Part defect detection method and system based on image model

Abstract

The invention relates to the technical field of image data processing, in particular to a part defect detection method and system based on an image model. The method comprises the following steps: acquiring a multi-source image of a part; performing image preprocessing on the multi-source image of the part to obtain a multi-source standard image of the part; constructing a neural topological structure of the parts based on the multi-source preprocessing data of the parts; constructing a feature extraction layer according to the neural topological structure of the part to obtain a part defect detection layer; performing non-parametric feature fusion based on the part defect detection layer, and constructing a part detection model; therefore, through multi-source data fusion, neural topology modeling and adaptive optimization, the problems of feature loss, insufficient detection precision, inaccurate defect positioning and the like in a traditional detection method are solved, and the reliability, precision and automation level of part defect detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a component defect detection method and system based on an image model. Background Art

[0002] Traditional component defect detection methods mostly rely on rule-based inspection techniques or manual visual inspection, which have various limitations. First, rule-based inspection methods typically determine defect types by setting fixed thresholds and rules. However, in real-world applications, component defects are often diverse and complex, making accurate identification difficult using simple rules. This makes rule-based methods prone to missed or false positives when faced with complex or subtle defects. While manual visual inspection methods can intuitively detect surface defects on components, they rely heavily on manual experience and are subject to significant subjectivity, resulting in low detection efficiency and unstable results. In production environments, manual inspection also struggles to adapt to the demands of efficient production lines, making it difficult to ensure accurate and consistent inspection of every component. Existing computer vision methods, despite incorporating image processing techniques, typically rely on manual image feature extraction and lack automated understanding of component defects in complex scenarios. Traditional image processing techniques primarily extract image features through methods such as edge detection and texture analysis. However, these methods lack robustness in complex backgrounds, varying lighting conditions, or diverse defect patterns, making them difficult to cope with the variability inherent in real-world production. Traditional defect detection systems generally lack the ability to comprehensively analyze multimodal data (such as images, thermal maps, and 3D scans). Component defect detection relies not only on surface images but also requires consideration of multiple factors, including internal defects, stress distribution, and heat conduction. However, existing technologies often only process a single type of data, failing to integrate multi-dimensional data and implement deep learning analysis. This results in insufficient comprehensiveness and accuracy in detection. Summary of the Invention

[0003] Based on this, it is necessary to provide a component defect detection method and system based on an image model to solve at least one of the above technical problems.

[0004] To achieve the above object, a component defect detection method based on an image model is provided, the method comprising the following steps:

[0005] Step S1: Acquire multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct a neural topology structure of components based on the multi-source preprocessed data of components;

[0006] Step S2: Construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and construct a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return them to the model to obtain a component detection optimization model;

[0007] Step S3: Acquire component measured data; use the component measured data as input to output component detection optimization results based on the component detection optimization model; perform defect comparison between the component measured data and the component detection optimization results to obtain component defect comparison data;

[0008] Step S4: performing bearing pitting anchoring confirmation on the component defect comparison data to obtain bearing pitting anchoring confirmation data; performing defect detection evaluation based on the bearing pitting anchoring confirmation data, and generating an image model component defect detection report.

[0009] The beneficial effect of the present invention is that by acquiring multi-source images of parts and performing image preprocessing, multi-source standard images are generated, ensuring the consistency of image quality and the integrity of feature information at the data level. Subsequently, a neural topology structure of the parts is constructed based on the preprocessed data, and by mapping the multi-dimensional feature relationships of the parts, the geometric shape, texture features and potential defect information are effectively integrated to achieve efficient expression of cross-modal features. The defect detection layer is constructed in the feature extraction layer, enabling the system to deeply mine defect features at different scales and complex backgrounds, and improve the robustness and generalization ability of detection through non-parametric feature fusion. The establishment of the parts detection model further relies on automated feature selection and multi-dimensional feature aggregation, effectively avoiding errors caused by manual intervention. After the detection results are output, the model is adaptively optimized through a hyperparameter update mechanism to form an iteratively optimized parts detection optimization model, ensuring continuous improvement in detection accuracy. In the actual detection process, the measured data of the parts are input into the optimized detection model to obtain accurate detection results, and the accuracy and stability of the detection are evaluated through defect comparison analysis. Furthermore, for bearing pitting, a typical defect type, the system utilizes a pitting anchoring confirmation method to precisely locate local defects. Combined with inspection and evaluation, the system further quantifies the degree of defects, providing a reliable basis for subsequent maintenance and quality control. Finally, component defect inspection reports are generated using image models, enabling intuitive presentation of inspection results and supporting decision-making. Therefore, through multi-source data fusion, neural topology modeling, and adaptive optimization, this invention addresses the problems of feature loss, insufficient detection accuracy, and inaccurate defect location in traditional inspection methods, thereby improving the reliability, accuracy, and automation level of component defect detection.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: setting the resolution of the 16K linear array industrial camera to 5 μm / pixel and the wavelength range to 450-650 nm;

[0012] Step S12: setting the Z-axis repeatability of the line laser 3D scanner to ±1 μm and the scanning width to 200 mm;

[0013] Step S13: Setting the thermal sensitivity of the uncooled infrared thermal imager to less than 30 mK;

[0014] Step S14: using a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager to collect multi-source data of components;

[0015] Step S15: performing data preprocessing on the component multi-source image data to obtain component multi-source image preprocessing data; and constructing a component neural topology structure based on the component multi-source preprocessing data.

[0016] This invention achieves multi-source data acquisition of parts and components through the coordinated application of a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager, ensuring the multidimensionality and integrity of image information at the data level. The industrial camera, with a high resolution of 5μm / pixel and a wavelength range of 450-650nm, provides detailed surface texture and optical property data, facilitating the detection of surface defects, microcracks, scratches, and other anomalies. The line laser 3D scanner, with its ±1μm Z-axis repeatability and 200mm scanning width, provides precise three-dimensional topography data, capturing the spatial contours and dimensional information of parts and components, providing a basis for subsequent shape deviation detection and dimensional measurement. The uncooled infrared thermal imager, with a thermal sensitivity of less than 30mK, captures the temperature field distribution on the surface of parts and components in real time, making it particularly suitable for identifying thermal anomalies and structural hazards within materials. In the data preprocessing stage, noise interference and displacement errors present in the data from different sensors are eliminated through methods such as denoising, calibration, and alignment, ensuring the consistency of multi-source data in space and time. Feature extraction methods are further employed to quantitatively describe key features in the image data, such as texture, edges, surface defects, and thermal distribution, generating standardized multi-source image preprocessing data. Subsequently, a component neural topology is constructed based on this multi-source preprocessing data. The data from different sensors is mapped and fused in a multidimensional space, establishing a network of node and edge relationships. This topological structure fully expresses the correlation between the component's surface information, topography, and temperature field data, providing a unified and efficient feature representation for subsequent defect detection and condition assessment.

[0017] Preferably, the data preprocessing of the multi-source image data of the component in step S1 includes:

[0018] Use 16K linear array industrial cameras, line laser 3D scanners, and uncooled infrared thermal imagers to collect multi-source data on parts, including visible light data images, 3D scanning data images, and thermodynamic data of parts;

[0019] Extracting component illumination texture from the component visible light data image to obtain component illumination texture data;

[0020] Perform point cloud geometry analysis on the 3D scanning data image of the parts to obtain the point cloud geometry data of the parts;

[0021] The thermodynamic data of components are divided into regional gradients and hot spot distribution is arranged to obtain component hot spot arrangement data.

[0022] This invention achieves comprehensive data collection from multiple sources of component data through the combined application of a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager, enhancing the multi-dimensional representation of component features at the data level. Visible light data images of components are acquired by the industrial camera, providing detailed information on surface texture, color distribution, and localized microscopic defects at a high resolution of 5μm / pixel. Light texture extraction methods are used to capture the microstructural features of the component surface, generating component light texture data that further reveals the uniformity and optical properties of the surface material and facilitates the identification of minor defects such as scratches and cracks. Regarding three-dimensional topography information, 3D scanned component data images captured by the line laser 3D scanner are analyzed through point cloud geometry to generate component point cloud geometry data. Point cloud data accurately depicts component geometric contours, edge features, and shape deviations in spatial coordinate form, offering significant advantages in dimensional inspection of complex curved surfaces or precision parts. Combined with point cloud geometry data, component shape restoration, flatness inspection, and assembly accuracy assessment can be performed. In the thermodynamic information extraction link, the thermodynamic data of the components collected by the uncooled infrared thermal imager is finely divided into temperature field changes through the regional gradient division method, and further formed into component hot spot arrangement data based on the hot spot distribution arrangement method. This data effectively characterizes the thermal distribution characteristics of the components and reveals the thermal anomaly areas inside the material. It plays an important role in the internal defect detection, material uniformity assessment and stress concentration area identification of components. Through the comprehensive application and feature extraction of these multi-source data, the present invention forms a multi-scale, multi-modal feature set from the surface to the inside, and constructs a complete component status information expression. Compared with the traditional single data source detection method, the present invention significantly improves the accuracy of defect recognition, the detection capability of complex shapes and the sensitivity of thermal anomaly detection at the data level.

[0023] Preferably, the step S1 of constructing the component neural topology structure includes:

[0024] Image feature extraction is performed on the component illumination texture data, that is, the directional gradient of oxidation spots is extracted from the component illumination texture data, where the directional gradient of oxidation spots ranges from 0.1 to 5.0 pixels / μm, and the first layer structure at the bottom of the surface is constructed; local convex hull curvature analysis is performed on the component illumination texture data, where the main curvature range is ±0.01 to 0.5 mm -1 , construct the surface primary mid-end second layer structure; extract the relative position assembly error between components based on the component illumination texture data, where the error range is ±50~500μm, and construct the macro high-level third layer structure;

[0025] The luminous flux neural topology map is constructed by constructing the first layer structure at the bottom of the surface, the second layer structure at the primary middle end of the surface, and the third layer structure at the macroscopic high level, where the luminous flux density ranges from 0.1 to 10 lx·μ -2 , obtain the component luminous flux texture structure;

[0026] Construct a component spectral clustering graph structure based on the 3D scanning data of the components;

[0027] Create three-dimensional heat conduction residual data of components based on the thermodynamic data of components;

[0028] Project the component's luminous flux texture structure onto the component's spectral clustering graph structure, mark the surface abnormal area, and obtain the component's abnormal projection area;

[0029] The abnormal response of the abnormal projection area of ​​the component is reversely modulated using the three-dimensional heat conduction residual data of the component, and multi-dimensional neural topology constraints are performed to obtain the neural topology structure of the component.

[0030] The present invention extracts the directional gradient of oxidation spots from the illumination texture data of components to form a first-layer structure at the bottom of the surface, effectively capturing local chemical corrosion traces and the distribution characteristics of microscopic oxidation spots on the material surface. The directional gradient of oxidation spots is set in the range of 0.1-0.5, which helps to identify structural defects such as topographic changes, scratches, and wear on the material surface. Furthermore, the macroscopic high-level third-layer structure is formed by extracting the relative position assembly errors between components. The optical feature distribution with an error range of 0.1-100lx forms the component luminous flux texture structure, thereby comprehensively characterizing the optical properties and defect areas of the material surface. On this basis, a spectral clustering graph structure is constructed using the 3D scanning data of the component, and spatial feature clustering is performed on the point cloud data to form a detailed three-dimensional contour feature, supporting further spatial anomaly detection. At the same time, three-dimensional heat conduction residual data is created based on the thermodynamic data of the component, revealing the thermal anomaly distribution and local thermal stress characteristics within the material. By projecting the luminous flux texture structure onto the spectral clustering graph structure, the surface anomaly area is accurately located, and the anomaly area is marked to form the component anomaly projection area. Finally, the three-dimensional heat conduction residual data is used to reversely modulate the abnormal response of the abnormal projection area, and the accuracy and completeness of abnormality recognition are further optimized through multi-dimensional neural topology constraints, and finally the neural topology structure of the component is obtained.

[0031] Preferably, step S2 includes the following steps:

[0032] Step S21: constructing a feature extraction layer based on the component neural topology structure to obtain a component defect detection layer;

[0033] Step S22: performing non-parametric feature fusion based on the component defect detection layer to obtain component field effect fusion data, wherein the non-parametric feature fusion includes equivalent charge density extraction, phonon spectrum heat flow path optimization, and material viscosity dissipation control;

[0034] Step S23: constructing a detection model for the component field effect fusion data to obtain a component detection model;

[0035] Step S24: Output the component detection results according to the component detection model; perform defect simulation verification based on the component detection result output to obtain component detection data; update hyperparameters based on the component detection data, and return the model to obtain a component detection optimization model.

[0036] The present invention constructs a feature extraction layer to analyze and reduce the multi-layer features in the neural topology step by step to form a component defect detection layer. In this process, the microscopic defect features, geometric distortion information and thermal anomaly distribution of the material surface are captured, thereby ensuring the multi-scale and multi-dimensional feature integrity. Subsequently, non-parametric feature fusion is performed based on the defect detection layer to further improve the accuracy and information content of the data expression. Through equivalent charge density extraction, the charge distribution anomalies on the surface and inside the component are identified, thereby revealing the existing material surface defects or local structural anomalies. At the same time, phonon spectrum heat flow path optimization is used to analyze the thermal conductivity characteristics inside the material, and locate potential microcracks or interface delamination by detecting abnormal changes in the phonon propagation path. Material viscosity dissipation control is based on the internal energy dissipation characteristics of the material to accurately identify stress concentration areas and effectively detect potential risks caused by material fatigue or stress corrosion. The component field effect fusion data formed by this series of feature fusions fully characterizes the comprehensive response characteristics of the component under the coupling of electrical, thermal and mechanical fields. During the inspection model construction phase, field-effect fusion data is used to train the component inspection model, enabling the model to perform deep learning and anomaly identification on multi-source heterogeneous data. The inspection model not only outputs component defect detection results but also further calibrates these results through defect simulation verification to ensure detection accuracy and reliability. Based on the inspection data, the model is dynamically optimized through a hyperparameter update mechanism to form an optimized component inspection model. This optimization process adjusts the model's parameter settings based on feedback from inspection data, continuously improving the model's adaptability to different operating conditions.

[0037] Preferably, step S21 includes the following steps:

[0038] Step S211: performing texture defect recognition on the component neural topology and forming a component curvature flow differential operator; extracting spatial abnormality features based on the component curvature flow differential operator and constructing a component surface defect layer;

[0039] Step S212: Determine the virtual sound pressure distribution of the component neural topology at a second harmonic frequency of 10 MHz, perform acoustic debonding analysis, and obtain a virtual acoustic signal of the component with a temperature rise of ΔT>1.2K; construct an internal defect layer of the component based on the virtual acoustic signal of the component;

[0040] Step S213: naturally transform the component surface defect layer and the component internal defect layer to obtain a component defect detection layer.

[0041] This method uses neural topology to identify texture defects and utilizes the curvature flow differential operator to capture local curvature variations in component surface topography. The curvature flow differential operator has strong spatial sensitivity and can effectively identify abnormal features such as tiny protrusions, depressions, cracks, and scratches on the material surface. At the data level, this method avoids misjudgments caused by noise or local shadows in traditional surface inspection, while ensuring accurate modeling of complex surface textures through the spatial continuity characteristics of curvature flow. Subsequently, for internal defect detection of components, acoustic debonding analysis is performed using a virtual sound pressure distribution at a second harmonic frequency of 10 MHz. The simulated virtual acoustic signal can identify hidden defects such as tiny cracks, delamination, debonding, and voids within the material. The sound pressure distribution data provides a comprehensive representation of the component's internal structural state. Combined with a temperature rise threshold of ΔT>1.2K, this further improves the sensitivity and specificity of the acoustic analysis. At the data level, the nonlinear characteristics of the acoustic signal and the temperature rise information together construct a virtual acoustic signal dataset for the component, providing a reliable basis for accurate internal defect detection. Ultimately, through a natural transformation method, the component surface defect layer and internal defect layer are non-destructively mapped to form a component defect detection layer. Natural transformation ensures seamless integration of feature information between different data layers, enabling comprehensive, multi-dimensional defect detection from the surface to the interior. At the data level, this approach effectively resolves the issues of traditional methods that separate surface and internal inspection results and fragment feature information, thereby improving the accuracy and consistency of defect detection.

[0042] Preferably, step S22 includes the following steps:

[0043] Step S221: Extract the characteristic equivalent charge density of the component defect detection layer, with a range of 1e -5 ~e -3 , and perform charge mapping to obtain the virtual charge density of the components;

[0044] Step S222: Optimize the phonon spectrum heat flow path of the component defect detection layer, where the thermal conductivity correction range is ±15% to 35%, and analyze the material lattice propagation direction, where the lattice constant accuracy is Obtain component material lattice data;

[0045] Step S223: Control the material viscosity dissipation of the component defect detection layer, where the viscosity coefficient range is 1e -3 ~e 3 Pa·s, to obtain the viscosity analysis data of component materials;

[0046] Step S224: fusion of component virtual charge density, component material lattice data, and component material viscosity analysis data through implicit physical verification to obtain component field effect fusion data.

[0047] The present invention detects the abnormal charge distribution on the surface and inside of the material during the extraction of the characteristic equivalent charge density, and identifies the defect characteristics caused by corrosion, cracks or material inhomogeneity. The range of equivalent charge density is 1e -5 ~e -3 Under this setting, the charge accumulation and local charge dissipation phenomena in the material microstructure are effectively captured. Through charge mapping technology, these charge characteristics are spatially projected to form virtual charge density data of components, providing a basis for further defect positioning. Secondly, through phonon spectrum heat flow path optimization, the thermal conductivity characteristics inside the components are analyzed, and the thermal conductivity correction range is ±15% Pa·s, so as to identify the energy dissipation characteristics inside the material. Viscous dissipation data reveals the internal damage evolution law of the material under stress, especially for defects such as fatigue cracks and local plastic deformation, providing an effective detection method. Finally, through implicit physical verification fusion, the virtual charge density of components, material lattice data and material viscosity analysis data are jointly analyzed in multiple physical fields to form component field effect fusion data. This fusion method not only enhances the correlation between different physical field characteristics, but also realizes comprehensive detection of cross-scale defects at the data level. Compared with traditional single-field detection methods, the present invention significantly improves the complementarity of multi-field data and the accuracy of defect detection.

[0048] Preferably, the defect comparison between the component measured data and the component inspection optimization result in step S3 includes:

[0049] Compare the defects of the actual measured data of parts and components with the optimized results of component inspection to obtain comparative defect data;

[0050] When the missed detection rate of the comparative defect data is greater than 0.15%, it is recorded as comparative defect marking data;

[0051] The inference speed is determined based on the comparison defect key mark data. If the inference speed is less than the 50ms production line beat, it is recorded as comparison defect production line data.

[0052] The comparison defect key marking data and the comparison defect production line data are weighted and fused to obtain the component defect comparison data.

[0053] This invention effectively identifies missed detections in the detection model by cross-validating detection optimization results with actual measurement results. When the missed detection rate of the comparison defect data exceeds 0.15%, it is recorded as comparison defect flag data. This threshold setting balances detection accuracy and error control, ensuring that even minor defects are effectively captured. At the data level, the missed detection flag data not only reflects local failure areas of the detection model but also provides a basis for targeted optimization. Subsequently, the inference speed during the detection process is evaluated using the comparison defect key flag data. When the inference speed is less than 50ms and meets the production line cycle requirements, comparison defect production line data is generated. This step ensures the stability and efficiency of the detection model in an industrial production environment through real-time inference speed monitoring. The dynamic inference speed control mechanism further reduces the risk of production interruptions caused by detection delays, ensuring the timeliness of the system at the data level. In the weight fusion stage, the comparison defect key flag data and the comparison defect production line data are fused to form component defect comparison data. This data fusion process comprehensively considers the severity of the missed detection flags and the actual performance of the inference speed. Through an adaptive weight adjustment mechanism, the fused data is more comprehensive and accurate. Compared with the traditional single accuracy evaluation method, the present invention achieves dual optimization of detection accuracy and reasoning efficiency at the data level, which not only improves the comprehensiveness of defect identification, but also effectively ensures the stable operation of the detection system on a high-beat production line.

[0054] Preferably, the defect detection and evaluation based on the component defect comparison data in step S4 includes:

[0055] Generate material fatigue curve based on bearing pitting anchor confirmation data;

[0056] Perform crack detection based on the material fatigue curve and calculate the material defect sensitivity coefficient to obtain the component defect coefficient;

[0057] The defect coefficient of the component is used to deduce the root cause of the defect using the preset defect coefficient judgment threshold. When the defect coefficient of the component is greater than or equal to the defect coefficient judgment threshold, it is determined to be a process root cause, that is, the component is tested for casting pressure to generate casting pressure test data; the component is tested for pore distribution to obtain pore distribution test data;

[0058] When the defect coefficient of a component is less than the defect coefficient judgment threshold, it is judged to be the root cause of material performance, that is, the hardness of the component is evaluated and calculated to obtain the component hardness performance test data; the abnormal crack extension path of the component is detected to obtain the abnormal crack detection data of the component.

[0059] The present invention establishes a fatigue life characterization model for materials under different stress amplitudes and cycle numbers by generating material SN fatigue curves. SN curve data contains the fatigue limit and fracture behavior characteristics of the material, which helps to reveal the fatigue damage mechanism of the material under actual working conditions. On this basis, crack detection is used to accurately identify cracks on the surface and inside the material, and combined with the calculation of the material defect sensitivity coefficient, the impact of defects on the structural integrity of the material is further quantified. The defect sensitivity coefficient data provides the amplification effect of defects under different stress states, thereby effectively judging the degree of danger of the defects. In the defect root cause deduction stage, the component defect coefficient is classified and processed using a preset defect coefficient judgment threshold. When the defect coefficient is greater than or equal to the judgment threshold, it is inferred that the defect originates from the process root cause, triggering casting pressure detection and pore distribution detection. The casting pressure detection data records the actual distribution state of pressure parameters during the casting process, which helps to locate process defects such as shrinkage cavities and slag inclusions caused by abnormal pressure. The pore distribution detection data is based on high-precision image analysis or CT scanning technology to quantify the spatial distribution and volume proportion of pores in the material, further verifying the stability of the process. On the other hand, when the defect coefficient is less than the judgment threshold, the defect is inferred to be rooted in material performance, and hardness assessment and abnormal crack propagation path detection are performed. Hardness performance test data provides quantitative indicators of material strength and durability based on the material's microstructure and work hardening state. Abnormal crack propagation path detection data uses fracture mechanics to analyze crack propagation direction and rate, revealing the crack propagation mechanism within the material and facilitating the identification of potential failure risks caused by material performance degradation.

[0060] The present invention further provides a component defect detection system based on an image model, which is used to execute the component defect detection method based on an image model as described above. The component defect detection system based on an image model comprises:

[0061] The data preprocessing and topology modeling module is used to obtain multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct the neural topology structure of components based on the multi-source preprocessed data of components;

[0062] The defect detection modeling module is used to construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and build a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return it to the model to obtain a component detection optimization model;

[0063] The detection and optimization module is used to obtain the actual measured data of the parts; use the actual measured data of the parts as input to output the parts detection optimization results based on the parts detection optimization model; compare the defects of the parts measured data and the parts detection optimization results to obtain parts defect comparison data;

[0064] The result verification and report generation module is used to perform bearing pitting anchor confirmation on the component defect comparison data to obtain bearing pitting anchor confirmation data; perform defect detection evaluation based on the bearing pitting anchor confirmation data, and generate an image model component defect detection report.

[0065] The beneficial effect of the present invention is that by acquiring multi-source images of parts and performing image preprocessing, multi-source standard images are generated, ensuring the consistency of image quality and the integrity of feature information at the data level. Subsequently, a neural topology structure of the parts is constructed based on the preprocessed data, and by mapping the multi-dimensional feature relationships of the parts, the geometric shape, texture features and potential defect information are effectively integrated to achieve efficient expression of cross-modal features. The defect detection layer is constructed in the feature extraction layer, enabling the system to deeply mine defect features at different scales and complex backgrounds, and improve the robustness and generalization ability of detection through non-parametric feature fusion. The establishment of the parts detection model further relies on automated feature selection and multi-dimensional feature aggregation, effectively avoiding errors caused by manual intervention. After the detection results are output, the model is adaptively optimized through a hyperparameter update mechanism to form an iteratively optimized parts detection optimization model, ensuring continuous improvement in detection accuracy. In the actual detection process, the measured data of the parts are input into the optimized detection model to obtain accurate detection results, and the accuracy and stability of the detection are evaluated through defect comparison analysis. Furthermore, for bearing pitting, a typical defect type, the system utilizes a pitting anchoring confirmation method to precisely locate local defects. Combined with inspection and evaluation, the system further quantifies the degree of defects, providing a reliable basis for subsequent maintenance and quality control. Finally, component defect inspection reports are generated using image models, enabling intuitive presentation of inspection results and supporting decision-making. Therefore, through multi-source data fusion, neural topology modeling, and adaptive optimization, this invention addresses the problems of feature loss, insufficient detection accuracy, and inaccurate defect location in traditional inspection methods, thereby improving the reliability, accuracy, and automation level of component defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of the steps of a component defect detection method based on an image model;

[0067] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0068] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0069] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative work are within the scope of protection of the present invention.

[0070] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0071] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0072] To achieve this, please refer to Figures 1 to 2 A component defect detection method based on an image model comprises the following steps:

[0073] Step S1: Acquire multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct a neural topology structure of components based on the multi-source preprocessed data of components;

[0074] Step S2: Construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and construct a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return them to the model to obtain a component detection optimization model;

[0075] Step S3: Acquire component measured data; use the component measured data as input to output component detection optimization results based on the component detection optimization model; perform defect comparison between the component measured data and the component detection optimization results to obtain component defect comparison data;

[0076] Step S4: performing bearing pitting anchoring confirmation on the component defect comparison data to obtain bearing pitting anchoring confirmation data; performing defect detection evaluation based on the bearing pitting anchoring confirmation data, and generating an image model component defect detection report.

[0077] The beneficial effect of the present invention is that by acquiring multi-source images of parts and performing image preprocessing, multi-source standard images are generated, ensuring the consistency of image quality and the integrity of feature information at the data level. Subsequently, a neural topology structure of the parts is constructed based on the preprocessed data, and by mapping the multi-dimensional feature relationships of the parts, the geometric shape, texture features and potential defect information are effectively integrated to achieve efficient expression of cross-modal features. The defect detection layer is constructed in the feature extraction layer, enabling the system to deeply mine defect features at different scales and complex backgrounds, and improve the robustness and generalization ability of detection through non-parametric feature fusion. The establishment of the parts detection model further relies on automated feature selection and multi-dimensional feature aggregation, effectively avoiding errors caused by manual intervention. After the detection results are output, the model is adaptively optimized through a hyperparameter update mechanism to form an iteratively optimized parts detection optimization model, ensuring continuous improvement in detection accuracy. In the actual detection process, the measured data of the parts are input into the optimized detection model to obtain accurate detection results, and the accuracy and stability of the detection are evaluated through defect comparison analysis. Furthermore, for bearing pitting, a typical defect type, the system utilizes a pitting anchoring confirmation method to precisely locate local defects. Combined with inspection and evaluation, the system further quantifies the degree of defects, providing a reliable basis for subsequent maintenance and quality control. Finally, component defect inspection reports are generated using image models, enabling intuitive presentation of inspection results and supporting decision-making. Therefore, through multi-source data fusion, neural topology modeling, and adaptive optimization, this invention addresses the problems of feature loss, insufficient detection accuracy, and inaccurate defect location in traditional inspection methods, thereby improving the reliability, accuracy, and automation level of component defect detection.

[0078] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a flow chart of a component defect detection method based on an image model according to the present invention. In this example, the component defect detection method based on an image model includes the following steps:

[0079] Step S1: Acquire multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct a neural topology structure of components based on the multi-source preprocessed data of components;

[0080] In an embodiment of the present invention, multi-source image data of components is acquired, including but not limited to data from different wavelengths, such as visible light images, infrared thermal images, and X-ray images, to comprehensively reflect the physical properties of components under different conditions. Multi-source image data fusion is key to improving inspection effectiveness. By integrating different data sources, more dimensional information can be obtained, effectively avoiding the bias introduced by a single data source, thereby providing a more accurate basis for component defect detection. Next, the multi-source image data of components undergoes image preprocessing, a key step in data processing. Its purpose is to eliminate noise and inconsistencies in the images from different sources through methods such as denoising, contrast enhancement, color balancing, and geometric correction, ensuring consistency and accuracy in the data used in subsequent analysis. Common image processing methods in this process include image filtering (such as Gaussian filtering and median filtering), edge detection (such as the Sobel operator and the Canny operator), deblurring, and image reconstruction. These preprocessing methods can effectively improve image quality and reduce errors in subsequent analysis. A neural topology structure for the components is then constructed based on the preprocessed multi-source image data. In this phase, deep learning and neural network methods are used to design the component's neural topology, enabling the network to extract useful feature information from the image. The construction of this neural network topology typically relies on deep learning algorithms such as convolutional neural networks (CNNs). By extracting hierarchical features from the image, the network gradually learns the image's high-level features, providing data support for subsequent component defect detection.

[0081] Step S2: Construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and construct a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return them to the model to obtain a component detection optimization model;

[0082] In an embodiment of the present invention, based on the component neural topology, a deep learning algorithm is used to extract features from image data, constructing a feature extraction layer. During this process, the neural network extracts low-level and high-level features from the original image based on training data, such as texture, shape, and edge features, which are crucial for defect detection. Through multi-layer convolution and pooling operations, the network effectively extracts these features and provides data support for subsequent defect detection. Next, a component defect detection layer is constructed based on the extracted features, a core step in the entire inspection process. In this layer, by integrating and analyzing multidimensional features, the network can determine component defects, including surface cracks, corrosion, and deformation. Then, using non-parametric feature fusion methods, different types of features are combined to construct a component inspection model. Non-parametric feature fusion typically employs statistical learning-based methods, such as principal component analysis (PCA) and independent component analysis (ICA). By eliminating inter-feature correlations, a more refined and effective feature set is obtained, further improving the accuracy and robustness of the model. Subsequently, the constructed inspection model is used to perform real-time defect inspection on components. The model automatically infers and judges the image data, outputting the component inspection results. This test result provides the data basis for the next step of optimization. To further improve the detection accuracy, this step implements hyperparameter updates based on the data feedback of the component inspection results and transmits the updated hyperparameters back to the detection model.

[0083] Step S3: Acquire component measured data; use the component measured data as input to output component detection optimization results based on the component detection optimization model; perform defect comparison between the component measured data and the component detection optimization results to obtain component defect comparison data;

[0084] In embodiments of the present invention, measured component data is obtained. This data is typically derived from physical measurements of components during actual production, covering information such as size, shape, surface quality, and material properties. This measured data serves as the basis for subsequent analysis and provides a reference standard for comparison. This measured data is then fed into a pre-optimized component inspection optimization model. This optimization model is a deep learning model trained and fine-tuned based on the neural topology and detection layer constructed in previous steps. It aims to more accurately infer component quality by learning from historical data and experience. The model analyzes the input measured component data, performs real-time inference, and outputs inspection results, determining whether a component has defects, as well as the specific type and location of the defects. Next, a comparative defect analysis is performed. Comparing the measured component data with the inspection optimization results generated by the model can identify components that were missed or incorrectly detected in actual production, thereby generating comparative component defect data. This process, by accurately matching and comparing measured results with model outputs, helps identify deficiencies in the inspection model, assesses its accuracy and generalization capabilities, and provides feedback for subsequent model optimization.

[0085] Step S4: performing bearing pitting anchoring confirmation on the component defect comparison data to obtain bearing pitting anchoring confirmation data; performing defect detection evaluation based on the bearing pitting anchoring confirmation data, and generating an image model component defect detection report.

[0086] In an embodiment of the present invention, the anchoring confirmation of bearing pitting is mainly carried out by analyzing the features related to bearing pitting in the defect comparison data, such as surface microstructure, corrosion location and corrosion depth. By comparing historical data and standard models, combined with image processing and feature recognition algorithms, the pitting areas existing on the parts can be accurately identified and anchoring confirmation can be performed. This process usually adopts advanced image processing algorithms, such as edge detection, image segmentation, feature point recognition and other technologies, combined with physical corrosion models to determine the specific location and morphology of bearing pitting. The anchoring confirmation data is the core output of this process, which provides a reliable basis for subsequent defect evaluation and detection report generation. Next, defect detection and evaluation are carried out based on the bearing pitting anchoring confirmation data, mainly by combining data fusion technology and machine learning algorithms to conduct a detailed analysis of the anchored defect area and evaluate its impact on the performance of the parts. This evaluation process usually involves the precise quantification of the defect area, analysis of its key parameters such as size, shape, depth, and prediction of the damage or failure mode caused by the defect. These evaluation results are combined with the defect characteristics to generate an image model part defect detection report.

[0087] Preferably, step S1 includes the following steps:

[0088] Step S11: setting the resolution of the 16K linear array industrial camera to 5 μm / pixel and the wavelength range to 450-650 nm;

[0089] Step S12: setting the Z-axis repeatability of the line laser 3D scanner to ±1 μm and the scanning width to 200 mm;

[0090] Step S13: Setting the thermal sensitivity of the uncooled infrared thermal imager to less than 30 mK;

[0091] Step S14: using a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager to collect multi-source data of components;

[0092] Step S15: performing data preprocessing on the component multi-source image data to obtain component multi-source image preprocessing data; and constructing a component neural topology structure based on the component multi-source preprocessing data.

[0093] In an embodiment of the present invention, by setting the resolution of the 16K linear array industrial camera to 5μm / pixel and determining its wavelength range to be 450-650nm, high-resolution capture of component surface details is achieved. This setting enables the camera to obtain high-definition image data of tiny defects and structural features on the surface of components, and provides high-quality raw data for subsequent image processing. The configuration of the line laser 3D scanner sets its Z-axis repeatability to ±1μm and the scanning width to 200mm, aiming to achieve accurate capture of the three-dimensional geometric shape of components. Through high-precision Z-axis scanning, fine three-dimensional point cloud data can be generated, which is crucial for subsequent defect detection and component geometric feature analysis. By setting the thermal sensitivity of the uncooled infrared thermal imager to less than 30mK, it is ensured that tiny temperature differences on the surface of components can be detected. This thermal sensitivity enables the infrared thermal imager to identify potential thermal anomalies under low temperature differences, thereby providing strong data support for heat conduction analysis and defect detection. By combining a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager, multi-source data on components is collected, including optical images, 3D scanning data, and thermal infrared images. This multimodal data provides rich input for subsequent analysis. Data preprocessing is performed on the acquired multi-source image data, including noise filtering, image enhancement, and thermal map correction, to improve data quality and usability. A neural topology structure of the components is constructed from this preprocessed data. This data can then be used to effectively identify and classify component defects through deep learning or other pattern recognition methods.

[0094] Preferably, the data preprocessing of the multi-source image data of the component in step S1 includes:

[0095] Use 16K linear array industrial cameras, line laser 3D scanners, and uncooled infrared thermal imagers to collect multi-source data on parts, including visible light data images, 3D scanning data images, and thermodynamic data of parts;

[0096] Extracting component illumination texture from the component visible light data image to obtain component illumination texture data;

[0097] Perform point cloud geometry analysis on the 3D scanning data image of the parts to obtain the point cloud geometry data of the parts;

[0098] The thermodynamic data of components are divided into regional gradients and hot spot distribution is arranged to obtain component hot spot arrangement data.

[0099] In this embodiment of the present invention, multi-source data acquisition is performed using a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager. The 16K linear array industrial camera captures visible light images of components, using its high-resolution sensor to detect minute surface texture features. After processing, this texture data can be used to extract illumination texture features on the component surface, generating component illumination texture data, which is crucial for detecting surface defects and microstructures. Next, the line laser 3D scanner scans the component surface, acquiring 3D morphological data and generating a point cloud image. This point cloud data accurately describes the component's geometry, providing a foundation for subsequent shape analysis. Point cloud geometry analysis involves calculating and evaluating the component's 3D structure. By analyzing geometric information such as the relative position, distance, and angle between points in the point cloud, component point cloud geometry data is generated. This data helps identify structural defects, morphological irregularities, and other issues. Finally, the uncooled infrared thermal imager collects thermodynamic data on the component, typically displayed as a temperature distribution image. By segmenting the thermodynamic data into regional gradients, the component surface can be segmented according to different temperature gradients, enabling further analysis of hotspot distribution.

[0100] Preferably, the step S1 of constructing the component neural topology structure includes:

[0101] Image feature extraction is performed on the component illumination texture data, that is, the directional gradient of oxidation spots is extracted from the component illumination texture data, where the directional gradient of oxidation spots ranges from 0.1 to 5.0 pixels / μm, and the first layer structure at the bottom of the surface is constructed; local convex hull curvature analysis is performed on the component illumination texture data, where the main curvature range is ±0.01 to 0.5 mm -1 , construct the surface primary mid-end second layer structure; extract the relative position assembly error between components based on the component illumination texture data, where the error range is ±50~500μm, and construct the macro high-level third layer structure;

[0102] The luminous flux neural topology map is constructed by constructing the first layer structure at the bottom of the surface, the second layer structure at the primary middle end of the surface, and the third layer structure at the macroscopic high level, where the luminous flux density ranges from 0.1 to 10 lx·μ -2 , obtain the component luminous flux texture structure;

[0103] Construct a component spectral clustering graph structure based on the 3D scanning data of the components;

[0104] Create three-dimensional heat conduction residual data of components based on the thermodynamic data of components;

[0105] Project the component's luminous flux texture structure onto the component's spectral clustering graph structure, mark the surface abnormal area, and obtain the component's abnormal projection area;

[0106] The abnormal response of the abnormal projection area of ​​the component is reversely modulated using the three-dimensional heat conduction residual data of the component, and multi-dimensional neural topology constraints are performed to obtain the neural topology structure of the component.

[0107] In an embodiment of the present invention, the illumination texture data of the parts is analyzed in detail, and by extracting the directional gradient of the oxidation spots and combining it with the set gradient range (0.1 to 5.0 pixels / μm), the surface details are accurately modeled to generate the first layer structure at the bottom of the surface. Then, the curvature of the surface of the parts is analyzed using the local convex hull curvature analysis technology to ensure that the tiny features of the surface changes are identified within the main curvature range, and the second layer structure of the primary and middle ends of the surface is further constructed. By extracting the relative position assembly error data between the components, the assembly accuracy of the parts can be accurately analyzed and optimized, and a macroscopic high-level third layer structure can be constructed based on this information. Then, these different levels of structures (the first layer structure at the bottom of the surface, the second layer structure at the primary and middle ends of the surface, and the third layer structure at the macroscopic high-level) are integrated and further constructed through the luminous flux neural topology map, using the range of luminous flux density to provide quantitative illumination information for the data, thereby generating the luminous flux texture structure of the parts. Next, based on the 3D scanning data of the parts, the spectral clustering graph structure of the parts is generated. This step helps to analyze and identify the geometric features of the parts and their clustering characteristics. In thermodynamic data processing, 3D heat conduction residual data of components is created to accurately simulate their thermodynamic behavior. Projecting the light flux texture structure onto the spectral clustering graph structure helps identify and mark surface anomalies. In subsequent processing steps, these anomalies are inversely modulated using the 3D heat conduction residual data and optimized using multidimensional neural topology constraints, resulting in an accurate neural topology for the component.

[0108] It is particularly important to build the component spectral clustering graph structure including:

[0109] Extract point cloud geometric features from 3D scanning data of parts to obtain 3D scanning point cloud data of parts;

[0110] Construct a component spatial similarity matrix based on the 3D scanning point cloud data of the components;

[0111] Calculate the characteristic vectors based on the component space similarity matrix and perform clustering to obtain the local geometric area characteristic data;

[0112] Based on the local geometric area feature data, the neighboring relationship is built to obtain the component spectral clustering graph structure.

[0113] In an embodiment of the present invention, point cloud geometric feature extraction is performed on 3D scanned data of a component. Point cloud processing techniques are then applied to optimize the raw data, remove noise, fill in missing parts, and accurately describe the spatial distribution characteristics of the point cloud. Using the three-dimensional spatial coordinate information, the geometric features of the component surface are extracted, resulting in 3D scanned point cloud data of the component. This data provides the necessary spatial description foundation for subsequent analysis. Next, based on this precisely extracted point cloud data, the spatial similarity of the components is calculated to construct a spatial similarity matrix. This matrix describes the relative positional relationships and geometric similarity between each data point in the point cloud, typically achieved by calculating the Euclidean distance or other similarity metrics between each point. This matrix reveals the overall spatial distribution characteristics of the component. Subsequently, eigenvectors are calculated on the spatial similarity matrix, and the data is reduced in dimensionality using eigenvalue decomposition, followed by further clustering. This clustering process divides the component's geometric features into multiple local regions based on the similarity between the data points. Each region has geometrically similar features, thus providing segmented data for subsequent regional analysis. Finally, based on the local geometric area feature data, proximity relationships are constructed and edges are generated. This process establishes connections between points with similar geometric features through the edge construction method in graph theory, thereby constructing the spectral clustering graph structure of parts, reflecting the spatial connections and similarities between different local areas.

[0114] It is particularly important to create 3D heat conduction residual data of components including:

[0115] Perform three-dimensional spatial mapping on the thermodynamic data of components to obtain thermodynamic mapping data of components;

[0116] Calculate the heat conduction characteristics of the component thermodynamic mapping data to obtain the component heat conduction characteristics data;

[0117] The residual calculation is performed based on the heat conduction characteristic data of the component to obtain the three-dimensional heat conduction residual data of the component.

[0118] In embodiments of the present invention, component thermodynamic data is mapped in three dimensions, projecting information such as heat distribution, temperature gradients, and heat flow paths into three-dimensional space. This process typically utilizes a heat conduction simulation algorithm, such as finite element analysis (FEA) or other numerical methods suitable for thermodynamic simulation, to map component thermodynamic data (e.g., temperature, heat flux, etc.) into a three-dimensional coordinate system, generating component thermodynamic mapping data. This data provides a spatial context for subsequent heat conduction analysis. Heat conduction characteristic calculations are then performed based on these thermodynamic mapping data. Heat conduction characteristic calculations typically consider factors such as thermal conductivity, the thermal diffusion properties of the material, and external heat sources. Heat conduction equations or models are used to solve the heat conduction process within the component, thereby obtaining component thermal conduction characteristic data. This calculation allows the component's thermal response, such as temperature distribution and thermal gradient, to be determined under different operating conditions. Finally, after obtaining the component thermal conduction characteristic data, residual calculations are performed to evaluate the discrepancy between the model's predicted heat conduction behavior and the actual measured data. The goal of residual calculation is to discover anomalies in the heat conduction process and reveal potential problems in the thermal response of components, such as uneven heat sources and low heat conduction efficiency.

[0119] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0120] Step S21: constructing a feature extraction layer based on the component neural topology structure to obtain a component defect detection layer;

[0121] Step S22: performing non-parametric feature fusion based on the component defect detection layer to obtain component field effect fusion data, wherein the non-parametric feature fusion includes equivalent charge density extraction, phonon spectrum heat flow path optimization, and material viscosity dissipation control;

[0122] Step S23: constructing a detection model for the component field effect fusion data to obtain a component detection model;

[0123] Step S24: Output the component detection results according to the component detection model; perform defect simulation verification based on the component detection result output to obtain component detection data; update hyperparameters based on the component detection data, and return the model to obtain a component detection optimization model.

[0124] In an embodiment of the present invention, a feature extraction layer is constructed based on the component neural topology, aiming to identify key features associated with component defects. Neural topology typically relies on deep learning or neural network methods. Through hierarchical feature extraction, complex multidimensional data is mapped into more manageable feature layers, effectively capturing potential component defects under different operating conditions. Next, based on the component defect detection layer, a non-parametric feature fusion approach is used to integrate multiple physical effects with the feature data. Non-parametric feature fusion includes equivalent charge density extraction, phonon spectroscopy heat flow path optimization, and material viscous dissipation control. Equivalent charge density extraction reveals charge distribution information on the material surface, phonon spectroscopy heat flow path optimization helps reveal the heat flow transfer process within the component, and material viscous dissipation control focuses on the material's energy loss behavior. These features work together to comprehensively evaluate the component's thermodynamic, electrical, and mechanical properties. Fusion of these multiple feature information generates component field effect fusion data, which reflects the component's comprehensive performance under complex physical environments. A detection model is then constructed based on this field effect fusion data. This phase uses machine learning algorithms (such as support vector machines and neural networks) combined with previously extracted feature data to build an efficient detection model capable of identifying component defects. Once the detection model is complete, it is used to inspect components, generating results. Defect simulation verification is then performed based on these results to further optimize the model's accuracy and reliability. Finally, the model's hyperparameters are updated based on the inspection data, and the optimized data is fed back to the model using a backpropagation mechanism, resulting in an optimized model for component inspection that enables more accurate defect detection and prediction in practical applications.

[0125] Preferably, step S21 includes the following steps:

[0126] Step S211: performing texture defect recognition on the component neural topology and forming a component curvature flow differential operator; extracting spatial abnormality features based on the component curvature flow differential operator and constructing a component surface defect layer;

[0127] Step S212: Determine the virtual sound pressure distribution of the component neural topology at a second harmonic frequency of 10 MHz, perform acoustic debonding analysis, and obtain a virtual acoustic signal of the component with a temperature rise of ΔT>1.2K; construct an internal defect layer of the component based on the virtual acoustic signal of the component;

[0128] Step S213: naturally transform the component surface defect layer and the component internal defect layer to obtain a component defect detection layer.

[0129] In embodiments of the present invention, texture defect recognition is performed on a component's neural topology to extract detailed features related to surface defects. Texture defect recognition typically uses image processing algorithms, such as convolutional neural networks (CNNs), to analyze the component's surface texture, thereby identifying potential surface defects such as cracks and potholes. Simultaneously, a curvature flow differential operator is constructed based on the identified texture features. Changes in curvature flow are typically analyzed through mathematical methods to further reveal the spatial distribution of surface defects. Based on the spatial anomaly features extracted by the curvature flow differential operator, a component surface defect layer is formed, which contains the distribution and characteristic information of various surface defects. By determining a virtual sound pressure distribution at a second harmonic frequency of 10 MHz for the component's neural topology and combining it with acoustic debonding analysis (analyzing the detachment process of surface adhesives), the component's performance in an acoustic environment is further explored. Under temperature rise conditions (ΔT > 1.2 K), the component's thermal response and acoustic signals can reveal potential internal defects such as pores and cracks. Based on these virtual acoustic signals, a component internal defect layer is constructed, which contains key information about the component's internal defects acquired through the acoustic signals. In step S213, the component surface defect layer and the component internal defect layer are finally naturally converted (through mathematical mapping or data fusion method) to obtain the component defect detection layer.

[0130] Preferably, step S22 includes the following steps:

[0131] Step S221: Extract the characteristic equivalent charge density of the component defect detection layer, with a range of 1e -5 ~e -3 , and perform charge mapping to obtain the virtual charge density of the components;

[0132] Step S222: Optimize the phonon spectrum heat flow path of the component defect detection layer, where the thermal conductivity correction range is ±15% to 35%, and analyze the material lattice propagation direction, where the lattice constant accuracy is Obtain component material lattice data;

[0133] Step S223: Control the material viscosity dissipation of the component defect detection layer, where the viscosity coefficient range is 1e -3 ~e 3 Pa·s, to obtain the viscosity analysis data of component materials;

[0134] Step S224: fusion of component virtual charge density, component material lattice data, and component material viscosity analysis data through implicit physical verification to obtain component field effect fusion data.

[0135] In an embodiment of the present invention, the characteristic equivalent charge density of the component defect detection layer is extracted. This process determines the charge density range of the component by analyzing the charge distribution on the surface and inside of the component. Usually, a mathematical model of electric field and charge distribution is used to calculate the equivalent charge density, and it is mapped to space to obtain the virtual charge density of the component. This process can reveal the behavioral characteristics of the component under the action of the electric field and provide a basis for the electrical analysis of defects. The phonon spectrum heat flow path optimization of the component defect detection layer is performed to analyze the characteristics of heat conduction inside the component. The heat flow path optimization aims to improve the conduction efficiency of the heat flow by adjusting the thermal conductivity of the material (correction range is ±15% to 35%), and analyze it in combination with the propagation direction of the material lattice. The accuracy of the lattice constant is controlled within Within, the arrangement and propagation characteristics of the lattice are precisely determined in order to accurately evaluate the thermal conduction behavior of the material, thereby obtaining the lattice data of the component material. The key to this step is to provide an in-depth understanding of the thermal performance of the component through the detailed modeling of the lattice structure and the optimization of thermal conductivity. The material viscosity dissipation of the component defect detection layer is controlled. Viscous dissipation refers to the energy lost due to internal friction when the material is subjected to external force. It is usually quantitatively analyzed by studying the viscosity coefficient of the material, which is set in the range of 1~Pa·s. This process helps to understand the energy dissipation characteristics of the material in the defect area and provides support for the viscosity analysis data of the component material. The component virtual charge density, component material lattice data and component material viscosity analysis data are fused through implicit physical verification, combined with the physical field and material mechanics model, to fuse the data from the above different sources to ensure that the obtained data takes into account the accuracy and consistency of electrical, thermal and mechanical properties, thereby obtaining component field effect fusion data.

[0136] Preferably, the defect comparison between the component measured data and the component inspection optimization result in step S3 includes the following:

[0137] Compare the defects of the actual measured data of parts and components with the optimized results of component inspection to obtain comparative defect data;

[0138] When the missed detection rate of the comparative defect data is greater than 0.15%, it is recorded as comparative defect marking data;

[0139] The inference speed is determined based on the comparison defect key mark data. If the inference speed is less than the 50ms production line beat, it is recorded as comparison defect production line data.

[0140] The comparison defect key marking data and the comparison defect production line data are weighted and fused to obtain the component defect comparison data.

[0141] In an embodiment of the present invention, a defect comparison is performed between the measured component data and the optimized component inspection results. A differential analysis method is used to compare the two and generate comparative defect data. This data comparison process typically uses quantitative metrics, such as the missed detection rate, and standard error assessment methods (such as the false negative rate) to assess the completeness of defect detection. If the missed detection rate exceeds 0.15%, the data is further recorded as comparative defect marking data, indicating that defect areas were not detected in a timely manner during actual inspection. Generating this marking data involves calculating an error matrix and identifying defect points with significant differences in a multidimensional data space, thereby ensuring that deficiencies in the inspection system are identified and tracked. Next, by determining the inference speed based on the comparative defect key marking data and combining it with the real-time inference capability of the component inspection system, it is determined whether the inference speed meets certain requirements. In particular, when the inference speed is less than 50ms, it indicates that the data can adapt to the production line cycle and provide real-time feedback. At this point, the data is recorded as comparative defect production line data, ensuring that the inspection system maintains efficient and accurate defect detection capabilities during actual production line operation. Finally, by comparing the defect key marking data and the defect production line data, weighted fusion is performed, and weighted calculation is performed based on the importance and correlation of the two to obtain the component defect comparison data.

[0142] Preferably, the defect detection and evaluation based on the component defect comparison data in step S4 includes:

[0143] Generate material fatigue curve based on bearing pitting anchor confirmation data;

[0144] Perform crack detection based on the material fatigue curve and calculate the material defect sensitivity coefficient to obtain the component defect coefficient;

[0145] The defect coefficient of the component is used to deduce the root cause of the defect using the preset defect coefficient judgment threshold. When the defect coefficient of the component is greater than or equal to the defect coefficient judgment threshold, it is determined to be a process root cause, that is, the component is tested for casting pressure to generate casting pressure test data; the component is tested for pore distribution to obtain pore distribution test data;

[0146] When the defect coefficient of a component is less than the defect coefficient judgment threshold, it is judged to be the root cause of material performance, that is, the hardness of the component is evaluated and calculated to obtain the component hardness performance test data; the abnormal crack extension path of the component is detected to obtain the abnormal crack detection data of the component.

[0147] In an embodiment of the present invention, based on the bearing pitting anchor confirmation data, it usually involves a quantitative analysis of the surface corrosion characteristics of the material in the actual working environment. This data can be generated by a statistical model based on the corrosion rate and the bearing usage time. The process of generating the material SN fatigue curve is to fit the fatigue life of the material at different stress levels through experimental data, and to determine the fatigue performance characteristics of the material through statistical analysis methods (such as regression analysis). It is usually necessary to consider the stress-life relationship of the material under different working conditions. Then, based on the generated material SN fatigue curve, crack detection is performed. This step is mainly achieved by modeling the relationship between fatigue life and crack propagation rate. Common crack propagation theories, such as the Paris formula, are used to evaluate the health status of the material by simulating the crack propagation process inside the material. The subsequent calculation of the material defect sensitivity coefficient mainly combines the crack propagation and the local stress state of the material structure, and based on the theory of stress concentration points, calculates the defect sensitivity coefficient of each area, thereby obtaining the overall defect coefficient of the component. Based on this defect coefficient, the root cause of the component defect is deduced using a preset threshold. When the defect coefficient of a component is greater than or equal to the defect coefficient judgment threshold, the system will determine that the defect is a process root cause. At this time, the casting pressure detection and pore distribution detection will be activated. By performing pressure analysis on the casting process of the casting, it is evaluated whether the casting has insufficient or uneven pressure during the manufacturing process. At the same time, the pore distribution is checked to confirm whether the pores of the component are too large or unevenly distributed. Conversely, when the defect coefficient of the component is less than the defect coefficient judgment threshold, it is judged to be a material performance root cause. The component will then be subjected to hardness assessment and abnormal crack propagation path detection. Hardness assessment is mainly carried out by testing the hardness of the material, using different hardness test methods such as Brinell hardness and Rockwell hardness to derive the fatigue resistance of the material; while abnormal crack propagation path detection uses advanced crack propagation analysis technology to monitor the crack propagation trajectory and combine finite element analysis technology to identify the potential crack propagation path to ensure the safety of the structure.

[0148] The present invention further provides an image model-based component defect detection system for executing the above-mentioned image model-based component defect detection method. The image model-based component defect detection system includes:

[0149] The data preprocessing and topology modeling module is used to obtain multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct the neural topology structure of components based on the multi-source preprocessed data of components;

[0150] The defect detection modeling module is used to construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and build a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return it to the model to obtain a component detection optimization model;

[0151] The detection and optimization module is used to obtain the actual measured data of the parts; use the actual measured data of the parts as input to output the parts detection optimization results based on the parts detection optimization model; compare the defects of the parts measured data and the parts detection optimization results to obtain parts defect comparison data;

[0152] The result verification and report generation module is used to perform bearing pitting anchor confirmation on the component defect comparison data to obtain bearing pitting anchor confirmation data; perform defect detection evaluation based on the bearing pitting anchor confirmation data, and generate an image model component defect detection report.

[0153] The beneficial effect of the present invention is that by acquiring multi-source images of parts and performing image preprocessing, multi-source standard images are generated, ensuring the consistency of image quality and the integrity of feature information at the data level. Subsequently, a neural topology structure of the parts is constructed based on the preprocessed data, and by mapping the multi-dimensional feature relationships of the parts, the geometric shape, texture features and potential defect information are effectively integrated to achieve efficient expression of cross-modal features. The defect detection layer is constructed in the feature extraction layer, enabling the system to deeply mine defect features at different scales and complex backgrounds, and improve the robustness and generalization ability of detection through non-parametric feature fusion. The establishment of the parts detection model further relies on automated feature selection and multi-dimensional feature aggregation, effectively avoiding errors caused by manual intervention. After the detection results are output, the model is adaptively optimized through a hyperparameter update mechanism to form an iteratively optimized parts detection optimization model, ensuring continuous improvement in detection accuracy. In the actual detection process, the measured data of the parts are input into the optimized detection model to obtain accurate detection results, and the accuracy and stability of the detection are evaluated through defect comparison analysis. Furthermore, for bearing pitting, a typical defect type, the system utilizes a pitting anchoring confirmation method to precisely locate local defects. Combined with inspection and evaluation, the system further quantifies the degree of defects, providing a reliable basis for subsequent maintenance and quality control. Finally, component defect inspection reports are generated using image models, enabling intuitive presentation of inspection results and supporting decision-making. Therefore, through multi-source data fusion, neural topology modeling, and adaptive optimization, this invention addresses the problems of feature loss, insufficient detection accuracy, and inaccurate defect location in traditional inspection methods, thereby improving the reliability, accuracy, and automation level of component defect detection.

[0154] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0155] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A component defect detection method based on an image model, characterized in that: The following steps are involved: Step S1: Acquire multi-source images of parts; perform image preprocessing on the multi-source images of parts to obtain multi-source standard images of parts; Constructing component neural topology based on multi-source pre-processed data of components; Step S2: Construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and construct a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return them to the model to obtain a component detection optimization model; Step S3: Acquire component measured data; use the component measured data as input to output component detection optimization results based on the component detection optimization model; perform defect comparison between the component measured data and the component detection optimization results to obtain component defect comparison data; Step S4: performing bearing pitting anchoring confirmation on the component defect comparison data to obtain bearing pitting anchoring confirmation data; performing defect detection evaluation based on the bearing pitting anchoring confirmation data, and generating an image model component defect detection report.

2. The component defect detection method based on image model according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: setting the resolution of the 16K linear array industrial camera to 5 μm / pixel and the wavelength range to 450-650 nm; Step S12: setting the Z-axis repeatability of the line laser 3D scanner to ±1 μm and the scanning width to 200 mm; Step S13: Setting the thermal sensitivity of the uncooled infrared thermal imager to less than 30 mK; Step S14: using a 16K linear array industrial camera, a line laser 3D scanner, and an uncooled infrared thermal imager to collect multi-source data of components; Step S15: performing data preprocessing on the component multi-source image data to obtain component multi-source image preprocessing data; and constructing a component neural topology structure based on the component multi-source preprocessing data.

3. The component defect detection method based on image model according to claim 1, characterized in that: The data preprocessing of the multi-source image data of the component in step S1 includes: Use 16K linear array industrial cameras, line laser 3D scanners, and uncooled infrared thermal imagers to collect multi-source data on parts, including visible light data images, 3D scanning data images, and thermodynamic data of parts; Extracting component illumination texture from the component visible light data image to obtain component illumination texture data; Perform point cloud geometry analysis on the 3D scanning data image of the parts to obtain the point cloud geometry data of the parts; The thermodynamic data of components are divided into regional gradients and hot spot distribution is arranged to obtain component hot spot arrangement data.

4. The component defect detection method based on image model according to claim 1, characterized in that: The construction of the component neural topology structure described in step S1 includes: Image feature extraction is performed on the component illumination texture data, that is, the directional gradient of oxidation spots is extracted from the component illumination texture data, where the directional gradient of oxidation spots ranges from 0.1 to 5.0 pixels / μm, and the first layer structure at the bottom of the surface is constructed; local convex hull curvature analysis is performed on the component illumination texture data, where the main curvature range is ±0.01 to 0.5 mm -1 , construct the surface primary mid-end second layer structure; extract the relative position assembly error between components based on the component illumination texture data, where the error range is ±50~500μm, and construct the macro high-level third layer structure; The luminous flux neural topology map is constructed by constructing the first layer structure at the bottom of the surface, the second layer structure at the primary middle end of the surface, and the third layer structure at the macroscopic high level, where the luminous flux density ranges from 0.1 to 10 lx·μ -2 , obtain the component luminous flux texture structure; Construct a component spectral clustering graph structure based on the 3D scanning data of the components; Create three-dimensional heat conduction residual data of components based on the thermodynamic data of components; Project the component's luminous flux texture structure onto the component's spectral clustering graph structure, mark the surface abnormal area, and obtain the component's abnormal projection area; The abnormal response of the abnormal projection area of ​​the component is reversely modulated using the three-dimensional heat conduction residual data of the component, and multi-dimensional neural topology constraints are performed to obtain the neural topology structure of the component.

5. The component defect detection method based on image model according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: constructing a feature extraction layer based on the component neural topology structure to obtain a component defect detection layer; Step S22: performing non-parametric feature fusion based on the component defect detection layer to obtain component field effect fusion data, wherein the non-parametric feature fusion includes equivalent charge density extraction, phonon spectrum heat flow path optimization, and material viscosity dissipation control; Step S23: constructing a detection model for the component field effect fusion data to obtain a component detection model; Step S24: Output the component detection results according to the component detection model; perform defect simulation verification based on the component detection result output to obtain component detection data; update hyperparameters based on the component detection data, and return the model to obtain a component detection optimization model.

6. The component defect detection method based on image model according to claim 1, characterized in that: Step S21 includes the following steps: Step S211: performing texture defect recognition on the component neural topology and forming a component curvature flow differential operator; extracting spatial abnormality features based on the component curvature flow differential operator and constructing a component surface defect layer; Step S212: Determine the virtual sound pressure distribution of the component neural topology at a second harmonic frequency of 10 MHz, perform acoustic debonding analysis, and obtain a virtual acoustic signal of the component with a temperature rise of ΔT>1.2K; construct an internal defect layer of the component based on the virtual acoustic signal of the component; Step S213: naturally transform the component surface defect layer and the component internal defect layer to obtain a component defect detection layer.

7. The component defect detection method based on image model according to claim 1, characterized in that: Step S22 includes the following steps: Step S221: Extract the characteristic equivalent charge density of the component defect detection layer, with a range of 1e -5 ~e -3 , and perform charge mapping to obtain the virtual charge density of the components; Step S222: Optimize the phonon spectrum heat flow path of the component defect detection layer, where the thermal conductivity correction range is ±15% to 35%, and analyze the material lattice propagation direction, where the lattice constant accuracy is Obtain component material lattice data; Step S223: Control the material viscosity dissipation of the component defect detection layer, where the viscosity coefficient range is 1e -3 ~e 3 Pa·s, to obtain the viscosity analysis data of component materials; Step S224: The component virtual charge density, component material lattice data, and component material viscosity analysis data are fused through implicit physical verification to obtain component field effect fusion data.

8. The component defect detection method based on image model according to claim 1, characterized in that: The defect comparison between the component measured data and the component inspection optimization results in step S3 includes: Compare the defects of the actual measured data of parts and components with the optimized results of component inspection to obtain comparative defect data; When the missed detection rate of the comparative defect data is greater than 0.15%, it is recorded as comparative defect marking data; The inference speed is determined based on the comparison defect key mark data. If the inference speed is less than the 50ms production line beat, it is recorded as comparison defect production line data. The comparison defect key marking data and the comparison defect production line data are weighted and fused to obtain the component defect comparison data.

9. The component defect detection method based on image model according to claim 1, characterized in that: The defect detection and evaluation based on the component defect comparison data described in step S4 includes: Generate material fatigue curve based on bearing pitting anchor confirmation data; Perform crack detection based on the material fatigue curve and calculate the material defect sensitivity coefficient to obtain the component defect coefficient; The defect coefficient of the component is used to deduce the root cause of the defect using the preset defect coefficient judgment threshold. When the defect coefficient of the component is greater than or equal to the defect coefficient judgment threshold, it is determined to be a process root cause, and the component is tested for casting pressure to generate casting pressure test data; the component is tested for pore distribution to obtain pore distribution test data; When the defect coefficient of a component is less than the defect coefficient judgment threshold, it is judged to be the root cause of material performance, and the hardness of the component is evaluated and calculated to obtain the component hardness performance test data; the abnormal crack extension path of the component is detected to obtain the abnormal crack detection data of the component.

10. A component defect detection system based on an image model, characterized in that: For executing the component defect detection method based on an image model according to claim 1, the component defect detection system based on an image model comprises: The data preprocessing and topology modeling module is used to obtain multi-source images of components; perform image preprocessing on the multi-source images of components to obtain multi-source standard images of components; and construct the neural topology structure of components based on the multi-source preprocessed data of components; The defect detection modeling module is used to construct a feature extraction layer based on the component neural topology to obtain a component defect detection layer; perform non-parametric feature fusion based on the component defect detection layer and build a component detection model; output component detection results based on the component detection model; update hyperparameters based on the component detection result data and return it to the model to obtain a component detection optimization model; The detection and optimization module is used to obtain the actual measured data of the parts; use the actual measured data of the parts as input to output the parts detection optimization results based on the parts detection optimization model; compare the defects of the parts measured data and the parts detection optimization results to obtain parts defect comparison data; The result verification and report generation module is used to perform bearing pitting anchor confirmation on the component defect comparison data to obtain bearing pitting anchor confirmation data; perform defect detection evaluation based on the bearing pitting anchor confirmation data, and generate an image model component defect detection report.

Images