Image recognition-based online detection method for coating defects of micro lithium battery pole pieces

Abstract

The application discloses a micro lithium battery pole piece coating defect online detection method based on image recognition and relates to the technical field of image detection. The pole piece continuous image data and production time sequence data containing the pole piece conveying speed value are synchronously collected to generate a to-be-detected image sequence, and the pole piece conveying speed value is embedded into a metadata area; after the image sequence is divided into blocks, spatial domain cycle offset, frequency domain offset and gray scale variance are extracted in parallel and fused to construct a defect response graph; a response tuple bound with the centroid coordinates and the defect response value is generated by extracting the connected region; a variable-length response track is constructed based on the pole piece conveying speed value, the defect stability is judged according to the sign consistency ratio of the difference sequence, and a defect feature vector is generated; the variable-length response track is matched with a standard track through dynamic time warping, and a process adjustment instruction code is output. The application realizes the integration of the online detection and process diagnosis of the coating defect, can effectively distinguish the periodic weak defect from the image noise and reduce the false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, specifically to an online detection method for coating defects in micro lithium battery electrodes based on image recognition. Background Technology

[0002] The coating quality of micro lithium-ion battery electrodes directly determines the battery's electrochemical performance and safety. Accurate detection and real-time diagnosis of coating defects on high-speed automated production lines are crucial for ensuring product yield. In recent years, with the evolution of high-resolution imaging equipment and machine vision algorithms, electrode inspection solutions have gradually moved towards digitalization and intelligence.

[0003] Existing technologies typically acquire electrode surface images using line-scan cameras and leverage spatial domain feature extraction algorithms, such as edge detection or regional grayscale contrast, to identify obvious defects like missed coatings, scratches, or particles, enabling real-time monitoring of coating quality. These methods are effective in identifying common single-morphological defects and can perform preliminary screening based on preset thresholds, demonstrating basic online detection capabilities.

[0004] However, in the fine coating scenarios of micro lithium batteries, the electrode surface is often accompanied by complex micro-textures and periodic background interference caused by the reciprocating motion of the equipment. This complex visual environment makes it easy for the minute fluctuations of the coating layer in the spatial domain to be confused with image noise. Existing technologies rely too heavily on single spatial domain feature discrimination, lacking consideration for the global frequency stability of the electrode texture and the analysis of regularity over time. When the production equipment experiences systematic and periodic minor defects due to process factors such as roller eccentricity and slurry pressure pulsation, single spatial domain detection often exhibits insufficient detection sensitivity or a high false alarm rate. Because existing technologies have failed to establish a deep coupling relationship between "spatial features - frequency domain distribution - time series," the system can only eliminate superficial defects and cannot further trace back and lock down the underlying process parameter anomalies based on the dynamic behavior characteristics of the defects. This disconnect between detection and diagnosis makes it difficult for the production line to make closed-loop adjustments to target the root cause of defects, increasing the risk of systemic batch scrap. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an online detection method for coating defects in micro lithium battery electrodes based on image recognition.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This invention discloses an online detection method for coating defects on micro lithium battery electrodes based on image recognition, comprising the following steps:

[0008] Synchronously acquire continuous image data of the electrode sheet and production time series data containing electrode sheet conveying speed values. Use the production time series data as an index to perform spatiotemporal binding on the continuous image data of the electrode sheet to generate a sequence of images to be tested, and embed the electrode sheet conveying speed values ​​into the metadata area of ​​the sequence of images to be tested.

[0009] The image sequence to be tested is divided into local image blocks by sliding window. For each local image block, the spatial period offset, frequency offset and gray scale variance are extracted in parallel. After normalization, they are fused according to preset weight coefficients to generate the defect response value of the local image block. The defect response map is constructed by arranging them according to their original spatial positions.

[0010] Extract connected regions in the defect response map whose defect response values ​​exceed a preset defect threshold, output their centroid coordinates and defect response values, and bind the centroid coordinates and defect response values ​​into structural response tuples;

[0011] Read the electrode transmission speed value from the metadata area, use the electrode transmission speed value to perform inter-frame displacement compensation on the centroid coordinates in the response tuple to generate a physical coordinate sequence, and use the defect response value in the response tuple as the data load to construct a variable-length response trajectory along the physical coordinate sequence within a preset frame window.

[0012] Calculate the difference sequence of defect response values ​​of adjacent frames in the variable-length response trajectory, determine the defect stability based on the sign consistency ratio of the difference sequence, and generate a defect feature vector containing the variable-length response trajectory and stability label;

[0013] The variable-length response trajectory of the defect feature vector is dynamically time-warped and matched with the standard trajectory in the pre-stored process library, and the process adjustment instruction code with the smallest matching distance is output.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] 1. This invention uses the electrode transmission speed value simultaneously for both spatiotemporal binding indexing and inter-frame displacement compensation calculations, allowing the same speed data to be reused at different stages of the detection process. This enables the construction of a physical coordinate mapping relationship between image frames without the need for additional sensors. Based on this, a variable-length response trajectory is constructed along the physical coordinate sequence using the defect response value as the data payload. This allows the confidence information output by visual inspection to be forcibly bound to the time-series analysis module in the form of a data payload, forming a unique and legitimate interface between the two modules at the architectural level. This avoids data format incompatibility issues caused by independent module design.

[0016] 2. This invention extracts spatial periodic offset, frequency offset, and grayscale variance in parallel from the same local image patch and then weights and fuses them to generate a defect response value. This couples features of three different dimensions—spatial texture periodic deviation, frequency domain dominant frequency drift, and local energy variation—at the same pixel location. When spatial and frequency domain anomalies occur simultaneously, the fused defect response value exhibits a superposition enhancement effect, effectively improving the detection sensitivity of systematic weak defects while remaining insensitive to isolated noise points. Furthermore, since the three features originate from the same local image patch and are output in a unified single-channel response value format, subsequent connected component extraction and temporal analysis do not require separate processing for different features, simplifying the data flow path.

[0017] 3. This invention calculates the difference sequence of defect response values ​​in adjacent frames within a variable-length response trajectory and determines defect stability based on the sign consistency ratio of the difference sequence. This upgrades the defect discrimination criterion from a single-frame absolute response value threshold to a cross-frame response change trend. When the defect response value at the same physical location exhibits a monotonically consistent sign change in consecutive frames, it indicates that the defect is persistent; when the response value fluctuates alternately between positive and negative, it is judged as occasional noise. This discrimination mechanism allows for a clear distinction between persistent and random defects, thus providing prior information on the stability dimension for subsequent process parameter matching. Based on this, the variable-length response trajectory is dynamically time-warped and matched with standard trajectories in a pre-stored process library, enabling direct comparison of defect evolution trajectories of different time lengths without needing to align to a fixed number of frames. Attached Figure Description

[0018] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0019] Figure 1 This is a flowchart illustrating the working principle of the present invention;

[0020] Figure 2 This is a timing flowchart of the present invention. Detailed Implementation

[0021] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0022] In existing technologies, the detection of coating defects in micro lithium-ion battery electrodes largely relies on spatial texture feature comparison or grayscale thresholding of single-frame images. This approach struggles to balance the sensitivity to detect periodic, weak defects with the ability to suppress random image noise. Traditional methods are particularly vulnerable when complex micro-textures exist on the electrode surface or when periodic background interference arises from equipment mechanical movement. Small spatial fluctuations can easily be confused with actual defects, leading to insufficient detection sensitivity or a high false alarm rate. Existing visual inspection systems cannot simultaneously perceive the dynamic behavior of defects over time. Especially when equipment exhibits systematic, periodic, weak defects due to roller eccentricity or slurry pressure pulsation, single-frame detection models may experience systematic missed detections, failing to meet the requirements for integrated online inspection and process diagnostics.

[0023] To address the aforementioned issues, the study revealed that the response intensity of electrode coating defects exhibits a specific evolution pattern across consecutive frames, with persistent defects and sporadic noise showing significant differences in the sign consistency of their response intensity difference sequences. Furthermore, it was found that by simultaneously serving as a spatiotemporal binding index and inter-frame displacement compensation, the electrode transmission speed value can achieve accurate tracing of the same physical location between adjacent frames without relying on additional positioning sensors. Further experimental verification demonstrated that dynamically time-warping and matching the variable-length response trajectory with standard trajectories in a pre-stored process library can overcome the matching obstacle of inconsistent trajectory lengths under different frame number windows, forming an end-to-end mapping mechanism from visual detection to process command output.

[0024] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Example:

[0026] like Figure 1 , Figure 2 As shown, an online detection method for coating defects in micro lithium-ion battery electrodes based on image recognition is applicable to online quality monitoring scenarios in high-speed electrode production lines. This method simultaneously introduces two signals—image and process parameters—at the data acquisition end. At the feature extraction end, it fuses spatial, frequency, and statistical features. At the trajectory analysis end, it achieves temporal characterization of defects based on physical coordinate mapping and differential stability evaluation. At the process diagnosis end, it completes end-to-end mapping from visual features to process instructions through dynamic time warping.

[0027] During the data acquisition phase, the system synchronously acquires continuous image data of the electrode sheets at a preset sampling frequency (e.g., 5000 lines / second) through the hardware trigger interface between the line scan camera and the coating equipment controller. Simultaneously, it reads production time-series data containing electrode conveying speed values ​​in real time through the communication interface with the coating equipment. These speed values ​​are sampled by the encoder at millisecond-level resolution, in mm / s, and typically range from 50 to 500 mm / s, corresponding to different production speed levels. The data processing module uses the timestamps in the production time-series data as a common index to perform spatiotemporal binding on each frame of the continuous electrode sheet image data. It writes the electrode conveying speed value at the corresponding moment into the metadata area of ​​that frame, forming the image sequence to be tested. The metadata area adopts a compact structure with a fixed byte offset, and the speed field is stored as a single-precision floating-point number (4 bytes), ensuring that subsequent processing modules can directly read the data through the field offset without additional parsing steps. The spatiotemporal binding operation forcibly associates speed information with image content at the data layer, providing unambiguous speed input for subsequent inter-frame displacement compensation calculations.

[0028] During the feature extraction stage, the system performs a sliding window block operation on each frame of the image sequence under test. The window size is set to W×H pixels (e.g., 64×64 pixels), and the step size is set to S pixels (e.g., 32 pixels), thereby obtaining several local image blocks. There is a 50% overlap between adjacent blocks to avoid missing boundaries. For each local image block, the system performs three types of feature extraction simultaneously in parallel computing: spatial period offset, frequency offset, and grayscale variance.

[0029] The process of extracting the spatial periodic offset is as follows: For each local image patch, extract the gray-level mean sequence along the coating direction (column direction), and use this mean sequence as the gray-level periodic function g(x). The pre-stored standard texture periodic library contains corresponding standard periodic function templates for the current electrode model. (x), the template is generated offline from defect-free samples using an averaging method, representing the periodic baseline of normal coating texture. The system compares g(x) with... (x) Perform normalized cross-correlation calculation, take the offset corresponding to the peak position of the cross-correlation function as the original phase difference, and take the absolute value of this phase difference as the spatial period offset. The value ranges from [0, T / 2], where T is the basic period length of the standard periodic function. The positive or negative sign of the phase difference is stored as an auxiliary label for the period offset direction, which is used to subsequently determine the directional characteristics of the defect.

[0030] The frequency domain offset extraction process is as follows: A two-dimensional fast Fourier transform is performed on each local image patch to obtain the frequency domain matrix F(u,v). Within the frequency domain matrix, the frequency coordinates of the dominant frequency component are determined by detecting the peak positions (excluding the DC component) in the amplitude spectrum |F(u,v)|. Calculate the dominant frequency coordinates and the center coordinates of the frequency domain matrix. The Euclidean distance is used as the frequency domain offset. Defined as: Simultaneously record the azimuth angle of the main frequency coordinates relative to the center coordinates. This is stored as a label for the frequency domain offset direction. The main frequency coordinates of a normal coating texture are concentrated near the standard position, and the frequency domain offset is close to zero; when the coating layer exhibits periodic anomalies, the main frequency component drifts, and the frequency domain offset increases.

[0031] The process of extracting grayscale variance is as follows: directly calculate the sample variance of grayscale values ​​of all pixels within the local image patch, denoted as . Gray-level variance reflects the degree of dispersion of gray levels within a local image patch and has a strong response to defects such as particles and scratches that cause sudden changes in local gray levels.

[0032] After the three types of features are extracted, the system performs independent normalization on each feature value. The normalization benchmark is determined by the historical statistical values ​​of defect-free samples: the 99th percentile of the defect-free samples in the corresponding feature dimension is used as the upper bound of the normalization. Divide each eigenvalue by The values ​​are then truncated to the [0,1] interval to obtain the normalized eigenvalues. , , According to the preset weighting coefficient ( , , ,satisfy The three types of normalized feature values ​​are weighted and fused to generate the defect response value R of the local image patch.

[0033] ;

[0034] The defect response values ​​are arranged according to the spatial location of each local image patch in the original image to construct a defect response map corresponding to the image resolution. Each pixel position in the defect response map corresponds to a defect response value, and high response areas correspond to the locations of suspected defects in the image.

[0035] During the connected component extraction phase, the system performs a process on the defect response map based on a preset defect threshold. The binarization process will identify response values ​​exceeding a preset defect threshold. The region in question is labeled as foreground, and the remaining regions are labeled as background. A connected component labeling algorithm (4-connected or 8-connected) is applied to the foreground region to extract each connected component, and the centroid coordinates of each connected component are calculated. and the maximum defect response value in this area The centroid coordinates and defect response values ​​are encapsulated into a response tuple using a fixed data structure, defined as follows: The defect response value field is stored immediately after the centroid coordinate field to ensure that the field order of the structure is unique and immutable. Downstream modules access each field with a fixed byte offset.

[0036] During the inter-frame displacement compensation stage, the system reads the electrode transfer speed value v (unit: mm / s) from the current frame data area and combines it with the camera sampling inter-frame time interval. (Unit: s) and image resolution scaling factor k (unit: pixels / mm), calculate the displacement compensation of the pole pieces in the image coordinate system between adjacent frames. :

[0037] ;

[0038] displacement compensation amount Inter-frame correction is performed on the centroid coordinates in the response tuples, transforming the centroid coordinates of the nth frame to their physical location in the image coordinate system of the 1st frame, generating a physical coordinate sequence. For response tuples with the same physical location within N consecutive frames, they are associated based on their proximity within a preset spatial neighborhood (radius r pixels) to correspond to the defect response value. For data load, along the physical coordinate sequence within a preset frame window (The width of this window is pre-calibrated based on the ratio of the coating equipment's response time to the electrode conveying speed, ranging from 10 to 30 frames.) A variable-length response trajectory is constructed within this window. When no valid response tuple is extracted at the corresponding physical location in a frame, the trajectory length at that location does not increase. Therefore, the trajectory length may be less than the preset frame number window. Therefore, it is called a variable-length response trajectory. Within a preset frame count window... If a frame does not have a valid response tuple associated with it in its corresponding physical neighborhood, then a defect response value of 0 is filled into the trajectory for that frame, thus constructing a tuple of length [missing information]. A complete sequence of response values ​​at equal time intervals, where the position and magnitude of non-zero values ​​characterize the spatiotemporal intensity of the defect.

[0039] During the defect stability assessment phase, the system responds to variable-length response trajectories. Differential sequence of defect response values ​​between adjacent frames Perform the calculation: Based on the sign consistency ratio of the difference sequences. Determine the stability of the defect:

[0040] ;

[0041] in for The number of >0, for The number of <0. The value range is [0.5, 1.0]. The closer the value is to 1.0, the more consistent the direction of change in the defect response value and the more stable the defect. A value close to 0.5 indicates alternating positive and negative response values, representing occasional noise. A preset stability threshold is used. (For example, 0.75) is used as a criterion: when When, assign a stability label label=1 (persistent defect); when At this time, label=0 (occasional noise). The system will then display a variable-length response trajectory. The stability label is encapsulated into a defect feature vector, which serves as the input for subsequent dynamic time warping matching.

[0042] Before performing dynamic time warping matching, if the actual number of frames n of the variable-length response trajectory is less than a preset frame window... Furthermore, the trajectory completeness η is lower than the trajectory completeness threshold. If the trajectory is correct, then trajectory completion is performed first, and the completed trajectory is used as the trajectory to be matched; otherwise, the original variable-length response trajectory is used directly for matching.

[0043] During the process matching stage, the system will have a variable-length response trajectory. Compared with the standard trajectories in the pre-stored process library Perform Dynamic Time Warping (DTW) matching on each of the j=1,2,...,M pairs and calculate the DTW distance. ( , The DTW algorithm constructs... The cumulative cost matrix ( The length of the trajectory to be matched. (where j is the length of the j-th standard trajectory), using Sakoe-Chiba band constraints (bandwidth). Limit the search range for alignment paths and use dynamic programming based on the recursive formula. (in The point-to-point distance is given by D(0,0)=0, and the boundary condition is... (j>0)) Search for the optimal alignment path, and the final DTW distance is This overcomes the matching obstacle of inconsistent trajectory lengths under different frame number windows, thereby enabling direct comparison between variable-length and standard trajectories.

[0044] Select the process adjustment command code corresponding to the standard trajectory that minimizes the DTW distance. As output, the output format is a predefined instruction code string, which is parsed by the coating equipment controller and the corresponding process parameter adjustment actions are executed.

[0045] Through the above steps, this application realizes a complete online detection process from image acquisition, feature extraction, connected region localization, inter-frame trajectory construction, stability judgment to process command output. Using the electrode conveying speed value as a common basis for spatiotemporal binding and displacement compensation ensures precise correspondence between image features and physical coordinates. Parallel extraction and weighted fusion of spatial, frequency, and grayscale features enable the system to have differentiated responses to periodic weak defects and random noise. The introduction of the differential sequence sign consistency ratio clearly distinguishes between persistent defects and occasional noise. The application of dynamic time warping solves the matching obstacle caused by inconsistent trajectory lengths at different speed levels. These mechanisms together constitute a closed-loop system integrating online detection of coating defects and process diagnosis.

[0046] It should be noted that the electrode transfer speed value *v* used in the above calculations may contain minor measurement errors from the encoder. In practical applications, evaluations have shown that the deviation in single-frame displacement calculation caused by this error is typically less than 0.1 pixels. Since the system performs inter-frame correlation within a preset spatial neighborhood (radius *r* pixels, typically *r* ≥ 2) based on the physical coordinate sequence when subsequently constructing the variable-length response trajectory, this neighborhood matching mechanism effectively absorbs coordinate drift of this magnitude, ensuring that the same physical defect is correctly correlated in consecutive frames. Therefore, the inherent error of the encoder is within the system's tolerance range designed in this scheme and does not affect the implementation of detection and diagnostic functions. For scenarios with higher accuracy requirements, error calibration and filtering techniques can be used to preprocess the speed value *v* to further optimize the displacement compensation effect.

[0047] In another embodiment, the defect response value in the defect response map exceeds a preset defect threshold. After connecting the regions, a centroid coordinate correction step is also included to improve the positioning accuracy of the centroid coordinates.

[0048] The system uses a preset sliding step size (e.g., 4 pixels) Traverse the defect response map. Within a local window (e.g., 16×16 pixels) centered on the current traversal position, extract the response peak and its position coordinates within that window to generate a peak distribution mask. The peak distribution mask has the same size as the defect response map. The mask value is set to the corresponding response value at the location of the response peak, and zero at other locations. For each extracted connected region, using the non-zero mask points in the peak distribution mask that fall within that connected region as weights, calculate the weighted average of the coordinates of all mask points within that region to obtain the corrected centroid coordinates. , ):

[0049] ;

[0050] in The peak value of the response at position (i,j) in the mask image. , The summation is limited to the set of non-zero mask points within the current connected region, using the corrected centroid coordinates ( ). , Replace the original centroid coordinates and store them as the centroid coordinates in the response tuple.

[0051] When defects exhibit asymmetrical distribution or multi-peak superposition, traditional centroid coordinates based on grayscale centroids may deviate from the actual energy center of the defect, leading to accumulated errors in inter-frame displacement compensation. By introducing a peak distribution mask for weighted correction, the centroid coordinates shift towards the concentrated location of the response energy, making the physical location recorded in the response tuple more accurately correspond to the actual defect center. The corrected centroid coordinates are used for subsequent physical coordinate sequence construction, effectively reducing coordinate drift in inter-frame displacement compensation, improving the positional consistency of variable-length response trajectories, and thus enhancing the accuracy of subsequent dynamic time warping matching.

[0052] This application further proposes the following detailed explanation of the spatial periodic offset extraction method: For each local image block, the gray-level mean is extracted column by column along the polarimetric coating direction (column direction), resulting in a gray-level mean sequence of length W, which serves as the gray-level periodic function g(x). The pre-stored standard texture periodicity library contains corresponding standard periodic function templates for each type of electrode. (x) The template is generated by averaging multiple frames of defect-free historical samples of this model, representing the grayscale periodic reference of normal coating texture. The template length is consistent with the window width W to ensure dimensional alignment for phase difference calculation.

[0053] right and After zero-mean normalization, the normalized cross-correlation function is calculated:

[0054] ;

[0055] in, This is the offset. , , respectively , The mean, and They are respectively and The standard deviation of W is the window width.

[0056] Take the cross-correlation function peak position As the original phase difference, with | |As spatial period offset The value range is limited to [0, T / 2], where T is the basic period length of the standard periodic function, ensuring the uniqueness of the offset (avoiding ambiguity caused by period folding). Phase difference The sign (positive or negative) serves as the label for the period offset direction and the spatial period offset. They are stored together. A positive sign indicates that the current texture cycle is offset forward relative to the standard template, and a negative sign indicates that it is offset backward, providing defect direction information for subsequent process diagnosis.

[0057] Phase difference calculation based on normalized cross-correlation is insensitive to changes in overall image brightness and can effectively address grayscale drift caused by light source fluctuations. The offset range is limited by an upper bound of T / 2 to ensure the spatial periodic offset of all electrode types. All data are based on a unified normalization benchmark, which facilitates subsequent weighted fusion with frequency domain offset and grayscale variance in a dimensionally consistent manner, reducing the risk of weight imbalance caused by differences in the value range of a single feature. When the coating speed changes, the system can adaptively adjust the window size according to the speed value, or resample the grayscale periodic function to maintain period alignment with the standard periodic template.

[0058] This application further proposes the following detailed description of the frequency domain offset extraction method: For each local image block, before performing a two-dimensional fast Fourier transform, a two-dimensional Hanning window is applied to the image block to suppress spectral leakage, and then a two-dimensional FFT (Fast Fourier Transform) is performed to obtain the frequency domain matrix F(u,v), where the values ​​of u and v range from... Calculate the amplitude spectrum |F(u,v)|, set the DC component F(0,0) to zero, and then search for the location of the maximum value in the amplitude spectrum. The frequency coordinates are used as the coordinates of the dominant frequency component. The center coordinates of the frequency domain matrix are defined as follows: (Round up).

[0059] Calculate the Euclidean distance between the dominant frequency coordinates and the center coordinates as the frequency domain offset. Simultaneously calculate the azimuth angle of the dominant frequency coordinates relative to the center coordinates. Stored as a frequency domain offset direction label. It belongs to (-π, π], in radians. The normalization upper bound is half the length of the diagonal of the frequency domain matrix. ,Will Divide by Normalized frequency offset , Belongs to [0,1]. Normal coating textures have stable spatial frequency components, with the dominant frequency coordinates distributed within a fixed neighborhood of the standard center position. The frequency is approximately zero; when the coating layer exhibits localized thickness anomalies or changes in stripe spacing, the dominant frequency in the frequency domain shifts. This increases accordingly. The frequency domain offset captures frequency stability information that is difficult to quantify directly from spatial domain features, and is related to the spatial period offset. When they complement each other and respond simultaneously, the fused defect response value produces a superimposed enhancement effect, which helps to improve the detection sensitivity of systematic weak defects.

[0060] In another embodiment, a preset weighting coefficient ( , , The calibration method is as follows to ensure that the weighted fusion of the three types of features maintains a low response in defect-free regions and maximizes the response gain in defective regions. The first step is to collect a calibration dataset of defect-free samples. Electrode images confirmed as qualified products by quality inspection are selected from historical production records. Local image blocks are extracted using the same sliding window segmentation method as online processing. The normalized values ​​of the three types of features for each image block are calculated, the distribution of the three types of features on defect-free samples is statistically analyzed, and their respective variances are calculated. The initial weighting coefficients are normalized using the inverse of the variance.

[0061] ;

[0062] The inverse variance weighting assigns lower initial weights to features that fluctuate more on defect-free samples, thereby reducing the interference of background noise on the fused response value.

[0063] The second step involves collecting a calibration sample set of known defect types, including typical defect types such as missing coating, scratches, and particles, with at least 50 local image patches for each defect type. Starting with the initial weight coefficients, a grid search is performed within a range of ±0.15 for each weight coefficient, with a step size of 0.05. For each candidate weight combination, the Fisher discriminant ratio of the fusion response value between positive defect samples and negative samples (no defects) in the calibration sample set is calculated. :

[0064] ;

[0065] in , These are the mean values ​​of the fused response values ​​of the positive and negative samples, respectively. , This corresponds to the standard deviation. Select a value that makes... The largest weight combination ( The final preset weight coefficients are normalized and stored in the system parameter file. When the electrode model or manufacturing process changes, the above two-step calibration process is re-executed to update the weight coefficients for the corresponding model. Initialization based on the inverse of the variance of defect-free samples ensures the consistency of the normalized dimensions of each feature; the grid search combined with the fine-tuning process of Fisher's discriminant ratio makes the weight combination converge towards maximizing discrimination, reducing the risk of systematic false alarms or false negatives caused by empirical weight settings.

[0066] This application further proposes that after constructing the variable-length response trajectory, a trajectory validity verification step is performed to address the issue of missing trajectory frames caused by occlusion, sampling jitter, or response values ​​briefly falling below a defect threshold. The system calculates the variable-length response trajectory. The actual number of frames n and the preset number of frames window The ratio of the two values ​​yields the trajectory completeness. The value range is (0,1). If Below the preset trajectory integrity threshold (For example If the value is 0.6, it is determined that there are significant frame missing frames in the current trajectory, and a completion operation needs to be performed.

[0067] The completion process is as follows: Based on the physical timestamps corresponding to the existing n response tuples, identify the set of missing frame indices in the trajectory. For each missing frame index Read the electrode transmission speed value of the corresponding frame from the metadata area. and sampling time interval Reverse calculation of the centroid coordinate interpolation points of the missing frames:

[0068] ;

[0069] in, `k` is the index of the nearest valid frame before the missing frame, `k` is the image resolution scaling factor, and `sign(·)` is the sign function. Placeholder tuples with a defect response value of 0 are inserted at the corresponding positions of the interpolation centroid coordinates (zero-padding) to generate the completed variable-length response trajectory. Window length is the preset number of frames The zero-padding method maintains the integrity of the trajectory's timeline structure while explicitly marking missing frames with zero values. This ensures that subsequent difference sequence calculations and DTW matching can detect the location and number of missing frames, preventing them from being misclassified as normal regions with zero response. When trajectory integrity is below a threshold, directly using short trajectories for DTW matching may result in the DTW distances of the short trajectories becoming similar to those of multiple standard trajectories, reducing the matching's discriminative power. The completion operation uses velocity-driven coordinate interpolation and zero-padding to unify the trajectory to a preset frame count window. The frame length maintains the consistency of the trajectory length without introducing spurious response values, providing a complete timeline alignment basis for process library matching.

[0070] This application further proposes that the pre-stored process library be constructed as follows to establish a mapping relationship between defect trajectory characteristics and process adjustment operations. Defect samples with clear process adjustment records are collected from historical production records. Each sample includes: (1) a variable-length response trajectory during the defect occurrence period; (2) the corresponding actual process adjustment record confirmed by the process engineer. The process adjustment record is stored in the form of an instruction code string. The sample set size requires that the number of samples corresponding to each type of process adjustment instruction code is not less than 30.

[0071] Hierarchical clustering (agglomerated, using the Ward connection criterion) is performed on all collected variable-length response trajectory samples using dynamic time-warped distance as the similarity metric. During the clustering process, an inter-cluster distance threshold is set. (Normalized DTW distance), when the minimum distance between two clusters exceeds Merging stops when the contour coefficient is reached. (in The average distance from a sample to other samples of the same type. To evaluate the clustering quality of different numbers of clusters K (the minimum average distance from a sample to a sample from another class), the K value corresponding to the largest silhouette coefficient that is not less than 0.6 is selected as the final number of clusters. If the maximum silhouette coefficient is less than 0.6, the local constraint width of the DTW distance is adjusted (the Sakoe-Chiba band width is gradually adjusted from 10 to 30) and then the clustering is re-organized.

[0072] After hierarchical clustering is completed, for each cluster, the distribution of process adjustment command codes corresponding to its internal samples is statistically analyzed. If a command code accounts for more than a preset association threshold (e.g., 70%) in the cluster, the cluster is associated with that command code. If no single command code accounts for more than the threshold, the cluster is marked as a pending confirmation cluster, which is then manually reviewed by process engineers to supplement the association. For each cluster with associated command codes, for example, the DTW centroid trajectory of all variable-length response trajectories within the cluster is calculated using the Dynamic Time Warping Centroid Averaging (DTWBarycenter Averaging) algorithm as a standard trajectory and stored in the process library. The length is uniformly set to a preset frame window. The process library storage structure is {instruction code string, standard trajectory (preset frame window)}. The data is represented by a floating-point array (DTW distance variance within clusters). When the accumulated new samples exceed the preset update trigger number (e.g., 200), clustering and standard trajectory calculation are re-executed to achieve periodic self-updating of the process library. Hierarchical clustering based on DTW distance ensures that samples with similar defect evolution patterns are grouped into the same category; using the DTW centroid trajectory as the standard trajectory reduces the interference of individual abnormal samples on the matching benchmark; and the statistical correlation method with actual process adjustment records provides traceable data basis for the establishment of the process library.

[0073] Preferably, after generating the defect feature vector, a temporal stability reassessment step is included to capture the dynamic evolution trend of the defect over time. The system will then use the variable-length response trajectory. = The trajectories are divided into a first segment (first n / 2 frames) and a second segment (last n / 2 frames) according to their frame numbers. The difference sequences of the first segment and the second segment are calculated independently, and the stability ratio of the first segment is calculated according to the same sign consistency ratio formula as in claim 1. Ratio of stability to the later stage .

[0074] Calculate the difference between the stability ratio of the latter segment and the stability ratio of the former segment. . A value of >0 indicates that the consistency of the defect response direction increases in the later part of the trajectory, and the defect shows an aggravating trend. <0 indicates that the consistency of the later stage is reduced, and the defects show a trend of decay or dissipation; A value close to 0 indicates that the defect remains stable throughout the entire trajectory. The system will | |Compared to the preset trend change threshold (For example, 0.2) Compare: If If so, the stability label will be updated, specifically according to the following rule: and If so, update the label to label=2 (aggravating defect); if and If so, the label will be updated to label=3 (attenuation defect). At the same time, As a parameter of variation, it is added to the defect feature vector, expanding the feature vector structure to { ,label, The stability label is output as a confidence parameter for the process adjustment command code: the command codes corresponding to label=1 (continuously stable) and label=2 (aggravating) have high confidence and are recommended to be executed immediately; the command codes corresponding to label=3 (decaying) have medium confidence and can be executed after further observation; label=0 (occasional) has the lowest confidence and is recommended to be recorded but not executed for the time being. By independently evaluating the stability ratio and calculating the difference in segments of the variable-length response trajectory, the dynamic evolution direction of defects can be captured based solely on the temporal structure of the defect response values ​​without introducing additional sensors. This allows the downstream controller to adopt priority-differentiated process response strategies based on the differences in the dynamic behavior of defects.

[0075] In another embodiment, the metadata area of ​​the image sequence under test also stores the slurry flow rate value and coating pressure value. Before dynamically time-warping and matching the variable-length response trajectory with the standard trajectory in the pre-stored process library, a multi-parameter fusion step is also included. The slurry flow rate value Q and the coating pressure value P are written to the metadata area by the process parameter acquisition module of the coating equipment with timestamps synchronized with the image frames, and the storage format is consistent with the velocity value field (each occupying 4 bytes of single-precision floating-point numbers). Before performing DTW matching, the system starts from a preset frame number window. The metadata area of ​​each frame reads the slurry flow rate value Q(i) and the coating pressure value P(i) in sequence to form a sequence of process parameters aligned with the time axis of the variable-length response trajectory.

[0076] The slurry flow rate sequence and coating pressure sequence were respectively subjected to Min-Max normalization:

[0077] ;

[0078] ;

[0079] in , , , The normalized boundary values ​​of the device range set for the system. , Belongs to [0,1]. The normalized process parameter sequence is compared with the variable-length response trajectory. After aligning along the time axis, vectors are concatenated to generate a multidimensional fused trajectory. :

[0080] ;

[0081] Multidimensional fusion trajectory This is a three-dimensional time series, where each time step includes the image defect response value and the normalized values ​​of two process parameters. Alternative one-dimensional variable-length response trajectory DTW matching is performed with the corresponding multidimensional standard trajectory in the pre-stored process library. The point-to-point distance in each dimension of the DTW cost function is weighted Euclidean distance.

[0082] ;

[0083] in , , Distance weights for each dimension (satisfying) The default value is (0.6, 0.2, 0.2), which can be further optimized through grid search based on the matching accuracy on the calibrated sample set. Under the same image defect response trajectory, different combinations of process parameters may correspond to different process adjustment instructions. After introducing slurry flow rate and coating pressure values ​​to form a multi-dimensional fusion trajectory, the differences in process parameter dimensions form additional discriminative power in the DTW cost function, making the matching results sensitive to the combination of process parameters, thereby outputting more accurate process adjustment instruction codes and improving the directionality and operability of process diagnosis.

[0084] Preferably, after outputting the process adjustment command code, the system further includes a closed-loop feedback and sampling adaptive optimization step to verify the actual effect of the process adjustment and dynamically optimize the detection system parameters. The system sends the process adjustment command code C* to the coating equipment controller via the equipment control interface, whereby the controller parses and executes the corresponding process parameter adjustment action. After the command is sent, the system uses a preset feedback observation window... (For example, 5 seconds) Continuously acquire and process the sequence of images to be tested. Internally, construct a sequence of defect response values ​​after execution for the same physical location, and calculate the mean response. , and the average response before the instruction was executed By comparison, the rate of change of the defect response value is defined as:

[0085] ;

[0086] A value of >0 indicates that the defect response decreased after the instruction was executed, and the process adjustment was effective; This indicates that the defect response has not decreased or worsened, and the current instruction code is ineffective. Simultaneously, the process parameter sequence within the feedback observation window is read from the metadata area, and the process fluctuation residual value is calculated. (Defined as the sum of the standard deviations of the slurry flow rate series and the coating pressure series).

[0087] Based on the rate of change of defect response value The system performs adaptive updates to the process library based on matching weights, depending on the size of the data. Exceeding the preset improvement threshold (For example, 0.3), then the matching weight of the current standard trajectory will be... By increment (e.g., 0.05) Enhancement; if If the value is negative (indicating worsening defects), the matching weight of the current standard trajectory is reduced to no less than the lower limit. (For example, 0.01) to prevent the weight from dropping to zero and causing the matching candidate set to degenerate. The updated process library selects the optimal instruction code in subsequent DTW matching using a weighted nearest neighbor method, and the hit probability of the effective instruction code increases with the number of closed-loop iterations.

[0088] when When the value is negative, the system additionally considers the residual value due to process fluctuations. Calculate the sampling frequency compensation coefficient :

[0089] ;

[0090] in It serves as a reference benchmark for the residual value of process fluctuations during normal system operation (derived from historical steady-state data statistics). The adjustment coefficient ( (belonging to [0.5, 2.0]), its specific value can be determined by calibrating on historical data, based on the balance between the timeliness of process fluctuation tracking and the system's computational load when adjusting the sampling frequency; Used to control the sensitivity of the sampling frequency to process fluctuations. The larger the value, the greater the adjustment range of the sampling frequency with process fluctuations. The specific value is determined by offline calibration based on the dynamic response characteristics of the coating equipment.

[0091] The calibration method is as follows: During the period when the coating equipment is running continuously and stably without defect alarms, collect no less than 1000 frames of image data at the standard sampling frequency. Calculate the standard deviation of the slurry flow rate value sequence and the coating pressure value sequence during this period and sum them. Take the average value of multiple measurements (no less than 5 times) as the standard deviation. .

[0092] in accordance with Real-time adjustment of the sampling frequency of continuous image data from the electrode: New sampling frequency , For standard sampling frequency, The upper limit is constrained by the camera hardware line frequency limit. When When the fluctuations are large (significant process variations), >1, sampling frequency increased; when the process fluctuation residual value converges to the preset stable range. The sampling frequency is restored to 1.0, returning to the baseline value, completing one closed-loop adaptive optimization cycle. The closed-loop feedback mechanism directly feeds back the process adjustment results to the process library matching weights through the defect response value change rate, while the sampling frequency compensation mechanism adaptively adjusts the observation density based on the degree of process fluctuation. The two work together to ensure that the detection system maintains a high accuracy rate in process diagnosis and detection stability during long-term operation.

[0093] The following is a specific embodiment of an online detection method for coating defects in micro lithium-ion battery electrodes based on image recognition:

[0094] A micro lithium battery manufacturer deployed the detection method described in this invention on its electrode coating production line. The production line operates at an electrode conveying speed of v = 200 mm / s, and the standard sampling frequency of the line scan camera is [not specified]. =2000 lines / second, image resolution scaling factor k=10 pixels / mm, window size W×H=64×64 pixels, step size S=32 pixels, preset frame rate window =20 frames, preset defect threshold =0.6, stability threshold 0.75, trajectory integrity threshold =0.6, trend change threshold =0.2, improving the threshold =0.3, feedback observation window =5 seconds. The process library pre-stores four types of process adjustment instruction codes, corresponding to four multi-dimensional standard trajectories.

[0095] After the system starts, the line scan camera uses... =2000 lines / second sampling, inter-frame time interval =0.0005s, corresponding to inter-frame shift Pixels. The electrode transmission speed value v=200mm / s is synchronously written to the metadata area of ​​each frame by the encoder, completing the spatiotemporal binding and entering the feature extraction stage.

[0096] Three types of features are extracted in parallel for a local image patch at coordinates (320, 128): the peak position of the cross-correlation between the gray-level periodic function and the standard template normalization. =-2 pixels, spatial period offset Pixels, T=16 pixels, after normalization =2 / 8=0.25; Main frequency coordinates after 2D FFT =(18,36), center coordinates , , After normalization ≈0.32; Gray-scale variance Normalized upper bound =800, =0.39. Based on the calibrated weighting coefficient ( , , =(0.35,0.40,0.25), fusion response value R=0.35×0.25+0.40×0.32+0.25×0.39=0.313< =0.6, this block is considered normal.

[0097] Perform the same processing on the local image patch at coordinates (480, 96): =0.62, =0.71, =0.68, fusion response value R=0.35×0.62+0.40×0.71+0.25×0.68=0.671> =0.6, marked as a suspected defect area. After extraction of the connected components, the centroid coordinates are (480.3, 96.7), and the maximum response value is... The values ​​are encapsulated as a response tuple {480.3, 96.7, 0.71}. After weighted correction of the peak distribution mask image, the centroid coordinates are corrected to (480.1, 96.5) and stored in the response tuple.

[0098] In 20 consecutive frames ( Within 20 frames, the defect response value sequence for this physical location (located in the neighborhood of (480.1±2.0 pixels) after inter-frame displacement compensation and conversion to the coordinate system of the first frame) is: {0.63, 0.67, 0.71, 0.74, 0.72, 0.76, 0.75, 0.78, 0.70, 0.73, 0.75, 0.77, 0.74, 0.79, 0.76, 0.78, 0.80, 0.77, 0.79, 0.82}, for a total of 20 frames. The trajectory completeness is... =20 / 20=1.0, no padding needed. The difference sequence contains 19 elements. , , The basic stability label is 0 (occasional noise).

[0099] Execution timing stability reassessment: Differential sequence of the initial trajectory (first 10 frames) The trajectory of the latter part (last 10 frames) Difference The stability of the later stage exceeds the threshold, and the stability label is updated to label=2 (aggravating defect). =0.222 is added to the defect feature vector, and the output is { ,label=2, =0.222}.

[0100] Read the average slurry flow rate value within 20 frames from the metadata area. =85.3 mL / min, average coating pressure value =0.42MPa, range boundary =60、 =120 (mL / min) =0.2、 =0.8 (MPa), after normalization =(85.3-60) / 60=0.422, =(0.42-0.2) / 0.6=0.367. This is used to generate a three-dimensional multidimensional fused trajectory by splicing it with the variable-length response trajectory. Perform DTW matching with the four multidimensional standard trajectories in the process library, with each DTW distance being [missing information]. minimum distance The corresponding process adjustment instruction code C*="PRESS_UP_2" (coating pressure increased by 2 levels), confidence parameter label=2 (aggravated type), is output to the coating equipment controller for execution.

[0101] After the controller executes the PRESS_UP_2 command, it enters the feedback observation window. =5 seconds). Mean response at the same physical location =0.48, before execution =0.74, rate of change The process adjustment was effective. The matching weight for the second standard trajectory was increased from 0.25 to 0.30, and the weights of the remaining trajectories were normalized and then updated. Process fluctuation residual values. =0.031, below the upper bound of the stability interval. =0.05, =1.0, sampling frequency maintained =2000 lines / second remains unchanged, this closed-loop optimization cycle ends.

[0102] This embodiment further verifies the following: when the process fluctuation residual value (Exceeding the upper bound of the stable interval) When taking Calculate the sampling frequency compensation coefficient Considering the camera's maximum line rate limit (up to 6000 lines / second), in practical applications... The sampling frequency was adjusted from the baseline of 2000 lines / second to 5000 lines / second, the sampling interval was shortened to 0.0002 seconds, and the inter-frame shift was adjusted. Reducing the pixel size from 1.0 to 0.1 significantly improves the inter-frame continuity of defect trajectories. When Restored to 0.03 (below) ), Restored to 1.0, the sampling frequency returned to the baseline value.

[0103] This application achieves precise tracking of inter-frame physical coordinates without the need for additional positioning sensors, provided the electrode conveying speed is known. Parallel extraction and fusion of spatial, frequency, and grayscale features enhance the system's response to periodic minor defects, resulting in superior detection sensitivity compared to single-feature methods. Segmented stability assessment and variation amplitude parameters imbue process diagnostic conclusions with dynamic trend information. The combination of multi-dimensional fusion trajectory and DTW matching enables precise differentiation of the same image response feature under different process parameters. A closed-loop feedback mechanism allows the process library to converge towards highly effective instruction codes during continuous operation, reducing both the false negative and false positive rates of coating defects. This achieves an integrated closed-loop system for online detection and process diagnosis of coating defects, providing a stable and reliable technical means for process quality control in micro lithium-ion battery electrode production lines.

[0104] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An online detection method for coating defects in micro lithium battery electrodes based on image recognition, characterized in that, Includes the following steps: Synchronously acquire continuous image data of the electrode sheet and production time series data containing electrode sheet conveying speed values. Use the production time series data as an index to perform spatiotemporal binding on the continuous image data of the electrode sheet to generate a sequence of images to be tested, and embed the electrode sheet conveying speed values ​​into the metadata area of ​​the sequence of images to be tested. The image sequence to be tested is divided into local image blocks by sliding window. For each local image block, the spatial period offset, frequency offset and gray scale variance are extracted in parallel. After normalization, they are fused according to preset weight coefficients to generate the defect response value of the local image block. The defect response map is constructed by arranging them according to their original spatial positions. Extract connected regions in the defect response map whose defect response values ​​exceed a preset defect threshold, output their centroid coordinates and defect response values, and bind the centroid coordinates and defect response values ​​into structural response tuples; Read the electrode transmission speed value from the metadata area, use the electrode transmission speed value to perform inter-frame displacement compensation on the centroid coordinates in the response tuple to generate a physical coordinate sequence, and use the defect response value in the response tuple as the data load to construct a variable-length response trajectory along the physical coordinate sequence within a preset frame window. Calculate the difference sequence of defect response values ​​of adjacent frames in the variable-length response trajectory, determine the defect stability based on the sign consistency ratio of the difference sequence, and generate a defect feature vector containing the variable-length response trajectory and stability label; The variable-length response trajectory of the defect feature vector is dynamically time-warped and matched with the standard trajectory in the pre-stored process library, and the process adjustment instruction code with the smallest matching distance is output. The specific methods for extracting the spatial period offset include: Extract a grayscale periodic function for each local image block, and calculate the phase difference between the grayscale periodic function and the corresponding template in the pre-stored standard texture periodic library; The absolute value of the phase difference is used as the spatial period offset, and the positive or negative sign of the phase difference is stored as the period offset direction label. The specific methods for extracting the frequency domain offset include: Perform a Fourier transform on each local image patch to extract the frequency coordinates of the dominant frequency component; Calculate the Euclidean distance between the frequency coordinates and the center coordinates of the transformed frequency domain matrix, and use the Euclidean distance as the frequency domain offset; Simultaneously, the azimuth angle of the frequency coordinates relative to the center coordinates is recorded as a frequency domain offset direction label; The grayscale variance extraction process is as follows: directly calculate the sample variance of all pixel grayscale values ​​within the local image block.

2. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, After extracting the connected regions in the defect response map where the defect response value exceeds a preset defect threshold, the process further includes a centroid coordinate correction step, specifically including: The defect response map is traversed with a preset sliding step size to extract the response peak value within a local window and generate a peak distribution mask map. The centroid coordinates of the connected region are weighted and corrected using the peak distribution mask image, and the corrected centroid coordinates are used as the centroid coordinates in the response tuple.

3. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, The calibration method for the preset weight coefficients specifically includes: Collect local image patches of defect-free samples and calculate the statistical distribution of spatial period offset, frequency offset and gray-level variance of each sample. The initial weighting coefficients are obtained by normalizing the variances of each distribution using the inverse of the variance. A calibration sample of known defect types is collected, and the initial weight coefficients are fine-tuned through grid search to maximize the discriminative power of the fused defect response value on the calibration sample.

4. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, After constructing the variable-length response trajectory, a trajectory validity verification step is also performed, which specifically includes: The length of the variable-length response trajectory is calculated as the ratio of the length of the preset frame window to obtain the trajectory integrity. If the trajectory integrity is lower than the preset trajectory integrity threshold, the centroid coordinate interpolation point of the missing frame is calculated in reverse based on the electrode transmission speed value, and the defect response value is filled in with zero at the interpolation point to generate the completed variable-length response trajectory.

5. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, The construction method of the pre-stored process library specifically includes: Collect the variable-length response trajectories of historical defect samples and the corresponding actual process adjustment records; Hierarchical clustering of the collected variable-length response trajectories is performed using dynamic time-warped distance as a similarity metric. The variable-length response trajectories within the same cluster are associated with the same process adjustment command code, and the cluster center trajectory is stored as the standard trajectory corresponding to the command code.

6. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, After generating the defect feature vector, the process also includes a temporal stability reassessment step, specifically including: The variable-length response trajectory is divided into a first segment trajectory and a second segment trajectory according to the time sequence; The sign consistency ratio of the difference sequences of the first segment trajectory and the second segment trajectory are calculated separately to obtain the stability ratio of the first segment and the stability ratio of the second segment. Calculate the difference between the stability ratio of the latter segment and the stability ratio of the former segment. If the difference exceeds a preset trend change threshold, update the stability label and add the difference as a change amplitude parameter to the defect feature vector. The stability label is output as a confidence parameter of the process adjustment instruction code.

7. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, The metadata area of ​​the image sequence to be tested also stores slurry flow rate and coating pressure values; before dynamically time-normalizing and matching the variable-length response trajectory with the standard trajectory in the pre-stored process library, a multi-parameter fusion step is also included, specifically including: After normalizing the slurry flow rate and coating pressure values, they are spliced ​​with the variable-length response trajectory to generate a multi-dimensional fusion trajectory. The multidimensional fusion trajectory is used to replace the variable-length response trajectory, and dynamic time warping is performed to match it with the standard trajectory in the pre-stored process library.

8. The online detection method for coating defects of micro lithium battery electrodes based on image recognition according to claim 1, characterized in that, After outputting the process adjustment instruction code, the process also includes closed-loop feedback and sampling adaptive optimization steps, specifically including: The process adjustment instruction code is sent to the coating equipment controller, and the image sequence to be tested is continuously monitored to receive the status feedback signal after the coating equipment executes the instruction; Analyze the status feedback signal to extract the rate of change of defect response value and process fluctuation residual value after the execution instruction; If the rate of change of the defect response value exceeds the preset improvement threshold, the matching weight of the current standard trajectory in the pre-stored process library is increased. If the rate of change of the defect response value is negative, the matching weight of the current standard trajectory is reduced, and the sampling frequency compensation coefficient is calculated based on the process fluctuation residual value. The sampling frequency of the continuous image data of the electrode is adjusted in real time according to the sampling frequency compensation coefficient until the process fluctuation residual value converges to the preset stable range.

Images