Graphite equipment weld defect intelligent detection and positioning system based on multispectral imaging

Abstract

The application discloses a graphite equipment weld defect intelligent detection and positioning system based on multispectral imaging, and particularly relates to the field of welding, and comprises a multi-band image acquisition unit, a spectral parameter analysis unit, a spectral anomaly identification unit, a weld defect positioning unit, a defect identification and grading unit and an element composition analysis unit.The multi-band image acquisition unit collects spectral image data of multiple bands and transmits the spectral image data to a multispectral data storage library.The spectral parameter analysis unit extracts characteristic parameters of a weld area, and establishes a spectral characteristic parameter library.The spectral anomaly identification unit locates suspected defect areas of spectral anomalies.The weld defect positioning unit marks defect boundaries and ranges.The defect identification and grading unit calculates a defect severity index.The element composition analysis unit detects suspected defect position samples, and analyzes types and contents of trace elements in the suspected defect position samples.The weld quality comprehensive evaluation unit integrates defect and composition data, generates a quality grade and rectification suggestions, solves the problem that existing technologies cannot locate defect causes, provides data support for subsequent maintenance and process optimization, and reduces equipment operation and maintenance safety hazards.

Description

Technical Field

[0001] This invention relates to the field of welding, and more specifically, to an intelligent detection and location system for weld defects in graphite equipment based on multispectral imaging. Background Technology

[0002] Currently, weld inspection in graphite equipment primarily relies on conventional visual inspection, ultrasonic inspection, infrared thermography, or single-band optical imaging technologies. Existing technologies typically employ manual visual inspection as an initial screening method, where personnel visually observe the weld surface or acquire surface images using magnifying glasses, industrial cameras, or other equipment to identify macroscopic defects such as weld beads and undercut. For internal defect detection, ultrasonic flaw detectors are commonly used, employing the transmission and reception of ultrasonic signals to determine the presence of abnormal areas such as porosity and lack of fusion within the weld. Some equipment utilizes infrared thermography to assist in inspection, locating potential defects by analyzing temperature field distribution differences during heating or cooling processes. The typical workflow involves: manual or robotic scanning of the weld, acquisition of single-band images or ultrasonic signals, extraction of surface defects using traditional filtering and edge detection algorithms, determination of whether manual re-inspection is necessary based on empirical thresholds, secondary ultrasonic inspection of suspected areas, and manual confirmation of defect types. This workflow relies heavily on limited optical information and human judgment to drive the entire inspection loop, representing the currently prevalent inspection path in the industry.

[0003] However, existing technologies still have significant limitations: First, single-band imaging struggles to identify subtle spectral changes in graphite materials caused by differences in crystal structure, failing to effectively distinguish hidden defects such as incomplete fusion, inclusions, and microcracks. Second, traditional image processing methods are sensitive to illumination, surface roughness, and environmental noise, leading to high false detection rates and poor repeatability. Third, ultrasonic testing relies on operator experience, often encountering severe echo attenuation and difficulty in accurately locating defects in complex structures or high-density graphite weld areas. Furthermore, existing technologies cannot correlate defect causes or identify potential trace impurities in welds through spectral or material analysis, making it difficult to quantify the severity of defects. Existing systems lack effective data storage, case reuse, and intelligent analysis mechanisms, hindering the formation of learnable detection models. Therefore, existing technologies cannot meet the stringent requirements for weld quality under the high-temperature, high-corrosion, and high-load conditions of graphite equipment, necessitating a novel detection system capable of multi-dimensional data fusion, intelligent identification, and defect cause analysis. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide an intelligent detection and location system for weld defects in graphite equipment based on multispectral imaging. The present invention provides the following technical solution:

[0005] A smart detection and location system for weld defects in graphite equipment based on multispectral imaging includes:

[0006] The multi-band image acquisition unit is used to perform omnidirectional scanning of the weld area of ​​the graphite equipment and transmit the acquired spectral image data of multiple bands to the spectral database.

[0007] The spectral parameter analysis unit is used to extract characteristic parameters such as spectral angle, main absorption peak position, local standard deviation, and Mahalanobis distance of the weld area from the spectral database, and to establish a spectral feature library of normal welds and abnormal areas.

[0008] The spectral anomaly identification unit is used to locate suspected defect areas with spectral anomalies by comparing the spectral parameters of the weld to be inspected with those of normal welds in the spectral feature library.

[0009] The weld defect location unit is used to locate suspected defects and mark the defect boundaries and extent by using a spectral anomaly mask combined with an image segmentation algorithm.

[0010] The defect identification and classification unit is used to classify the located defects and calculate the defect severity index.

[0011] The elemental composition analysis unit is used to perform trace element detection on samples from suspected defect locations when the defect severity index is greater than a preset threshold, and to analyze the types and contents of trace elements.

[0012] The impurity and composition assessment unit is used to compare the detected trace element data with the graphite equipment material standards to determine whether there are excessive impurities and their potential impact on weld performance.

[0013] The comprehensive weld quality assessment unit is used to combine defect identification results with trace element detection data to generate a quality grade and inspection report for the weld quality.

[0014] The inspection data management module includes a multispectral data repository, a weld defect case library, and a spectral characteristic parameter library.

[0015] The technical effects and advantages of this invention are as follows:

[0016] This invention, by introducing multi-band imaging and spectral parameter analysis technology, can comprehensively utilize the spectral characteristics of the weld area at different wavelengths to achieve simultaneous detection of surface defects and internal hidden defects. The multi-spectral information is modeled using unified features in the system of this invention, enabling defects such as lack of fusion, microcracks, and inclusions, which are difficult to distinguish using traditional single-band methods, to exhibit identifiable differences in the spectral dimension, thereby effectively improving the accuracy and comprehensiveness of defect identification.

[0017] This invention achieves automatic adaptation to imaging differences under different weld seams and environmental conditions by constructing a spectral feature library and a spectral anomaly recognition mechanism. The system employs multi-source feature collaborative judgment, effectively suppressing the effects of illumination changes, surface roughness, and environmental noise, resulting in more stable detection results. Simultaneously, this invention constructs a complete intelligent processing flow through defect segmentation, automatic location, and classification judgment, reducing reliance on human experience and improving overall detection efficiency and judgment consistency.

[0018] This invention combines spectral detection with trace element analysis. By analyzing the composition of suspected areas and comparing it with material standards, it enables source analysis of defect causes and, based on this, a comprehensive assessment of weld quality. The system's multi-level evaluation process generates structured quality grades and rectification suggestions, ensuring that the test results go beyond defect identification and further support equipment maintenance decisions and process optimization, thereby improving the scientific rigor and traceability of weld quality management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0020] Figure 2 This is a schematic diagram of the spectral feature extraction and anomaly identification process of the present invention.

[0021] Figure 3 This is a table of 15 sets of experimental data for the defect severity index of this invention.

[0022] Figure 4 This is a graph showing the relationship between the defect severity index and defect size in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] refer to Figures 1-4 The intelligent detection and location system for weld defects in graphite equipment based on multispectral imaging, as shown, includes the following steps in its specific implementation:

[0025] The multi-band image acquisition unit is used to perform omnidirectional scanning of the weld area of ​​the graphite equipment and transmit the acquired spectral image data of multiple bands to the multispectral data storage repository.

[0026] As a preferred feasible embodiment, it should be explained that the weld seam area of ​​graphite equipment is mostly irregular curves, annular or spliced ​​seams, and needs to cover the weld body, heat-affected zone and base material transition zone. The specific scanning method is as follows: The scanning method is selected according to the structural type of the graphite equipment weld seam: For annular weld seams of pipeline equipment, a spiral scanning method is adopted, controlling the multispectral imaging lens group to spirally advance along the circumference of the weld seam to ensure that each frame of image covers 1-2mm of weld seam width and avoids missed scans; For straight weld seams of flat plates, a serpentine reciprocating scanning method is adopted, and multi-angle supplementary scanning is assisted. That is, the lens group is driven by a robotic arm to adjust the pitch angle within ±15° to supplement the scan of easily missed areas such as the weld root and fusion line, so as to achieve all-round coverage; In terms of scanning control logic, laser positioning and visual guidance need to be combined. First, the real-time position coordinates of the weld seam are obtained through the laser contour sensor and fed back to the motion controller to dynamically correct the scanning trajectory. At the same time, the imaging resolution and step size are fixed and matched. If the spatial resolution is 50μm / pixel, the scanning step size needs to be set to 30-40μm.

[0027] Furthermore, in order to suppress the interference of surface roughness differences and local specular reflections of graphite welds on multi-band imaging, this embodiment sets a switchable polarizer assembly at the front end of the lens group and a coaxial polarizer at the end of the ring active light source, so that the illumination polarization direction and the analysis polarization direction at the imaging end are orthogonal to each other (cross polarization), thereby significantly reducing the high light saturation area and reflection artifacts.

[0028] Meanwhile, before each scan begins, the system automatically acquires the dark field image Idark and the standard white board image Iwhite. It performs radiometric calibration and reflectivity normalization on the original image Iraw for each band to obtain the normalized reflectivity image R=(Iraw-Idark) / (Iwhite-Idark). The system also binds and archives the correction parameters such as Idark, Iwhite, and polarization angle with the corresponding weld number and scan time.

[0029] The above processing makes the calculation of subsequent spectral angle, main absorption peak position, local standard deviation and Mahalanobis distance more robust to changes in illumination and surface reflection, and reduces the false detection rate.

[0030] The acquired spectral image data is a fusion of spatial and spectral information. Spatial geometric information includes the two-dimensional / three-dimensional pixel coordinates of the weld area, pixel-level measurements of weld width / height / excess height, and the spatial location and morphology of defects, which can be used to locate the weld and preliminarily determine whether the geometric dimensions comply with regulations. Spectral feature information covers the light intensity value of each pixel in different wavelength bands, spectral reflectance / absorption curves, characteristic peak wavelengths and intensities, enabling the identification of weld material uniformity and hidden defects. The selection of multiple wavelength bands needs to match the characteristics of the graphite material with the weld defect detection target, such as cracks, lack of fusion, and impurity inclusions. In this embodiment, the working spectral range of the multi-band image acquisition unit is 400-3000nm, primarily covering the visible light-near-infrared-short-wave infrared range, with the visible light band ranging from 400-760nm. Graphite has low reflectivity, but the reflectivity of the oxide layer on the weld surface varies significantly, making it suitable for detecting macroscopic defects such as surface cracks, weld beads, and undercut. The near-infrared band is 760-2500nm. The absorption characteristics of graphite for near-infrared light are related to its internal structure. In unfused areas, the absorption peak shifts due to insufficient bonding of graphite grains, and pores can cause sudden changes in light intensity, making it suitable for detecting internal pores and incomplete fusion. The short-wave infrared band is 2500-3000nm, which can penetrate the oxide layer on the weld surface, capturing microcracks and material inhomogeneities, and is less affected by ambient light. The actual number of bands needs to be adjusted according to the required detection accuracy. In this embodiment, multiple fixed band imaging channels are set within this range, with the preferred center wavelengths being 450nm, 800nm, 1500nm, 2000nm, and 2800nm, balancing detection efficiency and defect recognition rate.

[0031] This embodiment further provides an adaptive band enhancement mechanism: based on the initial spectral morphology of the local area of ​​the weld, the system calculates the signal-to-noise ratio and spectral gradient value of each band in real time, and dynamically adjusts the weight of different bands using the following formula:

[0032]

[0033] in: For the first Band signal-to-noise ratio; This represents the local gradient of the spectral curve in this band, used to reflect sensitivity to defects; This is the normalization coefficient.

[0034] The spectral parameter analysis unit is used to extract characteristic parameters of the weld area from the multispectral data repository and establish a spectral characteristic parameter library for normal welds and abnormal areas.

[0035] As a preferred feasible embodiment, it should be explained that the characteristic parameters refer to the spectral angle, the position of the main absorption peak, the local standard deviation, and the Mahalanobis distance.

[0036] Further explanation is needed regarding the spectral angle, which is the spectral similarity. It is calculated by comparing the angle between the spectral curve of the pixel under test and the reference spectral curve of the normal weld. The smaller the angle, the higher the spectral similarity between the pixel under test and the normal weld. The reference spectral curve of the normal weld is based on the average curve of samples of normal welds in the multispectral data repository. The position of the main absorption peak is the wavelength corresponding to the point with the lowest light intensity value found in the key band interval. The local standard deviation is the standard deviation of the light intensity values ​​of all bands calculated within an N×N small window centered on the target pixel. The value of N ranges from 3 to 5 and is dynamically adjusted according to the spatial resolution of multispectral imaging.

[0037] As attached Figure 2 As shown, the calculation process for Mahalanobis distance is as follows:

[0038] Step 1: Select a large number of known defect-free weld area multispectral image data from the multispectral data repository;

[0039] Step 2: Select three core bands for calculating the Mahalanobis distance. Each pixel is represented by a three-dimensional vector, such as [ ];

[0040] Step 3: Average the 3D vector values ​​of all normal pixels to obtain the mean vector u, such as [ ];

[0041] Step 4: Calculate the covariance between each pair of light intensity values ​​in the three bands, forming a 3×3 covariance matrix Σ, and calculate the inverse of this covariance matrix. ;

[0042] Step 5: Calculate the difference between the current pixel vector x and the normal mean vector to obtain the difference vector xu, such as [ ] and transpose the difference to obtain ;

[0043] Step 6: Perform matrix multiplication, then calculate the scalar b. The square root of b gives the Mahalanobis distance MD.

[0044] It should be noted that the spectral characteristic parameter library is a data set that stores the spectral characteristic parameters of the weld area. It includes a normal library and an abnormal library. The basic data framework of the two is consistent. Both are associated with weld identification and store four types of parameters: spectral angle, main absorption peak position, local standard deviation, and Mahalanobis distance, as well as calculation conditions such as band and window size. The data in the normal library comes from defect-free weld samples confirmed by manual re-inspection, while the abnormal library is classified and stored according to defect type, and defect-related attributes are labeled to provide a basis for subsequent comparative analysis.

[0045] It should be further explained that the parameters in the normal database conform to the normal weld seam pattern, with a small spectral angle, a stable main absorption peak position at the typical wavelength, a low local standard deviation, and a concentrated Mahalanobis distance; while the parameters in the abnormal database deviate from the normal weld seam pattern, with a large spectral angle, a shifted main absorption peak position, a high local standard deviation, and a discrete Mahalanobis distance that deviates from the normal statistics.

[0046] The spectral anomaly identification unit is used to locate suspected defect areas with spectral anomalies by comparing the spectral parameters of the weld to be inspected with those of normal welds in the feature library.

[0047] As a preferred feasible embodiment, it needs to be explained that the specific process of locating suspected defect areas with spectral anomalies is as follows: First, the Mahalanobis distance of each pixel in the image to be detected is calculated, and a threshold is set. If MD > MD threshold, then the pixel will be immediately marked as a strongly suspected anomaly area; if MD < MD threshold, then the pixel will be marked as a provisionally normal area. Second, the spectral anomaly identification unit will check three other features in parallel and compare them with their respective empirical thresholds: if the spectral angle > spectral angle threshold, it means that the material of the pixel does not match the standard normal material, which may be any type of defect or contamination; if the main absorption peak position shift is greater than the peak position shift threshold, it means that the lattice structure of graphite has changed, which may be internal defects such as lack of fusion or material doping; if the local standard deviation is greater than the local standard deviation threshold, it means that the texture around the pixel is uneven and the light intensity changes abruptly, which may be structural defects such as pores, cracks or inclusions.

[0048] To further enhance feature discrimination capabilities, this invention constructs an enhanced feature vector V from four types of spectral features using a cooperative matrix:

[0049]

[0050] An additional parameter R (feature coupling degree) is introduced, and its calculation method is as follows:

[0051]

[0052] This parameter can characterize the degree of coupling between spectral perturbations caused by defects. For example, cracks usually cause both the spectral angle and the local standard deviation to increase simultaneously. This degree of coupling does not exist in normal welds.

[0053] Further explanation is needed: the spectral anomaly identification unit determines the final state of a pixel based on the analysis results of all features. A pixel is considered a suspected defect if its MD exceeds the MD threshold and any of the remaining feature parameters exceeds the threshold. If the MD does not exceed the threshold, but the main absorption peak shift and local standard deviation both exceed the threshold, it is confirmed as a defect. A pixel is considered suspicious if its MD does not exceed the threshold, but only one other feature parameter exceeds the threshold. It is then marked as suspicious and requires manual review. The results of the manual review must be simultaneously entered into the detection data management module. If the suspicious pixel is confirmed to be normal, its spectral parameters are added to the normal spectral feature parameter library. A pixel is considered normal if all features are within the normal threshold range.

[0054] It needs to be further explained that after determining the final state of the pixel, the spectral anomaly recognition unit generates a binary mask image of the same size as the original image. White pixels represent suspected defect pixels, while black pixels represent the background and normal areas.

[0055] This embodiment introduces a dynamic threshold update method based on distribution shift detection: the system continuously performs KS (Kolmogorov-Smirnov) test on the feature distribution of newly added samples in the normal library. When the distribution shifts significantly, the system automatically updates: the MD threshold is set to the mean of the normal sample distribution + 2σ; the spectral angle threshold is set to the p95 value (95th percentile); and the peak shift threshold is calculated based on the local range.

[0056] The weld defect location unit is used to combine the suspected defect area and use image segmentation algorithms to accurately locate the suspected defect in the weld, and mark the defect boundary and range.

[0057] It needs to be explained that after receiving the binarized mask image transmitted by the spectral anomaly recognition unit, the weld defect location unit uses a median filtering algorithm to remove scattered and isolated white pixels. Next, it scans the entire image, identifies and marks all interconnected white pixel groups, and assigns a unique label to each independent group, resulting in a label image where background pixels are 0, all pixels in the first defect area are 1, the second is 2, and so on. After identifying the defects, the weld defect location unit calculates their location and shape: first, it uses a boundary tracking algorithm to trace along the outer edge of each connected region, recording the coordinates of each point forming the boundary, thus obtaining a list of contours. Each contour consists of a series of points, describing a polygon. Next, it calculates a series of features for each contour, generating a structured defect list, and finally outputs a visual defect location map: the contour and circumscribed rectangle of each defect are drawn on the original image, and a serial number is labeled.

[0058] This embodiment further introduces a spectral texture tensor T after the binarization mask is generated:

[0059]

[0060] The system performs region growth based on the gradient magnitude of T, enabling the same defect region to maintain contour consistency across different bands.

[0061] Furthermore, to ensure consistent localization results of defect boundaries across different spectral bands, this embodiment introduces cross-band contour consistency constraints based on region growing: First, the binarized mask output by the spectral anomaly recognition unit is used as the initial region Ω0; then, the gradient field of the spectral texture tensor T is calculated on the normalized reflectance images of each band, and a boundary optimization objective function J=α·Edata+β·Eedge+γ·Esmooth is constructed, where Edata is used to constrain the overlap between the segmentation result and the initial mask, Eedge is used to absorb the boundary where ∥∇T∥ is large, and Esmooth is used to penalize excessively jagged boundaries and isolated small connected regions; finally, iterative boundary refinement (including morphological closing operation, hole filling, and small area connected region removal) is used to obtain the final defect contour.

[0062] By using the above cross-band consistency constraints, boundary drift caused by single-band noise can be avoided, making the geometric quantities such as the outer rectangle, center coordinates and contour perimeter of the defect more stable, thereby improving the accuracy of subsequent defect size assessment and sampling location.

[0063] It should be explained that the defect list includes the smallest rectangle that can completely enclose the defect, the circumscribed rectangle, the area of ​​the defect, the center coordinates of the defect, and the perimeter of the outline.

[0064] The defect identification and classification unit is used to classify the located defects and calculate the defect severity index.

[0065] As a preferred feasible embodiment, it should be explained that the defect identification and classification unit directly inputs each located defect image block, along with its label, into a convolutional neural network for training. The trained model is integrated into the system. For each unknown defect sent by the weld defect location unit, the module inputs its image block into the model, and the model outputs a classification result and confidence probability, such as [Category: porosity, confidence level 92%].

[0066] It needs further explanation that the identified defect categories include porosity, cracks, lack of fusion, and inclusions.

[0067] It should be noted that the defect severity index W represents the defect type, which includes four types of defects: porosity, cracks, lack of fusion, and inclusions. It is calculated by weighting the severity of each type of defect. E represents the defect size score, which is obtained by comparing the actual physical size of the defect with industry acceptance standards and dividing the score into preset ranges. F represents the defect location coefficient, which shows that the severity of the same defect varies greatly in different locations. For example, the weld surface is the most dangerous area because it directly bears the load, while the base material is relatively less dangerous.

[0068] The elemental composition analysis unit is used to connect to the inductively coupled plasma mass spectrometry equipment to detect samples at suspected defect locations and analyze the types and contents of trace elements when the defect severity index is greater than a preset threshold.

[0069] As a preferred and feasible embodiment, it should be explained that the preset threshold is not a single fixed value, but integrates three levels of standards: First, strictly follow industry standards to clarify the threshold for visual judgment of weld defects; second, based on the weld defect case library, statistically determine the correlation threshold between trace element anomalies and defect deterioration; finally, through risk matrix assessment, combined with equipment service scenarios, a dynamic threshold system is formed.

[0070] After the defect severity index exceeds the dynamic threshold, the elemental composition analysis unit initiates automated linkage: First, the module receives the three-dimensional coordinates of the suspected defect location output by the weld defect location unit, controls the robotic arm equipped with a micro-sampling device to move to the target area, and cuts out a cylindrical sample with a diameter of 0.8-1mm and a depth of 0.5mm in the defect area using a fiber laser to avoid introducing external impurities during the sampling process; then, the sampling device sends the sample into the automated sample introduction system of the inductively coupled plasma mass spectrometer through a sealed pipeline, while transmitting relevant information such as defect number, sampling location, and equipment model, triggering the detection method package.

[0071] Furthermore, to ensure the consistency between the sampling location and the defect location, and to prevent the sample from being contaminated by external sources, this embodiment performs coordinate closed-loop verification before sampling: after the robotic arm is in position, the imaging end re-acquires multi-band local images of the defect area, and registers the defect contour in the real-time image with the historical positioning map. If the deviation exceeds the preset tolerance (e.g., 0.2mm), the position of the robotic arm end is automatically corrected before cutting is performed. During the sampling process, the micro-sampling device uses a disposable replaceable sampling head or an automatic cleaning process. Before cutting, the sampling area is purged with inert gas and a blank control sample is set to reduce cross-contamination caused by dust, oil, and environmental metal particles.

[0072] After sampling, the system generates a unique barcode / QR code for each sample and binds "defect number - 3D coordinates - band data index - sampling batch - operation log" into the data link to ensure that subsequent ICP-MS test results can be traced back to the corresponding defect location and original spectral evidence, thus realizing an evidence closed loop of component detection and defect localization.

[0073] This detection method package is optimized for the characteristics of graphite materials: it uses inductively coupled plasma mass spectrometry (ICP-MS) with a scanning mass range of 7-250 μU, and eliminates carbon-based interference through a collision reaction cell to ensure a detection limit as low as 1 ppb. After detection, the equipment outputs a quantitative analysis report of trace element types and contents, such as "Fe: 62 ppm, Ni: 18 ppm, S: 5 ppm". The method package will correlate this report with the defect severity index and spectral characteristic parameters to form a data chain.

[0074] It should be noted that the test results will be used for two core applications: first, defect cause tracing. For example, if a high content of Cu is detected, it can be located to contact contamination between the welding rod and the copper fixture during the welding process, providing a basis for process optimization; second, defect risk classification. Combining trace element diffusion kinetic models, such as the annual diffusion rate of S at 600℃ being 0.3 mm, the expansion trend of defects during the remaining service life of the equipment can be predicted. At the same time, all test data will be sent back to the multispectral data repository to optimize the parameters of the spectral parameter analysis unit, forming a closed-loop iteration of the detection system.

[0075] The impurity and composition assessment unit is used to compare the detected trace element data with the material standards of the graphite equipment to determine whether there are excessive impurities and their potential impact on weld performance.

[0076] As a preferred feasible embodiment, it should be explained that the impurity and composition evaluation unit has a built-in reference library to ensure that the comparison basis is fully adapted to the characteristics of graphite equipment: First, the limit values ​​of key impurities in the graphite matrix are clearly defined, such as Fe≤30ppm, Ni≤10ppm, S≤5ppm, and Cu≤8ppm. These elements will destroy the layered crystal structure of graphite and reduce its high-temperature resistance. Second, in conjunction with the supplementary standards of the application industry of graphite equipment, for welds in contact with corrosive media, Cl⁻≤2ppm is additionally limited to prevent the medium from reacting with Cl⁻ to generate corrosive gases, and Na⁺≤3ppm to prevent Na from causing embrittlement of graphite grain boundaries at high temperatures. Finally, for customized equipment, the special material standards provided by the equipment manufacturer are adopted, such as strictly controlling the limit value of element B to ≤0.1ppm.

[0077] Further explanation is needed regarding the method for identifying abnormal impurities: For a single trace element, if the detected value exceeds 120% of the corresponding reference standard limit, it is determined to be a single-element exceedance. For example, if the Fe detected value is 38 ppm, and the basic standard is 30 ppm, 38 > 30 × 120% = 36, so it is determined to be Fe exceedance. For scenarios where multiple elements coexist, if the detected values ​​of two or more impurities both exceed 80% of the limit, it is determined to be a synergistic exceedance. For example, if the Ni detected value is 8.5 ppm, the limit is 10 ppm, accounting for 85%, and the Cu detected value is 7 ppm, the limit is 8 ppm, accounting for 87.5%, the two will reduce the fatigue resistance of the weld, so it is determined to be a synergistic exceedance. At the same time, the module will automatically mark key exceedance items. If the exceedance element is an element such as S or Cl, which is extremely harmful to graphite welds, even if the exceedance is only 5%, it will be highlighted first.

[0078] The potential impact of impurities on weld performance will be graded: If Fe, Ni, or other metallic impurities exceed the standard, it will lead to a decrease in the bending strength and fracture toughness of the graphite weld, with an assessment conclusion of moderate impact; if S exceeds the standard, low-melting-point sulfides will form at the graphite grain boundaries at high temperatures, causing hot brittleness in the weld above 300℃, with an assessment conclusion of severe impact; for welds in chemical equipment, if Cl⁻ exceeds the standard, it will accelerate surface corrosion of the weld in acidic media, increasing the corrosion rate to 0.2mm / year, with an assessment conclusion of needing to repair the weld with an anti-corrosion coating; if Na⁺, K⁺, or other alkali metal ions exceed the standard, they will react with carbon in graphite to form carbides, damaging the weld seal, with an assessment conclusion of recommending replacement of the base material in the weld area; if B, Al, or other impurities exceed the standard, the upper limit of the high temperature resistance of the graphite weld will drop from 2000℃ to below 1600℃, with the assessment conclusion needing to be combined with the actual operating temperature of the equipment. If the operating temperature is >1500℃, it is determined that re-welding is required.

[0079] The comprehensive weld quality assessment unit is used to combine defect identification results with trace element detection data to comprehensively evaluate weld quality and generate quality grades and inspection reports.

[0080] As a preferred feasible embodiment, the comprehensive weld quality assessment unit uses defect identification results and trace element detection data as core inputs: First, the defect severity index (SI) is divided into four intervals: SI≤0.3 is slight, 0.3<SI≤0.6 is moderate, 0.6<SI≤0.8 is severe, and SI>0.8 is extremely severe. Then, the trace element exceedance is combined for correction. The correction rules are as follows: if there is an exceedance of high-risk elements such as S and Cl, or multiple elements exceed the limit together, the SI interval is increased by 1 level; if only a single low-risk element slightly exceeds the limit, the original SI interval is maintained. Then, the corresponding quality level is matched: SI≤0.3 and no exceeding impurities is excellent and can be put into use directly; 0.3<SI≤0.6 or only low-risk elements slightly exceed the limit is qualified and needs to be recorded and archived and routinely inspected; 0.6<SI≤0.8 or high-risk elements slightly exceed the limit is to be rectified and needs to be repaired and re-inspected; SI>0.8 or high-risk elements severely exceed the limit is unqualified and prohibited from service.

[0081] The final test report integrates the defect location map, trace element test data, and performance impact assessment conclusions, clearly indicating the quality level, rectification suggestions, and re-inspection deadline. The report is also synchronized to the equipment lifecycle management system to achieve traceability of quality data.

[0082] Furthermore, to achieve interpretable determination and actionable rectification of defect causes, this embodiment outputs defect cause probabilities and rectification suggestions while generating quality grades: The system inputs defect types (porosity / cracks / lack of fusion / inclusions), defect geometric quantities (area, perimeter, aspect ratio, etc.), key spectral features (spectral angle, peak shift, Mahalanobis distance, feature coupling degree R), and trace element spectra (Fe, Ni, S, Cu, Cl, etc.) into the cause discrimination model, establishes a "feature-composition-cause" association rule library, and outputs several candidate causes and their confidence levels, such as "local embrittlement caused by inclusions," "high-temperature embrittlement caused by sulfur impurities," and "lack of fusion caused by insufficient fusion," etc. At the same time, it automatically generates rectification suggestions based on the cause and risk matrix, including but not limited to: suggested repair weld length range, recommended re-inspection method and re-inspection time limit, and suggested process parameter adjustment direction (such as cleaning and dust removal, shielding gas purity control, welding heat input adjustment, material pretreatment and drying, etc.). Through the above mechanism, the test report not only provides the quality level, but also provides actionable rectification paths and traceability basis, thereby improving the efficiency of on-site handling.

[0083] The inspection data management module includes: a multispectral data storage library, a weld defect case library, and a spectral characteristic parameter library;

[0084] The multispectral data repository is used to store multispectral raw images and spectral feature data of weld seams in various graphite equipment.

[0085] It should be explained that the multispectral data repository consists of several layers. The bottom layer is the raw image storage layer, which uses a three-level folder structure based on device model, weld number, and scan time. Raw spectral images for each band (450nm, 800nm, 1500nm, 2000nm, and 2800nm) are saved in TIFF format, along with the scanning parameters for each frame to ensure traceability of the raw data. The middle layer is the spectral feature data layer, which associates the pixel coordinates of the raw images and stores preprocessed data such as the light intensity value, spectral reflectance / absorption curve, and characteristic peak wavelength and intensity for each pixel. This data is structured using an SQLite database and supports quick filtering by band range and weld region. For example, it allows for one-click retrieval of all characteristic peak data for a pipe weld in the 2500-3000nm shortwave infrared band. The top layer is the index layer, which uses an index table to locate the corresponding raw image and feature data by weld number or scan time, avoiding efficiency losses caused by cross-layer searches.

[0086] Furthermore, in addition to scanning parameters, this embodiment also writes the radiometric calibration parameters (dark field Idark, white field Iwhite), cross-polarization angle, exposure time / gain, lens temperature drift compensation coefficient, and image registration extrinsic parameter matrix of each band into the metadata file, and binds them to the unique index ID of the frame image; at the same time, the algorithm version number and model verification code (such as the version of the defect segmentation model and the defect classification model) are recorded in the index layer, as well as the associated key values ​​with the sample barcode, ICP-MS detection batch number, and final report number. Through the above data structure, any detection conclusion can be traced back to the complete chain of "original image - calibration parameters - processing version - sampling result - report output", meeting the needs of auditing and re-inspection reproduction.

[0087] The weld defect case library is used to record the types, characteristics, detection results, and handling solutions of historical defects.

[0088] It should be explained that each case in the weld defect case library first records basic defect information, including defect type, defect location, defect size, and corresponding equipment service scenario; secondly, it links complete detection data, covering multispectral image fragments, spectral feature parameters, trace element detection reports, and defect severity index; finally, it supplements the defect treatment results, such as repair process, re-inspection data, and subsequent equipment service status; in addition, the weld defect case library supports a similar case search function. By inputting key parameters such as the type, size, and spectral characteristics of the defect to be analyzed, it can automatically match historical cases in the library with a similarity of ≥80%, providing a reference for defect cause analysis and treatment plan formulation. For example, when a pipeline weld is found to have incomplete fusion and excessive sulfur content, the system can quickly retrieve the repair welding process parameters and re-inspection results of similar past cases.

[0089] The spectral feature parameter library is used to store feature parameter data of normal welds and abnormal areas.

[0090] It should be explained that both the normal and abnormal libraries in the spectral feature parameter library use the feature parameter template as a unified data structure. Each record contains weld identification, feature parameters, and calculation conditions. The data in the normal library comes from a large number of graphite weld samples that have been manually re-inspected and confirmed to be defect-free. The normal range of each feature parameter is determined through statistical analysis. New normal weld data is added every quarter to update the parameter distribution range and avoid judgment bias caused by batch differences in equipment. The abnormal library contains feature data of defect areas confirmed by inspection. It is stored according to defect type, and each data point is labeled with the corresponding defect level and trace element exceedance status for comparison by the spectral anomaly identification unit.

[0091] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0092] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart detection and location system for weld defects in graphite equipment based on multispectral imaging, characterized in that, include: The multi-band image acquisition unit is used to perform omnidirectional scanning of the weld area of ​​the graphite equipment and transmit the acquired spectral image data of multiple bands to the spectral database. The spectral parameter analysis unit is used to extract characteristic parameters such as spectral angle, main absorption peak position, local standard deviation, and Mahalanobis distance of the weld area from the spectral database, and to establish a spectral feature library of normal welds and abnormal areas. The spectral parameter analysis unit constructs an enhanced feature vector from the four types of spectral features—spectral angle, main absorption peak position, local standard deviation, and Mahalanobis distance—through a synergistic matrix, and additionally introduces parameters characterizing the degree of coupling of spectral perturbations to improve feature discrimination capability. The four types of spectral features are constructed into an enhanced feature vector V using a synergistic matrix: The additional feature coupling degree R is introduced, and its calculation method is as follows: This parameter characterizes the degree of coupling between spectral perturbations caused by defects; The spectral anomaly identification unit is used to locate the binary mask image of the suspected defect area with spectral anomalies by comparing the spectral parameters of the weld to be detected with those of normal welds in the spectral feature library. The weld defect location unit is used to locate suspected defects and mark the defect boundaries and extent by using a spectral anomaly mask combined with an image segmentation algorithm. The defect identification and classification unit is used to classify the located defects and calculate the defect severity index. The defect identification and grading unit inputs each defect image block into a convolutional neural network and outputs the defect category and confidence level. The defect category includes at least porosity, cracks, lack of fusion, and inclusions. The defect severity index is calculated by weighting the defect type weight W, the defect size score E, and the defect location coefficient F. The elemental composition analysis unit is used to perform trace element detection on samples from suspected defect locations when the defect severity index is greater than a preset threshold, and to analyze the types and contents of trace elements. The impurity and composition assessment unit is used to compare the detected trace element data with the graphite equipment material standards to determine whether there are excessive impurities and their potential impact on weld performance. The comprehensive weld quality assessment unit is used to combine defect identification results with trace element detection data to generate a quality grade and inspection report for the weld quality. The inspection data management module includes a multispectral data repository, a weld defect case library, and a spectral characteristic parameter library.

2. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 1, characterized in that: The multi-band image acquisition unit uses a combination of spiral scanning, serpentine reciprocating scanning and multi-angle supplementary scanning to perform a comprehensive scan of the weld body, heat-affected zone and base material transition zone, and dynamically corrects the scanning trajectory based on laser positioning and visual guidance.

3. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 1, characterized in that: The multi-band image acquisition unit operates in the spectral range of 400–3000 nm, including multiple fixed-band imaging channels of 450 nm, 800 nm, 1500 nm, 2000 nm, and 2800 nm.

4. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 1, characterized in that: The spectral anomaly identification unit classifies pixels based on multi-feature collaborative judgment rules, generates a binary mask image of the suspected defect area, and supports manual review of suspicious pixels; when the manual review confirms that the suspicious pixel is normal, its spectral parameters are added to the normal library of the spectral feature parameter library.

5. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 4, characterized in that: After generating the binarized mask image, the system introduces a spectral texture tensor T and performs region growing based on the gradient magnitude of T to ensure that the same defect region maintains contour consistency across different bands.

6. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 1, characterized in that: The weld defect location unit performs median filtering for noise reduction, connected component marking, and boundary tracking on the binarized mask image of the suspected defect area to obtain the circumscribed rectangle, area, center coordinates, and perimeter of the defect area, and generates a defect location map on the original image.

7. The intelligent detection and positioning system for weld defects in graphite equipment based on multispectral imaging according to claim 1, characterized in that: When the defect severity index exceeds the threshold, the elemental composition analysis unit receives the three-dimensional coordinates of the suspected defect location and controls the robotic arm equipped with a micro-sampling device to move to the target area. Micro-samples are obtained by fiber laser cutting and sent to the automated sample introduction system of the inductively coupled plasma mass spectrometry equipment through a sealed pipeline to complete the trace element detection. The detection results are associated with the defect number, sampling location and spectral characteristic parameters to form a data chain and are transmitted back to the detection data management module.

Images