Performance test method for PCB of AI server

Abstract

The invention belongs to the technical field of PCB testing, and discloses a performance testing method of an AI server PCB. Comprising the following steps: acquiring power-on operation temperature data of the PCB to construct a temperature distribution image, determining hot spot region characteristics based on the temperature distribution image to obtain heating test parameters, performing heating test on the PCB based on the heating test parameters, acquiring thermal resistance of each position of a hot spot region, drawing a thermal resistance characteristic curve, and analyzing the thermal resistance characteristic curve to obtain a thermal performance grade of the PCB. The bending rate of the bending test is determined according to the basic parameters of the PCB, the bending force is adjusted in real time according to the bending rate during the test, the test is stopped when the stress or the strain reaches a threshold value, the test data is analyzed to obtain the bending resistance grade of the PCB, and the PCB is divided into different grades by combining the thermal performance grade. According to the method, the efficiency and reliability of PCB testing are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of PCB board testing. More specifically, the present invention relates to a method for testing the performance of a PCB board of an AI server. Background Art

[0002] With the increasing demand for high-performance PCB boards in artificial intelligence servers, the thermal performance and anti-bending performance of PCB boards have gradually become key indicators of the reliability of PCB boards.

[0003] The patent application with the application publication number CN118858902A discloses a PCB board testing system, including a test acquisition module, a data analysis module, and an early warning beeping module. The present invention tests the board parts and components, screens out qualified board parts and components, prevents defective board parts and components used for welding from affecting the quality of the PCB, and determines the reasons for the unqualified board parts and components; detects the positioning of the board parts and the positioning of the components on the board parts, determines whether the components are accurately welded at the specified positions, and prevents the misalignment of component welding from affecting the quality of the PCB; detects the quality of the welding fluid, reduces the influence of the quality of the welding fluid on the welding result, and adjusts it in time before the welding fluid solidifies when the detection is unqualified, reducing losses.

[0004] However, the existing PCB board performance tests mostly focus on whether the components of the PCB board are normal, ignoring the thermal performance and anti-bending performance tests; in terms of thermal performance tests, traditional high-precision infrared thermal imagers are difficult to accurately obtain the global temperature distribution due to environmental interference, spatial resolution, and component occlusion problems; the resistance calculation technology is limited by the space of high-integration PCB boards, resulting in insufficient sensor coverage and the problem of easy distortion in temperature field reconstruction; and the traditional heating test mode is fixed, the thermal performance evaluation index is single, and the subjectivity is strong, and it is impossible to better evaluate the thermal performance of the PCB board; in the anti-bending performance test, the traditional method relies on experience to set parameters, has poor adaptability to different PCB boards, is prone to result deviation or damage to the board material, and only focuses on the maximum stress or strain of the PCB board, ignoring key mechanical parameters, and it is impossible to better evaluate the anti-bending performance of the PCB board.

[0005] In view of this, the present invention proposes a method for testing the performance of a PCB board of an AI server to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention provides the following technical solution: A method for testing the performance of a PCB board of an AI server, including:

[0007] S1. Obtain the temperature data of the PCB board under the energized operating state and generate a temperature distribution image;

[0008] S2. Determine the hot spot area based on the temperature distribution image, obtain the regional features within the hot spot area, and determine the heating test parameters based on the regional features;

[0009] S3. Conduct a heating test on the PCB board based on the heating test parameters, obtain the thermal resistance at each position in the hot spot area during the heating test, and obtain the thermal resistance characteristic curve; analyze the thermal resistance characteristic curve to obtain the thermal performance grade of the PCB board;

[0010] S4. Obtain the basic parameters of the PCB board, and determine the bending rate of the bending test based on the basic parameters; during the bending test, adjust the applied bending force in real time based on the bending rate, and record the test data and the real-time stress and real-time strain of the PCB board during the bending test;

[0011] S5. Preset the stress threshold and strain threshold; when the real-time stress of the PCB board reaches the stress threshold or the real-time strain reaches the strain threshold, stop the bending test, and analyze the test data to obtain the anti-bending performance grade of the PCB board;

[0012] S6. Classify the PCB board into first-class products, second-class products, third-class products or unqualified based on the thermal performance grade and anti-bending performance grade of the PCB board.

[0013] Further, the process of generating the temperature distribution image includes:

[0014] Construct a reference orthogonal copper wire grid that matches the size and structure of the PCB board based on the PCB design file;

[0015] Optimize the reference orthogonal copper wire grid through a fractal iteration algorithm to generate a fractal grid, and embed the fractal grid into the internal non-high-speed signal layer of the PCB board during the PCB manufacturing process; the intersections of the copper wires of the fractal grid are data acquisition nodes, and all data acquisition nodes form a temperature sensor network;

[0016] Collect the temperature data of the PCB board in the powered-on operating state through the temperature sensor network;

[0017] Based on the PCB design file, generate a heat source distribution map of the PCB board through thermal simulation software; construct a steady-state heat conduction equation of the PCB board based on the heat source distribution map;

[0018] Construct a physics-informed neural network based on the steady-state heat conduction equation of the PCB board; use the physics-informed neural network to reconstruct the collected temperature data to generate a temperature distribution image.

[0019] Further, the PCB design file includes a schematic file, a PCB layout file, a gerber file, and a drill file; the gerber file includes the material, thickness, number of layers, and size specifications of the PCB board;

[0020] In the fractal grid, a cross-bridging structure is adopted at the intersection of two copper wires, and the two copper wires are isolated by a dielectric layer.

[0021] Further, the process of collecting temperature data of the PCB board in the powered-on operating state through the temperature sensor network includes:

[0022] Place the PCB board in an incubator with a preset reference temperature, and use an impedance analyzer to obtain the reference resistance of each data acquisition node in the temperature sensor network;

[0023] Apply a rated working voltage and working current to the PCB board to make the temperature of the PCB board reach a steady state under working conditions; apply an excitation current to each copper wire in the temperature sensor network, and obtain the excitation voltage of the data acquisition node through the four-wire measurement method, and calculate the resistance of the data acquisition node;

[0024] Based on the reference resistance of the data acquisition node, the temperature coefficient of resistance of copper, and the resistance of the data acquisition node, obtain the temperature of the data acquisition node, and calculate the temperature change amount between the temperature of the data acquisition node and the reference temperature;

[0025] Construct a thermal coupling matrix matching the temperature sensor network through finite element simulation, and calculate the pseudo-inverse matrix of the thermal coupling matrix; use the pseudo-inverse matrix to solve the temperature change amount to obtain the initial temperature change amount between the temperature of the data acquisition node and the reference temperature, and then obtain the temperature of the data acquisition node; the temperatures of all data acquisition nodes constitute the temperature data of the PCB board in the powered-on operating state.

[0026] Further, the process of generating the temperature distribution image of the PCB board includes:

[0027] Perform normalization processing on the temperature data to generate an initial temperature field;

[0028] Incorporate the steady-state heat conduction equation into the loss function corresponding to the hidden layer of a pre-constructed neural network to obtain a reference physics-informed neural network;

[0029] Train the reference physics-informed neural network on an Adam or L-BFGS optimizer until the output of the reference physics-informed neural network meets the preset output requirements to obtain a physics-informed neural network;

[0030] Use the initial temperature field as the input of the physics-informed neural network to generate the temperature distribution image of the PCB board.

[0031] Further, the method of obtaining the thermal resistance characteristic curve includes:

[0032] Convert the temperature distribution image into a grayscale image by the weighted average method; extract hot spots from the grayscale image through clustering algorithms, edge detection algorithms, and connected component analysis to obtain hot spot regions;

[0033] Extract the regional features of the hot spot regions, where the regional features include geometric shapes and hot spot temperatures; the geometric shapes include the area, contour, and aspect ratio of the hot spot regions;

[0034] Define the heating method as follows: when the area of the hot spot region is greater than a preset maximum area threshold, use a distributed heating array to heat the PCB board;

[0035] When the aspect ratio of the hot spot region is greater than a preset aspect ratio, heat the PCB board through a bar heater, and the heaters are arranged along the long axis direction of the hot spot region;

[0036] Based on Fourier's law, the material and thickness of the PCB board, construct a heating power formula, and use the heating power formula to calculate the required heating power of the heater; the heating method and heating power together constitute the heating test parameters;

[0037] Conduct a heating test on the PCB board based on the heating test parameters; during the heating process, collect the temperature of the heating region in real time through an infrared thermal imager, and dynamically adjust the current heating power and the duration of the current heating power through the PID algorithm;

[0038] During the heating test process, collect the thermal resistance of each hot spot region in real time through resistance sensors distributed on the PCB board, and plot the thermal resistance change curve according to the time series; when the temperature change rate of the PCB board is less than a preset change rate threshold, plot the thermal resistance distribution curves at different positions; the thermal resistance change curve and the thermal resistance distribution curve together constitute the thermal resistance characteristic curve.

[0039] Furthermore, the method for obtaining the thermal performance level of the PCB board includes:

[0040] Extract the mathematical features of the thermal resistance change curve, where the mathematical features include the slope, inflection point, and smoothness of the curve; extract the spatio-temporal features of the thermal resistance distribution curve, where the spatio-temporal features include regional thermal resistance gradient and change rate;

[0041] Assign weights to different positions in the hot spot region through thermal sensitivity analysis, and the weights of all positions constitute a spatial weight matrix; construct a time decay factor based on the mathematical features and spatial features;

[0042] Based on the spatial weight matrix, time decay factor, and preset feature weights, construct a thermal performance comprehensive scoring formula; use the thermal performance comprehensive scoring formula to calculate the score of the hot spot region of the PCB board; based on the average value of the scores of all hot spot regions of the PCB board and the preset thermal performance scoring standard, obtain the thermal performance level of the PCB board.

[0043] Furthermore, the method for determining the initial bending force and bending rate of the bending test includes:

[0044] Construct a material database and a test database based on the basic parameters of the PCB board in the historical bending test and their corresponding bending test data; the basic parameters of the PCB board include material, thickness, number of layers, and size specifications;

[0045] The material database contains the material categories of the PCB board and the mechanical parameters corresponding to different material categories. The mechanical parameters include elastic modulus, Poisson's ratio, density, and coefficient of thermal expansion; the test database contains the historical test data of PCB boards with different thicknesses, numbers of layers, and size specifications;

[0046] Based on the data in the material database and the test database, use finite element simulation software to construct a 3D structure model of the PCB, and simulate the bending test scenario through the finite element simulation software to obtain simulation results;

[0047] Based on the simulation results and test requirements, identify the critical load through the failure criterion and then inversely deduce the initial bending force and bending rate required for the test to obtain the simulation data of the PCB board;

[0048] Use the orthogonal experimental design and scattering approximation algorithm to generate simulation data covering all parameters, and construct a simulation database based on the simulation data covering all parameters;

[0049] Extract the historical test data from the test database, and extract the corresponding simulation data from the simulation database; align the simulation data with the historical test data to obtain a data set, and perform mean imputation on the small amount of missing values in the data set;

[0050] Perform binary one-hot encoding on the material data in the data set, and perform normalization on the thickness and size data in the data set to obtain a test data set; randomly divide the test data set into a training set and a validation set according to a preset ratio;

[0051] Based on the training set and validation set data, use the method of 5-fold cross-validation combined with grid search to optimize the hyperparameters of the XGBoost model within the preset hyperparameter range; select the hyperparameter combination with the smallest mean absolute error and coefficient of determination on the validation set as the optimal parameters of the XGBoost model; when the proportion of new data in the data set reaches the preset requirement, trigger the incremental training of the model to dynamically update the model parameters;

[0052] Package the trained XGBoost model as an API service and integrate it into the bending test system; based on the input basic parameters of the PCB board, the bending test system gives the corresponding initial bending force and bending rate by calling the API service.

[0053] Further, the method for obtaining the anti-bending performance grade of the PCB board includes:

[0054] Perform a bending test on the PCB board based on the initial bending force and bending rate; during the bending test, use an adaptive PID control algorithm to drive the servo motor, take the error signal generated by the real-time feedback of the bending rate as the control input, and control the magnitude of the bending force by adjusting the rotation speed of the servo motor, thereby controlling the bending rate;

[0055] During the bending test, record relevant test data and simultaneously monitor the real-time stress and real-time strain of the PCB board; the test data includes data in four aspects: bending force, displacement, stress, and strain;

[0056] When the real-time stress of the PCB board reaches the preset stress threshold or the real-time strain reaches the preset strain threshold, stop the bending test; calculate the bending force per unit displacement, elastic modulus, and yield stress from the test data;

[0057] Construct a comprehensive anti-bending performance score formula based on the bending force per unit displacement, elastic modulus, and yield stress; calculate the score of the PCB board using the comprehensive anti-bending performance score formula; obtain the anti-bending performance grade of the PCB board based on the score of the PCB board and the preset thermal performance scoring standard.

[0058] Technical effects and advantages of the performance testing method for the AI server PCB board of the present invention:

[0059] In terms of the thermal performance testing of the PCB board, the present invention constructs a temperature sensor network with large-area coverage through a fractal grid, combines a cross-bridging structure to reduce parasitic interference, realizes high-precision temperature monitoring, and effectively solves the problems of low coverage rate of the traditional resistance method and high cost of infrared thermal imagers; uses a physics-informed neural network, integrates the steady-state heat conduction equation and the pseudo-inverse solution technology of the thermal coupling matrix, optimizes the temperature field reconstruction, controls the error within an extremely small range, reaches an extremely high level of resolution, and realizes the quantitative analysis of spatio-temporal characteristics such as thermal resistance gradients; adaptively selects the heating method based on the geometric characteristics of the hot spot area, combines PID closed-loop control and a high-precision infrared thermal imager to achieve extremely precise temperature control; constructs a comprehensive thermal performance scoring model based on the characteristics of the thermal resistance characteristic curve for quantitative grading, significantly reducing the deviation caused by human experience.

[0060] In terms of the anti-bending performance test, through the collaborative optimization of finite element simulation and the XGBoost model, combined with historical test data, it can intelligently predict the initial bending force and rate, significantly reducing the number of PCB boards required to determine the optimal test parameters; the adaptive PID algorithm is used to drive the servo motor, and high-precision sensors are used to achieve rapid dynamic adjustment of the bending force and rate, ensuring test safety and reliable data; by integrating core indicators such as elastic modulus, the weight coefficients are dynamically configured according to industry standards and the design requirements of the PCB board for multi-dimensional performance evaluation, and the anti-bending performance of the PCB board is accurately evaluated. Brief Description of the Drawings

[0061] Figure 1 Schematic diagram of a performance test method for an AI server PCB board of the present invention. Detailed Embodiments

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0063] Embodiment 1

[0064] Please refer to Figure 1 As shown, the performance test method for an AI server PCB board in this embodiment includes:

[0065] S1. Obtain the temperature data of the PCB board under the energized operating state and generate a temperature distribution image;

[0066] S2. Determine the hot spot area based on the temperature distribution image, obtain the regional characteristics within the hot spot area, and determine the heating test parameters based on the regional characteristics;

[0067] S3. Perform a heating test on the PCB board based on the heating test parameters, obtain the resistance at each position in the hot spot area during the heating test, and obtain a thermal resistance characteristic curve; analyze the thermal resistance characteristic curve to obtain the thermal performance grade of the PCB board;

[0068] S4. Obtain the basic parameters of the PCB board, determine the bending rate of the bending test based on the basic parameters; during the bending test, adjust the applied bending force in real time based on the bending rate, and record the test data and the real-time stress and real-time strain of the PCB board during the bending test;

[0069] S5. Preset stress threshold and strain threshold; when the real-time stress of the PCB board reaches the stress threshold or the real-time strain reaches the strain threshold, stop the bending test, analyze the test data, and obtain the anti-bending performance level of the PCB board.

[0070] S6. Classify the PCB boards into first-class products, second-class products, third-class products or unqualified ones based on the thermal performance level and anti-bending performance level of the PCB boards.

[0071] In the field of PCB board temperature acquisition, traditional methods have many limitations. Most traditional methods rely on high-precision infrared thermal imagers, but due to various factors, it is difficult to comprehensively and accurately obtain the temperature distribution of the PCB board, resulting in the inability to accurately reflect the hot spot area.

[0072] In addition to high-precision infrared thermal imagers, using resistors to calculate temperature is also a commonly used method in engineering; however, because the PCB board itself is a highly integrated product, it is impossible to reserve enough positions on the PCB board for placing resistors, and fewer resistors cannot comprehensively reflect the temperature of the PCB board, ultimately resulting in distortion of the temperature field reconstruction.

[0073] In addition, traditional heating tests use a fixed heating power and cannot be dynamically adjusted according to the characteristics of hot spots, which may cause problems such as local overheating or insufficient testing. In the thermal performance evaluation link, excessive reliance on a single index such as the highest temperature and the lack of quantitative analysis of spatio-temporal characteristics such as thermal resistance gradient and change rate result in strong subjectivity in rating and make it difficult to objectively and comprehensively evaluate the thermal performance of the PCB board.

[0074] To avoid the above problems, during the PCB manufacturing process, based on the PCB design file, design a reference orthogonal copper wire grid on the signal layer (usually the 3rd layer) inside the PCB board, and use the Hilbert fractal iteration algorithm to optimize the reference orthogonal copper wire grid to obtain a fractal grid.

[0075] The PCB design file includes a schematic file, a PCB layout file, a gerber file, and a drilling file; the gerber file includes the material, thickness, number of layers, and size specifications of the PCB board.

[0076] The line width of the fractal grid is 0.1 mm, and the line spacing is 0.5 mm; the intersection of the copper wires in the fractal grid is the data acquisition node, and all data acquisition nodes form a temperature sensor network, covering more than 98% of the PCB area; to reduce the cross-interference of wiring and the impact of parasitic capacitance on high-frequency impedance measurement, the data acquisition nodes at the intersection of copper wires in the fractal grid adopt a cross-bridging structure, and two copper wires are isolated by a 0.05 mm thick FR-4 dielectric layer to ensure that the insulation resistance is greater than 10 GΩ.

[0077] Place the PCB in an incubator at 25°C. When the temperature difference between the PCB and the incubator is 0.5°C, the temperature of the PCB reaches environmental steady state. Measure the resistance of the temperature sensor network at environmental steady state using an impedance analyzer to obtain the reference resistance R0 of each data acquisition node.

[0078] Apply the rated working current to the PCB to make it enter the working mode until it reaches steady state (the temperature change rate is less than 1°C / s). Apply the excitation current I m (usually 10 mA) to each copper wire using the Kelvin connection method to avoid errors caused by wire resistance and contact resistance.

[0079] Obtain the excitation voltage V of all data acquisition nodes through the four-wire measurement method. m Based on the excitation current and the excitation voltage of the data acquisition node, calculate the resistance R of the data acquisition node through Ohm's law.

[0080] Based on the reference resistance of the data acquisition node, the temperature coefficient of resistance of copper, and the resistance of the data acquisition node, calculate the temperature change ΔT of the data acquisition node. The calculation formula for the temperature change of the data acquisition node is: where α Cu is the temperature coefficient of resistance of copper, with a value of 0.0039 / °C; R (i) is the resistance of the i-th data acquisition node; is the reference resistance of the i-th data acquisition node measured at 25°C of the PCB.

[0081] Construct a thermal coupling matrix [H] that matches the temperature sensor network through finite element simulation. This matrix is used to describe the heat diffusion between data acquisition nodes. Among them, the matrix element where r ij is the Euclidean distance between the i-th intersecting data acquisition node and the j-th intersecting data acquisition node, and λ is the thermal conductivity of the material; i and j are both the intersection sequence numbers of the data acquisition nodes;

[0082] Calculate the pseudo-inverse matrix [H] + of the thermal coupling matrix, and use the pseudo-inverse matrix [H] + to solve the temperature change ΔT of the data acquisition node to obtain the initial temperature change ΔT initial of the data acquisition node, and then obtain the temperature T of the data acquisition node. The formula is: T = ΔT initial + T0, where represents the vector composed of the initial temperature changes of all data acquisition nodes obtained by the solution; where It represents the vector composed of the temperature changes of all data acquisition nodes; T0 is the reference temperature of 25°C; the temperatures of all data acquisition nodes constitute the temperature data of the PCB board in the powered-on operating state.

[0083] Normalize the temperature data to generate the initial temperature field.

[0084] Based on the PCB design file, extract the component power consumption information of the PCB board through thermal simulation software to generate the heat source distribution map of the PCB board.

[0085] Construct the steady-state heat conduction equation of the PCB board based on the thermodynamic principle and the heat source distribution map of the PCB board; the formula of the steady-state heat conduction equation is where ρ is the density of the material of the PCB board, representing the mass of the PCB board per unit volume; c p is the specific heat capacity at constant pressure of the material of the PCB board, referring to the heat required for a unit mass of the PCB board to increase by a unit temperature under constant pressure conditions; t is the time, reflecting the change of temperature over time; T is the node temperature; k is the thermal conductivity of the material of the PCB board, reflecting the ability to conduct heat. is the temperature operator, representing the spatial change rate of the temperature field. is the spatial divergence of the heat flux density, representing the rate of net heat outflow per unit volume; Q represents the heat source power density of the PCB board, which is calculated by combining the heat source distribution map with the Joule heat formula.

[0086] Incorporate the steady-state heat conduction equation into the loss function corresponding to the hidden layer of the pre-constructed neural network to obtain the benchmark physics-informed neural network.

[0087] Train the benchmark physics-informed neural network on the Adam or L-BFGS optimizer until the output of the benchmark physics-informed neural network meets the preset requirements (resolution reaches 0.1 0.1mm / pixel) to obtain the physics-informed neural network.

[0088] Input the initial temperature field into the physics-informed neural network to obtain the temperature distribution image of the PCB board with a resolution of 0.1mm / pixel.

[0089] It should be noted that the network architecture of the physics-informed neural network is divided into three parts: the input fusion layer, the physical constraint layer, and the super-resolution reconstruction layer.

[0090] The input fusion layer samples the input initial temperature field through the convolutional layer of the neural network to generate the temperature distribution feature map.

[0091] The physical constraint layer is the hidden layer of the neural network with the steady-state heat conduction equation added. The residual is obtained by calculating the difference between the predicted temperature and the steady-state heat conduction equation; and the calculated residual is used as part of the loss function, and the weights of each part in the loss function are corrected through backpropagation to make the output of the physics-informed neural network satisfy the conservation of energy.

[0092] The loss function is composed of weighted residual loss and boundary condition loss; the functional expressions of the residual loss and the boundary condition loss are both derived and solved by mathematical formulas, and the specific loss function formulas are not given here.

[0093] The image reconstruction layer gradually upsamples through transposed convolution to reconstruct the temperature distribution feature map into a temperature distribution image of 0.1 mm / pixel.

[0094] The temperature distribution image is converted into a grayscale image by the weighted average method.

[0095] The grayscale image is clustered by the K-means clustering algorithm, and similar temperature regions are grouped and clustered. If the temperature of the cluster center is higher than the preset temperature, it is a hot spot.

[0096] Edge detection algorithms such as Sobel and Canny are used to extract the boundaries of the hot spots, and the boundaries of the hot spots are determined through connected component analysis to obtain the hot spot regions.

[0097] The regional features of the hot spot regions are extracted. The regional features include geometric shapes and hot spot temperatures; among them, the geometric shapes include the area, contour, and aspect ratio of the hot spot regions, which are used to determine the heating range and heating method.

[0098] The heating method is defined as: when the area of the hot spot region is greater than the preset maximum area threshold, a distributed heating array is used to heat the PCB board.

[0099] When the aspect ratio of the hot spot region is greater than the preset aspect ratio, the PCB board is heated by a strip heater, and the heater is arranged along the long axis direction of the hot spot region.

[0100] The heating power formula is obtained by adding a heat capacity term to Fourier's law. The heating power formula is: where m is the mass of the hot spot region (calculated by area × thickness × density); c is the specific heat capacity of the PCB board material; t is the basic heating time of the target region (20 seconds); ΔT is the temperature difference between the target temperature and the current average temperature (°C); k is the thermal conductivity of the PCB board material; A is the hot spot area of the PCB board (m 2 ); d is the thickness of the PCB board, unit: m; substituting the material and thickness of the PCB board into the heating power function to obtain the heating power required by the heater.

[0101] Align the heater with the centroid of the hot spot area; when the distance between the centroid and the edge of the PCB is less than 5 mm, reduce the power on the edge side to 70% to avoid heat dissipation loss; if the distance between the centroids of multiple hot spots within the hot spot area is less than 20 mm, consider them as one heating area; the heating method and heating power together constitute the heating test parameters;

[0102] Conduct a heating test on the PCB based on the heating test parameters; during the heating test, collect the temperature field data of the heating area in real time through a high-precision infrared thermal imager (resolution ≤ 0.1 °C / pixel); during the heating process, adopt a PID closed-loop control strategy to dynamically adjust the heating power (adjustment range 0 - 300 W) and duration (accuracy 10 ms) based on real-time temperature feedback; PID control can effectively compensate for the thermal conduction lag effect through the proportional-integral-derivative algorithm, achieve the dynamic balance between heating power and heat dissipation, and finally make the PCB reach and stably maintain the target temperature state within the set time.

[0103] During the heating test, collect the thermal resistance of each hot spot area in real time through a resistance sensor, and synchronously record the timestamp and spatial coordinates; the resistance sensor uses a high-precision platinum thermal resistance (such as Pt100), whose resistance value has a linear relationship with temperature and can provide high-precision measurement within a wide temperature range; the signal output by the resistance sensor is converted into a standard voltage or current signal through a temperature transmitter;

[0104] To ensure the consistency of the output of the resistive sensor, it is necessary to perform three-level calibration on it, and the calibration steps are as follows:

[0105] 1. Establish a standard test environment with constant temperature (25 ± 0.1 °C) and constant pressure (5.000 ± 0.005 V);

[0106] 2. Synchronously collect 10 groups of step response data of the reference sensor and the sensor to be tested in the standard test environment;

[0107] 3. Construct a polynomial compensation model based on the step response data using the least squares method;

[0108] 4. Calculate the temperature drift coefficient of each channel using the compensation model and generate a calibration coefficient table.

[0109] Calculate the temperature data from the standard voltage or current signal collected by the resistance sensor, and convert the temperature data into thermal resistance values through the thermal resistance formula; plot the thermal resistance change curves (time dimension) of each position in the hot spot area according to the time series; when the temperature change rate of the PCB is less than 0.5 °C / s, plot the thermal resistance distribution curves at different positions; the thermal resistance characteristic curves include thermal resistance change curves and thermal resistance distribution curves;

[0110] The thermal resistance characteristic curves include the thermal resistance change curve and the thermal resistance distribution curve; these curves can clearly show the changing trends of thermal resistance over time and space, providing a basis for subsequent analysis.

[0111] Extract the mathematical and spatio-temporal characteristics of the thermal resistance characteristic curves; the mathematical characteristics include the slope, inflection points, and smoothness of the curves; the spatio-temporal characteristics include the regional thermal resistance gradient and the rate of change of thermal resistance over time;

[0112] Calculate the derivatives of each section of the thermal resistance change curve to obtain the corresponding slopes, which reflect the rate of change of thermal resistance; for example, a sharp increase in slope may indicate local overheating, which may be due to a decrease in the thermal conductivity of the material or excessive heat accumulation;

[0113] Precisely identify the inflection points of the curve by calculating the second derivative of the thermal resistance change curve or using a piecewise fitting algorithm; these inflection points mean that the thermal performance of the PCB enters a mutation stage, indicating that a phase change of the material occurs or the heat transfer path changes;

[0114] Quantify the smoothness of the thermal resistance change curve using statistical methods such as root mean square error; if the thermal resistance change curve shows abnormal fluctuations (the root mean square error exceeds the pre-set range), it indicates a risk of thermal runaway;

[0115] Based on the thermal resistance distribution curve, extract the thermal resistance values corresponding to different positions within the hot spot area; for these thermal resistance values, calculate the thermal resistance gradients in the x and y directions respectively; vectorially sum the thermal resistance gradients in these two directions, and the resulting value is the thermal resistance gradient at that position, and the direction of this thermal resistance gradient is determined by the vector synthesis rule;

[0116] Fit the thermal resistance gradients at each position within the hot spot area using radial basis functions (such as Gaussian functions, thin plate spline functions, etc.) to obtain the regional thermal resistance gradient of the hot spot area;

[0117] Based on the thermal resistance distribution curve, determine the time interval during which the thermal resistance within the hot spot area changes significantly. For the sequence data of the thermal resistance changing over time at each position within the hot spot area during this time interval, accurately calculate the rate of change of thermal resistance over time through derivative operations.

[0118] Calculate the characteristic parameters by calculating the curve characteristics and spatio-temporal characteristics of the hot spot area. The characteristic parameters include the absolute value of the slope, inflection point density, smoothness, modulus value of the regional thermal resistance gradient, and rate of change;

[0119] Assign weights to different positions within the hot spot area through thermal sensitivity analysis. The weights of all positions form a spatial weight matrix; the spatial weight matrix where d is the Euclidean distance from the current position to the heat source; d0 is the distance from the current position to the critical area (such as the heat dissipation boundary position); γ is the thermal attenuation coefficient (with a value ranging from 0.1 to 0.5);

[0120] Construct a time decay factor where α is the decay coefficient, with a value of 0.9 (usually taken from 0.8 to 0.95); t i is the time segment during which the abnormal feature persists (unit: seconds / minutes); n is the number of abnormal events that occur; n max is the preset maximum number of abnormal events that occur, with a value of 100, and its size can be adjusted according to the number of abnormal events that occur;.

[0121] It should be noted that the feature corresponding to when the feature parameter exceeds the preset parameter threshold is the abnormal feature; for example, the threshold of the slope is 2Ω / min, the threshold of the inflection point number is 3 / min, the threshold of the smoothness is 0.95, the threshold of the modulus of the thermal resistance gradient is 20℃ / 10mm, and the threshold of the change rate is 1℃ / s;

[0122] The abnormal feature shows instantaneous or short-term fluctuations in the data, with a short duration (in seconds); the abnormal event is caused by the persistence or combination of at least one abnormal feature, with a long duration (in minutes);

[0123] Construct a comprehensive thermal performance scoring formula based on the spatial weight matrix and the time decay factor: where μ(x j ) is the trapezoidal membership function of each feature, used for feature fusion; j is the feature index, and 1 to 5 correspond to the absolute value of the slope, the inflection point density, the smoothness, the modulus of the regional thermal resistance gradient, and the change rate respectively; ω j is the preset weight of the feature corresponding to the feature index, and the sum of the weights is 1; n max is the preset maximum number of abnormal events that occur.

[0124] Substitute the obtained feature parameters into the scoring formula after Z-score standardization processing to obtain the score of this hot spot area; calculate the scores of all hot spot areas on the PCB board, and then calculate the average score of the PCB board; if the average score is greater than 85, determine that the thermal performance level of the PCB board is "excellent"; if the average score is between 60 and 85, determine its thermal performance level as "good"; if the average score is less than 60, then determine the thermal performance level as "poor";

[0125] In the above thermal performance test process of the PCB board, a complete set of PCB thermal performance evaluation systems is constructed through multi-dimensional technological innovation: in the temperature monitoring link, the orthogonal copper wire grid uses fractal wiring to achieve temperature monitoring of a large area of the PCB board, with a very high coverage rate; the physics-informed neural network is used to fuse material properties and heat source distribution, and combined with the pseudo-inverse solution technology of the thermal coupling matrix, the error between the generated temperature distribution image of the PCB board and the actual temperature is controlled within a very small range. In terms of thermal performance testing, the PID closed-loop algorithm is used to dynamically regulate the heating power, combined with a high-precision infrared thermal imager, to achieve extremely precise temperature control accuracy. At the anti-interference design level, through the cross-bridging structure, the influence of parasitic parameters is effectively reduced. The characteristics of the PCB board generate synergistic effects through the weighted fusion mechanism, and finally a thermal performance scoring model is constructed, which can quantitatively reflect key indicators such as the thermal conduction efficiency and the uniformity of thermal resistance distribution of the PCB board. Its grading standard is highly consistent with the requirements of production line quality inspection, greatly reducing the human experience deviation in the traditional method, and providing a standardized solution for the industrial-grade PCB thermal performance evaluation.

[0126] The traditional bending test parameters (force or rate) rely on empirical settings and cannot give appropriate bending forces and bending rates for PCB boards with different materials, thicknesses, and sizes. Inappropriate bending forces and bending rates are likely to cause deviations in test results or damage to the PCB board, resulting in losses.

[0127] When evaluating the bending resistance performance of the PCB board, only the maximum stress or strain data of the PCB board are concerned, ignoring key parameters such as elastic modulus and yield stress, and the performance of the PCB board cannot be comprehensively evaluated.

[0128] To avoid the above problems, in the bending test, it is necessary to determine the initial bending force and rate based on the basic parameters of the PCB board, adjust the bending force in real time according to the bending rate during the test, automatically stop the test when the stress or strain reaches the critical value, and analyze the bending test data, and then obtain the bending resistance performance level; the process is as follows:

[0129] Construct a material database and a test database based on the basic parameters of the PCB board in the historical bending test and their corresponding bending test data; the basic parameters of the PCB board include material, thickness, number of layers, and size specifications;

[0130] The material database contains PCB material categories (FR-4, polyimide PI, composite base CEM series, etc.) and their corresponding mechanical parameters. The mechanical parameters include elastic modulus, Poisson's ratio, density, and coefficient of thermal expansion;

[0131] The test database contains historical test data of PCB boards with different thicknesses (1.0 mm to 3.0 mm), number of layers (single-layer to multi-layer boards), and size specifications (length, width, and aspect ratio).

[0132] Construct a 3D structural model of the PCB through finite element simulation software; take the material type of the PCB and its corresponding mechanical parameters, the thickness of the PCB board, the number of layers of the PCB board, the size specifications of the PCB board and boundary constraints (such as the full degree of freedom constraint of the fixed end) as input variables; simulate typical test scenarios such as four-point bending or cantilever beam bending through the finite element analysis method, and output simulation results including key parameters such as stress distribution cloud diagram, maximum deformation, bending stiffness, etc.; identify the critical load (such as the critical load when the maximum principal stress exceeds the yield strength of the material) through the failure criterion based on the simulation results and test requirements, and then infer the optimal initial bending force and bending rate required for the test to obtain the simulation data of the PCB board;

[0133] Through orthogonal experimental design and scattering approximation algorithm, simulation data covering all parameters (PCB boards of all materials, thicknesses, number of layers and size specifications) are generated, and a simulation database is built based on the simulation data.

[0134] Extract historical test data from the test database, and extract corresponding simulation data from the simulation database; align the simulation data and historical test data to obtain a data set, remove data that obviously deviates from physical laws (such as negative thickness or excessive bending force), and interpolate a small number of missing values using the mean of the same material;

[0135] The processed data is converted into a standard format as a data set for subsequent model training; for example, the sample data entry in the data set is: {material: FR-4, thickness: 1.6mm, number of layers: 8, size: 150mm×100mm, aspect ratio: 1.5}→{initial bending force: 12.5N, bending rate: 0.8mm / s}.

[0136] The PCB material in the data set is converted into binary one-hot encoding (for example, the FR-4 material encoding is 00010), and the thickness and size data in the data set are normalized to obtain the test data set; the test data set is randomly divided into a training set and a validation set in a ratio of 8:2.

[0137] Preset the hyperparameters of the XGBoost model. The preset hyperparameters are set as follows: number of trees: 30-70; tree depth: 3-6; learning rate: 0.05-0.2; regularization parameter: 0.1-1.0;

[0138] The XGBoost model is trained based on the data in the training set. During the training process, 5-fold cross-validation is used to evaluate the stability of the model to avoid the random interference caused by a single partition. At the same time, an early stopping mechanism is introduced to continuously monitor the loss of the validation set. When the loss of the validation set does not decrease after 10 consecutive rounds of training, the training process is terminated.

[0139] Traverse the hyperparameter combinations based on the grid search idea, and select the combination with the smallest mean absolute error and coefficient of determination on the validation set as the optimal parameters; when the newly added data accumulates to the preset requirement (the proportion reaches 10% of the dataset), trigger incremental training.

[0140] Package the trained model as an API service and embed it into the bending test system; the bending test system adopts an intelligent parameter acquisition mechanism, which supports automatically obtaining the basic parameters of the PCB board by manually inputting through the user interface or scanning the material code (QR code or barcode).

[0141] Based on the obtained PCB parameters, generate test parameters (initial bending force and bending rate) in real time and synchronize them to the bending test system;

[0142] During the bending test, adopt an adaptive PID control algorithm to drive the servo motor, use the real-time feedback error signal as the control input, and realize the precise control of the bending force by dynamically adjusting the rotation speed of the servo motor, and then realize the precise control of the bending rate.

[0143] When the bending rate deviates from the set value by ±10%, the bending test system will automatically trigger the dynamic adjustment mechanism: if the bending rate exceeds 20% of the set value, the bending test system will immediately trigger an emergency brake to stop the bending test to prevent damage to the PCB board;

[0144] The bending test adopts a real-time operating system to achieve high-frequency data acquisition exceeding 1kHz and a PID operation cycle not greater than 1ms, ensuring that the closed-loop response delay is less than 5ms and having a millisecond-level dynamic adjustment ability.

[0145] It should be noted that the bending test system includes four high-precision sensors:

[0146] A force sensor (range 0-200N, accuracy ±0.1%) is installed at the end of the force application arm to measure the bending force in real time;

[0147] The laser displacement sensor captures deformation displacements at the 0.01mm level with a sampling rate not less than 1kHz through the triangulation reflection method;

[0148] The laser velocimeter collects the actual bending rate in real time and compares it with the given bending rate to generate an error signal;

[0149] The strain gauge array is attached to the stress concentration area of the PCB (such as the edge of the PCB board or around the hole positions), and is converted into an electrical signal through a Wheatstone bridge to obtain strain data.

[0150] Set the stress threshold and strain threshold for the bending test, where:

[0151] Stress threshold σ t : σt = Yield strength of the material × Safety factor; The yield strength of the material is provided by the supplier; The value range of the safety factor is [0.7, 0.9]; For example, if the yield strength of material FR-4 is 400 MPa and the safety factor is 0.9, then the stress threshold is 360 MPa;

[0152] Strain threshold ε t : Based on the bending strain limit in IPC-6012 standard (generally 1.5% to 2.0%).

[0153] During the bending test, based on the real-time collected bending force and the geometric parameters of the PCB board, calculate the bending stress in real time; The stress calculation formula is: where F is the real-time collected bending force, unit: N; L is the distance between the support points of the test fixture, unit: m; d is the thickness of the PCB board, unit: m; h is the width of the PCB board, unit: m.

[0154] When the real-time stress of the PCB board reaches the stress threshold or the real-time strain reaches the strain threshold, immediately stop the bending test and record the bending force F at this time E , displacement X, stress and strain data; Based on the bending force F E and displacement X calculate the bending force per unit displacement

[0155] Based on the test data, draw the stress-strain curve; Select the initial linear segment of the stress-strain curve (usually the strain range of 0 - 0.2%); Perform least squares fitting on the data points of the initial linear segment to obtain the elastic modulus E;

[0156] Calculate the second derivative of the stress-strain curve; Find the first zero point where the second derivative changes from positive to negative, corresponding to the starting point of yield; Perform cubic polynomial fitting on the local data (0.05% strain before and after) of this zero point to obtain the yield stress σ y ;

[0157] Based on the PCB board design requirements and industry standards, assign different weights to the elastic modulus, yield stress and bending force per unit displacement, and calculate the comprehensive score of the anti-bending performance; The comprehensive score calculation formula of the anti-bending performance is: where E is the elastic modulus obtained by least squares fitting; E std is the theoretical value of the elastic modulus; σ y is the yield stress obtained by cubic polynomial fitting; σ t is the preset stress threshold; F X is the bending force per unit displacement; F std is the theoretical value of the bending force per unit displacement; ω1, ω2 and ω3 are the preset weight coefficients based on the design requirements and industry standards, and the sum of the three is 1.

[0158] Based on the comprehensive score of the bending resistance performance of the PCB board, its bending resistance performance can be divided into four grades: excellent, good, qualified, and unqualified. The specific evaluation criteria are as follows: a score of 90 or above is excellent; between 75 (inclusive) and 90 is good; between 60 (inclusive) and 75 is qualified; and a score below 60 is unqualified.

[0159] Based on the thermal performance grade and bending resistance performance grade of the PCB board, a comprehensive judgment is made on the PCB board:

[0160] If the thermal performance of the PCB board is "poor" or the bending resistance performance is "unqualified", it is judged as a non-conforming product:

[0161] If the thermal performance of the PCB board is "excellent" and the bending resistance performance is "excellent", it is judged as a first-class product;

[0162] If the thermal performance of the PCB board is "excellent" and the bending resistance performance is "good", or the thermal performance score is "good" and the bending resistance performance is "excellent" or "good", it is judged as a second-class product;

[0163] If the thermal performance of the PCB board is "excellent" and the bending resistance performance is "qualified", or the thermal performance score is "good" and the bending resistance performance is "qualified", it is judged as a third-class product.

[0164] In the above-mentioned bending resistance test process of the PCB board, through the efficient collaboration of finite element simulation and the XGBoost prediction model, a dynamic prediction system for bending test parameters is built. Compared with the traditional empirical setting method, the demand for test samples is greatly reduced. The incremental training mechanism relies on real-time data feedback to continuously optimize the model parameters, and shows strong scalability and adaptability in bending test scenarios for PCB boards of different specifications and materials. The performance evaluation system integrates three core indicators: elastic modulus, yield stress, and bending force per unit displacement. Its weight coefficients can be flexibly and dynamically configured according to various industry standards such as IPC-6012 and JEDEC. By establishing a four-level evaluation system of "excellent, good, qualified, unqualified", a clear corresponding relationship is established between the quantitative grading of bending resistance performance and product reliability, which not only ensures the detection accuracy but also greatly improves the quality inspection efficiency and effectively reduces the risk of product failure caused by mechanical performance defects.

[0165] In this embodiment, in terms of the thermal performance test of the PCB board, a temperature sensor network covering a large area is constructed through a fractal grid, and the cross-bridging structure is combined to reduce parasitic interference, realizing high-precision temperature monitoring, and effectively solving the problems of low coverage rate of the traditional resistance method and high cost of the infrared thermal imager; a physics-informed neural network is adopted to fuse the steady-state heat conduction equation and the pseudo-inverse solution technology of the thermal coupling matrix to optimize the temperature field reconstruction, control the error within an extremely small range, achieve an extremely high level of resolution, and realize the quantitative analysis of spatio-temporal characteristics such as thermal resistance gradient; the heating method is adaptively selected based on the geometric characteristics of the hot spot area, combined with PID closed-loop control and a high-precision infrared thermal imager, to achieve extremely precise temperature control; a comprehensive thermal performance scoring model is constructed based on the characteristics of the thermal resistance characteristic curve for quantitative grading, greatly reducing the human experience deviation. In terms of the anti-bending performance test, through the collaborative optimization of finite element simulation and the XGBoost model, combined with historical test data, it can intelligently predict the initial bending force and rate, greatly reducing the number of PCB boards required to determine the optimal test parameters; an adaptive PID algorithm is used to drive the servo motor, and a high-precision sensor is used to achieve rapid dynamic adjustment of the bending force and rate to ensure test safety and reliable data; the core indicators such as elastic modulus are fused, and the weight coefficients are dynamically configured according to industry standards and the design requirements of the PCB board for multi-dimensional performance evaluation to accurately evaluate the anti-bending performance of the PCB board.

[0166] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0167] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the presence of additional identical elements in the process, method, article or device including the said element.

[0168] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0169] In the description of the present invention, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0170] In the description of the present invention, "a number of" means one or more, and "a large number of" means two or more.

[0171] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0172] For the formulas in this specification, the dimensional quantities are removed and the numerical values are calculated. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold values in the formulas are set by those skilled in the art according to the actual situation.

[0173] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A performance testing method for an AI server PCB board, characterized in that, Including: S1. Obtain the temperature data of the PCB board under the energized operating state and generate a temperature distribution image; S2. Determine the hot spot area based on the temperature distribution image, obtain the regional characteristics within the hot spot area, and determine the heating test parameters based on the regional characteristics; S3. Conduct a heating test on the PCB board based on the heating test parameters, obtain the thermal resistance of each position in the hot spot area during the heating test, and obtain a thermal resistance characteristic curve; analyze the thermal resistance characteristic curve to obtain the thermal performance grade of the PCB board; S4. Obtain the basic parameters of the PCB board and determine the bending rate of the bending test based on the basic parameters; During the bending test, adjust the applied bending force in real time based on the bending rate, and record the test data during the bending test as well as the real-time stress and real-time strain of the PCB board; S5. Preset a stress threshold and a strain threshold; When the real-time stress of the PCB board reaches the stress threshold or the real-time strain reaches the strain threshold, stop the bending test, and analyze the test data to obtain the anti-bending performance grade of the PCB board; S6. Classify the PCB board into first-class products, second-class products, third-class products or unqualified based on the thermal performance grade and anti-bending performance grade of the PCB board.

2. The performance testing method of an AI server PCB board according to claim 1, characterized in that The process of generating the temperature distribution image includes: Construct a reference orthogonal copper wire grid that matches the size and structure of the PCB board based on the PCB design file; Optimize the reference orthogonal copper wire grid through a fractal iteration algorithm to generate a fractal grid, and embed the fractal grid into the internal non-high-speed signal layer of the PCB board during the PCB manufacturing process; the intersection of the copper wires of the fractal grid is a data acquisition node, and all data acquisition nodes form a temperature sensor network; Collect the temperature data of the PCB board under the energized operating state through the temperature sensor network; Based on the PCB design file, generate a heat source distribution map of the PCB board through thermal simulation software; construct a steady-state heat conduction equation of the PCB board based on the heat source distribution map; Construct a physics-informed neural network based on the steady-state heat conduction equation of the PCB board; use the physics-informed neural network to reconstruct the collected temperature data to generate a temperature distribution image.

3. The performance testing method of an AI server PCB board according to claim 2, characterized in that, The PCB design file includes a schematic file, a PCB layout file, a gerber file, and a drill file; the gerber file includes the material, thickness, number of layers, and size specifications of the PCB board; In the fractal grid, a cross-bridging structure is adopted at the intersection of two copper wires, and the two copper wires are isolated by a dielectric layer.

4. The performance testing method of an AI server PCB board according to claim 3, characterized in that The process of collecting the temperature data of the PCB board under the energized operating state through the temperature sensor network includes: Place the PCB board in a constant temperature oven with a preset reference temperature, and use an impedance analyzer to obtain the reference resistance of each data acquisition node in the temperature sensor network; Apply a rated working voltage and working current to the PCB board to make the temperature of the PCB board reach the working condition steady state; apply an excitation current to each copper wire in the temperature sensor network, and obtain the excitation voltage of the data acquisition node through the four-wire measurement method, and calculate the resistance of the data acquisition node; Obtain the temperature of the data acquisition node based on the reference resistance of the data acquisition node, the temperature coefficient of resistance of copper, and the resistance of the data acquisition node, and calculate the temperature change amount between the temperature of the data acquisition node and the reference temperature; Construct a thermal coupling matrix matching the temperature sensor network through finite element simulation, and calculate the pseudo-inverse matrix of the thermal coupling matrix; use the pseudo-inverse matrix to solve the temperature change amount to obtain the initial temperature change amount between the temperature of the data acquisition node and the reference temperature, and then obtain the temperature of the data acquisition node; the temperatures of all data acquisition nodes constitute the temperature data of the PCB board in the powered-on operating state.

5. The performance testing method of an AI server PCB board according to claim 4, characterized in that, The process of generating the temperature distribution image of the PCB board includes: Perform normalization processing on the temperature data to generate an initial temperature field; Incorporate the steady-state heat conduction equation into the loss function corresponding to the hidden layer of the pre-constructed neural network to obtain a reference physics-informed neural network; Train the reference physics-informed neural network on an Adam or L-BFGS optimizer until the output of the reference physics-informed neural network meets the preset output requirements to obtain a physics-informed neural network; Use the initial temperature field as the input of the physics-informed neural network to generate the temperature distribution image of the PCB board.

6. The performance testing method of an AI server PCB board according to claim 5, characterized in that, The method of obtaining the thermal resistance characteristic curve includes: Convert the temperature distribution image into a grayscale image by the weighted average method; extract hot spots from the grayscale image through clustering algorithms, edge detection algorithms, and connected component analysis to obtain hot spot regions; Extract the regional characteristics of the hot spot regions, and the regional characteristics include geometric shapes and hot spot temperatures; the geometric shapes include the area, contour, and aspect ratio of the hot spot regions; Define the heating method as: when the area of the hot spot region is greater than the preset maximum area threshold, use a distributed heating array to heat the PCB board; When the aspect ratio of the hot spot region is greater than the preset aspect ratio, heat the PCB board through a strip heater, and the heater is arranged along the long axis direction of the hot spot region; Based on Fourier's law, the material and thickness of the PCB board, construct a heating power formula, and use the heating power formula to calculate the required heating power of the heater; the heating method and the heating power together constitute the heating test parameters; Conduct a heating test on the PCB board based on the heating test parameters; during the heating process, use an infrared thermal imager to collect the temperature of the heating area in real time, and dynamically adjust the current heating power and the duration of the current heating power through a PID algorithm; During the heating test process, collect the thermal resistance of each hot spot region in real time through resistance sensors distributed on the PCB board, and plot the thermal resistance change curve according to the time series; when the temperature change rate of the PCB board is less than the preset change rate threshold, plot the thermal resistance distribution curves at different positions; the thermal resistance change curve and the thermal resistance distribution curve together constitute the thermal resistance characteristic curve.

7. A performance testing method for an AI server PCB board according to claim 6, characterized in that The method of obtaining the thermal performance level of the PCB board includes: Extract the mathematical characteristics of the thermal resistance change curve, and the mathematical characteristics include the slope, inflection point, and smoothness of the curve; extract the spatio-temporal characteristics of the thermal resistance distribution curve, and the spatio-temporal characteristics include regional thermal resistance gradient and change rate; Weights are assigned to different positions in the hot spot area through thermal sensitivity analysis, and the weights of all positions form a spatial weight matrix; a time decay factor is constructed based on mathematical features and spatial features; A comprehensive thermal performance scoring formula is constructed based on the spatial weight matrix, the time decay factor, and the preset feature weights; the hot spot area score of the PCB board is calculated using the comprehensive thermal performance scoring formula; the thermal performance grade of the PCB board is obtained based on the average value of the scores of all hot spot areas of the PCB board and the preset thermal performance scoring criteria.

8. The performance testing method of an AI server PCB board according to claim 7, characterized in that The methods for determining the initial bending force and bending rate of the bending test include: Construct a material database and a test database based on the basic parameters of the PCB board in the historical bending test and their corresponding bending test data; the basic parameters of the PCB board include material, thickness, number of layers, and size specifications; The material database contains the material categories of the PCB board and the mechanical parameters corresponding to different material categories. The mechanical parameters include elastic modulus, Poisson's ratio, density, and coefficient of thermal expansion; the test database contains the historical test data of PCB boards with different thicknesses, numbers of layers, and size specifications; Based on the data in the material database and the test database, use finite element simulation software to construct a 3D PCB structure model, and simulate the bending test scenario through the finite element simulation software to obtain the simulation results; Based on the simulation results and test requirements, identify the critical load through the failure criterion and then reverse deduce the initial bending force and bending rate required for the test to obtain the simulation data of the PCB board; Use the orthogonal experimental design and scattering approximation algorithm to generate simulation data covering all parameters, and construct a simulation database based on the simulation data covering all parameters; Extract the historical test data from the test database, and extract the corresponding simulation data from the simulation database; align the simulation data with the historical test data to obtain a data set, and perform mean imputation on the small amount of missing values in the data set; Perform binary one-hot encoding on the material data in the data set, and normalize the thickness and size data in the data set to obtain a test data set; randomly divide the test data set into a training set and a validation set according to a preset ratio; Based on the training set and validation set data, use the method of 5-fold cross-validation combined with grid search to optimize the hyperparameters of the XGBoost model within the preset hyperparameter range; select the hyperparameter combination with the smallest mean absolute error and determination coefficient on the validation set as the optimal parameters of the XGBoost model; when the proportion of new data in the data set reaches the preset requirement, trigger the incremental training of the model to dynamically update the model parameters; Package the trained XGBoost model as an API service and integrate it into the bending test system; based on the input basic parameters of the PCB board, the bending test system gives the corresponding initial bending force and bending rate by calling the API service.

9. The performance testing method of an AI server PCB board according to claim 8, characterized in that, The methods for obtaining the anti-bending performance grade of the PCB board include: Perform a bending test on the PCB board based on the initial bending force and bending rate; during the bending test, an adaptive PID control algorithm is used to drive the servo motor, and the error signal generated by the real-time feedback of the bending rate is used as the control input. By adjusting the speed of the servo motor, the magnitude of the bending force is controlled, and thus the bending rate is controlled. During the bending test, record the relevant test data, and at the same time monitor the real-time stress and real-time strain of the PCB board; the test data includes data on four aspects: bending force, displacement, stress, and strain. When the real-time stress of the PCB board reaches the preset stress threshold or the real-time strain reaches the preset strain threshold, stop the bending test; calculate the bending force per unit displacement, elastic modulus, and yield stress from the test data. Construct a comprehensive score formula for the bending resistance performance based on the bending force per unit displacement, elastic modulus, and yield stress; calculate the score of the PCB board using the comprehensive score formula for the bending resistance performance; obtain the bending resistance performance level of the PCB board based on the score of the PCB board and the preset bending resistance performance scoring standard.

Images