Defect detection method and device of secondary battery, terminal equipment and storage medium

Abstract

The invention discloses a defect detection method and system for a secondary battery and a storage medium, and relates to the technical field of intelligent manufacturing of batteries. The method comprises the following steps: accurately associating winding alignment degree (X-ray) and tab welding spot (visible light) data on a high-speed production line to a unique battery cell ID through a time axis synchronous triggering mechanism, and constructing a multi-modal feature sequence; the method focuses on critical morphology characteristics of which physical quantities are at qualified boundaries, the critical morphology characteristics are input into a deep learning model containing physical mechanism constraints, a nonlinear coupling relationship between morphology and electrochemical performance is analyzed, and failure probability distribution is output. And based on the risk identification result, physical interception is executed before capacity grading formation, or an adaptive drift compensation signal is sent to front-end manufacturing equipment. According to the method, the problem of high-speed data alignment is solved, and the spanning from microstructure detection to electrochemical performance estimation and process closed-loop self-healing is realized.

Description

Technical Field

[0001] This invention relates to the field of battery testing, and more particularly to a method, apparatus, terminal equipment, and storage medium for detecting defects in secondary batteries. Background Technology

[0002] With the explosive growth of the new energy vehicle and electrochemical energy storage industries, the manufacturing process of secondary batteries (especially lithium-ion power batteries) is evolving towards extremely high efficiency (production cycle ≥ 200" PPM) and extreme manufacturing precision.

[0003] However, existing high-speed production line quality control technologies still face significant technical bottlenecks. First, the problem of "spatiotemporal misalignment" in data association is severe. Due to the large physical span of processes such as winding and welding, and the extremely high speed of conveyor belts, traditional testing equipment often operates independently, making it difficult to accurately bind X-ray imaging data and visible light appearance data obtained from different workstations to the same cell's digital identifier (ID) within a millisecond-level time window, resulting in a broken data traceability chain.

[0004] Secondly, traditional AOI (Automated Optical Inspection) logic suffers from "survivorship bias." Existing technologies often employ a "hard threshold" judgment based on fixed rules (Pass / Fail), which can only eliminate obviously defective products with severely out-of-size dimensions. Traditional algorithms cannot identify "sub-healthy" cells (i.e., those with critical morphological characteristics) whose physical parameters are within acceptable range but are approaching critical boundaries or exhibiting abnormal statistical distributions. These cells are highly susceptible to electrochemical failures such as lithium plating and internal short circuits during subsequent long charge-discharge cycles. Summary of the Invention

[0005] This invention provides a method, apparatus, terminal device, and storage medium for defect detection of secondary batteries, enabling precise synchronization of multimodal data on high-speed production lines, in-depth pre-detection of electrochemical performance risks based on physical mechanisms, and adaptive closed-loop compensation for processes.

[0006] To achieve the above objectives, a first aspect of the present invention provides a defect detection method for secondary batteries, comprising:

[0007] The time-axis synchronous triggering mechanism is used to acquire the first imaging detection data of the secondary battery cell under test at the winding alignment station and the second imaging detection data at the tab solder joint detection station on the high-speed production line; the high-speed production line refers to a production line with a production rate greater than a preset threshold.

[0008] Based on the first imaging detection data, the second imaging detection data and the unique digital identifier of the secondary battery cell under test, a multimodal physical morphological feature sequence of the secondary battery cell under test is constructed.

[0009] The critical morphological features characterizing the performance fluctuation of the battery cell are extracted from the multimodal physical morphological feature sequence. The critical morphological features refer to the morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution.

[0010] The critical morphological features are input into a pre-trained performance correlation model; the performance correlation model is constructed based on the nonlinear coupling relationship between morphological features in historical production data and the true electrical performance after a preset cycle life stage, and the performance correlation model includes a deep learning architecture constrained by physical mechanisms.

[0011] The electrical performance prediction probability distribution of the secondary battery cell under test is calculated based on the performance correlation model, and the electrochemical failure risk of the secondary battery cell under test after the preset cycle life stage is identified based on the probability distribution.

[0012] Based on the risk identification results, before the secondary battery cell under test enters the capacity grading process, the production control system outputs a sorting and interception command or performs adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process.

[0013] In one possible implementation of the first aspect, the acquisition of first imaging detection data at the winding alignment station and second imaging detection data at the tab solder joint detection station on a high-speed production line via a time-axis synchronization triggering mechanism specifically includes:

[0014] An incremental encoder is installed on the main conveyor belt drive shaft of a high-speed production line.

[0015] The incremental encoder pulses are received through an FPGA-based real-time motion control card; the real-time motion control card includes a first trigger port and a second trigger port, the first trigger port being connected to an X-ray detector at the winding alignment station, and the second trigger port being connected to an industrial camera at the welding station.

[0016] When the feed sensor detects that the secondary battery cell to be tested has entered the conveyor belt, a unique digital identifier for the secondary battery cell to be tested is generated and written into the first address of the register;

[0017] The logical coordinates of the secondary battery cell under test in the register are updated in real time according to the increase of the incremental encoder pulse.

[0018] When the logical coordinates match the preset shooting point of the X-ray detector, the real-time motion control card outputs a first hardware trigger signal through the first trigger port; it collects the fluoroscopic image data of the X-ray detector and adds the current timestamp and the unique digital identifier, and uses the fluoroscopic image data as the first imaging detection data;

[0019] When the logical coordinates match the preset shooting point of the industrial camera, the real-time motion control card outputs a second hardware trigger signal through the second trigger port; it collects surface morphology image data from the industrial camera and adds the current timestamp and the unique digital identifier, using the surface morphology image data as the second imaging detection data.

[0020] In one possible implementation of the first aspect, the step of constructing a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data, and the unique digital identifier of the secondary battery cell under test specifically includes:

[0021] Using the edge of the secondary battery cell casing as a reference point, calculate the rotation and translation matrices, and calibrate the first imaging detection data and the second imaging detection data to the standard coordinate system;

[0022] The first imaging detection data and the second imaging detection data are subjected to ROI cropping and normalization processing;

[0023] The first imaging detection data and the second imaging detection data are stitched together in the channel dimension to form a multi-channel tensor;

[0024] After passing through the fully connected layer, the winding tension value and welding power value associated with the unique digital identifier are encoded into vectors and concatenated with the multi-channel tensor to obtain the multimodal physical morphological feature sequence of the secondary battery cell under test.

[0025] In one possible implementation of the first aspect, the ROI cropping and normalization processing of the first imaging detection data and the second imaging detection data specifically includes:

[0026] Scan along the grayscale gradient direction of the first imaging detection data to locate the boundary line between the positive and negative electrode sheets. Using the boundary line as the central axis, expand the preset pixel width to both sides to crop out a first rectangular image containing only the Overhang region.

[0027] Using template matching, the geometric center of the weld stamp in the second imaging detection data is located, and a second rectangular image covering the entire heat-affected zone is cropped out with the geometric center of the weld stamp as the center of the rectangle.

[0028] A bilinear interpolation algorithm is used to scale the first rectangular image and the second rectangular image to the same size, resulting in the processed first imaging detection data and the processed second imaging detection data.

[0029] In one possible implementation of the first aspect, the deep learning architecture constrained by the physical mechanism specifically refers to:

[0030] An electrochemical mechanism regularization term is introduced into the loss function of a deep learning network; the electrochemical mechanism regularization term includes potential constraints based on the Nernst equation, impedance constraints based on the equivalent circuit model, and concentration distribution constraints based on ion diffusion kinetics.

[0031] In one possible implementation of the first aspect, identifying the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the predicted electrical performance probability distribution specifically includes:

[0032] When the cumulative probability of exceeding the failure threshold in the electrical performance prediction probability distribution is greater than the preset risk tolerance, it is determined to have potential failure risk; the electrical performance prediction probability distribution includes the confidence interval of capacity retention rate and the probability density function of DC internal resistance growth.

[0033] In one possible implementation of the first aspect, the step of outputting a sorting and interception command through the production control system or adaptively compensating for the execution parameters of the corresponding equipment in the front-end manufacturing process before the secondary battery cell under test enters the capacity testing process, based on the risk identification result, specifically includes:

[0034] Based on the predicted probability distribution of electrical performance, the risk level of the secondary battery cell is determined; if the cumulative probability of the failure threshold is greater than the first failure threshold, the secondary battery cell is at level one risk; if the cumulative probability of the failure threshold is greater than or equal to the second failure threshold and less than or equal to the first failure threshold, the secondary battery cell is at level two risk; if the cumulative probability of the failure threshold is less than the second failure threshold, the secondary battery cell has no risk level; the first failure threshold is greater than the second failure threshold.

[0035] If the secondary battery cell is at level one risk, a sorting and interception command will be output through the production control system.

[0036] If the secondary battery cell is at level two risk, the average offset of the winding alignment of the most recent cells is calculated in real time, and a position compensation signal for the winding machine correction controller or a tension adjustment signal for the tension controller is generated; the shrinkage trend of the weld nugget area or weld depth is monitored in real time, and a laser power adjustment signal or welding pulse width modulation signal for the laser welding machine is generated.

[0037] A second aspect of the present invention provides a defect detection device for secondary batteries, comprising:

[0038] The imaging module is used to acquire the first imaging detection data of the secondary battery cell under test at the winding alignment station and the second imaging detection data at the tab solder joint detection station on the high-speed production line through a time axis synchronous triggering mechanism; the high-speed production line refers to a production line with a production rate greater than a preset threshold.

[0039] The sequence module is used to construct a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data and the unique digital identifier of the secondary battery cell under test.

[0040] The extraction module is used to extract critical morphological features characterizing cell performance fluctuations from the multimodal physical morphological feature sequence; the critical morphological features refer to morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution.

[0041] The input module is used to input the critical morphological features into a pre-trained performance correlation model; the performance correlation model is constructed based on the nonlinear coupling relationship between morphological features in historical production data and the true values ​​of electrical performance after a preset cycle life stage, and the performance correlation model includes a deep learning architecture constrained by physical mechanisms.

[0042] The detection module is used to calculate the predicted probability distribution of the electrical performance of the secondary battery cell under test according to the performance correlation model, and to identify the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the probability distribution.

[0043] The adjustment module is used to, based on the risk identification results, output sorting and interception commands through the production control system or to perform adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process before the secondary battery cell under test enters the capacity grading process.

[0044] A third aspect of the present invention provides a terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement a defect detection method for a secondary battery as described above.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a defect detection method for a secondary battery as described above.

[0046] At the data acquisition level, this invention employs a time-axis synchronous triggering mechanism. A high-precision encoder installed on the main shaft of the production line monitors the cell position in real time, and combined with FPGA hardware control, achieves microsecond-level synchronous triggering. This ensures that the imaging data from the two key workstations—winding alignment and tab soldering—precisely correspond to the same individual cell. At the feature extraction level, statistical criteria for critical morphological features are defined: morphological features whose boundary value difference from a preset acceptable range is less than or equal to a preset boundary threshold, and whose physical quantity value falls within a preset range of normal distribution, are considered potential risk indicators. At the predictive modeling level, a deep learning architecture constrained by physical mechanisms is constructed. Basic electrochemical principles (such as the Nernst equation and equivalent circuit models) are used as constraints for network training, enabling the model to not only fit historical data distributions but also conform to electrochemical physical laws, outputting a probability distribution of electrical performance predictions rather than a single, definitive value. At the control execution level, a closed-loop feedback mechanism is employed to dynamically generate sorting instructions or process compensation signals based on the risk level, achieving real-time quality control.

[0047] In summary, this invention completely solves the problem of accurate correlation of multi-source heterogeneous data in high-speed production environments through a hardware-level synchronization mechanism, providing a spatiotemporally consistent data foundation for quality analysis; the statistical definition of critical morphological features breaks through the limitation of traditional detection focusing only on obvious defects, and can accurately identify potentially failed cells that are "qualified but high-risk"; the deep learning model constrained by physical mechanisms significantly improves the reliability and interpretability of predictions, realizing cross-scale mapping from physical morphological features to long-term electrochemical performance; the adaptive drift compensation mechanism transforms quality control from passive interception to active prevention: after obtaining the pre-detection results, real-time process parameter optimization eliminates potential quality hazards from the source, greatly reducing resource waste and product failure risks in subsequent expensive processes. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating a defect detection method for a secondary battery according to an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of a defect detection device for secondary batteries according to an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To resolve the above issues, please refer to [link / reference]. Figure 1An embodiment of the present invention provides a defect detection method for secondary batteries, comprising:

[0052] S10. Through the time axis synchronization triggering mechanism, acquire the first imaging detection data of the winding alignment station and the second imaging detection data of the tab solder joint detection station of the secondary battery cell under test on the high-speed production line; the high-speed production line refers to a production line with a production rate greater than a preset threshold.

[0053] S11. Based on the first imaging detection data, the second imaging detection data, and the unique digital identifier of the secondary battery cell under test, construct a multimodal physical morphological feature sequence of the secondary battery cell under test.

[0054] S12. Extract critical morphological features characterizing cell performance fluctuations from the multimodal physical morphological feature sequence; the critical morphological features refer to morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution.

[0055] S13. Input the critical morphological features into a pre-trained performance correlation model; the performance correlation model is constructed based on the nonlinear coupling relationship between the morphological features in historical production data and the true values ​​of electrical performance after a preset cycle life stage, and the performance correlation model includes a deep learning architecture with physical mechanism constraints.

[0056] S14. Calculate the electrical performance prediction probability distribution of the secondary battery cell under test according to the performance correlation model, and identify the electrochemical failure risk of the secondary battery cell under test after the preset cycle life stage based on the probability distribution.

[0057] S15. Based on the risk identification results, before the secondary battery cell to be tested enters the capacity grading process, the production control system outputs a sorting and interception command or performs adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process.

[0058] Steps S10 and S11 involve the high-speed synchronous acquisition and sequence construction of multimodal data. The core of these two steps lies in achieving precise "spatiotemporal alignment" between the physical entity (battery cell) and digital information (image) under high-speed motion. An encoder is deployed on the drive shaft of the high-speed production line. When a battery cell is loaded, a unique digital identifier (ID) is generated and written to a virtual shift register. As the conveyor belt runs, the encoder emits pulses, and the logical coordinates of this ID in the register shift accordingly. When the logical coordinates reach the preset physical trigger points at the winding alignment station (usually an X-ray detector) and the tab solder joint station (usually an industrial camera), the system sends a hardware-level synchronization signal to acquire images. Finally, using this ID as an index, the system cleans, corrects, and stitches heterogeneous image data (perspective data and surface data) collected at different times and locations to construct a multimodal physical morphological feature sequence belonging to the battery cell.

[0059] This phase combines hardware-level pulse synchronization with software-level virtual tracking to effectively overcome the data silo effect and information mismatch problems caused by the large physical span of equipment and fast operating cycle in high-speed production lines. It ensures that the physical morphology data at the micro level can be accurately traced back to each individual cell, providing a high-quality multidimensional data foundation with zero latency, zero error and spatiotemporal consistency for subsequent refined analysis.

[0060] S12 is the critical morphology feature extraction process: A critical morphology feature is defined as follows: when the physical quantity value X satisfies |X - X_boundary| ≤ Δ_threshold and μ - kσ ≤ X ≤ μ + kσ, it is determined to be a critical feature. Here, X_boundary is the preset acceptable range boundary value, Δ_threshold is the preset boundary threshold (usually 5%-15% of the acceptable range width), μ is the historical data mean, σ is the historical data standard deviation, and k is the control coefficient (usually 2, corresponding to the 95% confidence interval). For example, for the diaphragm-electrode alignment deviation in winding alignment, if the acceptable range is 0.1-0.3 mm, when the measured value is 0.28 mm (close to the upper limit boundary of 0.3 mm) and falls within the 2σ range of historical data, it is identified as a critical feature.

[0061] This design breaks through the limitations of traditional testing that only focuses on obvious defects. It can accurately identify "qualified but high-risk" critical state cells, realizing the transformation from static quality judgment to dynamic risk prediction. It effectively captures those potentially problematic cells that are on the verge of being qualified but may fail prematurely during long-term use, providing a scientific basis for preventive quality control.

[0062] The S13-S14 performance pre-detection and risk identification process employs a deep learning architecture constrained by physical mechanisms. Fundamental electrochemical principles (such as the electrode potential relationship described by the Nernst equation, the impedance characteristics represented by the equivalent circuit model, and the ion diffusion behavior reflected by Fick's law) serve as constraints for network training. The model learns the nonlinear mapping relationship between morphological features in historical data and the true electrical performance values ​​after cycle life (e.g., capacity retention after 500 cycles), outputting a probability distribution of predicted electrical performance rather than a single, definitive value. Risk identification is based on this probability distribution; when the cumulative probability exceeding the failure threshold (e.g., capacity retention <80%) is greater than a preset risk tolerance, an electrochemical failure risk is identified.

[0063] The introduction of physical mechanism constraints significantly improves the model's generalization ability and physical consistency, avoids the risk of overfitting in small sample cases by pure data-driven methods, and makes the prediction results not only accurate and reliable, but also interpretable. It can identify potential failure risks that cannot be found in conventional detection in advance, and provide a scientific quantitative basis for quality decision-making.

[0064] S15 is the adaptive control execution process: based on closed-loop control theory, a hierarchical control strategy is implemented according to the risk identification results: for high-risk cells, a sorting and interception command is directly output; for medium-risk cells, the characteristic change trends of multiple consecutive cells are analyzed, the process deviation is calculated, and an adaptive drift compensation signal is generated. For example, when the winding alignment shows a systematic trend of being too large, the tension control parameters of the winding machine are automatically adjusted; when the weld point area coverage gradually decreases, the power parameters of the welding equipment are dynamically increased, forming a closed-loop control chain of "detection-analysis-compensation".

[0065] This stage extends quality inspection to the decision-making center of intelligent manufacturing, achieving the dual effects of prevention and process self-healing. By intercepting potential waste in the low-cost process stage, the waste of high-value materials and energy and safety hazards are greatly reduced. At the same time, through real-time fine-tuning and compensation of front-end equipment, process fluctuations can be dynamically suppressed to prevent batch quality accidents, thereby significantly improving the yield stability and overall equipment efficiency of the production line without human intervention.

[0066] In summary, the aforementioned secondary battery defect detection method constructs a complete "perception-analysis-prediction-control" closed-loop quality control system through four core components: multi-station synchronous data acquisition, intelligent identification of critical features, performance prediction driven by physical mechanisms, and adaptive process compensation. This system not only accurately identifies potential defects but also proactively optimizes manufacturing processes, fundamentally improving battery quality and production efficiency.

[0067] For example, the acquisition of first imaging detection data at the winding alignment station and second imaging detection data at the tab solder joint detection station on a high-speed production line via a time-axis synchronization triggering mechanism specifically includes:

[0068] An incremental encoder is installed on the main conveyor drive shaft of a high-speed production line.

[0069] The incremental encoder pulses are received via an FPGA-based real-time motion control card. The real-time motion control card includes a first trigger port and a second trigger port. The first trigger port is connected to an X-ray detector at the winding alignment station, and the second trigger port is connected to an industrial camera at the welding station.

[0070] When the feed sensor detects that the secondary battery cell to be tested has entered the conveyor belt, it generates a unique digital identifier for the secondary battery cell to be tested and writes the unique digital identifier into the first address of the register.

[0071] The logical coordinates of the secondary battery cell under test in the register are updated in real time according to the increase in the incremental encoder pulse.

[0072] When the logical coordinates match the preset shooting point of the X-ray detector, the real-time motion control card outputs a first hardware trigger signal through the first trigger port; it collects the fluoroscopic image data of the X-ray detector and adds the current timestamp and the unique digital identifier, and uses the fluoroscopic image data as the first imaging detection data.

[0073] When the logical coordinates match the preset shooting point of the industrial camera, the real-time motion control card outputs a second hardware trigger signal through the second trigger port; it collects surface morphology image data from the industrial camera and adds the current timestamp and the unique digital identifier, using the surface morphology image data as the second imaging detection data.

[0074] The implementation principle of this embodiment is based on precision motion control and hardware-level synchronization technology. An incremental encoder is mounted on the main conveyor belt drive shaft, generating pulse signals by detecting changes in rotation angle, accurately converting the physical position of the battery cell into a digital quantity. An FPGA-based real-time motion control card serves as the core control unit; its hardware parallel processing architecture ensures a microsecond-level response speed. Two independent trigger ports control imaging devices at different workstations, achieving precise synchronization at the hardware level. The system employs a register mapping mechanism. When the feed sensor detects a battery cell, it assigns a unique digital identifier and establishes an initial position mapping. As encoder pulses accumulate, the system calculates the battery cell's movement distance in real time and updates its logical coordinates in the register (i.e., its digital position relative to the production line start point). This logical coordinate system essentially maps physical spatial positions to a digital coordinate system. By matching the preset shooting point coordinates with the real-time logical coordinates, the corresponding imaging device is triggered. The dual-marking mechanism of timestamps and unique identifiers ensures that the acquired image data can be accurately traced to a specific battery cell and the acquisition time, forming a complete spatiotemporal data chain.

[0075] For example, when cell A enters the production line, the feed sensor assigns it the ID "A001" and sets its initial logical coordinate to 0. As the conveyor belt moves, the encoder generates 1000 pulses per revolution, and the system calculates the cell's movement distance based on the pulse count. When the logical coordinate reaches the preset value of the X-ray station (e.g., coordinate 1500), the FPGA immediately outputs the first trigger signal. Upon receiving the signal, the X-ray detector begins exposure and acquisition, and the system simultaneously timestamps the image "2024-07-15 14:30:25.123" and assigns the ID "A001". When the cell continues to move to logical coordinate 3000 (the preset position of the industrial camera), the FPGA outputs the second trigger signal, completing the synchronous acquisition and marking of the surface image.

[0076] This time-axis synchronous triggering mechanism completely solves the problem of spatiotemporal synchronization in multi-station data acquisition under high-speed production environments. Through hardware-level precise control and logical coordinate mapping, it ensures that detection data from different process stages and different modes can be accurately associated with the same individual cell, fundamentally eliminating the delay jitter and data misalignment problems existing in traditional software triggering methods. The dual binding mechanism of unique identifier and timestamp provides a complete and reliable data foundation for subsequent quality traceability and data analysis, making cross-dimensional correlation analysis from physical morphological characteristics to electrochemical performance possible. This lays a solid technical foundation for intelligent quality control in the battery manufacturing process. At the same time, the real-time and deterministic nature of hardware triggering significantly improves the system's anti-interference capability and long-term operational stability, adapting to the reliable operation requirements of complex electromagnetic environments in industrial sites.

[0077] For example, the step of constructing a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data, and the unique digital identifier of the secondary battery cell under test specifically includes:

[0078] The rotation and translation matrices are calculated using the edge of the secondary battery cell casing as a reference point, and the first imaging detection data and the second imaging detection data are calibrated to the standard coordinate system.

[0079] The first imaging detection data and the second imaging detection data are subjected to ROI cropping and normalization processing.

[0080] The first imaging detection data and the second imaging detection data are stitched together in the channel dimension to form a multi-channel tensor.

[0081] After passing through the fully connected layer, the winding tension value and welding power value associated with the unique digital identifier are encoded into vectors and concatenated with the multi-channel tensor to obtain the multimodal physical morphological feature sequence of the secondary battery cell under test.

[0082] In the calibration, cropping, and normalization process, the core issue is eliminating rigid body displacement noise caused by high-speed motion. In machine vision algorithms, firstly, the edges of the battery cell casing are extracted using edge detection algorithms, and the pose deviation of the current image coordinate system relative to the preset standard template coordinate system is calculated. Based on this deviation, an affine transformation matrix is ​​constructed. and using the formula Each pixel coordinate in the image is remapped. Then, the region of interest is cropped from the calibrated image, and the pixel grayscale values ​​are linearly normalized.

[0083] In the above formula, P=[ x , y , 1 ] T : Represents the homogeneous coordinate vector of any pixel in the original image, where These are the original x and y coordinates; P ' =[ x ' , y ' , 1 ] T : Represents the homogeneous coordinate vector of the corresponding pixel in the calibrated image, where The calibrated x and y coordinates; :express The affine transformation matrix, whose specific form is usually as follows: ; : Indicates the rotation angle deviation of the battery cell under test relative to the standard orientation; : These represent the translational distance deviations of the battery cell under test in the horizontal and vertical directions, respectively; : Represents the original pixel grayscale value before normalization (e.g., 0~255); : Represents the normalized pixel floating-point value (e.g., 0~1); : These represent the maximum and minimum grayscale values ​​allowed by the image sensor's measurement range, respectively.

[0084] By constructing a transformation matrix that includes rotation and translation factors, inverse geometric correction is performed on the high-speed acquired images, effectively eliminating random errors in imaging position caused by mechanical transmission vibration, fixture positioning gaps, or conveyor belt deviation. This ensures that each frame of image input to the neural network has a high degree of consistency in spatial distribution, enabling the model to focus on the morphological fluctuations of the battery cell itself rather than rigid body displacement noise, significantly improving the signal-to-noise ratio of feature extraction.

[0085] Then, a multi-channel tensor is formed by concatenating the data along the channel dimension. The winding tension value and welding power value are encoded as vectors and concatenated. This step aims to construct a holographic feature space. First, the X-ray fluoroscopic image and the visible light image are stacked along the channel dimension to preserve spatial structure information. Second, a fully connected layer is used as an encoder to map discrete process scalars (tension, power) into high-dimensional feature vectors, enabling them to be fused with image features. It should be noted that the winding tension value and welding power value can be obtained by extracting the production process parameters corresponding to the secondary battery cell under test.

[0086] : Represents the concatenated image feature tensor, whose dimension is denoted as . ,in This represents the total number of channels. Image height, Image width; : Represents the input process parameter vector, for example x scalar =[ v tension , v power ] T ,in This is the measured value of the winding tension. This is the measured value of welding power; : Represents the encoded high-dimensional feature vector (Embedding Vector); : Represents the weight matrix of the fully connected layer, used to learn the mapping weights between process parameters and image features; : Represents the bias vector of the fully connected layer; : Represents a non-linear activation function (such as ReLU), and the encoding process can be represented as .

[0087] This step achieves deep feature fusion of heterogeneous data, breaking down the data barrier between visual images and process parameters. By mapping low-dimensional process scalars to high-dimensional feature space through weight matrices and combining them with image tensors, the model can not only identify physical defects, but also uncover the deep nonlinear coupling relationship between "process parameter settings (such as tension and power)" and "final morphology formation", thereby greatly improving the ability to analyze the causes of complex defects and the robustness of performance prediction.

[0088] For example, the ROI cropping and normalization processing of the first imaging detection data and the second imaging detection data specifically includes:

[0089] Scan along the grayscale gradient direction of the first imaging detection data to locate the boundary line between the positive and negative electrode sheets. Using the boundary line as the central axis, expand the preset pixel width to both sides to crop out a first rectangular image containing only the Overhang region.

[0090] Using template matching, the geometric center of the weld stamp in the second imaging detection data is located, and a second rectangular image covering the entire heat-affected zone is cropped out with the geometric center of the weld stamp as the center of the rectangle.

[0091] A bilinear interpolation algorithm is used to scale the first rectangular image and the second rectangular image to the same size, resulting in the processed first imaging detection data and the processed second imaging detection data.

[0092] For the first imaging detection data (usually an X-ray fluoroscopic image), the system scans and analyzes along the direction of maximum gray-level gradient. The gray-level gradient reflects the degree of drastic change in pixel values. The gradient magnitude of each pixel is calculated. G(x,y) = √[( ∂ I / ∂ x) ² + ( ∂ I / ∂ y) ² ] (Where I is the pixel grayscale value, and x and y are the coordinate positions), the gradient peak position at the interface between the positive and negative electrode materials is located, thereby accurately identifying the interface line. Based on this interface line, a fixed pixel width is symmetrically extended to both sides to ensure that the clipping area completely covers the Overhang area (i.e., the key area where the negative electrode extends beyond the edge of the positive electrode). For the second imaging detection data (usually an optical surface image), a template matching algorithm is used. A standard solder joint shape template T(u,v) is predefined. By calculating the normalized cross-correlation function NCC(x,y) = Σ[T(u,v)·I(x+u,y+v)] / √[ΣT²(u,v)·ΣI²(x+u,y+v)] (where I is the current image, and u and v are the template coordinates), the position that best matches the template in the image is searched to accurately locate the geometric center of the solder joint. A rectangular area containing the complete heat-affected zone is clipped from this center.

[0093] This region positioning and trimming mechanism effectively focuses on key quality feature areas that affect battery performance. It accurately captures the material interface features of the internal structure through grayscale gradient analysis and accurately identifies the core areas of surface welding quality through template matching. This not only significantly reduces the amount of data and computational complexity in subsequent processing, but also ensures the relevance and effectiveness of feature extraction, avoids interference from irrelevant background information, and significantly improves the accuracy and efficiency of quality analysis, laying a solid foundation for the accurate identification of critical morphological features.

[0094] A bilinear interpolation algorithm is used to scale two rectangular images to the same size. X-ray detectors and industrial cameras typically have different pixel pitches, resulting in cropped images with corresponding physical dimensions but different pixel matrix dimensions. The bilinear interpolation algorithm achieves smooth image scaling by calculating the weighted average of the four nearest pixels surrounding the target pixel. For any pixel P(x,y) in the target image, its grayscale value I(x,y) is calculated from the grayscale values ​​of the four nearest pixels in the original image. The results were obtained through two linear interpolation calculations: First, interpolation in the x-direction was performed. and Then, interpolation is performed in the y-direction to obtain the final value P. This interpolation method effectively avoids the jagged effect caused by nearest neighbor interpolation and the computational complexity caused by bicubic interpolation while preserving image details, making it particularly suitable for industrial real-time processing scenarios.

[0095] This standardized processing mechanism unifies the data format of images with different imaging modalities and resolutions, ensuring the consistency and comparability of multi-source data in terms of spatial scale. It provides a standardized data foundation for subsequent feature fusion and model input. At the same time, bilinear interpolation achieves the best balance between computational efficiency and image quality while maintaining key image features, enabling the system to complete image preprocessing in real time in a high-speed production environment, ensuring the timeliness and reliability of the entire quality control process.

[0096] The ROI cropping and normalization implementation demonstrates an efficient and robust image preprocessing workflow through precise region localization, targeted feature extraction, and standardized data processing. Its innovation lies in the deep integration of edge detection, template matching, and spatial transformation techniques from computer vision. This achieves automated conversion from raw images to standardized feature inputs, significantly improving data processing efficiency and feature extraction accuracy while ensuring the complete preservation and accurate representation of key quality features. This provides high-quality data support for subsequent critical morphology feature recognition and electrical performance prediction.

[0097] For example, the deep learning architecture constrained by the physical mechanism specifically refers to:

[0098] An electrochemical mechanism regularization term is introduced into the loss function of a deep learning network; the electrochemical mechanism regularization term includes potential constraints based on the Nernst equation, impedance constraints based on the equivalent circuit model, and concentration distribution constraints based on ion diffusion kinetics.

[0099] The implementation principle of this embodiment is based on the Physical Information Neural Network (PINN) theory. By embedding fundamental electrochemical laws as regularization constraints into the loss function of the deep learning network, the model output not only conforms to the historical data distribution but also strictly follows the laws of electrochemical physics. Specifically, the total loss function is constructed as: L_total = L_data + λ1·L_nernst + λ2·L_impedance +λ3·L_diffusion, where L_data is the data fitting loss term (such as mean squared error). , , The regularization weight coefficient (usually 0.2 ≤ ≤0.5), respectively controlling the strength of each physical constraint.

[0100] The Nernst equation potential constraint is based on the principle of electrode electrochemical equilibrium, and its constraint terms are defined as follows: ,in The open-circuit voltage predicted by the model. The theoretical electric potential calculated for the Nernst equation. Parameter definitions: Standard electrode potential , Gas constant , absolute temperature , For electron transfer number, Faraday constant , It is the reaction quotient (reflecting the concentration ratio of the active substance).

[0101] The impedance constraint of the equivalent circuit model is based on the physical characterization of the battery's AC impedance spectrum, and the constraint terms are defined as follows: ,in The frequency response impedance predicted by the model. The theoretical impedance calculated for the equivalent circuit model. The equivalent circuit model typically includes: (Solution resistance) (charge transfer resistance) (Electrical double layer capacitor) Components such as (Warburg diffusion impedance), Calculations based on circuit theory ensure that the predicted impedance spectrum conforms to the physical processes of charge transfer and ion diffusion.

[0102] Ion diffusion kinetics constraints are based on Fick's diffusion law, and the constraint terms are defined as follows: ,in For lithium ion concentration distribution, For time, The diffusion coefficient is... This is the Laplace operator. This constraint requires that the concentration change rate predicted by the model must conform to the diffusion equation, reflecting the physical laws governing ion transport in the electrode material.

[0103] For example, when predicting the capacity retention rate of a battery cell after 500 cycles, if the model output voltage value deviates from the calculated value of the Nernst equation by more than a reasonable range, or the predicted impedance spectrum does not conform to the equivalent circuit model, or the concentration distribution violates the diffusion law, the corresponding regularization term will increase the total loss function, forcing the network parameters to be adjusted to meet the physical laws.

[0104] This physical mechanism constraint effectively solves the physical inconsistency problem of pure data-driven deep learning models in small sample and extrapolation scenarios. By embedding the fundamental laws of electrochemistry into the learning process, it significantly improves the model's generalization ability and interpretability, making the prediction results not only accurate in mathematical statistics but also reasonable in physical essence, avoiding the abnormal predictions that may occur in traditional black-box models that violate electrochemical principles. The three-dimensional constraints (potential balance, impedance characteristics, and diffusion kinetics) jointly regulate the model behavior from different physical levels, forming a mutually verifying constraint system, which greatly enhances the robustness and reliability of the prediction. This enables the system to accurately capture the intrinsic physical correlation between physical morphological characteristics and long-term electrochemical performance, providing a scientific basis for quality risk identification that combines data-driven accuracy and physical mechanism credibility.

[0105] For example, identifying the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the predicted probability distribution of electrical performance specifically includes:

[0106] When the cumulative probability of exceeding the failure threshold in the electrical performance prediction probability distribution is greater than the preset risk tolerance, it is determined to have potential failure risk; the electrical performance prediction probability distribution includes the confidence interval of capacity retention rate and the probability density function of DC internal resistance growth.

[0107] The probability distribution for predicting electrical performance uses a probability density function to describe the uncertainty of cell performance parameters. The confidence interval for capacity retention reflects the statistical range of capacity decay after long-term cycling, while the probability density function for the increase in DC internal resistance characterizes the probability distribution of internal resistance degradation. The core of risk identification lies in the calculation of cumulative probability, mathematically expressed as: when... At that time, it was determined that there was a potential risk of failure. Among them, For performance parameters (such as capacity retention rate or DC internal resistance growth rate). The failure threshold (e.g., capacity retention of 80% or internal resistance growth rate of 25%). The cumulative probability of exceeding this threshold. A preset risk tolerance level is set (usually between 0.3 and 0.5, reflecting the company's acceptance of quality risks). The system obtains the cumulative distribution function through the integral probability density function, calculates the probability that the performance parameters will exceed the failure threshold, and triggers a risk warning when this probability exceeds the tolerance threshold.

[0108] For example, for a certain battery cell under test, the model predicts that its capacity retention rate after 500 cycles follows a normal distribution N(85%, 3%²), and the failure threshold is set at 80%. Then, the cumulative probability of exceeding this threshold is P(X < 80%) = Φ((80-85) / 3) ≈ 0.047 (Φ is the standard normal cumulative distribution function), which is far below the risk tolerance of 0.4, and is judged to be risk-free. For another battery cell, the predicted capacity retention rate distribution is N(82%, 5%²), then P(X < 80%) = Φ((80-82) / 5) ≈ 0.345, which is close to but does not exceed the tolerance of 0.4, and is judged to be low risk. If the predicted distribution is N(78%, 4%²), then P(X < 80%) = Φ((80-78) / 4) ≈ 0.691, which is far beyond the tolerance of 0.4, and is judged to be high risk. The growth rate of DC internal resistance is assessed using a similar method. The two indicators are mutually verified to form a complete risk identification system.

[0109] For example, based on the risk identification results, before the secondary battery cell under test enters the capacity testing process, the production control system outputs a sorting and interception command or performs adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process, specifically including:

[0110] Based on the predicted probability distribution of electrical performance, the risk level of the secondary battery cell is determined; if the cumulative probability of the failure threshold is greater than the first failure threshold, the secondary battery cell is at level one risk; if the cumulative probability of the failure threshold is greater than or equal to the second failure threshold and less than or equal to the first failure threshold, the secondary battery cell is at level two risk; if the cumulative probability of the failure threshold is less than the second failure threshold, the secondary battery cell has no risk level; the first failure threshold is greater than the second failure threshold.

[0111] If the secondary battery cell is at level one risk, a sorting and interception command will be output through the production control system.

[0112] If the secondary battery cell is at level two risk, the average offset of the winding alignment of the most recent cells is calculated in real time, and a position compensation signal for the winding machine correction controller or a tension adjustment signal for the tension controller is generated; the shrinkage trend of the weld nugget area or weld depth is monitored in real time, and a laser power adjustment signal or welding pulse width modulation signal for the laser welding machine is generated.

[0113] In adaptive drift compensation control, this step, based on closed-loop control theory and adaptive adjustment algorithms, optimizes process parameters in real time for secondary risk cells. For the winding alignment problem, the system calculates the average winding alignment offset Δ_alignment = (1 / N)Σ(X_i - X_target) of the nearest N cells in real time, where X_i is the measured alignment of the i-th cell, X_target is the target alignment, and N is the sliding window size. Based on this offset, a PID control algorithm generates a position compensation signal: Δ_position = K_p·Δ_alignment + K_i·∫Δ_alignment dt + K_d·d(Δ_alignment) / dt, where K_p, K_i, and K_d are the proportional, integral, and derivative coefficients, respectively. To address welding quality issues, the system monitors the changing trend of the weld nugget area through time series analysis and calculates the shrinkage rate using first-order difference: rate_shrink = (A_t - A_{t-1}) / Δt, where A_t is the current weld nugget area, A_{t-1} is the area at the previous moment, and Δt is the time interval. When the shrinkage rate exceeds a threshold, a laser power adjustment signal is generated: Δ_power = α·|rate_shrink|, where α is the adjustment coefficient. These compensation signals are transmitted in real time to the controller of the corresponding device via industrial communication protocols (such as EtherCAT and Modbus).

[0114] This adaptive compensation mechanism achieves a fundamental shift in quality control from "post-event interception" to "pre-event prevention." By monitoring the drift trend of process parameters in real time, it automatically generates precise compensation signals, enabling manufacturing equipment to have self-adjustment and continuous optimization capabilities. This not only effectively eliminates the current quality risks of battery cells but also prevents batch quality problems of subsequent battery cells. It significantly reduces the frequency of manual intervention and process debugging costs, and significantly improves the intelligence level of the production line and the consistency of product quality, providing strong technical support for the high-quality development of battery manufacturing.

[0115] Compared to existing technologies, the above embodiments employ a time-axis synchronous triggering mechanism at the data acquisition level. By installing a high-precision encoder on the main shaft of the production line to monitor the cell position in real time, combined with FPGA hardware control, microsecond-level synchronous triggering is achieved. This ensures that the imaging data from the two key workstations—winding alignment and tab soldering—can accurately correspond to the same individual cell. At the feature extraction level, statistical criteria for critical morphological features are defined: morphological features whose boundary value difference from a preset acceptable range is less than or equal to a preset boundary threshold, and whose physical quantity value falls within a preset range of normal distribution, are considered potential risk indicators. At the predictive modeling level, a deep learning architecture constrained by physical mechanisms is constructed. Basic electrochemical principles (such as the Nernst equation and equivalent circuit models) are used as constraints for network training, enabling the model to not only fit historical data distributions but also conform to electrochemical physical laws, outputting a probability distribution of electrical performance predictions rather than a single, definitive value. At the control execution level, a closed-loop feedback mechanism is adopted to dynamically generate sorting instructions or process compensation signals based on the risk level, achieving real-time quality control.

[0116] In summary, the above embodiments completely solve the problem of accurate correlation of multi-source heterogeneous data in high-speed production environments through hardware-level synchronization mechanisms, providing a spatiotemporally consistent data foundation for quality analysis; the statistical definition of critical morphological features breaks through the limitation of traditional detection focusing only on obvious defects, and can accurately identify potentially failed cells that are "qualified but high-risk"; the deep learning model constrained by physical mechanisms significantly improves the reliability and interpretability of predictions, realizing cross-scale mapping from physical morphological features to long-term electrochemical performance; the adaptive drift compensation mechanism transforms quality control from passive interception to active prevention: after obtaining the pre-detection results, real-time process parameter optimization eliminates potential quality hazards from the source, greatly reducing the waste of resources and the risk of product failure in subsequent expensive processes.

[0117] See Figure 2 An embodiment of the present invention provides a defect detection device for a secondary battery, comprising: an imaging module 20, a sequence module 21, an extraction module 22, an input module 23, a detection module 24, and an adjustment module 25.

[0118] Imaging module 20 is used to acquire first imaging detection data of the secondary battery cell under test at the winding alignment station and second imaging detection data at the tab solder joint detection station on a high-speed production line through a time axis synchronous triggering mechanism; the high-speed production line refers to a production line with a production rate greater than a preset threshold.

[0119] The sequence module 21 is used to construct a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data and the unique digital identifier of the secondary battery cell under test.

[0120] Extraction module 22 is used to extract critical morphological features characterizing cell performance fluctuations from the multimodal physical morphological feature sequence; the critical morphological features refer to morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution.

[0121] Input module 23 is used to input the critical morphological features into a pre-trained performance correlation model; the performance correlation model is constructed based on the nonlinear coupling relationship between the morphological features in historical production data and the true values ​​of electrical performance after a preset cycle life stage, and the performance correlation model includes a deep learning architecture with physical mechanism constraints.

[0122] The detection module 24 is used to calculate the predicted probability distribution of the electrical performance of the secondary battery cell under test according to the performance correlation model, and to identify the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the probability distribution.

[0123] The adjustment module 25 is used to output sorting and interception commands through the production control system or to perform adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process before the secondary battery cell under test enters the capacity grading process, based on the risk identification results.

[0124] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the defect detection device for secondary batteries described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0125] Compared to existing technologies, this invention employs a time-axis synchronous triggering mechanism at the data acquisition level. A high-precision encoder installed on the main shaft of the production line monitors the cell position in real time, and combined with FPGA hardware control, achieves microsecond-level synchronous triggering. This ensures that the imaging data from the two key workstations—winding alignment and tab soldering—can accurately correspond to the same individual cell. At the feature extraction level, statistical criteria for critical morphological features are defined: morphological features whose boundary value difference from a preset acceptable range is less than or equal to a preset boundary threshold, and whose physical quantity value falls within a preset range of normal distribution, are considered potential risk indicators. At the predictive modeling level, a deep learning architecture constrained by physical mechanisms is constructed. Basic electrochemical principles (such as the Nernst equation and equivalent circuit models) are used as constraints for network training, enabling the model to not only fit historical data distributions but also conform to electrochemical physical laws, outputting a probability distribution of electrical performance predictions rather than a single, definitive value. At the control execution level, a closed-loop feedback mechanism is adopted to dynamically generate sorting instructions or process compensation signals based on the risk level, achieving real-time quality control.

[0126] In summary, the embodiments of this invention completely solve the problem of accurate correlation of multi-source heterogeneous data in high-speed production environments through hardware-level synchronization mechanisms, providing a spatiotemporally consistent data foundation for quality analysis; the statistical definition of critical morphological features breaks through the limitation of traditional detection focusing only on obvious defects, and can accurately identify potentially failed cells that are "qualified but high-risk"; the deep learning model constrained by physical mechanisms significantly improves the reliability and interpretability of predictions, realizing cross-scale mapping from physical morphological features to long-term electrochemical performance; the adaptive drift compensation mechanism transforms quality control from passive interception to active prevention: after obtaining the pre-detection results, real-time process parameter optimization eliminates potential quality hazards from the source, greatly reducing resource waste and product failure risks in subsequent expensive processes.

[0127] An embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a defect detection method for a secondary battery as described above.

[0128] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a defect detection method for a secondary battery as described above.

[0129] The computer device may be a smartphone, tablet, desktop computer, or cloud server, among other computing devices. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the figures are merely examples of computer devices and do not constitute a limitation on the computer device. It may include more or fewer components than illustrated, or a combination of certain components, or different components, such as input / output devices, network access devices, etc.

[0130] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0131] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0132] This invention provides a computer program product that, when run on a computer device, enables the computer device to execute the steps described in the various method embodiments above.

[0133] In the several embodiments provided by this invention, it will be understood that each block in the flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the figures. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved.

[0134] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A defect detection method for secondary batteries, characterized in that, include: The time-axis synchronous triggering mechanism is used to acquire the first imaging detection data of the secondary battery cell under test at the winding alignment station and the second imaging detection data at the tab solder joint detection station on the high-speed production line; the high-speed production line refers to a production line with a production rate greater than a preset threshold. Based on the first imaging detection data, the second imaging detection data and the unique digital identifier of the secondary battery cell under test, a multimodal physical morphological feature sequence of the secondary battery cell under test is constructed. The critical morphological features characterizing the performance fluctuation of the battery cell are extracted from the multimodal physical morphological feature sequence. The critical morphological features refer to the morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution. The critical morphological features are input into a pre-trained performance correlation model; The performance correlation model is constructed based on the nonlinear coupling relationship between the morphological features in historical production data and the true values ​​of electrical performance after a preset cycle life stage. The performance correlation model includes a deep learning architecture constrained by physical mechanisms. The electrical performance prediction probability distribution of the secondary battery cell under test is calculated based on the performance correlation model, and the electrochemical failure risk of the secondary battery cell under test after the preset cycle life stage is identified based on the probability distribution. Based on the risk identification results, before the secondary battery cell under test enters the capacity grading process, the production control system outputs a sorting and interception command or performs adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process.

2. The defect detection method for a secondary battery as described in claim 1, characterized in that, The acquisition of first imaging detection data at the winding alignment station and second imaging detection data at the tab solder joint detection station on the high-speed production line via a time-axis synchronization triggering mechanism specifically includes: An incremental encoder is installed on the main conveyor belt drive shaft of a high-speed production line. The incremental encoder pulses are received through an FPGA-based real-time motion control card; the real-time motion control card includes a first trigger port and a second trigger port, the first trigger port being connected to an X-ray detector at the winding alignment station, and the second trigger port being connected to an industrial camera at the welding station. When the feed sensor detects that the secondary battery cell to be tested has entered the conveyor belt, a unique digital identifier for the secondary battery cell to be tested is generated and written into the first address of the register; The logical coordinates of the secondary battery cell under test in the register are updated in real time according to the increase of the incremental encoder pulse. When the logical coordinates match the preset shooting point of the X-ray detector, the real-time motion control card outputs a first hardware trigger signal through the first trigger port; it collects the fluoroscopic image data of the X-ray detector and adds the current timestamp and the unique digital identifier, and uses the fluoroscopic image data as the first imaging detection data; When the logical coordinates match the preset shooting point of the industrial camera, the real-time motion control card outputs a second hardware trigger signal through the second trigger port; it collects surface morphology image data from the industrial camera and adds the current timestamp and the unique digital identifier, using the surface morphology image data as the second imaging detection data.

3. The defect detection method for a secondary battery as described in claim 1, characterized in that, The step of constructing a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data, and the unique digital identifier of the secondary battery cell under test specifically includes: Using the edge of the secondary battery cell casing as a reference point, calculate the rotation and translation matrices, and calibrate the first imaging detection data and the second imaging detection data to the standard coordinate system; The first imaging detection data and the second imaging detection data are subjected to ROI cropping and normalization processing; The first imaging detection data and the second imaging detection data are stitched together in the channel dimension to form a multi-channel tensor; After passing through the fully connected layer, the winding tension value and welding power value associated with the unique digital identifier are encoded into vectors and concatenated with the multi-channel tensor to obtain the multimodal physical morphological feature sequence of the secondary battery cell under test.

4. The defect detection method for a secondary battery as described in claim 3, characterized in that, The ROI cropping and normalization processing of the first imaging detection data and the second imaging detection data specifically includes: Scan along the grayscale gradient direction of the first imaging detection data to locate the boundary line between the positive and negative electrode sheets. Using the boundary line as the central axis, expand the preset pixel width to both sides to crop out a first rectangular image containing only the Overhang region. Using template matching, the geometric center of the weld stamp in the second imaging detection data is located, and a second rectangular image covering the entire heat-affected zone is cropped out with the geometric center of the weld stamp as the center of the rectangle. A bilinear interpolation algorithm is used to scale the first rectangular image and the second rectangular image to the same size, resulting in the processed first imaging detection data and the processed second imaging detection data.

5. The defect detection method for a secondary battery as described in claim 1, characterized in that, The deep learning architecture constrained by the physical mechanism specifically refers to: An electrochemical mechanism regularization term is introduced into the loss function of a deep learning network; the electrochemical mechanism regularization term includes potential constraints based on the Nernst equation, impedance constraints based on the equivalent circuit model, and concentration distribution constraints based on ion diffusion kinetics.

6. The defect detection method for a secondary battery as described in claim 1, characterized in that, The process of identifying the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the predicted probability distribution of electrical performance specifically includes: When the cumulative probability of exceeding the failure threshold in the electrical performance prediction probability distribution is greater than the preset risk tolerance, it is determined to have potential failure risk; the electrical performance prediction probability distribution includes the confidence interval of capacity retention rate and the probability density function of DC internal resistance growth.

7. The defect detection method for a secondary battery as described in claim 1, characterized in that, Based on the risk identification results, before the secondary battery cells under test enter the capacity testing process, the production control system outputs sorting and interception commands or performs adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process. Specifically, this includes: Based on the predicted probability distribution of electrical performance, the risk level of the secondary battery cell is determined; if the cumulative probability of the failure threshold is greater than the first failure threshold, the secondary battery cell is at level one risk; if the cumulative probability of the failure threshold is greater than or equal to the second failure threshold and less than or equal to the first failure threshold, the secondary battery cell is at level two risk; if the cumulative probability of the failure threshold is less than the second failure threshold, the secondary battery cell has no risk level; the first failure threshold is greater than the second failure threshold. If the secondary battery cell is at level one risk, a sorting and interception command will be output through the production control system. If the secondary battery cell is at level two risk, the average offset of the winding alignment of the most recent cells is calculated in real time, and a position compensation signal for the winding machine correction controller or a tension adjustment signal for the tension controller is generated; the shrinkage trend of the weld nugget area or weld depth is monitored in real time, and a laser power adjustment signal or welding pulse width modulation signal for the laser welding machine is generated.

8. A defect detection device for secondary batteries, characterized in that, include: The imaging module is used to acquire the first imaging detection data of the secondary battery cell under test at the winding alignment station and the second imaging detection data at the tab solder joint detection station on the high-speed production line through a time axis synchronous triggering mechanism; the high-speed production line refers to a production line with a production rate greater than a preset threshold. The sequence module is used to construct a multimodal physical morphological feature sequence of the secondary battery cell under test based on the first imaging detection data, the second imaging detection data and the unique digital identifier of the secondary battery cell under test. The extraction module is used to extract critical morphological features characterizing cell performance fluctuations from the multimodal physical morphological feature sequence; the critical morphological features refer to morphological features where the difference between the physical quantity value and the boundary value of the preset qualified interval is less than or equal to the preset boundary threshold, and the physical quantity value is within the preset interval of normal distribution. The input module is used to input the critical morphological features into a pre-trained performance correlation model; The performance correlation model is constructed based on the nonlinear coupling relationship between the morphological features in historical production data and the true values ​​of electrical performance after a preset cycle life stage. The performance correlation model includes a deep learning architecture constrained by physical mechanisms. The detection module is used to calculate the predicted probability distribution of the electrical performance of the secondary battery cell under test according to the performance correlation model, and to identify the electrochemical failure risk of the secondary battery cell under test after a preset cycle life stage based on the probability distribution. The adjustment module is used to, based on the risk identification results, output sorting and interception commands through the production control system or to perform adaptive drift compensation on the execution parameters of the corresponding equipment in the front-end manufacturing process before the secondary battery cell under test enters the capacity grading process.

9. A terminal device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a defect detection method for a secondary battery as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a defect detection method for a secondary battery as described in any one of claims 1 to 7.

Images