AI visual inspection method for battery manufacturing defects and self-adaptive correction system

Abstract

The invention relates to the technical field of battery detection, in particular to an AI visual detection method for battery manufacturing defects and a self-adaptive correction system.The AI visual detection method comprises the steps that multi-modal data of a to-be-detected battery is obtained, a defect correlation map is output according to the multi-modal data, and electrolyte and a coating of the to-be-detected battery are correlated to the defect correlation map; extracting material gene characteristics in the defect correlation map through an AI visual inspection model, and carrying out transfer calculation on the material gene characteristics to obtain root cause classification and defect position coordinates; according to the root cause classification and the defect position coordinates, executing equipment parameter adjustment operation, and obtaining a battery manufacturing defect detection result after equipment parameter adjustment; and inputting a battery manufacturing defect detection result into the AI visual detection model for optimization to obtain an optimized AI visual detection model so as to realize improvement of visual detection precision and response speed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery detection, in particular to an AI visual detection method and self-adaptive correction system for battery manufacturing defects. BACKGROUND

[0002] With the rapid development of electric vehicles, portable electronics and other fields, the safety and reliability of key components such as lithium ion batteries are increasingly important, because manufacturing defects such as tab offset, uneven coating and other problems may cause short circuits, thermal runaway and other serious consequences, threatening personal safety and causing economic losses.

[0003] Currently, the industry mainly relies on manual visual inspection or visual detection schemes based on conventional machine vision technology, but such schemes have the disadvantages of low detection efficiency, high false detection rate, and lack of real-time feedback capability on high-speed production lines, making it difficult to identify and handle defects in a timely manner.

[0004] Therefore, how to improve the accuracy and response speed of battery visual detection is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] In order to improve the accuracy and response speed of battery visual detection, the present application provides an AI visual detection method and self-adaptive correction system for battery manufacturing defects.

[0006] The AI visual detection method for battery manufacturing defects provided by the present application adopts the following technical scheme: An AI visual detection method for battery manufacturing defects, comprising: Obtaining multi-modal data of a battery to be detected, and outputting a defect correlation graph according to the multi-modal data, wherein the defect correlation graph is associated with the electrolyte and the coating of the battery to be detected; Extracting material gene features in the defect correlation graph through an AI visual detection model, performing transfer calculation on the material gene features, and obtaining root cause classification and defect position coordinates; Performing device parameter adjustment operation according to the root cause classification and the defect position coordinates, and obtaining battery manufacturing defect detection results after device parameter adjustment; Inputting the battery manufacturing defect detection results into the AI visual detection model for optimization, and obtaining an optimized AI visual detection model.

[0007] Further, the step of obtaining multi-modal data of a battery to be detected comprises: The tunable pulse is emitted to the battery to be detected, the tunable pulse penetrates the electrolyte layer of the battery to be detected through the battery surface of the battery to be detected, a molecular vibration attenuation signal in the electrolyte layer is captured through the quantum dot array, a micro-bubble distribution map is obtained, a battery surface diffraction pattern of the battery to be detected is acquired, an optical diffraction image is generated, a surface temperature of the battery to be detected is sampled, and a thermal field distribution video stream is generated according to the distribution of the surface temperature. The multi-modal data includes the micro-bubble distribution map, the optical diffraction image and the thermal field distribution video stream.

[0008] Further, the step of outputting the defect correlation atlas according to the multi-modal data includes: The multi-modal data is spatially registered and time-synchronously synchronized to obtain a multi-modal data set; The multi-modal data set is feature-extracted and reconstructed to obtain the defect correlation atlas.

[0009] Further, the step of spatially registering and time-synchronously synchronizing the multi-modal data to obtain the multi-modal data set includes: A unified spatial coordinate system is established based on a checkerboard calibration method; The micro-bubble distribution map, the optical diffraction image and the thermal field distribution video stream are aligned to the unified spatial coordinate system through an affine transformation matrix to obtain a registration matrix; After triggering time-synchronous synchronization of the micro-bubble distribution map, the optical diffraction image and the thermal field distribution video stream, the multi-modal data set is generated in combination with the registration matrix, wherein the multi-modal data set includes micro-bubble density distribution data, coating thickness deviation data and an electrolyte immersion state binary mask.

[0010] Further, the step of feature-extracting and reconstructing the multi-modal data set to obtain the defect correlation atlas includes: The micro-bubble density distribution data and the coating thickness deviation data are input into a fusion analysis module to identify a coupling relationship; In combination with a low-temperature non-immersed area marked by the electrolyte immersion state binary mask, the defect correlation atlas is obtained.

[0011] Further, the step of extracting a material gene feature in the defect correlation atlas and performing transfer calculation on the material gene feature to obtain a root cause classification and a defect position coordinate includes: After extracting a bond energy attenuation rate, a cracking rate and a thermal stress coefficient based on the defect correlation atlas, a local bonding failure probability is obtained according to the bond energy attenuation rate, an electrolyte decomposition risk value is obtained according to the cracking rate, and a current collector fatigue accumulation index is obtained according to the thermal stress coefficient; determining a relationship between the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index and a preset root cause judgment condition to obtain a root cause classification; and determining a defect position coordinate according to a spatial position mapped by the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index.

[0012] Further, the step of performing a device parameter adjustment operation according to the root cause classification and the defect position coordinate includes: determining a device to be adjusted and an adjustment amount according to the root cause classification and the defect position coordinate; performing a simulation adjustment operation based on the device to be adjusted and the adjustment amount to obtain a defect prediction recurrence rate, and determining whether the defect prediction recurrence rate is less than a preset defect prediction recurrence rate; if the defect prediction recurrence rate is less than the preset defect prediction recurrence rate, performing parameter adjustment on the device to be adjusted based on the adjustment amount; if the defect prediction recurrence rate is greater than the preset defect prediction recurrence rate, performing the step of determining a device to be adjusted and an adjustment amount according to the root cause classification and the defect position coordinate.

[0013] Further, the step of obtaining a battery manufacturing defect detection result after the device parameter adjustment includes: obtaining a battery defect occurrence rate after the device parameter adjustment, and determining whether the battery defect occurrence rate is less than a preset battery defect occurrence rate; if the battery defect occurrence rate is less than the preset battery defect occurrence rate, generating a battery manufacturing defect detection result according to the battery defect occurrence rate.

[0014] Further, the step of inputting the battery manufacturing defect detection result into an AI visual detection model for optimization to obtain an optimized AI visual detection model includes: locally training the battery manufacturing defect detection result to generate an encrypted gradient tensor, and performing blockchain notarization and federal aggregation on the encrypted gradient tensor to obtain an aggregated global gradient; optimizing the AI visual detection model through the aggregated global gradient to obtain an optimized AI visual detection model.

[0015] The application also provides a self-adaptive correction system applied to the AI visual detection method for battery manufacturing defects as described above, and the self-adaptive correction system includes: a data acquisition module configured to acquire multi-modal data of a battery to be detected, and output a defect correlation graph according to the multi-modal data, wherein the defect correlation graph is associated with electrolyte and coating of the battery to be detected; a data calculation module configured to extract material gene features in the defect correlation graph through an AI visual detection model, and perform transfer calculation on the material gene features to obtain a root cause classification and a defect position coordinate. a data detection module configured to perform a device parameter adjustment operation according to the root cause classification and the defect position coordinates, and obtain a battery manufacturing defect detection result after the device parameter adjustment; a model optimization module configured to input the battery manufacturing defect detection result into an AI visual detection model for optimization, and obtain an optimized AI visual detection model.

[0016] The beneficial effects achieved are: The present application provides an AI visual detection method for battery manufacturing defects. Specifically, multi-modal data of a battery to be detected is obtained, and a defect correlation graph is output according to the multi-modal data, wherein the defect correlation graph is associated with electrolyte and coating of the battery to be detected. Material gene features in the defect correlation graph are extracted by an AI visual detection model, and the material gene features are calculated by transmission to obtain root cause classification and defect position coordinates. A device parameter adjustment operation is performed according to the root cause classification and the defect position coordinates, and a battery manufacturing defect detection result after the device parameter adjustment is obtained. The battery manufacturing defect detection result is input into the AI visual detection model for optimization, and an optimized AI visual detection model is obtained.

[0017] That is, in the present application, by integrating multi-modal data acquisition, defect correlation graph analysis, material gene feature root cause tracing and equipment closed-loop control process, the visual detection accuracy and response speed are synergistically improved: at the data acquisition level, multi-modal data is simultaneously acquired, a defect correlation graph of electrolyte and coating is constructed by an AI visual detection model, the limitations of conventional single sensor are broken through, the defect morphology and material interaction are completely covered, and the completeness of abnormal feature recognition and the classification accuracy are significantly improved; at the defect diagnosis level, the defect source is accurately traced by graph network transmission calculation of material gene features, and only the appearance error is avoided, so that the root cause classification accuracy is improved; at the response control level, based on the real-time output of the defect position and the root cause classification, the device performs parameter adjustment, the conventional manual intervention is compressed to instantaneous closed loop of the production line, the self-adaptive ability of the AI visual detection model is dynamically optimized by each detection result, the false detection rate under new material process is continuously reduced and the calculation efficiency is improved, and finally the visual detection accuracy and speed are effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of an AI visual detection method for battery manufacturing defects according to the present application; Figure 2 is a flowchart of generating a defect correlation graph based on multi-modal data acquisition according to the present application; Figure 3 is a flowchart of defect correlation graph analysis according to the present application; Figure 4 A flowchart of a process for material gene eigenroot cause classification and equipment closed-loop control according to the present application; Figure 5 A module diagram of an adaptive correction system according to the present application. DETAILED DESCRIPTION

[0019] The following will be described in conjunction with the accompanying drawings Figure 1-5 The present application will be further described in detail.

[0020] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0021] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication between the two elements inside. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0022] The embodiments of the present application disclose an AI visual detection method for battery manufacturing defects.

[0023] Please refer to Figure 1 In an embodiment of the present application, an AI visual detection method for battery manufacturing defects includes steps S10-S40: Step S10, acquiring multi-modal data of the battery to be detected, and outputting a defect correlation graph according to the multi-modal data, wherein the defect correlation graph is associated with the electrolyte and the coating of the battery to be detected.

[0024] In the battery manufacturing process, strict defect detection must be carried out to ensure the safety and consistency of the battery. The current industry generally adopts manual visual inspection or conventional machine vision scheme. Manual visual inspection relies on the experience of operators to detect each piece, which is extremely low in efficiency and cannot detect invisible defects such as electrolyte micro-bubbles and internal coating micro-cracks. However, this detection method has a serious risk of missing detection. Although the conventional machine vision improves the detection speed, it can only collect surface visible light images and cannot penetrate the internal structure of the battery to perceive the physical relationship between the electrolyte flow state and the micro-morphology of the coating, resulting in a high false detection rate.

[0025] To solve the above problems, the embodiment obtains multi-modal data of the battery to be detected, integrates spectrum imaging and thermal field sensing technology that can penetrate the electrolyte layer, reveals the internal relationship between the electrolyte state and the coating deformation from different physical dimensions, and constructs the spatial and temporal mapping of the electrolyte micro-bubble distribution, the coating stress concentration area and the abnormal heat conduction to provide a data basis for the subsequent deep analysis of the physical field coupling.

[0026] The output defect correlation map intuitively presents the cause-and-effect chain of "electrolyte anomaly→coating damage" by fusing the spatial distribution characteristics of the electrolyte and the coating double parameters, such as mapping the bubble density gradient to the thickness reduction rate of the coating at the corresponding position. The map directly drives the correction of the coating pressure or vacuum injection parameters, and upgrades the characterization data in the conventional scheme to diagnostic evidence that can guide equipment adjustment.

[0027] In step S20, the material gene features in the defect correlation map are extracted by an AI visual detection model, and the material gene features are calculated by transmission to obtain root cause classification and defect position coordinates.

[0028] It should be noted that the AI visual detection model in the embodiment is a visual detection system for battery manufacturing defects based on a deep neural network. By pre-training to learn the internal laws of a large number of historical battery defect data, the AI visual detection model can analyze the multi-physical field defect correlation map.

[0029] Extracting the material gene features in the defect correlation map is a key step of mapping the observed macro-physical anomalies such as coating thinning and bubble aggregation to the micro-material failure mechanism. The material gene features reveal the essential laws such as field material molecular bond energy decay and solvent cracking reaction rate through physical field data reconstruction, which breaks through the limitation of conventional detection that can only perceive surface phenomena.

[0030] The transfer calculation of the extracted material genetic features is to simulate the dynamic propagation path of defects in the material microstructure, to realize cross-scale causal inference with a graph neural network (GNN) as a carrier, for example, a transfer chain of bubble pulse pressure to coating molecular chain fracture probability, and finally to obtain accurate judgment of the essence of failure such as adhesive dissociation or electrolyte decomposition, and the defect position coordinates lock the core target of process intervention, which together provide accurate positioning and essential diagnosis basis for real-time correction of device parameters, so that the battery manufacturing is changed from passive interception of defects to active elimination of failure mechanism, effectively improving the intelligence and convenience of battery detection.

[0031] In step S30, according to the root cause classification and the defect position coordinates, a device parameter adjustment operation is performed to obtain a battery manufacturing defect detection result after adjustment of the device parameters.

[0032] In the detection process of the battery, the essence of performing the device parameter adjustment operation according to the root cause classification and the defect position coordinates lies in breaking through the limitation of only recognizing the defect appearance in the conventional detection. Although the conventional battery detection scheme can discover phenomena such as coating crack and coating bubble anomaly, it cannot trace the essential mechanism of process failure such as adhesive molecular chain fracture and electrolyte solvent cracking, resulting in blind and lagging device parameter adjustment. For example, manual parameter adjustment relies on experience and trial and error, which is time-consuming and still cannot prevent the recurrence of similar defects.

[0033] However, the present embodiment can lock the root cause classification and the accurate defect position coordinates to direct the correction of specific parameters of specific devices and eliminate the defect generation mechanism from the source.

[0034] Further, by obtaining the battery manufacturing defect detection result after adjustment of the device parameters, a closed-loop verification link is constructed. When the new battery to be detected after parameter correction passes through online detection, the real-time feedback of the defect distribution change can verify the effectiveness of the device parameter adjustment and provide incremental data for dynamic optimization of the AI vision detection model, so that the system has a continuous and active optimization capability.

[0035] Compared with the conventional battery detection technology, the present embodiment changes the battery manufacturing from passive defect interception to active process immunity, locates the root cause classification and the defect position coordinates through real-time diagnosis, and performs directional device parameter adjustment, which effectively improves the accuracy and convenience of device parameter adjustment to a certain extent and avoids the inaccuracy and lag of manual experience adjustment.

[0036] In step S40, the battery manufacturing defect detection result is input into the AI vision detection model for optimization to obtain an optimized AI vision detection model.

[0037] The battery manufacturing defect detection result is input into the AI visual detection model for optimization, in order to solve the inherent defects that the conventional detection scheme cannot adapt to the dynamic interference such as material change and equipment aging in battery manufacturing. The conventional detection scheme relies on the model constructed by fixed training set, and when facing the false detection rate increase existing in new defects or process fluctuation, frequent shutdown and readjustment are required, and the response delay time is too long.

[0038] The embodiment performs online optimization on the AI visual detection model by inputting the battery manufacturing defect detection result, so that the AI visual detection model has autonomous adaptation capability, and upgrades the conventional passive interception to active learning. The optimized AI visual detection model not only can identify historical defects with higher accuracy, but also can predictively capture negative risks induced by process fluctuation, so as to change the detection function from post-screening of defective products to pre-preventing defects. Therefore, compared with the conventional detection scheme, the production rate of defective products can be greatly reduced, and the situation of cost increase caused by too many defective products can be avoided.

[0039] That is, the embodiment injects the AI visual detection model with the ability of real-time updating and optimization, so that the battery production line can still maintain relatively stable detection accuracy in variable environments such as material iteration and equipment wear.

[0040] Specifically, step S40 includes steps S41-S42: Step S41, the battery manufacturing defect detection result is locally trained to generate an encrypted gradient tensor, and the encrypted gradient tensor is stored in a blockchain and aggregated in a federation to obtain an aggregated global gradient.

[0041] Step S42, the AI visual detection model is optimized by the aggregated global gradient to obtain an optimized AI visual detection model.

[0042] In the embodiment, when the edge computing node locally trains the battery manufacturing defect detection result, an adversarial training mechanism is used to strengthen the model's ability to identify low-confidence samples and new defects: the generator creates defect morphology disturbance samples simulating process fluctuations, such as topological variants of coating cracks, and the discriminator learns subtle feature differences through a hybrid architecture of a visual converter and a U-Net++. In the training process, the gradient matrix of the model parameters is calculated, and the encrypted gradient tensor with node digital signature and timestamp is generated by the AES-256 encryption algorithm.

[0043] The encrypted gradient tensor is tamper-proofed by the blockchain storage system—SHA-256 hash operation is performed to generate a unique data fingerprint and write it into Hyperledger, and at the same time, the federal aggregation center triggers the aggregation of multi-source gradients according to the historical accuracy rate of the nodes, and generates an aggregated gradient (i.e., the aggregated global gradient) that integrates global knowledge.

[0044] Based on the polymeric gradient, the AI visual inspection model is optimized by first decrypting the gradient data and updating the convolution kernel weight parameters, then expanding the special recognition channel for the newly added defect type, and finally reconstructing the confidence threshold adaptive algorithm, outputting an AI visual inspection model with continuous optimization capability.

[0045] The optimized AI visual inspection model not only improves the historical defect recognition accuracy, but also incorporates new abnormalities caused by process fluctuations into the detection category through incremental learning mechanism, forming an intelligent closed loop of detection capability and dynamic collaboration with manufacturing environment.

[0046] Referring to Figure 2 As shown in the figure, regarding the feasible implementation of generating a defect correlation graph based on multi-modal data acquisition, the steps S11-S13 are as follows: Step S11, emit a tunable pulse to the battery to be detected, the tunable pulse penetrates the electrolyte layer of the battery to be detected through the battery surface of the battery to be detected, captures the molecular vibration attenuation signal in the electrolyte layer through the quantum dot array, obtains the micro-bubble distribution graph; and, obtains the diffraction pattern of the battery surface of the battery to be detected, generates an optical diffraction image; and, samples the surface temperature of the battery to be detected, generates a thermal field distribution video stream according to the distribution of the surface temperature; wherein, the multi-modal data includes the micro-bubble distribution graph, the optical diffraction image and the thermal field distribution video stream.

[0047] In this embodiment, the THz tunable emitter emits a tunable pulse to the battery to be detected in the detection range, the tunable pulse acts on the electrolyte layer after penetrating the battery shell and the electrode layer, the specific frequency band excites the electrolyte solvent molecule resonance, the quantum dot array in the quantum dot detector captures the attenuation intensity and phase shift of the molecular vibration signal in real time, obtains the molecular vibration attenuation signal, and reconstructs the three-dimensional spatial distribution graph of the micro-bubbles in the electrolyte layer (i.e. micro-bubble distribution graph).

[0048] Synchronous triggering of coherent light after Fresnel zone plate modulation illuminates the battery coating surface, the micro-topography of the coating surface modulates the incident light wavefront, forming a diffraction wave field containing topography information; a high-speed industrial camera captures the interference fringe image of the diffraction wave field on the imaging plane with a microsecond exposure time.

[0049] After obtaining the original diffraction pattern based on the diffraction wave field and interference fringe image, the optical diffraction image is reconstructed by an iterative constrained projection algorithm. Specifically, first, the captured interference fringe intensity distribution is simulated based on the angular spectrum propagation theory to simulate the propagation of light waves from the coating surface to the imaging plane. In the iterative calculation, the initial phase estimation value of the coating surface is continuously corrected until the simulated diffraction pattern matches the actual captured fringe with a matching degree of more than 99%. After outputting the corresponding phase distribution, the phase distribution is directly converted into the physical thickness of the coating to generate the optical diffraction image.

[0050] At the same time, the infrared thermal imager scans the surface temperature field of the battery at a sampling frequency of 30 Hz, combines with the environmental temperature compensation algorithm to eliminate interference, dynamically marks the areas below the average temperature, for example, marks the continuous low-temperature area below 0.5°C, and generates a time-space continuous thermal field distribution video stream.

[0051] The multi-modal data obtained through the above steps can provide a coupling data basis across physical fields for subsequent defect correlation analysis.

[0052] In step S12, the multi-modal data is spatially registered and time-synchronized to obtain a multi-modal data set.

[0053] The core significance of spatial registration and time synchronization of multi-modal data is to overcome the fragmented nature of multi-source sensing data in battery manufacturing and testing. The micro-bubble distribution map, optical diffraction image, and thermal field distribution video stream are in a state of spatial misalignment and temporal disconnection due to physical location differences and sampling time drift, causing the same physical event to be displayed as a temperature anomaly on the thermal imager while the optical system may not have captured the associated coating deformation response. Even the bubble aggregation is incorrectly associated with the coating cracking phenomenon in non-adjacent areas.

[0054] The spatial registration function eliminates sensor bias through coordinate system mapping technology to ensure that the three-dimensional positioning points of the micro-bubble clusters in the electrolyte are accurately aligned with the spatial coordinates of the coating directly above them, thereby revealing the true coupling relationship between bubble hydraulic impact and coating stress distortion. Time synchronization eliminates the causal inversion caused by sampling delays of different sensing units, and reconstructs the strict logical chain of multi-physical field changes. The multi-modal data set after space-time calibration constructs an association evidence chain for the electrolyte dynamic behavior corresponding to the micro-bubble distribution map, the coating mechanical response corresponding to the optical diffraction image, and the thermodynamic state corresponding to the thermal field distribution video stream. For example, when the data reveals that the peak value of micro-bubble density appears 0.5 seconds earlier than the thickness mutation of the coating directly above it, and the thermal field gradient change conforms to the fluid-solid energy transfer model, the system can lock the complete failure path of "liquid injection turbulence induced bubble aggregation → hydraulic impact leading to coating fatigue → increased thermal resistance causing local low temperature".

[0055] The generation of defect correlation atlas must rely on such strictly aligned multi-modal data sets to upgrade discrete abnormal signals to diagnosable process root causes, avoid isolated analysis of phenomena by conventional detection, and convert defect attribution from probabilistic guess to inevitable regularity, so as to change battery manufacturing from passive sampling inspection to active closed-loop control.

[0056] In a feasible implementation, the step S12 includes steps S121-S123: In step S121, a unified spatial coordinate system is established based on a checkerboard calibration method.

[0057] In step S122, the micro-bubble distribution map, the optical diffraction image and the thermal field distribution video stream are aligned to the unified spatial coordinate system through an affine transformation matrix to obtain a registration matrix.

[0058] In step S123, after time sequence synchronization of the micro-bubble distribution map, the optical diffraction image and the thermal field distribution video stream is triggered, the registration matrix is generated in combination, and the multi-modal data set includes micro-bubble density distribution data, coating thickness deviation data and electrolyte infiltration state binary mask.

[0059] A high-precision checkerboard calibration plate is installed at a fixed position of a battery conveying belt. When a battery passes through a detection position, a terahertz detection system, an optical diffraction camera and an infrared thermal imager synchronously capture battery images with the checkerboard. A computer vision algorithm extracts checkerboard corner point coordinates in images of each system. Taking the bubble coordinate system of the terahertz detection system as a reference, a spatial mapping relationship between the thickness image of the optical diffraction camera and the checkerboard feature points in the infrared thermal imager is calculated respectively, and a unified spatial coordinate system is established.

[0060] When the micro-bubble distribution map is aligned, the above unified spatial coordinate system is directly used as a reference space, the pixel coordinates of the thickness image corresponding to the optical diffraction image are mapped to the bubble coordinate system, a first affine matrix is calculated, the thickness data is resampled and registered through bilinear interpolation, the pixel of the thermal field image corresponding to the thermal field distribution video stream is mapped to the bubble coordinate system to generate a second affine matrix to realize frame-by-frame coordinate conversion, and finally the registration matrix of the three groups of data under the unified spatial grid is obtained.

[0061] Time sequence synchronization is realized through hardware triggering. A PLC controller sends an edge trigger signal at the moment when a battery enters a detection area, a terahertz tunable emitter starts a tunable pulse sequence, and a quantum dot detector records a time stamp start value. The trigger signal simultaneously turns on a laser to emit coherent light and starts a high-speed industrial camera to expose, and an infrared thermal imager starts scanning after a 0.3ms delay after receiving the trigger signal. The same trigger sequence ID and time stamp are embedded in the head of all data streams.

[0062] Finally, the registration matrix is applied to the original data stream, for example, each three-dimensional coordinate point in the micro-bubble distribution map is projected onto the two-dimensional coating thickness plane according to the affine transformation, and is superimposed with the thermal field frame data corresponding to the time stamp, to generate a multi-modal data set. The micro-bubble density distribution data is generated by the mapping relationship between the bubble volume and the coordinates, the coating thickness deviation data is derived from the difference calculation between the registered diffraction image and the reference thickness model, and the electrolyte immersion state binary mask is converted into a 0-1 matrix according to the spatial distribution of the continuous low temperature area below the regional average temperature, such as below 0.5℃. The three constitute a multi-modal data set strictly aligned in space and time for subsequent defect correlation atlas analysis.

[0063] In step S13, feature extraction and reconstruction are performed on the multi-modal data set to obtain a defect correlation atlas.

[0064] In this embodiment, the core value of feature extraction and reconstruction on the multi-modal data set is to solve the fundamental defect of isolated physical fields in conventional battery detection. When micro-bubble density, coating thickness deviation and electrolyte immersion state are used as independent representation quantities, they can only reflect local abnormal phenomena, such as simple coating thinning or local poor immersion, and cannot reveal the coupling failure chain that bubble aggregation causes coating stress distortion and then blocks electrolyte immersion.

[0065] Feature extraction is to map the micro-bubble density gradient to the hydraulic impact force field feature, reconstruct the coating thickness mutation to the adhesive molecular chain stress feature, and convert the immersion state binary mask to the thermodynamic gradient diffusion feature. The effect is to upgrade the original fragmented data to quantitative evidence that exposes the cross-physical field causality, for example, to prove that the thickness of the coating above the high-bubble density area must be accompanied by surrounding immersion deterioration. The generated electrolyte-coating defect correlation atlas is the explicit expression of this coupling mechanism, which locates the overlapping area of bubble-coating- immersion triple abnormalities through spatial superimposed thermodynamic map, such as the coincidence of bubble density peak and maximum coating thinning position and being surrounded by low-temperature immersion. It provides physical logic irrefutable input for subsequent diagnosis, for example, when the atlas shows that a certain area simultaneously presents high hydraulic impact feature, molecular chain high stress feature and thermal diffusion retardation feature, the system directly determines it as "connected failure induced by inaccurate liquid injection parameters", thereby accurately triggering the correction of the target device instead of blindly adjusting the device parameters.

[0066] In one possible implementation, step S13 includes steps S131-S132: In step S131, the micro-bubble density distribution data and the coating thickness deviation data are input into a fusion analysis module to identify the coupling relationship.

[0067] In step S132, the low-temperature unimmersed area marked by the electrolyte immersion state binary mask is combined to obtain a defect correlation atlas.

[0068] In this embodiment, when the micro-bubble density distribution data and the coating thickness deviation data are input into the fusion analysis module, the system automatically scans the gradient similarity of the two kinds of data through the spatial convolutional neural network (SCNN, Spatial Convolutional Neural Network), and when it is detected that the spatial coordinate coincidence degree of the outline boundary of the high-density micro-bubble area reflected by the micro-bubble density distribution data and the coating thickness sudden drop area reflected by the coating thickness deviation data exceeds a set threshold, such as the coating thinning amount surging within 2mm radius of the bubble aggregation center, it is determined that the two have a hydraulic impact coupling relationship.

[0069] At this time, the system performs three-dimensional probability fusion on the triple data in combination with the binary mask marked continuous low-temperature un-infiltrated area of the electrolyte infiltration state, and if the spatial distribution of the low-temperature un-infiltrated area covers the bubble density peak area and the coating thickness sudden drop area at the same time, a red warning layer is superimposed on the multi-layer probability graph; if the low-temperature un-infiltrated area only overlaps with a single factor, a yellow risk layer is superimposed; isolated abnormal points are marked with a blue monitoring mark, and finally a defect correlation atlas presented in a pseudo-color thermal map is generated, which is essentially a time-space coupling relationship of the bubble dynamic pressure-coating mechanical stress-thermal resistance effect, which is converted into a quantitative visual model that can drive root cause classification and diagnosis.

[0070] Referring to Figure 3 As shown in the figure, the feasible implementation of the defect correlation atlas analysis includes steps S21-S22 as follows: After extracting the bond energy attenuation rate, the cracking rate and the thermal stress coefficient based on the defect correlation atlas, the local bonding failure probability is obtained according to the bond energy attenuation rate, the electrolyte decomposition risk value is obtained according to the cracking rate, and the current collector fatigue accumulation index is obtained according to the thermal stress coefficient.

[0071] Firstly, the area representing the coating thickness mutation in the defect correlation atlas is mapped to the adhesive molecular dynamics model, and the bond energy attenuation amplitude is calculated according to the functional relationship between the coating abnormal deformation and the material yield strength; for the extraction of the cracking rate, the bubble volume gradient is converted into the cracking rate through the model of the electrolyte micro-bubble density distribution and the solvent molecule cracking activation energy in the defect correlation atlas; the thermal stress coefficient is derived from the coupling calculation of the electrolyte un-infiltrated area in the defect correlation atlas and the current collector expansion coefficient, and the lattice distortion accumulation is deduced by using the thermal conduction lag characteristic.

[0072] Subsequently, the process of obtaining the local bonding failure probability according to the bond energy attenuation rate is as follows: calling the molecular chain fracture probability function model, inputting the bond energy attenuation rate into the Monte Carlo simulation system, and outputting the local bonding failure probability according to the random sampling of the bond energy potential well depth.

[0073] The calculation of the electrolyte decomposition risk value is through the integral reaction equation of the cracking rate, combined with the solvent concentration half-life model to predict the irreversible decomposition threshold.

[0074] The generation of the current collector fatigue accumulation index adopts a lattice slip displacement algorithm, and when a thermal stress coefficient breaks through a material recrystallization critical condition, a dislocation density accumulation function outputs a fatigue index.

[0075] The numerical value obtained by the above calculation quantifies the micro-mechanism of the material genetic characteristics as a decision factor that can guide process correction, and provides physical logic support for subsequent root cause classification and parameter adjustment.

[0076] In step S22, the relationship between the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index and the preset root cause judgment condition is judged to obtain a root cause classification, and the defect position coordinates are determined according to the spatial position mapped by the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index.

[0077] In the judgment of the root cause classification, the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index are synchronously input into a root cause classification decision tree model, and the physical logic rules of the material failure mechanism are built in the model. If the bonding failure probability breaks through the molecular chain rupture threshold, it is directly judged as a binder failure; if the electrolyte decomposition risk value continuously exceeds the solvent chemical bond cleavage critical line, it is classified as electrolyte decomposition; if the current collector fatigue accumulation index reaches the lattice slip criterion, it points to the current collector fatigue; if all of the above exceed the standard, the dominant root cause is output according to the failure priority. This layered judgment mechanism ensures that the micro-mechanism of the material gene is converted into a process layer operable diagnostic conclusion without damage. The preset root cause judgment condition is the molecular chain rupture threshold, the solvent chemical bond cleavage critical line, and the lattice slip criterion.

[0078] For the determination of the defect position coordinates, the system maps the bond energy decay rate, the cleavage rate, and the fatigue index calculated by each spatial node back to the original detection coordinate system, and identifies the abnormal value aggregation area through a probability density clustering algorithm: a continuous spatial area with a failure probability greater than 95% is extracted, and its geometric center is calculated as the core coordinates. When there are multiple failure mechanisms superimposed in the same area, a three-dimensional thermal field diffusion simulation is automatically triggered, and the influence range is determined according to the probability decay boundary.

[0079] The spatial level precise positioning is provided for the equipment parameter adjustment, and the problems such as overlarge parameter adjustment range and overcompensation caused by the fuzzy spatial positioning in the conventional scheme are solved.

[0080] Referring to Figure 4 Fig. 1, a feasible implementation mode of the material genetic characteristic root cause classification tracing and equipment closed-loop control process is shown as follows, including steps S31-S35: In step S31, the equipment to be adjusted and the adjustment amount are determined according to the root cause classification and the defect position coordinates.

[0081] In step S32, a simulation adjustment operation is performed based on the to-be-adjusted device and the adjustment amount, and a defect prediction recurrence rate is obtained. It is determined whether the defect prediction recurrence rate is less than a preset defect prediction recurrence rate.

[0082] In step S33, if the defect prediction recurrence rate is less than the preset defect prediction recurrence rate, a parameter adjustment is performed on the to-be-adjusted device based on the adjustment amount.

[0083] In this embodiment, when the to-be-adjusted device and the adjustment amount are determined according to the root cause classification and the defect position coordinates, the system calls a preset device mapping rule library. Assuming that the root cause classification is adhesive failure, the coating machine is locked as the to-be-adjusted device, and the blade pressure compensation amount is automatically generated according to the spatial weight distribution of the defect position coordinates, such as the distance of the coating thinning area from the center of the doctor blade. When the root cause classification is electrolyte decomposition, the liquid injection machine is selected as the to-be-adjusted device, and the adjustment amount is dynamically calculated based on the integral value of the bubble density gradient at the defect coordinates. When the root cause classification is current collector fatigue, the roll press parameter correction is triggered, and the adjustment amount is inversely deduced from the decay characteristics of the fatigue index in space.

[0084] Subsequently, when the simulation adjustment operation is performed based on the to-be-adjusted device and the adjustment amount, the digital twin system loads the new device parameters and the battery production line kinetics model, simulates the operation of 200 consecutive production cycles, and counts the defect recurrence frequency. The adhesive failure recurrence rate is iteratively calculated by a molecular bond rupture probability model, and the electrolyte decomposition risk is predicted by a solvent concentration diffusion equation. Finally, the defect prediction recurrence rate is output by weighted aggregation.

[0085] If the defect prediction recurrence rate is lower than the preset defect prediction recurrence rate, an adjustment instruction is issued to the corresponding to-be-adjusted device. If the defect prediction recurrence rate exceeds the preset defect prediction recurrence rate, the combination of the to-be-adjusted device and the adjustment amount is re-matched, and the iteration is repeated until the defect prediction recurrence rate converges to the standard, forming a mistake-proofing mechanism of process parameter self-verification and self-optimization.

[0086] It should be noted that the preset defect prediction recurrence rate is an upper limit value of the acceptable defect prediction recurrence rate.

[0087] In step S34, the battery defect occurrence rate after the device parameter adjustment is obtained, and it is determined whether the battery defect occurrence rate is less than a preset battery defect occurrence rate.

[0088] In step S35, if the battery defect occurrence rate is less than the preset battery defect occurrence rate, a battery manufacturing defect detection result is generated according to the battery defect occurrence rate.

[0089] It should be noted that the preset battery defect occurrence rate is an upper limit value of the acceptable battery defect occurrence rate.

[0090] When the battery defect occurrence rate after the adjustment of the equipment parameters is obtained, the system starts targeted monitoring of 50 battery batches in continuous production, locks the same spatial area based on the defect position coordinates, and performs multimodal re-inspection, such as re-scanning the electrolyte layer at the coordinate point by the bubble detector and re-measuring the coating thickness by the optical system. According to the detection results, the number of recurrence of the same defects is counted, and the battery defect occurrence rate is calculated in combination with the total number of batch samples. The battery defect occurrence rate is compared with the preset battery defect occurrence rate in real time. If the battery defect occurrence rate is less than the preset battery defect occurrence rate, it is determined that the adjustment of the equipment parameters is effective, and the system automatically generates the battery manufacturing defect detection result including the defect distribution elimination rate. If the battery defect occurrence rate is greater than the preset battery defect occurrence rate, the iterative optimization of the root cause classification diagnosis-equipment parameter adjustment is triggered in reverse. According to the new defect characteristics captured in the re-inspection, such as the appearance of a new crack morphology in the coating, the root cause classification weight is updated, and the equipment to be adjusted and the adjustment amount are corrected according to the coordinate offset of the recurrence area, until the battery defect occurrence rate converges to the safe interval after adjustment, thereby fundamentally avoiding the blindness of conventional empirical parameter adjustment and effectively improving the battery detection accuracy and response speed.

[0091] The application also provides an adaptive correction system, as shown in Figure 5 The adaptive correction system comprises: A data acquisition module 10 is configured to acquire multimodal data of a battery to be detected, and output a defect correlation graph based on the multimodal data, wherein the defect correlation graph is associated with an electrolyte and a coating of the battery to be detected. A data calculation module 20 is configured to extract material gene features in the defect correlation graph by an AI visual detection model, and perform transfer calculation on the material gene features to obtain root cause classification and defect position coordinates. A data detection module 30 is configured to perform an equipment parameter adjustment operation according to the root cause classification and the defect position coordinates, and obtain a battery manufacturing defect detection result after the adjustment of the equipment parameters. A model optimization module 40 is configured to input the battery manufacturing defect detection result into the AI visual detection model for optimization to obtain an optimized AI visual detection model.

[0092] Optionally, the data acquisition module 10 is further configured to: emit a tunable pulse to the battery to be detected, the tunable pulse penetrates the electrolyte layer of the battery to be detected through the surface of the battery to be detected, captures molecular vibration attenuation signals in the electrolyte layer through a quantum dot array to obtain a micro-bubble distribution graph, acquires a diffraction pattern of the surface of the battery to be detected to generate an optical diffraction image, and samples the surface temperature of the battery to be detected to generate a thermal field distribution video stream according to the distribution of the surface temperature. The multimodal data comprises the micro-bubble distribution graph, the optical diffraction image, and the thermal field distribution video stream.

[0093] Optionally, the data acquisition module 10 is further configured to: spatially register and time-synchronize the multi-modal data to obtain a multi-modal data set; extract features and reconstruct the multi-modal data set to obtain a defect correlation atlas.

[0094] Optionally, the data acquisition module 10 is further configured to: establish a unified spatial coordinate system based on a checkerboard calibration method; align the micro-bubble distribution map, the optical diffraction image, and the thermal field distribution video stream to the unified spatial coordinate system through an affine transformation matrix to obtain a registration matrix; trigger time-synchronization of the micro-bubble distribution map, the optical diffraction image, and the thermal field distribution video stream, and generate a multi-modal data set in combination with the registration matrix, wherein the multi-modal data set includes micro-bubble density distribution data, coating thickness deviation data, and an electrolyte immersion state binary mask.

[0095] Optionally, the data acquisition module 10 is further configured to: input the micro-bubble density distribution data and the coating thickness deviation data into a fusion analysis module to identify a coupling relationship; obtain a defect correlation atlas in combination with a low-temperature unimmersed area marked by the electrolyte immersion state binary mask.

[0096] Optionally, the data calculation module 20 is further configured to: extract a bond energy attenuation rate, a cracking rate, and a thermal stress coefficient based on the defect correlation atlas, obtain a local adhesion failure probability according to the bond energy attenuation rate, obtain an electrolyte decomposition risk value according to the cracking rate, and obtain a current collector fatigue accumulation index according to the thermal stress coefficient; determine a relationship between the local adhesion failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index and a preset root cause judgment condition to obtain a root cause classification, and determine a defect position coordinate according to a spatial position mapped by the local adhesion failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index.

[0097] Optionally, the data detection module 30 is further configured to: determine a device to be adjusted and an adjustment amount according to the root cause classification and the defect position coordinate; perform a simulation adjustment operation based on the device to be adjusted and the adjustment amount to obtain a defect prediction recurrence rate, and determine whether the defect prediction recurrence rate is less than a preset defect prediction recurrence rate; if the defect prediction recurrence rate is less than the preset defect prediction recurrence rate, perform parameter adjustment on the device to be adjusted based on the adjustment amount; if the defect prediction recurrence rate is greater than the preset defect prediction recurrence rate, perform the step of determining the device to be adjusted and the adjustment amount according to the root cause classification and the defect position coordinate.

[0098] Optionally, the data detection module 30 is further configured to: obtain the battery defect occurrence rate after the adjustment of the device parameters, and determine whether the battery defect occurrence rate is less than a preset battery defect occurrence rate; if the battery defect occurrence rate is less than the preset battery defect occurrence rate, generate a battery detection result according to the battery defect occurrence rate.

[0099] Optionally, the model optimization module 40 is further configured to: perform local training on the battery manufacturing defect detection result to generate an encrypted gradient tensor, perform blockchain notarization and federal aggregation on the encrypted gradient tensor, and obtain an aggregated global gradient; optimize the AI visual detection model through the aggregated global gradient to obtain an optimized AI visual detection model.

[0100] The above are preferred embodiments of the present application, which do not limit the protection scope of the present application, and therefore: any equivalent changes made on the basis of the structure, shape, and principle of the present application should be covered within the protection scope of the present application.

Claims

1. An AI visual inspection method for battery manufacturing defects, characterized in that: include: Acquire multimodal data of a battery to be inspected, and output a defect correlation map based on the multimodal data, wherein the defect correlation map is associated with an electrolyte and a coating of the battery to be inspected; Through the AI ​​visual inspection model, the material gene features in the defect association map are extracted, and the material gene features are transferred and calculated to obtain the root cause classification and defect location coordinates; Performing an equipment parameter adjustment operation based on the root cause classification and the defect location coordinates, and obtaining a battery manufacturing defect detection result after the equipment parameter adjustment; The battery manufacturing defect detection results are input into the AI ​​visual inspection model for optimization to obtain an optimized AI visual inspection model.

2. The AI ​​visual inspection method for battery manufacturing defects according to claim 1, characterized in that: The step of obtaining multimodal data of the battery to be tested includes: Transmitting a tunable pulse to the battery to be tested, wherein the tunable pulse penetrates the electrolyte layer of the battery to be tested through the battery surface of the battery to be tested, and capturing the molecular vibration attenuation signal in the electrolyte layer through the quantum dot array to obtain a microbubble distribution map; and Acquiring a diffraction pattern on the battery surface of the battery to be tested to generate an optical diffraction image; and Sampling the surface temperature of the battery to be tested, and generating a thermal field distribution video stream based on the distribution of the surface temperature; Wherein, the multimodal data includes the microbubble distribution map, the optical diffraction image and the thermal field distribution video stream.

3. The AI ​​visual inspection method for battery manufacturing defects according to claim 2, characterized in that: The step of outputting a defect association map according to the multimodal data includes: Performing spatial registration and temporal synchronization on the multimodal data to obtain a multimodal dataset; Feature extraction and reconstruction are performed on the multimodal data set to obtain a defect association map.

4. The AI ​​visual inspection method for battery manufacturing defects according to claim 3, characterized in that: The step of performing spatial registration and temporal synchronization on the multimodal data to obtain a multimodal data set includes: Establish a unified spatial coordinate system based on the chessboard calibration method; Aligning the microbubble distribution map, the optical diffraction image, and the thermal field distribution video stream to the unified spatial coordinate system through an affine transformation matrix to obtain a registration matrix; After triggering the time synchronization of the microbubble distribution map, the optical diffraction image, and the thermal field distribution video stream, the multimodal dataset is generated in combination with the registration matrix, wherein the multimodal dataset includes microbubble density distribution data, coating thickness deviation data, and two masks of electrolyte infiltration state.

5. The AI ​​visual inspection method for battery manufacturing defects according to claim 4, characterized in that: The step of extracting and reconstructing features from the multimodal dataset to obtain a defect association map includes: Inputting the microbubble density distribution data and the coating thickness deviation data into a fusion analysis module to identify a coupling relationship; The defect correlation map is obtained by combining the low-temperature non-wetted areas marked by the second mask of the electrolyte wetting state.

6. The AI ​​visual inspection method for battery manufacturing defects according to claim 5, characterized in that: The step of extracting material gene features from the defect association map, performing transfer calculation on the material gene features, and obtaining root cause classification and defect location coordinates includes: After extracting the bond energy decay rate, cracking rate, and thermal stress coefficient based on the defect association map, the local bonding failure probability is obtained according to the bond energy decay rate, the electrolyte decomposition risk value is obtained according to the cracking rate, and the current collector fatigue accumulation index is obtained according to the thermal stress coefficient; Determining the relationship between the local bonding failure probability, the electrolyte decomposition risk value, the current collector fatigue accumulation index, and preset root cause judgment conditions to obtain the root cause classification; and The defect position coordinates are determined according to the spatial position mapped by the local bonding failure probability, the electrolyte decomposition risk value, and the current collector fatigue accumulation index.

7. The AI ​​visual inspection method for battery manufacturing defects according to claim 6, characterized in that: The step of performing equipment parameter adjustment operations based on the root cause classification and the defect location coordinates includes: Determining the equipment to be adjusted and the adjustment amount according to the root cause classification and the defect location coordinates; Performing a simulation adjustment operation based on the device to be adjusted and the adjustment amount to obtain a defect prediction recurrence rate, and determining whether the defect prediction recurrence rate is less than a preset defect prediction recurrence rate; If the defect predicted recurrence rate is less than the preset defect predicted recurrence rate, adjusting parameters of the device to be adjusted based on the adjustment amount; If the defect predicted recurrence rate is greater than the preset defect predicted recurrence rate, the step of determining the equipment to be adjusted and the adjustment amount based on the root cause classification and the defect location coordinates is performed.

8. The AI ​​visual inspection method for battery manufacturing defects according to claim 7, characterized in that: The step of obtaining the battery test result after the device parameters are adjusted includes: Obtaining a battery defect rate after the device parameters are adjusted, and determining whether the battery defect rate is less than a preset battery defect rate; If the battery defect occurrence rate is less than the preset battery defect occurrence rate, the battery detection result is generated according to the battery defect occurrence rate.

9. The AI ​​visual inspection method for battery manufacturing defects according to claim 8, characterized in that: The step of inputting the battery manufacturing defect detection result into the AI ​​visual inspection model for optimization to obtain the optimized AI visual inspection model includes: Performing local training on the battery manufacturing defect detection results to generate an encrypted gradient tensor, performing blockchain notarization and federated aggregation on the encrypted gradient tensor to obtain an aggregated global gradient; The AI ​​visual detection model is optimized by using the aggregated global gradient to obtain the optimized AI visual detection model.

10. An adaptive correction system, characterized in that: The adaptive correction system is applied to the AI ​​visual detection method for battery manufacturing defects according to any one of claims 1 to 9, comprising: a data acquisition module, configured to acquire multimodal data of the battery to be inspected, and output a defect association map based on the multimodal data, wherein the defect association map is associated with the electrolyte and coating of the battery to be inspected; A data calculation module is used to extract material gene features from the defect association map through an AI visual inspection model, perform transfer calculations on the material gene features, and obtain root cause classification and defect location coordinates; a data detection module, configured to perform an equipment parameter adjustment operation based on the root cause classification and the defect location coordinates, and obtain a battery manufacturing defect detection result after the equipment parameters are adjusted; A model optimization module is used to input the battery manufacturing defect detection results into the AI ​​visual inspection model for optimization to obtain an optimized AI visual inspection model.

Images