Flexible circuit board patch anomaly detection system and processing method based on multi-modal data

Abstract

The invention relates to a multi-modal data-based flexible circuit board patch anomaly detection system and a processing method, in particular to the field of flexible circuit board patch anomaly detection.According to the system, intelligent anomaly detection in a flexible manufacturing scene is realized through multi-module collaboration; an anti-interference stable foundation is laid; the real-time migration module realizes instantaneous parameter migration by means of feature similarity, so that the response efficiency of a new batch is remarkably improved; the dynamic optimization module adjusts and models parameters into a reinforcement learning process, and synchronously optimizes the detection precision and the system robustness; and the verification and knowledge evolution module senses distribution drift through a confrontation mechanism, safely injects new knowledge and avoids historical forgetting risks, finally forms a lifelong learning architecture from feature decoupling to closed-loop evolution, realizes cross breakthrough in the dimensions of inter-batch adaptive time consumption, manual maintenance frequency, small sample generalization ability and the like, and improves the learning efficiency. And continuous and reliable detection in a complex process environment is effectively supported.

Description

Technical Field

[0001] The present invention relates to the field of flexible circuit board patch abnormality detection. More specifically, the present invention relates to a flexible circuit board patch abnormality detection system and processing method based on multi-modal data. Background Art

[0002] At present, flexible circuit board patch abnormality detection systems are widely used in the production process of flexible circuit boards. They mainly rely on multi-modal data such as vision and sensor data to monitor the circuit boards in real time. By combining technologies such as image processing and signal analysis, these systems can detect the patch process of circuit boards on the production line and identify potential defects or abnormalities. Existing detection methods mostly use fixed algorithms, such as traditional image processing algorithms and machine learning models, to determine the quality based on features such as the appearance and texture of the circuit board, and can effectively detect many common defects, such as solder joint problems and component position deviations.

[0003] However, existing flexible circuit board patch detection systems face many challenges in practical applications. Especially when facing the slight differences between different batches of circuit boards, due to the possible slight production process differences in each batch of circuit boards, which lead to changes in their detection features, traditional algorithms are difficult to effectively adapt to these differences, resulting in a decrease in detection accuracy. Existing algorithms are usually optimized for a specific working condition and lack the ability to adapt to the continuous changes in the production process. When sensor data or patch features change, fixed algorithms often cannot cope and often require manual adjustment or retraining, increasing the complexity of system maintenance and resulting in inflexible responses. Therefore, the adaptive ability of the system is insufficient, unable to flexibly meet the detection requirements of different batches of circuit boards, affecting production efficiency and the stability of quality. Summary of the Invention

[0004] In view of the technical problems existing in the prior art, the present invention provides a flexible circuit board patch abnormality detection system and processing method based on multi-modal data, through an adaptive knowledge base construction module, a real-time migration module, a dynamic optimization module, and a verification and knowledge evolution module, to solve the problems raised in the above background art.

[0005] The technical solution of the present invention to solve the above technical problems is as follows. A flexible circuit board patch abnormality detection system based on multi-modal data specifically includes: an adaptive knowledge base construction module, a real-time migration module, a dynamic optimization module, and a verification and knowledge evolution module;

[0006] Adaptive Knowledge Base Construction Module: Based on the historical batch of circuit board image sets, defect labels, and process parameters, use a feature decoupling network to extract process-invariant features and process-sensitive features, and impose a mutual information minimization constraint to achieve feature separation; use the process-invariant features and corresponding defect labels to construct an adaptive knowledge base with dynamic clustering , and complete the initial knowledge reserve;

[0007] Real-time Migration Module: For the current batch of real-time images, extract their process-invariant features; based on the weighted nearest neighbor algorithm, in the constructed adaptive knowledge base match the historical optimal detection parameters; fuse the matching parameters according to the similarity weight to generate the initial detection parameters for the current batch;

[0008] Dynamic Optimization Module: Starting from the initial detection parameters generated by the real-time migration module, analyze the detection confidence and parameter fluctuation status in real time through a meta-policy network; based on a composite reward mechanism, use the proximal policy optimization algorithm to dynamically adjust the detection parameters and output the optimized defect detection results;

[0009] Verification and Knowledge Evolution Module: Based on the current batch detection results of the dynamic optimization module, calculate the feature distribution difference between it and the adaptive knowledge base through the adversarial domain adaptation algorithm; when the difference exceeds the threshold, screen high-confidence samples and use the knowledge distillation strategy to update the feature decoupling network; inject the verified new feature-label pairs into the adaptive knowledge base , and perform the update of the adaptive knowledge base ;

[0010] In a preferred embodiment, in the adaptive knowledge base construction module, the feature decoupling network uses a dual-branch structure to extract features. Branch one outputs process-invariant features, including representing the solder joint shape and component position. The process-invariant feature extraction formula is:

[0011] ;

[0012] where, represents the process-invariant feature, represents the feature encoder, represents the th input sample image, represents the encoder-specific parameter;

[0013] Branch two outputs process-sensitive features, including representing texture changes and local deformations; the process-sensitive feature extraction formula is:

[0014] ;

[0015] where, represents the process-sensitive feature, Indicates encoder-specific parameters.

[0016] In a preferred embodiment, the mutual information minimization constraint formula is:

[0017] ;

[0018] Wherein, represents the mutual information loss function, represents the mutual information calculation, represents the th sample's process condition label, represents the regularization strength coefficient.

[0019] In a preferred embodiment, in the real-time migration module, the formula for extracting invariant features of real-time image processes is:

[0020] ;

[0021] Wherein, represents the new input image of the current batch, represents the invariant features extracted from the new image , represents the feature encoder, represents the parameters of the encoder.

[0022] In a preferred embodiment, the specific calculation formula of the weighted nearest neighbor algorithm is:

[0023] ;

[0024] Wherein, represents the invariant features stored in the th entry of the adaptive knowledge base , represents the invariant features stored in the th storage of the adaptive knowledge base , represents the new feature and the feature in the adaptive knowledge base represents the squared Euclidean distance between them, represents a positive scaling parameter, represents the radial basis function, represents the value of all entries in the adaptive knowledge base represents the new sample feature and the th Features of each entry Normalized similarity between;

[0025] The expression of the initial detection parameter is:

[0026] ;

[0027] Among them, Represents the historical optimal detection parameter associated with the adaptive knowledge base In the th entry Related, Represents the initial detection parameter generated by the new sample .

[0028] In a preferred embodiment, the expression of the composite reward mechanism is:

[0029] ;

[0030] Among them, Represents the immediate reward at time , Represents the detection accuracy at time , Represents the precision reward weight coefficient, Represents the parameter at time , Represents the parameter at time , Represents the parameter fluctuation penalty weight, Represents whether manual intervention is required at time , Represents the manual intervention penalty weight;

[0031] The specific formula of the proximal policy optimization algorithm is:

[0032] ;

[0033] Among them, Represents the current policy to be optimized, Represents the copy of the policy before update, Represents the limit on the policy update amplitude, Represents the clipping range hyperparameter, Represents the estimated value of the advantage function at time .

[0034] In a preferred embodiment, the calculation formula of the adversarial domain adaptation algorithm is:

[0035] ;

[0036] Among them, Denote sampling historical data from the adaptive knowledge base Denote the detection data of the new batch Denote the feature encoder Denote the domain discriminator Denote the adversarial loss function;

[0037] If the discriminator accuracy > 85%, trigger incremental update; otherwise, do not trigger incremental update.

[0038] In a preferred embodiment, the high-confidence sample screening rule is:

[0039] ;

[0040] Wherein Denote the prediction result of the new batch of data The maximum probability value of the predicted label Denote the confidence threshold Denote the high-confidence sample set

[0041] In a preferred embodiment, the update formula of the knowledge distillation strategy is:

[0042] ;

[0043] Wherein Denote the frozen original encoder Denote the new encoder to be updated Denote the L2 distance;

[0044] The verification and knowledge evolution module updates the adaptive knowledge base The specific steps are as follows:

[0045] Inject the newly generated high-quality feature-label pairs Into the adaptive knowledge base And perform clustering compression on the adaptive knowledge base Merge similar features and only retain the representative cluster centers to complete the evolution of the adaptive knowledge base

[0046] This application also provides a processing method for a flexible circuit board patch abnormality detection system based on multi-modal data, specifically including the following steps:

[0047] Step S1: Based on historical images, defect labels, and process parameters, separate process-invariant features and sensitive features through a feature decoupling network, and construct a dynamic clustering knowledge base ;

[0048] ​​Step S2: Using the dynamic clustering knowledge base generated in Step S1 , extract the process-invariant features of the real-time image to perform weighted similarity matching, and fuse the historical optimal parameters to generate the initial detection parameters of the new sample;

[0049] Step S3: Starting from the initial parameters output in Step S2, construct a meta-reinforcement learning strategy in combination with real-time feedback, dynamically optimize the detection parameters, and output the detection data with labels;

[0050] Step S4: Analyze the distribution drift of the detection data generated in Step S3. When a significant change is detected, screen the high-confidence samples, update the network parameters through distillation constraints, and inject the new knowledge into the dynamic clustering knowledge base .

[0051] The beneficial effects of the present invention are as follows: The present invention realizes intelligent anomaly detection in a flexible manufacturing scenario through multi-module collaboration: The adaptive knowledge base construction module decouples the inter-batch difference features, laying a stable foundation for anti-interference; The real-time migration module realizes instantaneous parameter migration by means of feature similarity, significantly improving the response efficiency of the new batch; The dynamic optimization module models the parameter adjustment as a reinforcement learning process, synchronously optimizing the detection accuracy and system robustness; The verification and knowledge evolution module perceives the distribution drift through an adversarial mechanism, safely injects new knowledge and avoids the risk of historical forgetting, and finally forms a lifelong learning architecture from feature decoupling to closed-loop evolution, achieving a leapfrog breakthrough in dimensions such as inter-batch adaptive time-consuming, manual maintenance frequency, and small-sample generalization ability, effectively supporting continuous and reliable detection in a complex process environment. Description of the Drawings

[0052] Figure 1 is the flowchart of the method of the present invention;

[0053] Figure 2 is the block diagram of the system structure of the present invention. Detailed Embodiments

[0054] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0055] In the description of the present application, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of the described features. In the description of the present application, "a plurality" means two or more, unless otherwise specifically defined.

[0056] In the description of the present application, the term "for example" is used to mean "serving as an example, illustration, or explanation". Any embodiment described as "for example" in this application is not necessarily construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to implement and use the present invention. In the following description, details are set forth for purposes of explanation. It should be understood that those of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other instances, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but rather to be in line with the broadest scope consistent with the principles and features disclosed in this application.

[0057] Embodiment 1

[0058] This embodiment provides a processing method for a flexible circuit board patch anomaly detection system based on multi-modal data as shown in Figure 1 the following, which specifically includes the following steps:

[0059] Step S1: Based on historical images, defect labels, and process parameters, separate process-invariant features and sensitive features through a feature decoupling network, and construct a dynamic clustering knowledge base ;

[0060] Step S2: Utilize the dynamic clustering knowledge base generated in Step S1 to extract the process-invariant features of the real-time image, perform weighted similarity matching, and fuse the historical optimal parameters to generate the initial detection parameters of the new sample;

[0061] Step S3: Starting from the initial parameters output in Step S2, construct a meta-reinforcement learning strategy in combination with real-time feedback, dynamically optimize the detection parameters, and output the detection data with labels;

[0062] Step S4: Analyze the distribution drift of the detection data generated in Step S3. When a significant change is detected, screen the high-confidence samples, update the network parameters through distillation constraints, and inject the new knowledge into the dynamic clustering knowledge base .

[0063] Embodiment 2

[0064] This embodiment provides a flexible circuit board patch anomaly detection system based on multi-modal data as shown in Figure 2 the following, which specifically includes: an adaptive knowledge base construction module, a real-time migration module, a dynamic optimization module, and a verification and knowledge evolution module;

[0065] Adaptive Knowledge Base Construction Module: Based on the historical batch of circuit board image sets, defect labels, and process parameters, a feature decoupling network is used to extract process-invariant features and process-sensitive features, and a mutual information minimization constraint is imposed to achieve feature separation; the adaptive knowledge base for dynamic clustering is constructed using the process-invariant features and corresponding defect labels , to complete the initial knowledge reserve. This module is designed to decouple the root features of inter-batch differences for the first time, provide a stability basis for subsequent steps, and avoid noise interference in the migration process;

[0066] Real-time Migration Module: For the real-time image of the current batch, extract its process-invariant features; based on the weighted nearest neighbor algorithm, match the historical optimal detection parameters in the constructed adaptive knowledge base ; fuse the matching parameters according to the similarity weight to generate the initial detection parameters for the current batch. This module is designed to achieve zero-shot migration using feature similarity, with a response time < 50ms, significantly faster than traditional transfer learning fine-tuning (which takes minutes);

[0067] Dynamic Optimization Module: Starting from the initial detection parameters generated by the real-time migration module, the detection confidence and parameter fluctuation status are analyzed in real time through a meta-policy network; based on a composite reward mechanism (including detection quality reward, parameter stability penalty, and manual intervention penalty), the proximal policy optimization algorithm is used to dynamically adjust the detection parameters and output the optimized defect detection results. This module is designed to model the parameter adjustment as a continuous decision-making process and synchronously optimize the detection quality, system stability, and automation level through triple reward design;

[0068] Verification and Knowledge Evolution Module: Based on the current batch detection results of the dynamic optimization module, calculate the feature distribution difference between it and the adaptive knowledge base through the adversarial domain adaptation algorithm; when the difference exceeds the threshold, screen high-confidence samples and use the knowledge distillation strategy to update the feature decoupling network; inject the verified new feature-label pairs into the adaptive knowledge base , perform the update of the adaptive knowledge base , in this module, the "adversarial verification-distillation update" mechanism is first proposed, and the accuracy of knowledge base evolution is increased by 40%, while avoiding catastrophic forgetting

[0069] In this embodiment, specifically, it should be noted that in the adaptive knowledge base construction module, the feature decoupling network uses a dual-branch structure to extract features. Branch one outputs process-invariant features, including the characterization of solder joint shape and component position. The process-invariant features represent the key object characteristics in the image that are independent of production process conditions (such as equipment parameters, speed, temperature); for example, what shape a solder joint should be and where a component should be placed. These features should be stable and reliable and are the basis for judging the quality of the product itself. The formula for extracting process-invariant features is:

[0070] ;

[0071] Among them, represents process-invariant features, represents a feature encoder, and a neural network is preferably used in the application, represents the th input sample image, represents the encoder-specific parameter, which is used to control the extraction ability of process-invariant features. The function of this formula is: extract the core features (such as component position, solder joint shape) that are independent of process fluctuations from the input image for stable detection;

[0072] Branch two outputs process-sensitive features, including texture changes and local deformations. Process-sensitive features represent fluctuations or interferences in the image caused by changes in production process conditions; for example, texture differences caused by different welding temperatures, local deformations caused by slight vibrations or mechanical stresses, and light changes. These features reflect process fluctuations, but do not directly help in judging the inherent quality of the product (whether the target appears, the position is correct, the shape meets the standard), and may even cause interference; the formula for extracting process-sensitive features is:

[0073] ;

[0074] Among them, represents process-sensitive features, represents the encoder-specific parameter, which is used to control the extraction ability of process-sensitive features. The function of this formula is: extract noise features (such as texture changes, local deformations) affected by the process and separate them from process-invariant features to reduce interference;

[0075] Through the above decoupling, the following effects can be achieved:

[0076] S1. Improve robustness: mainly use for target detection, positioning or quality determination, and reduce the interference caused by even if the production process fluctuates;

[0077] S2. Better locate problems: observing helps to identify process fluctuations;

[0078] S3. Knowledge accumulation: store reliable into the adaptive knowledge base for subsequent query and comparison;

[0079] The formula for minimizing mutual information constraint is:

[0080] ;

[0081] Among them, represents the mutual information loss function, represents the mutual information calculation, represents the process condition label of the th sample, such as temperature range, machine number,

[0082] The adaptive knowledge base is constructed as:

[0083] ;

[0084] Among them, represents the adaptive knowledge base, represents the process invariant feature of the th sample, represents the quality label of the th sample, such as qualified or unqualified. The function of this formula is to store the process invariant features and their quality labels

[0085] of historical samples, providing a prototype library for online matching.

[0086] ;

[0087] Among them, represents the new input image of the current batch, such as a real-time captured circuit board image, represents the process invariant features extracted from the new image such as component position, solder joint shape, represents the feature encoder, which is the same encoder as the one extracted in the construction module of the adaptive knowledge base represents the parameters of the encoder, inherited from the parameters learned in the construction module of the adaptive knowledge base and fixed during the inference stage, only updated during fine-tuning. The function of this formula is to extract the process invariant features of new samples as the matching basis;

[0088] The specific calculation formula of the weighted nearest neighbor algorithm is:

[0089] ;

[0090] Among them, represents the adaptive knowledge base in the The process-invariant features stored in an entry represent the adaptive knowledge base the th stored process-invariant feature Traverse all entries in the adaptive knowledge base among all entries represent the new feature and the adaptive knowledge base the features in the squared Euclidean distance between them, measuring the degree of difference between the two represents a positive scaling parameter, also known as a hyperparameter. A larger will amplify the distance difference, making the most similar neighbors obtain higher weights represents the radial basis function. The smaller the distance the smaller the larger, indicating a higher degree of similarity represents the sum of the values of all entries in the adaptive knowledge base among all entries of this is the denominator part of the normalization, ensuring that the sum of all is 1 represents the feature of the new sample and the adaptive knowledge base the th entry the normalized similarity between the features, which is a value between 0 and 1 and satisfies , is the probability weight or contribution coefficient for the feature to be selected as a "neighbor" of the new sample. The neighbor with the smallest distance has the highest ;

[0091] The expression of the initial detection parameter is as follows:

[0092] ;

[0093] where represents the historical optimal detection parameter associated with the th entry in the adaptive knowledge base. These parameters are obtained through some optimization (such as the training of the object detection head or the best verification effect after fine-tuning) when processing samples similar to this entry (possibly training samples or verified inference samples) before, represents the initial detection parameter generated by the new sample which is the adaptive knowledge base represents the new sample generated, and it is the adaptive knowledge base The historically optimal parameters saved for all entries is the weighted average of;

[0094] The similarity weights calculated by this formula using the weighted nearest neighbor algorithm , for the new sample of the process invariant features and the adaptive knowledge base the features of the most similar historical samples in are associated, and the detection parameters with the best effects under their respective conditions corresponding to these historical samples are fused according to the similarity weights, so as to obtain the strong initial parameters suitable for the new sample , which is the core of "zero-shot transfer": reuse the historically optimal configuration according to feature similarity. In this embodiment, specifically, the dynamic optimization module is designed with a meta-policy network

[0095] , where the state , , , such as threshold increase or decrease;

[0096] The expression of the composite reward mechanism is:

[0097] ;

[0098] Among them, represents the immediate reward at time , represents the detection accuracy at time , represents the accuracy reward weight coefficient, represents the parameter at time , represents the parameter at time , represents the parameter fluctuation penalty weight, represents whether manual intervention is required at time , represents the manual intervention penalty weight;

[0099] The specific formula of the proximal policy optimization algorithm is:

[0100] ;

[0101] Among them, represents the current policy to be optimized, represents the copy of the policy before update, represents the limit on the policy update amplitude, represents the clipping range hyperparameter, usually with a value range of 0.1 - 0.3, The estimated value of the advantage function at a certain moment, which measures how much better an action is than the average: ; ;

[0102] The optimization objective of the proximal policy optimization algorithm is:

[0103] Indicates a good action, that is, restricting the growth of the policy not to exceed ;

[0104] If Indicates a bad action, that is, restricting the decline of the policy not to be lower than .

[0105] In this embodiment, specifically, for the verification and knowledge evolution module, the calculation formula of the adversarial domain adaptation algorithm is:

[0106] ;

[0107] Among them, Represents sampling historical data from the adaptive knowledge base , Represents a new batch of detection data, Represents the feature encoder (the dual-branch encoder of the shared adaptive knowledge base construction module), Represents the domain discriminator (a binary classification network that distinguishes "historical / new data features"), Represents the adversarial loss function. In addition, , ;

[0108] If the discriminator accuracy > 85%, an incremental update is triggered; otherwise, an incremental update is not triggered;

[0109] The high-confidence sample screening rule is:

[0110] ;

[0111] Among them, Represents the prediction result of the new batch of data, The maximum probability value (confidence) of the predicted label, Represents the confidence threshold, and its value range is 0.9 - 0.95, Represents the high-confidence sample set, and its screening logic is: only select samples with a confidence > to participate in the update to avoid contamination by incorrect knowledge;

[0112] The update formula of the knowledge distillation strategy is:

[0113] ;

[0114] Among them, represents the frozen original encoder, represents the new encoder to be updated, represents the L2 distance, that is, the feature difference metric;

[0115] The verification and knowledge evolution module updates the adaptive knowledge base The specific steps are as follows:

[0116] Inject the newly generated high-quality feature-label pairs into the adaptive knowledge base , and perform clustering compression on the adaptive knowledge base , merge similar features, and only retain the representative clustering centers to complete the evolution of the adaptive knowledge base .

[0117] It should be noted that in the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0118] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0119] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded computers, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for realizing the specified functions in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0120] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including instruction means, and the instruction means realizes the process described in the process Figure 1One or more processes and / or blocks Figure 1 The functions specified in one or more blocks.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 One or more processes and / or blocks Figure 1 The steps of the functions specified in one or more blocks.

[0122] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0123] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A flexible circuit board patch abnormality detection system based on multimodal data, characterized in that Specifically, it includes: An adaptive knowledge base construction module, a real-time migration module, a dynamic optimization module, and a verification and knowledge evolution module; Adaptive knowledge base construction module: Based on the historical batch of circuit board image sets, defect labels, and process parameters, use a feature decoupling network to extract process-invariant features and process-sensitive features, and apply the minimum mutual information constraint to achieve feature separation; use the process-invariant features and corresponding defect labels to construct an adaptive knowledge base with dynamic clustering , to complete the initial knowledge reserve; Real-time migration module: Extract the process-invariant features of the current batch of real-time images; Based on the weighted nearest neighbor algorithm, match the historical optimal detection parameters in the constructed adaptive knowledge base Match the historical optimal detection parameters in the constructed adaptive knowledge base; fuse the matching parameters according to the similarity weights to generate the initial detection parameters for the current batch; Dynamic optimization module: Starting from the initial detection parameters generated by the real-time migration module, analyze the detection confidence and parameter fluctuation status in real time through a meta-policy network; Based on a composite reward mechanism, adopt the proximal policy optimization algorithm to dynamically adjust the detection parameters and output the optimized defect detection results; Verification and Knowledge Evolution Module: Based on the current batch detection results of the dynamic optimization module, calculate the difference in feature distributions between it and the adaptive knowledge base through the adversarial domain adaptation algorithm ; when the difference exceeds the threshold, screen high-confidence samples and use the knowledge distillation strategy to update the feature decoupling network; inject the verified new feature-label pairs into the adaptive knowledge base , and perform the update of the adaptive knowledge base .

2. The flexible circuit board patch abnormality detection system based on multi-modal data according to claim 1, characterized in that: In the adaptive knowledge base construction module, the feature decoupling network uses a dual-branch structure to extract features. Branch 1 outputs the process-invariant features representing the shape of the solder joint and the position of the component; Branch 2 outputs the process-sensitive features representing texture changes and local deformations.

3. The flexible circuit board patch abnormality detection system based on multimodal data according to claim 2, wherein: The mutual information minimization constraint aims to maximize the independence between the process-invariant features and the process condition labels, and at the same time minimize the independence between the process-sensitive features and the process condition labels.

4. The flexible circuit board patch abnormality detection system based on multimodal data according to claim 3, wherein: In the real-time migration module, the process-invariant features of the current batch of real-time images are obtained through the branch responsible for extracting process-invariant features in the feature decoupling network.

5. The flexible circuit board patch abnormality detection system based on multimodal data according to claim 4, characterized in that: The weighted nearest neighbor algorithm calculates the normalized similarity weights using the radial basis function based on the Euclidean distance between the process-invariant features of the new sample and each feature entry in the adaptive knowledge base KB; the initial detection parameters are obtained by weighted fusion of the matched historical optimal detection parameters according to the similarity weights.

6. The flexible circuit board patch abnormality detection system based on multi-modal data according to claim 5, wherein: The composite reward mechanism includes a reward item determined by the detection accuracy, a penalty item determined by the parameter fluctuation range, and a penalty item determined by triggering manual intervention; when updating the parameters of the policy network, the proximal policy optimization algorithm uses gradient clipping technology to limit the amplitude of policy update.

7. The flexible circuit board patch abnormality detection system based on multimodal data according to claim 6, characterized in that: The adversarial domain adaptation algorithm trains a domain discriminator to distinguish whether the features come from the historical knowledge base or the new batch of data, and calculates the adversarial loss accordingly; if the classification accuracy of the discriminator for the new and old data exceeds the preset threshold, incremental update is triggered.

8. The flexible circuit board patch abnormality detection system based on multimodal data according to claim 7, wherein: The high-confidence sample screening rule is based on whether the maximum class probability value in the prediction results of the model for the new batch of data exceeds the preset confidence threshold.

9. The flexible circuit board patch abnormality detection system based on multi-modal data according to claim 8, characterized in that: The knowledge distillation strategy updates the network parameters by minimizing the distance between the features extracted by the feature decoupling network for high-confidence samples before and after update; when updating the adaptive knowledge base KB, add the corresponding new process-invariant features and their prediction labels of the high-confidence samples to the knowledge base, and perform clustering compression to merge similar features and only retain the representative clustering centers.

10. The processing method of the flexible circuit board patch abnormality detection system based on multimodal data according to any one of claims 1-9, characterized in that: Specifically, it includes the following steps: Step S1: Based on historical images, defect labels, and process parameters, separate process-invariant features and sensitive features through a feature decoupling network, and construct a dynamic clustering knowledge base ; Step S2: Using the dynamic clustering knowledge base generated in Step S1 , extract the process-invariant features of the real-time image to perform weighted similarity matching, and fuse the historical optimal parameters to generate the initial detection parameters of the new sample; Step S3: Starting from the initial parameters output in Step S2, construct a meta-reinforcement learning strategy in combination with real-time feedback, dynamically optimize the detection parameters and output the detection data with labels; Step S4: Analyze the distribution drift of the detection data generated in step S3. When a significant change is detected, filter out high-confidence samples, update the network parameters through distillation constraints, and inject new knowledge into the dynamic clustering knowledge base .

Images