Radiator surface temperature distribution detection method based on computer vision

Abstract

The invention relates to the technical field of computer vision and image sensing, and discloses a radiator surface temperature distribution detection method based on computer vision. The method comprises the following steps: collecting a multi-view visible light image sequence of the radiator in a working state; synchronously acquiring measured values of the contact temperature sensors at key positions on the surface; reconstructing a radiator three-dimensional geometric model based on the image; mapping the actually measured temperature into sparse observation points on the model; constructing a graph neural network fused with heat conduction physical prior, and iteratively solving whole-field temperature distribution by taking sparse observation as a boundary condition; and finally, carrying out back projection on the temperature field to generate a pseudo-color temperature map aligned with the visible light image pixel. The method comprises image acquisition, temperature sensing, three-dimensional modeling, temperature mapping, graph network modeling, full-field reconstruction and a visualization unit. According to the invention, through low-cost hardware and physically guided deep learning, high-precision, explainable and visual temperature distribution detection of the complex curved surface radiator is realized.

Description

Technical Field

[0002] This invention belongs to the field of computer vision and image sensing technology, specifically relating to a method for detecting the surface temperature distribution of a heat sink based on computer vision. Background Technology

[0003] With the rapid development of high-power electronic devices, new energy vehicles, and data centers, the accurate detection of surface temperature distribution in heat sinks, as key thermal management components, is crucial for system reliability and energy efficiency optimization. Traditional temperature monitoring methods mainly rely on infrared thermal imagers to capture infrared radiation from an object's surface and invert the temperature field. However, infrared thermal imaging technology is limited by high equipment costs, sensitivity to environmental reflection interference, and a strong dependence on material emissivity—especially on heat sink surfaces with complex geometries, multi-material splicing, or high reflectivity, where emissivity is difficult to accurately calibrate. This results in a significant deviation of the "apparent temperature" from the true temperature, significantly reducing measurement accuracy and engineering applicability.

[0004] Visual temperature measurement methods based on thermochromic materials have attracted attention in recent years. These methods utilize the physical property of material color changing with temperature, combined with optical imaging, to achieve non-contact temperature sensing. The core of this approach lies in establishing a mapping relationship between color response and real temperature, and overcoming the interference of factors such as ambient light, viewing angle changes, and material aging on color recognition. However, existing thermochromic schemes mostly use static color chart comparison or linear fitting models, which are difficult to handle nonlinear, multi-peak, or hysteretic temperature and color response behaviors, and lack the ability to reconstruct the spatial temperature field at high resolution, thus failing to meet the needs of accurate local hot spot location and dynamic heat flow analysis of radiators.

[0005] In existing technologies, infrared thermal imagers are difficult to popularize in small and medium-scale applications due to cost and emissivity correction challenges. On the other hand, while ordinary RGB cameras offer advantages in low cost and high resolution, they lack a reliable physical-visual coupling mechanism, making it impossible to directly output quantitative temperature distribution. Especially when the heatsink surface has complex conditions such as curvature, shading, oxidation, or uneven coating, traditional visual methods are prone to failure due to color distortion or lighting disturbances. Therefore, there is an urgent need for a novel temperature detection method that integrates intelligent responsive materials and deep learning decoding mechanisms. This method should avoid the inherent defects of infrared temperature measurement while overcoming the accuracy and robustness bottlenecks of traditional visual temperature measurement, achieving low-cost, high-precision, and adaptive reconstruction of the true temperature field of the heatsink surface. Summary of the Invention

[0006] This invention provides a computer vision-based method for detecting surface temperature distribution on heat sinks, aiming to solve the technical problems of traditional infrared thermal imagers being expensive, only able to acquire apparent temperature, and difficult to adapt to complex geometric surface structures. This invention constructs a low-cost, highly adaptable, and physically interpretable temperature field reconstruction mechanism by fusing visible light image information with limited-point contact temperature sensing data, achieving continuous and accurate visualization of temperature distribution on any complex surface of a heat sink.

[0007] According to one aspect of the present invention, a computer vision-based method for detecting surface temperature distribution of a heat sink is provided, comprising: acquiring a visible light image sequence of the heat sink in its working state; simultaneously acquiring measured temperature values ​​output by contact temperature sensors disposed at several key locations on the surface of the heat sink; generating a surface geometric model of the heat sink based on the visible light image sequence using a three-dimensional reconstruction algorithm; mapping the measured temperature values ​​to corresponding physical coordinate points in the surface geometric model to form a sparse temperature observation set; constructing a heat diffusion-guided graph neural network model based on the topological structure and thermal conduction physical priors of the surface geometric model; inputting the sparse temperature observation set as boundary conditions into the graph neural network model, and generating a continuous temperature distribution field covering the entire surface geometric model by iteratively solving the discrete form of the thermal conduction partial differential equation; and back-projecting the continuous temperature distribution field onto the original visible light image plane to generate a pseudo-color temperature distribution image aligned with the pixels of the visible light image.

[0008] Furthermore, the acquisition of visible light image sequences of the heat sink in its working state specifically includes: using at least two spatially fixed industrial-grade visible light cameras to simultaneously capture multiple frames of images of the heat sink in a stable thermal working state from different perspectives; the industrial-grade visible light cameras have a resolution of not less than 1920×1080 pixels, a frame rate of not less than 30 Hz, a lens focal length adjustable from 8 mm to 50 mm, and a photosensitive element that is a global shutter type complementary metal-oxide-semiconductor image sensor.

[0009] Preferably, the synchronous acquisition of the measured temperature values ​​output by contact temperature sensors located at several key positions on the surface of the heat sink specifically includes: arranging platinum resistance temperature sensors in the heat source concentration area, airflow inlet area, airflow outlet area, and geometric abrupt change area of ​​the heat sink; the platinum resistance temperature sensors conform to the industrial Class B accuracy standard, with a temperature measurement range of -50 degrees Celsius to +250 degrees Celsius and a response time of less than 1 second; all platinum resistance temperature sensors are connected to a multi-channel data acquisition card via shielded twisted-pair cables, the sampling frequency is set to 10 Hz, and hardware-level time synchronization is achieved with the trigger signal of the visible light camera.

[0010] Preferably, the step of generating the surface geometric model of the heat sink based on the visible light image sequence using a 3D reconstruction algorithm specifically includes: extracting and matching feature points from the visible light image sequence, and calculating the relative pose of the camera using a five-point algorithm; generating a sparse point cloud based on the camera pose and image matching points using a motion recovery structure algorithm; reconstructing the sparse point cloud using a Poisson surface to generate a triangular mesh model with closed watertightness; and performing Laplacian smoothing and normal consistency correction on the triangular mesh model to obtain the final surface geometric model; the number of vertices in the surface geometric model is controlled between 100,000 and 500,000, and the average size of the facets is no greater than 2 mm.

[0011] Preferably, mapping the measured temperature value to the corresponding physical coordinate point in the surface geometric model specifically includes: establishing a rigid body transformation matrix between the camera coordinate system and the world coordinate system using a hand-eye calibration method; using the rigid body transformation matrix to back-project the two-dimensional pixel coordinates of each contact temperature sensor in the image to three-dimensional space; searching for the grid vertex closest to the point in the three-dimensional space in the surface geometric model, marking the vertex as a temperature observation point, and assigning it a corresponding measured temperature value, thereby forming a sparse temperature observation set; the sparse temperature observation set contains no less than 5 observation points and uniformly covers the main thermal functional area of ​​the radiator.

[0012] Preferably, the construction of a heat diffusion-guided graph neural network model based on the topological structure and thermal conduction physics priors of the surface geometry model specifically includes: converting the triangular mesh of the surface geometry model into an undirected graph structure, where each mesh vertex is a graph node and each mesh edge is a graph edge; assigning a weight to each graph edge, the weight being determined by the Euclidean distance between two adjacent nodes and the local curvature, calculated using the following formula: ,in For nodes and The Euclidean distance between them The distance attenuation factor is set to 5 millimeters. For the edge The corresponding local mean curvature, The maximum curvature value of the entire model; a heat conduction operator is defined on the graph structure, its discrete form being: ,in It is an adjacency matrix. The degree matrix is ​​used; the graph neural network model contains 3 graph convolutional layers, and the propagation rule for each layer is as follows: ,in Let l be the node feature matrix of the l-th layer. For learnable parameter matrix, To modify the activation function of the linear unit.

[0013] Preferably, the step of inputting the sparse temperature observation set as a boundary condition into the graph neural network model and generating a continuous temperature distribution field covering the entire surface geometry model by iteratively solving the discrete form of the heat conduction partial differential equation specifically includes: initializing the temperature values ​​of all non-observation nodes to the ambient temperature; fixing the temperature values ​​of the observation nodes to their corresponding measured temperature values ​​and forcibly covering them after each forward propagation of the graph neural network; executing the forward propagation process of the graph neural network, with the number of iterations set to 20; in each iteration, the temperature values ​​of the non-observation nodes are updated according to the weighted average of their neighboring nodes, with the weights determined by the graph edge weights guided by the aforementioned heat diffusion; after the iteration is completed, each grid vertex is assigned a definite temperature value, forming a continuous temperature distribution field.

[0014] Preferably, the step of back-projecting the continuous temperature distribution field onto the original visible light image plane to generate a temperature distribution pseudo-color image aligned with the visible light image pixels specifically includes: selecting a visible light image from any viewpoint as a reference image; for each pixel in the reference image, projecting light onto the surface geometry model using the camera intrinsic parameter matrix, and solving for the intersection point of the light ray and the surface geometry model; if an intersection point exists, querying the temperature values ​​of the three vertices of the triangular facet where the intersection point is located, and obtaining the temperature value of the intersection point through centroid coordinate interpolation; if no intersection point exists, marking the temperature value of the pixel point as invalid; converting the temperature values ​​of all valid pixels into red, green, and blue three-channel color values ​​according to a preset color temperature mapping table, generating a temperature distribution pseudo-color image completely aligned with the original visible light image; the temperature range of the color temperature mapping table covers 40 degrees Celsius to 120 degrees Celsius, using a linearly interpolated rainbow color spectrum.

[0015] Preferably, a computer vision-based radiator surface temperature distribution detection system is provided, comprising: a multi-view visible light image acquisition unit for acquiring a sequence of visible light images of the radiator in operation; a distributed contact temperature sensing unit for synchronously acquiring measured temperature values ​​at several key locations on the radiator surface; a three-dimensional geometric modeling unit for generating a surface geometric model of the radiator based on the visible light image sequence; a temperature observation mapping unit for mapping the measured temperature values ​​to corresponding physical coordinate points in the surface geometric model to form a sparse temperature observation set; a thermal physics graph neural network modeling unit for constructing a heat diffusion-guided graph neural network model based on the topology and thermal conduction physics priors of the surface geometric model; a full-field temperature reconstruction unit for inputting the sparse temperature observation set as boundary conditions into the graph neural network model to generate a continuous temperature distribution field covering the entire surface geometric model; and a temperature field visualization unit for back-projecting the continuous temperature distribution field onto the original visible light image plane to generate a pseudo-color temperature distribution image aligned with the visible light image pixels.

[0016] Preferably, the multi-view visible light image acquisition unit consists of two or more industrial-grade visible light cameras. Each camera is connected to a central synchronization controller via a hard-trigger cable to ensure that the image acquisition time is strictly consistent. The industrial-grade visible light camera is equipped with a fixed-focus lens with an adjustable aperture range of F2.8 to F16, supports manual exposure control, and has an exposure time setting range of 1 millisecond to 100 milliseconds.

[0017] Preferably, the distributed contact temperature sensing unit includes no fewer than five platinum resistance temperature sensors, a multi-channel constant current source excitation circuit, a signal conditioning and amplification circuit, and a high-precision analog-to-digital converter; the platinum resistance temperature sensors are tightly attached to the heat sink surface with thermally conductive silicone grease and encapsulated and fixed with high-temperature resistant epoxy resin; the multi-channel constant current source excitation circuit provides a constant excitation current of 1 mA; the gain of the signal conditioning and amplification circuit is adjustable from 1 to 100 times, and the common-mode rejection ratio is not less than 80 dB; the high-precision analog-to-digital converter has a resolution of 24 bits and a sampling rate of not less than 1 kHz.

[0018] Preferably, the three-dimensional geometric modeling unit runs on a computing platform accelerated by a graphics processing unit, uses an open-source computer vision library to perform feature extraction and matching, uses a sparse bundleadjustment algorithm to optimize camera pose, uses a Poisson surface reconstruction algorithm to generate an initial mesh, and uses an iterative nearest-point algorithm to enhance the details of the mesh model.

[0019] Preferably, the thermal physical graph neural network modeling unit is implemented using a message-passing neural network framework, wherein the node feature dimension is 1, the edge weight calculation module is integrated into the message function, the aggregation function adopts a weighted summation method, and the update function is a linear transformation with a bias term followed by a modified linear unit activation; all parameters of the graph neural network model are pre-trained using a simulation dataset before deployment, and the simulation dataset is generated by finite element thermal simulation software and contains no less than 1000 sets of temperature field data under different radiator structures and different thermal boundary conditions.

[0020] Preferably, when the full-field temperature reconstruction unit performs temperature field reconstruction, it adopts an iterative solution strategy with fixed boundary conditions. That is, after each propagation of the graph neural network, the temperature value of the observed node is forcibly reset to the measured value to ensure that the physical constraints are strictly satisfied. The iteration termination condition is that the maximum number of iterations reaches 20 or the L2 norm of the full-field temperature change is less than 0.1 degrees Celsius.

[0021] Preferably, the temperature field visualization unit supports multi-view temperature field rendering, allowing users to select any acquisition viewpoint to generate a corresponding pseudo-color image; it also supports the generation of dynamic temperature evolution videos, by stitching together consecutive frames of pseudo-color images in chronological order, with the frame rate consistent with the original visible light image acquisition frame rate.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention abandons expensive infrared thermal imaging hardware and instead adopts a hybrid sensing architecture that combines a low-cost visible light camera array with a small number of high-precision contact temperature sensors, which significantly reduces the system hardware cost.

[0023] By introducing a three-dimensional reconstruction model based on the actual geometry of the radiator and using it as a carrier of thermophysical processes, this invention overcomes the shortcomings of traditional methods that cannot adapt to complex curved surfaces, and achieves accurate reconstruction of the temperature field on the surface of a radiator with arbitrary geometric shape.

[0024] The thermal diffusion guided graph neural network constructed in this invention explicitly encodes the physical laws of heat conduction into the network structure and message passing mechanism, making the temperature field reconstruction process have clear physical interpretability and avoiding the shortcomings of pure data-driven models in terms of physical consistency.

[0025] By embedding sparse measured temperature points as strict boundary conditions into the reconstruction process, the absolute accuracy of the reconstruction results at key locations is ensured. Simultaneously, the smoothness of heat conduction is used as a priori to reasonably infer the temperature distribution in unobserved areas. The resulting pseudo-color temperature distribution image is strictly pixel-aligned with the visible light image, allowing engineers to intuitively correlate abnormal temperature areas with specific physical structural features, greatly improving the efficiency and accuracy of radiator thermal design verification and fault diagnosis. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the heat diffusion-guided graph neural network model in this invention; Figure 3 This is a logical flowchart of the multi-view visible light image acquisition and three-dimensional geometric modeling in this invention; Figure 4 This is a flowchart illustrating the logical process of constructing and mapping sparse temperature observation sets in this invention. Figure 5 This is a flowchart illustrating the logical process of full-field temperature reconstruction and boundary condition iterative solution in this invention. Figure 6 This is a flowchart illustrating the logical process of generating temperature distribution field back projection and pseudo-color visualization in this invention. Detailed Implementation

[0027] This invention provides a computer vision-based method for detecting the surface temperature distribution of a heat sink. Its core lies in fusing visible light image information with limited-point contact temperature sensing data to construct a physically interpretable full-field temperature reconstruction mechanism. The method achieves continuous, accurate, and low-cost temperature distribution detection on the surface of heat sinks with arbitrary complex geometries through six main stages: multi-view visible light image acquisition, 3D geometric modeling, sparse temperature observation mapping, heat diffusion guided graph neural network modeling, iterative solution of the full-field temperature field, and temperature field visualization. The following will be combined with the appendix... Figure 1 To be continued Figure 6 This section provides a detailed implementation description of each functional module of the system, expanding upon it layer by layer.

[0028] The method first performs step S1: acquiring a sequence of visible light images of the heat sink in its operating state. This step is performed by at least two spatially fixed industrial-grade visible light cameras, each simultaneously capturing images of the heat sink in a stable thermal operating state from different perspectives. The industrial-grade visible light cameras employ global shutter type complementary metal-oxide-semiconductor image sensors with a resolution of at least 1920×1080 pixels, a frame rate of at least 30 Hz, and a lens focal length adjustable from 8 mm to 50 mm. All cameras are connected to a central synchronization controller via hard-trigger cables to ensure strict consistency in the acquisition time of each frame. Before image acquisition, a constant thermal load must be applied to the heat sink and it must be allowed to reach a thermal steady state, typically requiring continuous operation for at least 10 minutes to ensure the surface temperature distribution stabilizes. During acquisition, ambient lighting conditions are kept constant to avoid interference from strong direct light sources causing highlights or shadows. The exposure time for each camera is set between 1 ms and 100 ms, and the aperture value is manually adjusted between F2.8 and F16 to obtain clear images without motion blur. The acquired image sequences are stored in an uncompressed format on a high-speed solid-state storage device for subsequent 3D reconstruction.

[0029] Then, step S2 is executed: The measured temperature values ​​output by contact temperature sensors located at several key positions on the surface of the heat sink are acquired synchronously. Platinum resistance temperature sensors are arranged in the heat source concentration area, airflow inlet area, airflow outlet area, and areas of geometric abrupt changes in the heat sink, with a total number of no fewer than five. Each platinum resistance temperature sensor meets industrial Class B accuracy standards, with a temperature range of -50°C to +250°C and a response time of less than 1 second. The sensors are tightly attached to the heat sink surface using thermally conductive silicone grease and encapsulated and fixed using high-temperature resistant epoxy resin to ensure good thermal contact and long-term stability. All sensors are connected to a multi-channel data acquisition system via shielded twisted-pair cables. This system includes a multi-channel constant current source excitation circuit, a signal conditioning and amplification circuit, and a high-precision analog-to-digital converter. The constant current source provides a constant excitation current of 1 mA, the signal conditioning and amplification circuit has an adjustable gain range of 1 to 100 times, a common-mode rejection ratio of no less than 80 dB, and the analog-to-digital converter has a resolution of 24 bits and a sampling rate of no less than 1 kHz. The data acquisition system shares the same hardware trigger signal with the visible light camera to achieve time synchronization. The sampling frequency is set to 10 Hz to ensure that each frame of visible light image corresponds to a set of precisely synchronized temperature measurements. The acquired raw voltage signal is linearized and converted into temperature values, and the corresponding timestamp is recorded.

[0030] Next, step S3 is executed: Based on the visible light image sequence, a surface geometric model of the heat sink is generated using a 3D reconstruction algorithm. This process first preprocesses the images from all viewpoints with denoising and contrast enhancement. Then, a scale-invariant feature transform algorithm is used to extract feature points from each image, and feature matching is performed using the nearest neighbor distance ratio criterion. Successfully matched keypoint pairs are used to estimate the relative pose between cameras. A five-point algorithm combined with a random sampling consensus algorithm is used to eliminate false matches, and the initial camera pose is calculated. Based on this, a sparse 3D point cloud containing thousands to tens of thousands of spatial points is generated using a motion recovery structure algorithm, covering the main visible surface of the heat sink.

[0031] Subsequently, incremental bundle adjustment was used to jointly optimize the camera pose and point cloud coordinates, minimizing reprojection errors. The optimized sparse point cloud was used as input, and a Poisson surface reconstruction algorithm was employed to generate an initial triangular mesh model. This algorithm recovers surface details while maintaining watertightness by solving the Poisson equation for the indicator function. The generated initial mesh was further smoothed using Laplacian smoothing to eliminate high-frequency noise, and normal consistency correction was used to ensure that the normal directions of all facet elements were aligned. The final surface geometry model had between 100,000 and 500,000 vertices, with an average facet size of no more than 2 mm, sufficient to characterize complex geometric features such as heat sink fins, substrates, and holes. The entire 3D reconstruction process ran on a graphics processing unit-accelerated computing platform, utilizing open-source computer vision libraries for efficient parallel computation.

[0032] The S4 step then proceeds: mapping the measured temperature values ​​to the corresponding physical coordinates in the surface geometry model to form a sparse temperature observation set. This mapping process relies on precise hand-eye calibration. First, a world coordinate system is established, typically with the radiator mounting base as a reference. Then, the intrinsic parameters of each camera are calibrated using a checkerboard calibration plate to obtain parameters such as focal length, principal point, and distortion coefficient. Next, using the known 3D coordinates of the calibration points and their corresponding pixels in the image, the rigid body transformation matrix from the camera coordinate system to the world coordinate system is calculated. For each contact temperature sensor, its 2D pixel coordinates in a certain viewpoint image can be determined through template matching or manual annotation. Using these pixel coordinates, the camera intrinsic parameter matrix, and the rigid body transformation matrix, the 3D spatial position of the sensor in the world coordinate system is calculated through back projection.

[0033] Subsequently, the grid vertex closest to the given 3D point in the surface geometry model via Euclidean distance is searched. This vertex is marked as a temperature observation point and assigned a corresponding measured temperature value. If multiple sensors are mapped to the same vertex, the weighted average of the measured temperatures is taken, with the weight determined by the reciprocal of the distance from the sensor to the vertex. The resulting sparse temperature observation set contains at least five temperature-labeled grid vertices and uniformly covers the main thermal functional areas of the heat sink, including the heat source area, cooling area, and structural transition area.

[0034] Next, step S5 is executed: Based on the topological structure of the surface geometry model and the prior knowledge of heat conduction physics, a heat diffusion-guided graph neural network model is constructed. This model transforms the triangular mesh of the surface geometry model into an undirected graph structure, where each mesh vertex is a graph node and each mesh edge is a graph edge. The adjacency relationships of the graph are entirely determined by the mesh topology. A weight is assigned to each graph edge, reflecting the heat conduction efficiency, which is determined by the Euclidean distance between two adjacent nodes and the local curvature. The weight calculation formula is: in For nodes and The Euclidean distance between them The distance attenuation factor is set to 5 millimeters. For the edge The corresponding local mean curvature, This represents the maximum curvature value of the entire model. This weighting design reflects two physical properties of heat conduction: heat is more easily conducted along short paths, and it diffuses more easily in flat regions (low curvature) than at sharp edges (high curvature). Based on this weighted graph, a heat dissipation conduction operator is defined. ,in For an adjacency matrix, the elements are... , For a degree matrix, the diagonal elements The graph neural network model contains three graph convolutional layers, and the node feature propagation rule for each layer is as follows: in, Let be the node feature matrix of the l-th layer, initially... It is a column vector with dimensions equal to the number of nodes, and its elements are the initial node temperatures. The learnable parameter matrix has a dimension of 1×1; To correct the activation function of the linear unit, since the temperature field is a scalar field, the node feature dimension is always 1. Before deployment, the model is pre-trained using a simulation dataset generated by finite element thermal simulation software. This dataset contains no fewer than 1000 sets of steady-state temperature field data under different radiator structures and thermal boundary conditions to ensure the model has generalization ability.

[0035] Then, step S6 is executed: the sparse temperature observation set is input as a boundary condition into the graph neural network model. The discrete form of the heat conduction partial differential equation is solved iteratively to generate a continuous temperature distribution field covering the entire surface geometry model. During initialization, the temperature values ​​of all non-observed nodes are set to the ambient temperature, typically 25 degrees Celsius; the temperature values ​​of observed nodes are set to their corresponding measured temperatures. The forward propagation process of the graph neural network is executed, with 20 iterations. In each iteration, the new temperature value of each node is first calculated as a weighted average of the temperatures of its neighboring nodes, with the weights determined by the graph edge weights guided by the aforementioned heat diffusion. This is then multiplied by a learnable parameter. The process involves using an activation function. Crucially, after each forward propagation, the temperature values ​​of the observed nodes are forcibly reset to their original measured values ​​to ensure strict satisfaction of the physical boundary conditions. This operation constitutes an iterative solution strategy with fixed boundary conditions. The iteration terminates when the maximum number of iterations (20) is reached, or when the L2 norm of the overall temperature change is less than 0.1 degrees Celsius. After the iteration ends, each grid vertex is assigned a defined temperature value, forming a continuous, smooth, and physically consistent overall temperature distribution.

[0036] Finally, step S7 is executed: the continuous temperature distribution field is back-projected onto the original visible light image plane to generate a pseudo-color temperature distribution image aligned with the pixels of the visible light image. A visible light image from any viewpoint is selected as the reference image. For each pixel in this image, a ray originating from the camera center and passing through the pixel is constructed using the camera intrinsic parameter matrix. The intersection of this ray with the surface geometry model is calculated using an accelerated ray-triangle intersection test algorithm. If an intersection point exists, the triangular facet containing that intersection point is determined, and the temperature values ​​of the facet's three vertices are obtained. The temperature value at the intersection point is calculated using barycentric coordinate interpolation, with the interpolation formula being... in , , Let the coordinates be the centroid coordinates, satisfying If the ray does not intersect the model, the temperature value of that pixel is marked as invalid and filled with black in subsequent visualizations. The temperature values ​​of all valid pixels are converted to red, green, and blue three-channel color values ​​according to a preset color temperature mapping table. The color temperature mapping table covers 40 degrees Celsius to 120 degrees Celsius and uses a linear interpolation rainbow spectrum: 40 degrees Celsius corresponds to dark blue, 60 degrees Celsius to cyan, 80 degrees Celsius to green, 100 degrees Celsius to yellow, and 120 degrees Celsius to red. The generated pseudo-color image has the exact same resolution and pixel layout as the original visible light image, achieving strict alignment. Users can choose any acquisition angle to generate the corresponding pseudo-color image, or they can stitch consecutive frames in chronological order to generate a dynamic temperature evolution video, with the frame rate consistent with the original image acquisition frame rate.

[0037] The implementation of the above method relies on a complete detection system. This system includes a multi-view visible light image acquisition unit, a distributed contact temperature sensing unit, a 3D geometric modeling unit, a temperature observation mapping unit, a thermophysical map neural network modeling unit, a full-field temperature reconstruction unit, and a temperature field visualization unit. The multi-view visible light image acquisition unit consists of two or more industrial-grade visible light cameras, synchronized via hard triggering. The distributed contact temperature sensing unit includes a platinum resistance temperature sensor array, a constant current source, signal conditioning circuitry, and a high-precision analog-to-digital converter, ensuring the accuracy and temporal synchronization of temperature data. The 3D geometric modeling unit runs on a graphics processing unit acceleration platform, integrating feature extraction, pose estimation, point cloud generation, and mesh reconstruction modules. The temperature observation mapping unit achieves precise registration of sensor positions to the 3D model. The thermophysical map neural network modeling unit uses a message-passing neural network framework, embedding the physical laws of heat conduction into the network structure. The full-field temperature reconstruction unit performs iterative solutions with fixed boundary conditions. The temperature field visualization unit supports multi-view rendering and dynamic video generation. The entire system coordinates the work of each unit through central control software, achieving full automation from data acquisition to result output.

Claims

1. A method for detecting surface temperature distribution of a heat sink based on computer vision, characterized in that, include: Acquire visible light image sequences of the heat sink in its working state; The measured temperature values ​​output by contact temperature sensors located at several key positions on the surface of the heat sink are acquired simultaneously. Based on the visible light image sequence, a surface geometric model of the heat sink is generated using a three-dimensional reconstruction algorithm; The measured temperature values ​​are mapped to the corresponding physical coordinate points in the surface geometric model to form a sparse temperature observation set. Based on the topological structure and thermal conduction physics priors of the surface geometry model, a graph neural network model guided by thermal diffusion is constructed. The sparse temperature observation set is input into the graph neural network model as a boundary condition. By iteratively solving the discrete form of the heat conduction partial differential equation, a continuous temperature distribution field covering the entire surface geometry model is generated. The continuous temperature distribution field is back-projected onto the original visible light image plane to generate a pseudo-color temperature distribution image aligned with the pixels of the visible light image.

2. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 1, characterized in that, The sequence of visible light images of the heat sink in its working state includes: Use at least two spatially fixed industrial-grade visible light cameras to simultaneously capture multiple frames of images of the heat sink under stable thermal operating conditions from different perspectives; The industrial-grade visible light camera has a resolution of no less than 1920×1080 pixels, a frame rate of no less than 30 Hz, a lens focal length adjustable from 8 mm to 50 mm, and a global shutter type complementary metal-oxide-semiconductor image sensor.

3. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 2, characterized in that, The synchronous acquisition of measured temperature values ​​output by contact temperature sensors located at several key positions on the surface of the heat sink includes: Platinum resistance temperature sensors are placed in the heat source concentration area, airflow inlet area, airflow outlet area and geometric abrupt change area of ​​the radiator respectively. The platinum resistance temperature sensor conforms to the industrial Class B accuracy standard, with a temperature measurement range of -50 degrees Celsius to +250 degrees Celsius and a response time of less than 1 second. All platinum resistance temperature sensors are connected to a multi-channel data acquisition card via shielded twisted-pair cables. The sampling frequency is set to 10 Hz, and hardware-level time synchronization is achieved with the trigger signal of the visible light camera.

4. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 3, characterized in that, The step of generating a surface geometric model of the heat sink based on the visible light image sequence using a three-dimensional reconstruction algorithm includes: Feature points are extracted and matched from the visible light image sequence, and the relative pose of the camera is calculated using a five-point algorithm. Based on the camera pose and image matching points, a sparse point cloud is generated using the structure-reconstruction-motion algorithm. Poisson surface reconstruction is performed on the sparse point cloud to generate a triangular mesh model with closed watertightness. The triangular mesh model is smoothed using Laplacian smoothing and normal consistency correction to obtain the final surface geometry model. The number of vertices in the surface geometry model is controlled between 100,000 and 500,000, and the average size of the surface element is no greater than 2 millimeters.

5. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 4, characterized in that, The step of mapping the measured temperature value to the corresponding physical coordinate point in the surface geometric model includes: A rigid body transformation matrix between the camera coordinate system and the world coordinate system is established using a hand-eye calibration method. The rigid body transformation matrix is ​​used to back-project the two-dimensional pixel coordinates of each contact temperature sensor in the image to three-dimensional space. Search for the mesh vertex closest to the three-dimensional space point in the surface geometry model, mark the vertex as a temperature observation point, and assign it the corresponding measured temperature value to form a sparse temperature observation set. The sparse temperature observation set contains no fewer than 5 observation points and evenly covers the main thermal functional area of ​​the radiator.

6. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 5, characterized in that, The construction of a heat diffusion-guided graph neural network model based on the topological structure and prior knowledge of heat conduction of the surface geometry model includes: The triangular mesh of the surface geometry model is transformed into an undirected graph structure, where each mesh vertex is a graph node and each mesh edge is a graph edge; Each graph edge is assigned a weight, which is determined by the Euclidean distance between two adjacent nodes and the local curvature, calculated using the following formula: ,in For nodes and The Euclidean distance between them The distance attenuation factor is set to 5 millimeters. For the edge The corresponding local mean curvature, This represents the maximum curvature value of the entire model; A heat conduction operator is defined on the graph structure, and its discrete form is: ,in It is an adjacency matrix. It is a degree matrix; The graph neural network model contains three graph convolutional layers, and the propagation rule for each layer is as follows: ,in Let l be the node feature matrix of the l-th layer. For learnable parameter matrix, To modify the activation function of the linear unit.

7. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 6, characterized in that, The step of inputting the sparse temperature observation set as boundary conditions into the graph neural network model, and generating a continuous temperature distribution field covering the entire surface geometry model by iteratively solving the discrete form of the heat conduction partial differential equation, includes: Initialize the temperature values ​​of all non-observable nodes to the ambient temperature. The temperature value of the observed node is fixed to its corresponding measured temperature value, and this value is forcibly overwritten after each forward propagation of the graph neural network. The forward propagation process of the graph neural network is executed, with the number of iterations set to 20. In each iteration, the temperature value of the unobserved node is updated based on the weighted average of its neighboring nodes, with the weights determined by the graph edge weights guided by the aforementioned heat diffusion. After the iteration, each grid vertex is assigned a specific temperature value, forming a continuous temperature distribution field.

8. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 7, characterized in that, The step of back-projecting the continuous temperature distribution field onto the original visible light image plane to generate a pseudo-color temperature distribution image aligned with the pixels of the visible light image includes: A visible light image from any viewpoint is selected as the reference image; For each pixel in the reference image, a ray is projected onto the surface geometry model using the camera intrinsic parameter matrix, and the intersection point of the ray and the surface geometry model is solved. If an intersection point exists, query the temperature values ​​of the three vertices of the triangular facet containing the intersection point, and obtain the temperature value of the intersection point by interpolation using the centroid coordinates; If there is no intersection point, the temperature value of that pixel is marked as invalid; The temperature values ​​of all valid pixels are converted into red, green and blue three-channel color values ​​according to the preset color temperature mapping table to generate a temperature distribution pseudo-color image that is completely aligned with the original visible light image; The color temperature mapping table covers a temperature range of 40 degrees Celsius to 120 degrees Celsius and uses a linear interpolation rainbow chromatogram.

9. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 1, characterized in that, The heat diffusion-guided graph neural network model is pre-trained using a simulation dataset before deployment. The simulation dataset is generated by finite element thermal simulation software and contains no less than 1,000 sets of temperature field data under different radiator structures and different thermal boundary conditions.

10. The method for detecting heat sink surface temperature distribution based on computer vision according to claim 7, characterized in that, The iteration termination condition is that the maximum number of iterations reaches 20 or the L2 norm of the overall temperature change is less than 0.1 degrees Celsius.

Images