A method and system for detecting defects in liquid-cooled cold plates based on image processing

Abstract

This invention provides a method and system for detecting defects in liquid-cooled cold plates based on image processing, belonging to the field of liquid cooling technology. The method involves multi-source data acquisition from the liquid-cooled cold plate to obtain a multi-source dataset, followed by data preprocessing to generate a fused dataset. A liquid cooling defect detection model is then constructed, and the fused dataset is input into the model for pre-training. Simultaneously, model distillation is performed to obtain a lightweight liquid cooling defect detection model, which is then deployed to the inspection end of the liquid-cooled cold plate production line. Real-time monitoring of the liquid-cooled cold plate is conducted based on the lightweight model, and monitoring results are obtained. Defect verification is performed on the monitoring results according to a liquid cooling defect verification mechanism, and the defect verification results are obtained. Based on these results, the liquid cooling defect detection model is optimized, thereby improving the efficiency, accuracy, and timeliness of defect detection in liquid-cooled cold plates and their correlation with production processes.

Description

Technical Field

[0001] This invention relates to the field of liquid cooling technology, and more specifically, to a method and system for detecting defects in liquid cooling plates based on image processing. Background Technology

[0002] As a core component for efficient heat dissipation in data centers, the manufacturing quality of liquid-cooled cold plates directly affects the reliability and energy efficiency of the equipment. With the widespread application of liquid cooling technology in high-power-density scenarios, the complexity of the microchannel structure inside the cold plate has significantly increased. Traditional testing methods often have certain limitations in dealing with multi-dimensional defects such as surface defects, internal flow channel blockage, three-dimensional deformation, and abnormal thermal resistance.

[0003] On the one hand, traditional methods typically only detect single dimensions such as surfaces or flow channels, lacking simultaneous acquisition and correlation analysis of multi-physical field information such as polarization, flux, deformation, and thermal resistance, resulting in incomplete defect feature extraction. On the other hand, traditional methods rely on high-computing platforms to process massive amounts of data, often making it difficult to achieve real-time data parsing on edge devices, affecting production line inspection efficiency. Furthermore, static inspection rules cannot adapt to process fluctuations and new defect patterns, requiring frequent manual intervention and adjustments, increasing quality control costs. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a liquid-cooled cold plate defect detection method and system based on image processing. This addresses the technical issues in the prior art, where traditional detection methods lack simultaneous acquisition and correlation analysis of multi-physical field information, resulting in incomplete defect feature extraction, difficulties in real-time edge detection, and the inability of static detection rules to adapt to process fluctuations and new defect patterns.

[0005] The purpose and effectiveness of the image processing-based liquid-cooled cold plate defect detection method and system of the present invention are achieved by the following specific technical means:

[0006] A method for detecting defects in liquid-cooled cold plates based on image processing includes:

[0007] S1: Perform multi-source data acquisition on the liquid-cooled plate to obtain a multi-source dataset of the liquid-cooled plate, and perform data preprocessing on the multi-source dataset of the liquid-cooled plate to generate a fused dataset of the liquid-cooled plate.

[0008] S2: Construct a liquid cooling defect detection model and input the liquid cooling plate fusion dataset into the liquid cooling defect detection model for model pre-training;

[0009] S3: Perform model distillation on the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line;

[0010] S4: Real-time monitoring of the liquid cooling plate is performed based on the lightweight liquid cooling defect detection model to obtain the monitoring results of the liquid cooling plate, which include a liquid cooling plate defect coordinate diagram and a process correlation report.

[0011] S5: Verify the defects of the liquid cooling plate monitoring results according to the liquid cooling defect verification mechanism, obtain the defect verification results, and optimize the liquid cooling defect detection model based on the defect verification results.

[0012] As a further aspect of the present invention, the step of acquiring multi-source data from the liquid-cooled plate to obtain a multi-source dataset of the liquid-cooled plate, and preprocessing the multi-source dataset of the liquid-cooled plate to generate a fused dataset of the liquid-cooled plate, includes:

[0013] The reflective properties and internal structural features of the liquid-cooled plate surface are analyzed in real time, and data is collected from the liquid-cooled plate according to the preset acquisition compensation rules to obtain a multi-source dataset of the liquid-cooled plate. The multi-source dataset of the liquid-cooled plate includes surface polarization feature data, internal structural data, three-dimensional deformation data, and thermal resistance distribution field data.

[0014] Polarization difference operation and contrast enhancement are performed on the surface polarization feature data to eliminate metal reflection interference and extract subsurface texture defects, generating liquid-cooled polarization feature map;

[0015] The internal structural data is subjected to spectral noise reduction and channel segmentation based on short-wave infrared transmittance, and the microchannel flux distribution characteristics are quantified to generate a liquid cooling flux distribution map.

[0016] Perform point cloud statistical filtering and normal vector analysis on the three-dimensional deformation data to reconstruct the three-dimensional contour of the microchannel and calculate the depth deviation and fin warpage to generate a liquid cooling point cloud map.

[0017] The thermal resistance distribution field data is generated by calibrating infrared thermograms based on material emissivity and combining them with the Fourier heat conduction equation to invert the spatial distribution of thermal resistance coefficients.

[0018] Based on the liquid cooling point cloud map, the liquid cooling polarization feature map, liquid cooling flux distribution map and thermal resistance field map are mapped to the same feature space according to the affine transformation to generate a liquid cooling plate fusion dataset.

[0019] As a further aspect of the present invention, the preset acquisition compensation rule includes:

[0020] Dynamically adjust the spatial pose of the optical sensor to collect data from the highly reflective area on the surface of the liquid-cooled plate;

[0021] The internal structural data of the liquid-cooled plate was acquired by switching between short-wave infrared penetrating scanning and ultrasonic focusing detection combined modes.

[0022] As a further aspect of the present invention, the step of constructing a liquid cooling defect detection model, which involves inputting the liquid cooling plate fusion dataset into the liquid cooling defect detection model for model pre-training, includes:

[0023] Obtain defect-free normal liquid cooling plate samples and defective abnormal liquid cooling plate samples, and generate an adversarial network based on conditions to extract the structural features and defect features of the liquid cooling plate. Construct a liquid cooling defect training set based on the structural features and defect features of the liquid cooling plate.

[0024] The liquid cooling defect detection model is pre-trained based on the liquid cooling defect training set to obtain the liquid cooling plate defect rules. The liquid cooling plate defect rules include polarization deformation coupling relationship, flux thermal resistance mapping relationship and deformation thermal resistance coupling relationship.

[0025] The production process parameters of the liquid-cooled cold plate are obtained, which are represented by the welding current parameters, pressure parameters, and time parameters. The liquid-cooled cold plate defect detection model is pre-trained in the second stage by combining the liquid-cooled cold plate defect rules and the liquid-cooled cold plate fusion dataset to obtain the correlation and coupling relationship between liquid-cooled cold plate defects and production process.

[0026] As a further aspect of the present invention, the method further includes:

[0027] The polarization deformation coupling relationship is expressed as the coupling relationship between the surface defect state of the liquid-cooled plate and the polarization scattering intensity.

[0028] The flux-thermal resistance mapping relationship is expressed as the mapping relationship between the blockage cutoff rate caused by internal defects in the liquid-cooled plate and the change in thermal resistance.

[0029] The deformation thermal resistance coupling relationship is expressed as the coupling relationship between the deformation of the liquid-cooled cold plate and the abnormal thermal flow resistance of the liquid-cooled cold channel.

[0030] The associated coupling relationships include the coupling relationship between the current parameter offset and the deformation of the liquid-cooled plate, the correlation between the pressure parameter and the flow loss of the internal cooling channels of the liquid-cooled plate, and the mapping relationship between the time parameter deviation and the thermal flow resistance of the liquid-cooled plate.

[0031] As a further aspect of the present invention, the step of performing model distillation on the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and deploying the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line, includes:

[0032] Based on the pre-trained liquid cooling defect detection model, process features are extracted to obtain the process parameter correlation matrix. The process parameter correlation matrix is ​​represented as the correlation parameter matrix between the liquid cooling plate production process parameters and the liquid cooling plate physical field anomaly.

[0033] The physical field of the liquid-cooled plate includes the surface polarization, internal flux, three-dimensional deformation, and thermal resistance field of the liquid-cooled plate.

[0034] The process parameter correlation matrix of the teacher model and the student model is forced according to the mean square error, and multiple weights are applied to the key process defect samples for optimization and transfer. The key process defect samples are represented as process defect samples in the process parameter correlation matrix where the process parameters exceed the preset process threshold, and the physical field anomaly of the liquid cooling plate also exceeds the preset anomaly threshold.

[0035] The student model is trained based on the physical field inversion mechanism, which is represented by randomly shading the input channel corresponding to the physical field data of the liquid cooling plate. The student model is required to infer the missing features through the residual physical field data, obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line for real-time monitoring of liquid cooling plate defects.

[0036] As a further aspect of the present invention, the real-time monitoring of the liquid-cooled cold plate based on the lightweight liquid-cooling defect detection model to obtain the monitoring results of the liquid-cooled cold plate includes:

[0037] Based on the lightweight liquid cooling defect detection model, defect detection and process collaborative diagnosis are performed on the liquid cooling plates produced on the liquid cooling plate production line to generate liquid cooling plate monitoring results.

[0038] The defect coordinate diagram of the liquid-cooled plate is composed of the spatial coordinates of the defects in the liquid-cooled plate.

[0039] The process correlation report is a report on the correlation between defects in liquid-cooled cold plates and corresponding abnormalities in production process parameters.

[0040] As a further aspect of the present invention, the step of verifying the defects of the liquid-cooled cold plate monitoring results according to the liquid-cooling defect verification mechanism and obtaining the defect verification results includes:

[0041] The surface defect coordinates are extracted from the defect coordinate map of the liquid-cooled plate, and the matching between polarization scattering intensity and deformation height is verified to obtain the polarization deformation matching results.

[0042] The coordinates of internal defects in the defect coordinate map of the liquid-cooled cold plate are extracted, and the quantitative relationship between flux loss rate and thermal resistance increase is verified to obtain the flux-thermal resistance correlation matching results.

[0043] Map the process parameters in the process association report to the corresponding defect coordinates, verify the association parameters of the corresponding defect coordinates, and obtain the process defect verification results;

[0044] The polarization deformation matching results, flux thermal resistance correlation matching results, and process defect verification results are encapsulated to generate defect verification results.

[0045] As a further aspect of the present invention, the optimization of the liquid cooling defect detection model based on the defect verification results includes:

[0046] The defect verification results are analyzed and located to obtain the rules to be optimized corresponding to the liquid cooling defect detection model. The rules to be optimized are then optimized, and a corresponding training set is generated to verify the optimization of the rules.

[0047] A liquid-cooled cold plate defect detection system based on image processing, comprising:

[0048] The data acquisition module is used to acquire multi-source data of the liquid-cooled plate, obtain the multi-source dataset of the liquid-cooled plate, and perform data preprocessing to generate a fused dataset of the liquid-cooled plate.

[0049] The model training module is used to construct a liquid cooling defect detection model and pre-train the liquid cooling defect detection model based on the liquid cooling plate fusion dataset.

[0050] The model distillation module is used to distill the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and then deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line.

[0051] The defect monitoring module is used to monitor the liquid cooling plate in real time according to the lightweight liquid cooling defect detection model and obtain the monitoring results of the liquid cooling plate.

[0052] The verification and optimization module is used to verify the defects in the monitoring results of the liquid-cooled cold plate, obtain the defect verification results, and optimize the liquid-cooled defect detection model based on the defect verification results.

[0053] Based on the above aspects, the embodiments of this application realize multi-source data acquisition of liquid-cooled cold plates, obtain multi-source datasets of liquid-cooled cold plates, and perform data preprocessing on the multi-source datasets of liquid-cooled cold plates to generate fused datasets of liquid-cooled cold plates. By simultaneously acquiring the full-dimensional physical field information of the liquid-cooled cold plate, including surface polarization, internal microchannel flux, three-dimensional deformation, and thermal resistance, and by eliminating the influence of useless data through coordinate space alignment and data feature enhancement, high-quality input is provided for subsequent multi-physics coupling diagnosis, thereby improving the accuracy of defect location of liquid-cooled cold plates and the timeliness of process traceability.

[0054] A liquid cooling defect detection model is constructed. The fusion dataset of liquid cooling plates is input into the liquid cooling defect detection model for model pre-training. The pre-trained liquid cooling defect detection model is then distilled to obtain a lightweight liquid cooling defect detection model. This lightweight liquid cooling defect detection model is then deployed to the detection end of the liquid cooling plate production line. Through model pre-training, the defect feature rules and process association rules of the liquid cooling plate are learned. Combined with lightweight pruning to compress the model parameter scale, the multi-physics field data can be quickly parsed at the edge device, thereby improving the defect detection efficiency of liquid cooling plates.

[0055] A lightweight liquid cooling defect detection model is used to monitor the liquid cooling plate in real time, and the monitoring results are obtained. The monitoring results include a liquid cooling plate defect coordinate map and a process correlation report. The monitoring results are verified according to the liquid cooling defect verification mechanism, and the defect verification results are obtained. Based on the defect verification results, the liquid cooling defect detection model is optimized. The liquid cooling plate defect rules and parameter thresholds are continuously calibrated through a closed-loop feedback optimization mechanism, thereby improving the accuracy and generalization ability of liquid cooling plate defect diagnosis. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the execution flow of a liquid-cooled cold plate defect detection method based on image processing provided in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a liquid-cooled cold plate defect detection system based on image processing provided in an embodiment of the present invention. Detailed Implementation

[0058] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the technical solutions of the present invention, but should not be used to limit the scope of protection of the present invention. Example

[0059] As attached Figure 1 , Figure 2 As shown:

[0060] This invention provides an image processing-based method for detecting defects in liquid-cooled cold plates, applicable to liquid-cooling defect detection, comprising the following steps:

[0061] Step S1: Collect multi-source data from the liquid-cooled plate to obtain a multi-source dataset of the liquid-cooled plate, and preprocess the multi-source dataset of the liquid-cooled plate to generate a fused dataset of the liquid-cooled plate.

[0062] In this embodiment, step S1 includes:

[0063] Step S11: Perform multi-source data acquisition on the liquid-cooled plate to obtain a multi-source dataset of the liquid-cooled plate.

[0064] Specifically, multi-source data is collected according to the intelligent sensing collaboration mechanism. The intelligent sensing collaboration mechanism is to analyze the reflective characteristics and internal structural features of the surface of the liquid-cooled plate workpiece in real time, and to collect data from the liquid-cooled plate according to the preset acquisition compensation rules to obtain a multi-source dataset of the liquid-cooled plate.

[0065] Understandably, the preset acquisition compensation rules include:

[0066] The spatial pose of the optical sensor is dynamically adjusted to collect data on highly reflective areas on the surface of the liquid-cooled plate. For example, the robotic arm autonomously plans the scanning path based on the reinforcement learning algorithm and automatically tilts the polarization camera angle for highly reflective areas such as weld seams.

[0067] The system switches between short-wave infrared penetration scanning and ultrasonic focused detection modes to collect internal structural data of the liquid-cooled plate. For example, when the short-wave infrared transmittance is abnormal, it is determined to be a suspected microchannel blockage. At this time, the ultrasonic probe is automatically triggered to focus and scan, and the nature of the foreign object causing the microchannel blockage is identified by the difference in acoustic impedance.

[0068] Understandably, a high-precision polarization camera is used to simultaneously capture polarization state data from three spectral channels: visible light (wavelength range [400, 760] nm), near-infrared light (wavelength range [900, 1700] nm), and short-wave infrared light (wavelength range [1.7, 2.5] μm). The near-infrared band penetrates surface oil stains to reveal hidden scratches, while short-wave infrared light indirectly visualizes microchannel blockage defects through changes in medium transmittance. Simultaneously, by utilizing the difference in polarization characteristics between metal surface reflections and actual defects, Stokes vector analysis is used to eliminate reflective interference in real time. For the deformed fin area, a three-dimensional point cloud is constructed and spatially mapped with the infrared thermal image to capture the thermal flow field distortion characteristics caused by structural deformation.

[0069] The multi-source dataset for liquid-cooled cold plates includes surface polarization feature data, internal structure data, three-dimensional deformation data, and thermal resistance distribution field data;

[0070] The surface polarization feature data is represented as the surface texture data of the liquid cooling plate after polarization analysis.

[0071] The internal structure data is represented as the microchannel transmittance spectrum and the acoustic impedance distribution of the medium in the liquid-cooled plate.

[0072] The three-dimensional deformation data are represented as channel depth deviation, weld cross-sectional curvature, warpage, and surface fitting residual of the liquid-cooled plate.

[0073] The thermal resistance distribution field data is represented as the temperature gradient field and thermal resistance distribution of the liquid-cooled plate under working load.

[0074] Step S12: Perform data preprocessing on the multi-source dataset of liquid-cooled cold plates to generate a fused dataset of liquid-cooled cold plates.

[0075] Specifically, polarization difference operations and contrast enhancement are performed on the surface polarization feature data to eliminate metal reflection interference and extract subsurface texture defects, generating a liquid-cooled polarization feature map.

[0076] Understandably, highly reflective areas exhibit low differential values ​​due to their significant polarization characteristics, while genuine scratches and oxide spots maintain high differential responses. By combining dynamic adjustment of the incident angle and threshold segmentation, interference from metal reflections can be eliminated. For example, at a 60° incident angle, the polished area of ​​a weld seam on an aluminum cold-rolled plate has a grayscale value of 250 for horizontally polarized light and 15 for vertically polarized light, resulting in a calculated differential value of 235. The local standard deviation of this polished area is 5.2, the global differential value mean is 85, and the global standard deviation is 40. Based on the global differential value mean and global standard deviation, a differential separation threshold is dynamically calculated: 85 + 3 * 40 = 205. Since the differential value of the polished area of ​​the aluminum cold-rolled plate weld seam is 235, which is greater than the differential separation threshold, and the local standard deviation of 5.2 is lower than the preset standard deviation threshold of 10, it is determined to be interference. The grayscale value of the area is modified to 0. There are scratches on the surface of the aluminum cold plate. At a 60° incident angle, the horizontal polarization intensity of these scratches is 120, and the vertical polarization intensity is 100. The calculated difference is 20, and the local standard deviation of the scratch area is 28.7. Since the difference value of 20 is lower than the difference separation threshold, contrast enhancement is directly performed. Contrast enhancement means normalizing the difference value. Normalization means dividing the difference value by the average of the horizontal and vertical polarization intensities, i.e., 20 ÷ (120 + 100) / 2 = 0.18. The normalized difference value is 0.18. This normalized difference value is enhanced by 10 times and limited by 255, i.e., 0.18 * 10 * 255 = 459. Since 459 is much greater than 255, the output grayscale value is modified to 255.

[0077] Specifically, the internal structural data is subjected to spectral noise reduction and channel segmentation of short-wave infrared transmittance, the microchannel flux distribution characteristics are quantified, and a liquid cooling flux distribution map is generated.

[0078] Understandably, short-wave infrared transmittance spectral noise reduction is implemented based on non-uniformity correction to eliminate detector inherent noise. A five-level hard threshold filter using the sym8 wavelet basis is employed, where the threshold is three times the noise standard deviation, eliminating high-frequency shot noise. Anisotropic diffusion filtering suppresses fluid turbulence scattering noise. For example, in a liquid-cooled plate with a static coolant, a bright spot with a grayscale of 42 appears at a fixed position in the image. Analysis reveals that this bright spot originates from a defect in the detector's pixel circuitry. A dark current calibration image captured by a sealed lens is applied, and non-uniformity correction is performed to reduce the bright spot's grayscale to zero, thereby eliminating detector inherent noise and improving cooling efficiency. After the pump starts, the flow rate reaches 2.5 m / s, and the image suddenly shows snowflake-like noise. Five-level decomposition is performed using the sym8 wavelet basis, and a hard threshold of three times the noise standard deviation is used. If the noise standard deviation is 8 at this time, the corresponding hard threshold is 24. Snowflakes with gray levels below 24 are directly removed, thereby eliminating high-frequency shot noise. When the coolant flows at high speed through the bifurcation of a Y-shaped microchannel, if the microchannel width is 0.25 mm at this time, a vortex-like gray fog band is generated. Anisotropic diffusion filtering is used to smooth the turbulent area while preserving the sharp edge of the microchannel, that is, to preserve the sharp edge of 0.25 mm, and to homogenize the turbulent area inside the gray fog band, thereby suppressing fluid turbulence scattering noise.

[0079] Furthermore, the channel segmentation employs a combination of phase consistency edge detection and watershed algorithm to segment the microchannels within the liquid-cooled plate. A filter bank deployed in six directions scans the liquid-cooled plate image, calculating the phase consistency value of the Fourier components of each pixel to generate a phase edge map. Based on this phase edge map, a distance transform is calculated to generate a distance gradient field for the microchannel skeleton of the liquid-cooled plate. Marking points are set at local extrema, which are represented by ridges in the distance gradient field where the gradient is greater than 200 gray levels. Watershed flooding segmentation is then performed to obtain the segmentation results. These results are then validated, retaining connected components with a segmentation area greater than a preset area threshold (e.g., retaining connected components with a segmentation area greater than 0.05 square millimeters to eliminate scattered noise). Topological connectivity checks are applied to branch intersections, such as requiring the angle between adjacent channels to be less than 100°.

[0080] Specifically, point cloud statistical filtering and normal vector analysis are performed on the three-dimensional deformation data to reconstruct the three-dimensional contour of the microchannel and calculate the depth deviation and fin warpage to generate a liquid cooling point cloud map.

[0081] Understandably, point cloud statistical filtering involves calculating the average distance from each spatial point to its 50 nearest neighbors. If a point's distance exceeds three times the standard deviation of the average distance, it is considered a noise point and removed. The normal vector direction within a fixed-radius neighborhood of each point is calculated, and the local surface normal vector direction is obtained through principal component analysis. For example, the local surface normal vector direction within a neighborhood with a radius of 0.5 mm is obtained through principal component analysis. A normal vector field is constructed based on the local surface normal vector direction, and a Poisson surface is constructed to fill in the missing areas of the point cloud. The Z-axis depth deviation and fin warpage are calculated. The Z-axis depth deviation represents the depth deviation relative to the liquid-cooled plate production model, such as the depth deviation relative to C... The deviation of the AD model is expressed as the fin warpage amount, which is quantified by measuring the diagonal height difference using the four-point positioning method. The output is a liquid-cooled point cloud. For example, welding causes point cloud gaps and fin wave deformation in the Y-shaped channel intersection area. The core area of ​​wave deformation is located by point cloud filtering and normal vector analysis. The surface is reconstructed based on the normal vector field to fill in the missing point cloud area. The Z-axis depth deviation and fin warpage amount are calculated to be 0.18 mm and 0.15 mm per square millimeter, respectively. The output is a liquid-cooled point cloud. In the liquid-cooled point cloud, the depth exceeding the limit area with a depth greater than 0.1 mm is marked in red, and the warpage gradient is displayed in blue contour lines with a gradient of 0.05 mm.

[0082] Specifically, the thermal resistance distribution field data is generated by calibrating infrared thermograms based on material emissivity and combining them with the Fourier heat conduction equation to invert the spatial distribution of thermal resistance coefficients.

[0083] Understandably, the thermal resistance distribution field data is represented by infrared thermography calibrated based on the material emissivity. This is to address the differences in emissivity of the cold plate materials, such as the aluminum alloy body having an emissivity of 0.85 and the oxide layer only 0.45. The double blackbody reference method is used to correct the temperature measurement deviation, and the spatial distribution of the thermal resistance coefficient is represented by inverting the Fourier heat conduction equation. This is based on dividing the internal structure of the liquid-cooled cold plate into grid cells, tracing the heat flux density through the temperature gradient, and correlating the heat flux density with the temperature difference. For example, the measured heat flux density of a certain aluminum alloy fin area is 15kW / m², corresponding to a temperature difference of 3.2℃. Based on the heat flux density and temperature difference, the thermal resistance is back-calculated, and the thermal resistance value is obtained as 0.21℃ / W. This thermal resistance value is dynamically mapped to the internal structure of the liquid-cooled cold plate to divide the grid cells and generate a thermal resistance field map.

[0084] Understandably, the thermal resistance field diagram uses a color gradient from dark blue, light blue, green, yellow, orange to red to visually represent the spatial distribution of thermal resistance coefficients. The dark blue area represents thermal resistance values ​​not exceeding 0.15℃ / W, used to characterize the optimal heat dissipation efficiency zone, where heat is efficiently dissipated by the coolant. The light blue to green area represents thermal resistance values ​​within the range of (0.15, 0.25]℃ / W, used to characterize the acceptable heat dissipation efficiency zone, meeting heat dissipation standards. The yellow to orange area represents thermal resistance values ​​within the range of (0.25, 0.35]℃ / W, used to characterize the heat dissipation efficiency decay zone, i.e., the presence of blockages or insufficient coolant flow rate. The red area represents thermal resistance values ​​exceeding 0.35℃ / W, used to characterize the heat dissipation failure zone, i.e., heat accumulates and is difficult to dissipate.

[0085] Furthermore, based on the liquid cooling point cloud map, the liquid cooling polarization feature map, liquid cooling flux distribution map and thermal resistance field map are mapped to the same feature space according to the affine transformation to generate a liquid cooling plate fusion dataset.

[0086] Specifically, using preset physical markers in the liquid-cooled point cloud map as the spatial origin, such as the center of the welding positioning column or the corner of the cold plate mounting hole, the positions corresponding to the liquid-cooled polarization feature map and the thermal flux resistance field map are matched. The rotation matrix and translation vector are obtained through the least squares method. The rotation matrix is ​​used to eliminate the angular deviation between the coordinate systems corresponding to different data, and the translation vector is used to compensate for the error caused by mechanical displacement. The liquid-cooled polarization feature map, liquid-cooled flux distribution map, and thermal flux resistance field map are unified in spatial coordinate system based on the rotation matrix and translation vector. For example, the liquid-cooled polarization feature map corrects the camera pitch angle deviation based on the rotation matrix, so that the coordinates corresponding to the surface defects are accurately mapped to the surface of the point cloud fins; the liquid-cooled flux distribution map compensates for installation misalignment based on the translation vector, ensuring that the microchannel blockage points inside the liquid-cooled cold plate are bound to the spatial path of the point cloud channel; the thermal flux resistance field map eliminates the viewing angle tilt based on the rotation matrix and corrects the thermal imager offset based on the translation vector.

[0087] Understandably, obtaining the rotation matrix and translation vector using the least squares method involves selecting pre-defined physical marker points of the same name from the liquid cooling point cloud map, liquid cooling polarization feature map, liquid cooling flux distribution map, and thermal resistance field map. The rotation matrix and translation vector are then obtained by calculating the centroid alignment and spatial residual. For example, the point cloud coordinates of a server cold plate are A (101.2mm, 203.5mm), B (205.8mm, 198.1mm), and C (102.5mm, 305.7mm). The pixel coordinates corresponding to the liquid-cooled polarization feature map are A'(325px, 780px), B'(625px, 772px), and C'(330px, 1150px), with corresponding point cloud centroids of 136.5 / 235.8mm and liquid-cooled polarization feature map centroids of 426.7 / 900.7px. The alignment and spatial residual between the point cloud centroids and the liquid-cooled polarization feature map centroids are calculated, the rotation matrix and translation vector are obtained, and the two centroids are made coincident through rotation and translation.

[0088] Step S2: Construct a liquid cooling defect detection model by inputting the liquid cooling plate fusion dataset into the liquid cooling defect detection model for model pre-training.

[0089] Specifically, defect-free normal liquid cooling plate samples and defective abnormal liquid cooling plate samples are obtained, and an adversarial network is generated according to conditions to extract the structural features and defect features of the liquid cooling plate. A liquid cooling defect training set is constructed based on the structural features and defect features of the liquid cooling plate.

[0090] The liquid cooling defect detection model is pre-trained based on the liquid cooling defect training set to obtain the liquid cooling plate defect rules, which include polarization deformation coupling relationship, flux thermal resistance mapping relationship and deformation thermal resistance coupling relationship.

[0091] The polarization deformation coupling relationship refers to the coupling relationship between the surface defect state of the liquid-cooled plate and the polarization scattering intensity. That is, when scratches, oxidation or welding deformation occur on the surface of the cold plate, its micro-morphology will change the polarization scattering characteristics of light. For example, scratches with a depth of more than 0.1 mm will cause a sudden change in local curvature, which will cause a sharp increase in the intensity of vertically polarized light scattering.

[0092] The flux-thermal resistance mapping relationship represents the mapping relationship between the blockage loss rate caused by internal defects in the liquid-cooled plate and the change in thermal resistance. For example, blockages hinder the flow of coolant, leading to a decrease in thermal convection efficiency and thus an increase in thermal resistance.

[0093] The deformation-thermal resistance coupling relationship refers to the coupling relationship between the deformation of the liquid-cooled cold plate and the abnormal thermal flow resistance of the liquid-cooled cold channel. That is, when the cold plate is deformed by thermal stress or mechanical pressure, it will cause microchannel collapse or poor fin contact, which will eventually lead to abnormal thermal resistance. For example, the warping of the liquid-cooled cold plate fins reduces the heat dissipation contact area, thereby increasing the thermal resistance.

[0094] Furthermore, the production process parameters of the liquid-cooled cold plate are obtained, which are represented by the welding current parameters, pressure parameters, and time parameters. The correlation and coupling relationship between the liquid-cooled cold plate defects and the production process are obtained by combining the liquid-cooled cold plate defect rules and the liquid-cooled cold plate fusion dataset.

[0095] Specifically, the associated coupling relationships include:

[0096] The coupling relationship between current parameter offset and liquid-cooled plate deformation.

[0097] This is used to characterize the situation where, when the welding current exceeds the design threshold, a sudden increase in local heat input leads to a mismatch in the thermal expansion coefficient of the material, causing irreversible plastic deformation. For example, for every 10 amperes that the current deviates from the standard value, the warpage of the liquid-cooled plate fins increases by 0.035 mm.

[0098] The correlation between pressure parameters and the flux loss in the internal cooling channels of the liquid-cooled plate is used to characterize the situation where insufficient welding pressure leads to uneven penetration, residual porosity or incomplete welding in the weld zone, forming a core area of ​​flux loss. For example, for every 0.1 MPa decrease in pressure, the risk of microchannel blockage increases by 35%.

[0099] The mapping relationship between the time parameter deviation value and the thermal flow resistance of the liquid-cooled plate is used to characterize the situation where insufficient welding time causes the alloy layer to not diffuse sufficiently, resulting in microscopic air gaps that block the heat conduction path. For example, for every 0.1 seconds the time is shortened, the thermal resistance increases by 0.08℃ / W.

[0100] Understandably, since the liquid cooling point cloud map, liquid cooling polarization feature map, liquid cooling flux distribution map and thermal resistance field map have relatively complex mathematical relationships, and the coupling influence between them cannot be accurately described using current mathematical parameters, only fuzzy algorithms can be used. Based on this, iterative calculations can be performed using artificial intelligence to determine the relationships between them more accurately.

[0101] Specifically, a fuzzy rule basis matrix is ​​constructed to associate the features of the liquid cooling point cloud map, liquid cooling polarization feature map, liquid cooling flux distribution map, and thermal resistance field map. For example, "moderate deformation and high flux loss lead to a larger increase in thermal resistance." An initial confidence level of 0.6±0.2 is assigned, and model iterative optimization begins. Each round inputs 1,000 sets of cold plate samples. The interaction strength between the corresponding features of the liquid cooling point cloud map, liquid cooling polarization feature map, liquid cooling flux distribution map, and thermal resistance field map is calculated through fuzzy inference. For example, if a sample with a deformation of 0.12 mm belongs to the medium-high intensity fuzzy set, its flux loss is 0.14, and the confidence level of the corresponding rule is increased to 0.82, the convolutional kernel weights and fully connected layer parameters of the feature extraction network are adjusted accordingly to enhance the response sensitivity to coupling modes. Training stops when the confidence level of the coupling rule fluctuates below ±0.05 in 20 consecutive iterations, such as when the confidence level of the polarization deformation coupling relationship stabilizes at 0.93±0.03, and the accuracy of the validation set exceeds 98%.

[0102] Step S3: Perform model distillation on the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line.

[0103] Specifically, process features are extracted based on the pre-trained liquid cooling defect detection model to obtain a process parameter correlation matrix. The process parameter correlation matrix represents the correlation parameter matrix between the liquid cooling plate production process parameters and the physical field anomalies of the liquid cooling plate. The physical field of the liquid cooling plate includes the surface polarization, internal flux, three-dimensional deformation, and thermal resistance field of the liquid cooling plate.

[0104] The process parameter correlation matrix of the teacher model and the student model is forced according to the mean square error, and multiple weights are applied to the key process defect samples for optimization and transfer. The key process defect samples are represented as process defect samples in the process parameter correlation matrix where the process parameters exceed the preset process threshold, and the physical field anomaly of the liquid cooling plate also exceeds the preset anomaly threshold.

[0105] The student model is trained based on the physical field inversion mechanism, which is represented by randomly shading the input channel corresponding to the physical field data of the liquid cooling plate. The student model is required to infer the missing features through the residual physical field data, obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line for real-time monitoring of liquid cooling plate defects.

[0106] Step S4: Real-time monitoring of the liquid cooling plate is performed based on the lightweight liquid cooling defect detection model to obtain the monitoring results of the liquid cooling plate, which include the liquid cooling plate defect coordinate diagram and process correlation report.

[0107] Understandably, the lightweight liquid cooling defect detection model applies pre-trained liquid cooling plate defect rules and associated coupling relationships to the real-time acquired liquid cooling plate fusion dataset for layer-by-layer diagnosis. In surface defect analysis, it scans the liquid cooling polarization feature map based on the polarization deformation coupling relationship, extracts the boundary point coordinates of regions where continuous scattering values ​​exceed a set threshold, and generates a surface defect coordinate set. In internal channel state analysis, it processes the liquid cooling flux distribution map by combining the flux-thermal resistance mapping relationship, captures the spatial grid coordinates of transmittance drop blocks, and generates a flux loss coordinate cluster. In the deformation and heat dissipation collaborative analysis stage, it simultaneously calls the deformation-thermal resistance coupling relationship to parse the liquid cooling point cloud map and the thermal flow resistance field map. When the point cloud elevation offset is associated with a region where thermal resistance exceeds the standard, a composite defect coordinate set is generated. Based on the surface defect coordinate set, the flux loss coordinate cluster, and the composite defect coordinate set, a liquid cooling plate defect coordinate map is constructed.

[0108] Understandably, the lightweight liquid-cooled defect detection model matches the surface defect coordinate set with the welding current time trajectory, applies the current deformation association rule to calculate the spatial overlap weight, and generates the current deformation association result. For example, the contribution of current exceeding the limit in a certain deformation zone is 85%. The spatial association between the flux loss coordinate cluster and the pressure parameter record is dynamically weighted according to the flux-pressure rule. If the average pressure in the coordinate cluster distribution area is lower than 10% of the preset standard value, then insufficient pressure is confirmed as the main cause of blockage, and the responsibility weight is marked, generating the pressure-flux loss association result. For example, in a certain Y-shaped channel blockage, the responsibility for insufficient pressure accounts for 76%. The welding time deviation analysis in the thermal resistance anomaly domain converts the influence coefficient through the thermal resistance-time rule and generates the time-thermal resistance association result. For example, the thermal resistance of a certain annular weld is 0.41℃ / W: the responsibility weight for welding time difference is 68%. Based on the pressure-flux loss association result, the pressure-flux loss association result and the time-thermal resistance association result, a process association report is generated.

[0109] Step S5: Verify the liquid cooling plate monitoring results according to the liquid cooling defect verification mechanism, obtain the defect verification results, and optimize the liquid cooling defect detection model based on the defect verification results.

[0110] In this embodiment, step S5 includes:

[0111] Step S51: Perform defect verification on the monitoring results of the liquid cooling plate according to the liquid cooling defect verification mechanism, and obtain the defect verification results.

[0112] Specifically, the surface defect coordinates are extracted from the defect coordinate map of the liquid-cooled cold plate, and the matching between polarization scattering intensity and deformation height is verified to obtain the polarization deformation matching result. For example, the matching rule is that for every 0.1 mm increase in scratch depth, the polarization value should increase by 80±5 units. There is a scratch with a depth of 0.15 mm on the surface of the cold plate. The polarization value corresponding to the scratch is 185, and the reference polarization value of the cold plate is 30. The matching degree is calculated to be (185-30) / (0.15*80)≈99.4%, and the matching result is a high match.

[0113] The coordinates of internal defects in the defect coordinate map of the liquid-cooled plate are extracted, and the quantitative relationship between flux loss rate and thermal resistance increase is verified. The flux-thermal resistance correlation matching results are obtained. For example, the correlation rule is that every 10% flux loss should lead to a thermal resistance increase of at least 0.12℃ / W. The flux at a certain point inside the liquid-cooled plate is 0.42, and the corresponding thermal resistance is 0.48℃ / W. The flux loss rate at this point relative to the reference flux is 58%, and the theoretical thermal resistance increment is 0.696℃ / W. The thermal resistance reference is 0.18℃ / W, so the corresponding thermal resistance increment is 0.30℃ / W. Since 0.30℃ / W does not reach the expected value of 0.696℃ / W, it is judged as a correlation failure.

[0114] The process parameters in the process correlation report are mapped to the corresponding defect coordinates, and the correlation parameters of the corresponding defect coordinates are verified to obtain the process defect verification results. For example, when the current exceeds 210 amperes, the deformation probability will reach 90%. If the current at a certain point is 235 amperes and a deformation of 0.15 mm is produced, it is determined to meet the rules.

[0115] The polarization deformation matching results, flux thermal resistance correlation matching results, and process defect verification results are encapsulated to generate defect verification results.

[0116] Step S52: Optimize the liquid cooling defect detection model based on the defect verification results.

[0117] Specifically, the defect verification results are analyzed and located to obtain the rules to be optimized corresponding to the liquid cooling defect detection model, the rules to be optimized are optimized, and a corresponding training set is generated to verify the optimization of the optimized rules.

[0118] For example, when the defect verification results show that a certain batch of cold plates has misjudged oil film reflection and missed minor blockages, the liquid cooling defect detection model is optimized accordingly. Interference rules are added, and oil film judgment logic is added. If the polarization value is greater than 150 and the deformation is less than 0.05mm, the surface texture entropy is detected simultaneously. If the match is met, the alarm is blocked. The sensitivity of flux thermal resistance correlation is adjusted, and the thermal resistance threshold with low cutoff rate is reduced, thereby improving the sensitivity of minor blockages.

[0119] This invention provides an image processing-based liquid-cooled cold plate defect detection system, applicable to liquid-cooled defect detection, comprising:

[0120] The data acquisition module is used to acquire multi-source data of the liquid-cooled plate, obtain the multi-source dataset of the liquid-cooled plate, and perform data preprocessing to generate a fused dataset of the liquid-cooled plate.

[0121] The model training module is used to construct a liquid cooling defect detection model and pre-train the liquid cooling defect detection model based on the liquid cooling plate fusion dataset.

[0122] The model distillation module is used to distill the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and then deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line.

[0123] The defect monitoring module is used to monitor the liquid cooling plate in real time according to the lightweight liquid cooling defect detection model and obtain the monitoring results of the liquid cooling plate.

[0124] The verification and optimization module is used to verify the defects in the monitoring results of the liquid-cooled cold plate, obtain the defect verification results, and optimize the liquid-cooled defect detection model based on the defect verification results.

[0125] The specific usage and function of this embodiment are as follows:

[0126] First, multi-source data is acquired for the liquid-cooled cold plate to obtain a multi-source dataset. The multi-source dataset is then preprocessed to generate a fused dataset. By simultaneously acquiring the surface polarization, internal microchannel flux, three-dimensional deformation, and thermal resistance of the liquid-cooled cold plate, and by performing coordinate space alignment and data feature enhancement, the influence of abnormal data is eliminated. This provides high-quality input for subsequent multi-physics coupling analysis to diagnose defects, thereby improving the accuracy of defect location and the timeliness of process traceability.

[0127] Next, a liquid cooling defect detection model is constructed. The fusion dataset of liquid cooling plates is input into the liquid cooling defect detection model for model pre-training. The pre-trained liquid cooling defect detection model is then distilled to obtain a lightweight liquid cooling defect detection model. This lightweight liquid cooling defect detection model is then deployed to the detection end of the liquid cooling plate production line. By learning the defect feature rules and process association rules of liquid cooling plates through model pre-training and combining them with lightweight pruning, multi-physics field data can be efficiently parsed at the edge device node, thereby improving the efficiency of liquid cooling plate defect detection.

[0128] Finally, the liquid-cooled cold plate is monitored in real time based on the lightweight liquid-cooling defect detection model to obtain the monitoring results. The monitoring results include the liquid-cooled cold plate defect coordinate map and process correlation report. The monitoring results are verified according to the liquid-cooling defect verification mechanism to obtain the defect verification results. Based on the defect verification results, the liquid-cooling defect detection model is optimized. The liquid-cooling cold plate defect rules and parameter thresholds are continuously calibrated through a closed-loop feedback optimization mechanism, thereby improving the accuracy and generalization ability of liquid-cooled cold plate defect diagnosis.

[0129] Furthermore, embodiments of the present invention also provide an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods described above.

[0130] The following is a detailed introduction to the various components of the electronic device:

[0131] In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

[0132] The processor can perform various functions of an electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0133] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0134] The memory can be a real-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through an interface circuit of an electronic device; this embodiment of the invention does not specifically limit this.

[0135] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0136] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0137] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0138] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting defects in liquid-cooled cold plates based on image processing, characterized in that, The method includes: S1: Perform multi-source data acquisition on the liquid-cooled plate to obtain a multi-source dataset of the liquid-cooled plate, and perform data preprocessing on the multi-source dataset of the liquid-cooled plate to generate a fused dataset of the liquid-cooled plate. The multi-source data acquisition refers to the real-time analysis of the reflective properties and internal structural features of the liquid-cooled plate surface, and the acquisition of data from the liquid-cooled plate according to the preset acquisition compensation rules to obtain a multi-source dataset of the liquid-cooled plate. The multi-source dataset of the liquid-cooled plate includes surface polarization feature data, internal structural data, three-dimensional deformation data, and thermal resistance distribution field data. Data preprocessing is represented by performing polarization difference operations and contrast enhancement on surface polarization feature data, eliminating metal reflection interference and extracting subsurface texture defects to generate liquid-cooled polarization feature maps. The internal structural data is subjected to spectral noise reduction and channel segmentation based on short-wave infrared transmittance, and the microchannel flux distribution characteristics are quantified to generate a liquid cooling flux distribution map. Perform point cloud statistical filtering and normal vector analysis on the three-dimensional deformation data to reconstruct the three-dimensional contour of the microchannel and calculate the depth deviation and fin warpage to generate a liquid cooling point cloud map. The thermal resistance distribution field data is generated by calibrating infrared thermograms based on material emissivity and combining them with the Fourier heat conduction equation to invert the spatial distribution of thermal resistance coefficients. Based on the liquid cooling point cloud map, the liquid cooling polarization feature map, liquid cooling flux distribution map and thermal resistance field map are mapped to the same feature space according to the affine transformation to generate a liquid cooling plate fusion dataset. S2: Construct a liquid cooling defect detection model and input the liquid cooling plate fusion dataset into the liquid cooling defect detection model for model pre-training; The model pre-training means obtaining defect-free normal liquid cooling plate samples and defective abnormal liquid cooling plate samples, and extracting structural features and defect features of liquid cooling plates according to conditional generative adversarial networks, and constructing a liquid cooling defect training set based on the structural features and defect features of liquid cooling plates. The liquid cooling defect detection model is pre-trained based on the liquid cooling defect training set to obtain the liquid cooling plate defect rules. The liquid cooling plate defect rules include polarization deformation coupling relationship, flux thermal resistance mapping relationship and deformation thermal resistance coupling relationship. The polarization deformation coupling relationship is expressed as the coupling relationship between the surface defect state of the liquid-cooled plate and the polarization scattering intensity. The flux-thermal resistance mapping relationship is expressed as the mapping relationship between the blockage cutoff rate caused by internal defects in the liquid-cooled plate and the change in thermal resistance. The deformation thermal resistance coupling relationship is expressed as the coupling relationship between the deformation of the liquid-cooled cold plate and the abnormal thermal flow resistance of the liquid-cooled cold channel. The production process parameters of the liquid-cooled cold plate are obtained, which are represented by the current parameters, pressure parameters and time parameters for welding. The liquid-cooled cold plate defect rules and the liquid-cooled cold plate fusion dataset are combined to perform the second stage of model pre-training for the liquid-cooled defect detection model, and to obtain the correlation and coupling relationship between liquid-cooled cold plate defects and production process. The associated coupling relationships include the coupling relationship between the current parameter offset and the deformation of the liquid-cooled plate, the correlation between the pressure parameter and the flow loss of the internal cold channel of the liquid-cooled plate, and the mapping relationship between the time parameter deviation and the thermal flow resistance of the liquid-cooled plate. S3: Perform model distillation on the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line; S4: Real-time monitoring of the liquid cooling plate is performed based on the lightweight liquid cooling defect detection model to obtain the monitoring results of the liquid cooling plate, which include a liquid cooling plate defect coordinate diagram and a process correlation report. S5: Verify the defects of the liquid cooling plate monitoring results according to the liquid cooling defect verification mechanism, obtain the defect verification results, and optimize the liquid cooling defect detection model based on the defect verification results.

2. The method for detecting defects in liquid-cooled cold plates based on image processing according to claim 1, characterized in that, The preset acquisition compensation rules include: Dynamically adjust the spatial pose of the optical sensor to collect data from the highly reflective area on the surface of the liquid-cooled plate; The internal structural data of the liquid-cooled plate was acquired by switching between short-wave infrared penetrating scanning and ultrasonic focusing detection combined modes.

3. The method for detecting defects in liquid-cooled cold plates based on image processing according to claim 1, characterized in that, The step of performing model distillation on the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and deploying the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line, includes: Based on the pre-trained liquid cooling defect detection model, process features are extracted to obtain the process parameter correlation matrix. The process parameter correlation matrix is ​​represented as the correlation parameter matrix between the liquid cooling plate production process parameters and the liquid cooling plate physical field anomaly. The physical field of the liquid-cooled plate includes the surface polarization, internal flux, three-dimensional deformation, and thermal resistance field of the liquid-cooled plate. The process parameter correlation matrix of the teacher model and the student model is forced according to the mean square error, and multiple weights are applied to the key process defect samples for optimization and transfer. The key process defect samples are represented as process defect samples in the process parameter correlation matrix where the process parameters exceed the preset process threshold, and the physical field anomaly of the liquid cooling plate also exceeds the preset anomaly threshold. The student model is trained based on the physical field inversion mechanism, which is represented by randomly shading the input channel corresponding to the physical field data of the liquid cooling plate. The student model is required to infer the missing features through the residual physical field data, obtain a lightweight liquid cooling defect detection model, and deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line for real-time monitoring of liquid cooling plate defects.

4. The method for detecting defects in liquid-cooled cold plates based on image processing according to claim 1, characterized in that, The method of real-time monitoring of the liquid-cooled cold plate based on the lightweight liquid-cooling defect detection model to obtain the monitoring results of the liquid-cooled cold plate includes: Based on the lightweight liquid cooling defect detection model, defect detection and process collaborative diagnosis are performed on the liquid cooling plates produced on the liquid cooling plate production line to generate liquid cooling plate monitoring results. The defect coordinate diagram of the liquid-cooled plate is composed of the spatial coordinates of the defects in the liquid-cooled plate. The process correlation report is a report on the correlation between defects in liquid-cooled cold plates and corresponding abnormalities in production process parameters.

5. The method for detecting defects in liquid-cooled cold plates based on image processing according to claim 1, characterized in that, The step of verifying defects in the monitoring results of the liquid-cooled cold plate according to the liquid-cooling defect verification mechanism and obtaining the defect verification results includes: The surface defect coordinates are extracted from the defect coordinate map of the liquid-cooled plate, and the matching between polarization scattering intensity and deformation height is verified to obtain the polarization deformation matching results. The coordinates of internal defects in the defect coordinate map of the liquid-cooled cold plate are extracted, and the quantitative relationship between flux loss rate and thermal resistance increase is verified to obtain the flux-thermal resistance correlation matching results. Map the process parameters in the process association report to the corresponding defect coordinates, verify the association parameters of the corresponding defect coordinates, and obtain the process defect verification results; The polarization deformation matching results, flux thermal resistance correlation matching results, and process defect verification results are encapsulated to generate defect verification results.

6. The method for detecting defects in liquid-cooled cold plates based on image processing according to claim 1, characterized in that, The optimization of the liquid cooling defect detection model based on the defect verification results includes: The defect verification results are analyzed and located to obtain the rules to be optimized corresponding to the liquid cooling defect detection model. The rules to be optimized are then optimized, and a corresponding training set is generated to verify the optimization of the rules.

7. A liquid-cooled cold plate defect detection system based on image processing, used to implement the method according to any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire multi-source data of the liquid-cooled plate, obtain the multi-source dataset of the liquid-cooled plate, and perform data preprocessing to generate a fused dataset of the liquid-cooled plate. The model training module is used to construct a liquid cooling defect detection model and pre-train the liquid cooling defect detection model based on the liquid cooling plate fusion dataset. The model distillation module is used to distill the pre-trained liquid cooling defect detection model to obtain a lightweight liquid cooling defect detection model, and then deploy the lightweight liquid cooling defect detection model to the detection end of the liquid cooling plate production line. The defect monitoring module is used to monitor the liquid cooling plate in real time according to the lightweight liquid cooling defect detection model and obtain the monitoring results of the liquid cooling plate. The verification and optimization module is used to verify the defects in the monitoring results of the liquid-cooled cold plate, obtain the defect verification results, and optimize the liquid-cooled defect detection model based on the defect verification results.

Images