A machine learning-based PCB defect automatic detection system and method

Abstract

The application discloses a kind of machine learning PCB defect automatic detection system and method, mainly related to defect detection technical field.The system includes: deformation cloud chart sequence acquisition module, for data acquisition, obtains first PCB bending video image frame sequence, first flexible stress distribution diagram sequence and first deformation cloud chart sequence;Implicit trend vector acquisition module is used to obtain first crack implicit trend vector;Defect feature vector storage module is used to store first modified crack defect feature vector to defect memory bank;Update defect memory bank acquisition module is used to obtain updated defect memory bank;Defect detection result acquisition module is used to obtain the defect detection result of the flexible PCB to be measured.The beneficial effects of the present application are that the technical problems of low reliability of detection results caused by lack of dynamic defect detection of PCB in the prior art are solved, and the technical effect of dynamic and reliable defect detection of PCB is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect detection, and in particular to a machine learning-based PCB defect automatic detection system and method. BACKGROUND

[0002] With the development of flexible electronic technology, flexible printed circuit boards (PCBs) are widely used in wearable devices, flexible displays, and bending sensors. However, flexible PCBs often suffer from periodic bending, folding, and torsional loads during service, which can cause micro-cracks or hidden structural damage, seriously affecting their electrical performance and service life. Existing flexible PCB defect detection methods are mostly static detection, which is difficult to find subtle cracks or crack evolution trends caused by bending fatigue in time, thus leading to detection results susceptible to local noise interference, resulting in insufficient defect recognition accuracy.

[0003] The prior art has the technical problem of lack of dynamic defect detection of PCB, resulting in low reliability of detection results. SUMMARY

[0004] The present application provides a machine learning-based PCB defect automatic detection system and method for solving the technical problem of lack of dynamic defect detection of PCB in the prior art, resulting in low reliability of detection results.

[0005] In view of the above problems, the present application provides a machine learning-based PCB defect automatic detection system and method.

[0006] In a first aspect of the present application, a machine learning PCB defect automatic detection system is provided, comprising: a deformation cloud map sequence obtaining module, configured to apply a cycle bending load to a to-be-tested flexible PCB through a bending fatigue testing machine based on a preset bending load cycle parameter, and perform data acquisition on a bending process when a first bending load is applied by using an industrial camera, a flexible stress sensor and a deformation sensor, to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence and a first deformation cloud map sequence; an implicit trend vector obtaining module, configured to perform crack time sequence iterative trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector; a defect feature vector storage module, configured to correct the first crack implicit trend vector based on the first flexible stress distribution map sequence and the first deformation cloud map sequence to obtain a first corrected crack defect feature vector, and store the first corrected crack defect feature vector to a defect memory bank; an updated defect memory bank obtaining module, configured to perform bending crack defect testing on the to-be-tested flexible PCB again based on the preset bending load cycle parameter, and update the defect memory bank according to the testing result to obtain an updated defect memory bank; and a defect detection result obtaining module, configured to perform integrated correlation verification on the updated defect memory bank to obtain a to-be-tested flexible PCB defect detection result.

[0007] In a possible implementation, the bending load cycle parameter is selected from at least one of the following parameters: bending radius, bending frequency, bending angle, cycle number and load amplitude.

[0008] In a possible implementation, the implicit trend vector obtaining module is configured to perform the following steps: performing trend feature analysis on the first PCB bending video image frame sequence according to a crack multi-time sequence bandwidth set respectively, and sorting according to the corresponding crack multi-time sequence bandwidth from small to large to obtain a crack multi-time sequence trend vector sequence; extracting a first crack multi-time sequence trend vector and a second crack multi-time sequence trend vector from the crack multi-time sequence trend vector sequence to perform implicit trend enhancement analysis, and constructing a first implicit trend enhancement matrix; performing crack time sequence iterative trend enhancement on a second crack implicit trend vector by using the first implicit trend enhancement matrix to obtain a first-stage crack implicit trend vector; performing crack time sequence iterative trend enhancement on a third crack multi-time sequence trend vector in the crack multi-time sequence trend vector sequence based on the first-stage crack implicit trend vector, and similarly, obtaining the first crack implicit trend vector.

[0009] In a possible implementation, the implicit trend vector obtaining module is further configured to perform the following steps: based on the model of the flexible PCB to be tested, performing historical crack development data mining to obtain a historical crack development data set; dividing the historical crack development data set according to crack types, and performing crack development time length extraction on the division result to obtain M crack development time length sets, where M is a positive integer; extracting the maximum value and the minimum value in each of the M crack development time length sets respectively, and constructing the crack multi-time sequence bandwidth set.

[0010] In a possible implementation, the defect feature vector storage module is configured to perform the following steps: performing feature recognition on the first flexible stress distribution sequence and the first deformation cloud map sequence to determine a first flexible stress feature vector sequence and a first deformation cloud map feature vector sequence; performing trend analysis on the first flexible stress feature vector sequence and the first deformation cloud map feature vector sequence to determine a first flexible stress trend vector and a first deformation cloud map trend vector; and performing weighted correction on the first crack implicit trend vector based on the first flexible stress trend vector and the first deformation cloud map trend vector to obtain a first corrected crack defect feature vector.

[0011] In a possible implementation, the defect detection result obtaining module is configured to perform the following steps: extracting the first corrected crack defect feature vector and the second corrected crack defect feature vector from the updated defect memory bank; performing correlation verification on the second corrected crack defect feature vector and the first corrected crack defect feature vector by taking the crack position as a correlation constraint, and if the verification fails, removing the first corrected crack defect feature vector from the updated defect memory bank to obtain a stage cleaning updated defect memory bank; again extracting a third corrected crack defect feature vector from the stage cleaning updated defect memory bank to perform correlation verification on the second corrected crack defect feature vector, until the last corrected crack defect feature vector in the stage cleaning updated defect memory bank is reached to obtain a target cleaning updated defect memory bank; and performing identification on the target cleaning updated defect memory bank by using a detection result identifier to obtain the defect detection result of the flexible PCB to be tested.

[0012] In a possible implementation, the defect detection result obtaining module is further configured to perform the following steps: extracting first crack position center coordinates, a first crack contour region, second crack position center coordinates, and a second crack contour region of the first modified crack defect feature vector and the second modified crack defect feature vector; performing center position deviation distance calculation based on the first crack position center coordinates and the second crack position center coordinates, and determining whether a calculation result meets a preset position deviation threshold; if not, determining that the positions are inconsistent; calculating an intersection-over-union ratio based on the first crack contour region and the second crack contour region, and determining whether the intersection-over-union ratio meets a preset intersection-over-union ratio threshold; if not, determining that the crack contour diffusion is inconsistent; and when the positions are inconsistent or the crack contour diffusion is inconsistent, determining that the correlation verification result fails the verification.

[0013] In a second aspect of the present application, a machine learning-based PCB defect automatic detection method is provided, which includes: based on preset bending load cycle parameters, applying a cycle bending load to a to-be-tested flexible PCB by a bending fatigue testing machine, and collecting data of a bending process when a first bending load is applied by using an industrial camera, a flexible stress sensor, and a deformation sensor to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence, and a first deformation cloud image sequence; performing crack time sequence iteration trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector; performing crack defect correction on the first crack implicit trend vector according to the first flexible stress distribution map sequence and the first deformation cloud image sequence to obtain a first modified crack defect feature vector, and storing the first modified crack defect feature vector to a defect memory bank; again performing bending crack defect testing on the to-be-tested flexible PCB based on the preset bending load cycle parameters, and updating the defect memory bank according to a testing result to obtain an updated defect memory bank; performing integrated correlation verification on the updated defect memory bank to obtain a to-be-tested flexible PCB defect detection result.

[0014] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0015] This application, based on preset bending load cyclic parameters, applies cyclic bending loads to the flexible PCB under test using a bending fatigue testing machine. An industrial camera, flexible stress sensor, and deformation sensor are used to collect data on the bending process during the first application of the bending load, obtaining a first PCB bending video image frame sequence, a first flexible stress distribution map sequence, and a first deformation cloud map sequence. Then, crack time-series iterative trend analysis is performed on the first PCB bending video image frame sequence to obtain a first implicit crack trend vector. Subsequently, based on the first flexible stress distribution map sequence and the first deformation cloud map sequence, the first implicit crack trend vector is corrected for crack defects, obtaining a first corrected crack defect feature vector. This first corrected crack defect feature vector is stored in a defect memory bank. Then, the flexible PCB under test is tested again for bending crack defects based on the preset bending load cyclic parameters, and the defect memory bank is updated based on the test results, obtaining an updated defect memory bank. The updated defect memory bank is then integrated and correlated for verification to obtain the defect detection results of the flexible PCB under test. This achieves the technical effect of improving the reliability of defect detection by performing multiple detection verifications on the flexible PCB under test. Attached Figure Description

[0016] Appendix Figure 1 This is a schematic diagram of the structure of a machine learning-based automatic PCB defect detection system provided in an embodiment of the present invention.

[0017] Appendix Figure 2 This is a schematic diagram of a machine learning-based automatic PCB defect detection method provided in an embodiment of the present invention.

[0018] The labels shown in the attached diagram:

[0019] The module 11 obtains the deformation cloud map sequence, the module 12 obtains the implicit trend vector, the module 13 stores the defect feature vector, the module 14 updates the defect memory, and the module 15 obtains the defect detection result. Detailed Implementation

[0020] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims. It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices.

[0021] Example 1, as shown in the appendix Figure 1As shown, the present application provides a machine learning PCB defect automatic detection system, which comprises:

[0022] The deformation cloud sequence obtaining module 11 is configured to apply a cyclic bending load to the flexible PCB to be tested by a bending fatigue testing machine based on preset bending load cycle parameters, and collect data of a bending process when the first bending load is applied by using an industrial camera, a flexible stress sensor and a deformation sensor to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence and a first deformation cloud sequence.

[0023] Further, the bending load cycle parameters are selected from at least one of the following parameters: bending radius, bending frequency, bending angle, cycle number and load amplitude.

[0024] In one possible embodiment, the preset bending load cycle parameters refer to physical load conditions preset by a person skilled in the art before the system performs the bending fatigue test, and are selected from at least one of the following parameters: bending radius (bending curvature size), bending frequency (bending times per unit time), bending angle (bending amplitude), cycle number (total number of repeated bending) and load amplitude (size of applied load). The industrial camera is used to record dynamic image information in the bending process. The flexible stress sensor and the deformation sensor are embedded or attached to the surface of the flexible PCB to be tested, respectively, for real-time measurement of stress distribution and local deformation in the bending process, and output time sequence diagram data.

[0025] The bending fatigue testing machine applies continuous and controllable bending and resetting operations to the flexible PCB to be tested, simulating repeated bending of the product in actual application. When the first bending load is applied, the industrial camera synchronously collects the entire bending process to generate a continuous video image frame sequence, which is used for subsequent crack timing analysis. At the same time, the flexible stress sensor and the deformation sensor measure the stress distribution state and deformation of the PCB at different positions in real time, and output the first flexible stress distribution map sequence and the first deformation cloud sequence, respectively. By providing a multi-modal, time-aligned original data basis for the subsequent module, the data source comprehensiveness for subsequent crack implicit trend extraction, defect feature correction and memory verification is ensured, and the technical effect of improving the reliability of dynamic defect detection is achieved.

[0026] The implicit trend vector obtaining module 12 is configured to perform crack timing iterative trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector.

[0027] Further, the implicit trend vector obtaining module 12 is configured to perform the following steps:

[0028] respectively, and sorting according to the corresponding crack multi-time sequence bandwidths from small to large, to obtain a crack multi-time sequence trend vector sequence;

[0029] Performing implicit trend enhancement analysis on the first crack multi-time sequence trend vector and the second crack multi-time sequence trend vector extracted from the crack multi-time sequence trend vector sequence, to construct a first implicit trend enhancement matrix;

[0030] Performing crack time sequence iterative trend enhancement on the second crack implicit trend vector by using the first implicit trend enhancement matrix, to obtain a first-stage crack implicit trend vector;

[0031] Performing crack time sequence iterative trend enhancement on the third crack multi-time sequence trend vector in the crack multi-time sequence trend vector sequence based on the first-stage crack implicit trend vector, and so on, to obtain the first crack implicit trend vector.

[0032] Further, the implicit trend vector obtaining module 12 is further configured to perform the following steps:

[0033] Based on the specification model of the to-be-tested flexible PCB, performing historical crack development data mining to obtain a historical crack development data set;

[0034] Dividing the historical crack development data set according to crack types, and extracting crack development time lengths from the division results to obtain M crack development time length sets, wherein M is a positive integer;

[0035] Respectively extracting the maximum value and the minimum value in the M crack development time length sets, to construct the crack multi-time sequence bandwidth set.

[0036] In one possible embodiment, for the collected first PCB bending video image frame sequence, multi-scale trend feature analysis is performed according to the pre-constructed crack multi-time sequence bandwidth set, that is, different scale analysis is performed on the image frame sequence according to the development and evolution time lengths corresponding to different types of cracks. And the extraction results are sorted from small to large according to the size of the crack multi-time sequence bandwidth, to form a crack multi-time sequence trend vector sequence. Then, the crack multi-time sequence trend vector sequence is sequentially subjected to implicit trend enhancement analysis to determine a first implicit trend enhancement matrix. The first implicit trend enhancement matrix is used to optimize the second trend vector to form a first-stage crack implicit trend vector. Then, the module continues to iteratively enhance the remaining multi-time sequence trend vectors with the first-stage result as a reference, gradually improves the physical consistency and time continuity of the trend vector, and finally outputs the first crack implicit trend vector.

[0037] Preferably, in order to ensure the accuracy of the multi-time sequence bandwidth set, according to the specification and model of the flexible PCB to be tested, the historical crack development data of the flexible PCB of the same model is mined, different types of evolution time intervals are divided and extracted according to the crack types, and the maximum value and the minimum value are counted, so as to adaptively construct a bandwidth set suitable for the current test object. Through the time sequence consistent amplification and dynamic enhancement of the weak crack changes implied in the original video data, the sensitivity and accuracy of the subsequent defect correction and identification are significantly improved.

[0038] In an embodiment of the present application, the first element similarity set is obtained by using the cosine similarity calculation formula to perform one-to-one inner product mapping on the same type of elements in the first crack multi-time sequence trend vector and the second crack multi-time sequence trend vector. The first element similarity set reflects the similarity between the first crack multi-time sequence trend vector and the second crack multi-time sequence trend vector. Then, the first element similarity set is normalized, and the normalized value after normalization is filled into the initially empty matrix to obtain the first implicit trend enhancement matrix. The first implicit trend enhancement matrix reflects the implicit correlation between the first crack multi-time sequence trend vector and the second crack multi-time sequence trend vector. The first implicit trend enhancement matrix is multiplied with the second crack implicit trend vector, so as to suppress the noise in the crack evolution time sequence information and enhance the key data, and obtain the first stage crack implicit trend vector.

[0039] Preferably, based on the same obtaining principle as obtaining the first stage crack implicit trend vector, the third crack multi-time sequence trend vector in the crack multi-time sequence trend vector sequence is subjected to crack time sequence iterative trend enhancement by using the first stage crack implicit trend vector, and the remaining crack multi-time sequence trend vectors in the crack multi-time sequence trend vector sequence are iteratively enhanced in the same way, to obtain the first crack implicit trend vector.

[0040] In an embodiment of the present application, the first historical crack development data is extracted from the historical crack development data set, the remaining historical crack development data in the historical crack development data set is matched according to the crack type corresponding to the first historical crack development data, and the historical crack development data of the same crack type is added to the first division set. The first division set is excluded from the historical crack development data set, and then the remaining historical crack development data in the historical crack development data set is divided based on the same principle as obtaining the first division set, taking the crack type as the division basis, to obtain the M division sets. Wherein, M is a positive integer.

[0041] Further, the crack development time length of each historical crack development data in the M division sets is extracted to obtain M crack development time length sets. Further, the maximum value and the minimum value in each crack development time length set in the M crack development time length sets are extracted respectively to obtain an initial crack multi-time sequence bandwidth set. The initial crack multi-time sequence bandwidth set is unionized to obtain the crack multi-time sequence bandwidth set.

[0042] The defect feature vector storage module 13 is configured to correct the first crack implicit trend vector according to the first flexible stress distribution sequence and the first deformation cloud sequence to obtain a first corrected crack defect feature vector, and store the first corrected crack defect feature vector in the defect memory.

[0043] Further, the defect feature vector storage module 13 is configured to perform the following steps:

[0044] The first flexible stress feature vector sequence and the first deformation cloud feature vector sequence are determined by performing feature recognition on the first flexible stress distribution sequence and the first deformation cloud sequence.

[0045] The first flexible stress trend vector and the first deformation cloud trend vector are determined by performing trend analysis according to the first flexible stress feature vector sequence and the first deformation cloud feature vector sequence.

[0046] The first corrected crack defect feature vector is obtained by performing weighted correction on the first crack implicit trend vector based on the first flexible stress trend vector and the first deformation cloud trend vector.

[0047] In one possible embodiment, the first flexible stress feature vector sequence represents an ordered vector sequence composed of stress local or global features (such as maximum stress, stress gradient, stress concentration area, etc.) extracted from each frame of flexible stress distribution. The first deformation cloud feature vector sequence represents an ordered vector sequence composed of deformation, displacement direction, deformation concentration, etc. extracted from each frame of deformation cloud. The defect memory is a database or vector library for storing multiple corrected crack defect feature vectors, which is used for subsequent similarity comparison and dynamic learning.

[0048] In one possible embodiment, the first flexible stress distribution sequence and the first deformation cloud sequence are first processed in a traversal manner, and stress concentration, deformation peak and their spatial distribution are identified frame by frame, physical features of each frame are extracted, and the first flexible stress feature vector sequence and the first deformation cloud feature vector sequence are formed by summarizing. Then, time series trend analysis is performed on the two feature vector sequences respectively, and the overall trend vectors reflecting the stress evolution and deformation evolution in the entire bending cycle are obtained, which are referred to as the first flexible stress trend vector and the first deformation cloud trend vector. Subsequently, the two trend vectors are combined with the previously obtained first crack implicit trend vector, and frame-level correction is performed on the original trend vector by using a preset weighting coefficient (which can be a fixed value or self-adaptively learned through historical data), and the first corrected crack defect feature vector is output. Finally, the corrected feature vector is stored in the defect memory together with meta information (such as model, load parameter and timestamp), thereby providing traceable data support for subsequent defect evolution analysis, automatic similarity retrieval and multi-round verification.

[0049] In one possible embodiment, a plurality of sample flexible stress distribution maps and a plurality of sample flexible stress feature vectors are obtained as training data, the framework based on the feedforward neural network is supervised trained by using the training data, the network parameters of the framework are revised according to the accuracy of the output data of the framework during the training process, and the feature recognizer after the revision is completed is obtained. The feature recognizer is used to perform feature recognition on the first flexible stress distribution sequence, and the first flexible stress feature vector sequence is obtained. Based on the same obtaining principle as that of the first flexible stress feature vector sequence, the first deformation cloud sequence is subjected to feature recognition, and the first deformation cloud feature vector sequence is obtained.

[0050] Further, a trend analyzer is obtained, wherein the trend analyzer is a functional module constructed by using a plurality of sample feature vector sequences and a plurality of sample trend vectors as a training data set and by using a convolutional neural network as a framework. The trend analyzer is used to perform trend analysis on the first flexible stress feature vector sequence and the first deformation cloud feature vector sequence respectively, and the first flexible stress trend vector and the first deformation cloud trend vector are obtained. Further, the first crack implicit trend vector is weighted and corrected according to the weight preset by a person skilled in the art by using the first flexible stress trend vector and the first deformation cloud trend vector, and the first corrected crack defect feature vector is obtained.

[0051] By coupling and correcting the pure visual crack trend result with the actual physical field data, the technical effect of greatly improving the physical consistency of crack detection and the reliability of subsequent identification is achieved.

[0052] The updating defect memory base obtaining module 14 is configured to perform a bending crack defect test on the to-be-tested flexible PCB based on the preset bending load cycle parameters again, and update the defect memory base according to a test result to obtain an updated defect memory base.

[0053] In one possible embodiment, based on the same principle as that of obtaining the first modified crack defect feature vector, a corresponding modified crack defect feature vector is obtained by performing a bending crack defect test on the to-be-tested flexible PCB based on the preset bending load cycle parameters until a cycle number of the preset bending load cycle parameters is met, and a test result, that is, the modified crack defect feature vector obtained each time is stored in the defect memory base to update the defect memory base, so as to obtain the updated defect memory base. By continuously accumulating multiple rounds of crack evolution data for the PCB model, a high-quality and dynamically evolving defect memory base is formed, which provides data guarantee for subsequent integrated correlation verification and final defect identification.

[0054] The defect detection result obtaining module 15 is configured to perform integrated correlation verification on the updated defect memory base to obtain a defect detection result of the to-be-tested flexible PCB.

[0055] Further, the defect detection result obtaining module 15 is configured to perform the following steps:

[0056] The first modified crack defect feature vector and the second modified crack defect feature vector are extracted from the updated defect memory base;

[0057] The second modified crack defect feature vector and the first modified crack defect feature vector are used for correlation verification with the crack position as a correlation constraint, if the verification fails, the first modified crack defect feature vector is removed from the updated defect memory base to obtain a stage-cleaning updated defect memory base;

[0058] The third modified crack defect feature vector is extracted from the stage-cleaning updated defect memory base again to perform correlation verification on the second modified crack defect feature vector until the last modified crack defect feature vector of the stage-cleaning updated defect memory base is reached to obtain a target-cleaning updated defect memory base;

[0059] The target-cleaning updated defect memory base is identified by using a detection result identifier to obtain the defect detection result of the to-be-tested flexible PCB.

[0060] Further, the defect detection result obtaining module 15 is further configured to perform the following steps:

[0061] The first crack position center coordinates, the first crack contour area, the second crack position center coordinates and the second crack contour area of the first modified crack defect feature vector and the second modified crack defect feature vector are extracted.

[0062] a center position deviation distance is calculated based on the first crack position center coordinates and the second crack position center coordinates, and it is determined whether the calculation result satisfies a preset position deviation threshold value, and if not, it is determined that the positions are inconsistent;

[0063] an intersection-over-union ratio is calculated based on the first crack profile region and the second crack profile region, and it is determined whether the intersection-over-union ratio satisfies a preset intersection-over-union ratio threshold value, and if not, it is determined that the crack profile diffusion is inconsistent;

[0064] When the positions are inconsistent or the crack profile diffusion is inconsistent, the associated verification result is that the verification fails.

[0065] In one possible embodiment, by performing pairwise verification of the position and shape of adjacent or consecutive crack correction vectors in the updated defect memory library, redundant features with abnormal physical positions or trends are removed, and the reliability of the final flexible PCB defect detection result is improved. The crack position center coordinates refer to the crack geometric center point coordinates calculated by a crack shape segmentation or positioning algorithm, which are used for subsequent center point deviation analysis. The crack profile region refers to a closed region describing the range of the crack shape, which can be expressed by a rectangular bounding box, a convex hull or a pixel set, and is mainly used to calculate the intersection-over-union (IOU) to determine whether the crack diffusion is consistent. The center position deviation distance refers to the Euclidean distance of the two crack center points in space, which is used to constrain the spatial consistency of the crack position. The intersection-over-union (IOU) is the ratio of the intersection area to the union area of the two crack profile regions, which is used to measure the consistency and spatial overlap of the two crack shapes.

[0066] First, adjacent first correction crack defect feature vectors and second correction crack defect feature vectors are extracted from the updated defect memory library in chronological order, and position and profile correlation verification is performed based on the pair of vectors. Specifically, the crack position center coordinates of the two vectors are extracted, the center position deviation distance is calculated, and compared with the preset position deviation threshold value preset by the person skilled in the art. If the deviation is too large, it does not meet the requirements, and it is determined that the crack position is inconsistent, and the verification is marked as failed. At this time, it indicates that the first correction crack defect feature vector may be a false crack defect caused by accidental error or noise interference.

[0067] Meanwhile, the crack profile regions of the first modified crack defect feature vector and the second modified crack defect feature vector are further extracted, and an intersection-over-union (IOU) is calculated. Preferably, the intersection-over-union is obtained by calculating the ratio of the intersection to the union of the crack profile regions of the first modified crack defect feature vector and the second modified crack defect feature vector. If the intersection-over-union is lower than a preset intersection-over-union threshold value preset by the person skilled in the art, it is determined that the crack morphology diffusion is inconsistent, and it is considered that the verification fails. Once a pair of verification fails, the previous crack modification vector is immediately removed from the updated defect memory library, and a stage cleaning updated defect memory library is formed.

[0068] Subsequently, the next pair of modified vectors (such as the second and third vectors) is continuously extracted from the stage-cleaning library, and the position and profile association verification and removal operation is repeatedly performed until the last modified vector in the library is reached, and the stage cleaning is completed. Finally, the module calls the detection result recognizer to perform pattern clustering or confidence analysis on the effective modified vectors in the target cleaning updated defect memory library, and outputs the final defect detection result of the flexible PCB, including whether there is a crack, the crack position and evolution trend, and the reliability score of the detection.

[0069] Preferably, a plurality of sample target cleaning updated defect memory libraries and a plurality of sample defect detection results of the to-be-tested flexible PCB are obtained as a training data set, the plurality of sample defect detection results of the to-be-tested flexible PCB in the training data set are labeled, and the plurality of sample target cleaning updated defect memory libraries are input into a framework based on a feedforward neural network for training to obtain outputted plurality of verification defect detection results of the to-be-tested flexible PCB. The plurality of verification defect detection results of the to-be-tested flexible PCB are compared with the labeled plurality of sample defect detection results of the to-be-tested flexible PCB, the success rate of the comparison is counted, and if the success rate of the comparison is greater than or equal to a value preset by the person skilled in the art, the verification passes, and the trained detection result recognizer is obtained.

[0070] Through the multi-stage spatial and morphological consistency verification and dynamic removal, a physically self-consistent, clean and traceable defect feature vector set is constructed, so that the detection result has high robustness and low misjudgment rate under dynamic load.

[0071] In summary, the embodiments of the present application have at least the following technical effects:

[0072] 1. The present application obtains video image frame sequences, stress distribution graph sequences and deformation cloud atlas sequences covering a complete bending cycle through fatigue tests based on preset bending load cycle parameters, multi-modal synchronous acquisition of industrial cameras, flexible stress sensors and deformation sensors, which significantly improves the real-time detection capability of crack implicit growth and local abnormalities compared with traditional single vision or single point measurement, and effectively makes up for the problem that existing static detection means cannot track dynamic fatigue cracks.

[0073] 2. The application sufficiently superimposes flexible stress and deformation evolution information in the crack feature vector through crack time sequence iterative trend analysis and multi-modal physical trend weighted correction, greatly suppressing noise and false signals in pure visual trends. Meanwhile, the modified crack defect feature vector is continuously written into the defect memory bank to form a crack evolution sample library that can be repeatedly searched and dynamically updated, providing consistent physical, historical and traceable data support for subsequent detection.

[0074] 3. The application can continuously verify and update the crack correction feature vector in the memory bank through repeated bending fatigue testing of the flexible PCB to be tested, and combine crack position constraints with intersection over union (IOU) verification to remove inconsistent or noisy samples in multiple stages, significantly reducing cumulative error and single detection failure risk.

[0075] Embodiment two, based on the same inventive concept as the machine learning-based PCB defect automatic detection system in the preceding embodiment, as shown in FIG. 2, the application provides a machine learning-based PCB defect automatic detection method. The system and method embodiments in the application embodiment are based on the same inventive concept. The method comprises: Figure 2

[0076] Based on the preset bending load cycle parameters, a bending fatigue testing machine is used to apply a cyclic bending load to the flexible PCB to be tested, and an industrial camera, a flexible stress sensor and a deformation sensor are used to collect data during the bending process when the first bending load is applied, to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence and a first deformation cloud map sequence;

[0077] Performing crack time sequence iterative trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector;

[0078] According to the first flexible stress distribution map sequence and the first deformation cloud map sequence, the first crack implicit trend vector is corrected for crack defects to obtain a first modified crack defect feature vector, and the first modified crack defect feature vector is stored in a defect memory bank;

[0079] Again based on the preset bending load cycle parameters, the flexible PCB to be tested is subjected to bending crack defect testing, and the defect memory bank is updated according to the test results to obtain an updated defect memory bank;

[0080] Performing integrated correlation verification on the updated defect memory bank to obtain a flexible PCB defect detection result.

[0081] ​Further, the bending load cycle parameters are selected from at least one of the following parameters: bending radius, bending frequency, bending angle, cycle number, load amplitude.

[0082] Further, the method further comprises:

[0083] The first PCB bending video image frame sequence is respectively analyzed according to the crack multi-time sequence bandwidth set, and is sorted according to the corresponding crack multi-time sequence bandwidth from small to large, to obtain a crack multi-time sequence trend vector sequence;

[0084] The first crack multi-time sequence trend vector and the second crack multi-time sequence trend vector are extracted from the crack multi-time sequence trend vector sequence for implicit trend enhancement analysis, to construct a first implicit trend enhancement matrix;

[0085] The second crack implicit trend vector is subjected to crack time sequence iterative trend enhancement by using the first implicit trend enhancement matrix, to obtain a first-stage crack implicit trend vector;

[0086] The third crack multi-time sequence trend vector in the crack multi-time sequence trend vector sequence is subjected to crack time sequence iterative trend enhancement based on the first-stage crack implicit trend vector, and the first crack implicit trend vector is obtained by analogy.

[0087] Further, the method further comprises:

[0088] Based on the specification model of the to-be-tested flexible PCB, historical crack development data mining is performed to obtain a historical crack development data set;

[0089] The historical crack development data set is divided according to crack types, and crack development time length extraction is performed on the division result to obtain M crack development time length sets, wherein M is a positive integer;

[0090] The maximum value and the minimum value in the M crack development time length sets are respectively extracted to construct the crack multi-time sequence bandwidth set.

[0091] Further, the method further comprises:

[0092] Feature recognition is performed on the first flexible stress distribution map sequence and the first deformation cloud map sequence to determine a first flexible stress feature vector sequence and a first deformation cloud map feature vector sequence;

[0093] Trend analysis is performed according to the first flexible stress feature vector sequence and the first deformation cloud map feature vector sequence to determine a first flexible stress trend vector and a first deformation cloud map trend vector;

[0094] The first crack implicit trend vector is weighted and corrected based on the first flexible stress trend vector and the first deformation cloud trend vector, to obtain a first corrected crack defect feature vector.

[0095] Further, the method further comprises:

[0096] The first corrected crack defect feature vector and the second corrected crack defect feature vector are extracted from the updated defect memory base;

[0097] The second corrected crack defect feature vector and the first corrected crack defect feature vector are associated and verified with the crack position as an associated constraint, if the verification fails, the first corrected crack defect feature vector is removed from the updated defect memory base, to obtain a stage cleaning updated defect memory base;

[0098] The third corrected crack defect feature vector is extracted from the stage cleaning updated defect memory base to associate and verify the second corrected crack defect feature vector, until the last corrected crack defect feature vector of the stage cleaning updated defect memory base is reached, to obtain a target cleaning updated defect memory base;

[0099] The target cleaning updated defect memory base is identified by using a detection result identifier, to obtain the defect detection result of the flexible PCB to be detected.

[0100] Further, the method further comprises:

[0101] The first crack position center coordinates, the first crack contour area, the second crack position center coordinates and the second crack contour area of the first corrected crack defect feature vector and the second corrected crack defect feature vector are extracted;

[0102] The center position deviation distance is calculated based on the first crack position center coordinates and the second crack position center coordinates, and it is judged whether the calculation result meets a preset position deviation threshold, if not, it is determined that the positions are inconsistent;

[0103] The intersection and union ratio is calculated based on the first crack contour area and the second crack contour area, and it is judged whether the intersection and union ratio meets a preset intersection and union ratio threshold, if not, it is determined that the contour diffusion is inconsistent;

[0104] When the positions are inconsistent or the contour diffusion is inconsistent, the associated verification result is failed.

[0105] It should be noted that the above-mentioned embodiment sequences of the present application are merely for description only, but not for representing the advantages and disadvantages of the embodiments. And the above-mentioned embodiments of the present specification have been described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0106] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0107] The specification and drawings are merely exemplary of the present application, and any and all modifications, variations, combinations or equivalents that are within the scope of the present application should be included. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A machine learning based PCB defect auto-detection system, characterized in that, The system comprises: The deformation cloud sequence obtaining module is configured to apply a cyclic bending load to the flexible PCB to be tested by a bending fatigue testing machine based on preset bending load cycle parameters, and collect data on a bending process when the first bending load is applied by using an industrial camera, a flexible stress sensor, and a deformation sensor, to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence, and a first deformation cloud sequence; The implicit trend vector obtaining module is configured to perform crack timing iterative trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector; The defect feature vector storage module is configured to correct the first crack implicit trend vector based on the first flexible stress distribution map sequence and the first deformation cloud sequence, to obtain a first corrected crack defect feature vector, and store the first corrected crack defect feature vector in a defect memory bank; The updated defect memory bank obtaining module is configured to perform bending crack defect testing on the flexible PCB to be tested again based on the preset bending load cycle parameters, and update the defect memory bank based on a testing result, to obtain an updated defect memory bank; The defect detection result obtaining module is configured to perform integrated correlation verification on the updated defect memory bank, to obtain a defect detection result of the flexible PCB to be tested. The implicit trend vector obtaining module is configured to perform the following steps: The first PCB bending video image frame sequence is analyzed for trend features according to a crack multi-time-series bandwidth set, and is sorted according to the corresponding crack multi-time-series bandwidth in descending order, to obtain a crack multi-time-series trend vector sequence; First and second crack multi-time-series trend vectors are extracted from the crack multi-time-series trend vector sequence for implicit trend enhancement analysis, to construct a first implicit trend enhancement matrix; The second crack implicit trend vector is subjected to crack timing iterative trend enhancement by using the first implicit trend enhancement matrix, to obtain a first-stage crack implicit trend vector; The third crack multi-time-series trend vector in the crack multi-time-series trend vector sequence is subjected to crack timing iterative trend enhancement based on the first-stage crack implicit trend vector, and the same is repeated to obtain the first crack implicit trend vector; Historical crack development data are mined based on the specifications and models of the flexible PCB to be tested, to obtain a historical crack development data set; The historical crack development data set is divided according to crack types, and crack development time lengths are extracted from the division result, to obtain M crack development time length sets, where M is a positive integer; The maximum and minimum values in the M crack development time length sets are extracted respectively, to construct the crack multi-time-series bandwidth set.

2. The machine learned PCB defect auto-detection system of claim 1, wherein, The bending load cycle parameters are selected from at least one of the following parameters: bending radius, bending frequency, bending angle, cycle number, and load amplitude.

3. The machine learned PCB defect auto-detection system of claim 1, wherein, The defect feature vector storage module is configured to perform the following steps: Feature recognition is performed by traversing the first flexible stress distribution map sequence and the first deformation cloud sequence, to determine a first flexible stress feature vector sequence and a first deformation cloud feature vector sequence; performing trend analysis on the first flexible stress feature vector sequence and the first deformation nephogram feature vector sequence to determine a first flexible stress trend vector and a first deformation nephogram trend vector; performing weighted correction on the first crack implicit trend vector based on the first flexible stress trend vector and the first deformation nephogram trend vector to obtain a first modified crack defect feature vector.

4. The machine learned PCB defect auto-detection system of claim 1, wherein, The defect detection result obtaining module is configured to perform the following steps: extracting the first modified crack defect feature vector and the second modified crack defect feature vector from the updated defect memory bank; performing correlation verification on the second modified crack defect feature vector and the first modified crack defect feature vector with the crack position as a correlation constraint, and if the verification fails, removing the first modified crack defect feature vector from the updated defect memory bank to obtain a stage cleaning updated defect memory bank; performing correlation verification on the second modified crack defect feature vector and the third modified crack defect feature vector extracted from the stage cleaning updated defect memory bank again until the last modified crack defect feature vector in the stage cleaning updated defect memory bank is reached to obtain a target cleaning updated defect memory bank; performing identification on the target cleaning updated defect memory bank by using a detection result identifier to obtain the defect detection result of the to-be-tested flexible PCB.

5. The machine learned PCB defect auto-detection system of claim 4, wherein, The defect detection result obtaining module is further configured to perform the following steps: extracting the first crack position center coordinates, the first crack contour region, the second crack position center coordinates and the second crack contour region of the first modified crack defect feature vector and the second modified crack defect feature vector; performing center position deviation distance calculation based on the first crack position center coordinates and the second crack position center coordinates, and determining whether the calculation result meets a preset position deviation threshold value, and if not, determining that the positions are inconsistent; calculating an intersection and union ratio based on the first crack contour region and the second crack contour region, and determining whether the intersection and union ratio meets a preset intersection and union ratio threshold value, and if not, determining that the crack contour diffusion is inconsistent; if the positions are inconsistent or the crack contour diffusion is inconsistent, the correlation verification result is a failure.

6. A machine learning-based automatic PCB defect detection method, characterized by, The method is implemented by using the machine learning-based PCB defect automatic detection system of any one of claims 1-5, and the method comprises the following steps: applying a cyclic bending load to the to-be-tested flexible PCB by using a bending fatigue testing machine based on preset bending load cycle parameters, and collecting data of a bending process when the first bending load is applied by using an industrial camera, a flexible stress sensor and a deformation sensor to obtain a first PCB bending video image frame sequence, a first flexible stress distribution map sequence and a first deformation nephogram sequence; performing crack timing iteration trend analysis on the first PCB bending video image frame sequence to obtain a first crack implicit trend vector; performing crack defect correction on the first crack implicit trend vector based on the first flexible stress distribution map sequence and the first deformation nephogram sequence to obtain a first modified crack defect feature vector, and storing the first modified crack defect feature vector in a defect memory bank. Again, based on the preset bending load cycle parameters, the flexible PCB to be tested is subjected to bending crack defect test, and the defect memory library is updated according to the test result to obtain an updated defect memory library; The updated defect memory library is subjected to integrated correlation verification to obtain a defect detection result of the flexible PCB to be tested.

Images