A 3D vision-based real-time monitoring and defect identification system for battery cell welding

Abstract

The application discloses a kind of real-time monitoring and defect identification system in welding based on 3D vision, specifically relates to industrial vision detection technical field, corresponding relationship of welding stage and acquisition time is established by synchronous trigger, continuous acquisition welding point space topography data, calculate oscillation characteristic factor and time sequence matching factor, and regulate collection rhythm acquisition time sequence and acquisition mode according to this, three-dimensional data is formed on the basis of regulation, and the three-dimensional geometric feature set of welding point is constructed, and then the determination output of welding defect type and severity is completed, and complete welding real-time monitoring and processing flow logic closed loop structure is formed;The application realizes the alignment of welding stage and three-dimensional acquisition time by synchronous trigger, adjusts acquisition rhythm, acquisition time sequence and acquisition mode by combining dynamic evaluation and closed loop regulation, obtains high consistency three-dimensional data, constructs the three-dimensional geometric feature set of welding point, outputs welding defect type and defect severity, and improves battery welding quality.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and more specifically, to a real-time monitoring and defect identification system for battery cell welding based on 3D vision. Background Technology

[0002] In the battery cell manufacturing process, welding processes such as tab welding and connector welding are typically completed under high-speed, high-energy-density conditions. These welding processes are characterized by short time scales, rapid changes in the molten pool, and significant fluctuations in the spatial morphology of the weld joint within milliseconds. During the welding process, the weld joint may exhibit dynamic behaviors such as molten pool oscillation, local collapse, or weld misalignment. These behaviors have a significant impact on welding quality, but after welding is completed, this information is often difficult to fully reflect through the static appearance characteristics of the weld joint alone.

[0003] In existing technologies, post-weld inspection or static imaging based on two-dimensional vision is typically used to analyze the weld surface in order to monitor welding quality. While these technologies are simple to implement and mature in welding quality assessment, they primarily focus on the final result after welding and struggle to capture the process information of how the spatial morphology of the weld changes over time during welding. Therefore, in scenarios where the dynamic characteristics of the welding process are significant, their ability to reflect certain welding states remains limited. Meanwhile, some technologies introduce three-dimensional vision to inspect welds, enhancing the ability to acquire spatial morphology information. However, these three-dimensional data acquisition methods often use fixed time windows or fixed sampling rhythms, typically without establishing a clear temporal correspondence with the welding stages in the welding process.

[0004] In actual industrial welding scenarios, the surface of the weld pool may exhibit periodic oscillation characteristics. To improve stability or reduce noise, some 3D vision acquisition methods often employ sampling averaging or integration in the time dimension. This results in the output 3D height data focusing more on reflecting overall morphological features, thus weakening the ability to express transient spatial morphological changes during welding. Furthermore, when the welding cycle time changes due to equipment operating status, process switching, or production rhythm adjustments, 3D vision acquisition methods based on fixed sampling parameters may struggle to consistently cover critical stages of the welding process under certain circumstances, affecting the match between the acquired 3D data and the actual welding state. Therefore, this invention proposes a real-time monitoring and defect identification system for battery cell welding based on 3D vision to address the aforementioned problems. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A real-time monitoring and defect identification system for battery cell welding based on 3D vision, comprising:

[0007] The synchronous triggering module is used to start the three-dimensional vision monitoring process based on the welding stage signal output by the welding equipment during the cell welding process, so as to establish a time correspondence between the three-dimensional data acquisition process and the welding stage.

[0008] The 3D acquisition module is used to continuously acquire data of the welding area based on the time correspondence provided by the synchronous triggering module, forming 3D data containing information on the change of the spatial morphology of the weld point over time.

[0009] The dynamic evaluation module is used to calculate the oscillation characterization factor, which reflects the intensity of the change in the spatial morphology of the weld point in the time dimension, and the time matching factor, which reflects the degree of matching between the three-dimensional data acquisition time and the welding stage, based on the three-dimensional data acquired by the three-dimensional acquisition module.

[0010] The closed-loop control module is used to control the acquisition rhythm, acquisition timing, and acquisition method of the three-dimensional acquisition module according to the magnitude of the oscillation characterization factor and the magnitude of the timing matching factor, so that the three-dimensional data used in subsequent processing maintains a corresponding relationship with the welding process.

[0011] The feature construction module is used to construct a set of three-dimensional geometric features of the weld point based on the regulated three-dimensional data after the closed-loop control module has completed the control.

[0012] The defect determination module is used to analyze the welding quality of weld points based on a set of three-dimensional geometric features, and output the type and severity of welding defects.

[0013] In a preferred embodiment, during the cell welding process, the current welding stage is determined based on the welding stage signal, and a three-dimensional visual monitoring process is started at a preset time node corresponding to the determined welding stage, so that the start time of the three-dimensional visual monitoring process is correlated with the welding stage.

[0014] In a preferred embodiment, the three-dimensional acquisition module determines the actual start time of three-dimensional data acquisition based on the correspondence between welding stages and time established by the synchronous triggering module, and continuously acquires data on the welding area at the determined start time to obtain the spatial morphology information of the weld point during the welding process.

[0015] During continuous acquisition, the changes in the spatial morphology of the solder joints are recorded in chronological order. The spatial morphology information of the solder joints obtained by continuous acquisition is indexed and associated with the corresponding time information to form three-dimensional data reflecting the changes in the spatial morphology of the solder joints over time.

[0016] In a preferred embodiment, the generation logic of the oscillation characterization factor is as follows:

[0017] The spatial height values ​​of the solder joints at each acquisition time are obtained in chronological order, and the spatial height values ​​at adjacent acquisition times are subtracted to obtain multiple spatial height changes.

[0018] The squares of each spatial height change are then accumulated. The accumulated result is divided by the number of spatial height changes to obtain the energy accumulation value that characterizes the cumulative degree of change in the spatial morphology of the weld point.

[0019] Based on the positive and negative directions of the spatial height change at adjacent acquisition times, the spatial height change process is divided into an upward change state and a downward change state, and the proportion of the two types of change states in the total spatial height change is calculated.

[0020] The proportions of the two types of change states are multiplied and accumulated with the absolute values ​​of their corresponding spatial height changes, and the sum of the accumulated results is divided by the sum of the absolute values ​​of spatial height changes to obtain the distribution complexity value, which characterizes the distribution complexity of the spatial morphology change states of the weld point.

[0021] The cumulative energy value and the distribution complexity value are multiplied, and the square root of the product is applied to obtain the oscillation characterization factor that reflects the intensity of the comprehensive change in the spatial morphology of the weld point over time.

[0022] In a preferred embodiment, the generation logic of the time-series matching factor is as follows:

[0023] During the 3D data acquisition process, the acquisition time of each frame of 3D data is recorded, and an acquisition trigger sequence is generated based on the acquisition time. The acquisition trigger sequence takes a value of one at the acquisition time point and a value of zero at non-acquisition time points.

[0024] The welding stage signal output by the welding equipment is acquired, and the stage edge time point of the welding stage signal is extracted. A stage edge sequence is generated based on the stage edge time point. The stage edge sequence takes a value of one at the stage edge time point and a value of zero at non-stage edge time points.

[0025] Within a predetermined time range, the acquisition trigger sequence and the stage edge sequence are slidably aligned. For each sliding offset, the corresponding product of the two sequences at that offset is calculated and accumulated to obtain the correlation value. The maximum value of the correlation value is divided by the number of times the acquisition trigger sequence takes the value of one to obtain the alignment correlation value.

[0026] The welding cycle period is determined based on the time interval between adjacent stage edge time points. The phase value is obtained by taking the remainder of the acquisition time of each frame of three-dimensional data with respect to the welding cycle period. All phase values ​​are divided into multiple phase intervals and the proportion of each phase interval is calculated. The maximum proportion is taken as the phase concentration value.

[0027] The alignment correlation value and the phase concentration value are multiplied to obtain the time-series matching factor, which reflects the degree of matching between the 3D data acquisition time and the welding stage.

[0028] In a preferred embodiment, taking the remainder of the acquisition time of each frame of 3D data modulo the welding cycle period means:

[0029] Ignoring the specific welding cycle number to which the acquisition time belongs, only the relative time position of the acquisition time within a single welding cycle is retained to characterize the stage position of the acquisition time within the welding cycle.

[0030] In a preferred embodiment, the control logic of the closed-loop control module is as follows:

[0031] The oscillation characterization factor and timing matching factor output by the dynamic evaluation module are input into the fuzzy logic device and mapped into fuzzy quantities that reflect the strength of changes in the spatial morphology of the solder joint and the degree of timing matching of the three-dimensional data acquisition, respectively.

[0032] Based on the pre-established fuzzy rules, the fuzzy quantities corresponding to the oscillation characterization factor and the fuzzy quantities corresponding to the time sequence matching factor are combined and inferred to generate the fuzzy inference result of the acquisition control intensity.

[0033] The fuzzy inference results are clarified to obtain the control quantities used to regulate the 3D acquisition module;

[0034] The acquisition rhythm of the 3D acquisition module is adjusted based on the control variables to change the time interval between adjacent acquisitions.

[0035] The acquisition timing of the 3D acquisition module is adjusted based on the control variables to change the correspondence between the acquisition start time and the welding stage.

[0036] The acquisition method of the 3D acquisition module is adjusted based on the control variables to change the combination of acquisition parameters used when the 3D data is generated.

[0037] In a preferred embodiment, the closed-loop logic of the closed-loop control module is as follows:

[0038] During the cell welding production process, multiple sets of historical three-dimensional data of completed welding are collected, and the corresponding welding quality judgment results are obtained simultaneously.

[0039] The corresponding oscillation characterization factor and time-series matching factor are calculated based on historical three-dimensional data, and the calculation results are correlated with the welding quality judgment results.

[0040] Based on the value ranges of the oscillation characterization factor and the timing matching factor, the historical welding process is divided into multiple state combinations;

[0041] Based on the welding quality judgment results under different state combinations, the correspondence between each state combination and the acquisition rhythm, acquisition sequence, and acquisition method control direction is determined, and the determined correspondence is stored as fuzzy rules of the fuzzy logic unit.

[0042] In a preferred embodiment, after the closed-loop control module adjusts the subsequent three-dimensional data acquisition parameters based on the oscillation characterization factor and the timing matching factor, it acquires the three-dimensional data formed under the adjusted acquisition rhythm, acquisition timing and acquisition method.

[0043] Spatial alignment and data processing are performed on the acquired 3D data to determine the effective spatial area corresponding to the solder joint;

[0044] Based on the effective spatial region, extract the spatial height, spatial contour and spatial distribution information of the weld points;

[0045] The extracted spatial height, spatial contour, and spatial distribution information are combined to construct a set of three-dimensional geometric features of the weld points for subsequent welding quality analysis.

[0046] In a preferred embodiment, the defect determination module determines the welding quality of the weld point based on one of a pre-trained threshold discrimination model, support vector machine determination model, or neural network determination model.

[0047] The technical effects and advantages of this invention are as follows:

[0048] This invention initiates a 3D visual monitoring process during cell welding based on welding stage signals output by the welding equipment via a synchronous trigger module. This establishes a time correspondence between the 3D data acquisition process and the welding stage, binding the start point and continuous process of 3D data acquisition to the key time nodes of the welding stage under the same time reference. This avoids deviations in the monitored object caused by delayed, premature, or time-shifted start-up of the 3D visual monitoring process, which may result in the acquisition of pre- or post-weld segments. Simultaneously, the time correspondence provides a stable temporal constraint for subsequent continuous acquisition, ensuring that the acquired information on the spatial morphology of the weld point changes over time has a clear stage affiliation. This facilitates consistent comparison and tracing of the stage characteristics of the welding process, thereby improving the usability and reliability of 3D data in real-time monitoring scenarios during welding from the source.

[0049] This invention uses a dynamic evaluation module to calculate oscillation characterization factors and timing matching factors based on the 3D data acquired by the 3D acquisition module. A closed-loop control module then regulates the acquisition rhythm, timing, and method of the 3D acquisition module according to the magnitudes of these factors, ensuring that the 3D data used in subsequent processing maintains a correspondence with the welding process. This forms a closed-loop chain of "evaluation-control-reacquisition," allowing the acquisition rhythm to be dynamically adjusted based on the intensity of changes in the weld point's spatial morphology, the acquisition timing to be corrected based on the time correspondence of the welding stages, and the acquisition method to be switched or adjusted based on the parameter combinations in the 3D data formation process. This ensures that even when there are transient fluctuations, rhythm changes, or changes in acquisition conditions during the welding process, 3D data corresponding to the welding process can still be continuously acquired. This reduces the risk of missed, incorrect, or erroneous acquisitions due to fixed acquisition parameters, and improves the stability, continuity, and consistency of real-time monitoring and defect identification in continuous production.

[0050] After the closed-loop control module completes its control, this invention constructs a three-dimensional geometric feature set of the weld joint based on the controlled three-dimensional data. The defect judgment module then analyzes the welding quality of the weld joint based on this three-dimensional geometric feature set, outputting the type and severity of welding defects. This establishes the evaluation of weld joint welding quality on a three-dimensional data foundation with a time-corresponding relationship and optimized by the control module. This makes the spatial geometric information of the weld joint obtained in the feature construction stage more reflective of the true shape and trend of the weld joint, reducing feature deviations introduced by data distortion or stage misalignment. Furthermore, the defect judgment module can output the defect type and severity under a unified three-dimensional geometric feature set input, realizing a structured expression and graded evaluation of the welding quality of the weld joint. This facilitates the formation of clear quality judgment results and handling basis during the production process, improves the objectivity, repeatability, and comparability of defect identification, and provides a stable data foundation for subsequent process improvement and quality management. Attached Figure Description

[0051] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0052] Figure 1 This is a schematic diagram of a real-time monitoring and defect identification system for battery cell welding based on 3D vision, as described in this invention. Detailed Implementation

[0053] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0054] Reference Figure 1 The following examples were obtained:

[0055] Example 1: A real-time monitoring and defect identification system for battery cell welding based on 3D vision, comprising:

[0056] The synchronous trigger module ensures that the 3D visual monitoring process no longer operates independently of the welding process. Instead, it directly uses the welding stage signal output by the welding equipment as the time reference, thus guaranteeing the consistency of the 3D data acquisition process with the welding stage in the time dimension. By establishing a time correspondence between the 3D data acquisition process and the welding stage, the mixing of pre-weld or post-weld data due to uncertain acquisition start times can be avoided, providing an accurate time anchor point basis for subsequent dynamic analysis of weld points.

[0057] Based on the time correspondence established by the synchronous triggering module, the 3D acquisition module continuously acquires data on the welding area, thereby obtaining complete process data on the spatial morphology of the weld joint as it changes over time. This module, through continuous acquisition rather than single-frame acquisition, ensures that the formation, changes, and stabilization of the weld joint during the welding process are fully recorded, providing a reliable data source for the subsequent dynamic evaluation module to analyze the intensity of weld joint morphology changes and the rationality of the acquisition timing.

[0058] The dynamic evaluation module analyzes the 3D data acquired by the 3D acquisition module, calculating the oscillation characterization factor and the time-series matching factor to reflect the dynamic characteristics of the welding process from different dimensions. The oscillation characterization factor characterizes the intensity of changes in the spatial morphology of the weld point over time, avoiding misjudgments of weld point stability based solely on data from a single moment. The time-series matching factor characterizes the degree of matching between the 3D data acquisition time and the welding stage, identifying potential risks arising from inconsistencies between the acquisition timing and the welding cycle, thus providing a quantitative basis for subsequent control.

[0059] The closed-loop control module, based on the oscillation characterization factor and timing matching factor output by the dynamic evaluation module, regulates the acquisition rhythm, acquisition sequence, and acquisition method of the 3D acquisition module, enabling the 3D data acquisition process to dynamically adjust according to changes in the welding state. This module can correct subsequent acquisition parameters when the weld point spatial morphology changes drastically or the acquisition sequence deviates from the welding stage, thereby ensuring that the 3D data used in subsequent processing always corresponds to the actual welding process and improving the reliability of the monitoring results.

[0060] After the closed-loop control module completes its control, the feature construction module processes the controlled 3D data to construct a set of 3D geometric features of the weld joint. By constructing features under the condition that the acquisition parameters have been corrected, feature distortion caused by unreasonable acquisition rhythm or timing can be avoided, ensuring that the constructed 3D geometric features can truly reflect the spatial morphology of the weld joint and provide stable and reliable feature input for subsequent welding quality assessment.

[0061] The defect assessment module analyzes the welding quality of weld points based on the three-dimensional geometric feature set formed by the feature construction module, and outputs the type and severity of welding defects. By systematically assessing the geometric features of weld points, this module can classify, identify, and evaluate the severity of welding defects, transforming welding quality evaluation from subjective experience-based judgment to objective judgment results based on three-dimensional data, thereby providing a basis for welding quality control and process optimization.

[0062] In one embodiment, during the cell welding process, the current welding stage is determined based on the welding stage signal, and a three-dimensional visual monitoring process is initiated at a preset time node corresponding to the determined welding stage, thus establishing a correspondence between the start time of the three-dimensional visual monitoring process and the welding stage. Specifically, the welding stage signal can be divided into three categories: preheating stage, main welding stage, and arc termination stage, and preset time nodes are configured for each stage. For example, the preheating stage is initiated 0.02 seconds after the start of welding, the main welding stage is initiated 1.2 seconds after the start of welding, and the arc termination stage is initiated 1.8 seconds after the start of welding. When a welding stage switch is determined, the preset time node is recalibrated using the edge of the stage switch as a reference point, so that the three-dimensional visual monitoring process can maintain stage alignment even when there are slight jitters in the welding rhythm, avoiding the misinclusion of pre-welding or post-welding segments into the welding data range.

[0063] The 3D acquisition module, based on the correspondence between welding stages and time established by the synchronous triggering module, determines the actual start time of 3D data acquisition. At this determined start time, continuous acquisition of the welding area is performed to obtain spatial morphology information of the weld joint during the welding process. Specifically, the actual start time can be set with a safety delay, such as 0.01 to 0.03 seconds, to avoid saturation interference caused by the initial transient strong light during welding on the 3D visual monitoring process. Simultaneously, the welding area is limited to a spatial window including the overlap boundary of the tab and the connecting piece, such as an observation range of 8 mm long and 5 mm wide, ensuring that the continuously acquired spatial morphology information covers the main areas of weld joint bulges, collapses, and edge burrs. The sampling interval for continuous acquisition can be set to 0.005 to 0.02 seconds, allowing for the formation of a spatial morphology sequence of 50 to 200 frames within approximately one second during the main welding stage, ensuring that the morphology changes of the weld joint during formation and solidification are completely captured.

[0064] During continuous acquisition, the changes in the spatial morphology of the weld joints are recorded sequentially over time. The continuously acquired spatial morphology information is then indexed and associated with the corresponding time information to form three-dimensional data reflecting the changes in the spatial morphology of the weld joints over time. Specifically, a time stamp can be written to each frame of spatial morphology information, and a data organization method with the welding stage as the primary index and the acquisition time as the secondary index can be established. This allows the changes in spatial morphology within the same welding stage to be directly read in time series form. At the same time, key quantities in the spatial morphology information are recorded over time, such as the curves of the weld joint's highest point height, the weld joint boundary contour area, and the estimated weld joint volume. These are used to distinguish between slow changes and transient abrupt changes in subsequent processing. For example, if a sudden change occurs between 0.5 and 0.7 seconds, with the height jumping from 0.1 mm to 0.25 mm, the corresponding frame interval can be directly located in the three-dimensional data using an index, enabling rapid tracing of dynamic anomaly segments. Based on the three-dimensional data formed by the aforementioned index association, the welding stage, preset time node, and actual start time are uniformly written into the record item under the same time base, so that any frame of spatial morphology information can be traced back to its corresponding welding stage and acquisition starting point.

[0065] In one embodiment, the generation logic of the oscillation characterization factor is as follows: obtain the spatial height values ​​of the solder joint at each acquisition time in chronological order, and subtract the spatial height values ​​of adjacent acquisition times to obtain multiple spatial height changes. To ensure consistency of the time order, the acquisition times can be arranged in millisecond increments. For example, 21 acquisition times are obtained in 0.005-second intervals from 0.5 to 0.6 seconds, with corresponding spatial height values ​​of 0.12 mm, 0.14 mm, 0.13 mm, 0.17 mm up to 0.20 mm. By subtracting adjacent values, a sequence of spatial height changes is obtained, such as positive 0.02 mm, negative 0.01 mm, positive 0.04 mm, and positive 0.03 mm. This makes the transient fluctuations of the solder joint's spatial morphology in the time dimension explicit in the form of spatial height changes, avoiding misjudgment of solder joint stability caused by observing only the height of a single frame.

[0066] Each change in spatial height is squared and then summed. The summation result is divided by the number of changes in spatial height to obtain the energy accumulation value, which characterizes the cumulative degree of change in the spatial morphology of the weld joint. For example, squaring a positive 0.02 mm yields 0.0004 square millimeters, squaring a negative 0.01 mm yields 0.0001 square millimeters, and summing all the squared values ​​yields 0.0012 square millimeters. Dividing this by the number of changes in spatial height (20) yields the mean value of 0.00006 square millimeters. This mean value is the basic quantitative result of the energy accumulation value, used to reflect the overall intensity of the change in spatial height within the time window. By squaring, larger changes in spatial height can have a stronger impact on the summation result, making the energy accumulation value more sensitive to drastic fluctuations in the spatial morphology of the weld joint, thus providing a reliable amplitude accumulation scale for subsequent change state analysis.

[0067] Based on the positive and negative directions of spatial height changes at adjacent acquisition times, the spatial height change process is divided into rising and falling states, and the proportion of each state in all spatial height changes is statistically analyzed. For example, in twenty spatial height changes, the rising state occurs twelve times and the falling state occurs eight times, so the proportion of the rising state is 0.6% and the proportion of the falling state is 0.4%. During the weld formation process, if the rising and falling states alternate frequently, it often corresponds to periodic oscillations or disturbances on the surface of the molten pool. In this case, the energy accumulation value alone may not be able to distinguish between "unidirectional slow rise" and "high-frequency reciprocating fluctuations". Therefore, by introducing information on the directional dimension through the proportion of the rising and falling states, the structural characteristics of the spatial height change process can be further revealed.

[0068] The distribution complexity value, which characterizes the complexity of the spatial morphology changes of the solder joint, is obtained by multiplying and summing the proportions of the two types of change states with the absolute values ​​of their corresponding spatial height changes. The sum of these sums is then divided by the total sum of the absolute values ​​of spatial height changes. The energy accumulation value is then multiplied by the distribution complexity value, and the square root of the product is applied to obtain the oscillation characterization factor, which reflects the overall intensity of the changes in the spatial morphology of the solder joint over time. For example, the proportion of rising change states (0.6%) is multiplied by the absolute value of the spatial height change corresponding to each rising change state and then summed; the proportion of falling change states (0.4%) is multiplied by the absolute value of the spatial height change corresponding to each falling change state. The absolute values ​​of the changes are multiplied and accumulated sequentially. The sum of these two values ​​is then divided by the total absolute value of all spatial height changes, which is 0.48 mm, to obtain a distribution complexity value of 0.52. The cumulative energy value of 0.00006 square millimeters is then multiplied by the distribution complexity value of 0.52 to obtain 0.0000312 square millimeters. The square root of this product is then applied to obtain 0.0056 millimeters, which is the oscillation characterization factor. By using this calculation method that integrates the degree of amplitude accumulation with the distribution complexity of the change state, the oscillation characterization factor can simultaneously characterize "how drastic" and "how complex" the spatial morphology changes of the weld point are, making it more suitable for identifying risk scenarios in the welding process where millisecond-level oscillations are masked by time averaging.

[0069] In one embodiment, the generation logic of the timing matching factor is as follows: during the 3D data acquisition process, the acquisition time of each frame of 3D data is recorded, and an acquisition trigger sequence is generated based on the acquisition time. The acquisition trigger sequence takes a value of one at the acquisition time point and a value of zero at non-acquisition time points. To ensure comparability with the calculation of the oscillation characterization factor, the statistics and calculation of the acquisition time can be limited to the same predetermined time range as the oscillation characterization factor. For example, a time window with a length of 0.2 seconds is selected within the main welding stage, and the acquisition time point of each frame of 3D data is recorded within this time window, such as 0.01 seconds, 0.02 seconds, 0.04 seconds, and 0.05 seconds. Then, the sequence position after being discretized with a finer time granularity is marked with a value of one, and the other positions are marked with a value of zero, so that the acquisition trigger sequence can objectively reflect whether the acquisition is continuous within the predetermined time range, whether there is any missed acquisition, and whether the acquisition is premature or delayed.

[0070] The welding stage signal output by the welding equipment is acquired, and the stage edge time points of the welding stage signal are extracted. A stage edge sequence is generated based on the stage edge time points. The stage edge sequence takes a value of 1 at the stage edge time points and a value of 0 at non-stage edge time points. For example, within the same predetermined time range, the welding stage signal may have an edge time point of 0.00 seconds when switching from the preheating stage to the main welding stage, and an edge time point of 0.20 seconds when switching from the main welding stage to the arc termination stage. These stage edge time points are mapped to a time scale consistent with the acquisition trigger sequence and marked as 1. The other positions are marked as 0, so that the stage edge sequence becomes a time reference for the changes in the welding stage, thereby providing a unified comparison benchmark for subsequent judgment on whether the acquisition trigger sequence is truly aligned with the changes in the welding stage.

[0071] Within a predetermined time range, the acquisition trigger sequence and the stage edge sequence are slidably aligned. For each sliding offset, the two sequences are multiplied in the same position at that offset and the results are accumulated to obtain a correlation value. The maximum correlation value is then divided by the number of times the acquisition trigger sequence takes the value of 1 to obtain the alignment correlation value. For example, the acquisition trigger sequence is moved forward by multiple offsets of 0.005 seconds, 0.01 seconds, and 0.015 seconds. After each move, the moved acquisition trigger sequence and the stage edge sequence are multiplied in the same position and the results are accumulated. If the correlation value is the largest at an offset of 0.01 seconds, and this maximum correlation value is 8, while the acquisition trigger sequence takes the value of 1 ten times, then the alignment correlation value is 0.8. Through this sliding alignment and normalization process, even when there are slight pauses or trigger delay fluctuations in the welding cycle, the optimal alignment between the acquisition trigger and the stage edge can still be found globally, avoiding a one-sided judgment on timing matching based solely on a single time deviation.

[0072] The welding cycle period is determined based on the time interval between adjacent stage edge time points. The phase value is obtained by taking the remainder of the acquisition time of each frame of 3D data divided by the welding cycle period. All phase values ​​are divided into multiple phase intervals, and the proportion of each phase interval is calculated. The largest proportion is taken as the phase concentration value. The alignment correlation value and the phase concentration value are multiplied to obtain a temporal matching factor reflecting the degree of matching between the 3D data acquisition time and the welding stage. Taking the remainder of the acquisition time of each frame of 3D data divided by the welding cycle period means that, ignoring the specific welding cycle number to which the acquisition time belongs, only the relative time position of the acquisition time within a single welding cycle is retained, which is used to characterize the stage position of the acquisition time within the welding cycle. For example, if the time interval between adjacent stage edge time points is 0.20 seconds, then the welding cycle period is 0.20 seconds. The relative time position of the acquisition time of 0.23 seconds within a single welding cycle is 0.03 seconds, and the relative time position of the acquisition time of 0.39 seconds within a single welding cycle is 0.19 seconds. The relative time positions are divided into ten phase intervals and their proportions are statistically analyzed. If the maximum proportion is 0.7, then the phase concentration value is 0.7. Combined with the aforementioned alignment correlation value of 0.8, the product of the two is 0.56, which is the timing matching factor. By simultaneously using the alignment correlation value to characterize the alignment degree between the acquisition trigger and the stage edge within the same predetermined time range as the oscillation characterization factor, and using the phase concentration value to characterize the stability of the stage position of the acquisition within the welding cycle, the timing matching factor can more comprehensively reflect the matching degree between the acquisition time and the welding stage. It is suitable for identifying the stage misalignment risk caused by fixed time window sampling under welding cycle fluctuation conditions.

[0073] It should be noted that the time-series matching factor and the oscillation characterization factor use the same predetermined time range to ensure that both factors characterize the welding process under the same time reference, thus making them comparable and consistent in subsequent analysis and control. By simultaneously calculating the intensity of changes in the spatial morphology of the weld point in the time dimension and the degree of matching between the three-dimensional data acquisition time and the welding stage within the same predetermined time range, it can be ensured that the dynamic characteristics of the spatial morphology reflected by the oscillation characterization factor and the acquisition time sequence characteristics reflected by the time-series matching factor originate from the same welding process. This avoids situations where one factor corresponds to the early stage of welding and the other to the middle and late stages of welding due to inconsistent time windows, thereby improving the accuracy and stability of joint judgment and closed-loop control based on the two types of factors.

[0074] In one embodiment, the closed-loop control module is used to regulate the acquisition rhythm, acquisition sequence, and acquisition method of the three-dimensional acquisition module according to the magnitude of the oscillation characterization factor and the magnitude of the timing matching factor, so that the three-dimensional data used in subsequent processing maintains a correspondence with the welding process; and inputs the oscillation characterization factor and timing matching factor output by the dynamic evaluation module into the fuzzy logic device, and maps them into fuzzy quantities that reflect the strength of the change in the spatial morphology of the weld point and the degree of timing matching of the three-dimensional data acquisition, respectively. In practice, the oscillation characterization factor and the temporal matching factor are first normalized, mapping the oscillation characterization factor to the interval of zero to one, and the temporal matching factor to the interval of zero to one. For example, the oscillation characterization factor is linearly compressed to zero to one in the range of 0.0 to 0.1 millimeters, while the temporal matching factor remains unchanged in the range of zero to one. After normalization, three fuzzy quantity levels are constructed respectively: the oscillation characterization factor corresponds to "weak", "medium" and "strong", and the temporal matching factor corresponds to "poor", "medium" and "excellent". Each level is assigned a value using a triangular membership function. The triangular membership function determines the membership degree change with three nodes. For example, "weak" uses 0, 0, and 0.5 as nodes, "medium" uses 0.25, 0.5, and 0.75 as nodes, and "strong" uses 0.5, 1, and 1 as nodes, thereby ensuring that each input value can obtain a clear membership degree value and enter the subsequent inference.

[0075] Based on pre-established fuzzy rules, the fuzzy quantities corresponding to the oscillation characterization factor and the fuzzy quantities corresponding to the time-series matching factor are combined and inferred to generate the fuzzy inference result of the acquisition control intensity. The inference operation can adopt the maximum-minimum inference method, where the "AND" relationship is achieved by taking the minimum value, the "OR" relationship is achieved by taking the maximum value, and the rule aggregation is achieved by taking the maximum value, ensuring that the inference process is reproducible and the amount of computation is controllable. The fuzzy rules can be partitioned according to the acquisition risk, for example, "if the oscillation characterization factor is strong and the time-series matching factor is poor, the acquisition control intensity is high", "if the oscillation characterization factor is strong and the time-series matching factor is medium, the acquisition control intensity is relatively high", "if the oscillation characterization factor is medium and the time-series matching factor is poor, the acquisition control intensity is relatively high", and "if the oscillation characterization factor is weak and the time-series matching factor is excellent, the acquisition control intensity is low". For example, if the normalized oscillation characterization factor is 0.8 and the time-series matching factor is 0.3, then the membership degree of "strong" is 0.6 and the membership degree of "poor" is 0.4. According to the minimum value rule, the triggering intensity of this rule is 0.4. Then, the triggering results of other rules are aggregated according to the maximum value to form the fuzzy inference result of the acquisition control intensity, so that the control decision is simultaneously constrained by the strength of spatial morphological changes and the quality of time-series matching.

[0076] The fuzzy inference results are then processed to obtain the control quantity used to regulate the 3D acquisition module. The declarative processing can employ the centroid method, dividing the fuzzy output of the acquisition regulation intensity into three levels: "low," "medium," and "high." Similarly, a triangular membership function is used to define the output membership degree distribution; for example, "low" covers 0 to 0.4, "medium" covers 0.2 to 0.8, and "high" covers 0.6 to 1. The aggregated fuzzy output is then plotted on the output domain from 0 to 1, forming a truncated shape. The centroid position is obtained by summing the products of the output domain position and the membership degree, and then dividing by the sum of the membership degrees. This centroid position is the control quantity. For example, if the truncated height of "high" after aggregation is 0.4 and the truncated height of "medium" is 0.2, the control quantity obtained by the centroid method will fall around 0.7. This transforms the fuzzy inference results into a directly executable continuous value, ensuring that the regulation action is adjustable rather than only having discrete levels.

[0077] The acquisition rhythm of the 3D acquisition module is adjusted based on control variables to change the time interval between adjacent acquisitions; the acquisition timing of the 3D acquisition module is adjusted based on control variables to change the correspondence between the acquisition start time and the welding stage; the acquisition method of the 3D acquisition module is adjusted based on control variables to change the combination of acquisition parameters used when forming 3D data. To ensure a closed loop with the results of previous inference, the control quantity can be mapped to three specific adjustment quantities: First, the acquisition rhythm adjustment quantity is positively correlated with the control quantity. For example, when the control quantity is 0.7, the time interval between adjacent acquisitions is shortened from 0.01 seconds to 0.006 seconds, allowing for higher acquisition density in segments with rapid changes in spatial morphology. Second, the acquisition timing adjustment quantity is negatively correlated with or segmentally correlated with the control quantity. For example, when the control quantity exceeds 0.6, the acquisition start time is finely adjusted by 0.01 seconds towards the forefront of the critical time interval of the welding stage to offset the stage misalignment caused by the beat delay. Third, the acquisition mode adjustment quantity is based on the control quantity trigger threshold. For example, when the control quantity exceeds 0.8, a high dynamic range acquisition parameter combination is switched; when the control quantity is between 0.4 and 0.8, a high frame rate acquisition parameter combination is switched; and when the control quantity is below 0.4, the conventional acquisition parameter combination is maintained. This ensures that the 3D data used in subsequent processing maintains a correspondence with the welding process and provides a more reliable data foundation for the next round of dynamic evaluation module to continue outputting oscillation characterization factors and timing matching factors.

[0078] In one embodiment, the closed-loop logic of the closed-loop control module is as follows: During the cell welding production process, multiple sets of historical three-dimensional data of completed welding are collected, and the corresponding welding quality judgment results are obtained simultaneously. To ensure that the historical data covers typical working conditions, historical three-dimensional data formed under different welding beats and different welding energy combinations can be selected. For example, in one hundred sets of welding samples, three types of working conditions are included, with welding beats of 0.15 seconds, 0.20 seconds, and 0.25 seconds, respectively, and the welding quality judgment results are recorded simultaneously as qualified, minor defects, and serious defects. The historical three-dimensional data can include information on the change of the spatial morphology of the weld point over time, and is truncated within a predetermined time range after the welding stage is aligned, so that each set of historical three-dimensional data and the welding quality judgment results are consistent in time scale, thereby providing a unified data entry and comparison basis for subsequent calculation of corresponding indicators and establishment of rules based on historical three-dimensional data.

[0079] Based on historical 3D data, the corresponding oscillation characterization factor and temporal matching factor are calculated, and the calculation results are correlated with the welding quality judgment results. In the specific implementation, for each set of historical 3D data, the oscillation characterization factor is first calculated according to the same predetermined time range, and then the temporal matching factor is calculated according to the same predetermined time range, so that the values ​​of the two types of factors correspond to the same key stage of the welding process. Subsequently, an association record table is established, and the oscillation characterization factor value, temporal matching factor value, and welding quality judgment result of each set of samples are written into the same record. For example, one set of samples has an oscillation characterization factor of 0.05 mm, a temporal matching factor of 0.56, and a welding quality judgment result of minor defects, while another set of samples has an oscillation characterization factor of 0.09 mm, a temporal matching factor of 0.32, and a welding quality judgment result of severe defects. Through this one-to-one correspondence, the subsequent state combination division can directly use the welding quality judgment result as the basis for rule construction.

[0080] Based on the value ranges of the oscillation characterization factor and the timing matching factor, the historical welding process is divided into multiple state combinations. In specific implementation, the oscillation characterization factor can be divided into low, medium, and high ranges according to its numerical value. For example, the low range is 0.0 to 0.03 mm, the medium range is 0.03 to 0.07 mm, and the high range is 0.07 to 0.12 mm. At the same time, the timing matching factor can be divided into poor, medium, and excellent ranges according to its numerical value. For example, the poor range is 0 to 0.4, the medium range is 0.4 to 0.7, and the excellent range is 0.7 to 1. Based on this, nine state combinations are formed, and each associated record is mapped to the corresponding state combination. For example, records where the oscillation characterization factor is in the high range and the timing matching factor is in the poor range are grouped into the same state combination. This ensures that the state combination reflects both the strength of the change in the spatial morphology of the weld point and the degree of timing matching of the three-dimensional data acquisition, providing a structured grouping basis for determining the subsequent control direction.

[0081] Based on the welding quality judgment results under different state combinations, the correspondence between each state combination and the acquisition rhythm, acquisition timing, and acquisition mode control direction is determined. This determined correspondence is stored as fuzzy rules for a fuzzy logic unit. In specific implementation, the distribution of welding quality judgment results for each state combination can be statistically analyzed. For example, in a state combination with a high oscillation characterization factor and a poor timing matching factor, if the proportion of severe defects reaches 70%, then the acquisition rhythm control direction for that state combination is determined to increase the acquisition density, the acquisition timing control direction is determined to correct the acquisition start time towards the forefront of the critical time interval of the welding stage, and the acquisition mode control direction is determined to cut... The acquisition parameter combination is then switched to one that emphasizes anti-interference or high dynamic range. For example, in a state combination where the oscillation characterization factor is in the low range and the timing matching factor is in the high range, if the qualified rate reaches 90%, the acquisition rhythm control direction is determined to maintain the conventional acquisition density, the acquisition timing control direction is determined to maintain the current correspondence, and the acquisition method control direction is determined to maintain the conventional acquisition parameter combination. The correspondence between each state combination and the three control directions is written into the rule base to form fuzzy rules that can be directly called by the fuzzy logic device. This enables the subsequent closed-loop control module to output stable and traceable control decisions based on the combination of the oscillation characterization factor and the timing matching factor during the actual welding process.

[0082] Acquisition rhythm controls the sampling density of 3D data in the time dimension, acquisition timing controls the correspondence between the start time of 3D data acquisition and the welding stage, and acquisition method controls the parameter combination and imaging configuration used in a single 3D data acquisition process. The three factors act on three different dimensions: time density, time alignment, and imaging configuration, and are independent of each other yet work together.

[0083] In one embodiment, after the closed-loop control module adjusts the subsequent 3D data acquisition parameters based on the oscillation characterization factor and the timing matching factor, it acquires 3D data formed under the adjusted acquisition rhythm, acquisition timing, and acquisition mode. The acquisition rhythm determines the time interval between adjacent acquisitions, the acquisition timing determines the correspondence between the acquisition start time and the welding stage, and the acquisition mode determines the combination of acquisition parameters used when forming the 3D data. For example, when the subsequent acquisition rhythm is adjusted from 0.01 seconds to 0.006 seconds, and the acquisition start time is shifted forward by 0.01 seconds relative to the critical time interval of the welding stage, while the acquisition mode is switched to a higher dynamic range parameter combination, the resulting set of 3D data can contain information on the spatial morphology of the weld points changing over time across 150 consecutive frames within the main welding stage. This allows subsequent welding quality analysis to be based on continuous observations after stage alignment, rather than relying on the accidental state of a single frame.

[0084] The acquired 3D data undergoes spatial alignment and data processing to determine the effective spatial region corresponding to the weld point. Spatial alignment can employ rigid registration based on the welding fixture positioning reference, ensuring comparability of 3D data from different frames under the same spatial coordinates. Data processing includes removing outliers, filling small gaps, and maintaining the geometric continuity of the weld point edges. The effective spatial region can be expanded outward from the weld point center to form a fixed window, for example, a spatial range of 8 mm in the length direction, 5 mm in the width direction, and 0.5 mm in the height direction along the tab. Within this range, the main weld point region is further determined based on height protrusions and edge gradients, thereby avoiding the inclusion of spatter isolated points or fixture background in the weld point analysis. For example, after spatial alignment, the positioning deviation of the weld point center can be controlled within 0.02 mm, enabling stable superposition of weld point contours from different frames during subsequent feature extraction and reducing morphological drift errors caused by micro-vibrations of the viewing angle.

[0085] Based on the effective spatial region, the spatial height, spatial contour, and spatial distribution information of the weld point are extracted. These extracted spatial height, spatial contour, and spatial distribution information are then combined to construct a three-dimensional geometric feature set for subsequent welding quality analysis. Spatial height may include the highest point height, average height, collapse depth, and height fluctuation range; spatial contour may include contour closure, contour area, aspect ratio, and edge roughness; and spatial distribution information may include the volume distribution centroid, volume concentration, and symmetry. For example, a weld point within the effective spatial region may yield a highest point height of 0.22 mm, a collapse depth of 0.06 mm, a contour area of ​​1.8 square millimeters, a contour closure of 0.93, a volume distribution centroid offset from the design center of 0.12 mm, and a volume concentration of 0.78. These quantities are then combined in a predetermined order to form a three-dimensional geometric feature set for the weld point. This ensures that the geometric fullness, offset, and collapse risk of the weld point can be uniformly expressed within the same feature set, providing a consistent input format for subsequent model judgment.

[0086] The defect assessment module determines the welding quality of weld joints based on one of the following pre-trained models: a threshold discrimination model, a support vector machine (SVM) model, or a neural network model. In the threshold discrimination model, key features in the weld joint's three-dimensional geometric feature set are compared item by item with preset threshold intervals. For example, a height below 0.15 mm indicates insufficient energy risk; a volume distribution centroid offset exceeding 0.2 mm indicates weld misalignment risk; and a collapse depth exceeding 0.08 mm indicates collapse risk. The severity of the defect is then output as slight, moderate, or severe based on the number and magnitude of the exceedances. In the SVM model, the weld joint's three-dimensional geometric feature set is mapped to a discrimination space. A pre-trained classification hyperplane is used to classify defect types, such as weld misalignment, over-welding, and cold welds, and the distance from the sample to the classification boundary is used as a severity measure; a larger distance indicates a more obvious and severe defect. In the neural network judgment model, the three-dimensional geometric feature set of the weld point is input into a pre-trained multi-layer structure. The coupling relationship between features is learned through nonlinear mapping. For example, the combination pattern of spatial height and spatial distribution is associated with the internal defect proxy features, and the probability distribution of the defect type is output. The category corresponding to the highest probability is taken as the defect type, and the severity of the defect is characterized by the highest probability value or the concentration of the probability distribution. For example, when the probability of weld deviation in the output corresponding to the weld point feature set is 0.85 and the probability of collapse is 0.12, the defect type can be output as weld deviation and the severity can be given as moderate to severe. Thus, the interpretability of rules, boundary discrimination ability and complex pattern recognition ability are taken into account in the same judgment process.

[0087] Threshold discrimination model, support vector machine discrimination model and neural network discrimination model are all mature technical routes widely used in the fields of welding quality inspection, industrial vision inspection and quality classification. This invention does not improve the structure of the model itself, but introduces it into the welding quality judgment scenario of weld point based on three-dimensional geometric feature set, and applies it in combination with the highly consistent three-dimensional data obtained by the previous closed-loop control, thereby improving the reliability of the judgment of welding defect type and defect severity.

[0088] The threshold discrimination model is a rule-based judgment model. Its training process mainly relies on the statistical analysis results of historical welding samples. By manually or automatically judging the quality of completed welding samples, the distribution of values ​​of each three-dimensional geometric feature under different welding quality levels is statistically analyzed to determine the threshold range used to distinguish normal welds from defective welds. In one embodiment, 500 historical samples of completed welding can be selected, and the set of three-dimensional geometric features of the weld point can be extracted for each sample, while the welding quality judgment result is obtained simultaneously. For example, in qualified weld samples, the maximum spatial height is concentrated between 0.18 mm and 0.25 mm, while in poor weld samples, the maximum spatial height is mostly below 0.15 mm; in over-welded samples, the collapse depth is mostly greater than 0.08 mm. Based on the above statistical results, a spatial height below 0.15 mm can be set as the poor weld threshold, a collapse depth greater than 0.08 mm can be set as the over-welded threshold, and a volume distribution centroid offset exceeding 0.2 mm can be set as the weld deviation threshold. After setting the thresholds in this way, a threshold discrimination model can be formed for rapid and interpretable defect judgment of subsequent weld quality.

[0089] Support Vector Machine (SVM) classification models belong to supervised learning. Their training process is based on labeled sample data, and by constructing optimal classification boundaries in the feature space, they distinguish between different types of welding defects. This model is widely used in industrial scenarios with limited sample sizes and high feature dimensionality. In one embodiment, the three-dimensional geometric feature set of each weld point can be represented as a multi-dimensional feature vector, including features such as maximum height, collapse depth, contour area, contour closure, volume distribution centroid offset, and volume concentration. Each feature vector is labeled with a corresponding welding quality category, such as qualified, off-center weld, incomplete weld, or over-welded. During the training phase, three hundred sets of feature vectors are input into the SVM model, and the classification boundary is iteratively optimized to separate different welding quality categories as much as possible in the feature space. For example, after training, off-center weld samples are mainly distributed in areas with large volume distribution centroid offsets, while incomplete weld samples are mainly distributed in areas with low spatial heights. The model uses support vectors to determine the boundaries between these areas. During the determination phase, a new set of three-dimensional geometric features of the weld point is mapped into the feature space. The type of welding defect is determined by its position relative to the classification boundary, and the distance from the sample to the classification boundary is used as a measure of the severity of the defect.

[0090] The neural network judgment model belongs to the deep learning model and can automatically learn the complex relationships between three-dimensional geometric features through a multi-layer nonlinear structure. It has strong discrimination ability in scenarios with complex welding defect morphologies and obvious coupling between features. In one embodiment, the neural network judgment model can adopt a structure containing convolutional layers. The convolutional layers are used to extract local features from the spatial distribution of the three-dimensional geometric features of the weld point. For example, convolution operations are performed on the weld point height distribution map, contour projection map, or volume distribution map to extract local protrusions, collapsed edges, or asymmetric structures. For example, the first convolutional layer can use multiple convolutional kernels to perform local scanning of the weld point height distribution map and extract height change gradient features; the second convolutional layer further combines the feature maps output by the previous layer to extract higher-level morphological patterns; subsequently, pooling operations are used to reduce the feature dimensionality and enhance robustness.

[0091] During the training phase, the 3D geometric feature representations of a large number of labeled weld point samples are input into the neural network model, with welding defect types used as supervisory signals. For example, the spatial height distribution of weld points is converted into a 2D height map, and the contour information is converted into an edge map. These maps are then fed as multi-channel inputs into convolutional layers. After multiple convolutions and nonlinear mappings, the probability values ​​of each welding defect type are output through fully connected layers. During training, the convolution kernel parameters and connection weights are continuously adjusted to gradually approximate the model's output probability distribution to the true welding quality label. After training, the model can determine the new set of 3D geometric features of weld points, outputting the probability distribution of welding defect types. The category corresponding to the highest probability is taken as the welding defect type, while the magnitude of the probability value or the concentration of the probability distribution is used as a quantitative indicator of defect severity.

[0092] The above-mentioned models or function formulas are all dimensionless and numerical calculations. The models or function formulas are obtained by software simulation based on a large amount of collected data to obtain the most recent real situation. The preset parameters in the models or function formulas are set by those skilled in the art according to the actual situation.

[0093] 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.

[0094] 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.

[0095] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0096] 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 real-time monitoring and defect identification system for battery cell welding based on 3D vision, characterized in that, include: The synchronous triggering module is used to start the three-dimensional vision monitoring process based on the welding stage signal output by the welding equipment during the cell welding process, so as to establish a time correspondence between the three-dimensional data acquisition process and the welding stage. The 3D acquisition module is used to continuously acquire data of the welding area based on the time correspondence provided by the synchronous triggering module, forming 3D data containing information on the change of the spatial morphology of the weld point over time. The dynamic evaluation module is used to calculate the oscillation characterization factor, which reflects the intensity of the change in the spatial morphology of the weld point in the time dimension, and the time matching factor, which reflects the degree of matching between the three-dimensional data acquisition time and the welding stage, based on the three-dimensional data acquired by the three-dimensional acquisition module. The closed-loop control module is used to control the acquisition rhythm, acquisition timing, and acquisition method of the three-dimensional acquisition module according to the magnitude of the oscillation characterization factor and the magnitude of the timing matching factor, so that the three-dimensional data acquired after control maintains a time correspondence with the welding process. The feature construction module is used to construct a set of three-dimensional geometric features of the weld point based on the regulated three-dimensional data after the closed-loop control module has completed the control. The defect determination module is used to analyze the welding quality of weld points based on a three-dimensional geometric feature set, and output the welding defect type and defect severity. The generation logic of the oscillation characterization factor is as follows: The spatial height values ​​of the solder joints at each acquisition time are obtained in chronological order, and the spatial height values ​​at adjacent acquisition times are subtracted to obtain multiple spatial height changes. The squares of each spatial height change are then accumulated. The accumulated result is divided by the number of spatial height changes to obtain the energy accumulation value that characterizes the cumulative degree of change in the spatial morphology of the weld point. Based on the positive and negative directions of the spatial height change at adjacent acquisition times, the spatial height change process is divided into an upward change state and a downward change state, and the proportion of the two types of change states in the total spatial height change is calculated. The proportions of the two types of change states are multiplied and accumulated with the absolute values ​​of their corresponding spatial height changes, and the sum of the accumulated results is divided by the sum of the absolute values ​​of spatial height changes to obtain the distribution complexity value, which characterizes the distribution complexity of the spatial morphology change states of the weld point. The cumulative energy value and the distribution complexity value are multiplied, and the square root of the product is applied to obtain the oscillation characterization factor that reflects the intensity of the comprehensive change in the spatial morphology of the weld point over time. The generation logic of the time-series matching factor is as follows: During the 3D data acquisition process, the acquisition time of each frame of 3D data is recorded, and an acquisition trigger sequence is generated based on the acquisition time. The acquisition trigger sequence takes a value of one at the acquisition time point and a value of zero at non-acquisition time points. The welding stage signal output by the welding equipment is acquired, and the stage edge time point of the welding stage signal is extracted. A stage edge sequence is generated based on the stage edge time point. The stage edge sequence takes a value of one at the stage edge time point and a value of zero at non-stage edge time points. Within a predetermined time range, the acquisition trigger sequence and the stage edge sequence are slidably aligned. For each sliding offset, the corresponding product of the two sequences at that offset is calculated and accumulated to obtain the correlation value. The maximum value of the correlation value is divided by the number of times the acquisition trigger sequence takes the value of one to obtain the alignment correlation value. The welding cycle period is determined based on the time interval between adjacent stage edge time points. The phase value is obtained by taking the remainder of the acquisition time of each frame of three-dimensional data with respect to the welding cycle period. All phase values ​​are divided into multiple phase intervals and the proportion of each phase interval is calculated. The maximum proportion is taken as the phase concentration value. The alignment correlation value and the phase concentration value are multiplied to obtain the time-series matching factor, which reflects the degree of matching between the 3D data acquisition time and the welding stage.

2. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 1, characterized in that, The control logic of the closed-loop control module is as follows: The oscillation characterization factor and timing matching factor output by the dynamic evaluation module are input into the fuzzy logic device and mapped into fuzzy quantities that reflect the strength of changes in the spatial morphology of the solder joint and the degree of timing matching of the three-dimensional data acquisition, respectively. Based on the pre-established fuzzy rules, the fuzzy quantities corresponding to the oscillation characterization factor and the fuzzy quantities corresponding to the time sequence matching factor are combined and inferred to generate the fuzzy inference result of the acquisition control intensity. The fuzzy inference results are clarified to obtain the control quantities used to regulate the 3D acquisition module; The acquisition rhythm of the 3D acquisition module is adjusted based on the control variables to change the time interval between adjacent acquisitions. The acquisition timing of the 3D acquisition module is adjusted based on the control variables to change the time correspondence between the acquisition start time and the welding stage. The acquisition method of the 3D acquisition module is adjusted based on the control variables to change the combination of acquisition parameters used when the 3D data is generated.

3. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 1, characterized in that, The 3D acquisition module determines the actual start time of 3D data acquisition based on the correspondence between welding stages and time established by the synchronous triggering module. At the determined start time, the welding area is continuously acquired to obtain the spatial morphology information of the weld points during the welding process. During continuous data acquisition, the changes in the spatial morphology of the solder joints are recorded in chronological order. The spatial morphology information of the solder joints obtained through continuous acquisition is indexed and associated with the corresponding time information to form three-dimensional data reflecting the changes in the spatial morphology of the solder joints over time.

4. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 3, characterized in that, Taking the remainder of the acquisition time of each frame of 3D data divided by the welding cycle period means: Ignoring the specific welding cycle number to which the acquisition time belongs, only the relative time position of the acquisition time within a single welding cycle is retained to characterize the stage position of the acquisition time within the welding cycle.

5. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 4, characterized in that, The closed-loop logic of the closed-loop control module is as follows: During the cell welding production process, multiple sets of historical three-dimensional data of completed welding are collected, and the corresponding welding quality judgment results are obtained simultaneously. The corresponding oscillation characterization factor and time-series matching factor are calculated based on historical three-dimensional data, and the calculation results are correlated with the welding quality judgment results. Based on the value ranges of the oscillation characterization factor and the timing matching factor, the historical welding process is divided into multiple state combinations; Based on the welding quality judgment results under different state combinations, the correspondence between each state combination and the acquisition rhythm, acquisition sequence, and acquisition method control direction is determined, and the determined correspondence is stored as fuzzy rules of the fuzzy logic unit.

6. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 5, characterized in that, After the closed-loop control module adjusts the subsequent three-dimensional data acquisition parameters based on the oscillation characterization factor and the timing matching factor, it acquires the three-dimensional data formed under the adjusted acquisition rhythm, acquisition timing and acquisition mode. Spatial alignment and data processing are performed on the acquired 3D data to determine the effective spatial area corresponding to the solder joint; Based on the effective spatial region, extract the spatial height, spatial contour and spatial distribution information of the weld points; The extracted spatial height, spatial contour, and spatial distribution information are combined to construct a set of three-dimensional geometric features of the weld points for subsequent welding quality analysis.

7. The real-time monitoring and defect identification system for battery cell welding based on 3D vision according to claim 6, characterized in that, The defect determination module determines the welding quality of weld points based on one of the following: a pre-trained threshold discrimination model, a support vector machine determination model, or a neural network determination model.

Images