Battery tab welding quality inspection system and method

Abstract

This invention belongs to the field of battery testing technology. It discloses a battery tab welding quality inspection system and method. The method includes: synchronously acquiring X-ray images and electrical performance data using a preset global clock signal to construct a time-series dataset sorted by timestamps; generating a geometric transformation matrix between adjacent X-ray image frames using an initial feature point tracking algorithm; aligning the spatial positions of the X-ray image time-series data in the time-series dataset based on the geometric transformation matrix to obtain aligned X-ray image time-series data; and extracting time-series features from the electrical performance time-series data in the time-series dataset to generate an electrical performance time-series feature vector. The overall solution significantly improves the accuracy of dynamic defect detection, solves the problem of missed detection in the traditional single-frame analysis mode, and provides full-cycle, multi-dimensional detection capabilities for lithium-ion battery quality control.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and more specifically, to a battery tab welding quality testing system and method. Background Technology

[0002] The welding quality of lithium-ion battery tabs directly affects the reliability and safety of the battery. Existing detection systems mainly rely on single-frame X-ray image analysis to examine the static structural features of the battery tab welding area, such as the shape of the weld, and the single-frame imaging of porosity or cracks. However, this single-frame analysis mode has significant limitations:

[0003] First, the existing system lacks a mechanism for temporal correlation of multi-dimensional data. X-ray images and electrical performance data are acquired independently by different devices. Due to factors such as sampling frequency and clock synchronization errors, the time coordinate systems of the two types of data are independent of each other, making it impossible to form a unified basis for time series analysis. For example, although a single frame of X-ray image can show the shape of a solder joint, it cannot correlate with the internal resistance fluctuation at the same moment, making it difficult to identify dynamic defects caused by abnormal internal resistance.

[0004] Secondly, existing technologies do not utilize the temporal correlation of continuous detection data to analyze the evolution of defects. Bar welding defects typically exhibit gradual changes, with their harmfulness accumulating with the number of cycles. However, single-frame image analysis can only capture the instantaneous state and cannot track the evolution trajectory of defects across multiple frame sequences. For example, microcracks are difficult to identify in a single frame image due to their small size, but the resulting abnormal internal resistance requires trend analysis of multiple frames of electrical performance data for effective detection. Existing systems, lacking temporal correlation processing, are prone to missing such dynamic defects.

[0005] Furthermore, existing detection methods do not integrate the temporal characteristics of multimodal data. Electrical performance data is more sensitive to early defects than single-frame images, but existing systems do not combine it with the temporal changes of X-ray images. This results in inherent defects in detection modes that rely solely on images or electrical performance data: analyzing images alone may miss electrical contact defects that have not yet formed obvious structural damage, while relying solely on electrical performance data cannot locate the physical location of defects.

[0006] In summary, existing lithium-ion battery tab welding quality inspection systems lack the ability to detect dynamic defects due to the absence of a multi-dimensional data time alignment mechanism, the failure to utilize the temporal correlation of continuous detection data, and the lack of integration of multi-modal temporal features.

[0007] In view of this, the present invention proposes a battery tab welding quality inspection system and method to solve the above problems. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a method for detecting the welding quality of battery tabs, comprising:

[0009] X-ray images and electrical performance data are acquired synchronously using a preset global clock signal to construct a time-series dataset sorted by timestamps;

[0010] The geometric transformation matrix between adjacent X-ray image frames is generated using an initial feature point tracking algorithm.

[0011] Based on the geometric transformation matrix, the time series data of X-ray images in the time series dataset are spatially aligned to obtain the time series data of X-ray images after alignment.

[0012] Time-series features are extracted from electrical performance time-series data in the time-series dataset to generate electrical performance time-series feature vectors.

[0013] The time-series data of the aligned X-ray images and the time-series feature vectors of electrical properties are input into a preset neural network model, which outputs the probability of predicted defect development and diagnoses the welding quality.

[0014] Furthermore, methods for constructing time-series datasets sorted by timestamps include:

[0015] A global clock signal is generated using a synchronous clock to synchronously trigger the acquisition actions of the X-ray imaging device and the electrical performance tester, thereby obtaining X-ray images and electrical performance data, respectively. The X-ray image data includes the pixel coordinates and grayscale values ​​of each frame of the X-ray image. The electrical performance data includes solder joint voltage, solder joint current, electrode temperature, and dynamic resistance.

[0016] Embed timestamps for each frame of X-ray image and each set of electrical performance data;

[0017] For each frame of X-ray image, within a preset time search range, the corresponding data group with the nearest timestamp in the electrical performance data is searched and matched to obtain the initial time series dataset;

[0018] When the sampling frequency of electrical performance data is higher than the frame rate of X-ray images, electrical performance data at the corresponding time is generated by linear interpolation for X-ray image frames without directly matching data.

[0019] The initial time series dataset is arranged in ascending order of timestamps to construct a time series dataset sorted by timestamps.

[0020] Furthermore, methods for generating geometric transformation matrices between adjacent X-ray image frames include:

[0021] Feature detection is performed on the first frame of the X-ray image to extract initial feature points of the battery tab welding area; the initial feature points include the coordinates of the edge corner points of the tab welding area and the coordinates of the center of the welding point.

[0022] Based on the optical flow method, the initial feature points in the X-ray image are tracked to generate the coordinate sequence of the initial feature points in each frame of the X-ray image, and the motion trajectory sequence of the feature points is obtained.

[0023] Based on the sequence of motion trajectories of feature points, the least squares method is used to calculate the translation matrix, rotation matrix, and scaling factor between adjacent X-ray image frames, which are then combined to form the geometric transformation matrix between adjacent X-ray image frames.

[0024] Furthermore, methods for obtaining position-aligned X-ray image time-series data include:

[0025] Let index variable k = 1, with a value range of [1, K]; index variable q = 1, with a value range of [1, Q]; K is the number of X-ray image frames, and Q is the number of pixels in each X-ray image frame;

[0026] S101: For the qth new pixel in the kth frame X-ray image of the time series dataset after the geometric transformation matrix is ​​applied, calculate the sampling position of the qth new pixel before the geometric transformation matrix is ​​applied based on the inverse matrix of the geometric transformation matrix, and mark it as the original sampling position.

[0027] S102: Determine the S nearest neighbor pixels with integer coordinates centered at the original sampling position of the qth new pixel, and calculate the gray value of the S pixels.

[0028] S103: Calculate the weighting coefficient based on the relative distance between the qth new pixel and the S pixels at the original sampling position;

[0029] S104: Multiply the gray values ​​of S pixels by their corresponding weight coefficients and sum them to obtain the gray value of the qth pixel after the geometric transformation matrix.

[0030] S105: If q is less than Q, then increment q by 1 and execute S101 to S105; if q is greater than or equal to Q and k is less than k, then increment q by 1 and execute S101 to S105; if q is greater than or equal to Q and k is greater than or equal to k, then complete the spatial position transformation of the X-ray image time series data and obtain the position-aligned X-ray image time series data.

[0031] Furthermore, methods for generating time-series feature vectors of electrical performance include:

[0032] Calculate the mean and standard deviation of each data category in the electrical performance time series data;

[0033] The standardized electrical performance time series data is generated by subtracting the mean of the corresponding category from each data point in each category of the electrical performance time series data and dividing by the standard deviation of the corresponding category.

[0034] Determine the sliding window size and sliding step size, and starting from the first data group of the standardized electrical performance time series data, extract data segments of the sliding window size in sequence according to the sliding step size to generate multiple electrical performance time series data segments.

[0035] Perform a Fourier transform on each electrical performance time-series data segment to generate a frequency domain feature vector;

[0036] Determine the timestamp corresponding to each frequency domain feature vector;

[0037] By concatenating all frequency domain feature vectors in chronological order according to their timestamps, the electrical performance time-series feature vector is obtained.

[0038] Furthermore, training methods for neural network models include:

[0039] Obtain a training dataset, which includes time-series X-ray image data samples with timestamps and aligned positions, corresponding electrical performance time-series data samples, and manually labeled welding quality label vectors.

[0040] A dual-branch neural network model is constructed, which includes an image feature extraction branch, a temporal fusion branch, and a fully connected classification layer.

[0041] Input the time-series data of the aligned X-ray image into the image feature extraction branch, and output the image feature vector;

[0042] The image feature vector and electrical performance time-series data are input into the time-series fusion branch, and a fused feature vector is generated through a splicing operation.

[0043] The fused feature vector is input into the fully connected classification layer, which outputs a defect probability vector.

[0044] Calculate the binary cross-entropy loss value for each dimension of the defect probability vector and take the average value to obtain the comprehensive loss value;

[0045] The trainable parameters of the two-branch neural network model are updated using the backpropagation algorithm until the comprehensive loss value converges, resulting in a well-trained neural network model.

[0046] Furthermore, methods for predicting defect development probability and diagnosing weld quality include:

[0047] The time-series data of the aligned X-ray images are input into the image feature extraction branch of the preset neural network model, and the output image feature vector is generated.

[0048] The image feature vector and the electrical performance time-series feature vector are input into the time-series fusion branch of the preset neural network model, and the fused feature vector is output.

[0049] The fused feature vector is input into the fully connected classification layer of the preset neural network model, and the defect probability vector is output.

[0050] Based on the defect probability vector, the defect type is determined by a preset defect threshold judgment rule, and a welding quality diagnosis result is generated.

[0051] Furthermore, the defect types include incomplete solder joints, burn-through, solder joint cracks, electrode tab misalignment, uneven welding, and abnormal solder joint dimensions;

[0052] The defect threshold judgment rule is to compare the probability value of each element in the defect probability vector with the preset defect threshold of the corresponding defect type. If the probability value of an element in the defect probability vector is greater than the corresponding defect threshold, it is determined that there is a defect of that type. The severity level of the defect is determined according to the relationship between the probability values ​​of each element in the defect probability vector.

[0053] Furthermore, methods for determining the severity level of defects include:

[0054] Sort each element whose probability value exceeds the preset defect threshold for the corresponding defect type from high to low according to its probability value;

[0055] If there are multiple defect types that exceed the defect threshold, the probability value of each defect type is multiplied by its corresponding pre-set impact factor and then summed to obtain the comprehensive impact value; the severity level is determined based on the interval in which the comprehensive impact value falls.

[0056] If only a single defect type exceeds the defect threshold, the severity level is directly determined based on the degree of deviation between the probability value of that defect type and the threshold; the severity level includes mild, moderate and severe.

[0057] Compared with existing technologies, the technical effects and advantages of the battery tab welding quality inspection system and method of the present invention are as follows:

[0058] By aligning multi-dimensional data in a spatiotemporal manner, a time-series dataset including X-ray images and electrical performance data was constructed. A dual-branch neural network model was used to extract features from the X-ray images and electrical performance data. A fully connected classification layer was used to obtain a defect probability vector, and welding quality diagnostic results were generated based on this vector. This approach achieved spatiotemporal alignment and deep fusion of multi-modal data, enabling the capture of the evolution trajectory of electrode welding defects in continuous X-ray image frames. Temporal feature extraction and dynamic fusion improved the detection system's sensitivity to early micro-defects and its anti-interference capability under complex operating conditions. The overall solution significantly improved the accuracy of dynamic defect detection in lithium battery electrode welding, solved the missed detection problem in traditional single-frame analysis, and provided full-cycle, multi-dimensional detection capabilities for lithium-ion battery quality control. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of a battery tab welding quality inspection system according to an embodiment of the present invention;

[0060] Figure 2 This is a flowchart of the battery tab welding quality inspection method according to an embodiment of the present invention;

[0061] Figure 3 This is a flowchart illustrating a method for constructing a time-series dataset sorted by timestamps, according to an embodiment of the present invention.

[0062] Figure 4 This is a flowchart of a method for obtaining position-aligned X-ray image time-series data according to an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of the feature point motion trajectory sequence according to an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0065] Example 1

[0066] Please see Figure 1 As shown, this embodiment discloses a battery tab welding quality inspection system, including a data acquisition module, a geometric transformation module, a position alignment module, a feature extraction module, and a prediction and diagnosis module. Each module is connected by wires and / or wirelessly to realize data transmission.

[0067] The data acquisition module synchronously acquires X-ray images and electrical performance data through a preset global clock signal, and constructs a time-series dataset sorted by timestamp.

[0068] Please see Figure 3 As shown, methods for constructing time-series datasets sorted by timestamps include:

[0069] A global clock signal is generated using a synchronous clock to synchronously trigger the acquisition actions of the X-ray imaging device and the electrical performance tester, thereby obtaining X-ray images and electrical performance data, respectively. The X-ray image data includes the pixel coordinates and grayscale values ​​of each frame of the X-ray image. The electrical performance data includes solder joint voltage, solder joint current, electrode temperature, and dynamic resistance.

[0070] Embed timestamps for each frame of X-ray image, each set of electrical performance data, and each environmental parameter;

[0071] For each frame of X-ray image, within a preset time search range, the corresponding data group with the nearest timestamp in the electrical performance data is searched and matched to obtain the initial time series dataset;

[0072] When the sampling frequency of electrical performance data is higher than the frame rate of X-ray images, electrical performance data at the corresponding time is generated by linear interpolation for X-ray image frames without directly matching data.

[0073] The initial time series dataset is arranged in ascending order of timestamps to construct a time series dataset sorted by timestamps.

[0074] For example, this embodiment provides a method for constructing a time-series dataset sorted by timestamps in a battery tab welding quality inspection system, as detailed below:

[0075] A global clock signal is generated using a synchronous clock to synchronously trigger the acquisition actions of the X-ray imaging equipment, electrical performance tester, and environmental sensors, thereby obtaining X-ray images and electrical performance data, respectively.

[0076] Embed timestamps for each frame of X-ray image and each set of electrical performance data.

[0077] For each frame of X-ray image, the method of matching the nearest corresponding data group with the adjacent timestamp in the electrical performance data within a preset time search range to obtain the initial time series dataset includes:

[0078] Read the precise timestamp of the current X-ray image frame to be matched, and perform a bidirectional search in the stored electrical performance data;

[0079] The retrieval process centers on the timestamp of the current X-ray image to be matched. For example, the preset time search range is set to ±10 milliseconds, which can be adjusted according to the actual system accuracy. All electrical performance data groups falling within the time window of [image timestamp - 10ms, image timestamp + 10ms] are filtered out.

[0080] When multiple data groups exist within a time window, a specific matching strategy will be adopted: the single data group whose timestamp is less than the timestamp of the current X-ray image to be matched and is closest to the acquisition time of the current X-ray image to be matched will be selected as the matching item; if there is only a data group with a timestamp greater than the timestamp of the current X-ray image to be matched within the time window, the matching of that frame image will be automatically determined to be unsuccessful.

[0081] This matching process ensures that each frame of image is strictly associated with the actual operating condition data of the moment most recently before its capture time. For example, when the electrical performance sensor samples at a frequency of 1 kHz (1 ms interval) and X-rays are imaged at 10 fks (100 ms interval), a single frame of X-ray image can usually match 20 sets of electrical performance data within a preset 20 ms time window. At this time, the electrical performance sampling point that is immediately adjacent to the X-ray image acquisition time in the timestamp sorting is accurately selected, which effectively avoids future data interference and conforms to the causality principle of industrial inspection.

[0082] After matching is completed, the current X-ray image frame is combined with the selected electrical performance data set to form an initial time-series dataset.

[0083] When the sampling frequency of electrical performance data is higher than the frame rate of X-ray images, methods for generating electrical performance data at corresponding times using linear interpolation for X-ray image frames without directly matching data include:

[0084] Locate the two most valid data sets closest to each other before and after the timestamp of the current X-ray image to be processed in the electrical performance data: data set A with the previous timestamp T1 and data set B with the next timestamp T2. The time difference between T1 and T2 and the image timestamp T0 is required to be less than the preset maximum allowable interpolation interval (e.g., 50ms).

[0085] If two valid data sets exist, linear interpolation is performed based on time weights:

[0086] The calculation method for the solder joint voltage V0 in the electrical performance data is as follows:

[0087] V0 = V A +(V B -V A )×(T0-T1) / (T2-T1);

[0088] In the formula, V0 is the solder joint voltage in the electrical performance data corresponding to the current X-ray image timestamp obtained by interpolation; VA The solder joint voltage in the electrical performance data corresponding to data group A; V B The solder joint voltage is the electrical performance data corresponding to data group B.

[0089] The calculation method for the solder joint current I0 in the electrical performance data is as follows:

[0090] I0 = I A +(I B -I A )×(T0-T1) / (T2-T1);

[0091] In the formula, I0 is the weld current in the electrical performance data corresponding to the current X-ray image timestamp obtained by interpolation; I A The solder joint current in the electrical performance data corresponding to data group A; I B The solder joint current is the electrical performance data corresponding to data group B.

[0092] The calculation method for electrode temperature θ0 in electrical performance data is as follows:

[0093] θ0=θ A +(θ B -θ A )×(T0-T1) / (T2-T1);

[0094] In the formula, θ0 is the electrode temperature in the electrical performance data corresponding to the current X-ray image timestamp obtained by interpolation; θ A θ represents the electrode temperature in the electrical performance data corresponding to data set A. B The electrode temperature is the value in the electrical performance data corresponding to data group B.

[0095] The calculation method for the dynamic resistance R0 in the electrical performance data is as follows:

[0096] R0 = R A +(R B -R A )×(T0-T1) / (T2-T1);

[0097] In the formula, R0 is the dynamic resistance in the electrical performance data corresponding to the current X-ray image timestamp obtained by interpolation; R A R is the dynamic resistance in the electrical performance data corresponding to data group A; B This refers to the dynamic resistance in the electrical performance data corresponding to data group B.

[0098] For example, when electrical performance data is sampled at 1 kHz (1 ms interval), X-ray image is formed at T = 105.5 ms, and the previous matching point is T = 100 ms and the next sampling point is T = 106 ms, the solder joint voltage at T = 100 ms and T = 106 ms is automatically taken, and the solder joint voltage interpolated at T = 105.5 ms is calculated according to the time ratio.

[0099] The initial time series dataset is arranged in ascending order of timestamps to construct a time series dataset sorted by timestamps.

[0100] The global clock signal setting ensures strict consistency of the three types of data in the time coordinate system, avoiding data misalignment caused by asynchronous device clocks, and forming the basis for subsequent time-series analysis. The time window for electrical performance data is completely consistent with the time window for X-ray images, ensuring strict correspondence of multi-dimensional data in the time dimension. This provides an accurate data foundation for subsequent analysis of the correlation between defect evolution, electrical performance response, and environmental interference. This time alignment method significantly improves the detection system's ability to capture dynamic defects, accurately identifying characteristic changes in the battery tab welding area within a short period, making the detection results more reliable.

[0101] The fundamental reason for using a combination of X-ray image data and electrical performance data for computational processing is that these data comprehensively characterize the core dynamic characteristics of battery tab welding from two complementary dimensions: spatial physical state and energy transfer process. For example, the pixel coordinates of the X-ray image are mapped to the positional changes of pixels in the tab welding area through a geometric transformation matrix, such as tab offset and weld nugget size; while the difference in grayscale value directly reflects the material density distribution. For example, the low-density grayscale value of the tab poor weld area drops by about 15-30 units, which can accurately capture hidden defects such as internal cracks and pores in the weld point; in the electrical performance data, the real-time interaction between the weld point voltage and current quantifies the energy input stability, the dynamic resistance reveals the state of the tab contact interface through voltage-current ratio calculation, and the electrode temperature reflects the heat dissipation efficiency.

[0102] The geometric transformation module uses an initial feature point tracking algorithm to generate geometric transformation matrices between adjacent X-ray image frames.

[0103] Methods for generating geometric transformation matrices between adjacent X-ray image frames include:

[0104] Feature detection is performed on the first frame of the X-ray image to extract initial feature points of the battery tab welding area; the initial feature points include the coordinates of the edge corner points of the tab welding area and the coordinates of the center of the welding point.

[0105] Based on the optical flow method, the initial feature points in the X-ray image are tracked to generate the coordinate sequence of the initial feature points in each frame of the X-ray image, and the motion trajectory sequence of the feature points is obtained.

[0106] Based on the sequence of motion trajectories of feature points, the least squares method is used to calculate the translation matrix, rotation matrix, and scaling factor between adjacent X-ray image frames, which are then combined to form the geometric transformation matrix between adjacent X-ray image frames.

[0107] For example, this embodiment provides a method for generating a geometric transformation matrix between adjacent X-ray image frames in a battery tab welding quality inspection system, as detailed below:

[0108] Methods for extracting initial feature points of the battery tab welding area by performing feature detection on the first frame of the X-ray image include:

[0109] The X-ray images were preprocessed, with a median filter used to eliminate noise, and histogram equalization was used to enhance the contrast of the tab welding area.

[0110] For the extraction of edge corner points in the electrode welding area, the Harris corner detection algorithm is used: the gradients in the X and Y directions of the first frame X-ray image are calculated, an autocorrelation matrix is ​​constructed, and the corner response of each edge corner point is calculated using the corner response function; for example, the corner response threshold is set to 0.8 × 10⁻⁶. 4 (Empirical value) Pixels corresponding to edge corners with corner responses greater than the corner response threshold are selected on the electrode welding edge contour and included in the corner candidate set;

[0111] Based on the grayscale difference between the weld nugget area and the background, morphological closing operations are applied to each weld point in the electrode welding area to fill the internal pores of the weld point; the weld nugget area is a locally bright area that appears as an approximately circular shape in the X-ray image;

[0112] Adjacent pixels in a solder joint whose grayscale value is greater than a set grayscale threshold are grouped into a single connected component.

[0113] Calculate the gray-level weighted centroid of all pixels within each connected component to obtain the solder joint center coordinates (x, y, y) with sub-pixel precision. c y c ):

[0114]

[0115] In the formula, (x p ,y p Let I(x) be the integer coordinates of the p-th pixel within the connected component; p ,y p ) is the coordinate (x p ,y p The normalized gray values ​​of ) range from [0,1]; ∑(x p ×I(x p ,y p)) represents the gray-level weighted summation of the x-coordinates of all pixels within the connected component; ∑(y p ×I(x p y p )) represents the gray-level weighted summation of the y-coordinates of all pixels within the connected component; ∑I(x p y p ) represents the sum of the gray values ​​of all pixels within the connected component.

[0116] After the initial inspection is completed, feature verification is performed to obtain the coordinates of the edge corner points of the welding area. For example, edge corner points located within 5 pixels of the edge of the first frame X-ray image are removed from the corner point candidate set, and invalid weld points with a diameter of less than 3 pixels or a grayscale mean lower than 1.5 times the standard deviation of the background are filtered out.

[0117] The output includes the initial feature points containing the coordinates of the edge corners of the welding area and the center coordinates of the weld point.

[0118] Please see Figure 5 As shown, based on the optical flow method, the method for tracking initial feature points in X-ray images and generating a coordinate sequence of initial feature points in each frame of X-ray images to obtain the feature point motion trajectory sequence includes:

[0119] Using the first frame of the X-ray image as the reference frame, the reference frame and the subsequent k-th frame are combined to form an image pair, where k is a positive integer greater than or equal to 2;

[0120] For each initial feature point of an image pair, a three-layer image pyramid is constructed, including the original resolution layer (layer 0), the 1 / 2 downsampling layer (layer 1), and the 1 / 4 downsampling layer (layer 2).

[0121] Initialize the motion vector at the highest layer (layer 2), and calculate the displacement increment of the highest layer by iteratively solving the optical flow equation;

[0122] The displacement increment of the highest layer is passed to the next layer and the coordinates of the displacement increment of the highest layer are corrected. Finally, sub-pixel level displacement vectors are obtained in the original resolution layer (layer 0).

[0123] Update the predicted coordinates of the initial feature points in the k-th frame X-ray image based on sub-pixel level displacement vectors;

[0124] Using the predicted coordinates of the k-th frame X-ray image as a reference, the coordinates of the corresponding initial feature points in the (k-1)-th frame X-ray image are traced backward to obtain the backward tracing coordinates.

[0125] The Euclidean distance between the reverse tracking coordinates and the original corresponding coordinates of the (k-1)th frame X-ray image is calculated as the position deviation. For example, if the position deviation of the initial feature point does not exceed 0.01 pixels, the initial feature point is determined to be a valid tracking point.

[0126] Record the coordinates (x, y) of the valid tracking point in the kth frame. k ,y k And associated with the initial coordinates (x0, y0) of the reference frame, forming the motion trajectory sequence of the initial feature point {(x0, y0), (x0, y1), ..., (x0, y2)}. k ,y k )}.

[0127] Based on the feature point motion trajectory sequence, the least squares method is used to calculate the translation matrix, rotation matrix, and scaling factor between adjacent X-ray image frames, and the methods for combining them to form the geometric transformation matrix between adjacent X-ray image frames include:

[0128] Suppose there are n valid tracking points in the k-th frame of the X-ray image, and i represents the ith valid tracking point. Then the coordinates of the ith valid tracking point are: The coordinates of the i-th valid tracking point in the (k+1)-th frame are The following objective function is solved using the least squares method:

[0129]

[0130] The parameters to be determined include the scaling factor s, the rotation angle θ, and the translation amount t. x and t y To simplify the calculation, the solution process is broken down into the following steps:

[0131] The centroid offset between the effective tracking points in the k-th frame and the (k+1)-th frame in the feature point motion trajectory sequence is taken as the initial estimate of the translation in the k-th frame.

[0132]

[0133] In the formula, The average coordinates of the effective tracking points in the k-th frame; The average coordinates of the (k+1)th valid tracking point;

[0134] The decentralized coordinates of the effective tracking point in the k-th frame are obtained by subtracting the centroid of the effective tracking point in the k-th frame from the coordinates of the effective tracking point in the feature point motion trajectory sequence.

[0135]

[0136] Construct a 2×2 covariance matrix H:

[0137]

[0138] In the formula, element H 11 The covariance of motion in the X direction; element H 22 The covariance of motion in the Y direction; element H12 Indicates the degree of XY motion coupling; element H 21 Indicates the degree of coupling between the Y and X motions;

[0139] Perform singular value decomposition on H:

[0140]

[0141] In the formula, U and V are 2×2 orthogonal matrices; Σ is a singular value matrix used to characterize the transformation intensity;

[0142] The rotation matrix R and the scaling factor s are:

[0143]

[0144] In the formula, ||Σ|| F For Frobenius norm; det(UV) T The value is ±1, which is used to ensure the correctness of the rotation direction;

[0145] The solved parameters are integrated into a 3×3 homogeneous coordinate transformation matrix T. k→k+1 :

[0146]

[0147] Output T k→k+1 As the geometric transformation matrix between adjacent frames.

[0148] This study quantifies the temporal changes in physical deformation during the lithium-ion battery tab welding process by tracking the motion trajectory of feature points using optical flow. Dynamic compensation of the initial feature point displacement in X-ray images reflects the positional shift of the lithium-ion battery tabs due to thermal expansion or mechanical vibration during welding. Simultaneously, a geometric transformation matrix maps multiple X-ray images to the same coordinate system, aligning the images and allowing for comparison of the same physical location at different times. Without proper alignment, the same weld point may experience coordinate drift in consecutive frames, for example, from (100, 200) to (105, 203), leading the neural network to misjudge it as a tab misalignment defect.

[0149] The position alignment module performs spatial position alignment on the X-ray image time series data in the time series dataset based on the geometric transformation matrix, resulting in position-aligned X-ray image time series data.

[0150] Please see Figure 4 As shown, methods for obtaining position-aligned X-ray image time-series data include:

[0151] Let index variable k = 1, with a value range of [1, k]; index variable q = 1, with a value range of [1, Q]; k is the number of X-ray image frames, and Q is the number of pixels in each X-ray image frame;

[0152] S101: For the qth new pixel in the kth frame X-ray image of the time series dataset after the geometric transformation matrix is ​​applied, calculate the sampling position of the qth new pixel before the geometric transformation matrix is ​​applied based on the inverse matrix of the geometric transformation matrix, and mark it as the original sampling position.

[0153] S102: Determine the S nearest neighbor pixels with integer coordinates and their gray values, centered on the sampling position corresponding to the qth new pixel in the original image;

[0154] S103: Calculate the weighting coefficient based on the relative distance between the sampling position of the qth new pixel in the original image and the S integer coordinate points;

[0155] S104: Multiply the gray values ​​of S pixels by their corresponding weight coefficients and sum them to obtain the gray value of the qth pixel after the geometric transformation matrix.

[0156] S105: If q is less than Q, then increment q by 1 and execute S101 to S105; if q is greater than or equal to Q and k is less than k, then increment q by 1 and execute S101 to S105; if q is greater than or equal to Q and k is greater than or equal to k, then complete the spatial position transformation of the X-ray image time series data and obtain the position-aligned X-ray image time series data.

[0157] For example, this embodiment provides a method for obtaining time-series X-ray image data after position alignment in a battery tab welding quality inspection system, as detailed below:

[0158] For the q-th new pixel in the k-th frame X-ray image of a time-series dataset after the geometric transformation matrix, the method for calculating the sampling position of the q-th new pixel before the geometric transformation matrix is ​​based on the inverse of the geometric transformation matrix includes:

[0159] Perform pixel-by-pixel inverse sampling on the k-th frame X-ray image in the time-series dataset. Let the homogeneous coordinates of the q-th pixel in the k-th frame X-ray image be... Where (u k，q ,v k，q ) represents the integer grid coordinates of the q-th pixel in the k-th frame of the X-ray image, for example, q=1 corresponds to (0,0), q=2 corresponds to (0,1), etc.

[0160] The inverse matrix of the geometric transformation matrix Calculate the original sampling position of the q-th pixel.

[0161]

[0162] Methods for determining the S nearest neighbor pixels and their grayscale values ​​centered on the original sampling position of the q-th new pixel include:

[0163] Original sampling position obtained Coordinates are represented as Calculate the reference point at the top left corner of the original sampling location:

[0164]

[0165] In the formula, For example, the floor function.

[0166] For example, the nearest S = 4 integer coordinate points of the original sampling position are fixedly selected. This is the standard configuration of bilinear interpolation. The reference point is (98,195). Then the 4 selected integer coordinate points are: A(98,195), B(99,195), C(98,196), D(99,196).

[0167] The pixel grayscale value at integer coordinates can be obtained by directly accessing the X-ray image memory index.

[0168] Methods for calculating weighting coefficients based on the relative distance between the original sampling position of the q-th new pixel and the S integer coordinate points include:

[0169] Calculate the normalized horizontal offset dx and vertical offset dy of the original sampling position relative to the reference point A:

[0170]

[0171] For example, when And when A(98,195), dx = 0.2, dy = 0.7; according to the bilinear interpolation rule, the weight coefficients of the four integer coordinate points are determined by their geometric distance from the sampling position:

[0172] Weight of point A (bottom left corner): w A = (1-dx)×(1-dy); This reflects that the influence of point A on the sampling position decreases as the horizontal and vertical distances increase;

[0173] Weight of point B (bottom right corner): w B = dx × (1 - dy); This reflects that the contribution of point B is positively correlated with the horizontal distance and negatively correlated with the vertical distance.

[0174] Weight of point C (top left corner): w C= (1-dx)×dy; reflects that the contribution of point C is positively correlated with the vertical distance and negatively correlated with the horizontal distance;

[0175] Weight of point D (top right corner): w D =dx×dy; This reflects that the contribution of point D is positively correlated with both the horizontal and vertical distances.

[0176] The gray values ​​of the S pixels are multiplied by their corresponding weight coefficients and then summed to obtain the gray value of the q-th pixel after the geometric transformation matrix.

[0177] Methods for performing grayscale normalization and edge enhancement processing on the transformed X-ray image time series data to obtain position-aligned X-ray image time series data include:

[0178] Iterate through all pixels in each frame of the X-ray image and calculate the mean gray value μ of the current frame k. k and standard deviation σ k A linear transformation is used to map the gray values ​​I(k,q) of the X-ray image to the target interval [0,255], resulting in the gray value I after gray-level normalization. norm (k,q):

[0179]

[0180] In the formula, α is the mean of the global standard deviation. β is the global grayscale mean. Ensure consistent grayscale distribution across frames;

[0181] Spatial convolution is performed using a 3×3 Sobel operator, and the horizontal gradient G is calculated respectively. x With vertical gradient G y :

[0182]

[0183] Synthetic gradient magnitude Then, edge-enhanced grayscale values ​​I are generated using a nonlinear enhancement function. edge (p,q):

[0184] I edge (p,q)=I norm (p,q)+λ×tanh(γG);

[0185] In the formula, λ is set to 0.8 to control the enhancement intensity; γ is set to 0.5 to adjust the edge sharpness; and the tanh(·) function is used to suppress noise amplification.

[0186] The output is the time series data of the X-ray image after position alignment.

[0187] The original sampling position of the new pixel is calculated using the inverse of the geometric transformation matrix. Combined with bilinear interpolation, the grayscale values ​​of the nearest neighbor pixels are weighted and summed to ensure precise alignment of each frame in spatial coordinates. This avoids defect feature shifts or misjudgments caused by positional deviations between different frames, laying a spatial foundation for subsequent temporal alignment of time-series data. Grayscale normalization and edge enhancement processes eliminate grayscale differences and noise interference across frames, ensuring consistency in grayscale distribution and feature representation across different time points. This guarantees the comparability of X-ray image time-series data in the temporal dimension after alignment, allowing for accurate tracking of dynamic changes in defects such as crack propagation and weld morphology evolution between frames. This addresses the issue of ineffective temporal feature correlation due to spatial misalignment in misaligned images. Furthermore, precise spatial alignment enables the synchronization of X-ray image time-series data with electrical performance time-series data via timestamps, further achieving spatiotemporal unification of spatial and electrical performance features, providing a prerequisite for subsequent multimodal time-series feature fusion. If images are not aligned, their extracted spatial features will not match the electrical performance features in the spatiotemporal dimension, rendering multimodal fusion meaningless. This image spatial alignment, achieved through geometric transformation, interpolation calculation, grayscale normalization, and edge enhancement, fundamentally solves the problem of the lack of a multidimensional data temporal alignment mechanism. It provides a precise spatial benchmark for temporal correlation analysis and multimodal feature fusion, thereby enhancing the detection capability of dynamic defects. This enables the system to accurately capture the evolution of defects in the temporal dimension and output reliable detection results.

[0188] The feature extraction module is used to extract time-series features from the electrical performance time-series data in the time-series dataset and generate electrical performance time-series feature vectors.

[0189] Methods for generating time-series feature vectors of electrical performance include:

[0190] Calculate the mean and standard deviation of each data category in the electrical performance time series data;

[0191] The standardized electrical performance time series data is generated by subtracting the mean of the corresponding category from each data point in each category of the electrical performance time series data and dividing by the standard deviation.

[0192] Determine the sliding window size and sliding step size, and starting from the first data group of the standardized electrical performance time series data, extract data segments of the sliding window size in sequence according to the sliding step size to generate multiple electrical performance time series data segments.

[0193] Perform a Fourier transform on each electrical performance time-series data segment to generate a frequency domain feature vector;

[0194] Determine the timestamp corresponding to each frequency domain feature vector;

[0195] By concatenating all frequency domain feature vectors in chronological order according to their timestamps, the electrical performance time-series feature vector is obtained.

[0196] For example, this embodiment provides a method for generating electrical performance time-series feature vectors in a battery tab welding quality inspection system, as detailed below:

[0197] Calculate the mean and standard deviation of each data category in the electrical performance time series data.

[0198] The standardized electrical performance time series data is generated by subtracting the mean of the corresponding category from each data point in each category of the electrical performance time series data and dividing by the standard deviation.

[0199] The method for determining the sliding window size and sliding step size, and generating multiple electrical performance time series data segments by sequentially extracting data segments of the sliding window size according to the sliding step size, starting from the first data group of the standardized electrical performance time series data, includes:

[0200] For example, the obtained standardized electrical performance time series data has 100 data sets. Based on the time series analysis method, the sliding window size is set to 10 data sets and the sliding step size is set to 5 data sets.

[0201] Starting from the first data group of the electrical performance time series data, i.e., index 0, data segments are extracted sequentially according to the sliding step size: first, data groups with indices 0 to 9 are extracted to form the first electrical performance time series data segment; then, the sliding step size is moved by 5 data groups to extract data groups with indices 5 to 14 to form the second electrical performance time series data segment; this process continues to extract data groups with indices 10 to 19, 15 to 24, etc., to generate subsequent segments in sequence; when the window moves to the end of the data, if the remaining data groups are less than the window size (for example, only 5 points with indices 95 to 99), the insufficient part is ignored;

[0202] Multiple complete electrical performance time-series data segments are generated, each segment consisting of 10 data groups for subsequent processing.

[0203] Perform a Fourier transform on each electrical performance time-series data segment to generate a frequency domain feature vector.

[0204] Determine the timestamp corresponding to each frequency domain feature vector.

[0205] By concatenating all frequency domain feature vectors in chronological order according to their timestamps, the electrical performance time-series feature vector is obtained.

[0206] Frequency domain correlation features are extracted using Fourier transform to capture periodic patterns in the welding process. For example, incomplete weld defects are characterized by a large proportion of low-frequency energy in the frequency domain feature vector, while burn-through defects are characterized by a surge in high-frequency abrupt energy in the frequency domain feature vector. Each frequency domain feature vector is associated with a unique timestamp, ensuring strict synchronization between the frequency domain evolution of electrical performance and X-ray images, thus establishing a correlation between electrical performance and X-ray images.

[0207] The prediction and diagnosis module is used to input the time-series data of the position-aligned X-ray image and the time-series feature vector of electrical performance into a preset neural network model, and output the predicted probability of defect development and diagnose the welding quality.

[0208] Training methods for neural network models include:

[0209] Obtain a training dataset, which includes time-series X-ray image data samples with timestamps and aligned positions, corresponding electrical performance time-series data samples, and manually labeled welding quality label vectors.

[0210] A dual-branch neural network model is constructed, which includes an image feature extraction branch, a temporal fusion branch, and a fully connected classification layer.

[0211] Input the time-series data of the aligned X-ray image into the image feature extraction branch, and output the image feature vector;

[0212] The image feature vector and the electrical performance time-series feature vector are input into the time-series fusion branch, and a fused feature vector is generated through a concatenation operation.

[0213] The fused feature vector is input into the fully connected classification layer, which outputs a defect probability vector.

[0214] Calculate the cross-entropy loss value of each dimension of the defect probability vector and take the average value to obtain the comprehensive loss value;

[0215] The trainable parameters of the two-branch neural network model are updated using the backpropagation algorithm until the comprehensive loss value converges, resulting in a well-trained neural network model.

[0216] For example, this embodiment provides a method for training a neural network model in a battery tab welding quality inspection system, as detailed below:

[0217] Obtain a training dataset, which includes time-series X-ray image data samples with timestamps and aligned positions, corresponding electrical performance time-series data samples, and manually labeled welding quality label vectors.

[0218] Methods for constructing two-branch neural network models include:

[0219] An image feature extraction branch is constructed, which adopts a convolutional neural network structure and contains multiple convolutional and pooling layers. It is used to process the input position-aligned X-ray image time series data, extract spatial features through convolution operations and output image feature vectors.

[0220] A temporal fusion branch is constructed, which includes a long short-term memory network layer to receive the image feature vector and the electrical performance temporal feature vector output by the image feature extraction branch, and then input the two feature vectors into the long short-term memory network layer to generate a fused feature vector.

[0221] A fully connected classification layer is constructed, which consists of multiple fully connected layers, wherein the last fully connected layer uses the sigmoid activation function and outputs a defect probability vector.

[0222] The time-series data of the aligned X-ray image is input into the image feature extraction branch, and the image feature vector is output.

[0223] The image feature vector and the electrical performance time-series feature vector are input into the time-series fusion branch, and a fused feature vector is generated through a concatenation operation.

[0224] Methods that input fused feature vectors into a fully connected classification layer and output defect probability vectors include:

[0225] Use the fused feature vector output from the temporal fusion branch as input data;

[0226] For example, the fully connected classification layer consists of three fully connected layers: the first fully connected layer contains 256 neurons, which receive the fused feature vector and then perform a nonlinear transformation through the ReLU activation function; the second fully connected layer contains 128 neurons, which receive the output of the previous layer and then perform a ReLU activation function; the third fully connected layer contains 6 neurons, which receive the output of the previous layer and then perform a sigmoid activation function to generate a defect probability vector; the output value of each dimension of the defect probability vector ranges from 0 to 1, corresponding to the predicted probability of six defect types: cold solder joint, burn-through, cracked solder joint, tab offset, uneven welding, and abnormal solder joint size.

[0227] Methods for calculating the cross-entropy loss value of each dimension of the defect probability vector and then averaging it to obtain the comprehensive loss value include:

[0228] Obtain the defect probability vector output by the fully connected classification layer and the manually labeled welding quality label vector of the corresponding sample in the training dataset;

[0229] For each type of defect, perform the following operations independently:

[0230] Extract the predicted probability value corresponding to this type from the defect probability vector, and simultaneously extract the true label value (0 or 1) corresponding to this type from the manually labeled welding quality label vector, and calculate the binary cross-entropy loss value for this dimension.

[0231] The method for calculating the binary cross-entropy loss value of this dimension is as follows: if the true label value is 1, then the binary cross-entropy loss value is the negative logarithmic function applied to the predicted probability value; if the true label value is 0, then the loss value is the negative logarithmic function applied to 1 minus the predicted probability value.

[0232] After calculating the binary cross-entropy loss value for each dimension of the defect type, the binary cross-entropy loss values ​​are summed and divided by the number of defect types to obtain the comprehensive loss value; the comprehensive loss value reflects the overall prediction bias of the neural network model for the defect type.

[0233] The trainable parameters of the two-branch neural network model are updated using the backpropagation algorithm until the comprehensive loss value converges, resulting in a well-trained neural network model.

[0234] Methods for predicting defect development probability and diagnosing weld quality include:

[0235] The time-series data of the aligned X-ray images are input into the image feature extraction branch of the preset neural network model, and the output image feature vector is generated.

[0236] The image feature vector and the electrical performance time-series feature vector are input into the time-series fusion branch of the preset neural network model, and the fused feature vector is output.

[0237] The fused feature vector is input into the fully connected classification layer of the preset neural network model, and the defect probability vector is output.

[0238] Based on the defect probability vector, the defect type is determined by a preset defect threshold judgment rule, and a welding quality diagnosis result is generated.

[0239] The defect types include incomplete solder joints, burn-through, cracked solder joints, offset electrode tabs, uneven welding, and abnormal solder joint dimensions.

[0240] The defect threshold judgment rule is to compare the probability value of each element in the defect probability vector with the preset defect threshold of the corresponding defect type. If the probability value of an element in the defect probability vector is greater than the corresponding defect threshold, it is determined that there is a defect of that type. The severity level of the defect is determined according to the relationship between the probability values ​​of each element in the defect probability vector.

[0241] The method for determining the severity level of a defect includes:

[0242] Sort each element whose probability value exceeds the preset defect threshold for the corresponding defect type from high to low according to its probability value;

[0243] If there are multiple defect types that exceed the defect threshold, the probability value of each defect type is multiplied by its corresponding pre-set impact factor and then summed to obtain the comprehensive impact value; the severity level is determined based on the interval in which the comprehensive impact value falls.

[0244] If only a single defect type exceeds the defect threshold, the severity level is directly determined based on the degree of deviation between the probability value of that defect type and the defect threshold; the severity level includes mild, moderate and severe.

[0245] For example, this embodiment provides a method for predicting the probability of defect development and determining the severity level of defects in welding quality diagnosis, as follows:

[0246] In this embodiment, the defect threshold is set to 0.8, based on the confidence distribution of historical defect samples. For example, analyzing 1000 confirmed defect samples reveals that when the neural network model predicts a probability greater than or equal to 0.8, the accuracy of manual re-inspection confirming defects reaches over 95%. The influence factors for each defect type are set according to the actual severity: 0.5 for incomplete solder joint, 0.9 for burn-through, 0.7 for cracked solder joint, 0.6 for offset electrode, 0.4 for uneven welding, and 0.6 for abnormal solder joint size. The defect probability vector output by the neural network is [0.92, 0.95, 0.88, 0.75, 0.65, 0.82].

[0247] Defect types exceeding the defect threshold were selected, including incomplete solder joints (0.92), burn-through (0.95), cracked solder joints (0.88), and abnormal solder joint dimensions (0.82).

[0248] The probabilities, ranked from highest to lowest, are: weld penetration (0.95) > weld crack (0.88) > incomplete weld (0.92) > abnormal weld size (0.82);

[0249] Due to the existence of multiple defects exceeding the standard, the comprehensive impact value is calculated as follows:

[0250] The impact value of weld burn-through = 0.95 × 0.9 = 0.855; the impact value of weld crack = 0.88 × 0.7 = 0.616; the impact value of incomplete weld = 0.92 × 0.5 = 0.460; the impact value of dimensional anomaly = 0.82 × 0.6 = 0.492; the comprehensive impact value = 0.855 + 0.616 + 0.460 + 0.492 = 2.423;

[0251] Based on maintenance costs and production risks, the intervals are divided into the following rules: a comprehensive impact value < 1.5 is considered mild; 1.5 ≤ comprehensive impact value < 2.5 is considered moderate; and a comprehensive impact value ≥ 2.5 is considered severe. Since 2.423 falls within the moderate interval (1.5-2.5), it is determined to be a moderate defect. The interval division rules can be adaptively adjusted according to actual production needs.

[0252] If only the probability value of weld penetration exceeds the limit, for example, the probability vector is [0.75, 0.91, 0.78, 0.65, 0.72, 0.79], then it is handled according to the single defect rule:

[0253] The deviation is calculated as (0.91-0.8) / 0.2 = 0.55; the denominator 0.2 is a preset range coefficient, reflecting the effective discrimination range of the probability value, which can be adaptively adjusted according to the actual situation.

[0254] The deviation level is set as follows: deviation < 0.3 is mild; 0.3 ≤ deviation < 0.6 is moderate; deviation ≥ 0.6 is severe; since 0.55 falls in the moderate range, the weld penetration defect is judged to be moderately severe; the deviation level can be adjusted adaptively according to actual production needs.

[0255] By constructing a temporal fusion branch in a two-branch neural network model containing a long short-term memory (LSTM) layer, the temporal correlation of continuous data is fully utilized. This allows the model to capture the dynamic changes of electrode welding defects over time, enabling it to predict the development trend of these defects and overcoming the previous limitation of not being able to utilize temporal information for dynamic defect detection. The aligned X-ray image temporal data is input into the image feature extraction branch to obtain image feature vectors. These vectors are then concatenated and fused with the electrical performance temporal feature vectors in the temporal fusion branch, achieving the fusion of multimodal temporal features. This allows the neural network model to acquire complementary information from both image and electrical performance temporal data, comprehensively assessing welding quality and overcoming the limitations of single-modal data detection.

[0256] Example 2

[0257] Please see Figure 2 As shown, this embodiment provides a method for inspecting the welding quality of battery tabs, including:

[0258] X-ray images and electrical performance data are acquired synchronously using a preset global clock signal to construct a time-series dataset sorted by timestamps;

[0259] The geometric transformation matrix between adjacent X-ray image frames is generated using an initial feature point tracking algorithm.

[0260] Based on the geometric transformation matrix, the time series data of X-ray images in the time series dataset are spatially aligned to obtain the time series data of X-ray images after alignment.

[0261] Time-series features are extracted from electrical performance time-series data in the time-series dataset to generate electrical performance time-series feature vectors.

[0262] The time-series data of the aligned X-ray images and the time-series feature vectors of electrical properties are input into a preset neural network model, and the output is the predicted probability of defect development and the diagnosis of welding quality.

[0263] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0264] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery tab welding quality detection method, characterized in that: X-ray images and electrical performance data are synchronously collected through a preset global clock signal to construct a time series dataset sorted by timestamps; an initial feature point tracking algorithm is used to generate a geometric transformation matrix between adjacent X-ray image frames, specifically including: feature detection is performed on the first frame of X-ray image to extract initial feature points of the battery tab welding area; the initial feature points include edge corner coordinates and welding point center coordinates of the tab welding area; based on the optical flow method, the initial feature points in the X-ray image are tracked to generate a coordinate sequence of the initial feature points in each frame of X-ray image, and a feature point motion trajectory sequence is obtained; based on the feature point motion trajectory sequence, a least squares method is used to calculate a translation matrix, a rotation matrix and a scaling factor between adjacent X-ray image frames to form a geometric transformation matrix between adjacent X-ray image frames; the X-ray image time series data in the time series dataset are spatially aligned based on the geometric transformation matrix to obtain position-aligned X-ray image time series data, specifically including: let index variable k = 1, the value range of k is [1, K]; index variable q = 1, the value range of q is [1, Q]; K is the number of X-ray image frames, and Q is the number of pixel points of each X-ray image; S101: for the qth new pixel point of the kth frame of X-ray image in the time series dataset after being affected by the geometric transformation matrix, the sampling position corresponding to the qth new pixel point before being affected by the geometric transformation matrix is calculated according to the inverse matrix of the geometric transformation matrix, and is marked as the original sampling position; S102: the S nearest integer coordinate pixel points are determined with the qth new pixel point at the original sampling position as the center, and the gray values of the S pixel points are calculated; S103: the weight coefficients are calculated according to the relative distances between the qth new pixel point at the original sampling position and the S pixel points; S104: the gray values of the S pixel points are multiplied by the corresponding weight coefficients respectively, and then summed to obtain the gray value of the qth new pixel point after being affected by the geometric transformation matrix; S105: if q is less than Q, then q is increased by 1, and S101 to S105 are executed; if q is greater than or equal to Q, and k is less than k, then q = 0, k is increased by 1, and S101 to S105 are executed; if q is greater than or equal to Q, and k is greater than or equal to k, then the spatial position transformation of the X-ray image time series data is completed, and the position-aligned X-ray image time series data is obtained; time series features are extracted from the electrical performance time series data in the time series dataset to generate an electrical performance time series feature vector; the position-aligned X-ray image time series data and the electrical performance time series feature vector are input into a preset neural network model to output a predicted defect development probability, and the welding quality is diagnosed; the training method of the neural network model includes: obtaining a training dataset, the training dataset including timestamped position-aligned X-ray image time series data samples, corresponding electrical performance time series data samples and artificially labeled welding quality label vectors; constructing a double-branch neural network model, the double-branch neural network model including an image feature extraction branch, a time series fusion branch and a fully connected classification layer; The position-aligned X-ray image time series data is input into an image feature extraction branch to output an image feature vector; The image feature vector and the electrical performance time series data are input into a time series fusion branch to generate a fusion feature vector through a splicing operation; The fusion feature vector is input into a fully connected classification layer to output a defect probability vector; The binary cross-entropy loss value of each dimension of the defect probability vector is calculated and averaged to obtain a comprehensive loss value; The trainable parameters of the double-branch neural network model are updated using a backpropagation algorithm until the comprehensive loss value converges, and a trained neural network model is obtained.

2. The battery tab weld quality detection method of claim 1, wherein, The method for constructing a time series dataset sorted by timestamps comprises: A global clock signal is generated using a synchronous clock to synchronously trigger the acquisition actions of the X-ray imaging device and the electrical performance tester to obtain X-ray images and electrical performance data respectively; the X-ray image data includes the pixel point coordinates of each X-ray image and the grayscale values corresponding to the pixel points; the electrical performance data includes the voltage of the solder joint, the current of the solder joint, the temperature of the electrode, and the dynamic resistance; A timestamp is embedded for each X-ray image and each set of electrical performance data; For each X-ray image, the nearest corresponding data set with a timestamp in the electrical performance data within a pre-set time search range is searched for matching to obtain an initial time series dataset; When the sampling frequency of the electrical performance data is higher than the frame rate of the X-ray images, for the X-ray image frames without direct matching data, the electrical performance data at the corresponding time is generated through a linear interpolation method; The initial time series dataset is arranged in ascending order of timestamps to construct a time series dataset sorted by timestamps.

3. The battery tab weld quality detection method of claim 1, wherein, The method for generating an electrical performance time series feature vector comprises: The mean and standard deviation of each type of data in the electrical performance time series data are calculated; Each data of each type of data in the electrical performance time series data is subtracted by the mean of the corresponding type and divided by the standard deviation of the corresponding type to generate standardized electrical performance time series data; The size of the sliding window and the sliding step are determined, and starting from the first data set of the standardized electrical performance time series data, data segments of the size of the sliding window are sequentially intercepted according to the sliding step to generate multiple electrical performance time series data segments; Fourier transform is performed on each electrical performance time series data segment to generate a frequency domain feature vector; The timestamp corresponding to each frequency domain feature vector is determined; According to the order of the timestamps, all the frequency domain feature vectors are connected end to end to obtain an electrical performance time series feature vector.

4. The battery tab weld quality detection method of claim 1, wherein, The method for predicting defect development probability and diagnosing welding quality comprises: The position-aligned X-ray image time series data is input into the image feature extraction branch of a pre-set neural network model to output an image feature vector; The image feature vector and the electrical performance time series feature vector are input into the time series fusion branch of the pre-set neural network model to output a fusion feature vector; The fusion feature vector is input into the fully connected classification layer of the pre-set neural network model to output a defect probability vector; Based on the defect probability vector, the defect type is determined through a pre-set defect threshold judgment rule to generate a welding quality diagnosis result.

5. The battery tab welding quality detection method according to claim 4, characterized in that: The defect types include virtual welding of the welding spot, welding through, crack of the welding spot, tab deviation, uneven welding, and abnormal size of the welding spot; The defect threshold judging rule is to compare the probability value of each element in the defect probability vector with the preset defect threshold of the corresponding defect type, and if the probability value of an element in the defect probability vector is greater than the corresponding defect threshold, it is determined that there is a defect of this type; According to the size relationship of the probability values of the elements in the defect probability vector, the severity level of the defect is determined.

6. The battery tab weld quality detection method of claim 5, wherein, The method for determining the severity level of the defect includes: The elements whose probability values exceed the preset defect threshold of the corresponding defect type are sorted in descending order of the probability values; If there are multiple defect types exceeding the defect threshold, multiply the probability values of the defect types by the corresponding preset influence factors and sum them up to obtain a comprehensive influence value; the severity level is determined according to the interval in which the comprehensive influence value is located; If only a single defect type exceeds the defect threshold, the severity level is directly divided according to the deviation degree of the probability value of the defect type from the threshold; the severity level includes mild, moderate, and severe.

7. A battery tab welding quality detection system for implementing the battery tab welding quality detection method according to any one of claims 1 to 6, characterized in that, It includes: A data acquisition module acquires X-ray images and electrical performance data synchronously through a preset global clock signal, and constructs a time series data set sorted by time stamp; A geometric transformation module generates a geometric transformation matrix between adjacent X-ray image frames using an initial feature point tracking algorithm; A position alignment module performs spatial position alignment on the X-ray image time series data in the time series data set based on the geometric transformation matrix, and obtains position-aligned X-ray image time series data; A feature extraction module is used for time series feature extraction on the electrical performance time series data in the time series data set, and generates an electrical performance time series feature vector; A prediction and diagnosis module is used for inputting the position-aligned X-ray image time series data and the electrical performance time series feature vector into a preset neural network model, and outputting a predicted defect development probability and diagnosing the welding quality.

Images