Lithium electrode cloth surface defect detection method and system based on charge transfer and latent image detection

Abstract

The application discloses a lithium electrode cloth surface defect detection method and system based on charge transfer and latent image detection, comprising the following steps: S1, generating fine gaseous charge probes on the surface of a lithium electrode to be detected through a controllable nitrogen corona device; S2, transferring and locking the charge distribution; S3, generating a charge latent image; and S4, double-mode physical guiding defect detection and defect quantization output. The application generates controllable fine gaseous charge probes on the surface of an electrode cloth, transfers the non-uniform charge distribution to an elastic insulating medium with charge traps, acquires charge spatial distribution data, analyzes the data in combination with a double-mode physical guiding defect detection model, realizes high-precision identification and physical quantization of defects under small sample conditions, and outputs defect categories, positions, outlines, areas and depth information. The method can not only accurately and quickly detect small surface defects, but also embeds electrostatic physical laws into a neural network structure, reduces learnable parameters by more than 95%, and realizes small sample training.

Description

Technical Field

[0001] This invention relates to the field of lithium electrode cloth surface inspection technology, and in particular to a method and system for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection. Background Technology

[0002] In the existing technology, the detection methods for surface defects of lithium electrode cloth usually rely on visual inspection, optical microscopy and other traditional surface inspection techniques. These methods often rely on manual or machine vision systems to capture and analyze images of surface defects, or to use contact measuring instruments to perform physical measurements on the surface. Although these detection methods are effective under certain conditions, they have a number of limitations. Visual inspection is affected by lighting conditions, surface reflectivity and surface texture, which leads to reduced detection accuracy. In particular, it is difficult to identify small or transparent surface defects. The detection speed of the existing technology is relatively slow, and it is difficult to efficiently cope with the rapidly changing production lines in actual production environments.

[0003] However, existing technologies for detecting surface defects in lithium electrode fabrics suffer from limitations. Traditional visual inspection techniques lack sufficient accuracy in identifying surface defects, especially when defect samples are scarce. They struggle to reliably identify and locate minute, intricate defects. Current surface inspection methods often rely on image capture or contact scanning, failing to effectively capture detailed defect information, resulting in low detection accuracy. Furthermore, visual inspection exhibits poor generalization ability in environments with limited training samples, restricting its ability to identify and locate defects of different types and shapes. These shortcomings limit the application of traditional methods in the efficient and accurate detection of surface defects in lithium electrode fabrics. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for detecting surface defects in lithium electrode cloth based on charge transfer and latent image detection. The method involves generating a fine, controllable gaseous charge probe on the electrode cloth surface and transferring the surface charge to an elastic insulating medium with charge traps (for long-term charge locking). A high-precision non-contact voltmeter is then used to acquire all charge potential data, forming a charge latent image. This latent image is input into a pre-trained dual-modal physical-guided defect detection model for analysis. This achieves high-precision identification and physical quantification of defects under small sample conditions, and outputs information on defect type, location, contour, area, and depth.

[0005] The objective of this invention is achieved in part by providing a method for detecting surface defects in lithium electrode cloth based on charge transfer and latent image detection, comprising the following steps:

[0006] A method for detecting surface defects in lithium electrode cloth based on charge transfer and latent image detection, characterized by comprising the following steps:

[0007] S1) On the surface of the lithium electrode cloth to be tested, a fine gaseous charge probe is generated by a controllable nitrogen corona device and uniformly deposited on the surface of the lithium electrode cloth to be tested.

[0008] S2) Charge distribution transfer and locking steps: Within the relaxation time window before the gaseous charge probe completely decays, the elastic insulating receiving medium with charge trapping characteristics is conformally pressed and quickly peeled off from the surface of the lithium electrode cloth, so that the charge on the surface of the electrode cloth is transferred and locked to the receiving medium, forming a charge distribution that is not easily decayed.

[0009] S3) Charge latent image generation steps: A non-contact electrostatic voltmeter array is used to scan and measure the charge on the surface of the receiving medium to obtain charge potential data at all points. Combined with adaptive threshold wavelet denoising based on noise energy estimation, feature extraction and normalization processing, a charge latent image that can be used for identification and analysis is formed.

[0010] S4) The charge latent image is input into the pre-trained dual-modal physical guided defect detection model. The dual-modal physical guided defect detection model extracts spatial geometric features and potential physical features through a dual-branch deformable coding module. The model then performs dual-channel long-range correlation coding, bottleneck aggregation, and multi-level scale adaptive fusion decoding through a hierarchical fusion coding and decoding module. Finally, the model outputs the defect category, location, contour, area, and depth information in parallel through a multi-task output head.

[0011] Furthermore, the generation of the gaseous charge probe in step 1) is controlled by a high-voltage pulsed gas discharge system. The discharge system operates in pulse mode with a pulse width of 50-200 μs and a repetition frequency preferably of 1-10 kHz. The relaxation time window is controlled by a timing control module, and the time interval from the completion of charge deposition to the start of transfer pressing is no more than 100 ms.

[0012] Furthermore, in step 2), the elastic insulating receiving medium is a polydimethylsiloxane composite material doped with nano-zinc oxide particles as charge trapping centers. The nano-zinc oxide particles have a particle size of 20-50 nm and a doping amount of 2-5 wt%. The receiving medium is a high resistivity insulating composite material with a bulk resistivity of 10 Ω·cm. 12 -10 15 The thickness is on the order of Ω·cm and ranges from 50 to 200 μm.

[0013] Furthermore, the surface of the receiving medium is covered with a barrier layer, which is aluminum oxide or tantalum pentoxide, and has a thickness of 2-10 nm.

[0014] Furthermore, the conformal pressing described in step 2) is achieved by a pneumatic roller with a pressure of 0.3-0.8 MPa, a contact time of 50-200 ms, a peeling angle of 30-45°, and a peeling speed of 10-50 mm / s.

[0015] Furthermore, step 3) specifically includes: measuring charge potential data through a dynamic scanning measurement system, which consists of multiple micrometer-scale electrode arrays with a spacing of less than 1 mm between adjacent electrodes;

[0016] An adaptive threshold wavelet denoising and minimum-maximum normalization method based on noise energy estimation was adopted. The detection was carried out in a low humidity environment with an ambient humidity not exceeding 5%RH.

[0017] Adaptive threshold wavelet denoising based on noise energy estimation employs an improved threshold function:

[0018]

[0019] Where λ is the adaptive threshold, derived from the noise energy estimation function. calculate, Let N be the noise standard deviation and N be the signal length. Normalization is performed using a modified min-max formula:

[0020]

[0021] Where μ is the mean and σ is the standard deviation, the 3σ principle is used to achieve more robust data distribution scaling.

[0022] Furthermore, the dual-modal physical-guided defect detection model described in step 4) includes a dual-branch deformable coding module and a layered fusion coding and decoding module;

[0023] The dual-branch deformable coding module includes a spatial geometry coding branch and a potential physics coding branch, which are used to extract features from the two dimensions of geometric shape and potential physics, respectively.

[0024] The spatial geometry coding branch receives the latent image of the charge to be measured. The data is sequentially processed through a first downsampling convolutional layer (3×3 kernel, stride 2, 16 output channels), a feature extraction layer (3×3 kernel, stride 1, 32 output channels), a data-driven deformable convolutional layer, and a channel adjustment layer (1×1 kernel, 128 output channels) to extract the geometric features of the defects and output a spatial feature map. The offset field and kernel parameters of the deformable convolutional layer are obtained through backpropagation end-to-end learning, which can adapt to irregular defect shapes such as scratches and bubbles.

[0025] The potential physical coding branch receives the latent image of the charge to be measured. The potential gradient is calculated sequentially through the gradient calculation layer:

[0026]

[0027] The physical offset calculation layer generates the sampling offset according to the following formula:

[0028]

[0029] Where λ is a learnable parameter and ε is a numerical stability constant, the learnable parameter λ is jointly optimized through backpropagation and other parameters of the encoding branch; then it is sampled along the electric field line direction by a physically driven deformable sampling layer, and finally encoded by a feature encoding layer (3×3 convolution kernel, 4 stride, 128 output channels) to output a potential feature map. ;

[0030] The layered fusion encoding and decoding module receives the spatial feature map and the potential feature map, including a dual-channel encoding path, a shared bottleneck layer, a fusion decoding path, and a multi-task output head;

[0031] The dual-channel encoding path includes a spatial encoding path and a potential encoding path, each containing three downsampling encoding layers. Each encoding layer contains a bidirectional selective scanning mechanism, which updates state variables through the following state-space equation to capture long-range correlations in charge distribution:

[0032]

[0033] in , The learnable state transition parameters are updated through backpropagation; the first-scale encoded features are output after processing by the first encoding layer, the second-scale encoded features are output after processing by the second encoding layer, and the third-scale encoded features are output after processing by the third encoding layer; the spatial dimension is reduced between layers through downsampling;

[0034] The shared bottleneck layer is located at the bottom of the encoder-decoder structure. It receives the underlying encoding features of the spatial encoding path and the potential encoding path, aggregates the dual-modal global information, and outputs the bottleneck features.

[0035] The fusion decoding path comprises three upsampling decoding layers, each containing a bidirectional selective scanning mechanism. The fusion decoding path receives spatially encoded features and potential-encoded features of the corresponding scale via skip connections, and then fuses them using a hierarchical adaptive fusion mechanism before inputting them into the corresponding decoder layer. The hierarchical adaptive fusion mechanism integrates the spatially encoded features of the l-th layer... With potential coding features Perform dual-modal physical residual gated fusion; first, calculate the dual-modal physical residual:

[0036]

[0037] The physical residual represents the degree of consistency between geometric and potential features at their corresponding spatial locations; the physical residual is then processed through 3×3 convolutional encoding and gated nonlinear activation to generate spatially varying fusion weights. Calculate the fusion features using the following formula:

[0038]

[0039] in This represents element-wise multiplication; the fusion weights Dynamic adjustment based on dual-modal physical consistency: in the region of geometric and potential response consistency ( Small) tends to 1 based on the geometric characteristics of the trust space, in regions where the potential changes abruptly but the space is relatively flat ( The 3×3 convolution weights are learned through backpropagation and tend to 0 to trust the physical characteristics of the potential.

[0040] The multi-task output head includes a defect detection head, a defect segmentation head, and a defect quantization head, which are used to receive the final layer decoding features and output defect detection results and defect quantization parameters in parallel.

[0041] The defect detection head uses global average pooling and a fully connected structure to classify and regress the final layer decoding features, and outputs the defect category probability and defect location information to achieve global discrimination and localization of defect types.

[0042] The defect segmentation head uses a fully convolutional structure to upsample and convolutionally classify the features decoded in the last layer, and outputs a pixel-level defect segmentation mask after Softmax normalization, so as to achieve fine extraction of defect contours.

[0043] The defect quantization head uses a convolutional structure to upsample the decoding features of the final layer, and based on electrostatic principles, calculates the defect depth according to the following potential-depth physical relationship formula within a range where defect depth and potential decay are linearly related. (unit ):

[0044]

[0045] in The calibration coefficients are learnable and are set based on prior physical calibration values ​​during initialization, and fine-tuned through backpropagation during end-to-end training. This is the potential in the normal region. This represents the potential of the defect region.

[0046] Calculate the defect area using the following area calculation formula. (unit ):

[0047]

[0048] in This represents the number of pixels in the defect area. pixel size (unit) This enables the quantitative measurement of defect geometric parameters.

[0049] Another aspect of the objective of this invention is achieved as follows: a lithium electrode cloth surface defect detection system based on charge transfer and latent image detection, comprising a gaseous charge probe generation module, a transfer module, a charge latent image generation module, and an intelligent recognition module.

[0050] The gaseous charge probe generation module is used to generate and deposit gaseous charge probes on the surface of the lithium electrode cloth to be tested.

[0051] The transfer module is used to transfer the charge distribution from the lithium electrode cloth to the elastic insulating receiving medium before the charge decays, so as to achieve stable preservation of charge information.

[0052] The charge latent image generation module is used to scan, measure, and preprocess the charge potential data on the surface of the receiving medium.

[0053] The intelligent recognition module is used to achieve high-precision recognition and physical quantification of defects under small sample conditions based on a dual-modal physical guided defect detection model, and outputs defect category, location, contour, area and depth information.

[0054] Furthermore, the intelligent recognition module includes a dual-branch deformable coding submodule and a layered fusion coding and decoding submodule;

[0055] The dual-branch deformable coding submodule includes a spatial geometry coding unit and a potential physical coding unit. The spatial geometry coding unit extracts geometric features sequentially through a downsampling convolutional layer, a feature extraction layer, a data-driven deformable convolutional layer, and a channel adjustment layer, and outputs a spatial feature map. The potential physical coding unit extracts potential physical features sequentially through a gradient calculation layer, a physical offset calculation layer, a physical-driven deformable sampling layer, and a feature coding layer, and outputs a potential feature map.

[0056] The hierarchical fusion encoding and decoding submodule includes a dual-channel encoding unit, a shared bottleneck layer, a fusion decoding unit, and a multi-task output unit. The dual-channel encoding unit performs multi-layer bidirectional selective scanning encoding and downsampling on the spatial feature map and the potential feature map, respectively. The shared bottleneck layer aggregates the dual-channel bottom-level encoded features. The fusion decoding unit includes multiple upsampling decoding layers. Each layer receives spatial encoded features and potential encoded features of the corresponding scale through skip connections. After dynamic weighted fusion of the dual-modal encoded features of the layer through a hierarchical adaptive fusion mechanism, the features are input into the decoder of the layer, achieving global receptive field modeling with linear complexity.

[0057] The multi-task output unit includes a defect detection head, a defect segmentation head, and a defect quantization head, which are respectively responsible for defect category discrimination and localization, pixel-level contour extraction, and depth and area quantization.

[0058] Furthermore, the transfer module includes a receiving medium supply mechanism, a pressing roller, a peeling mechanism, and a timing controller, which ensures that the time interval from the completion of charge deposition to the start of transfer is <100ms.

[0059] Compared with existing technologies, the advantages of this invention are as follows: This invention generates controllable fine gaseous charge probes on the surface of the electrode cloth and transfers their non-uniform charge distribution to an elastic insulating medium with charge traps. Then, a high-precision non-contact voltmeter is used to acquire the spatial distribution data of the charge. This data is then analyzed using a dual-modal physically guided defect detection model based on a physical information neural network. This achieves high-precision identification and physical quantification of defects under small sample conditions, and outputs information on defect type, location, contour, area, and depth. This method not only enables high-precision and rapid detection of minute surface defects, but also reduces the number of learnable parameters by more than 95% compared to a pure data-driven convolutional neural network of the same number of layers, achieving small-sample training (requiring only 50-100 labeled samples). Simultaneously, the output depth and area have clear physical traceability and interpretability, meeting the metrological requirements of industrial testing. Furthermore, this invention effectively reduces environmental humidity interference, enhances signal stability and resolution, and significantly improves the reliability and practicality of lithium electrode cloth surface defect detection through charge transfer and locking, dynamic multi-point measurement, and multi-task output with physical constraints. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0061] Figure 1 A flowchart of the method of the present invention.

[0062] Figure 2 Schematic diagram of the charge transfer and locking process in this invention;

[0063] (a) Schematic diagram of charge deposition state (b) Schematic diagram of conformal pressing state (c) Schematic diagram of peeling and locking state.

[0064] Figure 3 A schematic diagram of the layered structure of the receiving medium in this invention.

[0065] Figure 4 A schematic diagram comparing the charge distribution in the normal region and the defect region in this invention.

[0066] Figure 5 The curve showing the effect of relaxation time on signal retention rate in this invention.

[0067] Figure 6 Schematic diagram of the dual-modal physical-guided defect detection model architecture in this invention;

[0068] (a) Schematic diagram of spatial geometric coding branch structure (b) Schematic diagram of potential physical coding branch structure (c) Schematic diagram of layered fusion coding and decoding module architecture. Detailed Implementation

[0069] 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.

[0070] like Figure 1 The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection includes the following steps:

[0071] S1) On the surface of the lithium electrode cloth to be tested, a fine gaseous charge probe is generated by a controllable nitrogen corona device and uniformly deposited on the surface of the lithium electrode cloth to be tested. The deposition amount and distribution of the gaseous charge probe can be controlled according to the characteristics of the sample to be tested.

[0072] The generation of the gaseous charge probe is controlled by a high-voltage pulsed gas discharge system. This system operates in pulse mode with a pulse width of 50-200 μs and a preferred repetition frequency of 1-10 kHz. The effect of relaxation time on signal retention is as follows: Figure 5 As shown, the relaxation time window is controlled by the timing control module, and the time interval from the completion of charge deposition to the start of transfer bonding is no more than 100 ms, preferably 10-50 ms.

[0073] like Figure 4As shown, defects such as pores, scratches, foreign matter, uneven coating, and local contamination on the surface of lithium electrode cloth can change the local dielectric properties, surface resistance, roughness, and charge discharge and migration paths. Therefore, under the same conditions, the charge deposited in different micro-regions will form different spatial distribution characteristics. Normal areas usually show a relatively continuous and gentle potential distribution, while defective areas show characteristics such as potential abrupt changes, abnormal gradients, local peaks and valleys, or discontinuous boundaries. Thus, the material or structural differences of defects are mapped into detectable differences in charge distribution.

[0074] S2) Charge distribution transfer and locking steps: Within the relaxation time window before the gaseous charge probe completely decays, the elastic insulating receiving medium with charge trapping characteristics is conformally pressed and quickly peeled off from the surface of the lithium electrode cloth, so that the charge on the surface of the electrode cloth is transferred and locked to the receiving medium, forming a charge distribution that is not easily decayed.

[0075] like Figure 2 As shown, the charge gradually decays in air and on the material surface. Therefore, this method completes the transfer and locking operations within the relaxation time window after charge deposition, i.e., the time interval from the completion of charge deposition to the start of transfer pressing is controlled within 100ms, preferably 10-50ms. Subsequently, the elastic insulating receiving medium with charge trapping characteristics is conformally pressed onto the surface of the lithium electrode cloth, and rapidly peeled off at the set peel angle and peel speed, so that the non-uniform charge distribution on the surface of the electrode cloth is transferred to the surface of the receiving medium. Since the receiving medium is a cross-linked solid elastomer and the transfer process does not involve liquid developer, the contact time is short and the peeling speed is fast. Therefore, this process is mainly an instantaneous physical transfer process, which helps to reduce the risk of adhesion residues on the surface of the lithium electrode cloth.

[0076] like Figure 3 As shown, the elastic insulating receiving medium is a polydimethylsiloxane composite material doped with nano-zinc oxide particles as charge trapping centers. The particle size of the nano-zinc oxide particles is 20-50 nm, and the doping amount is 2-5 wt%. The receiving medium is a high resistivity insulating composite material with a bulk resistivity of 10 Ω·cm. 12 -10 15 The thickness is on the order of Ω·cm and ranges from 50 to 200 μm. The surface of the receiving medium is covered with a barrier layer, which is made of alumina or tantalum pentoxide and has a thickness of 2-10 nm, to prevent surface leakage of charge after transfer. Conformal pressing is achieved by a pneumatic roller with a pressure of 0.3-0.8 MPa, a contact time of 50-200 ms, a peel angle of 30-45°, and a peel speed of 10-50 mm / s.

[0077] S3) Charge latent image generation steps: A non-contact electrostatic voltmeter array is used to scan and measure the charge on the surface of the receiving medium to obtain charge potential data at all points. Combined with adaptive threshold wavelet denoising based on noise energy estimation, feature extraction and normalization processing, a charge latent image that can be used for identification and analysis is formed.

[0078] All potentials are scanned using a dynamic multi-point measurement system, which consists of at least 32 micrometer-level electrode arrays with a spacing of less than 1 mm between adjacent electrodes. The model training employs end-to-end joint optimization, and the total loss function is a weighted combination of defect classification loss, defect segmentation loss, defect detection loss, and defect quantization loss. The defect quantization loss is the mean square error loss between the predicted depth / area and the actual measured value. In this embodiment, the dynamic multi-point measurement system uses a non-contact electrostatic voltmeter, which consists of multiple micrometer-level sensing electrodes. A high-precision micro-displacement mechanism adjusts the distance between the electrodes and the surface of the lithium electrode cloth under test, ensuring that the spatial resolution reaches the sub-millimeter level.

[0079] An adaptive threshold wavelet denoising and minimum-maximum normalization method based on noise energy estimation was adopted. The detection was carried out in a low humidity environment with an ambient humidity not exceeding 5%RH.

[0080] Adaptive threshold wavelet denoising based on noise energy estimation employs an improved threshold function:

[0081]

[0082] Where λ is the adaptive threshold, derived from the noise energy estimation function. calculate, Let N be the noise standard deviation and N be the signal length. Normalization is performed using a modified min-max formula:

[0083]

[0084] Where μ is the mean and σ is the standard deviation, the 3σ principle is used to achieve more robust data distribution scaling.

[0085] S4) The charge latent image is input into the pre-trained dual-modal physical guided defect detection model. The dual-modal physical guided defect detection model extracts spatial geometric features and potential physical features through a dual-branch deformable coding module. The model then performs dual-channel long-range correlation coding, bottleneck aggregation, and multi-level scale adaptive fusion decoding through a hierarchical fusion coding and decoding module. Finally, the model outputs the defect category, location, contour, area, and depth information in parallel through a multi-task output head.

[0086] The dual-modal physical-guided defect detection model includes a dual-branch deformable coding module and a layered fusion coding and decoding module;

[0087] The dual-branch deformable coding module includes a spatial geometry coding branch and a potential physics coding branch, which are used to extract features from the two dimensions of geometric shape and potential physics, respectively.

[0088] like Figure 6 As shown in (a), the spatial geometric coding branch receives the latent image of the charge to be measured. The data is sequentially processed through a first downsampling convolutional layer (3×3 kernel, stride 2, 16 output channels), a feature extraction layer (3×3 kernel, stride 1, 32 output channels), a data-driven deformable convolutional layer, and a channel adjustment layer (1×1 kernel, 128 output channels) to extract the geometric features of the defects and output a spatial feature map. The offset field and kernel parameters of the deformable convolutional layer are obtained through backpropagation end-to-end learning, which can adapt to irregular defect shapes such as scratches and bubbles.

[0089] like Figure 6 As shown in (b), the potential physical coding branch receives the latent image of the charge to be measured. The potential gradient is calculated sequentially through the gradient calculation layer:

[0090]

[0091] The physical offset calculation layer generates the sampling offset according to the following formula:

[0092]

[0093] Where λ is a learnable parameter and ε is a numerical stability constant, the learnable parameter λ is jointly optimized through backpropagation and other parameters of the encoding branch; then it is sampled along the electric field line direction by a physically driven deformable sampling layer, and finally encoded by a feature encoding layer (3×3 convolution kernel, 4 stride, 128 output channels) to output a potential feature map. ;

[0094] like Figure 6 As shown in (c), the layered fusion encoding and decoding module receives the spatial feature map and the potential feature map, including a dual-channel encoding path, a shared bottleneck layer, a fusion decoding path, and a multi-task output head;

[0095] The dual-channel encoding path includes a spatial encoding path and a potential encoding path, each containing three downsampling encoding layers. Each encoding layer contains a bidirectional selective scanning mechanism, which updates state variables through the following state-space equation to capture long-range correlations in charge distribution:

[0096]

[0097] in , The learnable state transition parameters are updated through backpropagation; the first-scale encoded features are output after processing by the first encoding layer, the second-scale encoded features are output after processing by the second encoding layer, and the third-scale encoded features are output after processing by the third encoding layer; the spatial dimension is reduced between layers through downsampling;

[0098] The shared bottleneck layer is located at the bottom of the encoder-decoder structure. It receives the underlying encoding features of the spatial encoding path and the potential encoding path, aggregates the dual-modal global information, and outputs the bottleneck features.

[0099] The fusion decoding path comprises three upsampling decoding layers, each containing a bidirectional selective scanning mechanism. The fusion decoding path receives spatial and potential coding features of the corresponding scale via skip connections, and after fusion by a hierarchical adaptive fusion mechanism, inputs them into the corresponding layer's decoder. The hierarchical adaptive fusion mechanism will... Layer spatial coding features With potential coding features Perform dual-modal physical residual gated fusion. First, calculate the dual-modal physical residual:

[0100]

[0101] The physical residual represents the degree of consistency between geometric and potential features at their corresponding spatial locations; the physical residual is then processed through 3×3 convolutional encoding and gated nonlinear activation to generate spatially varying fusion weights. Calculate the fusion features using the following formula:

[0102]

[0103] in This represents element-wise multiplication; the fusion weights Dynamic adjustment based on dual-modal physical consistency: in the region of geometric and potential response consistency ( Small) tends to 1 based on the geometric characteristics of the trust space, in regions where the potential changes abruptly but the space is relatively flat ( The 3×3 convolution weights are learned through backpropagation and tend to 0 to trust the physical characteristics of the potential.

[0104] The multi-task output head includes a defect detection head, a defect segmentation head, and a defect quantization head, which are used to receive the final layer decoding features and output defect detection results and defect quantization parameters in parallel.

[0105] The defect detection head uses global average pooling and a fully connected structure to classify and regress the features decoded in the last layer, and outputs the probability of defect category and defect location information to achieve global discrimination and localization of defect type.

[0106] The defect segmentation head uses a fully convolutional structure to upsample and convolutionally classify the features decoded in the last layer, and outputs a pixel-level defect segmentation mask after Softmax normalization, so as to achieve fine extraction of defect contours.

[0107] The defect quantization head uses a convolutional structure to upsample the decoding features of the final layer, and based on electrostatic principles, calculates the defect depth according to the following potential-depth physical relationship formula within a range where defect depth and potential decay are linearly related. (unit ):

[0108]

[0109] Where k is a learnable calibration coefficient, which is set according to the prior physical calibration value during initialization and fine-tuned through backpropagation during end-to-end training; This is the potential in the normal region. This represents the potential of the defect region.

[0110] Calculate the defect area using the following area calculation formula. (unit ):

[0111]

[0112] Where N is the number of pixels in the defect area. pixel size (unit) This enables the quantitative measurement of defect geometric parameters.

[0113] To further illustrate the advantages of the method of the present invention over existing detection methods in terms of micro-defect identification, signal stability and environmental adaptability, the detection effects of different detection methods are compared, and the results are shown in Table 1.

[0114] Table 1

[0115]

[0116] As shown in Table 1, compared with manual visual inspection and ordinary machine vision methods, the method of this invention, by embedding prior knowledge such as gradient calculation and potential-depth physical relationship into the network structure, maintains a detection accuracy of over 95% even with a 90% reduction in training samples compared to purely data-driven deep learning models. Furthermore, it can complete model training without requiring a large number of defect samples, significantly reducing data collection costs in industrial settings. It also outperforms other indicators such as minimum detectable defect size and false detection rate, while exhibiting higher signal stability and lower humidity sensitivity.

[0117] A surface defect detection system for lithium electrode cloth based on charge transfer and latent image detection includes a gaseous charge probe generation module, a transfer module, a charge latent image generation module, and an intelligent recognition module.

[0118] The gaseous charge probe generation module is used to generate and deposit gaseous charge probes on the surface of the lithium electrode cloth to be tested;

[0119] The transfer module is used to transfer the charge distribution from the lithium electrode cloth to the elastic insulating receiving medium before the charge decays, thereby saving the charge information. The transfer module includes a receiving medium supply mechanism, a pressing roller, a peeling mechanism, and a timing controller. The timing controller ensures that the time interval from the completion of charge deposition to the start of transfer is <100ms.

[0120] The charge latent image generation module is used to scan and measure all points on the surface of the receiving medium and generate a charge latent image;

[0121] The intelligent recognition module is used to achieve high-precision recognition and physical quantification of defects under small sample conditions based on a dual-modal physical guided defect detection model, and outputs defect category, location, contour, area and depth information;

[0122] The intelligent recognition module includes a dual-branch deformable coding submodule and a layered fusion coding and decoding submodule;

[0123] The dual-branch deformable coding submodule includes a spatial geometry coding unit and a potential physical coding unit. The spatial geometry coding unit extracts geometric features sequentially through a downsampling convolutional layer, a feature extraction layer, a data-driven deformable convolutional layer, and a channel adjustment layer, outputting a spatial feature map. The potential physical coding unit extracts potential physical features sequentially through a gradient calculation layer, a physical offset calculation layer, a physical-driven deformable sampling layer, and a feature coding layer, outputting a potential feature map. The hierarchical fusion coding and decoding submodule includes a dual-channel coding unit, a shared bottleneck layer, a fusion decoding unit, and a multi-task output unit. The dual-channel coding unit performs multi-layer bidirectional selective scanning coding and downsampling on the spatial feature map and the potential feature map, respectively. The shared bottleneck layer aggregates the dual-channel bottom-level coding features. The fusion decoding unit contains multiple upsampling decoding layers. Each layer receives spatial coding features and potential coding features of the corresponding scale through skip connections. After dynamic weighted fusion of the dual-modal coding features of the layer through a hierarchical adaptive fusion mechanism, the data is input into the decoder of that layer to achieve global receptive field modeling with linear complexity.

[0124] The multi-task output unit includes a defect detection head, a defect segmentation head, and a defect quantization head, which are responsible for defect category discrimination and localization, pixel-level contour extraction, and depth and area quantization, respectively.

[0125] This lithium electrode cloth surface defect detection system uses a gaseous charge probe generation module to deposit controllable charges on the electrode cloth surface, and then a transfer module to transfer the non-uniform charge distribution to an elastic insulating medium with charge traps, thus achieving stable preservation of charge information. The charge latent image generation module scans all points with high precision and generates charge latent images. The intelligent recognition module is based on a dual-modal physical guided defect detection model. It extracts geometric morphological features and potential physical features through a dual-branch deformable coding submodule. Through a hierarchical fusion encoding and decoding submodule, it performs dual-channel long-range correlation encoding, bottleneck aggregation, and multi-level scale adaptive fusion decoding. Finally, through a multi-task output unit, it outputs the defect category, pixel-level segmentation mask, and quantitative parameters such as depth and area in parallel, achieving high-sensitivity, high-resolution, fast, and reliable lithium electrode cloth surface defect detection.

[0126] like Figure 1 and Figure 4 As shown, this method first utilizes a controllable nitrogen corona device to ionize nitrogen gas under a high-frequency, high-voltage electric field, forming a fine gaseous charge probe, which is then uniformly and tunably deposited onto the surface of a lithium electrode cloth. Since defects on the lithium electrode cloth surface, such as pores, scratches, foreign matter, uneven coating, and localized contamination, alter local dielectric properties, surface resistance, roughness, and charge dissipation and migration paths, the charges deposited in different micro-regions under the same conditions will exhibit different spatial distribution characteristics. Normal regions typically show a relatively continuous and gentle potential distribution, while defective regions exhibit characteristics such as potential abrupt changes, abnormal gradients, local peaks and valleys, or discontinuous boundaries, thus mapping the material or structural differences of the defects into detectable differences in charge distribution.

[0127] like Figure 2 , Figure 3 and Figure 5 As shown, charges gradually decay in air and on material surfaces. Therefore, this method completes the transfer and locking operations within a relaxation time window after charge deposition, i.e., the time interval from the completion of charge deposition to the start of transfer pressing is controlled within 100 ms, preferably 10–50 ms. Subsequently, an elastic insulating receiving medium with charge trapping properties is conformally pressed onto the surface of the lithium electrode cloth, and rapidly peeled off at a set peel angle and peel speed. This transfers the non-uniform charge distribution on the electrode cloth surface to the receiving medium surface. Since the receiving medium is a cross-linked solid elastomer and the transfer process does not involve liquid developer, the contact time is short and the peeling speed is fast. Therefore, this process is mainly an instantaneous physical transfer process, which helps to reduce the risk of residual adhesion to the lithium electrode cloth surface. The receiving medium uses a polydimethylsiloxane composite material, with nano-zinc oxide particles doped inside to form charge trap centers, and an ultra-thin high-barrier layer covering the surface to suppress surface charge leakage, thereby locking the originally easily decaying transient charge information into a stable charge latent image.

[0128] Next, in the charge latent image generation step, a high-precision non-contact electrostatic voltmeter array is used to scan and measure the charge on the surface of the receiving medium, acquiring the potential data and spatial distribution information of each measurement point. This data is then combined with adaptive threshold wavelet denoising based on noise energy estimation, feature extraction, and normalization processing to form a charge latent image that can be used for identification and analysis. Preferably, the detection process is carried out in a low-humidity environment with an ambient humidity not exceeding 5%RH to reduce the interference of moisture on the charge signal and improve signal stability.

[0129] like Figure 6 As shown, in the final identification and analysis step, the charge latent image is input into a pre-trained dual-modal physical-guided defect detection model. Spatial geometric features and potential physical features are extracted by a dual-branch deformable coding module, respectively. These are then processed by a hierarchical fusion encoding and decoding module for dual-channel long-range correlation encoding, bottleneck aggregation, and multi-layer scale adaptive fusion decoding. Finally, a multi-task output head outputs pixel-level segmentation masks, defect category probabilities, and quantization parameters such as defect depth and area in parallel. Specifically, the defect segmentation head achieves fine extraction of the defect contour, the defect detection head achieves global discrimination and localization of the defect type, and the defect quantization head calculates the defect geometric parameters based on the potential-depth physical relationship.

[0130] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for detecting surface defects in lithium electrode cloth based on charge transfer printing and latent image detection, characterized in that, Includes the following steps: S1) On the surface of the lithium electrode cloth to be tested, a fine gaseous charge probe is generated by a controllable nitrogen corona device and uniformly deposited on the surface of the lithium electrode cloth to be tested. S2) Charge distribution transfer and locking steps: Within the relaxation time window before the gaseous charge probe completely decays, the elastic insulating receiving medium with charge trapping characteristics is conformally pressed and quickly peeled off from the surface of the lithium electrode cloth, so that the charge on the surface of the electrode cloth is transferred and locked to the receiving medium, forming a charge distribution that is not easily decayed. S3) Charge latent image generation steps: A non-contact electrostatic voltmeter array is used to scan and measure the charge on the surface of the receiving medium to obtain charge potential data at all points. Combined with adaptive threshold wavelet denoising based on noise energy estimation, feature extraction and normalization processing, a charge latent image for identification and analysis is formed. S4) The charge latent image is input into the pre-trained dual-modal physical guided defect detection model. The dual-modal physical guided defect detection model extracts spatial geometric features and potential physical features through a dual-branch deformable coding module. The model then performs dual-channel long-range correlation coding, bottleneck aggregation, and multi-level scale adaptive fusion decoding through a hierarchical fusion coding and decoding module. Finally, the model outputs the defect category, location, contour, area, and depth information in parallel through a multi-task output head.

2. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 1, characterized in that, The generation of the gaseous charge probe in step 1) is controlled by a high-voltage pulsed gas discharge system. The discharge system operates in pulse mode with a pulse width of 50-200 μs and a repetition frequency preferably of 1-10 kHz. The relaxation time window is controlled by a timing control module, and the time interval from the completion of charge deposition to the start of transfer bonding is no more than 100 ms.

3. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 1, characterized in that, In step 2), the elastic insulating receiving medium is a polydimethylsiloxane composite material doped with nano-zinc oxide particles as charge trapping centers. The nano-zinc oxide particles have a particle size of 20-50 nm and a doping amount of 2-5 wt%. The receiving medium is a high resistivity insulating composite material with a bulk resistivity of 10 Ω·cm. 12 -10 15 The thickness is on the order of Ω·cm and ranges from 50 to 200 μm.

4. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 3, characterized in that, The surface of the receiving medium is covered with a barrier layer, which is aluminum oxide or tantalum pentoxide, and has a thickness of 2-10 nm.

5. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 1, characterized in that, The conformal pressing described in step 2) is achieved by a pneumatic roller with a pressure of 0.3-0.8 MPa, a contact time of 50-200 ms, a peeling angle of 30-45°, and a peeling speed of 10-50 mm / s.

6. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 1, characterized in that, Step 3) specifically includes: generating a charge latent image through a dynamic multi-point measurement system, which consists of multiple micrometer-scale electrode arrays with a spacing of less than 1 mm between adjacent electrodes; An adaptive threshold wavelet denoising and minimum-maximum normalization method based on noise energy estimation was adopted. The detection was carried out in a low humidity environment with an ambient humidity not exceeding 5%RH. Adaptive threshold wavelet denoising based on noise energy estimation employs an improved threshold function: ； Where λ is the adaptive threshold, derived from the noise energy estimation function. calculate, Let N be the noise standard deviation and N be the signal length. Normalization is performed using a modified min-max formula: ； Where μ is the mean and σ is the standard deviation, the 3σ principle is used to achieve more robust data distribution scaling.

7. The method for detecting surface defects of lithium electrode cloth based on charge transfer and latent image detection according to claim 1, characterized in that, The dual-modal physical-guided defect detection model described in step 4) includes a dual-branch deformable coding module and a layered fusion coding and decoding module; The dual-branch deformable coding module includes a spatial geometry coding branch and a potential physics coding branch, which are used to extract features from the two dimensions of geometric shape and potential physics, respectively. The spatial geometry coding branch receives the latent image of the charge to be measured. The data is sequentially passed through a first downsampling convolutional layer, a feature extraction layer, a data-driven deformable convolutional layer, and a channel adjustment layer to extract the geometric morphological features of the defects and output a spatial feature map. The offset field and kernel parameters of the deformable convolutional layer are obtained through backpropagation end-to-end learning, which can adapt to the irregular defect shapes of scratches and bubbles. The potential physical coding branch receives the latent image of the charge to be measured. The potential gradient is calculated sequentially through the gradient calculation layer: ； The physical offset calculation layer generates the sampling offset according to the following formula: ； Where λ is a learnable parameter and ε is a numerical stability constant, the learnable parameter λ is jointly optimized through backpropagation and other parameters of the encoding branch; then it is sampled along the electric field line direction by a physically driven deformable sampling layer, and finally encoded by a feature encoding layer to output a potential feature map. ; The layered fusion encoding and decoding module receives spatial feature maps and potential feature maps, including a dual-channel encoding path, a shared bottleneck layer, a fusion decoding path, and a multi-task output head; The dual-channel encoding path includes a spatial encoding path and a potential encoding path, each containing three downsampling encoding layers. Each encoding layer contains a bidirectional selective scanning mechanism, which updates state variables through the following state-space equation to capture long-range correlations in charge distribution: ； in , These are learnable state transition parameters, updated via backpropagation. The first coding layer processes the first-scale coded features, the second coding layer processes the second-scale coded features, and the third coding layer processes the third-scale coded features; the spatial dimension is reduced between layers through downsampling. The shared bottleneck layer is located at the bottom of the encoder-decoder structure. It receives the underlying encoding features of the spatial encoding path and the potential encoding path, aggregates the dual-modal global information, and outputs the bottleneck features. The fusion decoding path comprises three upsampling decoding layers, each containing a bidirectional selective scanning mechanism. The fusion decoding path receives spatially encoded features and potential-encoded features of the corresponding scale via skip connections, and then fuses them using a hierarchical adaptive fusion mechanism before inputting them into the corresponding decoder layer. The hierarchical adaptive fusion mechanism integrates the spatially encoded features of the l-th layer... With potential coding features Perform dual-modal physical residual gated fusion; first, calculate the dual-modal physical residual: ； Physical residuals characterize the consistency between geometric and potential features at their corresponding spatial locations; the physical residuals are then processed through 3×3 convolutional encoding and gated nonlinear activation to generate spatially varying fusion weights. ; Calculate the fusion features using the following formula: ； in, This represents element-wise multiplication; the fusion weights Dynamic adjustment based on bimodal physical consistency: tends to 1 in regions where geometry and potential response are consistent to trust spatial geometric features, and tends to 0 in regions where potential changes abruptly but space is smooth to trust potential physical features; 3×3 convolution weights are learned through backpropagation; The multi-task output head includes a defect detection head, a defect segmentation head, and a defect quantization head, which are used to receive the final layer decoding features and output defect detection results and defect quantization parameters in parallel. The defect detection head uses global average pooling and a fully connected structure to classify and regress the features decoded in the last layer, and outputs the probability of defect category and defect location information to achieve global discrimination and localization of defect type. The defect segmentation head uses a fully convolutional structure to upsample and convolutionally classify the features decoded in the last layer, and outputs a pixel-level defect segmentation mask after Softmax normalization, so as to achieve fine extraction of defect contours. The defect quantization head uses a convolutional structure to upsample the decoding features of the final layer, and based on electrostatic principles, calculates the defect depth according to the following potential-depth physical relationship formula within a range where defect depth and potential decay are linearly related. : ； Where k is a learnable calibration coefficient, which is set according to the prior physical calibration value during initialization and fine-tuned through backpropagation during end-to-end training; This is the potential in the normal region. This represents the potential of the defect region. Calculate the defect area S using the following area calculation formula: ； Where N is the number of pixels in the defect area and p is the pixel size, enabling quantitative measurement of the defect's geometric parameters.

8. A surface defect detection system for lithium electrode cloth based on charge transfer printing and latent image detection, characterized in that, It includes a gaseous charge probe generation module, a transfer module, a charge latent image generation module, and an intelligent recognition module; The gaseous charge probe generation module is used to generate and deposit gaseous charge probes on the surface of the lithium electrode cloth to be tested. The transfer module is used to transfer the charge distribution from the lithium electrode cloth to the elastic insulating receiving medium before the charge decays, thereby saving the charge information. The charge latent image generation module is used to scan and measure all points on the surface of the receiving medium and generate a charge latent image; The intelligent recognition module is used to achieve high-precision recognition and physical quantification of defects under small sample conditions based on a dual-modal physical guided defect detection model, and outputs defect category, location, contour, area and depth information.

9. The lithium electrode cloth surface defect detection system based on charge transfer and latent image detection according to claim 8, characterized in that, The intelligent recognition module includes a dual-branch deformable coding submodule and a layered fusion coding and decoding submodule; The dual-branch deformable coding submodule includes a spatial geometric coding unit and a potential physical coding unit; The spatial geometry encoding unit extracts geometric features sequentially through a downsampling convolutional layer, a feature extraction layer, a data-driven deformable convolutional layer, and a channel adjustment layer, and outputs a spatial feature map; the potential physical encoding unit extracts potential physical features sequentially through a gradient calculation layer, a physical offset calculation layer, a physical-driven deformable sampling layer, and a feature encoding layer, and outputs a potential feature map. The layered fusion encoding and decoding submodule includes a dual-channel encoding unit, a shared bottleneck layer, a fusion decoding unit, and a multi-task output unit; the dual-channel encoding unit performs multi-layer bidirectional selective scanning encoding and downsampling on the spatial feature map and the potential feature map, respectively; the shared bottleneck layer aggregates the dual-channel bottom-level encoded features. The fusion decoding unit comprises multiple upsampling decoding layers. Each layer receives spatial coding features and potential coding features of the corresponding scale through skip connections. The dual-modal coding features of the layer are dynamically weighted and fused by a hierarchical adaptive fusion mechanism before being input into the decoder of the layer, so as to achieve global receptive field modeling with linear complexity. The multi-task output unit includes a defect detection head, a defect segmentation head, and a defect quantization head, which are used for defect category discrimination and localization, pixel-level contour extraction, and depth and area quantization, respectively.

10. The lithium electrode cloth surface defect detection system based on charge transfer and latent image detection according to claim 8, characterized in that, The transfer module includes a receiving medium supply mechanism, a pressing roller, a peeling mechanism, and a timing controller, which ensures that the time interval from the completion of charge deposition to the start of transfer is <100ms.

Images