Quality assessment method, apparatus, device

Abstract

The application discloses a quality evaluation method and device, and equipment. The method comprises the following steps: acquiring a cutting area image corresponding to a workpiece to be processed; performing pretreatment and data enhancement processing on the cutting area image to obtain an enhanced image sample set; inputting the enhanced image sample set into a pre-trained quality evaluation model, performing downsampling processing and fusion on the enhanced image sample set through a feature extraction branch to obtain a multi-scale fusion feature map, wherein the quality evaluation model comprises the feature extraction branch and a classification branch; performing classification processing on the multi-scale fusion feature map through the classification branch to obtain a quality detection result corresponding to the workpiece to be processed; and matching the quality detection result with a preset process parameter database to generate a target process parameter recommendation scheme corresponding to the workpiece to be processed. The application is suitable for all laser cutting parts and improves the quality evaluation accuracy and efficiency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a quality assessment method, apparatus, or device. Background Technology

[0002] With industrial development, the precision requirements for various components are becoming increasingly stringent. Accurate and efficient inspection and process optimization of laser cutting quality have become core industry needs. However, laser cutting quality is influenced by a combination of multiple process parameters, including laser power, cutting speed, focal point position, type and pressure of auxiliary gas, and material properties. A complex nonlinear mapping relationship exists between these process parameters and the final quality result, posing a significant challenge to precise quality control and rapid optimization of process parameters.

[0003] Currently, quality assessment of laser-cut parts is mainly based on machine vision-based defect detection methods. For example, by acquiring cross-sectional images of laser-cut workpieces and labeling defect categories and locations, a target detection model is trained, which then outputs the defect category, location, and confidence level.

[0004] However, most of the above methods use a single feature extraction method or a special model designed for specific materials and working conditions. The cutting defects (slag, cracks, excessive roughness, etc.) of different materials such as stainless steel, carbon steel, and aluminum alloys have different manifestations and discrete feature distributions, and cannot be applied to all laser-cut parts. Summary of the Invention

[0005] The main objective of this application is to provide a quality assessment method, apparatus, and equipment that is applicable to all laser-cut parts and improves the accuracy and efficiency of quality assessment.

[0006] To achieve the above objectives, in a first aspect, this application provides a quality assessment method, comprising: Acquire the image of the cutting area corresponding to the workpiece to be processed, wherein the image of the cutting area is an RGB image obtained by acquiring the cutting cross-section of the workpiece to be processed; The image of the cut region is preprocessed and data augmented to obtain an enhanced image sample set; The enhanced image sample set is input into a pre-trained quality assessment model. The enhanced image sample set is downsampled and fused through the feature extraction branch to obtain a multi-scale fused feature map. The quality assessment model includes a feature extraction branch and a classification branch. The multi-scale fused feature map is classified by classification branch to obtain the quality inspection result of the workpiece to be processed. The system matches the quality inspection results with a pre-set database of process parameters to generate a recommended scheme of target process parameters for the workpiece to be processed.

[0007] In one embodiment, acquiring an image of the cutting area corresponding to the workpiece to be processed includes: Fix the workpiece to be processed at the inspection station; The image acquisition device is controlled to acquire images of the cut section of the workpiece to be processed according to preset acquisition parameters, so as to obtain the initial cutting image; The initial cut image is scaled up according to a preset size to obtain the cut area image.

[0008] In one embodiment, the image of the cut region is preprocessed and data augmented to obtain an enhanced image sample set, including: The image of the cut area is subjected to grayscale correction processing to obtain a grayscale corrected image; The grayscale-corrected image is subjected to noise filtering to obtain a preprocessed image; Perform at least one or more of random rotation, mirroring, and cropping on the preprocessed image to obtain an enhanced image sample set.

[0009] In one embodiment, the enhanced image sample set is downsampled and fused through a feature extraction branch to obtain a multi-scale fused feature map, including: The enhanced image sample set is downsampled multiple times through a feature extraction branch to extract shallow-level and deep-level features respectively. By fusing shallow and deep features through element-wise addition, a multi-scale fused feature map is obtained.

[0010] In one embodiment, the enhanced image sample set is downsampled multiple times through a feature extraction branch to extract shallow-level features and deep-level features, including: The enhanced image sample set is downsampled for the first time to extract the cross-sectional edge features in the enhanced image sample set, thus obtaining the first shallow feature map; The first shallow feature map is downsampled a second time to extract the cross-sectional geometric features from the enhanced image sample set, thus obtaining the second shallow feature map. The second shallow feature map is downsampled a third time to extract the defect texture features from the enhanced image sample set to obtain a deep feature map. The defect texture features include at least one or more of the following: slag texture features, crack texture features, overheated texture features, and drag line texture features.

[0011] In one embodiment, the multi-scale fused feature map is classified by a classification branch to obtain the quality inspection result corresponding to the workpiece to be processed, including: The multi-scale fused feature map is input into the segmentation sub-network in the classification branch to perform pixel-level localization of the defect region in the cut surface of the workpiece to be processed, and obtain the defect segmentation mask. The multi-scale fused feature map is input into the classification sub-network in the classification branch to classify and identify each defect region in the defect segmentation mask, and obtain the defect category label corresponding to each defect region. The quality inspection results are generated based on the defect category labels corresponding to each defect area.

[0012] In one embodiment, the process involves matching the quality inspection results with a preset process parameter database to generate a recommended scheme of target process parameters for the workpiece to be processed, including: Based on the defect category labels in the quality inspection results, query the associated data in the preset process parameter database to obtain the candidate process parameter adjustment schemes corresponding to the defect category labels. The associated data is the mapping data between defect categories and process parameters. Obtain information on the material type, thickness, and laser source type of the workpiece to be processed; Using material type information, thickness information, and laser source type information as filtering criteria, a target process parameter adjustment scheme that matches the workpiece to be processed is selected from the candidate process parameter adjustment schemes. A recommended scheme for the target process parameters is generated based on the adjustment scheme for the target process parameters.

[0013] In one embodiment, the method further includes: In response to the defect category label of slag adhesion defect in the quality inspection results, a first process adjustment instruction is generated based on the associated data in the preset process parameter database to reduce laser power or increase cutting speed. In response to the defect category label of excessive roughness in the quality inspection results, a second process adjustment command is generated based on the associated data in the preset process parameter database to adjust the coke depth or the auxiliary gas pressure. The first or second process adjustment command is sent to the PLC controller to perform closed-loop adjustment of the real-time cutting parameters of the workpiece to be processed.

[0014] In one embodiment, the method further includes: Acquire the image sequence of the historical cutting area corresponding to each processing moment in the continuous processing of the workpiece; Obtain the historical quality detection results corresponding to each historical cut region image in the historical cut region image sequence; Determine the trend of quality change of the cut section of the workpiece to be processed based on historical quality inspection results; Acquire historical working data of the cutting head during continuous processing and / or historical power attenuation data of the laser during continuous processing; Based on the trend of the quality change of the cutting section and historical working data and / or historical power attenuation data, a loss prediction correlation model is established, and the loss state of the cutting head and / or laser is predicted based on the loss prediction correlation model to obtain the loss prediction results. A process adjustment prediction scheme is generated based on the loss prediction results.

[0015] Secondly, embodiments of this application provide a quality assessment apparatus, comprising: The image acquisition module is used to acquire the image of the cutting area corresponding to the workpiece to be processed. The image of the cutting area is an RGB image obtained by acquiring the cutting cross-section of the workpiece to be processed. The image processing module is used to preprocess and augment the image of the cut area to obtain an augmented image sample set. The feature extraction module is used to input the enhanced image sample set into the pre-trained quality assessment model. The enhanced image sample set is downsampled and fused through the feature extraction branch to obtain a multi-scale fused feature map. The quality assessment model includes a feature extraction branch and a classification branch. The feature classification module is used to classify the multi-scale fused feature map through classification branches to obtain the quality inspection result corresponding to the workpiece to be processed; The parameter recommendation module is used to match the quality inspection results with the preset process parameter database to generate a target process parameter recommendation scheme for the workpiece to be processed.

[0016] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods described above.

[0018] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the methods described above.

[0019] This application provides a quality assessment method, apparatus, and device, including: acquiring an image of a cutting area corresponding to a workpiece to be processed; preprocessing and data augmentation of the cutting area image to obtain an enhanced image sample set; inputting the enhanced image sample set into a pre-trained quality assessment model, and performing downsampling and fusion processing on the enhanced image sample set through a feature extraction branch to obtain a multi-scale fusion feature map, wherein the quality assessment model includes a feature extraction branch and a classification branch; performing classification processing on the multi-scale fusion feature map through the classification branch to obtain a quality inspection result corresponding to the workpiece to be processed; and matching the quality inspection result with a preset process parameter database to generate a target process parameter recommendation scheme corresponding to the workpiece to be processed. This application effectively expands the defect sample library under different material, thickness, and laser light source combinations through image preprocessing and data augmentation techniques, overcoming the problem of overfitting due to small sample training. Combined with multi-scale downsampling and fusion strategies in the feature extraction branch, it accurately captures multi-level features such as cross-sectional edges, geometric contours, and deep defect textures while ensuring lightweight computation. This achieves high-precision classification and localization of defects such as slag, cracks, and excessive roughness in the cut sections of materials such as stainless steel, carbon steel, and aluminum alloys, resulting in high defect recognition accuracy. Furthermore, this method intelligently matches quality inspection results with a preset process parameter database, quickly selecting the optimal process adjustment scheme based on defect category, material type, thickness, and light source information. This is applicable to all laser-cut parts and improves the accuracy and efficiency of quality assessment. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 This is a schematic diagram of the structure of a quality assessment system provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a quality assessment method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a quality assessment model provided in an embodiment of this application; Figure 4 This is a schematic diagram of a three-dimensional correlation framework of parameters, features, and quality provided in an embodiment of this application; Figure 5 This is a schematic diagram of a visual interface provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a quality assessment device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein.

[0023] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0024] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0025] It should be understood that in this application, "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product or device.

[0026] It should be understood that in this application, "multiple" refers to two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, "and / or B" can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "Contains A, B, and C", "Contains A, B, and C" means that all three A, B, and C are contained; "Contains A, B, or C" means that one of A, B, and C is contained; "Contains A, B, and / or C" means that any one, two, or three of A, B, and C are contained.

[0027] It should be understood that in this application, "B corresponding to A", "B corresponding to A", "A corresponds to B", or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Matching A and B is defined as a similarity between A and B that is greater than or equal to a preset threshold.

[0028] Depending on the context, "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection."

[0029] The data involved in this application may be data authorized by the tester or fully authorized by all parties. The collection, dissemination, and use of the data shall comply with the relevant laws, regulations and standards of the relevant countries and regions. The implementation methods / executives of this application may be combined with each other.

[0030] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0031] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0032] Laser cutting, as a high-efficiency, high-precision, and highly flexible processing method, has wide applications in various fields such as automobiles, ships, and aerospace. The quality of laser cutting, especially surface quality indicators such as the roughness of the cut surface and slag adhesion, directly affects the subsequent deep processing of the workpiece and ultimately determines the performance and quality of the product.

[0033] With industrial development, the precision requirements for various components are becoming increasingly stringent. Accurate and efficient inspection and process optimization of laser cutting quality have become core industry needs. However, laser cutting quality is influenced by a combination of multiple process parameters, including laser power, cutting speed, focal point position, type and pressure of auxiliary gas, and material properties. A complex nonlinear mapping relationship exists between these process parameters and the final quality result, posing a significant challenge to precise quality control and rapid optimization of process parameters.

[0034] Currently, the detection of laser cutting quality and optimization of process parameters mainly rely on the following methods. The first is manual judgment based on experience and consulting process manuals, where operators visually observe the quality of the cut surface and manually adjust parameters based on personal experience or by referring to process manuals. The second is machine vision-based defect detection methods. For example, by acquiring cross-sectional images of the laser-cut workpiece and labeling defect categories and locations to train a target detection model, then outputting defect categories, locations, and confidence levels; or using a coaxial vision sensing system to acquire visual images of the cutting front and cutting spark clusters, extracting image feature signals of cutting defects and changes in surface roughness. The third is an automatic process parameter generation method based on image recognition. This involves acquiring images of the cut surface with a preset cutting quality and constructing a laser cutting physical model. Process parameters are selected from a process library, and after cutting, the workpiece surface image is acquired and compared with the preset image. If the quality does not meet the requirements, an error model is established to inversely solve for parameter correction. The fourth is a method based on traditional machine learning, such as using the LBP algorithm to extract laser-cut cross-sectional features and using an SVM model to classify cutting parameters and cutting quality.

[0035] The aforementioned existing technical solutions still have many shortcomings and deficiencies in practical industrial applications. First, methods relying on human experience are highly subjective, have low accuracy, and are time-consuming and labor-intensive. Novice operators find it difficult to quickly learn, resulting in high process debugging costs and low efficiency. Second, existing methods based on machine vision or image recognition mostly employ single feature extraction techniques or dedicated models designed for specific materials and working conditions. However, the cutting defects (slag buildup, cracks, excessive roughness, etc.) of different materials such as stainless steel, carbon steel, and aluminum alloys exhibit significant differences in form and have discrete feature distributions, making it difficult to cover all scenarios with a single feature extraction method. Third, due to the high cost and difficulty in collecting and labeling defect samples, existing models generally face the problem of small sample sizes during training, easily leading to overfitting and insufficient generalization ability in real industrial scenarios.

[0036] Furthermore, existing solutions often struggle to achieve an effective balance between accuracy and computational speed, failing to simultaneously meet the requirements of high-precision detection and real-time performance on industrial equipment. Meanwhile, the nonlinear mapping relationships between process parameters, material properties, and quality results are extremely complex. Most existing technologies lack effective means to organically integrate physical theories such as energy efficiency models, fluid models, and thermal action models with deep learning models, making it difficult to establish accurate process parameter recommendation models. Finally, existing solutions lack sufficient visualization, failing to transform abstract algorithmic predictions into intuitive and actionable process guidance, and failing to form a complete closed-loop control system of "detection, analysis, and adjustment," thus limiting their practical application value in intelligent manufacturing scenarios.

[0037] To address the aforementioned problems, this application provides a quality assessment method.

[0038] Please see Figure 1 , Figure 1 This is a schematic diagram of the structural domain of a quality assessment system provided in an embodiment of this application. Figure 1 As shown, it includes the following: The system consists of a process data layer, an AI process algorithm layer, and an AI process software layer, arranged from bottom to top. These three layers interact bidirectionally through standardized data interfaces and calling protocols.

[0039] The process data layer is the foundational support layer of the system, responsible for managing all raw data and processing knowledge. This layer contains a basic experimental materials database, storing historical cutting process parameters, corresponding cut section images, and labeled quality tags for different materials such as stainless steel, carbon steel, and aluminum alloys under various thicknesses (1mm to 50mm) and different laser light source combinations (pulsed / continuous). Simultaneously, the data layer also includes structured information such as 3D cross-sectional morphology data, cross-sectional feature extraction results, and defect diagnosis records, as well as model parameters obtained through backpropagation and neural network training. This layer not only provides the raw data required for training the upper-layer AI algorithms but also continuously receives feedback data from the process guidance software (such as user adoption or correction of recommended solutions), enabling dynamic updates and iterative optimization of the database.

[0040] The AI ​​process algorithm layer is the core decision-making layer of the system, integrating various deep learning models and process theory models. This layer includes basic AI detection algorithms and models (such as improved CNN segmentation models and ResNet classification models), lightweight detection models that have undergone deep optimization (such as using depthwise separable convolution and model quantization techniques to balance accuracy and computational speed), as well as theoretical models and acquisition models (covering energy efficiency models, fluid models, and thermal action models).

[0041] Furthermore, the algorithm layer incorporates a built-in debugging skills database, gathering a wealth of experience rules from process experts. This includes specialized process strategies for 10mm stainless steel laser cutting, general process guidelines for common thicknesses of stainless steel, and a comprehensive process knowledge base covering common thicknesses of materials such as stainless steel, carbon steel, and aluminum alloys. Through joint reasoning between neural networks and theoretical models, this layer maps the multi-scale fused feature maps output from the feature extraction branch into specific quality inspection results and defect category labels. It further combines energy density threshold formulas (such as the power density conditions required for melting and cutting) and fluid characteristic parameters to generate preliminary process adjustment strategies, providing prediction results and model parameters for the upper-layer software.

[0042] The AI ​​process software layer serves as the system's user interaction and business integration layer, comprising three main functional modules: image preprocessing software, data analysis software, and process guidance software. The image preprocessing software handles image acquisition control, size scaling, grayscale correction, noise filtering, and data augmentation, transforming the original cut images into an enhanced image sample set. The data analysis software utilizes the detection model in the AI ​​process algorithm layer to perform feature extraction, defect segmentation, and classification, and statistically analyzes the output quality inspection results, generating a visualized defect distribution report and quality trend charts. The process guidance software, based on matching logic, queries relevant data in the process data layer to generate recommended solutions for target process parameters, presenting them to operators in an intuitive graphical format via a human-machine interface. It also supports one-click distribution to the PLC controller for closed-loop adjustment. This software layer is also responsible for transmitting the analyzed feature data and user feedback data back to the process data layer, driving the continuous expansion of the basic experimental material database and the retraining of the model database, thus forming a complete data closed loop of "data, algorithm, application, and feedback."

[0043] The calling relationships between the layers are as follows: The process guidance software uses defined interfaces (such as...) Figure 1 The system sends inference requests to the algorithm layer via strategies such as "image-to-process" and "model-to-process" (as indicated by annotations). The algorithm layer then returns prediction results and model parameters. Simultaneously, the process guidance software provides updated data (such as newly added "defect-parameter" mapping records or revised process schemes) to the data layer based on user operations and on-site execution results. The data layer responds to data query and training data requests from the software layer. The entire architecture fully embodies a modular and scalable design philosophy, supporting standardized process recommendations for existing stainless steel, carbon steel, and aluminum alloy materials, while also facilitating future expansion to laser processing scenarios for other metals or non-metals. This enables cross-scenario adaptive feature extraction and parameter optimization, meeting the industrial needs of multi-variety, small-batch production.

[0044] Please see Figure 2 , Figure 2 This is a flowchart illustrating a quality assessment method provided in an embodiment of this application. Figure 2 As shown, it includes the following steps: Step S201: Obtain the image of the cutting area corresponding to the workpiece to be processed. Among them, the cutting area image is an RGB image obtained by acquiring the cut surface of the workpiece to be processed.

[0045] In one embodiment, acquiring an image of the cutting area corresponding to the workpiece to be processed includes: fixing the workpiece to be processed at a detection station; controlling an image acquisition device to acquire images of the cutting cross-section of the workpiece to be processed according to preset acquisition parameters to obtain an initial cutting image; and scaling the initial cutting image according to a preset size to obtain a cutting area image.

[0046] Specifically, first, the workpiece to be processed is fixed at the inspection station, which is equipped with a positioning fixture. The positioning fixture automatically adjusts the clamping force and positioning position according to the material type (stainless steel, carbon steel or aluminum alloy) and size of the workpiece to ensure that the cutting surface faces the lens of the image acquisition device.

[0047] Subsequently, the image acquisition device is controlled to acquire images of the cut surface of the workpiece to be processed according to preset acquisition parameters. The image acquisition device uses an industrial CCD camera in conjunction with a ring light source or a coaxial light source. The preset acquisition parameters include resolution, exposure time, gain, and light source intensity, among which the resolution is set to 1920×1080 pixels, and the exposure time is adaptively adjusted within the range of 10μs to 100μs according to the reflective characteristics of the material surface. After acquiring the initial cut image, the initial cut image is scaled according to a preset size. The preset size is determined according to the input layer requirements of the quality assessment model, for example, the image is uniformly scaled to 512×512 pixels or 224×224 pixels. During the scaling process, a bicubic interpolation algorithm is used to preserve the detailed information of the cross-sectional texture and defect edges. The image of the cut area is obtained after scaling.

[0048] Step S202: Preprocess and data augmentation are performed on the image of the cut region to obtain an enhanced image sample set.

[0049] In one embodiment, preprocessing and data augmentation of the cut region image to obtain an enhanced image sample set includes: performing grayscale correction processing on the cut region image to obtain a grayscale corrected image; performing noise filtering processing on the grayscale corrected image to obtain a preprocessed image; and performing at least one or more of random rotation, mirroring, and shearing on the preprocessed image to obtain an enhanced image sample set.

[0050] Specifically, firstly, grayscale correction is performed on the image of the cut area. Grayscale correction employs histogram equalization or adaptive gamma correction methods. Based on the illumination distribution characteristics of cross-sectional images of different materials such as stainless steel, carbon steel, and aluminum alloy, the overall brightness and contrast of the image are adjusted to ensure that slag-covered areas, roughness textures, and drag lines in the cross-section have significant distinguishability in the grayscale space, resulting in a grayscale-corrected image.

[0051] Then, noise filtering is performed on the grayscale corrected image. Noise filtering uses a combination of median filtering and Gaussian filtering. Median filtering is used to remove salt-and-pepper noise and isolated noise points, while Gaussian filtering is used to smooth high-frequency noise in the image. The filter kernel size is set to 3×3 or 5×5 according to the image resolution. The preprocessed image is obtained after filtering.

[0052] Finally, data augmentation is performed on the preprocessed image. Data augmentation includes at least one or more of random rotation, mirroring, and cropping: random rotation refers to rotating the image at a random angle within the range of -15° to +15°; mirroring refers to flipping the image horizontally or vertically around its vertical or horizontal center line; cropping refers to randomly translating and scaling the image to simulate image changes under different acquisition angles and distances.

[0053] Through the above data augmentation operations, the defect sample library for different materials such as stainless steel, carbon steel, and aluminum alloys is expanded.

[0054] Step S203: Input the enhanced image sample set into the pre-trained quality assessment model, and perform downsampling and fusion on the enhanced image sample set through the feature extraction branch to obtain a multi-scale fused feature map.

[0055] The quality assessment model includes a feature extraction branch and a classification branch.

[0056] In one embodiment, the enhanced image sample set is downsampled and fused through a feature extraction branch to obtain a multi-scale fused feature map, including: performing multiple downsampling processes on the enhanced image sample set through the feature extraction branch to extract shallow-level features and deep-level features respectively; and fusing the shallow-level features and deep-level features by adding them element by element to obtain a multi-scale fused feature map.

[0057] Specifically, the enhanced image sample set is downsampled multiple times through a feature extraction branch to extract shallow-level and deep-level features, including: performing a first downsampling process on the enhanced image sample set to extract cross-sectional edge features from the enhanced image sample set, resulting in a first shallow feature map; performing a second downsampling process on the first shallow feature map to extract cross-sectional geometric features from the enhanced image sample set, resulting in a second shallow feature map; and performing a third downsampling process on the second shallow feature map to extract defect texture features from the enhanced image sample set, resulting in a deep-level feature map. The defect texture features include at least one or more of the following: slag texture features, crack texture features, overheated texture features, and drag line texture features.

[0058] Specifically, the quality assessment model includes a feature extraction branch and a classification branch. The feature extraction branch of the quality assessment model employs an improved lightweight convolutional neural network architecture. This architecture uses MobileNetV3 or ShuffleNetV2 as the backbone network and reduces computational cost through depthwise separable convolutions.

[0059] First, the enhanced image sample set is downsampled multiple times through a feature extraction branch to extract shallow and deep features. The first downsampling process uses a convolution operation with a stride of 2 to reduce the image resolution to half of its original value, extracting the cross-sectional edge features from the enhanced image sample set to obtain the first shallow feature map. The second downsampling process continues to perform a convolution operation with a stride of 2 on the first shallow feature map, further reducing the resolution to 1 / 4, extracting cross-sectional geometric features (such as cross-sectional contour shape, cut width, etc.) to obtain the second shallow feature map. The third and subsequent downsampling processes continue to extract deep features, including defect texture features (slag texture features, crack texture features, overheated texture features, and drag line texture features, etc.), to obtain deep feature maps.

[0060] After completing the multi-level feature extraction, shallow-level features and deep-level features are fused by adding them element by element. That is, the feature maps of each level are aligned in the channel dimension and then added pixel by pixel to obtain a multi-scale fused feature map.

[0061] like Figure 3 As shown, the network structure of the quality assessment model, from input to output, includes an input layer, a spatial transformation layer, a max pooling layer, an average pooling layer, a batch normalization layer (BN layer), a linear rectified layer (ReLU layer), a fully connected layer, and an output layer, ultimately generating a laser cutting parameter adjustment scheme report.

[0062] Specifically, the input layer receives the cut area image after preprocessing and data augmentation in steps S201 and S202. Its size is 480×480×3 (RGB three channels), a size that balances the preservation of cross-sectional defect details with the efficiency of subsequent network computation. The spatial transformation layer performs a spatial affine transformation on the input image, achieving automatic calibration and standardization of the cut section position in the image. This eliminates the spatial displacement and rotation effects caused by workpiece clamping angle deviations or camera viewing angle differences, enhancing the network's robustness to cross-sectional images under different placement postures.

[0063] Subsequently, the network alternately uses max pooling and average pooling layers for downsampling. Max pooling layers are used to extract the most significant activation responses in defect areas of the image (such as local extrema features like slag edges and crack tips), while average pooling layers are used to obtain global background information and texture statistics of the cross-section. The combination of both yields a multi-scale fused feature map, allowing the network to focus on both local discriminative details and overall contextual information. After each pooling layer, a batch normalization (BN) layer is introduced to normalize the feature maps of each batch, accelerating training convergence, mitigating the vanishing gradient problem, and providing a regularization effect to prevent overfitting, especially suitable for small-sample defect data scenarios. Following the BN layer, a ReLU layer provides a non-linear activation function, effectively mitigating gradient saturation through sparse activation characteristics while reducing computational overhead, meeting the lightweight requirements of real-time inference in industrial settings.

[0064] Step S204: Classify the multi-scale fused feature map through classification branches to obtain the quality inspection result corresponding to the workpiece to be processed.

[0065] In one embodiment, the quality inspection result corresponding to the workpiece to be processed is obtained by classifying the multi-scale fused feature map through a classification branch. This includes: inputting the multi-scale fused feature map into the segmentation sub-network in the classification branch to perform pixel-level localization of the defect region in the cut surface of the workpiece to be processed, thereby obtaining a defect segmentation mask; inputting the multi-scale fused feature map into the classification sub-network in the classification branch to classify and identify each defect region in the defect segmentation mask, thereby obtaining a defect category label corresponding to each defect region; and generating a quality inspection result based on the defect category label corresponding to each defect region.

[0066] Specifically, firstly, the multi-scale fused feature map is input into the segmentation sub-network. This sub-network employs a semantic segmentation model based on either the U-Net or DeepLabV3+ architecture to perform pixel-level localization of defect regions in the cut surface of the workpiece, outputting a defect segmentation mask. In the defect segmentation mask, different pixel values ​​correspond to different defect categories; for example, slag-covered areas are labeled as the first pixel value, cracked areas as the second pixel value, overheated areas as the third pixel value, and drag lines as the fourth pixel value. The segmentation sub-network addresses the vanishing gradient problem in deep network training through ResNet's residual module and introduces dilated convolution to expand the receptive field for accurate capture of small defects such as slag-covered areas and micro-cracks. On the validation set, the segmentation model achieves an Intersection over Union (IoU) of 96.7% and a global accuracy of 99.7%, meeting the requirements for accurate localization of defect regions in the cut surface.

[0067] Then, the multi-scale fused feature map and the defect segmentation mask are input into the classification sub-network. The classification sub-network is constructed based on multilayer perceptron (MLP) or global average pooling plus fully connected layers to classify and identify each defect region in the defect segmentation mask and output the defect category label corresponding to each defect region (such as "slag", "crack", "overheating", "excessive roughness" etc.).

[0068] Finally, based on the defect category labels and defect location information corresponding to each defect area, a comprehensive quality inspection result is generated.

[0069] like Figure 3 As shown, after multiple pooling and activation layers, the high-dimensional feature map, such as the multi-scale fusion feature map, is flattened into a one-dimensional feature vector and input to a fully connected layer. This fully connected layer employs a multilayer perceptron (MLP) structure to map the extracted abstract visual features into specific process parameter adjustment amounts. The output dimension of the fully connected layer is determined according to the type of process parameter to be adjusted, such as continuous values ​​like laser power adjustment percentage, cutting speed adjustment percentage, focal position offset, and auxiliary gas pressure adjustment. The output layer is responsible for integrating these values ​​into structured quality inspection results, such as a "Laser Cutting Parameter Adjustment Scheme Report." This report not only includes specific adjustment recommendations but may also include confidence scores and adjustment priority rankings.

[0070] Step S205: Match the quality inspection results with the preset process parameter database to generate a recommended scheme of target process parameters for the workpiece to be processed.

[0071] In one embodiment, a target process parameter recommendation scheme is generated by matching the quality inspection results with a preset process parameter database. This includes: querying associated data in the preset process parameter database based on the defect category tags in the quality inspection results to obtain candidate process parameter adjustment schemes corresponding to the defect category tags, wherein the associated data is mapping data between defect categories and process parameters; obtaining material type information, thickness information, and laser source type information of the workpiece to be processed; using the material type information, thickness information, and laser source type information as filtering conditions, selecting the target process parameter adjustment scheme that matches the workpiece to be processed from the candidate process parameter adjustment schemes; and generating a target process parameter recommendation scheme based on the target process parameter adjustment scheme.

[0072] Specifically, the process first queries the pre-defined process parameter database based on the defect category labels in the quality inspection results to obtain candidate process parameter adjustment schemes corresponding to the defect category labels. The process parameter database pre-stores a mapping relationship between "defect category - process parameter adjustment scheme". For example, when the defect category is "slag buildup", the associated candidate adjustment schemes include reducing laser power (adjustment range 5%-20%), increasing cutting speed (adjustment range 5%-15%), and adjusting the focal point position, etc.; when the defect category is "roughness exceeds the standard", the associated candidate adjustment schemes include adjusting the depth of focus, adjusting the auxiliary gas pressure (pressure and flow rate of oxygen / nitrogen / air), and adjusting the cutting speed, etc.

[0073] Then, obtain the material type information (stainless steel / carbon steel / aluminum alloy), thickness information (1-50mm), and laser source type information (pulsed / continuous) of the workpiece to be processed. Using the material type information, thickness information, and laser source type information as filtering criteria, select the target process parameter adjustment scheme that matches the specific working conditions of the current workpiece from the candidate process parameter adjustment schemes.

[0074] Finally, based on the selected target process parameter adjustment schemes and combined with the current initial process parameters, a recommended target process parameter scheme is generated, which includes parameters such as laser power, cutting speed, focal position, and auxiliary gas type and pressure.

[0075] like Figure 4As shown in the embodiments of this application, a three-dimensional correlation analysis framework of "parameters, features, and quality" is further constructed. This framework systematically explains the inherent mapping logic from laser processing input parameters to the image features of the cutting section, and then to the final quality assessment result. It provides clear theoretical and physical constraints for feature extraction in step S203, quality classification in step S204, and process parameter recommendation in step S205. This framework divides all variables involved in the entire processing into three major categories of key characteristics: key input characteristics (KICs), key measurement characteristics (KMCs), and key performance characteristics (KPCs). The causal transmission relationship among these three constitutes the core link of quality assessment and process optimization.

[0076] Key Input Characteristics (KICs) refer to the set of process parameters that can be actively controlled and preset before processing, namely, parameters related to energy density and fluid properties. Specifically, these include: laser power (directly determining the energy density acting on the material surface), processing speed (affecting the interaction time between the laser and the material), pulse frequency and duration (for pulsed laser modes, determining the energy and peak power of a single pulse), depth of focus (affecting spot size and energy distribution), material properties (including intrinsic parameters such as density, specific heat capacity, melting temperature, vaporization temperature, latent heat of fusion, thermal conductivity, and reflectivity to laser for stainless steel, carbon steel, and aluminum alloys), as well as fluid control parameters such as auxiliary gas type, pressure, and nozzle diameter. These KICs are variables that operators can directly set or adjust before or during processing, and are the core components of the recommended target process parameters in step S205.

[0077] Key measurement characteristics (KMCs) refer to observable physical quantities captured in real time by optical sensing devices during the processing. These serve as intermediate variables, directly reflecting the immediate response state of the material after KICs act upon it. The KMCs shown in the figure mainly include: the morphology of the molten pool and plume (reflecting the dynamic process of material melting and vaporization), radiation intensity (characterizing the strength of plasma or thermal radiation), and temperature intensity (temperature field distribution in the cutting area). These measurement characteristics are acquired online by optical sensing devices (such as high-speed cameras, photodiodes, infrared thermometers, etc.). However, in this application, considering the stability and real-time requirements of the industrial site, the preferred implementation is to indirectly characterize the cumulative effect of KMCs by acquiring cross-sectional images after cutting. That is, the cutting area image acquired in step S201 itself contains all the visual imprints of KMCs on the final result of the processing.

[0078] Key performance characteristics (KPCs) are the final output cutting quality indicators, which are the target objects directly judged by the quality assessment model. The diagram subdivides KPCs into two main categories: "form" and "defect." The former includes macroscopic morphological indicators such as surface roughness (Ra value) and texture features (e.g., uniformity of stripe distribution, dragging line trend), while the latter covers specific defect categories such as slag adhesion, cracks, and overheating. These KPCs correspond one-to-one with the defect segmentation mask and defect category labels output by the segmentation sub-network in step S204, and are also the direct basis for matching the process parameter adjustment scheme in step S205.

[0079] The technical significance of this framework lies in establishing a full-chain quantitative mapping relationship between "KICs, KMCs, and KPCs": On the one hand, the quality assessment model analyzes the texture trends and slag morphology of cross-sectional images (KPCs) to infer anomalies in KMCs (such as excessive energy input leading to over-boiling of the molten pool, or insufficient cooling rate leading to slag accumulation), and then traces the deviation sources of KICs (such as excessively high laser power or excessively low cutting speed). On the other hand, theoretical models (such as energy density threshold formulas and fluid dynamics equations) provide physical prior constraints for the nonlinear relationship between KICs and KPCs in this framework, so that the shallow-level features (cross-sectional edges, geometric contours) and deep-level features (defect textures, drag lines) captured by the feature extraction branch in step S203 form semantic correspondences with KMCs and KPCs, respectively, thereby effectively reducing the model's dependence on pure data fitting and improving the generalization ability under small sample conditions. Meanwhile, this framework also provides a multi-dimensional matching dimension for the screening of candidate process parameter adjustment schemes in step S205. When a certain type of KPCs (such as slag) exceeds the standard in the quality inspection results, the system can use this framework to reverse the causal chain to locate the most likely KICs adjustment direction (reduce power or increase speed), ensuring that the recommended scheme is not only relevant at the data level, but also interpretable and operable at the physical mechanism level, thereby truly realizing the quantitative correlation between "predicted results and process adjustment".

[0080] like Figure 5 As shown in the figure, after completing inference, the quality assessment model in this embodiment outputs detailed quality inspection results and process adjustment suggestions through a visual interface. This example illustrates the model's specific prediction output for a laser cutting cross-section image, intuitively demonstrating the classification processing results of step S204 and the process parameter matching recommendation effect of step S205. It represents the final presentation of the system's complete "image acquisition, AI analysis, parameter recommendation" chain on the user side.

[0081] The prediction results interface clearly displays the cutting conditions (cutting speed 5 m / min) corresponding to the input image. The main defect category output by the model is "Heavy slag adhesion," with a confidence level as high as 99.57%. This result corresponds perfectly to the defect category label output by the classification sub-network in step S204, indicating that the quality detection result of the cut section is a slag adhesion defect, and the model has a very high confidence level in this judgment. Meanwhile, the right side of the interface provides a recommended range of defocusing amount (4 to 0) under this speed condition. This value is derived from the optimal defocusing amount range records for the same or similar conditions in the process parameter database.

[0082] More importantly, the interface automatically generates specific process suggestions listed in bullet points based on the quality inspection results, including: ① Increase defocusing amount (adjust in the positive direction); ② Check if the gas pressure is sufficient; ③ Appropriately reduce the cutting speed; ④ Check if the nozzle is clean; ⑤ Confirm the laser power setting. These suggestions are consistent with the target process parameter adjustment scheme generated in step S205 based on the defect category label query related data. Specifically, for the defect of "severe slag adhesion," the association rules in the process parameter database indicate that slag adhesion is usually caused by factors such as excessive laser power, excessively slow cutting speed, or improper defocusing amount settings. Therefore, the system comprehensively recommends a multi-dimensional coordinated adjustment strategy, including increasing defocusing amount, reducing cutting speed, and checking gas pressure and nozzle cleanliness, fully reflecting the technical approach of multivariate coupling optimization.

[0083] Furthermore, the bottom of the interface displays the model's predicted probability distribution for all defect categories in the form of a probability distribution bar chart. "Severe slag adhesion" dominates with an absolute advantage of 99.57%, while the probabilities of the remaining categories (slight slag adhesion, rough surface without slag (negative / positive defocus direction), rough surface without slag (layered stripes), and smooth surface without slag) are all below 1%. This panoramic display of multi-category probabilities not only enhances the operator's confidence in the model's judgment but also provides a reference for subsequent manual review and fine-tuning. When the model's confidence level for a certain category is within the threshold boundary (e.g., 70%-90%), the operator can make a more prudent decision by combining the probability distribution of other categories in the interface.

[0084] This interactive prediction interface corresponds entirely to the implementation method in step S205, which involves "displaying the recommended target process parameters to the operator through a visual interface." It also supports a closed-loop execution function of "one-click distribution to the PLC controller." Operators can directly adopt the system's recommended process adjustment scheme or fine-tune the recommended parameters based on their own experience before distribution, achieving intelligent process optimization through human-machine collaboration. The system exhibits a parameter adjustment response delay of ≤0.2 seconds during stable processing and a defect identification accuracy rate exceeding 99%, comprehensively replacing over 80% of manual process debugging workload. The prediction result interface shown in this figure is a typical manifestation of these technical effects in practical applications.

[0085] In one embodiment, the method further includes: In response to the defect category label of slag adhesion defect in the quality inspection results, a first process adjustment instruction is generated based on the associated data in the preset process parameter database to reduce laser power or increase cutting speed. In response to the defect category label of excessive roughness in the quality inspection results, a second process adjustment command is generated based on the associated data in the preset process parameter database to adjust the coke depth or the auxiliary gas pressure. The first or second process adjustment command is sent to the PLC controller to perform closed-loop adjustment of the real-time cutting parameters of the workpiece to be processed.

[0086] Specifically, in response to the defect category label "slag adhesion defect" in the quality inspection results, a first process adjustment instruction is generated based on the associated data in the preset process parameter database to reduce laser power or increase cutting speed. For example, when a slag adhesion defect is detected on the stainless steel cut surface, the process parameter database is queried to determine that the slag adhesion defect is usually caused by excessive laser power or excessive cutting speed. Based on this, a first process adjustment instruction is generated to reduce the laser power by 10% or increase the cutting speed by 8%.

[0087] In response to a defect category label indicating excessive roughness in the quality inspection results, a second process adjustment command is generated based on associated data in a pre-set process parameter database to adjust the coke depth or the auxiliary gas pressure. For example, when excessive roughness is detected on the cut surface of carbon steel, the process parameter database is queried to determine that excessive roughness is usually caused by the focal point position deviating from the optimal value or insufficient auxiliary gas pressure. Based on this, a second process adjustment command is generated to move the focal point position upward by 0.5 mm or increase the auxiliary gas pressure by 15%.

[0088] Subsequently, either the first or second process adjustment command is sent to the PLC controller. The PLC controller adjusts the laser cutting equipment's cutting parameters in real time, such as laser power, cutting speed, focal position, or auxiliary gas pressure, based on the received command, achieving closed-loop control of "image acquisition, AI analysis, parameter recommendation, and equipment execution." After adjustment, the system re-acquires the cut surface image and performs another quality assessment to verify the adjustment effect. If defects are still not eliminated, iterative adjustments are made until the cutting quality meets the preset standards.

[0089] In one embodiment, the method further includes: Acquire the image sequence of the historical cutting area corresponding to each processing moment in the continuous processing of the workpiece; Obtain the historical quality detection results corresponding to each historical cut region image in the historical cut region image sequence; Determine the trend of quality change of the cut section of the workpiece to be processed based on historical quality inspection results; Acquire historical working data of the cutting head during continuous processing and / or historical power attenuation data of the laser during continuous processing; Based on the trend of the quality change of the cutting section and historical working data and / or historical power attenuation data, a loss prediction correlation model is established, and the loss state of the cutting head and / or laser is predicted based on the loss prediction correlation model to obtain the loss prediction results. A process adjustment prediction scheme is generated based on the loss prediction results.

[0090] Specifically, the image sequence of the historical cutting area corresponding to each processing moment of the workpiece to be processed is obtained. The image sequence is arranged in the order of processing time and covers the complete processing process from the start of the equipment to the current moment.

[0091] The system acquires historical quality inspection results for each historical cut area image in the historical cut area image sequence, i.e., the defect category label and quality score output by each image after processing by the quality assessment model. Based on the historical quality inspection results, it determines the trend of quality change in the cut surface of the workpiece to be processed, such as statistically analyzing the change curve of slag formation frequency over time and the decay trend of roughness score. Simultaneously, it acquires historical operating data of the cutting head during continuous processing (such as cutting head temperature, vibration amplitude, and focusing lens contamination level) and / or historical power attenuation data of the laser during continuous processing (such as the change curve of laser output power over time and pump source attenuation data).

[0092] Then, based on the trend of cut surface quality changes and historical operating data and / or historical power attenuation data, a loss prediction correlation model is established. This correlation model uses time-series prediction models such as Long Short-Term Memory Networks (LSTM) or Gated Recurrent Units (GRUs) to jointly model historical image quality sequences with equipment operating state sequences. The loss prediction correlation model is used to predict the loss state of the cutting head and / or laser, yielding loss prediction results, such as predicting that the focusing lens of the cutting head needs to be replaced after 200 cuts remaining, and that the laser pump source's power attenuates below a threshold after 500 hours of operation remaining.

[0093] Finally, based on the loss prediction results, a process adjustment prediction scheme is generated. For example, when the laser power decays to a certain level, the initial laser power setting value is automatically increased to compensate for the decay, or maintenance is arranged in advance before the contamination of the cutting head lens reaches the threshold.

[0094] This application provides a quality assessment method, including: acquiring an image of the cutting area corresponding to the workpiece to be processed; preprocessing and data augmentation of the cutting area image to obtain an enhanced image sample set; inputting the enhanced image sample set into a pre-trained quality assessment model, and performing downsampling and fusion processing on the enhanced image sample set through a feature extraction branch to obtain a multi-scale fusion feature map, wherein the quality assessment model includes a feature extraction branch and a classification branch; performing classification processing on the multi-scale fusion feature map through the classification branch to obtain a quality detection result corresponding to the workpiece to be processed; and matching the quality detection result with a preset process parameter database to generate a target process parameter recommendation scheme corresponding to the workpiece to be processed. This application effectively expands the defect sample library under different material, thickness, and laser light source combinations through image preprocessing and data augmentation techniques, overcoming the problem of overfitting due to small sample training. Combined with multi-scale downsampling and fusion strategies in the feature extraction branch, it accurately captures multi-level features such as cross-sectional edges, geometric contours, and deep defect textures while ensuring lightweight computation. This achieves high-precision classification and localization of defects such as slag, cracks, and excessive roughness in the cut sections of materials such as stainless steel, carbon steel, and aluminum alloys, resulting in high defect recognition accuracy. Furthermore, this method intelligently matches quality inspection results with a preset process parameter database, quickly selecting the optimal process adjustment scheme based on defect category, material type, thickness, and light source information. This is applicable to all laser-cut parts and improves the accuracy and efficiency of quality assessment.

[0095] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0096] The following are device embodiments of this application. For details not described in detail, please refer to the corresponding method embodiments described above.

[0097] Figure 6 This diagram illustrates a structural schematic of a quality assessment device according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. The quality assessment device includes: The image acquisition module 601 is used to acquire the image of the cutting area corresponding to the workpiece to be processed, wherein the image of the cutting area is an RGB image obtained by acquiring the cutting cross-section of the workpiece to be processed; Image processing module 602 is used to preprocess and data augmentation process the image of the cut area to obtain an enhanced image sample set; The feature extraction module 603 is used to input the enhanced image sample set into the pre-trained quality assessment model. The enhanced image sample set is downsampled and fused through the feature extraction branch to obtain a multi-scale fused feature map. The quality assessment model includes a feature extraction branch and a classification branch. The feature classification module 604 is used to classify the multi-scale fused feature map through classification branches to obtain the quality inspection result corresponding to the workpiece to be processed; The parameter recommendation module 605 is used to match the quality inspection results with the preset process parameter database to generate a target process parameter recommendation scheme for the workpiece to be processed.

[0098] In one embodiment, the image acquisition module 601 is further configured to fix the workpiece to be processed at the detection station; The image acquisition device is controlled to acquire images of the cut section of the workpiece to be processed according to preset acquisition parameters, so as to obtain the initial cutting image; The initial cut image is scaled up according to a preset size to obtain the cut area image.

[0099] In one embodiment, the image processing module 602 is further configured to perform grayscale correction processing on the image of the cut area to obtain a grayscale corrected image; The grayscale-corrected image is subjected to noise filtering to obtain a preprocessed image; Perform at least one or more of random rotation, mirroring, and cropping on the preprocessed image to obtain an enhanced image sample set.

[0100] In one embodiment, the feature extraction module 603 is further configured to perform multiple downsampling processes on the enhanced image sample set through the feature extraction branch, and extract shallow-level features and deep-level features respectively; By fusing shallow and deep features through element-wise addition, a multi-scale fused feature map is obtained.

[0101] In one embodiment, the feature extraction module 603 is further configured to perform a first downsampling process on the enhanced image sample set, extract the cross-sectional edge features in the enhanced image sample set, and obtain a first shallow feature map; The first shallow feature map is downsampled a second time to extract the cross-sectional geometric features from the enhanced image sample set, thus obtaining the second shallow feature map. The second shallow feature map is downsampled a third time to extract the defect texture features from the enhanced image sample set to obtain a deep feature map. The defect texture features include at least one or more of the following: slag texture features, crack texture features, overheated texture features, and drag line texture features.

[0102] In one embodiment, the feature classification module 604 is further used to input the multi-scale fused feature map into the segmentation sub-network in the classification branch, to perform pixel-level localization of the defect region in the cut section of the workpiece to be processed, and to obtain a defect segmentation mask. The multi-scale fused feature map is input into the classification sub-network in the classification branch to classify and identify each defect region in the defect segmentation mask, and obtain the defect category label corresponding to each defect region. The quality inspection results are generated based on the defect category labels corresponding to each defect area.

[0103] In one embodiment, the parameter recommendation module 605 is further configured to query associated data in a preset process parameter database based on the defect category label in the quality inspection result, and obtain candidate process parameter adjustment schemes corresponding to the defect category label, wherein the associated data is the mapping data between defect categories and process parameters; Obtain information on the material type, thickness, and laser source type of the workpiece to be processed; Using material type information, thickness information, and laser source type information as filtering criteria, a target process parameter adjustment scheme that matches the workpiece to be processed is selected from the candidate process parameter adjustment schemes. A recommended scheme for the target process parameters is generated based on the adjustment scheme for the target process parameters.

[0104] In one embodiment, the apparatus further includes a closed-loop control module, which is configured to generate a first process adjustment command for reducing laser power or increasing cutting speed based on associated data in a preset process parameter database in response to a defect category label of slag adhesion defect in the quality inspection result. In response to the defect category label of excessive roughness in the quality inspection results, a second process adjustment command is generated based on the associated data in the preset process parameter database to adjust the coke depth or the auxiliary gas pressure. The first or second process adjustment command is sent to the PLC controller to perform closed-loop adjustment of the real-time cutting parameters of the workpiece to be processed.

[0105] In one embodiment, the apparatus further includes: a loss prediction module, which is used to acquire a sequence of historical cutting area images corresponding to each processing moment of the workpiece during continuous processing; Obtain the historical quality detection results corresponding to each historical cut region image in the historical cut region image sequence; Determine the trend of quality change of the cut section of the workpiece to be processed based on historical quality inspection results; Acquire historical working data of the cutting head during continuous processing and / or historical power attenuation data of the laser during continuous processing; Based on the trend of the quality change of the cutting section and historical working data and / or historical power attenuation data, a loss prediction correlation model is established, and the loss state of the cutting head and / or laser is predicted based on the loss prediction correlation model to obtain the loss prediction results. A process adjustment prediction scheme is generated based on the loss prediction results.

[0106] This application provides a quality assessment device, specifically used for: acquiring an image of the cutting area corresponding to the workpiece to be processed; preprocessing and data augmentation of the cutting area image to obtain an enhanced image sample set; inputting the enhanced image sample set into a pre-trained quality assessment model, and performing downsampling and fusion processing on the enhanced image sample set through a feature extraction branch to obtain a multi-scale fusion feature map, wherein the quality assessment model includes a feature extraction branch and a classification branch; classifying the multi-scale fusion feature map through the classification branch to obtain a quality detection result corresponding to the workpiece to be processed; and matching the quality detection result with a preset process parameter database to generate a target process parameter recommendation scheme corresponding to the workpiece to be processed. This application effectively expands the defect sample library under different material, thickness, and laser light source combinations through image preprocessing and data augmentation techniques, overcoming the problem of overfitting due to small sample training. Combined with multi-scale downsampling and fusion strategies in the feature extraction branch, it accurately captures multi-level features such as cross-sectional edges, geometric contours, and deep defect textures while ensuring lightweight computation. This achieves high-precision classification and localization of defects such as slag, cracks, and excessive roughness in the cut sections of materials such as stainless steel, carbon steel, and aluminum alloys, resulting in high defect recognition accuracy. Furthermore, this method intelligently matches quality inspection results with a preset process parameter database, quickly selecting the optimal process adjustment scheme based on defect category, material type, thickness, and light source information. This is applicable to all laser-cut parts and improves the accuracy and efficiency of quality assessment.

[0107] This application Figure 7 A schematic diagram of a computer device is provided. (Example) Figure 7 As shown, the computer device 7 in this embodiment includes a processor 701, a memory 702, and a computer program 707 stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program 707, it implements the steps in the various quality assessment method embodiments described above, for example... Figure 2 Steps 201 to 205 are shown. Alternatively, when processor 701 executes computer program 707, it implements the functions of each module / unit in the above-described quality assessment device embodiments, for example... Figure 6 The functions of modules / units 601 to 605 shown.

[0108] This application also provides a readable storage medium storing a computer program, which, when executed by a processor, is used to implement the quality assessment methods provided in the various embodiments described above.

[0109] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0110] This application also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the execution of the executable instructions by the at least one processor causes the device to implement the quality assessment methods provided in the various embodiments described above.

[0111] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0112] The above 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.

Claims

1. A quality assessment method, characterized in that, include: Obtain the image of the cutting area corresponding to the workpiece to be processed, wherein the image of the cutting area is an RGB image obtained by acquiring the cutting cross-section of the workpiece to be processed; The image of the cut region is preprocessed and data augmented to obtain an enhanced image sample set; The enhanced image sample set is input into a pre-trained quality assessment model. The enhanced image sample set is downsampled and fused through the feature extraction branch to obtain a multi-scale fused feature map. The quality assessment model includes a feature extraction branch and a classification branch. The multi-scale fused feature map is classified by the classification branch to obtain the quality inspection result corresponding to the workpiece to be processed. The quality inspection results are matched with a preset process parameter database to generate a recommended scheme of target process parameters for the workpiece to be processed.

2. The method according to claim 1, characterized in that, The step of acquiring the image of the cutting area corresponding to the workpiece to be processed includes: The workpiece to be processed is fixed at the inspection station; The image acquisition device is controlled to acquire images of the cut section of the workpiece to be processed according to preset acquisition parameters, so as to obtain an initial cutting image; The initial cut image is scaled according to a preset size to obtain the cut area image.

3. The method according to claim 1, characterized in that, The preprocessing and data augmentation of the image of the cut region to obtain an enhanced image sample set includes: The image of the cut area is subjected to grayscale correction processing to obtain a grayscale corrected image; The grayscale corrected image is subjected to noise filtering to obtain the preprocessed image; The preprocessed image is subjected to one or more of random rotation, mirroring, and cropping to obtain the enhanced image sample set.

4. The method according to claim 1, characterized in that, The step of downsampling and fusing the enhanced image sample set through the feature extraction branch to obtain a multi-scale fused feature map includes: The enhanced image sample set is downsampled multiple times through the feature extraction branch to extract shallow-level features and deep-level features respectively; The multi-scale fused feature map is obtained by fusing the shallow-level features and the deep-level features through element-wise addition.

5. The method according to claim 4, characterized in that, The step of performing multiple downsampling processes on the enhanced image sample set through the feature extraction branch to extract shallow-level features and deep-level features respectively includes: The enhanced image sample set is subjected to a first downsampling process to extract the cross-sectional edge features in the enhanced image sample set, thereby obtaining a first shallow feature map; The first shallow feature map is downsampled a second time to extract the cross-sectional geometric features from the enhanced image sample set, thus obtaining the second shallow feature map. The second shallow feature map is subjected to a third downsampling process to extract the defect texture features from the enhanced image sample set, thereby obtaining the deep feature map. The defect texture features include at least one or more of the following: slag texture features, crack texture features, overheated texture features, and drag line texture features.

6. The method according to claim 1, characterized in that, The step of classifying the multi-scale fused feature map through the classification branch to obtain the quality inspection result corresponding to the workpiece to be processed includes: The multi-scale fused feature map is input into the segmentation sub-network in the classification branch to perform pixel-level localization of the defect region in the cut section of the workpiece to be processed, thereby obtaining a defect segmentation mask. The multi-scale fused feature map is input into the classification sub-network in the classification branch to classify and identify each defect region in the defect segmentation mask, and obtain the defect category label corresponding to each defect region. The quality inspection results are generated based on the defect category labels corresponding to each defect area.

7. The method according to claim 1, characterized in that, The step of matching the quality inspection results with a preset process parameter database to generate a recommended target process parameter scheme for the workpiece to be processed includes: Based on the defect category label in the quality inspection result, query the associated data in the preset process parameter database to obtain the candidate process parameter adjustment scheme corresponding to the defect category label, wherein the associated data is the mapping data between defect category and process parameter; Obtain the material type information, thickness information, and laser source type information of the workpiece to be processed; Using the material type information, the thickness information, and the laser source type information as filtering criteria, a target process parameter adjustment scheme that matches the workpiece to be processed is selected from the candidate process parameter adjustment schemes. Based on the target process parameter adjustment scheme, a recommended scheme for the target process parameters is generated.

8. The method according to claim 1, characterized in that, The method further includes: In response to the defect category label of slag adhesion defect in the quality inspection result, a first process adjustment instruction for reducing laser power or increasing cutting speed is generated based on the associated data in the preset process parameter database. In response to the defect category label of excessive roughness in the quality inspection result, a second process adjustment command is generated based on the associated data in the preset process parameter database to adjust the coke depth or the auxiliary gas pressure. The first process adjustment command or the second process adjustment command is sent to the PLC controller to perform closed-loop adjustment of the real-time cutting parameters of the workpiece to be processed.

9. The method according to claim 1, characterized in that, The method further includes: Obtain the historical cutting area image sequence corresponding to each processing moment during the continuous processing of the workpiece; Obtain the historical quality detection results corresponding to each historical cut region image in the historical cut region image sequence; The trend of quality change of the cut section of the workpiece to be processed is determined based on the historical quality inspection results. Acquire historical operating data of the cutting head during the continuous processing and / or historical power attenuation data of the laser during the continuous processing; Based on the trend of the quality change of the cutting section and the historical working data and / or the historical power attenuation data, a loss prediction correlation model is established, and the loss state of the cutting head and / or the laser is predicted based on the loss prediction correlation model to obtain the loss prediction result. A process adjustment prediction scheme is generated based on the loss prediction results.

10. A quality assessment device, characterized in that, include: The image acquisition module is used to acquire the image of the cutting area corresponding to the workpiece to be processed, wherein the image of the cutting area is an RGB image obtained by acquiring the cutting cross-section of the workpiece to be processed; The image processing module is used to preprocess and augment the image of the cut area to obtain an augmented image sample set. The feature extraction module is used to input the enhanced image sample set into a pre-trained quality assessment model, and to perform downsampling and fusion on the enhanced image sample set through the feature extraction branch to obtain a multi-scale fused feature map. The quality assessment model includes a feature extraction branch and a classification branch. The feature classification module is used to classify the multi-scale fused feature map through the classification branch to obtain the quality inspection result corresponding to the workpiece to be processed. The parameter recommendation module is used to match the quality inspection results with a preset process parameter database to generate a target process parameter recommendation scheme for the workpiece to be processed.

11. A computer device, characterized in that, Includes a memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the quality assessment method as described in any one of claims 1 to 9.

Images