Intelligent defect detection method and device for soft package lithium battery pole piece

Abstract

The invention provides an intelligent defect detection method and device for a soft package lithium battery pole piece, and relates to the technical field of lithium battery detection.The method comprises the steps that pole piece manufacturing process parameters and material microcosmic features are read, a backtracking data set is established, and defect probability prediction of the lithium battery pole piece is conducted according to the backtracking data set and a historical defect data set; a monitoring sensor is started; a multi-modal data set is established; space-time alignment of the multi-modal data set is executed; defect attention identification is carried out by using the prediction result and alignment feature representation, a defect attention identification result is established, and an intelligent defect detection result is established in combination with the prediction result. The technical problems of low detection accuracy and low reliability in the defect detection of the soft package lithium battery pole piece in the prior art are solved. The technical effects of improving the defect detection accuracy and reliability of the soft package lithium battery pole piece, reducing the false alarm rate and improving the overall quality control level of the lithium battery pole piece are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery detection, in particular to a defect intelligent detection method and device for soft-pack lithium battery pole piece. BACKGROUND

[0002] With the rapid development of new energy vehicles and electronic equipment, the demand for lithium batteries is increasing, among which soft-pack lithium batteries are widely used due to their high energy density and good safety. However, various defects such as scratches, cracks, uneven coating, and abnormal porosity on the surface of the soft-pack lithium battery pole piece may occur during production, which seriously affects the performance and safety of the battery. The existing defect detection method for soft-pack lithium battery pole piece mainly relies on machine vision and manual re-inspection mode. Machine vision detection and manual inspection have insufficient ability to identify internal defects of the pole piece, and it is difficult to find small defects inside the pole piece, which may lead to missed detection. In addition, manual re-inspection is easily affected by subjective factors, resulting in inconsistency and reliability problems of the detection results, thereby affecting the overall quality of the battery.

[0003] The existing technology has the technical problem of low detection accuracy and reliability in detecting defects of soft-pack lithium battery pole piece, which leads to insufficient quality control precision of lithium battery pole piece. SUMMARY

[0004] The purpose of the present application is to provide a defect intelligent detection method and device for soft-pack lithium battery pole piece, which solves the technical problem of low detection accuracy and reliability in detecting defects of soft-pack lithium battery pole piece in the prior art, which leads to insufficient quality control precision of lithium battery pole piece.

[0005] In view of the above problems, the present application provides a defect intelligent detection method and device for soft-pack lithium battery pole piece.

[0006] In a first aspect, the present application provides a defect intelligent detection method for soft-pack lithium battery pole piece, which comprises: scanning the unique identification code of the lithium battery pole piece, reading the pole piece manufacturing process parameters and material micro features according to the unique identification code, and establishing a backtracking data set; reading a historical defect data set, predicting the defect probability of the lithium battery pole piece according to the backtracking data set and the historical defect data set, and establishing a prediction result; starting a monitoring sensor at the time node when the lithium battery pole piece reaches the target detection area, the monitoring sensor comprising a visual sensor and an infrared sensor, and establishing a multi-modal data set; performing spatial and temporal alignment of the multi-modal data set, and establishing an aligned feature representation of the same physical area; performing defect attention identification using the prediction result and the aligned feature representation, establishing a defect attention identification result, and establishing a defect intelligent detection result according to the defect attention identification result and the prediction result.

[0007] Optionally, sensor calibration of visual images and infrared thermal images in the multi-modal data set is performed, a mapping relationship of each sensor coordinate system to a physical coordinate system of the production line is established, distortion correction of the multi-modal data set is performed, and standardized modal data in the same coordinate system is output. After the MARK points and texture features of the lithium battery pole piece are configured, the standardized modal data is projected to the same physical region using the MARK points and the texture features, and the spatially aligned modal feature map is output. In the same coordinate system, the pole piece conveying speed and the sampling time stamp of each sensor are read. Using a sliding time window and speed compensation, the different modal data frames in the modal feature map are corresponded to the same physical region based on the pole piece conveying speed and the sampling time stamp of each sensor, and the time-aligned multi-modal feature sequence is output. The multi-modal feature sequence is output as an aligned feature representation of the same physical region.

[0008] Optionally, the prediction result and the aligned feature representation are sent to a dynamic joint identification network. After receiving the prediction result, the priori layer of the dynamic joint identification network is called, the defect occurrence probability in the prediction result is projected to the physical coordinate system, and the priori defect risk map is established. After configuring the dynamic attention weight of the dynamic joint identification network using the priori defect risk map, attention feature identification of the aligned feature representation is performed, and the defect attention identification result is established. The defect attention identification result and the prediction result are jointly authenticated and fused to establish a joint defect confidence, and the joint defect confidence is output as a defect intelligent detection result.

[0009] Optionally, the identification source data of the defect attention identification result is obtained, the data quality of the identification source data is evaluated, and the source data trust degree is established. If the joint authentication trust degree in the joint authentication fusion process of the defect attention identification result and the prediction result is lower than a first preset threshold, and the corresponding source data trust degree is lower than a second preset threshold, an additional acquisition instruction is triggered. The attention area is configured according to the additional acquisition instruction, and additional data acquisition of the attention area is performed. The joint defect confidence is updated according to the additional data acquisition result.

[0010] Optionally, the backtracking data set is parsed, and a traceability parameter table is established. The manufacturing process parameters in the traceability parameter table include rolling pressure, coating speed, baking temperature, and roll pressure tension. The material micro features include particle distribution, coating uniformity, and porosity characteristics. The historical defect data set is parsed, time series defect probability prediction is performed according to the historical defect data set, and a first prediction result is established. Causal association analysis is performed based on the historical defect data set and the traceability parameter table, and a second prediction result is established. The first prediction result and the second prediction result are fused to establish the prediction result.

[0011] Optionally, the defect intelligent detection result is used for defect classification early warning level matching to establish a classification early warning signal; after the classification early warning signal is used for configuration and transmission of the conveyor belt shunting parameters, automatic lithium battery pole piece shunting and transmission management is performed.

[0012] Optionally, the monitoring sensor further comprises an ultrasonic sensor; the ultrasonic sensor is used for ultrasonic monitoring of the lithium battery pole piece to establish ultrasonic feedback data; the ultrasonic feedback data is used for trust authentication of the defect intelligent detection result; and the defect intelligent detection result is updated.

[0013] Optionally, the defect intelligent detection result is used for construction of a device linkage feedback of a production device; the device linkage feedback is recorded in a compensation space of the production device; when the device linkage feedback energy in the compensation space is higher than a preset energy threshold, device control compensation of the production device is performed.

[0014] Optionally, if the defect confidence in the defect intelligent detection result is lower than a confidence threshold, a retest instruction is generated; the lithium battery pole piece is backtracked and transmitted according to the retest instruction, and is marked as a retest workpiece.

[0015] In a second aspect of the present application, a defect intelligent detection device for soft package lithium battery pole pieces is provided, which comprises: a data obtaining module, which is used for scanning a unique identification code of a lithium battery pole piece, reading pole piece manufacturing process parameters and material microscopic characteristics according to the unique identification code, and establishing a backtracking data set; a defect prediction module, which is used for reading a historical defect data set, performing defect probability prediction of the lithium battery pole piece according to the backtracking data set and the historical defect data set, and establishing a prediction result; a sensor starting module, which is used for starting a monitoring sensor at a time node when the lithium battery pole piece reaches a target detection area, the monitoring sensor comprising a visual sensor and an infrared sensor, and establishing a multi-modal data set; a data alignment module, which is used for performing spatial and temporal alignment of the multi-modal data set, and establishing an aligned feature representation of a same physical area; a defect detection result establishing module, which is used for performing defect attention identification using the prediction result and the aligned feature representation, establishing a defect attention identification result, and establishing a defect intelligent detection result according to the defect attention identification result and the prediction result.

[0016] One or more technical solutions provided in the present application have at least the following technical effects or advantages: The method provided by the embodiment of the application comprises the following steps: scanning a unique identification code of a lithium battery pole piece, reading pole piece manufacturing process parameters and material microscopic characteristics according to the unique identification code, establishing a traceability data set, reading a historical defect data set, performing defect probability prediction of the lithium battery pole piece according to the traceability data set and the historical defect data set, establishing a prediction result, starting a monitoring sensor at a time node when the lithium battery pole piece reaches a target detection area, the monitoring sensor comprising a visual sensor and an infrared sensor, establishing a multi-modal data set, performing spatial and temporal alignment of the multi-modal data set, establishing aligned feature representation of a same physical area, performing defect attention identification by using the prediction result and the aligned feature representation, establishing a defect attention identification result, and establishing a defect intelligent detection result according to the defect attention identification result and the prediction result. The technical effect of improving defect detection accuracy and reliability of the soft package lithium battery pole piece, reducing false positive rate and improving overall quality control level of the lithium battery pole piece is achieved.

[0017] The above description is only a summary of the technical solutions of the application. In order to enable one skilled in the art to better understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to enable the above and other purposes, characteristics and advantages of the application to be more obvious and easy to understand, the following specific embodiments of the application are described. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the application, nor is it intended to limit the scope of the application. Other features of the application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only exemplary, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0019] Figure 1 A flowchart of a defect intelligent detection method for soft package lithium battery pole pieces provided by the application.

[0020] Figure 2 A structural diagram of a defect intelligent detection device for soft package lithium battery pole pieces provided by the application.

[0021] Legend of the drawings: data obtaining module 11, defect prediction module 12, sensor starting module 13, data alignment module 14, defect detection result establishing module 15. DETAILED DESCRIPTION

[0022] This application provides an intelligent defect detection method and apparatus for pouch lithium battery electrodes, addressing the technical problem of low accuracy and reliability in existing pouch lithium battery electrode defect detection technologies, which leads to insufficient quality control precision of lithium battery electrodes. It achieves the technical effect of improving the accuracy and reliability of pouch lithium battery electrode defect detection, reducing false alarm rates, and improving the overall quality control level of lithium battery electrodes.

[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0024] Example 1, as Figure 1 As shown, this application provides a method for intelligent defect detection of electrode sheets in pouch lithium batteries. The method includes: Scan the unique identification code of the lithium battery electrode, read the electrode manufacturing process parameters and material microstructure based on the unique identification code, and establish a retrospective dataset.

[0025] Specifically, when the lithium battery pole piece flows to the starting point of the target detection area, the unique identification code on the surface of the battery pole piece is automatically positioned and scanned by a high-resolution industrial code reading device, such as a CCD code reader or a laser scanning gun. For example, the unique identification code is a two-dimensional code or a DPM code engraved by laser or inkjet, which has high anti-pollution and wear-resistant characteristics, ensuring that it can be quickly and accurately read throughout the production process. After reading the unique identification code, a data retrieval instruction is triggered, which connects to the production database that stores the entire process data of lithium battery pole piece raw material procurement and manufacturing. By using the unique identification code as a retrieval keyword, the manufacturing process parameters and material micro features of the lithium battery pole piece are accurately extracted from the production database, and a backtracking data set is formed according to the obtained manufacturing process parameters and material micro features, wherein the manufacturing process parameters include rolling pressure, coating speed, baking temperature, rolling tension, etc., which are related to the physical properties and quality stability of the lithium battery pole piece. For example, if the rolling pressure is too large or too small, it will affect the thickness uniformity and internal structure of the pole piece. Material micro features include particle distribution, coating uniformity, porosity characteristics, etc., which are used to reflect the quality and micro changes in the processing of the lithium battery pole piece. For example, uneven particle distribution can cause stress concentration in the pole piece, leading to defects, and poor coating uniformity can affect the performance consistency of the battery, and porosity characteristics affect the charge and discharge efficiency of the battery. By scanning the unique identification code and establishing the backtracking data set of the lithium battery pole piece, accurate and reliable raw data are provided for lithium battery pole piece defect detection, thereby improving the accuracy and effectiveness of defect detection.

[0026] The historical defect data set is read, and the defect probability of the lithium battery pole piece is predicted based on the backtracking data set and the historical defect data set to establish a prediction result.

[0027] Specifically, the historical defect data set of the soft-pack lithium battery pole piece is read from the enterprise-level quality management system through a data interface, and the historical defect data set is a time series stored database containing various defect records of historical lithium battery pole pieces in the production process, including defect type, occurrence process, manufacturing process parameter, material micro feature, defect position, severity, and corresponding pole piece identification code information. Then, based on the backtracking data set and the historical defect data set, the backtracking data set of the current lithium battery pole piece is taken as an input feature, which is compared and learned with the historical defect data set to realize the defect probability prediction of the lithium battery pole piece and obtain a prediction result.

[0028] Further, the defect probability prediction of the lithium battery pole piece according to the traceability data set and the historical defect data set, and the establishment of the prediction result, include: analyzing the traceability data set, establishing a traceability parameter table, the manufacturing process parameters in the traceability parameter table include rolling pressure, coating speed, baking temperature, and rolling tension, and the material micro features include particle distribution, coating uniformity, and porosity characteristics; analyzing the historical defect data set, performing time series defect probability prediction according to the historical defect data set, and establishing a first prediction result; performing causal correlation analysis based on the historical defect data set and the traceability parameter table, and establishing a second prediction result; and fusing the first prediction result and the second prediction result to establish the prediction result.

[0029] Specifically, the traceability data set is analyzed by an ETL tool to extract current lithium battery pole piece key fields, including manufacturing process parameters and material micro features, the manufacturing process parameters in the traceability parameter table include rolling pressure, coating speed, baking temperature, and rolling tension, and the material micro features include particle distribution, coating uniformity, and porosity characteristics. Then, the extracted manufacturing process parameters and material micro features are subjected to data cleaning and standardization processing, and the process parameters and micro feature parameters are associated and stored with the pole piece unique identification code as an index to establish a traceability parameter table. The historical defect data set is obtained and analyzed, and for the historical defect data set, a time series analysis technique such as an autoregressive integrated moving average model is used to perform time series defect probability prediction. For example, the historical defect data set is analyzed, the historical defect records are sorted by production time stamp, and feature words such as defect type, occurrence process, defect location, and severity are extracted, and the autoregressive integrated moving average model is used to capture the periodicity of defect occurrence, such as defect fluctuation caused by equipment start and stop at daily shift handover, to generate a first prediction result, the first prediction result is a defect probability distribution graph based on time trend, reflecting high-incidence defect types and their probability values at different time periods.

[0030] Based on the historical defect data set and the traceability parameter table, a causal correlation analysis is performed, which refers to analyzing the causal relationship between the historical defect data and the manufacturing process parameters and the material microscopic characteristics, for identifying which manufacturing process parameters and material microscopic characteristics have a significant impact on the occurrence of the pole piece defect. For example, the historical defect data set and the data in the traceability parameter table are obtained, and then the SHAP value algorithm is used to analyze the contribution of multiple traceability parameters to the occurrence of the battery pole piece defect. Through calculation, it is found that when the rolling pressure exceeds 80 MPa, the occurrence rate of the edge crack of the pole piece is significantly increased, indicating that the rolling pressure is an important risk factor for the occurrence of the edge crack. Similarly, when the porosity is less than 25%, the risk of coating exposure increases, indicating that the porosity is a key influencing factor for the coating exposure. Based on the analysis results, a causal correlation is established between each parameter in the traceability parameter table and the defect occurrence phenomenon. Further, based on the causal correlation analysis, a probability prediction of the occurrence of the lithium battery pole piece defect is performed, and a second prediction result is obtained, which reflects the specific defect and its probability caused by the abnormality of the specific parameters of the current pole piece. Finally, a fusion strategy such as weighted average and voting mechanism is used to integrate the first and second prediction results to obtain the prediction result. For example, the accuracy of the time series prediction and the causal analysis prediction is evaluated by calculating the AUC value through cross-validation, and dynamic weights are assigned to the first and second prediction results. If the AUC value of the time series prediction is low, indicating that its prediction accuracy is relatively low, a lower weight such as 0.4 can be given, and if the AUC value of the causal analysis is high, indicating that the prediction accuracy is good, a higher weight such as 0.6 can be given. According to the weights, the two prediction results are weighted and fused, and finally a comprehensive prediction result is obtained, so as to more accurately predict the probability of the occurrence of the lithium battery pole piece defect, and improve the defect detection accuracy and reliability of the soft package lithium battery pole piece.

[0031] The monitoring sensor is started at the time node when the lithium battery pole piece reaches the target detection area, and the monitoring sensor includes a visual sensor and an infrared sensor, and a multi-modal data set is established.

[0032] Specifically, when the lithium battery pole piece flows to the starting point of the target detection area, the time node is accurately captured by the laser positioning or photoelectric sensor installed on the production line, and the monitoring sensor is automatically started. In order to ensure the accuracy and reliability of data acquisition, high-precision synchronization technology is adopted, based on accurate clock synchronization mechanism, to ensure that the visual sensor and infrared sensor are started synchronously at the moment when the pole piece reaches the target detection area. The monitoring sensor includes a visual sensor and an infrared sensor. The visual sensor, such as a high-resolution industrial camera or a CMOS line array camera, is used to collect the appearance image of the pole piece, providing detailed visual information of the pole piece surface, including scratches, cracks, edge burrs, color changes, etc. At the same time, the infrared sensor collects infrared radiation data of the pole piece at the same time and the same physical position by measuring the change of the physical parameters of the pole piece. The infrared radiation data is converted into a temperature distribution map, reflecting the changes in the thermal characteristics of the pole piece surface caused by coating thickness, porosity, residual solvent, etc. The visual sensor data and infrared sensor data are integrated to obtain a multi-modal data set, which includes visual image data and infrared thermal image data, providing information about the pole piece from different angles. The visual image focuses on surface features, and the infrared thermal image focuses on internal thermal distribution. By combining the two kinds of data, defects can be more accurately identified and located, and the accuracy and reliability of detection can be improved.

[0033] The spatial and temporal alignment of the multi-modal data set is performed to establish an aligned feature representation of the same physical area.

[0034] Further, the spatial and temporal alignment of the multi-modal data set is performed to establish an aligned feature representation of the same physical area, including: performing sensor calibration of the visual image and infrared thermal image in the multi-modal data set, establishing a mapping relationship from each sensor coordinate system to the physical coordinate system of the production line, and performing distortion correction of the multi-modal data set to output standardized modal data in the same coordinate system; after configuring the MARK point and texture features of the lithium battery pole piece, projecting the standardized modal data to the same physical area using the MARK point and the texture features, and outputting the modal feature map after spatial alignment; in the same coordinate system, reading the pole piece conveying speed and the sampling time stamp of each sensor; using a sliding time window and speed compensation, based on the pole piece conveying speed and the sampling time stamp of each sensor, corresponding different modal data frames in the modal feature map to the same physical area, outputting the multi-modal feature sequence after time alignment, and outputting the multi-modal feature sequence as the aligned feature representation of the same physical area.

[0035] Specifically, in the high-speed continuous production scene of soft-pack lithium battery pole pieces, due to the high-speed movement of the pole pieces on the conveyor belt, the time-space data mismatch problem will occur due to the difference in sampling frequency between the visual sensor and the infrared sensor. For example, the visual sensor industrial camera collects the surface data of the pole piece at a high frequency of 2 ms / frame, and the infrared sensor is limited by the thermal response time to collect temperature distribution at a frequency of 10 ms, resulting in different time points for the same physical area on the pole piece in the sensor data. If the collected data is directly compared, the time-space dislocation of defect geometric features and thermal features will occur, that is, the visual camera captures the defect at point A, while the infrared camera corresponds to the thermal image at point B, which is not the same physical area, reducing the accuracy of defect diagnosis. Therefore, in the detection process of lithium battery pole pieces, in order to ensure that the multi-modal data accurately reflects the true state of the pole piece, the spatial and temporal alignment of the multi-modal data set is performed. First, through sensor calibration technology, the visual images and infrared thermal images in the multi-modal data set are calibrated, the mapping relationship of each sensor coordinate system to the production line physical coordinate system is established, and the original data in the multi-modal data set is corrected by using the calibration parameters. Distortion correction refers to processing the collected image or thermal image data to remove distortion caused by the sensor's own characteristics or external environmental factors. For example, the distortion correction of the visual image can use image processing algorithms to correct lens distortion, so that the straight lines in the image remain straight after processing, thereby improving the accuracy of the data. After calibration and distortion correction, the standardized modal data in the same coordinate system is output. The MARK points and texture features of the lithium battery pole piece are pre-configured at the edge or specific position of the pole piece. The MARK point is a specific marker point on the pole piece, which is used to establish a spatial reference between different modal data, and the texture feature refers to the texture information on the surface of the pole piece. By matching the MARK points and texture features, the corresponding points in the visual image and infrared thermal image can be accurately aligned by using image registration technology, and then the standardized modal data is projected to the same physical area, outputting the spatially aligned modal feature map, ensuring the spatial consistency of different modal data.

[0036] On the basis of spatial alignment, further time alignment is performed to solve the space-time misalignment problem of multi-modal data caused by differences in sampling frequency. In the same production line coordinate system, the conveying speed of the pole piece and the sampling time stamp of each sensor are synchronously read, the sampling time stamp refers to the specific time point at which the sensor collects data, and the conveying speed refers to the speed at which the pole piece moves on the production line, which is obtained through encoder feedback. The time stamp and the conveying speed are coordinate-converted, and the physical position of each data frame is calculated through the formula position = speed x (current time stamp-reference time stamp). For example, the reference point is the coordinate origin, and when the visual frame is collected at T1 = 10:00:00.002, the conveying speed is 3 m / s, and the frame corresponds to the pole piece coordinates (0, 3x0.002) = (0, 6 mm). In view of the difference in sampling frequency between the visual sensor and the infrared sensor, a joint strategy of sliding time window and speed compensation technology is used, based on the conveying speed of the pole piece and the sampling time stamp of each sensor, the different modal data frames in the modal feature map are corresponded to the same physical region. The sliding time window is a time series analysis technology, which processes the data by segmenting through a fixed length time window on the time axis. The speed compensation is to adjust the time alignment of different modal data according to the conveying speed of the pole piece, to ensure the consistency of the data in time. For example, set the time window length to 20 ms, slide the window on the time axis based on the visual frame, calculate the pole piece position corresponding to each time point in the window through speed compensation, and generate temperature frames matched with the physical position of the visual frame in the window through cubic spline interpolation on the infrared data. For the missing data area, such as the blank of infrared data caused by high-frequency sampling of vision, the missing data area is filled through dynamic adjustment and interpolation of the sliding time window, and the time-aligned multi-modal feature sequence is output, and the multi-modal feature sequence is output as the aligned feature representation of the same physical region.

[0037] Through space-time alignment, false positives and false negatives caused by different sensor perspectives and sampling times are eliminated, thereby greatly reducing the probability of false positives and false negatives, and making the defect detection result of the soft pack lithium battery pole piece more accurate and reliable.

[0038] The prediction result and the aligned feature representation are used for defect attention identification, a defect attention identification result is established, and a defect intelligent detection result is established according to the defect attention identification result and the prediction result.

[0039] Further, the defect attention identification result is established by sending the prediction result and the aligned feature representation to a dynamic joint identification network, calling a priori layer of the dynamic joint identification network, projecting defect occurrence probability in the prediction result to a physical coordinate system after receiving the prediction result to establish a priori defect risk map, performing attention feature identification of the aligned feature representation after configuring a dynamic attention weight of the dynamic joint identification network by using the priori defect risk map to establish a defect attention identification result, and performing joint authentication fusion of the defect attention identification result and the prediction result to establish a joint defect confidence degree, and outputting the joint defect confidence degree as a defect intelligent detection result.

[0040] Specifically, the dynamic joint identification network architecture comprises a priori layer, a feature extraction layer, a dynamic weight configuration layer and a fusion layer. The priori layer receives the prediction result. The feature extraction layer analyzes the aligned multi-modal data by using a deep learning algorithm such as a convolutional neural network. The dynamic weight configuration layer performs weight distribution according to the priori layer result. The fusion layer combines the defect attention identification result and the prediction result, calculates a joint defect confidence degree by a fusion algorithm, and outputs the joint defect confidence degree as the final defect detection result. The whole network is trained end-to-end to constantly optimize the parameters of each layer to improve the accuracy and reliability of defect detection.

[0041] The prediction result and the aligned feature representation are input into the dynamic joint identification network. After receiving the prediction result, the prior layer in the dynamic joint identification network projects the defect occurrence probability in the prediction result to the physical coordinate system through the spatial transformation network to generate a prior defect risk map. The prior defect risk map is a two-dimensional or three-dimensional mapping diagram showing the defect occurrence probability at each position on the pole piece. For example, if the prediction result shows that a certain area has a high defect probability, this area will be marked as a high-risk area on the prior defect risk map. Using the prior defect risk map generated during production, the dynamic joint identification network configures a dynamic attention weight. The dynamic attention weight is dynamically adjusted according to the probability distribution in the prior defect risk map, so that the dynamic joint identification network analyzes the aligned multi-modal data when processing the aligned feature representation, identifies possible defect features, and obtains a defect attention identification result. Guided by the dynamic attention weight, the dynamic joint identification network can more efficiently identify defects in high-risk areas, improving the accuracy and efficiency of detection. Finally, the defect attention identification result and the prediction result are fused by a fusion algorithm such as weighted averaging or Bayesian fusion to generate an accurate and reliable joint defect confidence, which comprehensively reflects the probability value of the existence of defects, and the joint defect confidence is output as the defect intelligent detection result. By using the prediction result and the aligned feature representation for defect attention identification, defects on the pole piece can be more accurately identified and located, not only improving the accuracy of soft-pack lithium battery pole piece defect detection, but also improving the intelligence and efficiency of soft-pack lithium battery pole piece detection through the configuration of the dynamic attention weight.

[0042] Further, the joint defect confidence is established by fusing the defect attention identification result and the prediction result, including: obtaining identification source data of the defect attention identification result, performing data quality evaluation on the identification source data, and establishing source data trust degree; if the joint authentication trust degree in the process of fusing the defect attention identification result and the prediction result is lower than a first preset threshold, and the corresponding source data trust degree is lower than a second preset threshold, an additional acquisition instruction is triggered; a focus area is configured according to the additional acquisition instruction, and additional data acquisition of the focus area is performed, and the joint defect confidence is updated according to the additional data acquisition result.

[0043] Specifically, the recognition source data used to generate the defect attention recognition result is obtained, and the recognition source data is original feature information extracted from the multi-modal feature sequence, including visual image, infrared thermal image and other data. The data quality of the recognition source data is evaluated, and the evaluation factors include image sharpness, contrast, noise level, and signal-to-noise ratio of infrared thermal image, etc. The source data trust degree is calculated to measure the credibility of the sensor data itself. For example, for visual images, the Laplace operator is used to calculate the image sharpness, and the peak-to-valley difference of the gray level histogram is used to quantify the contrast, which reflects the distinction between the defect and the background. At the same time, the local variance is used to analyze the noise level. For infrared thermal images, the signal-to-noise ratio of the thermal map is calculated, and the signal-to-noise ratio = thermal signal mean value / noise standard deviation, wherein the thermal signal is the temperature mean value of the defect area, and the noise is estimated by the temperature fluctuation of the defect-free area. Finally, the indicators are normalized and weighted summed, and the weights are determined by the entropy weight method to obtain the source data trust degree. If the joint authentication trust degree in the joint authentication and fusion process of the defect attention recognition result and the prediction result is lower than the first preset threshold, and the corresponding source data trust degree is also lower than the second preset threshold, it indicates that the current detection result has a large uncertainty and cannot make a high confidence decision. At this time, the additional acquisition instruction is triggered. The additional acquisition instruction reconfigures the sensor parameters according to the attention area indicated in the prediction and recognition results, such as adjusting the exposure time of the visual camera and the gain of the infrared thermal imager, and performs additional data acquisition on the attention area when the lithium battery pole piece passes through the next available detection station. The attention area refers to a specific area on the lithium battery pole piece that needs additional data acquisition and detailed detection according to the additional acquisition instruction, which is based on the place that shows a higher defect risk or has data uncertainty in the preliminary detection result. Finally, the high-quality additional data acquisition result obtained this time is re-injected into the fusion decision-making process to update and output a higher joint defect confidence as the final defect intelligent detection result. For example, if the preliminary detection result shows that a certain area may have defects, but the joint authentication trust degree is low and the source data trust degree is low, more detailed information is obtained through additional data acquisition to more accurately determine whether the area has defects. Through the dynamic adjustment and verification mechanism, the electrode defect detection can adapt to different production environments and pole piece states, improving the flexibility and accuracy of defect detection. In addition, by updating the joint defect confidence, a more reliable and accurate detection result is provided, which guarantees the output stability and accuracy of the entire intelligent detection process from the root, greatly reducing the risk of misjudgment.

[0044] Further, after the defect intelligent detection result is established according to the defect attention recognition result and the prediction result, the method further comprises: utilizing the defect intelligent detection result to perform defect classification and early warning level matching to establish a classification early warning signal; and after configuring the classification early warning signal to transfer the belt, automatically performing the split conveying management of the lithium battery pole piece.

[0045] Specifically, the defect types, positions and severity in the defect intelligent detection result are analyzed according to the preset defect knowledge base and process standard, and the defect classification warning level is matched to establish a classification warning signal, which is used to indicate the severity and processing priority of the defects on the pole piece. For example, the warning levels are divided into three categories: serious determination first-level warning, general defect second-level warning and qualified product third-level warning. After obtaining the classification warning signal by matching, the classification warning signal is transmitted to the programmable logic controller (PLC), and the PLC configures the conveyor belt shunting parameters using the classification warning signal to automatically calculate the action timing and target channel of the shunting execution mechanism. For example, the first-level warning corresponds to the waste box channel, the second-level warning corresponds to the re-inspection buffer channel, and the qualified product corresponds to the main channel flowing into the next process. Through driving the shunting baffle, push rod or cross belt execution mechanism, the automatic shunting and conveying management of the lithium battery pole piece is realized, and the rapid and accurate physical separation of the pole pieces with different quality states is realized.

[0046] Through automatic execution of the shunting and conveying management, the fine classification of the defective battery pole pieces is realized, and the entire battery pole piece quality sorting process does not require manual intervention, which greatly improves the automation degree and production rhythm of the production line, reduces the inconsistency of shunting caused by factors such as manual fatigue and subjective judgment, and ensures the stability of product quality and production efficiency.

[0047] Further, the monitoring sensor further includes an ultrasonic sensor, which is used to perform ultrasonic monitoring of the lithium battery pole piece, establish ultrasonic feedback data, and use the ultrasonic feedback data to perform trust authentication on the defect intelligent detection result, and update the defect intelligent detection result.

[0048] Specifically, in the detection process of lithium battery pole pieces, in addition to visual sensors and infrared sensors, monitoring sensors also include ultrasonic sensors, which are installed at appropriate positions on the production line. When lithium battery pole pieces pass through, ultrasonic sensors are automatically started and perform ultrasonic monitoring at the same time. By emitting high-frequency ultrasonic pulses, the pulses propagate and reflect inside the pole pieces. By analyzing the echo amplitude, time of flight and frequency change of the reflected waves, ultrasonic feedback data is generated, which reflects the internal structural characteristics of the pole pieces, such as the thickness of the delamination defect and the size of the air bubble. Then the collected ultrasonic feedback data is preprocessed, including signal filtering, feature extraction and data normalization, etc. to remove noise and extract useful feature information, ensuring the quality and consistency of the data. And use ultrasonic feedback data to trust authentication of defect intelligent detection results, trust authentication refers to cross-validation of comparative ultrasonic feedback data and defect intelligent detection results to verify the reliability and accuracy of the detection results. By establishing a defect feature correlation matrix, which is a pre-defined knowledge base, a mapping relationship between defects and ultrasonic feedback data is established, such as coating peeling defect corresponding to ultrasonic echo amplitude enhancement. The ultrasonic feedback data of ultrasonic detection and the defect intelligent detection results are spatio-temporally aligned, and consistent comparison is made according to the feature correlation matrix. If the ultrasonic feedback data and the defect intelligent detection results are consistent, it means that the detection result is reliable, if there is a difference, it means that the defect anomaly may be a false defect, such as dust, water stains and other interference, or the ultrasonic feedback data is affected by interference or noise. In this case, its confidence is lowered, and the difference area is marked to trigger the re-inspection process or directly marked as no defect. Finally, through fusion algorithms such as D-S evidence theory, Bayesian reasoning or weighted decision, the ultrasonic feedback data and the existing detection results are comprehensively analyzed to generate more accurate defect intelligent detection results.

[0049] By introducing ultrasonic sensors, internal structural defects of the pole pieces can be detected, enhancing the comprehensiveness and accuracy of battery pole piece defect detection. Moreover, through the fusion and verification mechanism of multi-modal data, not only the credibility of the detection results is improved, but also the possibility of false positives and missed detections is reduced, thereby improving the overall quality control level of lithium battery pole pieces.

[0050] Further, after establishing the defect intelligent detection result according to the defect attention recognition result and the prediction result, the method further includes: constructing a device linkage feedback of a production device according to the defect intelligent detection result; recording the device linkage feedback in a compensation space of the production device; and when the device linkage feedback energy in the compensation space is higher than a preset energy threshold, performing device control compensation of the production device.

[0051] Specifically, after generating the defect intelligent detection result, the process root of the defect is traced back according to the type, position and severity of the defect, and a device linkage feedback of the production equipment is constructed, the device linkage feedback refers to feeding back the detected defect information to the production equipment, so that the production equipment can adjust its operation parameters according to the defect information. For example, if the detection result shows that there are many crack defects in the pole piece of a batch, it is analyzed that the rolling pressure of the production equipment is too high. At this time, a feedback signal is generated, indicating that the production equipment needs to adjust the rolling pressure. The device linkage feedback is recorded in the compensation space of the production equipment in real time, and the compensation space is a first-in first-out data buffer area for accumulating process adjustment suggestions, which is essentially a dynamic weight queue. Each feedback obtains an energy value according to the defect confidence and severity level, and each feedback includes specific content and a time stamp, such as December 18, 2024, 14:00, the rolling pressure is too high, reduce 5%. The device linkage feedback energy in the compensation space is continuously monitored, and the device linkage feedback energy is a quantitative index for measuring the intensity of the accumulated feedback information in the compensation space. A preset energy threshold is set based on the statistical process control (SPC) principle. When it is found that the feedback energy of the device is higher than the preset energy threshold, it means that the process deviation of the device has exceeded the normal fluctuation range, and the device control compensation of the production equipment is automatically executed. The operation parameters of the production equipment are adjusted according to the device linkage feedback, so as to actively correct the process deviation in the next production cycle and prevent the reoccurrence of defects from the source. By constructing the device linkage feedback and executing the device control compensation, the closed-loop control of defect detection and battery pole piece production is realized, the automation degree and efficiency of battery pole piece production are improved, and the consistency and reliability of soft package lithium battery pole piece production quality are ensured.

[0052] Further, if the defect confidence in the defect intelligent detection result is lower than the confidence threshold, a retest instruction is generated, the lithium battery pole piece is backtracked and transmitted according to the retest instruction, and is marked as a retest workpiece.

[0053] Specifically, the defect confidence level in the intelligent defect detection results is evaluated. If the defect confidence level is lower than a confidence threshold, it indicates significant uncertainty in the current detection results. In this case, a retest command is generated. The retest command is a control signal that instructs the lithium battery electrode to be re-inspected. According to the retest command, the lithium battery electrode is transferred from its current detection location back to the previous detection area for re-inspection. After the electrode returns to the detection area, it is marked as a retest workpiece, indicating that the electrode needs to be re-inspected. The retest workpiece mark can be a physical tag or an electronic record, used to distinguish between ordinary electrodes and electrodes requiring retesting. By generating retest commands and transmitting them retrospectively, uncertain detection results can be re-verified, ensuring the accuracy and reliability of electrode detection results and improving the overall quality and safety of lithium battery electrodes.

[0054] Example 2, based on the same inventive concept as the intelligent defect detection method for soft-pack lithium battery electrodes in the foregoing examples, such as... Figure 2 As shown, this application provides a smart defect detection device for pouch lithium battery electrode sheets, wherein the smart defect detection device for pouch lithium battery electrode sheets includes: The data acquisition module 11 is used to scan the unique identification code of the lithium battery electrode, read the electrode manufacturing process parameters and material micro-characteristics based on the unique identification code, and establish a retrospective dataset; the defect prediction module 12 is used to read the historical defect dataset, predict the defect probability of the lithium battery electrode based on the retrospective dataset and the historical defect dataset, and establish a prediction result; the sensor activation module 13 is used to activate the monitoring sensor at the time node when the lithium battery electrode reaches the target detection area, the monitoring sensor including a visual sensor and an infrared sensor, and establish a multimodal dataset; the data alignment module 14 is used to perform spatial and temporal alignment of the multimodal dataset and establish an alignment feature representation of the same physical area; the defect detection result establishment module 15 is used to use the prediction result and the alignment feature representation to identify defects of concern, establish a defect concern identification result, and establish a defect intelligent detection result based on the defect concern identification result and the prediction result.

[0055] Further, the data alignment module 14 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: performing sensor calibration of visual images and infrared thermal images in the multi-modal data set, establishing a mapping relationship from each sensor coordinate system to a production line physical coordinate system, and performing distortion correction of the multi-modal data set to output standardized modal data in the same coordinate system; after configuring MARK points and texture features of the lithium battery pole piece, projecting the standardized modal data to the same physical area by using the MARK points and the texture features, and outputting a modal feature map after spatial alignment; reading a pole piece conveying speed and a sampling time stamp of each sensor in the same coordinate system; using a sliding time window and speed compensation, corresponding different modal data frames in the modal feature map to the same physical area based on the pole piece conveying speed and the sampling time stamp of each sensor, and outputting a multi-modal feature sequence after time alignment, and outputting the multi-modal feature sequence as an aligned feature representation of the same physical area.

[0056] Further, the defect detection result establishing module 15 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: sending the prediction result and the aligned feature representation to a dynamic joint identification network; calling a priori layer of the dynamic joint identification network, projecting a defect occurrence probability in the prediction result to a physical coordinate system after receiving the prediction result, and establishing a priori defect risk map; after configuring a dynamic attention weight of the dynamic joint identification network by using the priori defect risk map, performing attention feature identification of the aligned feature representation, and establishing a defect attention identification result; performing joint authentication fusion of the defect attention identification result and the prediction result, establishing a joint defect confidence, and outputting the joint defect confidence as a defect intelligent detection result.

[0057] Further, the defect detection result establishing module 15 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: acquiring identification source data of the defect attention identification result, performing data quality evaluation on the identification source data, and establishing source data trustworthiness; if a joint authentication trustworthiness in the joint authentication fusion process of the defect attention identification result and the prediction result is lower than a first preset threshold, and the corresponding source data trustworthiness is lower than a second preset threshold, triggering an additional acquisition instruction; configuring an attention area according to the additional acquisition instruction, and performing additional data acquisition of the attention area, and updating the joint defect confidence according to the additional data acquisition result.

[0058] Further, the defect prediction module 12 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: analyzing the backtracking data set, establishing a traceability parameter table, the manufacturing process parameters in the traceability parameter table including rolling pressure, coating speed, baking temperature, and rolling tension, and the material microscopic features including particle distribution, coating uniformity, and porosity characteristics; analyzing the historical defect data set, performing time series defect probability prediction according to the historical defect data set, and establishing a first prediction result; performing causal association analysis based on the historical defect data set and the traceability parameter table, and establishing a second prediction result; and fusing the first prediction result and the second prediction result to establish the prediction result.

[0059] Further, the defect detection result establishing module 15 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: performing defect classification early warning level matching by using the defect intelligent detection result, and establishing a classification early warning signal; and after the classification early warning signal is used to configure the conveying belt shunting parameters, automatically performing lithium battery pole piece shunting conveying management.

[0060] Further, the sensor starting module 13 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: the monitoring sensor further includes an ultrasonic sensor, the ultrasonic sensor is used to perform ultrasonic monitoring of the lithium battery pole piece, ultrasonic feedback data is established, the ultrasonic feedback data is used to perform trust authentication on the defect intelligent detection result, and the defect intelligent detection result is updated.

[0061] Further, the defect detection result establishing module 15 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: constructing a device linkage feedback of a production equipment according to the defect intelligent detection result; recording the device linkage feedback into a compensation space of the production equipment, and when the device linkage feedback energy in the compensation space is higher than a preset energy threshold, performing device control compensation of the production equipment.

[0062] Further, the defect detection result establishing module 15 in the defect intelligent detection device for soft package lithium battery pole piece is further used for: if the defect confidence in the defect intelligent detection result is lower than a confidence threshold, generating a retest instruction, backtracking the lithium battery pole piece according to the retest instruction, and marking the lithium battery pole piece as a retest workpiece.

[0063] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. Figure 1The method for intelligent defect detection of soft package lithium battery pole piece in embodiment one and the specific examples are also applicable to the device for intelligent defect detection of soft package lithium battery pole piece in this embodiment. Through the foregoing detailed description of the method for intelligent defect detection of soft package lithium battery pole piece, those skilled in the art can clearly understand the device for intelligent defect detection of soft package lithium battery pole piece in this embodiment. Therefore, for the sake of brevity of the specification, no further detailed description is given here.

[0064] The above description of disclosed embodiments enables one skilled in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0065] Obviously, for those skilled in the art, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A method for intelligent detection of defects of a soft-pack lithium battery pole piece, characterized in that, The method comprises: Scanning the unique identification code of the lithium battery pole piece, reading the pole piece manufacturing process parameters and material micro features according to the unique identification code, and establishing a traceability data set; Reading a historical defect data set, performing defect probability prediction of the lithium battery pole piece according to the traceability data set and the historical defect data set, and establishing a prediction result; Starting a monitoring sensor at a time node when the lithium battery pole piece reaches a target detection area, the monitoring sensor comprising a visual sensor and an infrared sensor, and establishing a multi-modal data set; Performing spatial and temporal alignment of the multi-modal data set, and establishing an aligned feature representation of the same physical area; Performing defect attention recognition using the prediction result and the aligned feature representation, establishing a defect attention recognition result, and establishing a defect intelligent detection result according to the defect attention recognition result and the prediction result.

2. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, The execution of the spatial and temporal alignment of the multi-modal data set and the establishment of the aligned feature representation of the same physical area comprises: Performing sensor calibration of visual images and infrared thermal images in the multi-modal data set, establishing a mapping relationship from each sensor coordinate system to the production line physical coordinate system, and performing distortion correction of the multi-modal data set to output standardized modal data in the same coordinate system; After configuring the MARK point and the texture feature of the lithium battery pole piece, projecting the standardized modal data to the same physical area using the MARK point and the texture feature, and outputting the modal feature map after spatial alignment; Reading the pole piece conveying speed and the sampling time stamp of each sensor in the same coordinate system; Using a sliding time window and speed compensation, corresponding different modal data frames in the modal feature map to the same physical area based on the pole piece conveying speed and the sampling time stamp of each sensor, outputting the multi-modal feature sequence after temporal alignment, and outputting the multi-modal feature sequence as the aligned feature representation of the same physical area.

3. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, The defect attention recognition using the prediction result and the aligned feature representation and the establishment of the defect attention recognition result comprise: Sending the prediction result and the aligned feature representation to a dynamic joint recognition network; Calling the priori layer of the dynamic joint recognition network, projecting the defect occurrence probability in the prediction result to the physical coordinate system after receiving the prediction result, and establishing a priori defect risk map; After configuring the dynamic attention weight of the dynamic joint recognition network using the priori defect risk map, performing attention feature recognition of the aligned feature representation, and establishing a defect attention recognition result; Jointly authenticating and fusing the defect attention recognition result and the prediction result, establishing a joint defect confidence, and outputting the joint defect confidence as a defect intelligent detection result.

4. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 3, characterized in that, The joint authentication and fusion of the defect attention recognition result and the prediction result and the establishment of the joint defect confidence comprise: Obtaining the recognition source data of the defect attention recognition result, performing data quality evaluation on the recognition source data, and establishing a source data trust degree; If the joint authentication trust degree in the joint authentication fusion process of the defect attention recognition result and the prediction result is lower than a first preset threshold, and the corresponding source data trust degree is lower than a second preset threshold, an additional acquisition instruction is triggered; According to the additional acquisition instruction, a region of interest is configured, and additional data acquisition of the region of interest is performed, and the joint defect confidence is updated according to the additional data acquisition result.

5. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, The defect probability prediction of the lithium battery pole piece according to the backtracking data set and the historical defect data set establishes a prediction result, including: The backtracking data set is analyzed to establish a traceability parameter table, the manufacturing process parameters in the traceability parameter table include rolling pressure, coating speed, baking temperature and rolling tension, and the material micro features include particle distribution, coating uniformity and porosity characteristics; The historical defect data set is analyzed, and a time sequence defect probability prediction is performed according to the historical defect data set to establish a first prediction result; Based on the historical defect data set and the traceability parameter table, a causal correlation analysis is performed to establish a second prediction result; The first prediction result and the second prediction result are fused to establish the prediction result.

6. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, After the defect intelligent detection result is established according to the defect attention recognition result and the prediction result, including: The defect intelligent detection result is used for defect classification early warning level matching to establish a classification early warning signal; After the classification early warning signal is used to configure the conveyor belt shunt parameters, the shunt conveying management of the lithium battery pole piece is automatically performed.

7. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, The monitoring sensor further includes an ultrasonic sensor, and the ultrasonic sensor is used to perform ultrasonic monitoring of the lithium battery pole piece to establish ultrasonic feedback data, and the ultrasonic feedback data is used for trust authentication of the defect intelligent detection result to update the defect intelligent detection result. 8.The method for intelligent defect detection of soft-pack lithium battery pole piece according to claim 1, characterized in that, After the defect intelligent detection result is established according to the defect attention recognition result and the prediction result, further including: According to the defect intelligent detection result, a device linkage feedback of the production equipment is constructed; The device linkage feedback is recorded in the compensation space of the production equipment, and when the device linkage feedback energy in the compensation space is higher than a preset energy threshold, a device control compensation of the production equipment is performed.

9. The method for intelligent detection of defects of soft package lithium battery pole piece according to claim 1, characterized in that, If the defect confidence in the defect intelligent detection result is lower than a confidence threshold, a retest instruction is generated, the lithium battery pole piece is backtracked according to the retest instruction, and is marked as a retest workpiece.

10. A device for intelligent detection of defects of a soft-pack lithium battery pole piece, characterized in that, The steps of the defect intelligent detection method for soft package lithium battery pole piece in any one of claims 1 to 9, including: A data acquisition module is configured to scan the unique identification code of the lithium battery pole piece, read the pole piece manufacturing process parameters and material micro features according to the unique identification code, and establish a backtracking data set; A defect prediction module is configured to read a historical defect data set, perform defect probability prediction of the lithium battery pole piece according to the backtracking data set and the historical defect data set, and establish a prediction result; A sensor starting module is configured to start a monitoring sensor when the lithium battery pole piece reaches a target detection area, the monitoring sensor includes a visual sensor and an infrared sensor, and a multi-modal data set is established. a data alignment module configured to perform spatial and temporal alignment of the multi-modal data set to establish aligned feature representations of the same physical region; a defect detection result establishment module configured to perform defect attention identification using the prediction result and the aligned feature representations, to establish a defect attention identification result, and to establish a defect intelligent detection result according to the defect attention identification result and the prediction result.

Images