constraints of key feature sets.

[0036] To ensure consistent scale and object consistency during S2 calls, the main edge of the electrode and the midline of the diaphragm texture are first resampled at equal intervals according to arc length. The resampling interval is provided by the workstation baseline template and remains unchanged within the batch. Then, the coordinates of the resampled points are normalized to a unified coordinate range according to the image width and height. Subsequently, the resampled electrode main edge coordinate sequence, the resampled diaphragm texture midline coordinate sequence, workstation action status, effective electrode area, and effective electrode area percentage are merged into a key feature set. The order and naming of the key feature set fields are fixed. After the key feature set is written, it is directly handed over to S2 for sequential mapping inverse ratio calculation and normal distance spectrum clustering calculation.

[0037] Step S1 completes the construction of single-piece input records and the encapsulation of key feature sets. The output includes the coordinate sequence of the main edge line of the electrode, the coordinate sequence of the midline of the diaphragm texture, and the workstation action status. The input sources are clear, the naming is unified, and the standards are consistent, which can support the subsequent construction of mapping sequences and deviation quantification analysis.

[0038] Dual-line features alone cannot directly guide the handling actions; assembly deviations need to be transformed from geometric relationships into actionable conclusions. S2 establishes a unified reference around the main edge line of the electrode, projects the diaphragm texture centerline into mapping point pairs, and then extracts stable indicators from two evidence chains: order relationship and normal fluctuation, to avoid misleading results from a single indicator under complex working conditions.

[0039] Step S2 selects only the sequence mapping inversion rate and normal spectrum clustering for comprehensive analysis because they respectively characterize two complementary and most critical types of interference information. The former reflects the structural mismatch after the mapping order is disturbed, while the latter reflects the clustering intensity of normal fluctuations in a specific frequency band. Together, they cover the main manifestations of assembly interference from geometric relationships to texture morphology, while distinguishing them from illumination fluctuations and overall translation. This avoids the correlation superposition and judgment drift caused by introducing redundant parameters. Based on this, the interference confidence coefficient is calculated, which can maintain the stability of the judgment caliber and the simplicity of the threshold calibration, while also improving the interpretability of the results and the consistency of workstation execution, making it easier for subsequent handling instructions to achieve accurate direction, accurate position, and accurate triggering.

[0040] S201. Input verification and mapping preparation.

[0041] To ensure that the mapping calculation always corresponds to the same workpiece record, the coordinate sequence of the electrode main edge line, the coordinate sequence of the diaphragm texture centerline, the station action status, the effective area of ​​the electrode, and the proportion of the effective area of ​​the electrode are first read from the key feature set. Then, input verification is performed. The verification conditions include that the two coordinate sequences have the same length and are not less than the minimum sampling length, the coordinate sequences are sorted in ascending order of arc length, the proportion of the effective area of ​​the electrode is in the closed interval between zero and one, and the station action status fields are complete and the timestamps are consistent. After the verification is passed, continuous sampling positions are established on the electrode main edge line coordinate sequence according to the arc length, and the local tangential direction and local normal direction are calculated. The boundary sampling positions use one-sided difference, the middle sampling positions use two-sided difference, the direction vector is uniformly normalized, and the mapping index uses the continuous integer sequence number from the first sampling position to the last sampling position.

[0042] S202. Unified reference mapping and sequential mapping inverse rate calculation.

[0043] To transform geometric relationships into ordered relationships, the projection point with the smallest normal distance is searched in the coordinate sequence of the membrane texture along the local normal direction of each sampling point. When the minimum normal distances are tied, the absolute values ​​of tangential offsets are compared first, and the candidate point with the smallest absolute value of tangential offset is selected. When the absolute values ​​of tangential offsets are still tied, the candidate point with the smallest mapping index value is selected, forming a unique mapping point pair sequence and a unique mapping sequence. Then, an order consistency check is performed, and the number of inversion pairs is counted according to the order of the mapping indexes. The inversion rate is normalized to the total number of comparable point pairs. The inversion rate ranges from zero to one. An increase in the inversion rate indicates a stronger degree of order disorder. In the embodiment, when the edge of the electrode is pulled, the mapping sequence produces an index reversal in a local segment, and the inversion rate increases synchronously and is consistent with the direction of the on-site misalignment.

[0044] S203. Calculation of normal distance sequence and normal spectral clustering.

[0045] To distinguish between overall translation and local folding, the normal distance is first calculated point by point according to the sampling sequence number based on the mapping point pair to obtain a normal distance sequence. Then, a linear baseline is formed by the first and last sampling points, and baseline subtraction is performed on each sampling point to obtain a detrended sequence. Subsequently, the preset quantile of the absolute value sequence of the detrended sequence is taken as the truncation threshold, and amplitude limiting is performed on the amplitude exceeding the threshold to form a truncated sequence. Then, discrete spectrum decomposition is performed on the truncated sequence to calculate the spectral energy corresponding to each discrete frequency index. The spectral energy is aggregated according to the folded frequency band index set preset in the current workstation baseline template version, and the normal distance spectral aggregation degree is defined by the ratio of folded frequency band energy to full frequency band energy. When the full frequency band energy is zero, the normal distance spectral aggregation degree is defined as zero. The normal distance spectral aggregation degree takes values ​​from zero to one. The folded frequency band index set remains unchanged within the same batch and is updated when the template version is switched.

[0046] S204. Dual Evidence Fusion and Bias Cause Labeling Generation.

[0047] S204-1. Support distribution construction and mapping caliber fixation.

[0048] The input includes the sequential inversion rate and the normal distance spectrum clustering, while simultaneously reading two sets of monotonic segmented mapping relationships from the workstation baseline template. The first set of mapping relationships converts the sequential inversion rate into the first evidence support of each label within the deviation cause label set, and the second set of mapping relationships converts the normal distance spectrum clustering into the second evidence support of each label within the deviation cause label set. The monotonic segmented mapping relationships are defined using segmented intervals and interval endpoint support. When the sequential inversion rate or normal distance spectrum clustering falls into a certain interval, the support is obtained by linear interpolation based on the interval endpoint support; when it falls outside the interval, the value is taken from the nearest endpoint support. Both the first and second evidence support are limited to the range of zero to one. The deviation cause label set is fixed as overall translation type, local traction type, and composite perturbation type. The support distribution includes three support items, all of which maintain the same field order and are bound to the template version.

[0049] S204-2. Quality allocation generation and definition of unallocated sets.

[0050] To transform the support distribution into a form that can be processed by evidence fusion, the sum of support for the first piece of evidence and the sum of support for the second piece of evidence are calculated separately. When the sum of support is zero, the default support vector in the workstation baseline template is read and replaced before the calculation continues. Then, the unassigned smoothing amount in the workstation baseline template is introduced. The quality of each deviation cause label is defined as the corresponding support divided by the sum of the support and the unassigned smoothing amount. The quality of the unassigned set is defined as the unassigned smoothing amount divided by the sum of the support and the unassigned smoothing amount. The unassigned set is used to express the part of the evidence that cannot clearly point to any deviation cause label. The quality of the unassigned set is also limited to the range of zero to one. The quality allocation satisfies that the sum of the quality of the three types of deviation cause labels and the quality of the unassigned set is one, thereby ensuring that the two sets of quality allocations can be directly combined under the same dimension.

[0051] S204-3. Calculation of Conflict Quantity and Rules for Conflict Limitation

[0052] To suppress the amplification of conflict when two pieces of evidence point to different deviation cause labels, the conflict quantity is first calculated. This conflict quantity is obtained by multiplying and summing the quality of any deviation cause label in the first evidence quality allocation and the quality of different deviation cause labels in the second evidence quality allocation. The conflict quantity is limited to a left-closed, right-open range of zero to one. Next, the upper limit of conflict in the workstation baseline template is read, and the smaller value between the conflict quantity and the upper limit is taken as the amplitude-limited conflict quantity. This amplitude-limited conflict quantity is used in the denominator of the subsequent fusion normalization. The upper limit of conflict is managed according to the template version to avoid introducing conflict amplification paths into the fusion results from different batches.

[0053] S204-4. Integration of redistribution, unique bias cause label, and trust level generation.

[0054] To generate a unique and executable conclusion, the fusion quality is calculated according to the conflict suppression evidence fusion rules. For each deviation cause label, the cross-quality of labels with the same name is calculated first, then the cross-quality of the first evidence label quality and the second evidence unassigned set quality is calculated, and then the cross-quality of the first evidence unassigned set quality and the second evidence label quality is calculated again. The three cross-quality values ​​are added together and divided by one minus the amplitude limit conflict amount to obtain the label fusion quality. The unassigned set fusion quality is obtained by multiplying the two sets of unassigned set qualities and dividing by one minus the amplitude limit conflict amount. The unique deviation cause label is the one with the highest fusion quality. In case of a tie, the label priority order defined by the workstation baseline template is used. The confidence level is jointly determined by the unique deviation cause label fusion quality and the unassigned set fusion quality. The confidence level increases with the unique deviation cause label fusion quality and decreases with the unassigned set fusion quality. The confidence level is ultimately limited to the range of zero to one and is bound to the template version to ensure that different workstations and different batches use the same calculation method.

[0055] The trust level can be determined in the following ways: We can define it using the same criteria: first calculate the original credibility, then perform interval truncation. Specifically, the original credibility = unique bias cause label fusion quality / (1 - unassigned set fusion quality + smoothing term). The credibility level is the smaller value between the original credibility and one. For example, when the unique bias cause label fusion quality is 0.72, the unassigned set fusion quality is 0.18, and the smoothing term is 0.02, the credibility level is approximately 0.85. If the unique bias cause label fusion quality drops to 0.52 and the unassigned set fusion quality rises to 0.38, the credibility level is approximately 0.81 under the same smoothing term. This achieves the common constraint that the more concentrated the label evidence, the higher the credibility level, and the greater the unassigned uncertainty, the lower the credibility level.

[0056] S204-5. Generate and encapsulate the deviation direction marker and deviation position segment for output.

[0057] The deviation direction marker is derived from the overall sign trend of the normal distance sequence. First, the mean of the normal distance sequence is calculated, and then the zero-interval threshold of the direction in the workstation baseline template is read. When the mean falls within the zero-interval threshold, a neutral value is output; when the mean is greater than the upper limit of the zero-interval threshold, a positive value is output; and when the mean is less than the lower limit of the zero-interval threshold, a negative value is output. The deviation direction marker value set is fixed as negative, neutral, and positive. The deviation position segment is derived from the salient position set of the normal distance sequence. First, the absolute value sequence of the normal distance is taken, and then the salient quantiles in the workstation baseline template are read. The quantile amplitude of the absolute value sequence is used as the salient threshold. The sampling numbers with absolute values ​​not less than the salient threshold constitute the salient position set. When the salient position set is empty, a neutral position segment is output and the upper limit of the confidence level is simultaneously reduced. When the salient position set is not empty, the minimum and maximum sampling numbers of the salient position set are taken and normalized according to the total number of samples to form the deviation position segment. The deviation position segment is limited to the range of zero to one. Step S204 finally outputs the unique deviation cause label, confidence level, deviation direction mark, deviation location segment, sequence mapping inversion rate, and normal distance spectrum clustering in a fixed field order. Step S3 directly reads the deviation direction mark and deviation location segment to complete the direction consistency verification and location consistency verification. The field naming and field order remain unchanged when the template version is switched.

[0058] S205. Output encapsulation and step S3 call constraints.

[0059] To ensure that the consistency check in step S3 can be directly invoked, the deviation cause label, confidence level, deviation direction mark, deviation location segment, sequence mapping inversion rate, and normal distance spectrum clustering are output in a fixed field order. The deviation direction mark value set is fixed as negative, neutral, and positive. The deviation location segment is expressed using a normalized interval and limited to the range of zero to one. The deviation location segment is extracted using the minimum normalized position and the maximum normalized position of the high amplitude normal distance point set. When the high amplitude point set is empty, the neutral location segment is output and the low confidence state is marked simultaneously. Step S3 performs direction consistency check based on the deviation direction mark, position consistency check based on the deviation location segment, and trigger condition consistency check based on the workstation action status.

[0060] Step S2 outputs a unique deviation cause label and confidence level, and simultaneously provides deviation direction markers and deviation location segments, completing the transformation from image geometry to processing semantics. The result has single-valuedness, interpretability, and callability, and can directly enter the consistency verification and action generation stage.

[0061] Deviation identification results can only be converted into reliable workstation actions after passing through execution gating. If there is any inconsistency in the direction of action, the location of action, or the triggering condition, erroneous or ineffective actions may occur on-site. S3 indexes the action template with a unique deviation cause label and performs three consistency checks before the action is taken to ensure that the instruction and the evidence are in the same direction, position, and condition.

[0062] S301. Input the acceptance and unique template selection.

[0063] To ensure a single and reproducible handling path, the system first receives the deviation cause label, confidence level, deviation direction marker, deviation location segment, sequence mapping inversion rate, and normal distance spectrum clustering from step S2. Then, it receives the workstation action status and performs a field integrity check. After the field integrity is verified, the system performs a first-round screening in the handling template library based on the deviation cause label. In the first-round results, a second-round screening is performed based on the confidence level proximity rule. In the second-round results, a final-round screening is performed based on the deviation location segment overlap priority rule. If parallel templates still exist, the latest template is selected in the order of template version age. Finally, a unique handling instruction set is obtained, and the template version is locked. The confidence level maintains a normalized interval expression, the deviation direction marker is limited to a negative, neutral, and positive three-value set, and the deviation location segment is limited to the normalized start point to the normalized end point, with the start point no later than the end point.

[0064] S302. Three-item consistency check and consistency score calculation.

[0065] To prevent malfunctions caused by mismatched direction and position, a deviation direction consistency check is first performed. When the deviation direction is marked as neutral, only neutral handling directions are accepted. When the deviation direction is marked as negative or positive, the handling directions must be in the same direction. Then, a deviation position consistency check is performed using an interval overlap ratio algorithm. The interval overlap ratio is calculated in a fixed order: first, the intersection length of the deviation position segment and the handling action segment is calculated; then, the union length of the two segments is calculated; finally, the position consistency score is obtained by dividing the intersection length by the union length. When the union length is zero, the position consistency score is defined as zero. The position consistency score is expressed using a normalized interval. Subsequently, a trigger condition consistency check is performed, and the template trigger conditions are matched item by item according to the workstation action status. When all trigger items are satisfied, the trigger condition consistency is passed.

[0066] The handling action zone is a continuous positional interval within which the actuator corresponding to the handling command can generate effective corrections, based on the sampling reference of the electrode's main edge. Its value falls within a normalized interval. In practice, the handling action zone and the deviation position zone are used for positional consistency checks; only when they overlap sufficiently is entry into the execution branch permitted. If a handling command set contains multiple actions, the handling action zone uses the combined interval of the action ranges of these actions.

[0067] Example: The deviation direction marker shows a positive direction and the deviation position segment is concentrated at the edge. The handling template library gives the same-direction edge handling instruction. The workstation action status is in the template triggering interval. After all three consistencys are passed, the execution branch is entered.

[0068] S303. Gating decision and re-inspection rollback.

[0069] To ensure that the execution action only occurs when the evidence is closed, a hard-gating logic is used to summarize the consistency results of deviation direction, deviation position, and trigger condition. When all three results pass, a unique set of handling instructions is issued and written to the execution start marker. When any result fails, a re-inspection instruction is generated and written to the interception marker, along with the failure reason code and failure field index. The failure reason code is limited to three categories: inconsistency of direction, inconsistency of position, and inconsistency of trigger. The re-inspection instruction is sent back to the collection node while keeping the current template version unchanged.

[0070] S304. Execution record keeping and handover with step S4.

[0071] To ensure traceability of the alignment and template recursive update in step S4, both the execution branch and the re-inspection branch output records with a unified structure. The record fields are fixed as follows: template version, unique handling instruction set, workstation action status, deviation cause label, credibility level, deviation direction mark, deviation location segment, execution time sequence, image evidence index, branch mark, and failure reason code. The field order and naming are fixed. Step S4 reads the records according to the fixed field order and conducts re-inspection and verification result determination with the same caliber. In the example, when the position consistency fails, the branch mark remains in the re-inspection state, and the failure reason code directly points to the position inconsistency. On-site reviewers can directly locate and intercept the cause and re-collect the data.

[0072] Step S3 forms a unique set of handling instructions and a traceable execution record. When the verification is successful, the handling is executed. When the verification fails, a re-inspection instruction is output and the data collection node is rolled back. The action branches are clear and the responsibility boundaries are well-defined, which ensures both the effectiveness of the handling and the verifiability of the problem.

[0073] The completion of the action does not mean that the deviation has converged. It must be re-inspected under the same work station, viewpoint, and caliber to determine whether the action has truly changed the deviation state. Step S4 takes over the execution record and conducts a re-inspection and recalculation. By aligning and comparing before and after the action, a verification result is formed, and a decision is made on whether to update or retain the template based on this result.

[0074] S401. Re-inspection trigger and data collection consistency confirmation.

[0075] The execution log provides the handling actions and evidence index. During the re-inspection phase, it is first confirmed that the re-inspection screen and the execution screen belong to the same workstation and the same viewing angle. Then, image acquisition is initiated. The processing logic is to first read the workstation action status at the execution time and the workstation action status at the re-inspection time, and then calculate the normalized amount of the lateral position deviation, the normalized amount of the longitudinal position deviation, and the normalized amount of the posture deviation. All three normalized amounts are obtained by dividing the corresponding original deviation by the same normalized scale in the workstation baseline template. Then, the square root of the sum of the squares of the three normalized amounts is used to obtain the pose consistency amount. The pose consistency amount is not less than zero. The pose consistency amount is compared with the pose consistency threshold in the workstation baseline template. If the threshold requirement is met, the same source image of the re-inspection workstation is acquired. If the threshold requirement is not met, a re-inspection instruction is output and a pose inconsistency mark is written.

[0076] S402. Reconstruction and recalculation of the same caliber.

[0077] The re-examined image enters the same caliber process as before execution. The verification stage only compares the differences before and after, without repeating the details extracted in the previous stage. The processing logic is as follows: first, reconstruct the key feature set according to the same caliber as in step S1; then, recalculate the sequence mapping inversion rate, normal distance spectral clustering, deviation cause label, confidence level, deviation direction marker, and deviation location segment according to the same caliber as in step S2; and finally, form a comparison package of the results before and after re-examination according to a fixed field order. The field names and field order are consistent. The sequence mapping inversion rate, normal distance spectral clustering, and confidence level are all expressed in normalized intervals, and the deviation location segment is expressed from the normalized start point to the normalized end point.

[0078] S403. Verification result calculation and pass / fail determination.

[0079] The goal of the verification phase is to determine whether the action has resulted in reproducible improvement. The judgment logic uses parallel comparison of cause-preservation paths and cause-transfer paths. The processing logic calculates the changes in two parameters, the normal distance amplitude ratio, the convergence of key features, and the label stability, and synthesizes the verification coefficients. The changes in two parameters are composed of the inversion rate of the sequence mapping after re-examination minus the inversion rate of the sequence mapping before treatment and the normal distance spectrum clustering after re-examination minus the normal distance spectrum clustering before treatment, and are mapped to the improvement score according to the template change threshold. The normal distance amplitude ratio is obtained by dividing the absolute mean of the normal distance after re-examination by the absolute mean of the normal distance before treatment, and is mapped to the improvement score after amplitude limiting. The convergence of key features is obtained by subtracting the normalized value of the geometric difference between the first and second lines and mapping it to the improvement score after amplitude limiting. The label stability is obtained by jointly mapping the label change state of the deviation cause and the change state of the credibility level. The verification coefficients are synthesized by the geometric mean of the four improvement scores and limited to the normalized interval.

[0080] The determination is made by comparing two paths in parallel. The first path is the cause-preservation path, which requires that the deviation cause label remains consistent, the confidence level does not decrease, and the verification coefficient reaches the template threshold. The second path is the cause-transfer path, which requires that the deviation cause label migrates, the confidence level increases, the verification coefficient reaches the transfer threshold, the deviation direction mark is consistent with the treatment direction, and the deviation location segment is consistent with the treatment effect segment. If either path is satisfied, a verification pass mark is output.

[0081] Example: Before the treatment, the deviation cause label showed a composite perturbation type. After the treatment, the deviation cause label migrated to an overall translation type, the confidence level increased, the edge wrinkles disappeared, the verification coefficient reached the migration threshold, and the verification passed mark was set to pass.

[0082] S404. Workstation baseline template conditions update.

[0083] Template updates only accept validated samples, and the update criteria remain consistent in one direction to prevent abnormal samples from contaminating the template. The processing logic is as follows: first, the re-examination post-order mapping inverse rate and the re-examination post-normal distance spectrum clustering are written into the historical queue in chronological order and a fixed window is maintained; then, the key feature baseline interval is updated according to the quantile boundary of the historical queue; subsequently, order-preserving regression is performed with the two-parameter trajectory and deviation cause label of validated samples as input to obtain a new evidence fusion monotonic mapping relationship; then, the consistency verification threshold is updated according to the consistency score distribution of validated samples; finally, a new template version is generated and written back to the template repository and the effective time is recorded; when validation fails, the template version remains unchanged and the failure reason code is appended.

[0084] S405. Interchange packaging and link closure.

[0085] The handover phase is responsible for transforming the verification conclusions into callable data for the next task. Field naming and order must be fixed. The processing logic is as follows: first, encapsulate the verification pass flag, verification coefficient, deviation cause label change status, confidence level change status, key feature convergence status, and template version; then, encapsulate the branch flag and failure reason code; finally, send the handover package into the next task's call chain. Step S1 reads the collection caliber field from the template version; step S2 reads the baseline interval and evidence fusion mapping relationship from the template version; and step S3 reads the consistency verification threshold field from the template version. The on-site action chain and analysis chain maintain the same version caliber.

[0086] Step S4 completes the generation of verification results and template version management, transforming the re-inspection conclusions into workstation baseline template update rules that can be used sustainably, realizing an integrated closed loop of detection, handling, review, and updating, so that subsequent batches can maintain stable judgment and stable execution under the same technical standards.

[0087] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0088] All the thresholds or preset parameters mentioned above can be pre-calibrated through offline simulation testing, or set to fixed values ​​according to the on-site operating procedures.

[0089] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0090] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A visual positioning and detection method for battery assembly process, characterized in that, Including the following steps: S1: When the workpiece is stationary after lamination and pressing, collect the same source image of the station and read the station action status, complete the workpiece identification binding, extract the main edge line of the electrode and the center line of the diaphragm texture, and form a key feature set. S2: Based on the main edge of the pole piece, establish a sampling reference and generate a mapping sequence. Extract the mapping order mismatch characterization and normal fluctuation clustering characterization. Map the two types of characterizations to support and perform conflict suppression evidence fusion. Output the unique bias cause label and confidence level. S3: Search the treatment template library by the unique deviation cause label and generate a unique treatment instruction set. Perform three consistency checks. If the checks pass, issue the treatment instruction set. If the checks fail, output a re-check instruction and write it into the traceable execution record. S4: After treatment, re-extract the same source image from the same work station and from the same perspective, reconstruct the key feature set according to the same caliber and recalculate the results, align with the results before treatment to generate verification results, update the work station baseline template or retain the original template based on the verification results and output the re-inspection instruction.

2. The visual positioning and detection method for a battery assembly process according to claim 1, characterized in that, Step S1 includes: When the lamination and pressing are completed and the workpiece is stationary, the same source image of the station is acquired and the station action status is read synchronously. The binding is completed according to the same workpiece identifier. The effective area of ​​the electrode is constructed according to the action constraint window and the acquisition aperture is fixed to eliminate background noise and reflection disturbance.

3. The visual positioning and detection method for a battery assembly process according to claim 2, characterized in that, Step S1 also includes: Continuous edges are extracted within the effective area of ​​the electrode and the main edge line of the electrode is selected. Texture ridges are extracted within the same area and the midline of the diaphragm texture is selected. The coordinate sequence of the main edge line of the electrode, the coordinate sequence of the midline of the diaphragm texture, and the station action status are uniformly encapsulated into a key feature set.

4. The visual positioning and detection method for a battery assembly process according to claim 3, characterized in that, Step S2 includes: A unified reference is established using the coordinate sequence of the main edge line of the electrode and a normal sampling line is generated. The coordinate sequence of the middle line of the diaphragm texture is projected onto the normal sampling line to form a mapping point pair sequence and a mapping sequence. Mapping conflicts are resolved according to the rules of tangential offset priority and sequence number priority. The mapping index is limited to the sampling sequence number range.

5. The visual positioning and detection method for a battery assembly process according to claim 4, characterized in that, Step S2 also includes: The mapping sequence is subjected to order consistency check and inversion relationship is extracted to generate the order-mapped inversion rate. The normal distance sequence is extracted from the mapping point pair and detrending processing and spectral decomposition are performed. The energy is gathered according to the fold correlation frequency band defined by the workstation baseline template to generate the normal distance spectrum clustering. The mapping order mismatch characterization is quantified by the order-mapped inversion rate, and the normal fluctuation clustering characterization is quantified by the normal distance spectrum clustering.

6. The visual positioning and detection method for a battery assembly process according to claim 5, characterized in that, Step S2 also includes: The sequence inversion rate and normal distance spectrum clustering are mapped to the bias cause label support distribution. Conflict suppression evidence fusion is used to generate unique bias cause labels and confidence levels, while also outputting bias direction markers and bias location segments.

7. The visual positioning and detection method for a battery assembly process according to claim 6, characterized in that, Step S3 includes: In the template library, templates are filtered by deviation cause tags, and template versions are determined by the proximity of the credibility level and the overlap of the deviation location segment. A unique set of handling instructions is generated and the template version and field order are locked.

8. The visual positioning and detection method for a battery assembly process according to claim 7, characterized in that, Step S3 also includes: The system performs consistency checks on deviation direction, deviation location, and trigger condition based on the unique set of instructions. If the consistency check passes, the system issues a handling action and writes it into the execution record. If the consistency check fails, the system outputs a re-check instruction and writes the failure reason code and image evidence index.

9. The visual positioning and detection method for a battery assembly process according to claim 8, characterized in that, Step S4 includes: After the handling action is completed, the work station and viewpoint are locked according to the execution record and the same source image of the re-inspection work station is collected. The key feature set is reconstructed according to the same caliber in step S1 and the sequence inversion rate, normal distance spectrum clustering, deviation cause label and confidence level are recalculated according to the same caliber in step S2. At the same time, deviation direction mark and deviation location segment are generated and the comparison results before and after the handling are formed.

10. The visual positioning and detection method for a battery assembly process according to claim 9, characterized in that, Step S4 also includes: The comparison results before and after the treatment are written into the verification process and combined with the verification coefficient, the change status of the deviation cause label, the change status of the confidence level, and the convergence status of the key features to generate a verification pass mark. When the verification passes, the workstation baseline template is updated and the template version is released. When the verification fails, the original template version is maintained and a re-inspection instruction is issued. The handover fields are called by steps S1 to S3 in a fixed order.

Images