A semi-supervised continual learning PCBA target detection method based on improved YOLO11
By improving the YOLO11 object detection network and combining the HGBlock module, GhostConv module, teacher-student semi-supervised learning, and EWC incremental learning, the accuracy and stability issues of small-sized, dense object detection in complex PCBA scenarios were solved. This enabled continuous model learning and effective decomposition planning of detection results, improving the practicality and reliability of automated PCBA decomposition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG JIANZHU UNIVERSITY
- Filing Date
- 2026-02-14
- Publication Date
- 2026-06-05
AI Technical Summary
In complex PCBA scenarios, existing technologies struggle to effectively improve the ability of component inspection models to represent small-sized, densely packed target features. Furthermore, the models are prone to performance degradation and catastrophic forgetting during continuous updates, resulting in insufficient inspection accuracy and stability.
An improved YOLO11 object detection network is adopted, which combines the HGBlock module, GhostConv module, teacher-student semi-supervised learning mechanism and EWC incremental learning loss function. Through multi-scale feature enhancement, lightweight convolutional structure, semi-supervised learning and continuous learning strategies, the model's feature representation ability for small-sized and dense targets is improved, and the risk of forgetting during model update is reduced.
It improves the accuracy and stability of PCBA component testing, enhances the model's adaptability in complex industrial environments, realizes the effective transformation of testing results into dismantling plans, and improves the practicality and reliability of automated PCBA dismantling and intelligent recycling of electronic waste.
Smart Images
Figure CN122156736A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and industrial target detection technology, and in particular to a semi-supervised continuous learning PCBA target detection method based on improved YOLO11. Background Technology
[0002] Over the past few decades, with the rapid development of electronic manufacturing and computer vision technologies, the integration and complexity of electronic products have continuously increased. Printed Circuit Board Assemblies (PCBAs) have been widely used in communication equipment, industrial control systems, smart terminals, and various electronic products. In the process of electronic manufacturing and maintenance, the inspection and disassembly of components on PCBAs are indispensable and crucial steps.
[0003] In existing industrial applications, the inspection and identification of PCBA components typically rely on manual visual inspection or traditional image processing methods. Although related technologies have been applied to some extent in actual production, with the continuous increase in the number of components and the continuous reduction in size, manual inspection methods have gradually revealed problems such as low efficiency, poor stability, and susceptibility to subjective experience; while traditional image processing methods based on rules and thresholds have limited adaptability in situations such as complex backgrounds, changes in lighting, and component occlusion.
[0004] With the development of computer vision technology, deep learning-based target detection methods have gradually become an important research direction for PCBA component identification and have been applied in fields such as industrial inspection, quality control, and automated production. Depending on the detection target and application scenario, PCBA component detection tasks can involve problems such as multi-category component identification, small-sized target detection, and accurate localization in densely distributed target scenarios.
[0005] PCBAs contain a wide variety of electronic components with significant size differences, dense arrangement, and complex structure, making them a typical scenario for detecting small and dense targets. In actual detection, the model not only needs to focus on the spatial location features of the components but also needs to effectively distinguish between different types of components with subtle differences in appearance. This places higher demands on the feature representation capabilities of the target detection network.
[0006] Furthermore, in PCBA testing applications, the number of labeled samples is usually limited due to labeling costs and data acquisition conditions, and a large amount of unlabeled data is difficult to fully utilize. At the same time, with the continuous introduction of new PCBA models, the types and layouts of components are constantly changing, and existing models are prone to forgetting existing knowledge during the update process, which affects the stability of the test results.
[0007] Therefore, in complex PCBA scenarios, how to improve the ability of component inspection models to express small-sized, dense target features, make full use of limited labeled data, and maintain the stability of inspection performance during continuous model updates remains an urgent problem to be solved in the current field of PCBA component inspection. Summary of the Invention
[0008] To address the issues of insufficient component detection accuracy, low utilization of labeled data, and performance degradation of models during continuous updates in the aforementioned existing technologies in complex PCBA scenarios, this invention provides a semi-supervised continuous learning PCBA target detection method based on improved YOLO11. This method can improve the accuracy and stability of PCBA component detection in complex industrial environments and generate executable disassembly guidance results.
[0009] This invention discloses a semi-supervised continuous learning PCBA object detection method based on improved YOLO11, comprising: S1: Construct a PCBA component image dataset and perform partial annotation; S2: In the backbone network of the YOLO11 object detection network model, the standard Conv module is replaced with the HGBlock module, and the standard Conv module is replaced with the GhostConv module in the feature fusion network to obtain the improved YOLO11 object detection network model. S3: An improved YOLO11 object detection network model is trained using a teacher-student semi-supervised learning mechanism. High-confidence pseudo-labels are generated for unlabeled PCBA image samples, and the model is trained based on consistency constraints. S4: Based on the Fisher information matrix, the importance weights of the model parameters are calculated. The EWC incremental learning loss function is used as the loss function for incremental training of the improved YOLO11 object detection network model to constrain the historical learning parameters, so as to achieve continuous learning of PCBA component categories and reduce catastrophic forgetting. S5: Acquire the PCBA image to be detected, and perform component target detection based on the trained and improved YOLO11 target detection network model, outputting the spatial relationship of the components, including category labels, bounding box position information and corresponding detection confidence. S6: Based on the spatial relationship of PCBA components obtained from the detection and the preset disassembly rules, construct the component disassembly priority relationship and generate the component disassembly sequence that satisfies the disassembly constraints; S7: Combining the retrieval enhancement generation method, relevant knowledge retrieval and text generation are performed on the disassembly sequence to output structured PCBA disassembly guidance results.
[0010] Further, step S1 includes: Construct a PCBA component image dataset, and perform component feature segmentation and annotation on the PCBA images in the PCBA component image dataset; Place the obtained PCBA image and the labeled image under the imgs module and the labels module, respectively.
[0011] Furthermore, step S2 specifically includes the following steps: Modify the configuration file of the YOLO11 object detection network model. In yolo11.yaml, replace the standard Conv module in the backbone network with the HGBlock module, and replace the standard Conv module in the feature fusion network with the GhostConv module. Add the HGBlock and GhostConv modules to the module definition file corresponding to the YOLO11 object detection network model; Modify the YOLO11 object detection network model build file to support the parsing and calling of the newly added modules; Furthermore, step S2 also includes: The improved YOLO11 object detection network model was tested using test images to verify its performance.
[0012] Furthermore, the teacher-student semi-supervised learning mechanism in step S3 specifically involves: using the teacher model to detect the weakly enhanced unlabeled PCBA image and generating pseudo-labels with confidence levels higher than a preset threshold; and using the pseudo-labels to perform consistency constraint training on the corresponding strongly enhanced image samples. Specifically, the improved YOLO11 object detection network model was initialized as the teacher model and supervised warm-up training was performed on labeled PCBA images to obtain stable feature representation capabilities; then, the lighter-weight improved YOLO11 object detection network model was used as the student model to carry out semi-supervised optimization on unlabeled PCBA data. During training, the teacher model performs forward inference on weakly augmented input samples and generates high-confidence pseudo-labels when the prediction confidence is higher than a preset threshold. These pseudo-labels are used to supervise the student model's learning on the corresponding strongly augmented samples, thereby constructing a weak-strong consistency constraint mechanism. Meanwhile, to mitigate the instability caused by false label noise, the teacher model parameters... The student model is updated incrementally using an exponential moving average method, making its output a smoother and more reliable supervision signal; The loss function for the student model is defined as: ; in, This represents the overall optimization objective of the student model during the semi-supervised phase. This is a supervised loss calculated based on a small number of labeled samples, used to maintain the model's ability to discriminate on basic detection tasks; The consistency constraint loss is used to measure the difference between the prediction results of the student model under strongly augmented input and the high-confidence pseudo-labels generated by the teacher model under weakly augmented input. This represents the current set of parameters for the student model; This is the unsupervised weight coefficient, used to balance the relative importance of supervised learning and consistency constraints in the overall training process. Its value is gradually increased in the early stages of training to avoid the model being interfered with by unreliable pseudo-labels in the early stages.
[0013] Furthermore, the EWC incremental learning loss function in step S4 is expressed as follows: ; in, The overall optimization objective after introducing EWC constraints; This is the original loss function for the current task, used to drive the model to fit new data; This is the set of model parameters for the current incremental training phase. Indicates the first One parameter; The optimal parameters obtained after training the old task are the first one. One parameter; This is an approximation of the diagonal elements of the Fisher information matrix, used to characterize the first... The importance of each parameter to the old task A larger value indicates that the parameter is more critical to the performance of the old task, and therefore, deviations from it during incremental training should be avoided. The harsher the punishment; This is a hyperparameter used to balance the weighting of "learning new tasks" and "retaining old tasks". The larger the value, the stronger the retention of old knowledge and the more conservative the adaptation to new tasks; conversely, the smaller the value, the more conducive it is to learning new tasks but the risk of forgetting increases. This indicates that all parameter dimensions are accumulated, thereby achieving flexible constraints on the overall parameter space.
[0014] Further, step S6 includes: Based on the spatial relationship of each component in the component testing results and the preset disassembly and occlusion rules, a component disassembly priority relationship diagram is constructed. Based on the priority relationship diagram, the components are topologically sorted to obtain the component disassembly order.
[0015] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the semi-supervised continuous learning PCBA target detection method based on improved YOLO11.
[0016] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the semi-supervised continuous learning PCBA target detection method based on improved YOLO11.
[0017] Compared with the prior art, the present invention has at least the following beneficial effects: This invention addresses the shortcomings of existing target detection networks in industrial PCBA scenarios, such as insufficient ability to represent the features of small-sized, densely distributed components, and the heavy reliance of traditional supervised learning methods on large amounts of manually labeled data and the catastrophic forgetting that occurs during incremental learning. It proposes a YOLO11 target detection method that integrates the HGBlock module, the GhostConv module, a semi-supervised learning mechanism, and a continuous learning strategy. By synergistically improving the target detection network structure and training strategy, this method effectively enhances detection accuracy, generalization ability, and continuous learning capability in complex industrial scenarios while reducing model computational complexity.
[0018] This invention enhances the ability of the object detection network to represent the features of small-sized, densely distributed PCBA components by introducing the HGBlock multi-scale feature enhancement module and the GhostConv lightweight convolutional structure. It improves the utilization efficiency of unlabeled data by employing a teacher-student semi-supervised learning mechanism to generate high-confidence pseudo-labels for unlabeled samples and participate in joint model training. Furthermore, it reduces the risk of catastrophic forgetting by introducing a continuous learning mechanism to achieve incremental learning of new component categories and mitigate performance degradation during incremental updates. Combined with a retrieval enhancement generation method, it effectively transforms detection results into disassembly planning, improving the practicality and reliability of PCBA component detection and disassembly applications. This invention organically integrates object detection, semi-supervised continuous learning, disassembly relationship modeling, and disassembly scheme generation, enhancing the system's adaptability and scalability in complex PCBA scenarios and improving its engineering applicability and practical value in automated PCBA disassembly and intelligent electronic waste recycling applications.
[0019] Other beneficial effects of the present invention will be described in detail in the Detailed Description of the Embodiments section. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A schematic diagram of an improved YOLO11 target detection network model structure is disclosed in a preferred embodiment of the present invention.
[0022] Figure 2 This is a schematic diagram of the structure of the HGBlock module disclosed in a preferred embodiment of the present invention.
[0023] Figure 3 This is a schematic diagram of a semi-supervised learning training framework based on a teacher-student architecture, as disclosed in a preferred embodiment of the present invention.
[0024] Figure 4 This is a schematic diagram of the continuous learning mechanism based on Elastic Weight Consolidation (EWC) disclosed in a preferred embodiment of the present invention.
[0025] Figure 5 This is a schematic diagram of the target detection results of PCBA components disclosed in a preferred embodiment of the present invention.
[0026] Figure 6 This is an implementation state diagram of the trained YOLO11 object detection network model disclosed in a preferred embodiment of the present invention.
[0027] Figure 7 This is a diagram illustrating the operational process of a preferred embodiment of the present invention.
[0028] Figure 8 This is an implementation state diagram of the modified model disclosed in the preferred embodiment of the present invention.
[0029] Figure 9 This is a schematic diagram of the overall operation interface of the disassembly planning system based on enhanced PCBA component target detection and retrieval, as disclosed in a preferred embodiment of the present invention.
[0030] Figure 10 This is a schematic diagram of the generation of PCBA component disassembly sequence and resource evaluation results based on the detection results disclosed in a preferred embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0032] This invention discloses a semi-supervised continuous learning PCBA object detection method based on improved YOLO11, specifically including the following steps: S1: Construct a PCBA component image dataset and perform partial annotation; Step S1 specifically includes the following steps: Construct a PCBA component image dataset, and perform component feature segmentation and annotation on the PCBA images in the PCBA component image dataset; Place the obtained PCBA image and the labeled image under the imgs module and the labels module respectively; Specifically, see Figure 6 and Figure 8 By constructing a PCBA component image dataset and completing YOLO-format annotation, the dataset is divided into training, validation, and test sets. The dataset description file yolo11.yaml is configured to complete the setting of class definitions and image paths.
[0033] S2: Introduce the HGBlock module into the backbone network of the YOLO11 object detection network model, and replace the standard Conv module with the GhostConv module in the feature fusion network to obtain an improved YOLO11 object detection network model, such as... Figure 1 As shown; Step S2 specifically includes the following steps: Modify the configuration file of the YOLO11 object detection network model. In yolo11.yaml, replace the standard Conv module in the backbone network with the HGBlock module, and replace the standard Conv module in the feature fusion network with the GhostConv module. Add the HGBlock module and GhostConv module to the module definition file corresponding to the YOLO11 object detection network model. Modify the YOLO11 object detection network model build file to support the parsing and calling of the newly added modules; Specifically, the steps for replacing the standard Conv module in the backbone network of the YOLO11 object detection network model with the HGBlock module, and replacing the standard Conv module in the feature fusion network with the GhostConv module, are as follows: ① In the network structure configuration file yolo11.yaml, replace the original standard Conv module with the HGBlock module at the corresponding layer position of the backbone network, and keep the stride parameter unchanged; ② Implement the network structure of the GhostConv module and HGBlock module in the model module definition file, and complete the module registration in the nn / modules module file; ③ Modify the model parsing and construction logic, and add a parsing and registration mechanism for the new module type during the network construction process, so that the model construction process can identify and call the new module; During model training, the improved YOLO11 object detection network model is constructed by loading the modified yolo11.yaml network structure configuration file, and then trained to complete the PCBA component detection task.
[0034] Step S2 further includes: The improved YOLO11 object detection network model was tested using test images to verify its performance. S3: An improved YOLO11 object detection network model is trained using a teacher-student semi-supervised learning mechanism. High-confidence pseudo-labels are generated for unlabeled PCBA image samples, and the model is trained based on consistency constraints. The teacher-student semi-supervised learning mechanism in step S3 is as follows: the teacher model is used to detect the weakly enhanced unlabeled PCBA image and generate pseudo-labels with confidence higher than a preset threshold; the pseudo-labels are used to perform consistency constraint training on the corresponding image samples after strong enhancement. Specifically, the improved YOLO11 object detection network model was initialized as the teacher model and supervised warm-up training was performed on labeled PCBA images to obtain stable feature representation capabilities. Subsequently, the improved YOLO11 object detection network model with a lighter structure was used as the student model to carry out semi-supervised optimization on unlabeled PCBA data. During training, the teacher model performs forward inference on weakly augmented input samples and generates high-confidence pseudo-labels when the prediction confidence is higher than a preset threshold (e.g., 0.90). These pseudo-labels are used to supervise the student model's learning on the corresponding strongly augmented samples, thereby constructing a weak-strong consistency constraint mechanism. Meanwhile, to mitigate the instability caused by false label noise, the teacher model parameters are gradually updated by the student model using an exponential moving average method, making its output a smoother and more reliable supervision signal; The overall training objective of the student model consists of supervised loss and consistency loss, and its loss function is defined as: ; in, This represents the overall optimization objective of the student model during the semi-supervised phase. This is a supervised loss calculated based on a small number of labeled samples, used to maintain the model's ability to discriminate on basic detection tasks; The consistency constraint loss is used to measure the difference between the prediction results of the student model under strongly augmented input and the high-confidence pseudo-labels generated by the teacher model under weakly augmented input. This represents the current set of parameters for the student model; This is the unsupervised weight coefficient, used to balance the relative importance of supervised learning and consistency constraints in the overall training process. Its value is gradually increased in the early stages of training to avoid the model being interfered with by unreliable pseudo-labels in the early stages.
[0035] Through the above mechanism, the student model can fully explore the structural and semantic information in a large amount of unlabeled PCBA data while maintaining detection stability, thereby gradually improving the detection performance of small-sized, densely distributed components and enhancing the overall detection capability of the model. Semi-supervised learning training mechanisms based on a teacher-student architecture, such as... Figure 3 As shown; S4: Based on the Fisher information matrix, the importance weights of the model parameters are calculated. The EWC incremental learning loss function is used as the loss function for incremental training of the improved YOLO11 object detection network model to constrain historical learning parameters, thereby achieving continuous learning of PCBA component categories and reducing catastrophic forgetting. The continuous learning mechanism based on Elastic Weight Consolidation (EWC) is as follows: Figure 4 As shown; The EWC incremental learning loss function in step S4 is expressed as follows: ; in, The overall optimization objective after introducing EWC constraints; This is the original loss function for the current task, used to drive the model to fit new data; This is the set of model parameters for the current incremental training phase. Indicates the first One parameter; The optimal parameters obtained after training the old task are the first one. These parameters serve as anchors for existing knowledge. This is an approximation of the diagonal elements of the Fisher information matrix, used to characterize the first... The importance of each parameter to the old task A larger value indicates that the parameter is more critical to the performance of the old task, and therefore, deviations from it during incremental training should be avoided. The harsher the punishment; This is a hyperparameter used to balance the weighting of "learning new tasks" and "retaining old tasks". The larger the value, the stronger the retention of old knowledge and the more conservative the adaptation to new tasks; conversely, the smaller the value, the more conducive it is to learning new tasks but the risk of forgetting increases. This indicates that all parameter dimensions are accumulated, thereby achieving flexible constraints on the overall parameter space; S5: Obtain the PCBA image to be detected, and perform component target detection based on the improved YOLO11 target detection network model trained in steps S3 and S4, and output the spatial relationship of the components, including category labels, bounding box position information and corresponding detection confidence. Specifically, the PCBA image input during the identification and detection phase: see [link to relevant documentation] Figure 5 and Figure 7 The PCBA image to be detected is input into a trained, improved YOLO11 object detection network model for component object detection. The model first extracts multi-scale features from the input image through the backbone network, enhancing the feature representation capability of small-sized and densely distributed components. Then, a feature fusion network fuses and enhances feature maps at different scales. Finally, the detection head outputs the bounding box coordinates, class label, and corresponding confidence score for each candidate target. The detection results are then processed by non-maximum suppression to generate the final detection boxes, forming a structure as shown below. Figure 7 The image shown is a visualization of the detection results.
[0036] S6: Based on the detected spatial relationship of components and the preset disassembly rules, construct the component disassembly priority relationship and generate the component disassembly sequence that satisfies the disassembly constraints; PCBA image from disassembly planning input: See Figure 9 Based on the component detection results obtained from the improved YOLO11 target detection network model, including category labels, bounding box location information and detection confidence, a spatial and occlusion relationship model between components is constructed, and disassembly sequence planning results are generated.
[0037] The process involves constructing a disassembly priority relationship and generating a disassembly order: Based on the target detection results (component spatial relationships), the spatial relative positional relationships and overlap relationships between the bounding boxes of each component are extracted. Combined with preset disassembly occlusion rules and process constraints, the occlusion and dependency relationships between components are determined, thereby constructing a component disassembly priority relationship graph. This graph uses components as nodes and occlusion or dependency relationships as directed edges to describe the sequential constraints of different components during the disassembly process. Further, the disassembly priority relationship graph is subjected to topological sorting to generate a PCBA component disassembly order that satisfies the disassembly constraints, as shown in the figure. Figure 10 As shown.
[0038] S7: Combining the retrieval enhancement generation method, relevant knowledge retrieval and text generation are performed on the disassembly sequence to output structured PCBA disassembly guidance results.
[0039] Specifically, disassembly guidance results are generated based on the disassembly order: Based on the generated disassembly order, the Retrieval-Augmented Generation (RAG) method is used to retrieve disassembly operation procedures, safety specifications, and tool information corresponding to the target components from a pre-built disassembly knowledge base according to the disassembly order. The retrieval results are then fused, reasoned, and text-generated to form structured and executable PCBA component disassembly guidance results, which are used to guide subsequent automated or semi-automated disassembly operations.
[0040] By adopting the above disassembly planning method, a disassembly sequence of PCBA components that meets the disassembly constraints can be generated based on the target detection results, and structured disassembly guidance information can be output, thereby improving the feasibility and reliability of the PCBA automated disassembly process.
[0041] The target detection network model used in this invention employs the YOLO11 architecture. In the embodiments of this invention, the YOLO11 target detection network model used is not a single, fixed structure model, but rather a highly efficient detection network evolved from the YOLO series of single-stage target detection frameworks. The YOLO11 target detection network model includes an input module, a backbone network, a feature fusion network (Neck), and a head. These components work together to complete the tasks of target feature extraction, feature fusion, and target prediction.
[0042] This invention introduces the GhostConv module (lightweight convolutional structure) and the HGBlock module (multi-scale feature enhancement module) into the YOLO11 object detection network model to improve the network's ability to represent the features of small-sized, densely distributed components in complex industrial PCBA scenarios. By structurally improving the traditional convolutional feature extraction method, the network can more fully mine key features related to component detection while reducing computational complexity, thereby improving overall detection performance and model practicality.
[0043] (1) Input: The input is used to preprocess and augment the original input image, including normalizing the image size and augmenting the input sample by combining random scaling and random cropping to improve the model’s adaptability to targets of different scales.
[0044] (2) Backbone: The backbone is used to extract multi-level feature information from the input image and is the core component of the object detection network. The YOLO11 backbone extracts features from the input image step by step through a multi-layer convolutional structure, forming feature maps with different resolutions and semantic levels, so as to provide basic feature representations for subsequent feature fusion and object detection.
[0045] The backbone network of the improved YOLO11 target detection network model of this invention includes the following modules connected in sequence: HGStem module, first HGBlock module, first DWConv module, second HGBlock module, second DWConv module, third HGBlock module, fourth HGBlock module, fifth HGBlock module, third DWConv module, sixth HGBlock module, SPPF module, and PSA module.
[0046] The HGBlock module is a feature extraction module used to enhance multi-scale feature representation capabilities. Its structural diagram is shown below. Figure 2 As shown, its design goal is to solve the detection difficulties caused by the large differences in component sizes and the large number and dense distribution of small-sized targets in complex PCBA scenarios. In actual industrial inspection, single-scale features often cannot simultaneously take into account global semantic information and local spatial details, easily leading to the features of small-sized components being submerged by background information. The HGBlock module introduces a multi-branch feature extraction structure to process the input feature map in parallel at different receptive field scales and fuse features of different scales, thereby achieving effective aggregation of cross-scale feature information. In this way, the semantic information contained in high-level features can be fully combined with the spatial detail information in low-level features, enabling the network to significantly enhance its ability to capture local fine-grained features while maintaining global semantic understanding capabilities.
[0047] (3) Feature Fusion Network (Neck): The feature fusion network of the YOLO11 network is used to integrate features from different levels of the backbone network. Through the multi-scale feature fusion structure, the semantic information contained in the high-level features is effectively combined with the spatial detail information in the low-level features, thereby enhancing the network's ability to detect targets of different sizes and improving the overall stability and accuracy of detection.
[0048] The feature fusion network (Neck) in the improved YOLO11 target detection network model of this invention includes a first Upsample module, a first Concat module, a first C3K2 module, a second Upsample module, a second Concat module, a second C3K2 module, a first GhostConv module, a third Concat module, a third C3K2 module, a second GhostConv module, a fourth Concat module, and a fourth C3K2 module connected in sequence.
[0049] The multi-scale high-dimensional feature maps output by the backbone network are input into the feature fusion network and processed as follows: the output features of the PSA module in the backbone network are input into the first Concat module after passing through the first Upsample module. In the first Concat module, they are concatenated with the output features of the fifth HGBlock module. The features fused by the first Concat module first pass through the first C3K2 module, then through the second Upsample module, and then into the second Concat module. In the second Concat module, they are concatenated with the output features of the second HGBlock module. The features concatenated by the second Concat module are processed by the second C3K2 module and then input into the first GhostConv module and the detection head.
[0050] The output features of the second C3K2 module are input into the third Concat module after passing through the first GhostConv module. In the third Concat module, they are fused with the output features of the first C3K2 module. The features fused by the third Concat module are then processed by the third C3K2 module and input into the second GhostConv module and the detection head.
[0051] The output features of the third C3K2 module are input into the fourth Concat module after passing through the second GhostConv module. In the fourth Concat module, they are concatenated with the output features of the PSA module. The concatenated features in the fourth Concat module are then processed by the fourth C3K2 module and input into the detection head.
[0052] Traditional convolutional operations often require a large number of convolutional kernels to generate feature maps, and the features generated by different kernels generally exhibit strong redundancy, leading to a significant increase in model parameters and computational overhead. To address these issues, this invention introduces the GhostConv module. The core idea of GhostConv (GhostConvolution) is that it's a lightweight convolutional structure designed to reduce the computational complexity of convolutional neural networks. It divides the feature generation process into two stages using an "intrinsic feature generation + inexpensive linear transformation" approach. First, a set of representative intrinsic feature maps is generated using a small number of standard convolutional operations. Then, through low-cost operations such as depthwise convolution and linear transformation, more redundant feature maps are generated from the intrinsic feature maps to simulate the feature distribution produced by traditional convolution. This approach significantly reduces the number of model parameters and floating-point operations (FLOPs) while maintaining the network's effective representation of target features, thereby improving the model's deployment adaptability in real-time industrial detection tasks.
[0053] (4) Detection Head: The detection head is used to predict the target based on the fused features, outputting the target's category information and corresponding spatial location information. During target localization, bounding box regression is used to accurately predict the target location, and a loss function based on intersection-over-union (IoU) is introduced during the training phase to optimize the prediction results. In the post-processing stage, non-maximum suppression (NMS) is used to filter and suppress overlapping prediction boxes to remove redundant detection results and obtain the final detection output.
[0054] To address the common challenges in complex industrial PCBA scenarios, such as strong background interference, dense component distribution, and the ease with which small targets are missed, this invention introduces the HGBlock and GhostConv modules into the YOLO11 target detection network model. The HGBlock and GhostConv modules work synergistically as key components. The HGBlock module, through a multi-scale feature enhancement mechanism, assigns greater weight to key component features in the feature mapping, effectively mitigating the insensitivity of traditional target detection networks to feature differences in complex industrial PCBA scenarios. The GhostConv module primarily reduces redundant computation during feature extraction, improving model inference efficiency and running speed. By adaptively enhancing and filtering feature vectors, important features related to component detection occupy a larger proportion in the feature mapping, thereby improving the network's responsiveness to key target regions and reducing the probability of false positives and false negatives. The combination of these two modules achieves an effective balance between model complexity and detection performance while ensuring detection accuracy.
[0055] In summary, the semi-supervised continuous learning PCBA object detection method based on improved YOLO11 of the present invention includes two parts: model training process and system inference process.
[0056] During the model training phase, the constructed PCBA image dataset and its corresponding annotation information are input into the YOLO11 object detection network model improved by the HGBlock and GhostConv modules for training. Based on the teacher-student semi-supervised learning mechanism, high-confidence pseudo-labels are generated for unlabeled samples and participate in the joint training of the model. During the model update process, the Elastic Weight Consolidation (EWC) continuous learning mechanism is introduced to achieve incremental learning of newly added component categories and reduce the risk of catastrophic forgetting.
[0057] During the model inference phase, the trained network is used to perform object detection on the test images to verify the improvement in detection performance. The PCBA image to be detected is acquired and input into the improved YOLO11 object detection network model, which outputs the category and location information of the components. Based on the detection results, spatial occlusion and dependency relationships between components are constructed. Then, using the Retrieval-Augmented Generation (RAG) method, corresponding disassembly rules and safety constraint information are retrieved from the professional knowledge base to generate a component disassembly sequence and operation guidance result that meets the actual constraints.
[0058] To further verify the effectiveness of the structural improvement proposed in this invention, a comparative experiment was conducted between the YOLO11 object detection network model without the HGBlock and GhostConv modules (the basic YOLO11 model) and the improved YOLO11 object detection network model of this invention, under the same test dataset and experimental conditions. By comprehensively analyzing metrics such as precision, recall, mean average precision (mAP), and inference time, the differences in overall detection performance and small-sized component detection capabilities between the different models were evaluated. The experimental results are shown in Table 1.
[0059] The mean accuracy (mAP) is the average of the AP values for all target categories, used to measure the overall performance of the model in multi-class target detection tasks. This invention achieves a stable improvement in mAP without significantly increasing model complexity by introducing the HGBlock and GhostConv modules.
[0060] Recall R measures a model's ability to cover the true target, and is calculated as the ratio of the number of correctly predicted positive samples to the total number of actual positive samples. This invention improves the model while maintaining high accuracy, simultaneously enhancing recall, further validating the effectiveness and practicality of the proposed method in complex industrial PCBA scenarios.
[0061] Table 1 Comparison of Experimental Results
[0062] As shown in Table 1, compared with the YOLO11 target detection network model that does not introduce the HGBlock and GhostConv modules, the improved YOLO11 target detection network model of this invention has achieved significant improvements in key indicators such as accuracy, recall, and average detection precision. This indicates that the proposed structural improvement can effectively enhance the overall detection capability of multi-scale targets in complex PCBA scenarios while ensuring the computational efficiency of the model.
[0063] Furthermore, to analyze the differences in detection performance of the model across different categories of small-sized components, statistical analysis was performed on the detection accuracy of each category. The experimental results are shown in Table 2.
[0064] Table 2 Testing Capabilities of Small-Sized Components
[0065] As shown in Table 2, the improved model proposed in this invention has achieved high detection or classification accuracy in most small-sized component categories, indicating that the introduction of the HGBlock module and GhostConv module can effectively improve the model's ability to distinguish different types of small targets and reduce the problems of missed detection and false detection caused by the small size and dense distribution of targets.
[0066] The above description is only a partial embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent changes or substitutions made by those skilled in the art to the above method steps without departing from the technical concept of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A semi-supervised continuous learning PCBA object detection method based on improved YOLO11, characterized in that, include: S1: Construct a PCBA component image dataset and perform partial annotation; S2: In the backbone network of the YOLO11 object detection network model, the standard Conv module is replaced with the HGBlock module, and the standard Conv module is replaced with the GhostConv module in the feature fusion network to obtain the improved YOLO11 object detection network model. S3: An improved YOLO11 object detection network model is trained using a teacher-student semi-supervised learning mechanism. High-confidence pseudo-labels are generated for unlabeled PCBA image samples, and the model is trained based on consistency constraints. S4: Based on the Fisher information matrix, the importance weights of the model parameters are calculated. The EWC incremental learning loss function is used as the loss function for incremental training of the improved YOLO11 object detection network model to constrain the historical learning parameters, so as to achieve continuous learning of PCBA component categories and reduce catastrophic forgetting. S5: Acquire the PCBA image to be detected, and perform component target detection based on the trained and improved YOLO11 target detection network model, and output the spatial relationship of the components; S6: Based on the detected spatial relationship of components and the preset disassembly rules, construct the component disassembly priority relationship and generate the component disassembly sequence that satisfies the disassembly constraints; S7: Combining the retrieval enhancement generation method, relevant knowledge retrieval and text generation are performed on the disassembly sequence to output structured PCBA disassembly guidance results.
2. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, Step S1 includes: Perform component feature segmentation and annotation on PCBA images in the PCBA component image dataset; Place the obtained PCBA image and its corresponding label under the imgs module and the labels module, respectively.
3. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, Step S2 includes: Modify the configuration file of the YOLO11 object detection network model. In yolo11.yaml, replace the standard Conv module in the backbone network with the HGBlock module, and replace the standard Conv module in the feature fusion network with the GhostConv module. Add the HGBlock and GhostConv modules to the module definition file corresponding to the YOLO11 object detection network model; Modify the YOLO11 object detection network model build file to support the parsing and calling of newly added modules.
4. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, Step S2 further includes: The improved YOLO11 object detection network model was tested using test images to verify its performance.
5. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, Step S3 includes: The improved YOLO11 object detection network model was initialized as the teacher model and trained in a supervised manner on labeled PCBA images to obtain stable initial feature representation capabilities. Subsequently, the improved YOLO11 object detection network model with a lighter structure was used as the student model and semi-supervised optimization was carried out on unlabeled PCBA images. During training, the teacher model performs forward inference on weakly augmented input samples and generates high-confidence pseudo-labels when the prediction confidence is higher than a preset threshold. These labels are used to supervise the student model's learning on the corresponding strongly augmented samples, thereby constructing a weak-strong consistency constraint mechanism. Meanwhile, to mitigate the instability caused by false label noise, the teacher model parameters... The student model is updated incrementally using an exponential moving average method, making its output a smoother and more reliable supervision signal; The loss function for the student model is defined as: ； in, This represents the overall optimization objective of the student model during the semi-supervised phase. This is a supervised loss calculated based on a small number of labeled samples, used to maintain the model's ability to discriminate on basic detection tasks; The consistency constraint loss is used to measure the difference between the prediction results of the student model under strongly augmented input and the high-confidence pseudo-labels generated by the teacher model under weakly augmented input. This represents the current set of parameters for the student model; These are the unsupervised weighting coefficients.
6. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, The EWC incremental learning loss function in step S4 is expressed as follows: ； in, The overall optimization objective after introducing EWC constraints; This is the original loss function for the current task, used to drive the model to fit new data; This is the set of model parameters for the current incremental training phase. Indicates the first One parameter; The optimal parameters obtained after training the old task are the first one. One parameter; This is an approximation of the diagonal elements of the Fisher information matrix, used to characterize the first... The importance of each parameter to the old task A larger value indicates that the parameter is more critical to the performance of the old task, and therefore, deviations from it during incremental training should be avoided. The harsher the punishment; This is a hyperparameter used to balance the weighting of "learning new tasks" and "retaining old tasks". The larger the value, the stronger the retention of old knowledge and the more conservative the adaptation to new tasks; conversely, the smaller the value, the more conducive it is to learning new tasks but the risk of forgetting increases. This indicates that all parameter dimensions are accumulated, thereby achieving flexible constraints on the overall parameter space.
7. The semi-supervised continuous learning PCBA target detection method based on improved YOLO11 according to claim 1, characterized in that, Step S6 includes: Based on the spatial relationship of each component in the component testing results and the preset disassembly and occlusion rules, a component disassembly priority relationship diagram is constructed. Based on the priority relationship diagram, the components are topologically sorted to obtain the component disassembly order.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the semi-supervised continuous learning PCBA target detection method based on improved YOLO11 as described in any one of claims 1 to 7.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it is used to implement the semi-supervised continuous learning PCBA target detection method based on improved YOLO11 as described in any one of claims 1 to 7.