Machine vision intelligent detection control method based on 3D convolution and reinforcement learning

Abstract

The application relates to a machine vision intelligent detection control method based on 3D convolution and reinforcement learning, which comprises the following steps: acquiring image data of a welding area; performing feature extraction processing on an input data set based on a 3D convolutional neural network to generate corresponding multi-dimensional feature maps; constructing a multi-scale target detection model to generate target detection results corresponding to welds and defects; generating corresponding target contour information and spatial position information through a segmentation regression model; performing fusion processing on the target detection results, the target contour information and the spatial position information to generate corresponding fusion decision results; acquiring corresponding execution result data; constructing a feedback sample data set, and updating model parameters according to the feedback sample data set. The method realizes the collaborative improvement of recognition accuracy, control real-time performance and system self-adaptive capability in a complex industrial welding scene, and solves the problems of insufficient multi-scale recognition, missing spatial information, rigid control mode and lack of continuous optimization capability.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent detection and control in machine vision, and in particular to a machine vision intelligent detection and control method based on 3D convolution and reinforcement learning. Background Technology

[0002] Currently, in the production process of lithium-ion battery module PACKs, the welding quality between the battery terminals and connecting pieces directly affects the consistency and safety of the product. Therefore, machine vision technology is usually required to identify, track, and detect defects in the weld seams. Existing technologies commonly employ template matching, edge detection, and traditional machine learning methods. Template matching typically involves pre-constructing a weld seam template and matching it with a real-time acquired image to determine the weld seam location. However, this method is susceptible to changes in lighting conditions or fluctuations in weld seam morphology, leading to poor recognition stability. Edge detection methods use operators such as Canny and Sobel to extract edge information from the image to locate the weld seam area. However, in complex welding scenarios or when the contrast between the weld seam and the background is low, the edge information is not obvious, easily resulting in false positives or false negatives. Summary of the Invention

[0003] To address the problem of false detections or missed detections in complex welding scenarios, this application provides a machine vision intelligent detection and control method based on 3D convolution and reinforcement learning.

[0004] A machine vision intelligent detection and control method based on 3D convolution and reinforcement learning, comprising:

[0005] Acquire image data of the welding area and construct an input dataset based on the image data;

[0006] Based on a 3D convolutional neural network, feature extraction processing is performed on the input dataset to generate corresponding multidimensional feature maps;

[0007] Based on multi-dimensional feature maps, a multi-scale target detection model is constructed, and the target detection results of corresponding welds and defects are generated based on the multi-scale target detection model.

[0008] Based on multidimensional feature maps, a segmentation regression model is used to perform fine characterization processing on the detection region corresponding to the target detection result, generating corresponding target contour information and spatial location information.

[0009] The target detection results, target contour information, and spatial location information are fused to generate corresponding fusion decision results;

[0010] A reinforcement learning control model is constructed based on the fusion decision results to generate corresponding optimal welding control parameters. The welding actuator is then driven to perform welding actions based on the optimal welding control parameters, thereby obtaining the corresponding execution result data.

[0011] A feedback sample dataset is constructed based on the execution result data. The model parameters in the 3D convolutional neural network, multi-scale object detection model, and segmentation regression model are updated based on the feedback sample dataset.

[0012] By adopting the above technical solutions, and by constructing a complete processing chain from image data acquisition, 3D feature extraction, multi-scale detection, fine segmentation and localization, fusion decision-making, reinforcement learning control and feedback optimization, weld recognition, defect detection and welding control form a closed-loop collaboration, thereby achieving high-precision recognition, real-time response and adaptive optimization capabilities in complex welding scenarios, effectively improving welding quality stability and system intelligence.

[0013] Preferably, the step of acquiring image data of the welding area and constructing an input dataset based on the image data includes:

[0014] Based on an industrial camera, images of the welding area are acquired and processed to generate corresponding raw image data.

[0015] Based on the pose sensor, the spatial pose of the target corresponding to the welding area is collected and processed to generate the corresponding original pose data.

[0016] The original image data and original pose data are preprocessed to generate corresponding preprocessed data. The preprocessing operations include at least image processing operations such as denoising and contrast enhancement of the original image data, and outlier removal and coordinate transformation of the original pose data.

[0017] Based on preset normalization rules, the preprocessed data is standardized to generate corresponding standardized data.

[0018] The various standardizations are cached and integrated to generate the corresponding input dataset.

[0019] By adopting the above technical solution, and by introducing multi-source data acquisition methods from industrial cameras and pose sensors, combined with denoising, enhancement and normalization processing, the input data is unified in terms of spatial information integrity and numerical consistency, thereby significantly reducing the impact of environmental interference on the recognition results and improving the stability and robustness of subsequent model processing.

[0020] Preferably, the step of performing feature extraction processing on the input dataset based on a 3D convolutional neural network to generate the corresponding multidimensional feature map includes:

[0021] The input dataset is parsed and processed using the established data receiving interface, and the corresponding parsed data is output.

[0022] Based on the queue mechanism, the parsed data is processed in an ordered manner to generate the corresponding data to be processed.

[0023] The data to be processed is input into a 3D convolutional neural network for multi-layer convolution operations to generate corresponding intermediate feature data.

[0024] The intermediate feature data is processed by pooling to generate the corresponding dimensionality-reduced feature data.

[0025] Multi-level feature mapping is performed on the dimensionality-reduced feature data to generate corresponding multi-dimensional feature maps, which include at least spatial structure information and semantic feature information.

[0026] By adopting the above technical solution, constructing a data parsing and queue scheduling mechanism, and combining it with the multi-layer feature extraction structure of a three-dimensional convolutional network, the input data can be processed in an orderly manner and expressed in a deep feature dimension in both temporal and spatial dimensions, thereby improving the ability to capture complex weld structures and subtle defect features.

[0027] Preferably, the step of inputting the data to be processed into a 3D convolutional neural network for multi-layer convolution operations to generate corresponding intermediate feature data includes:

[0028] Based on the preset 3D convolution kernel weights and bias parameters in the 3D convolutional neural network, a 3D convolution operation is performed on the data to be processed to generate corresponding intermediate feature data. The 3D convolution operation satisfies the following relationship:

[0029] F = Conv3D(I; W, b)

[0030] Where I represents the data to be processed, W represents the 3D convolution kernel weights, b represents the bias parameter, Conv3D represents the 3D convolution operation, and F represents the intermediate feature data.

[0031] By adopting the above technical solution and clarifying the relationship between the kernel weights and bias parameters in the three-dimensional convolution operation, the feature extraction process has a quantifiable computational basis, thereby enhancing the consistency and controllability of feature expression and improving the accuracy of the model in modeling spatial structural information.

[0032] Preferably, the step of constructing a multi-scale target detection model based on a multi-dimensional feature map, and generating corresponding target detection results for welds and defects based on the multi-scale target detection model, includes:

[0033] Multi-dimensional feature maps are used to perform multi-level feature decomposition to generate corresponding feature maps at multiple scales.

[0034] Based on feature maps at multiple scales, a feature pyramid network is constructed, and the feature pyramid network is used to perform top-down and horizontal fusion processing on the feature maps at each scale to generate corresponding multi-scale fused feature maps.

[0035] An attention mechanism region proposal network is constructed based on the multi-scale fused feature map, and feature enhancement processing is performed on the multi-scale fused feature map based on the attention mechanism region proposal network to generate the corresponding enhanced feature map.

[0036] Based on the enhanced feature maps and the output of the feature pyramid network, corresponding object detection proposals at various scales are generated. The object detection proposals at each scale satisfy the following relationship:

[0037] Pi=Attention-RPN(Fi)+FPN(Fi);

[0038] Fi represents the feature map of the i-th layer, Attention-RPN represents the attention mechanism region proposal network, FPN represents the feature pyramid network, and Pi represents the object detection proposal at the i-th scale.

[0039] Target screening is performed based on target detection proposals at various scales to generate corresponding target detection results for welds and defects.

[0040] By adopting the above technical solution, and constructing a multi-scale detection structure that combines a feature pyramid network and an attention mechanism region proposal network, the model can simultaneously focus on targets of different scales and automatically focus on key regions, thereby improving the detection accuracy and completeness of welds and minor defects in complex backgrounds.

[0041] Preferably, the step of refining the detection region corresponding to the target detection result through a segmentation regression model based on a multi-dimensional feature map to generate corresponding target contour information and spatial location information includes:

[0042] Based on the target detection results, the corresponding detection region feature data is extracted from the multi-dimensional feature map to generate the corresponding region feature map.

[0043] The region feature map is input into the segmentation regression model to construct the segmentation branch and the coordinate regression branch;

[0044] Based on the segmentation branch, the region feature map is processed by convolution operation to generate the corresponding segmentation mask data;

[0045] Based on the coordinate regression branch, regression calculations are performed on the regional feature map to generate corresponding spatial coordinate data;

[0046] By jointly processing segmentation mask data and spatial coordinate data, corresponding target contour information and spatial location information are generated.

[0047] By adopting the above technical solution and introducing a combined processing method of segmentation branch and coordinate regression branch on the basis of detection, the system can not only obtain the precise contour of the target, but also obtain its spatial position information, thereby achieving high-precision positioning and morphological characterization of the weld target and meeting the requirements of refined welding control.

[0048] Preferably, the step of fusing the target detection results, target contour information, and spatial location information to generate the corresponding fusion decision result includes:

[0049] Determine the corresponding detection confidence parameters based on the target detection results;

[0050] Determine the corresponding contour accuracy parameters based on the target contour information;

[0051] Determine the corresponding coordinate error parameters based on spatial location information;

[0052] The detection confidence parameter, contour accuracy parameter, and coordinate error parameter only perform feature alignment processing without outputting the corresponding alignment feature data;

[0053] Based on preset fusion weights, each aligned feature data is weighted and fused to generate the corresponding fusion decision result.

[0054] By adopting the above technical solution, feature alignment and weighted fusion processing are performed on detection confidence, contour accuracy and coordinate error, so that information from multiple sources can be comprehensively evaluated in a unified decision space, thereby reducing the risk of misjudgment caused by single result bias and improving the reliability and stability of the overall decision results.

[0055] Preferably, the step of constructing a feedback sample dataset based on the execution result data includes:

[0056] Extract real sample data based on the execution results data;

[0057] Construct a generative adversarial network (GAN) with a least-squares mechanism. The GAN consists of a generator and a discriminator.

[0058] Real sample data is input into the generative adversarial network to train the generator and discriminator, and then to generate corresponding synthetic sample data.

[0059] Based on real sample data and synthetic sample data, a corresponding feedback sample dataset is generated.

[0060] By adopting the above technical solution and introducing generative adversarial networks to expand real samples, a feedback sample dataset containing real and synthetic data is constructed, thereby effectively improving data diversity in industrial scenarios where samples are scarce and providing sufficient data support for continuous model optimization.

[0061] Preferably, the step of inputting real sample data into a generative adversarial network to train the generator and discriminator, and then generating corresponding synthetic sample data, includes:

[0062] Input random noise data into the generator to generate corresponding preliminary sample data;

[0063] Based on the discriminator, the real sample data and the preliminary sample data are processed to generate the corresponding synthetic sample data;

[0064] The generation of synthetic sample data satisfies the following relationship:

[0065] minGmaxDV(D, G)=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1−D(G(z)))];

[0066] Where G is the generator, D is the discriminator, x is the real sample, Z is random noise, pdata(x) is the distribution of the real sample, pz(z) is the noise distribution, and V(D,G) is the GAN loss function.

[0067] By adopting the above technical solution and introducing an adversarial training mechanism into the generative adversarial network, the generator continuously approximates the real sample distribution, while the discriminator improves the sample discrimination ability, thereby enhancing the authenticity and effectiveness of the synthesized data and further improving the quality of the data augmentation process.

[0068] Preferably, the step of updating the model parameters in the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model based on the feedback sample dataset includes:

[0069] The feedback sample dataset is input into the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model respectively, so that each model generates the model prediction results of the base.

[0070] Determine the true annotation results based on the real sample data in the feedback sample dataset;

[0071] Based on the loss function, the difference features between the model prediction results and the ground truth labeling results are calculated. Then, gradient updates are performed on the model parameters of each model according to these difference features, generating updated model parameters. The updated model parameters satisfy the following relationship:

[0072] θ′=θ−η⋅∇θ L (y^, y);

[0073] Where θ represents the original model parameters, η represents the learning rate, y^ represents the model prediction result, y represents the ground truth labeling result, L(y^, y) represents the difference features of the loss function between the predicted y^ and the ground truth y, and θ represents the updated model parameters.

[0074] By adopting the above technical solution, and by jointly training and updating the parameters of each model based on the feedback sample dataset, the model can continuously correct the prediction deviation according to the actual execution results, thereby continuously improving the recognition accuracy and control effect, and realizing the adaptive optimization capability of the system in the actual operation process.

[0075] In summary, this application includes at least one of the following beneficial technical effects:

[0076] This application constructs a closed-loop processing mechanism consisting of spatial feature modeling, multi-scale target recognition, fine representation, fusion decision-making, adaptive control, and feedback re-optimization, transforming the previously separate processing of detection and control into an integrated linkage. Specifically, firstly, 3D convolution is used to extract spatial structure features from the input data, enabling the model to simultaneously perceive weld morphological changes and subtle defect features, thus overcoming the limitations of traditional methods that rely solely on two-dimensional information. Based on this, a multi-scale detection structure combined with an attention mechanism is used to uniformly model welds and defect targets of different sizes and focus on key areas, achieving stable recognition even in complex environments. Subsequently, through a joint processing method of segmentation and coordinate regression, the targets are finely represented at both the contour and position levels, elevating the detection results from coarse-grained judgments to precise spatial information usable for control. Furthermore, by fusing the detection results, contour information, and position information, a comprehensive evaluation of multi-source information is achieved within the same decision space, thereby improving the stability and reliability of the decision results. On this basis, a reinforcement learning mechanism is introduced to map the fused decision results to welding control parameters, and through feedback-based strategy optimization, the control process can adaptively adjust to changes in operating conditions, avoiding the limitations of manually set control strategies. Simultaneously, by feeding back the execution results and combining sample expansion and model updates, the system continuously corrects its recognition and control capabilities during actual operation, forming a continuously optimized closed-loop mechanism. This achieves a synergistic improvement in recognition accuracy, real-time control, and system adaptability in complex industrial welding scenarios, effectively solving the problems of insufficient multi-scale recognition, lack of spatial information, rigid control methods, and lack of continuous optimization capabilities in existing technologies. Attached Figure Description

[0077] Figure 1 This is a flowchart of a machine vision intelligent detection and control method based on 3D convolution and reinforcement learning in one embodiment of this application.

[0078] Figure 2This is a flowchart illustrating a specific process of a machine vision intelligent detection and control method based on 3D convolution and reinforcement learning in one embodiment of this application. Detailed Implementation

[0079] The present application will be further described in detail below with reference to the accompanying drawings.

[0080] In one embodiment, such as Figures 1-2 As shown, this application discloses a machine vision intelligent detection and control method based on 3D convolution and reinforcement learning, including:

[0081] S10. Obtain image data of the welding area and construct an input dataset based on the image data;

[0082] S20. Based on a 3D convolutional neural network, feature extraction processing is performed on the input dataset to generate corresponding multi-dimensional feature maps;

[0083] S30. Based on the multi-dimensional feature map, construct a multi-scale target detection model, and generate target detection results for corresponding welds and defects based on the multi-scale target detection model;

[0084] S40. Based on multi-dimensional feature maps, the detection region corresponding to the target detection result is finely characterized by a segmentation regression model to generate the corresponding target contour information and spatial location information.

[0085] S50. The target detection results, target contour information and spatial location information are fused to generate the corresponding fusion decision results;

[0086] S60. Construct a reinforcement learning control model based on the fusion decision results to generate the corresponding optimal welding control parameters. Drive the welding actuator to perform welding actions according to the optimal welding control parameters, and then obtain the corresponding execution result data.

[0087] S70. Construct a feedback sample dataset based on the execution result data, and update the model parameters in the 3D convolutional neural network, multi-scale object detection model, and segmentation regression model according to the feedback sample dataset.

[0088] In this embodiment, the input dataset is first constructed by acquiring image data of the welding area. The image data can be obtained by industrial cameras in real time from the welding process. Essentially, it is to convert the visual information of the actual welding site into a digital signal that can be processed by calculation. In this process, the input dataset can include not only single frame images, but also continuous frame data or time-related sequence information, so as to provide a basis for subsequent three-dimensional feature modeling, thereby enabling the system to reflect the dynamic change characteristics of the weld during the processing.

[0089] After obtaining the input dataset, a three-dimensional convolutional neural network is used to extract features from the input data. The three-dimensional convolution can be understood as a convolution operation that introduces depth or time dimensions on the basis of traditional two-dimensional convolution. This allows the model to extract spatial structural features of the image while also capturing the changing relationships between different frames, thereby generating a multi-dimensional feature map containing spatial structural information and semantic information. This multi-dimensional feature map is essentially a high-dimensional feature representation used to characterize the morphological features, texture distribution, and potential defect features of the weld area, providing a unified data foundation for subsequent identification and analysis.

[0090] When constructing a multi-scale target detection model based on the multi-dimensional feature map, the feature map is processed in layers to form feature representations of different resolutions. The model is then fused to achieve unified modeling of targets at different scales. This enables the system to identify both large-scale weld structures and small-sized defects. In this process, the multi-scale implementation can be understood as analyzing the same target at different "observation scales," thereby avoiding a decrease in recognition ability due to changes in target size. Finally, the corresponding target detection results for welds and defects are generated, which typically include the target's location and its corresponding confidence level.

[0091] After obtaining the target detection results, the detected area is further refined through a segmentation regression model. The segmentation process is used to divide the detected target area into pixels, that is, to determine whether each pixel belongs to the weld or defect area, thereby obtaining the precise contour information of the target. At the same time, the regression process is used to perform numerical calculations on the spatial position of the target, which can be understood as outputting the specific position parameters of the target in the image or spatial coordinate system. The combination of the two enables the system not only to determine the existence of the target, but also to accurately describe its shape and position, thereby meeting the requirements of high-precision welding control.

[0092] After obtaining target detection results, contour information, and spatial location information, a unified decision result is generated through fusion processing. This fusion process can be understood as a comprehensive evaluation of information from different sources. For example, by uniformly measuring the reliability of detection results, the accuracy of contour description, and the deviation of location calculation, multiple pieces of information can work together in the same decision space, thereby reducing the impact of inaccurate single information and improving the stability of the overall judgment.

[0093] After generating the fusion decision results, a reinforcement learning control model is introduced to adjust the welding process. Reinforcement learning can be understood as a feedback-based learning mechanism. The system learns better control methods by continuously trying different control strategies and obtaining evaluations based on the execution results. In this embodiment, the fusion decision results are used as the current state input. The model generates corresponding control parameters based on the state and performs actual actions through the welding actuator, so that the welding path or process parameters can be dynamically adjusted according to the detection results, thereby achieving adaptive optimization of the control process.

[0094] After the actuator completes its action, the system acquires execution result data and constructs a feedback sample dataset. This allows the system to record the relationship between the actual execution effect and the corresponding input. The feedback sample dataset not only reflects the real situation under the current working conditions but also provides a data foundation for subsequent model optimization. Furthermore, by using this feedback sample to update the parameters of the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model, the model can continuously correct its parameters based on the actual operating results, thereby gradually improving the recognition accuracy and control effect. Ultimately, this forms a closed-loop operating mechanism from data acquisition to model optimization, enabling the system to continuously improve itself during long-term operation.

[0095] Specifically, after obtaining the fusion decision result, it is used as the input for the control phase to transform the detection result into an execution action. Specifically, a reinforcement learning control model is first constructed based on the fusion decision result. The fusion decision result characterizes the current welding state and can serve as the state input in reinforcement learning, driving the model to evaluate the current working condition. During model construction, a deep Q-network structure is used. By estimating the value corresponding to different control actions, the system can select the optimal solution from multiple candidate control strategies. The control action can be understood as different combinations of motion control parameters of the welding actuator, such as motion path adjustment, speed change, or energy output regulation. During training, a feedback mechanism based on execution performance is introduced to evaluate various control actions, enabling the model to gradually learn the optimal control strategy under different states, thereby generating the corresponding optimal welding control parameters. The decision-making process can satisfy the relationship a_t=argmax_a Q(s_t,a;θ), where s_t represents the current state determined by the fusion decision result, a represents the candidate control action, Q(s_t,a;θ) represents the value evaluation of the combination of state and action under parameter θ, and a_t represents the finally selected optimal control action.

[0096] After generating the optimal welding control parameters, these parameters are sent to the welding actuator to drive the actual action. In a specific implementation, the welding actuator may include a servo motor or a robotic arm joint, etc., and a motion control module performs signal conversion and amplification processing on the control parameters, transforming them into drive signals to control the welding head to complete the welding operation according to a predetermined trajectory or process parameters. In this way, the system can directly map the front-end recognition and decision-making results into physical actions, achieving closed-loop linkage from perception to execution.

[0097] After the actuator completes the welding action, the execution results are collected by sensing units installed on the welding actuator to obtain data reflecting the current execution state. These sensing units may include position sensors and force sensors, used to acquire information such as the displacement accuracy of the welding head and the applied force during the welding process. The collected execution result data is used to characterize the difference between the actual execution effect and the expected control, and is transmitted back to the system as feedback data. Through the continuous accumulation and utilization of this feedback data, the system can evaluate and correct the control strategy, thereby providing data support for subsequent model optimization. Ultimately, a closed-loop control mechanism is formed, driven by the fused decision results and fed back by the execution result data, enabling the welding process to have dynamic adaptive adjustment capabilities.

[0098] Furthermore, the step of acquiring image data of the welding area and constructing the input dataset based on the image data includes:

[0099] S101. Based on an industrial camera, perform image acquisition and processing on the welding area to generate corresponding raw image data;

[0100] S102. Based on the pose sensor, the spatial pose of the target corresponding to the welding area is collected and processed to generate the corresponding original pose data.

[0101] S103. Perform preprocessing operations on the original image data and the original pose data to generate corresponding preprocessed data. The preprocessing operations include at least image processing operations such as denoising and contrast enhancement of the original image data, and outlier removal and coordinate transformation of the original pose data.

[0102] S104. Based on the preset normalization processing rules, the preprocessed data is standardized to generate the corresponding standardized data.

[0103] S105. The various standardizations are cached and integrated to generate the corresponding input dataset.

[0104] In this embodiment, the process of acquiring image data of the welding area and constructing the input dataset first involves using an industrial camera to acquire and process images of the welding area to obtain raw image data reflecting the welding process. Simultaneously, a pose sensor is used to synchronously acquire the spatial pose of the target object within the welding area. This spatial pose characterizes the target's position and orientation in space, thereby generating corresponding raw pose data. Through this method, the input data includes not only image information but also its corresponding spatial location information, thus providing a multi-dimensional data foundation for subsequent processing.

[0105] After acquiring the original image data and the original pose data, preprocessing operations are performed on both to eliminate the impact of environmental noise and data anomalies on subsequent analysis. Specifically, for the original image data, Gaussian filtering or median filtering is used for denoising to reduce random noise interference, and histogram equalization is used for contrast enhancement to improve the recognizability of the weld area in the image. For the original pose data, outlier removal is performed to eliminate abnormal points acquired during the acquisition process, and coordinate transformation is used to unify data from different sources to the same coordinate system, ensuring a consistent spatial reference standard. Through these preprocessing operations, corresponding preprocessed data is generated, unifying the image and pose information in terms of both data quality and representation.

[0106] After preprocessing, the preprocessed data is standardized based on a preset normalization rule. This standardization maps data with different dimensions and value ranges to a specified numerical interval, facilitating stable training and inference of the model. The standardization process satisfies the following relationship: Xnorm = (X − μ) / σ, where X represents the original image data or original pose data, μ represents the mean of the corresponding dataset, σ represents the standard deviation of the corresponding dataset, and Xnorm represents the standardized data. This process ensures consistency in numerical scale across data from different sources, thereby improving the model's efficiency in learning features.

[0107] After obtaining the standardized data, it is cached and integrated to construct the corresponding input dataset. Temporary storage can be achieved using local memory or a high-speed cache database to ensure data continuity and integrity during processing. Simultaneously, the standardized data is transmitted via industrial Ethernet or 5G communication to achieve low-latency data flow. Through this caching and transmission mechanism, the input dataset can be stably and promptly provided to subsequent processing stages, thus supporting the real-time performance and reliability of the entire system.

[0108] Furthermore, the step of performing feature extraction processing on the input dataset based on a 3D convolutional neural network to generate the corresponding multidimensional feature map includes:

[0109] S201. The input dataset is parsed and processed through the established data receiving interface, and the corresponding parsed data is output.

[0110] S202. Based on the queue mechanism, perform ordered Fanta processing on the parsed data to generate corresponding data to be processed.

[0111] S203. Input the data to be processed into a 3D convolutional neural network for multi-layer convolutional operations to generate corresponding intermediate feature data.

[0112] S204. Perform pooling operations on the intermediate feature data to generate the corresponding dimensionality-reduced feature data.

[0113] S205. Perform multi-layer feature mapping processing based on the dimensionality reduction feature data to generate corresponding multi-dimensional feature maps. The multi-dimensional feature maps include at least spatial structure information and semantic feature information.

[0114] In this embodiment, the process of feature extraction from the input dataset based on a 3D convolutional neural network first involves establishing a data receiving interface to parse the input dataset, adapting it to the unified processing requirements of different data types. The input dataset may include standardized image data and corresponding pose data. The data receiving interface identifies and parses the formats of various data types, transforming them into a unified data structure that can be processed by subsequent models, thereby generating corresponding parsed data. Based on this, a queue mechanism is introduced to schedule the parsed data, ensuring that various types of data are distributed in an orderly manner according to a preset order, thus generating corresponding data to be processed. This queue mechanism can be implemented using a message queue to ensure that the processing order and stability of data remain even under high concurrency or continuous input conditions.

[0115] After obtaining the data to be processed, it is input into a 3D convolutional neural network for multi-layer convolutional operations. The 3D convolutional neural network performs convolution operations on the data in the spatial and depth dimensions, enabling the model to simultaneously capture the structural information of the image and the correlation features across layers or frames, thereby generating corresponding intermediate feature data. Subsequently, pooling operations are performed on the intermediate feature data to reduce the dimensionality of the features and retain key response information, thereby generating corresponding dimensionality-reduced feature data. On this basis, multi-layer feature mapping is further performed on the dimensionality-reduced feature data, so that the features are gradually abstracted and strengthened at different levels, and finally a corresponding multi-dimensional feature map is generated. The multi-dimensional feature map includes at least structural information reflecting the spatial structure of the weld and semantic feature information used to distinguish different target categories, thereby providing basic feature support for subsequent target detection and fine representation processing.

[0116] Furthermore, the step of inputting the data to be processed into a 3D convolutional neural network for multi-layer convolution operations to generate corresponding intermediate feature data includes:

[0117] Based on the preset 3D convolution kernel weights and bias parameters in the 3D convolutional neural network, a 3D convolution operation is performed on the data to be processed to generate corresponding intermediate feature data. The 3D convolution operation satisfies the following relationship:

[0118] F = Conv3D(I; W, b)

[0119] Where I represents the data to be processed, W represents the 3D convolution kernel weights, b represents the bias parameter, Conv3D represents the 3D convolution operation, and F represents the intermediate feature data.

[0120] Furthermore, the steps of constructing a multi-scale target detection model based on multi-dimensional feature maps, and generating corresponding target detection results for welds and defects based on the multi-scale target detection model, include:

[0121] S301. Perform multi-level feature decomposition based on multi-dimensional feature maps to generate corresponding multi-scale feature maps;

[0122] S302. Based on feature maps at multiple scales, construct a feature pyramid network and perform top-down and horizontal fusion processing on the feature maps at each scale through the feature pyramid network to generate corresponding multi-scale fused feature maps.

[0123] S303. Construct an attention mechanism region proposal network based on the multi-scale fused feature map, and perform feature enhancement processing on the multi-scale fused feature map based on the attention mechanism region proposal network to generate the corresponding enhanced feature map.

[0124] S304. Based on the enhanced feature maps and combined with the output of the feature pyramid network, corresponding target detection proposals at each scale are generated. The target detection proposals at each scale satisfy the following relationship:

[0125] Pi=Attention-RPN(Fi)+FPN(Fi);

[0126] Fi represents the feature map of the i-th layer, Attention-RPN represents the attention mechanism region proposal network, FPN represents the feature pyramid network, and Pi represents the object detection proposal at the i-th scale.

[0127] S305. Based on the target detection proposals at various scales, perform target screening processing to generate corresponding target detection results for welds and defects.

[0128] In this embodiment, the process of constructing a multi-scale target detection model based on the multi-dimensional feature map involves uniformly modeling targets of different scales based on the aforementioned feature extraction results. Specifically, firstly, multi-level feature decomposition is performed based on the multi-dimensional feature map, dividing the feature information at different levels according to resolution and semantic expressive power, thereby generating corresponding multiple scale feature maps. The lower-level features focus more on detailed information, while the higher-level features focus more on semantic information, in order to adapt to the detection needs of targets of different sizes.

[0129] After obtaining multiple scale feature maps, a feature pyramid network is constructed to fuse the features at each scale, enabling feature information at different levels to interact within a unified structure. The feature pyramid network introduces high-level semantic information into low-level features through top-down information transmission and lateral connections, thereby generating a multi-scale fused feature map containing multi-scale information. This allows the model to simultaneously represent large-size weld structures and small-size defect targets.

[0130] Based on this, an attention mechanism region proposal network is further constructed according to the multi-scale fusion feature map. By weighting the feature responses, the model can enhance the representation of potential target regions and suppress the interference of background regions, thereby generating corresponding enhanced feature maps. The attention mechanism can be understood as giving higher weights to key regions, making the feature distribution more concentrated in the weld and defect regions.

[0131] Subsequently, based on the enhanced feature maps and the output of the feature pyramid network, target proposal generation is performed on the feature maps at each scale to form corresponding target detection proposals at each scale. These proposals satisfy the relationship Pi = Attention-RPN(Fi) + FPN(Fi), where Fi represents the i-th layer feature map, Attention-RPN represents the attention mechanism region proposal network, FPN represents the feature pyramid network, and Pi represents the target detection proposal at the i-th scale. This approach ensures that target information at each scale can be effectively extracted through the combined effect of fused features and attention enhancement.

[0132] Finally, based on the target detection proposals of each scale, the corresponding target detection results of welds and defects are obtained, so that the system can still achieve stable and accurate detection results in complex backgrounds and the coexistence of multiple scale targets.

[0133] Furthermore, the step of refining the detection region corresponding to the target detection result through a segmentation regression model based on the multi-dimensional feature map to generate the corresponding target contour information and spatial location information includes:

[0134] S401. Based on the target detection results, extract the corresponding detection region feature data from the multi-dimensional feature map and generate the corresponding region feature map.

[0135] S402. Input the region feature map into the segmentation regression model to construct the segmentation branch and the coordinate regression branch;

[0136] S403. Perform convolution operations on the region feature map based on the segmentation branch to generate the corresponding segmentation mask data;

[0137] S404. Perform regression calculation on the regional feature map based on the coordinate regression branch to generate corresponding spatial coordinate data;

[0138] S405. Based on the segmentation mask data and spatial coordinate data, perform joint processing to generate the corresponding target contour information and spatial location information.

[0139] In this embodiment, after target detection is completed, the detection results are not directly used for subsequent control. Instead, the detection area indicated by the target detection results is further refined. Specifically, based on the target range determined by the detection results, feature information of the corresponding region is extracted from the multi-dimensional feature map. During this extraction process, information originally distributed in the global feature map can be converted into a regional feature map containing only target-related content through region clipping or feature mapping alignment. This allows subsequent processing to focus on the weld or defect itself and reduces the interference of background information on the results.

[0140] After obtaining the region feature map, it is input into the segmentation regression model for further analysis and processing. This model includes a segmentation branch and a coordinate regression branch, which model the morphological and positional features of the target, respectively. In the segmentation branch, multi-layer convolutional operations progressively enhance the boundary responses in the region features, enabling the model to distinguish pixel differences between the target and the background, thereby achieving pixel-level segmentation of the target region and forming segmentation mask data reflecting the target's contour structure. In this process, a lightweight convolutional neural network structure can be used to reduce computational complexity and improve processing speed; for example, branch networks can be built based on MobileNet or ShuffleNet, allowing the system to maintain good segmentation accuracy while meeting real-time requirements.

[0141] Meanwhile, in the coordinate regression branch, feature compression and mapping are performed on the region feature map to transform high-dimensional feature information into low-dimensional numerical parameters. These parameters describe the target's positional relationship in the image coordinate system or spatial coordinate system, such as the coordinates of the target's center point or the spatial distribution of key location points, thereby generating corresponding spatial coordinate data. Essentially, this regression process establishes a mapping relationship between feature responses and spatial locations, enabling the model to directly infer the target's specific location from the feature level.

[0142] After the two branches are processed separately, the segmentation mask data and the spatial coordinate data are jointly processed to express the target's shape and position information in a unified result, thereby generating corresponding target contour information and spatial position information. During this joint process, feature alignment or data fusion can be used to match the two types of information in the same coordinate system, ensuring the consistency of contour boundaries and spatial positions. Simultaneously, the overall output relationship of the segmentation regression model can be expressed as y^=CNN(F)+Reg(F), where F represents the region feature map, CNN represents the processing procedure corresponding to the segmentation branch, Reg represents the processing procedure corresponding to the coordinate regression branch, and y^ represents the final prediction result. Through the above processing, the system, based on target detection, further obtains fine structural information and precise position information that can be used for control, thus providing reliable data support for subsequent fusion decisions and welding control.

[0143] Furthermore, the step of fusing the target detection results, target contour information, and spatial location information to generate the corresponding fusion decision result includes:

[0144] S501. Determine the corresponding detection confidence parameters based on the target detection results;

[0145] S502. Determine the corresponding contour accuracy parameters based on the target contour information;

[0146] S503. Determine the corresponding coordinate error parameters based on the spatial location information;

[0147] S504: The detection confidence parameter, contour accuracy parameter and coordinate error parameter are only processed for feature alignment without outputting the corresponding alignment feature data;

[0148] S505. Based on the preset fusion weights, perform weighted fusion processing on each aligned feature data to generate the corresponding fusion decision result.

[0149] In this embodiment, after obtaining the target detection result, the target contour information, and the spatial location information, no single result is directly used as the final judgment criterion. Instead, the three are fused together to generate a more stable and reliable fusion decision result. Specifically, firstly, corresponding evaluation parameters are extracted based on the target detection result, the target contour information, and the spatial location information. For the target detection result, the corresponding detection confidence parameter can be determined based on the probability distribution or classification response of the model output to characterize the credibility of the target detection. For the target contour information, the corresponding contour accuracy parameter can be determined based on the matching degree or boundary continuity between the segmentation mask and the actual contour to characterize the accuracy of the target boundary description. For the spatial location information, the corresponding coordinate error parameter is determined by calculating the degree of deviation between the predicted coordinates and the reference position to reflect the deviation of the position estimation.

[0150] After obtaining the aforementioned parameters, feature alignment processing is performed to transform data from different sources under a unified scale and expression. For example, normalization or standardization maps parameters of different dimensions to a unified range, thereby generating corresponding aligned feature data to ensure the comparability of various information in the subsequent fusion process. Based on this, each aligned feature data is weighted according to a preset fusion weight, allowing different information sources to participate in the final decision according to their reliability or importance. The fusion process satisfies the relationship D = α·Dseg + (1−α)·Ddet, where Dseg can be understood as the representation result obtained by combining the target contour information and spatial location information, Ddet corresponds to the target detection result, and α represents the fusion weight satisfying 0≤α≤1. By adjusting this weight, the influence of various information types can be balanced under different operating conditions.

[0151] Through the above fusion process, the detection results, contour information and location information work together in a unified decision space to generate corresponding fusion decision results. These fusion decision results can, to some extent, offset the uncertainty of a single information source, reduce the risk of misjudgment, and provide a more stable and reliable input basis for subsequent reinforcement learning control models.

[0152] Furthermore, the step of constructing a feedback sample dataset based on the execution result data includes:

[0153] S7011. Extract real sample data based on the execution result data;

[0154] S7012. Construct a least-squares generative adversarial network, which includes a generator and a discriminator.

[0155] S7013. Input real sample data into the generative adversarial network to train the generator and discriminator, and then generate corresponding synthetic sample data.

[0156] S7014. Generate the corresponding feedback sample dataset based on the real sample data and the synthetic sample data.

[0157] In this embodiment, after acquiring the execution result data, to enable the system to continuously optimize, it is necessary to reuse the effective information from the execution process. Specifically, the execution result data is first filtered and parsed to extract data content that reflects the actual welding effect, such as feature information related to weld quality, defect type, or control response. This data is then associated and labeled with the corresponding input states to form real sample data with practical significance. This type of data can directly reflect the actual performance of the model and control strategy under the current working conditions.

[0158] After obtaining the real sample data, a generative adversarial network (GAN) is constructed to expand the sample data. This GAN is designed based on a least squares mechanism and includes a generator and a discriminator. The generator generates new sample data based on the input data, while the discriminator judges the authenticity of the input data, thereby continuously optimizing the performance of both during adversarial training. In specific implementation, the real sample data is input into the GAN. By alternately training the generator and discriminator, the generator gradually learns the data distribution characteristics of the real samples and outputs synthetic sample data that is statistically similar to the real samples. Simultaneously, the discriminator continuously improves its ability to distinguish between real and generated data, thus forming a stable adversarial training process.

[0159] After completing the adversarial training described above, the synthetic sample data output by the generator is integrated with the real sample data. By uniformly organizing and labeling samples from different sources, consistency in data structure and labeling system is achieved, thus forming a corresponding feedback sample dataset. In this process, the introduction of synthetic sample data effectively expands the coverage of the original data, compensating for the limited number of actual collected samples. This ensures that the feedback sample dataset not only contains real-world working condition information but also possesses richer feature distributions, thereby providing more comprehensive data support for subsequent model parameter updates.

[0160] Furthermore, the step of inputting real sample data into a generative adversarial network to train the generator and discriminator, and then generating corresponding synthetic sample data, includes:

[0161] S70131. Input random noise data into the generator to generate corresponding preliminary sample data;

[0162] S70132. Based on the discriminator, the real sample data and the preliminary sample data are processed to generate the corresponding synthetic sample data;

[0163] The generation of synthetic sample data satisfies the following relationship:

[0164] minGmaxDV(D, G)=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1−D(G(z)))];

[0165] Where G is the generator, D is the discriminator, x is the real sample, Z is random noise, pdata(x) is the distribution of the real sample, pz(z) is the noise distribution, and V(D,G) is the GAN loss function.

[0166] In this embodiment, after introducing the real sample data into the generative adversarial network (GAN), in order for the generator to learn the distribution characteristics of the real data, the generator and discriminator need to be collaboratively optimized through an adversarial training process. Specifically, in the initial training phase, random noise data is input into the generator. This random noise can be understood as a set of unstructured initial inputs, which drive the generator to output sample data with a certain structure. This process generates corresponding preliminary sample data, which typically differs significantly from the real samples in the initial stage.

[0167] After generating preliminary sample data, the real sample data and the preliminary sample data are simultaneously input into the discriminator for discrimination processing. The discriminator is used to determine whether the input data comes from the real sample distribution and evaluates the generation quality of the generator by outputting the discrimination result. In this process, the discriminator continuously improves its ability to distinguish between real data and generated data, while the generator adjusts its own parameters according to the discrimination result, so that the generated data gradually approaches the distribution characteristics of the real samples, thus forming an adversarial relationship between the two.

[0168] Specifically, the loss function essentially constructs an adversarial optimization objective for the generator G and the discriminator D, making the generated data gradually approximate the distribution of the real data. Here, E_{x~p_{data}(x)}[log D(x)] represents the expected value of the discriminator D giving a higher discrimination probability to the real sample x when the real sample x follows the real distribution p_{data}(x). Therefore, maximizing this term drives the discriminator to improve its ability to recognize real data. Conversely, E_{z~p_z(z)}[log(1−D(G(z)))] represents the probability that the generated sample G(z) mapped by the generator G is identified as "non-real" by the discriminator when the random noise z follows the distribution p_z(z). Therefore, maximizing this term enhances the discriminator's ability to distinguish generated samples. Meanwhile, the generator G optimizes by minimizing the overall objective, adjusting the mapping relationship to maximize the value of D(G(z)). This ensures that the generated samples appear closer to the real samples in the discriminator's view, thus gradually approximating the true data distribution corresponding to p_{data}(x). This adversarial optimization structure of min_G max_D allows both the generator and discriminator to improve together during the game, making the final generated samples statistically closer to the real samples, achieving effective modeling of complex data distributions.

[0169] Furthermore, the step of updating the model parameters in the 3D convolutional neural network, multi-scale object detection model, and segmentation regression model based on the feedback sample dataset includes:

[0170] S7021. Input the feedback sample dataset into the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model respectively, so that each model generates the model prediction results of the base.

[0171] S7022. Determine the true annotation results based on the real sample data in the feedback sample dataset;

[0172] S7023. Based on the loss function, calculate the difference features between the model prediction results and the true labeled results. Then, perform gradient update processing on the model parameters of each model according to the difference features to generate updated model parameters. The updated model parameters satisfy the following relationship:

[0173] θ′=θ−η⋅∇θ L (y^, y);

[0174] Where θ represents the original model parameters, η represents the learning rate, y^ represents the model prediction result, y represents the ground truth labeling result, L(y^, y) represents the difference features of the loss function between the predicted y^ and the ground truth y, and θ represents the updated model parameters.

[0175] In this embodiment, after obtaining the feedback sample dataset, the model parameters need to be updated to enable each model to continuously optimize based on actual operating conditions. Specifically, the feedback sample dataset is first input into the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model, respectively, so that each model performs forward computation on the input data under the current parameter conditions, thereby outputting the corresponding model prediction results. During this process, each model generates its own prediction output for feature extraction, object detection, and fine representation tasks, and these outputs reflect the current model's processing capability under real-world operating data.

[0176] After obtaining the model prediction results, corresponding real annotation information is further extracted based on the real sample data in the feedback sample dataset. This real annotation information is used to characterize actual conditions such as weld location, defect type, or contour boundary, thus serving as a reference for model optimization. Subsequently, a loss function is constructed to quantify the difference between the model prediction results and the real annotation results. This difference can be understood as the degree of deviation between the model's current output and the actual situation.

[0177] In this embodiment, the parameter update relationship θ′=θ−η⋅∇_θ L(y^,y) is used to describe the optimization direction and adjustment magnitude of the model during training. Here, θ represents the current parameter state of the model, which can be understood as the set of weights and biases within the model, determining how the model responds to input data. θ′ represents the new parameter state obtained after this round of updates, which has been adjusted towards reducing error compared to θ. L(y^,y) represents the loss function, used to measure the degree of difference between the model's predicted result y^ and the true labeled result y. The closer the prediction is to the true value, the smaller the value of this loss function; therefore, its trend reflects the quality of the model's performance.

[0178] Based on this, ∇_θ L(y^, y) represents the gradient of the loss function with respect to the model parameters θ. Essentially, it represents the sensitivity of each parameter to changes in error under the current state; that is, how the loss function changes when a parameter undergoes a small change. This gradient vector indicates the direction in which the loss function increases. Therefore, during the update process, by subtracting η⋅∇_θ L(y^, y) from the original parameters θ—that is, adjusting the parameters in the opposite direction of the gradient—the model is optimized in the direction of reducing the loss function, thereby gradually improving the consistency between the predicted and actual results.

[0179] Here, η is the learning rate, used to control the step size of each parameter update. When η is large, the parameter update amplitude is large, and the model converges faster but may be unstable; when η is small, the update process is more stable but the convergence speed is relatively slower. Through the above update mechanism, the model continuously corrects its own parameters in multiple iterations, thereby gradually reducing the prediction error and improving the model's recognition accuracy and generalization ability in actual welding scenarios.

[0180] In another embodiment, without changing the overall processing flow and closed-loop control logic, equivalent alternative structures can be used for different functional modules to improve the system's adaptability in different application scenarios. Specifically, in the process of feature extraction processing of the input dataset based on 3D convolutional neural networks, in addition to using a three-dimensional convolutional structure to model the spatial features of the data, other deep convolutional network structures can also be used to extract features from the input dataset. For example, classic convolutional networks such as ResNet or VGG can be selected, and feature encoding of the input data can be performed through multi-layer convolution and residual connection structures to generate corresponding multi-dimensional feature maps, which also have the ability to express weld structure and defect features.

[0181] In the process of constructing a multi-scale target detection model based on multi-dimensional feature maps, in addition to using a feature pyramid structure combined with an attention mechanism region proposal structure to achieve multi-scale target detection, a single-stage target detection structure can also be used to directly process the multi-dimensional feature maps. For example, the YOLO series or SSD structure can be used to complete the target localization and classification tasks simultaneously in a single network, thereby generating the corresponding target detection results for welds and defects, achieving rapid target identification while ensuring detection efficiency.

[0182] In the process of refining the detection region corresponding to the target detection result through the segmentation regression model, in addition to using a lightweight convolutional network as the segmentation branch, other semantic segmentation network structures can also be used to process the feature map of the region. For example, U-Net or DeepLab structures can be used. Through the feature fusion mechanism of the encoder and decoder or the dilated convolution structure, the target region can be divided into pixels to generate the corresponding segmentation mask data, which together with the spatial coordinate data output by the coordinate regression branch constitute the target contour information and spatial location information.

[0183] In the process of constructing the feedback sample dataset based on the execution result data, in addition to using a generative adversarial network based on the least squares mechanism to expand the real samples, other generative adversarial network structures can also be used to generate samples, such as WGAN or StyleGAN structures. Synthetic sample data with the same distribution as the real samples can be generated through adversarial training mechanisms, thereby expanding the diversity of sample data and constructing the feedback sample dataset together with the real sample data for subsequent model optimization.

[0184] In the process of constructing a reinforcement learning control model based on the fusion decision results to generate optimal welding control parameters, in addition to using a deep Q-network structure to learn the control strategy, other reinforcement learning algorithms can also be used to optimize the control strategy, such as PPO or A3C algorithms. Through policy gradient or multi-threaded parallel learning mechanisms, different control actions are evaluated and updated to generate corresponding optimal welding control parameters, enabling the welding actuator to achieve adaptive adjustment according to different working conditions.

[0185] By employing the aforementioned alternative implementation methods, the system can flexibly select specific implementation structures based on different computing power conditions, real-time requirements, or application scenarios, while maintaining overall functional consistency, thereby improving the system's versatility and scalability.

[0186] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A machine vision intelligent detection and control method based on 3D convolution and reinforcement learning, characterized in that, include: Acquire image data of the welding area, and construct an input dataset based on the image data; Based on a 3D convolutional neural network, feature extraction processing is performed on the input dataset to generate a corresponding multidimensional feature map. Based on the multi-dimensional feature map, a multi-scale target detection model is constructed, and target detection results for corresponding welds and defects are generated based on the multi-scale target detection model. Based on the multidimensional feature map, the detection region corresponding to the target detection result is finely characterized by a segmentation regression model to generate the corresponding target contour information and spatial location information. The target detection result, the target contour information, and the spatial location information are fused to generate a corresponding fusion decision result; Based on the fusion decision results, a reinforcement learning control model is constructed to generate corresponding optimal welding control parameters. The welding actuator is driven to perform welding actions according to the optimal welding control parameters, thereby obtaining the corresponding execution result data. A feedback sample dataset is constructed based on the execution result data. The model parameters in the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model are updated according to the feedback sample dataset.

2. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of acquiring image data of the welding area and constructing an input dataset based on the image data includes: Based on an industrial camera, images of the welding area are acquired and processed to generate corresponding raw image data. Based on the pose sensor, the spatial pose of the target corresponding to the welding area is collected and processed to generate the corresponding original pose data. The original image data and the original pose data are preprocessed to generate corresponding preprocessed data. The preprocessing operation includes at least image processing operations such as denoising and contrast enhancement of the original image data, and outlier removal and coordinate transformation of the original pose data. Based on preset normalization rules, the preprocessed data is standardized to generate corresponding standardized data. The various standardizations are cached and integrated to generate the corresponding input dataset.

3. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of performing feature extraction processing on the input dataset based on a 3D convolutional neural network to generate corresponding multidimensional feature maps includes: The input dataset is parsed and processed using the established data receiving interface, and corresponding parsed data is output. Based on the queue mechanism, the parsed data is processed in an ordered manner to generate corresponding data to be processed; The data to be processed is input into a 3D convolutional neural network for multi-layer convolution operations to generate corresponding intermediate feature data. The intermediate feature data is processed by pooling to generate corresponding dimensionality-reduced feature data; Based on the dimensionality-reduced feature data, multi-layer feature mapping is performed to generate a corresponding multi-dimensional feature map, which includes at least spatial structure information and semantic feature information.

4. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 3, characterized in that, The step of inputting the data to be processed into a 3D convolutional neural network for multi-layer convolution operations to generate corresponding intermediate feature data includes: Based on the preset 3D convolution kernel weights and bias parameters in the 3D convolutional neural network, a 3D convolution operation is performed on the data to be processed to generate corresponding intermediate feature data. The 3D convolution operation satisfies the following relationship: F = Conv3D(I; W, b) Wherein, I is the data to be processed, W is the 3D convolution kernel weight, b is the bias parameter, Conv3D is the three-dimensional convolution operation, and F is the intermediate feature data.

5. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of constructing a multi-scale target detection model based on the multi-dimensional feature map, and generating target detection results for corresponding welds and defects based on the multi-scale target detection model, includes: Based on the multidimensional feature map, multi-level feature decomposition is performed to generate corresponding multiple scale feature maps. Based on multiple scale feature maps, a feature pyramid network is constructed, and the feature pyramid network is used to perform top-down and horizontal fusion processing on each scale feature map to generate a corresponding multi-scale fused feature map. An attention mechanism region proposal network is constructed based on the multi-scale fused feature map, and feature enhancement processing is performed on the multi-scale fused feature map based on the attention mechanism region proposal network to generate a corresponding enhanced feature map. Based on the enhanced feature map and the output of the feature pyramid network, corresponding target detection proposals at various scales are generated. The target detection proposals at various scales satisfy the following relationship: Pi=Attention-RPN(Fi)+FPN(Fi); Fi represents the feature map of the i-th layer, Attention-RPN represents the attention mechanism region proposal network, FPN represents the feature pyramid network, and Pi represents the target detection proposal at the i-th scale. Based on the target detection proposals for each scale, target screening is performed to generate corresponding target detection results for welds and defects.

6. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of performing fine-grained characterization processing on the detection region corresponding to the target detection result based on the multi-dimensional feature map and generating corresponding target contour information and spatial location information through a segmentation regression model includes: Based on the target detection results, the corresponding detection region feature data is extracted from the multi-dimensional feature map to generate the corresponding region feature map. The region feature map is input into the segmentation regression model to construct segmentation branches and coordinate regression branches; Based on the segmentation branch, the region feature map is processed by convolution operation to generate corresponding segmentation mask data; Based on the coordinate regression branch, regression calculations are performed on the regional feature map to generate corresponding spatial coordinate data; Based on the segmentation mask data and the spatial coordinate data, the corresponding target contour information and spatial location information are generated through joint processing.

7. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of fusing the target detection result, the target contour information, and the spatial location information to generate a corresponding fusion decision result includes: Determine the corresponding detection confidence parameters based on the target detection results; Determine the corresponding contour accuracy parameters based on the target contour information; Determine the corresponding coordinate error parameters based on the spatial location information; The detection confidence parameter, contour accuracy parameter, and coordinate error parameter only undergo feature alignment processing without outputting the corresponding alignment feature data; Based on preset fusion weights, each of the aligned feature data is weighted and fused to generate a corresponding fusion decision result.

8. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 1, characterized in that, The step of constructing a feedback sample dataset based on the execution result data includes: Extract real sample data based on the execution result data; A generative adversarial network based on a least squares mechanism is constructed, wherein the generative adversarial network includes a generator and a discriminator; The real sample data is input into the generative adversarial network to train the generator and the discriminator, thereby generating corresponding synthetic sample data; Based on the real sample data and the synthetic sample data, a corresponding feedback sample dataset is generated.

9. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 8, characterized in that, The step of inputting the real sample data into the generative adversarial network to train the generator and the discriminator, and then generating the corresponding synthetic sample data, includes: Input random noise data into the generator to generate corresponding preliminary sample data; Based on the discriminator, the real sample data and the preliminary sample data are processed to generate corresponding synthetic sample data; The generation of the synthetic sample data satisfies the following relationship: minGmaxDV(D, G)=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1−D(G(z)))]; Where G is the generator, D is the discriminator, x is the real sample, Z is random noise, pdata(x) is the distribution of the real sample, pz(z) is the noise distribution, and V(D,G) is the GAN loss function.

10. The machine vision intelligent detection and control method based on 3D convolution and reinforcement learning according to claim 8, characterized in that, The step of updating the model parameters in the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model based on the feedback sample dataset includes: The feedback sample dataset is input into the 3D convolutional neural network, the multi-scale object detection model, and the segmentation regression model, respectively, so that each model generates a base model prediction result; The true annotation result is determined based on the real sample data in the feedback sample dataset; Based on the loss function, the difference features between the model prediction results and the ground truth labeling results are calculated. Then, gradient update processing is performed on the model parameters of each model according to these difference features to generate updated model parameters. The updated model parameters satisfy the following relationship: θ′=θ−η⋅∇θ L (y^, y); Where θ represents the original model parameters, η represents the learning rate, y^ represents the model prediction result, y represents the ground truth labeling result, L(y^, y) represents the difference features of the loss function between the predicted y^ and the ground truth y, and θ represents the updated model parameters.

Images